1 /*
2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
28 #include "gc_implementation/g1/heapRegion.hpp"
29 #include "gc_interface/collectedHeap.hpp"
30 #include "memory/barrierSet.hpp"
31 #include "memory/cardTableModRefBS.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/graphKit.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "runtime/deoptimization.hpp"
41 #include "runtime/sharedRuntime.hpp"
42
43 //----------------------------GraphKit-----------------------------------------
44 // Main utility constructor.
45 GraphKit::GraphKit(JVMState* jvms)
46 : Phase(Phase::Parser),
47 _env(C->env()),
48 _gvn(*C->initial_gvn())
49 {
50 _exceptions = jvms->map()->next_exception();
51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
52 set_jvms(jvms);
53 }
54
55 // Private constructor for parser.
56 GraphKit::GraphKit()
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn())
60 {
61 _exceptions = NULL;
62 set_map(NULL);
63 debug_only(_sp = -99);
64 debug_only(set_bci(-99));
65 }
66
67
68
69 //---------------------------clean_stack---------------------------------------
70 // Clear away rubbish from the stack area of the JVM state.
71 // This destroys any arguments that may be waiting on the stack.
72 void GraphKit::clean_stack(int from_sp) {
73 SafePointNode* map = this->map();
74 JVMState* jvms = this->jvms();
75 int stk_size = jvms->stk_size();
76 int stkoff = jvms->stkoff();
77 Node* top = this->top();
78 for (int i = from_sp; i < stk_size; i++) {
79 if (map->in(stkoff + i) != top) {
80 map->set_req(stkoff + i, top);
81 }
82 }
83 }
84
85
86 //--------------------------------sync_jvms-----------------------------------
87 // Make sure our current jvms agrees with our parse state.
88 JVMState* GraphKit::sync_jvms() const {
89 JVMState* jvms = this->jvms();
90 jvms->set_bci(bci()); // Record the new bci in the JVMState
91 jvms->set_sp(sp()); // Record the new sp in the JVMState
92 assert(jvms_in_sync(), "jvms is now in sync");
93 return jvms;
94 }
95
96 //--------------------------------sync_jvms_for_reexecute---------------------
97 // Make sure our current jvms agrees with our parse state. This version
98 // uses the reexecute_sp for reexecuting bytecodes.
99 JVMState* GraphKit::sync_jvms_for_reexecute() {
100 JVMState* jvms = this->jvms();
101 jvms->set_bci(bci()); // Record the new bci in the JVMState
102 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
103 return jvms;
104 }
105
106 #ifdef ASSERT
107 bool GraphKit::jvms_in_sync() const {
108 Parse* parse = is_Parse();
109 if (parse == NULL) {
110 if (bci() != jvms()->bci()) return false;
111 if (sp() != (int)jvms()->sp()) return false;
112 return true;
113 }
114 if (jvms()->method() != parse->method()) return false;
115 if (jvms()->bci() != parse->bci()) return false;
116 int jvms_sp = jvms()->sp();
117 if (jvms_sp != parse->sp()) return false;
118 int jvms_depth = jvms()->depth();
119 if (jvms_depth != parse->depth()) return false;
120 return true;
121 }
122
123 // Local helper checks for special internal merge points
124 // used to accumulate and merge exception states.
125 // They are marked by the region's in(0) edge being the map itself.
126 // Such merge points must never "escape" into the parser at large,
127 // until they have been handed to gvn.transform.
128 static bool is_hidden_merge(Node* reg) {
129 if (reg == NULL) return false;
130 if (reg->is_Phi()) {
131 reg = reg->in(0);
132 if (reg == NULL) return false;
133 }
134 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
135 }
136
137 void GraphKit::verify_map() const {
138 if (map() == NULL) return; // null map is OK
139 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
140 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
141 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
142 }
143
144 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
145 assert(ex_map->next_exception() == NULL, "not already part of a chain");
146 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
147 }
148 #endif
149
150 //---------------------------stop_and_kill_map---------------------------------
151 // Set _map to NULL, signalling a stop to further bytecode execution.
152 // First smash the current map's control to a constant, to mark it dead.
153 void GraphKit::stop_and_kill_map() {
154 SafePointNode* dead_map = stop();
155 if (dead_map != NULL) {
156 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
157 assert(dead_map->is_killed(), "must be so marked");
158 }
159 }
160
161
162 //--------------------------------stopped--------------------------------------
163 // Tell if _map is NULL, or control is top.
164 bool GraphKit::stopped() {
165 if (map() == NULL) return true;
166 else if (control() == top()) return true;
167 else return false;
168 }
169
170
171 //-----------------------------has_ex_handler----------------------------------
172 // Tell if this method or any caller method has exception handlers.
173 bool GraphKit::has_ex_handler() {
174 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
175 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
176 return true;
177 }
178 }
179 return false;
180 }
181
182 //------------------------------save_ex_oop------------------------------------
183 // Save an exception without blowing stack contents or other JVM state.
184 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
185 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
186 ex_map->add_req(ex_oop);
187 debug_only(verify_exception_state(ex_map));
188 }
189
190 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
191 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
192 Node* ex_oop = ex_map->in(ex_map->req()-1);
193 if (clear_it) ex_map->del_req(ex_map->req()-1);
194 return ex_oop;
195 }
196
197 //-----------------------------saved_ex_oop------------------------------------
198 // Recover a saved exception from its map.
199 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
200 return common_saved_ex_oop(ex_map, false);
201 }
202
203 //--------------------------clear_saved_ex_oop---------------------------------
204 // Erase a previously saved exception from its map.
205 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
206 return common_saved_ex_oop(ex_map, true);
207 }
208
209 #ifdef ASSERT
210 //---------------------------has_saved_ex_oop----------------------------------
211 // Erase a previously saved exception from its map.
212 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
213 return ex_map->req() == ex_map->jvms()->endoff()+1;
214 }
215 #endif
216
217 //-------------------------make_exception_state--------------------------------
218 // Turn the current JVM state into an exception state, appending the ex_oop.
219 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
220 sync_jvms();
221 SafePointNode* ex_map = stop(); // do not manipulate this map any more
222 set_saved_ex_oop(ex_map, ex_oop);
223 return ex_map;
224 }
225
226
227 //--------------------------add_exception_state--------------------------------
228 // Add an exception to my list of exceptions.
229 void GraphKit::add_exception_state(SafePointNode* ex_map) {
230 if (ex_map == NULL || ex_map->control() == top()) {
231 return;
232 }
233 #ifdef ASSERT
234 verify_exception_state(ex_map);
235 if (has_exceptions()) {
236 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
237 }
238 #endif
239
240 // If there is already an exception of exactly this type, merge with it.
241 // In particular, null-checks and other low-level exceptions common up here.
242 Node* ex_oop = saved_ex_oop(ex_map);
243 const Type* ex_type = _gvn.type(ex_oop);
244 if (ex_oop == top()) {
245 // No action needed.
246 return;
247 }
248 assert(ex_type->isa_instptr(), "exception must be an instance");
249 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
250 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
251 // We check sp also because call bytecodes can generate exceptions
252 // both before and after arguments are popped!
253 if (ex_type2 == ex_type
254 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
255 combine_exception_states(ex_map, e2);
256 return;
257 }
258 }
259
260 // No pre-existing exception of the same type. Chain it on the list.
261 push_exception_state(ex_map);
262 }
263
264 //-----------------------add_exception_states_from-----------------------------
265 void GraphKit::add_exception_states_from(JVMState* jvms) {
266 SafePointNode* ex_map = jvms->map()->next_exception();
267 if (ex_map != NULL) {
268 jvms->map()->set_next_exception(NULL);
269 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
270 next_map = ex_map->next_exception();
271 ex_map->set_next_exception(NULL);
272 add_exception_state(ex_map);
273 }
274 }
275 }
276
277 //-----------------------transfer_exceptions_into_jvms-------------------------
278 JVMState* GraphKit::transfer_exceptions_into_jvms() {
279 if (map() == NULL) {
280 // We need a JVMS to carry the exceptions, but the map has gone away.
281 // Create a scratch JVMS, cloned from any of the exception states...
282 if (has_exceptions()) {
283 _map = _exceptions;
284 _map = clone_map();
285 _map->set_next_exception(NULL);
286 clear_saved_ex_oop(_map);
287 debug_only(verify_map());
288 } else {
289 // ...or created from scratch
290 JVMState* jvms = new (C) JVMState(_method, NULL);
291 jvms->set_bci(_bci);
292 jvms->set_sp(_sp);
293 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms));
294 set_jvms(jvms);
295 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
296 set_all_memory(top());
297 while (map()->req() < jvms->endoff()) map()->add_req(top());
298 }
299 // (This is a kludge, in case you didn't notice.)
300 set_control(top());
301 }
302 JVMState* jvms = sync_jvms();
303 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
304 jvms->map()->set_next_exception(_exceptions);
305 _exceptions = NULL; // done with this set of exceptions
306 return jvms;
307 }
308
309 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
310 assert(is_hidden_merge(dstphi), "must be a special merge node");
311 assert(is_hidden_merge(srcphi), "must be a special merge node");
312 uint limit = srcphi->req();
313 for (uint i = PhiNode::Input; i < limit; i++) {
314 dstphi->add_req(srcphi->in(i));
315 }
316 }
317 static inline void add_one_req(Node* dstphi, Node* src) {
318 assert(is_hidden_merge(dstphi), "must be a special merge node");
319 assert(!is_hidden_merge(src), "must not be a special merge node");
320 dstphi->add_req(src);
321 }
322
323 //-----------------------combine_exception_states------------------------------
324 // This helper function combines exception states by building phis on a
325 // specially marked state-merging region. These regions and phis are
326 // untransformed, and can build up gradually. The region is marked by
327 // having a control input of its exception map, rather than NULL. Such
328 // regions do not appear except in this function, and in use_exception_state.
329 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
330 if (failing()) return; // dying anyway...
331 JVMState* ex_jvms = ex_map->_jvms;
332 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
333 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
334 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
335 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
336 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
337 assert(ex_map->req() == phi_map->req(), "matching maps");
338 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
339 Node* hidden_merge_mark = root();
340 Node* region = phi_map->control();
341 MergeMemNode* phi_mem = phi_map->merged_memory();
342 MergeMemNode* ex_mem = ex_map->merged_memory();
343 if (region->in(0) != hidden_merge_mark) {
344 // The control input is not (yet) a specially-marked region in phi_map.
345 // Make it so, and build some phis.
346 region = new (C) RegionNode(2);
347 _gvn.set_type(region, Type::CONTROL);
348 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
349 region->init_req(1, phi_map->control());
350 phi_map->set_control(region);
351 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
352 record_for_igvn(io_phi);
353 _gvn.set_type(io_phi, Type::ABIO);
354 phi_map->set_i_o(io_phi);
355 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
356 Node* m = mms.memory();
357 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
358 record_for_igvn(m_phi);
359 _gvn.set_type(m_phi, Type::MEMORY);
360 mms.set_memory(m_phi);
361 }
362 }
363
364 // Either or both of phi_map and ex_map might already be converted into phis.
365 Node* ex_control = ex_map->control();
366 // if there is special marking on ex_map also, we add multiple edges from src
367 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
368 // how wide was the destination phi_map, originally?
369 uint orig_width = region->req();
370
371 if (add_multiple) {
372 add_n_reqs(region, ex_control);
373 add_n_reqs(phi_map->i_o(), ex_map->i_o());
374 } else {
375 // ex_map has no merges, so we just add single edges everywhere
376 add_one_req(region, ex_control);
377 add_one_req(phi_map->i_o(), ex_map->i_o());
378 }
379 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
380 if (mms.is_empty()) {
381 // get a copy of the base memory, and patch some inputs into it
382 const TypePtr* adr_type = mms.adr_type(C);
383 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
384 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
385 mms.set_memory(phi);
386 // Prepare to append interesting stuff onto the newly sliced phi:
387 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
388 }
389 // Append stuff from ex_map:
390 if (add_multiple) {
391 add_n_reqs(mms.memory(), mms.memory2());
392 } else {
393 add_one_req(mms.memory(), mms.memory2());
394 }
395 }
396 uint limit = ex_map->req();
397 for (uint i = TypeFunc::Parms; i < limit; i++) {
398 // Skip everything in the JVMS after tos. (The ex_oop follows.)
399 if (i == tos) i = ex_jvms->monoff();
400 Node* src = ex_map->in(i);
401 Node* dst = phi_map->in(i);
402 if (src != dst) {
403 PhiNode* phi;
404 if (dst->in(0) != region) {
405 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
406 record_for_igvn(phi);
407 _gvn.set_type(phi, phi->type());
408 phi_map->set_req(i, dst);
409 // Prepare to append interesting stuff onto the new phi:
410 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
411 } else {
412 assert(dst->is_Phi(), "nobody else uses a hidden region");
413 phi = dst->as_Phi();
414 }
415 if (add_multiple && src->in(0) == ex_control) {
416 // Both are phis.
417 add_n_reqs(dst, src);
418 } else {
419 while (dst->req() < region->req()) add_one_req(dst, src);
420 }
421 const Type* srctype = _gvn.type(src);
422 if (phi->type() != srctype) {
423 const Type* dsttype = phi->type()->meet_speculative(srctype);
424 if (phi->type() != dsttype) {
425 phi->set_type(dsttype);
426 _gvn.set_type(phi, dsttype);
427 }
428 }
429 }
430 }
431 }
432
433 //--------------------------use_exception_state--------------------------------
434 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
435 if (failing()) { stop(); return top(); }
436 Node* region = phi_map->control();
437 Node* hidden_merge_mark = root();
438 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
439 Node* ex_oop = clear_saved_ex_oop(phi_map);
440 if (region->in(0) == hidden_merge_mark) {
441 // Special marking for internal ex-states. Process the phis now.
442 region->set_req(0, region); // now it's an ordinary region
443 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
444 // Note: Setting the jvms also sets the bci and sp.
445 set_control(_gvn.transform(region));
446 uint tos = jvms()->stkoff() + sp();
447 for (uint i = 1; i < tos; i++) {
448 Node* x = phi_map->in(i);
449 if (x->in(0) == region) {
450 assert(x->is_Phi(), "expected a special phi");
451 phi_map->set_req(i, _gvn.transform(x));
452 }
453 }
454 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
455 Node* x = mms.memory();
456 if (x->in(0) == region) {
457 assert(x->is_Phi(), "nobody else uses a hidden region");
458 mms.set_memory(_gvn.transform(x));
459 }
460 }
461 if (ex_oop->in(0) == region) {
462 assert(ex_oop->is_Phi(), "expected a special phi");
463 ex_oop = _gvn.transform(ex_oop);
464 }
465 } else {
466 set_jvms(phi_map->jvms());
467 }
468
469 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
470 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
471 return ex_oop;
472 }
473
474 //---------------------------------java_bc-------------------------------------
475 Bytecodes::Code GraphKit::java_bc() const {
476 ciMethod* method = this->method();
477 int bci = this->bci();
478 if (method != NULL && bci != InvocationEntryBci)
479 return method->java_code_at_bci(bci);
480 else
481 return Bytecodes::_illegal;
482 }
483
484 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
485 bool must_throw) {
486 // if the exception capability is set, then we will generate code
487 // to check the JavaThread.should_post_on_exceptions flag to see
488 // if we actually need to report exception events (for this
489 // thread). If we don't need to report exception events, we will
490 // take the normal fast path provided by add_exception_events. If
491 // exception event reporting is enabled for this thread, we will
492 // take the uncommon_trap in the BuildCutout below.
493
494 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
495 Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
496 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
497 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
498
499 // Test the should_post_on_exceptions_flag vs. 0
500 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
501 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
502
503 // Branch to slow_path if should_post_on_exceptions_flag was true
504 { BuildCutout unless(this, tst, PROB_MAX);
505 // Do not try anything fancy if we're notifying the VM on every throw.
506 // Cf. case Bytecodes::_athrow in parse2.cpp.
507 uncommon_trap(reason, Deoptimization::Action_none,
508 (ciKlass*)NULL, (char*)NULL, must_throw);
509 }
510
511 }
512
513 //------------------------------builtin_throw----------------------------------
514 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
515 bool must_throw = true;
516
517 if (env()->jvmti_can_post_on_exceptions()) {
518 // check if we must post exception events, take uncommon trap if so
519 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
520 // here if should_post_on_exceptions is false
521 // continue on with the normal codegen
522 }
523
524 // If this particular condition has not yet happened at this
525 // bytecode, then use the uncommon trap mechanism, and allow for
526 // a future recompilation if several traps occur here.
527 // If the throw is hot, try to use a more complicated inline mechanism
528 // which keeps execution inside the compiled code.
529 bool treat_throw_as_hot = false;
530 ciMethodData* md = method()->method_data();
531
532 if (ProfileTraps) {
533 if (too_many_traps(reason)) {
534 treat_throw_as_hot = true;
535 }
536 // (If there is no MDO at all, assume it is early in
537 // execution, and that any deopts are part of the
538 // startup transient, and don't need to be remembered.)
539
540 // Also, if there is a local exception handler, treat all throws
541 // as hot if there has been at least one in this method.
542 if (C->trap_count(reason) != 0
543 && method()->method_data()->trap_count(reason) != 0
544 && has_ex_handler()) {
545 treat_throw_as_hot = true;
546 }
547 }
548
549 // If this throw happens frequently, an uncommon trap might cause
550 // a performance pothole. If there is a local exception handler,
551 // and if this particular bytecode appears to be deoptimizing often,
552 // let us handle the throw inline, with a preconstructed instance.
553 // Note: If the deopt count has blown up, the uncommon trap
554 // runtime is going to flush this nmethod, not matter what.
555 if (treat_throw_as_hot
556 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
557 // If the throw is local, we use a pre-existing instance and
558 // punt on the backtrace. This would lead to a missing backtrace
559 // (a repeat of 4292742) if the backtrace object is ever asked
560 // for its backtrace.
561 // Fixing this remaining case of 4292742 requires some flavor of
562 // escape analysis. Leave that for the future.
563 ciInstance* ex_obj = NULL;
564 switch (reason) {
565 case Deoptimization::Reason_null_check:
566 ex_obj = env()->NullPointerException_instance();
567 break;
568 case Deoptimization::Reason_div0_check:
569 ex_obj = env()->ArithmeticException_instance();
570 break;
571 case Deoptimization::Reason_range_check:
572 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
573 break;
574 case Deoptimization::Reason_class_check:
575 if (java_bc() == Bytecodes::_aastore) {
576 ex_obj = env()->ArrayStoreException_instance();
577 } else {
578 ex_obj = env()->ClassCastException_instance();
579 }
580 break;
581 }
582 if (failing()) { stop(); return; } // exception allocation might fail
583 if (ex_obj != NULL) {
584 // Cheat with a preallocated exception object.
585 if (C->log() != NULL)
586 C->log()->elem("hot_throw preallocated='1' reason='%s'",
587 Deoptimization::trap_reason_name(reason));
588 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
589 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
590
591 // Clear the detail message of the preallocated exception object.
592 // Weblogic sometimes mutates the detail message of exceptions
593 // using reflection.
594 int offset = java_lang_Throwable::get_detailMessage_offset();
595 const TypePtr* adr_typ = ex_con->add_offset(offset);
596
597 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
598 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
599 // Conservatively release stores of object references.
600 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
601
602 add_exception_state(make_exception_state(ex_node));
603 return;
604 }
605 }
606
607 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
608 // It won't be much cheaper than bailing to the interp., since we'll
609 // have to pass up all the debug-info, and the runtime will have to
610 // create the stack trace.
611
612 // Usual case: Bail to interpreter.
613 // Reserve the right to recompile if we haven't seen anything yet.
614
615 assert(!Deoptimization::reason_is_speculate(reason), "unsupported");
616 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
617 if (treat_throw_as_hot
618 && (method()->method_data()->trap_recompiled_at(bci(), NULL)
619 || C->too_many_traps(reason))) {
620 // We cannot afford to take more traps here. Suffer in the interpreter.
621 if (C->log() != NULL)
622 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
623 Deoptimization::trap_reason_name(reason),
624 C->trap_count(reason));
625 action = Deoptimization::Action_none;
626 }
627
628 // "must_throw" prunes the JVM state to include only the stack, if there
629 // are no local exception handlers. This should cut down on register
630 // allocation time and code size, by drastically reducing the number
631 // of in-edges on the call to the uncommon trap.
632
633 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
634 }
635
636
637 //----------------------------PreserveJVMState---------------------------------
638 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
639 debug_only(kit->verify_map());
640 _kit = kit;
641 _map = kit->map(); // preserve the map
642 _sp = kit->sp();
643 kit->set_map(clone_map ? kit->clone_map() : NULL);
644 Compile::current()->inc_preserve_jvm_state();
645 #ifdef ASSERT
646 _bci = kit->bci();
647 Parse* parser = kit->is_Parse();
648 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
649 _block = block;
650 #endif
651 }
652 PreserveJVMState::~PreserveJVMState() {
653 GraphKit* kit = _kit;
654 #ifdef ASSERT
655 assert(kit->bci() == _bci, "bci must not shift");
656 Parse* parser = kit->is_Parse();
657 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
658 assert(block == _block, "block must not shift");
659 #endif
660 kit->set_map(_map);
661 kit->set_sp(_sp);
662 Compile::current()->dec_preserve_jvm_state();
663 }
664
665
666 //-----------------------------BuildCutout-------------------------------------
667 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
668 : PreserveJVMState(kit)
669 {
670 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
671 SafePointNode* outer_map = _map; // preserved map is caller's
672 SafePointNode* inner_map = kit->map();
673 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
674 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
675 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
676 }
677 BuildCutout::~BuildCutout() {
678 GraphKit* kit = _kit;
679 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
680 }
681
682 //---------------------------PreserveReexecuteState----------------------------
683 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
684 assert(!kit->stopped(), "must call stopped() before");
685 _kit = kit;
686 _sp = kit->sp();
687 _reexecute = kit->jvms()->_reexecute;
688 }
689 PreserveReexecuteState::~PreserveReexecuteState() {
690 if (_kit->stopped()) return;
691 _kit->jvms()->_reexecute = _reexecute;
692 _kit->set_sp(_sp);
693 }
694
695 //------------------------------clone_map--------------------------------------
696 // Implementation of PreserveJVMState
697 //
698 // Only clone_map(...) here. If this function is only used in the
699 // PreserveJVMState class we may want to get rid of this extra
700 // function eventually and do it all there.
701
702 SafePointNode* GraphKit::clone_map() {
703 if (map() == NULL) return NULL;
704
705 // Clone the memory edge first
706 Node* mem = MergeMemNode::make(C, map()->memory());
707 gvn().set_type_bottom(mem);
708
709 SafePointNode *clonemap = (SafePointNode*)map()->clone();
710 JVMState* jvms = this->jvms();
711 JVMState* clonejvms = jvms->clone_shallow(C);
712 clonemap->set_memory(mem);
713 clonemap->set_jvms(clonejvms);
714 clonejvms->set_map(clonemap);
715 record_for_igvn(clonemap);
716 gvn().set_type_bottom(clonemap);
717 return clonemap;
718 }
719
720
721 //-----------------------------set_map_clone-----------------------------------
722 void GraphKit::set_map_clone(SafePointNode* m) {
723 _map = m;
724 _map = clone_map();
725 _map->set_next_exception(NULL);
726 debug_only(verify_map());
727 }
728
729
730 //----------------------------kill_dead_locals---------------------------------
731 // Detect any locals which are known to be dead, and force them to top.
732 void GraphKit::kill_dead_locals() {
733 // Consult the liveness information for the locals. If any
734 // of them are unused, then they can be replaced by top(). This
735 // should help register allocation time and cut down on the size
736 // of the deoptimization information.
737
738 // This call is made from many of the bytecode handling
739 // subroutines called from the Big Switch in do_one_bytecode.
740 // Every bytecode which might include a slow path is responsible
741 // for killing its dead locals. The more consistent we
742 // are about killing deads, the fewer useless phis will be
743 // constructed for them at various merge points.
744
745 // bci can be -1 (InvocationEntryBci). We return the entry
746 // liveness for the method.
747
748 if (method() == NULL || method()->code_size() == 0) {
749 // We are building a graph for a call to a native method.
750 // All locals are live.
751 return;
752 }
753
754 ResourceMark rm;
755
756 // Consult the liveness information for the locals. If any
757 // of them are unused, then they can be replaced by top(). This
758 // should help register allocation time and cut down on the size
759 // of the deoptimization information.
760 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
761
762 int len = (int)live_locals.size();
763 assert(len <= jvms()->loc_size(), "too many live locals");
764 for (int local = 0; local < len; local++) {
765 if (!live_locals.at(local)) {
766 set_local(local, top());
767 }
768 }
769 }
770
771 #ifdef ASSERT
772 //-------------------------dead_locals_are_killed------------------------------
773 // Return true if all dead locals are set to top in the map.
774 // Used to assert "clean" debug info at various points.
775 bool GraphKit::dead_locals_are_killed() {
776 if (method() == NULL || method()->code_size() == 0) {
777 // No locals need to be dead, so all is as it should be.
778 return true;
779 }
780
781 // Make sure somebody called kill_dead_locals upstream.
782 ResourceMark rm;
783 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
784 if (jvms->loc_size() == 0) continue; // no locals to consult
785 SafePointNode* map = jvms->map();
786 ciMethod* method = jvms->method();
787 int bci = jvms->bci();
788 if (jvms == this->jvms()) {
789 bci = this->bci(); // it might not yet be synched
790 }
791 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
792 int len = (int)live_locals.size();
793 if (!live_locals.is_valid() || len == 0)
794 // This method is trivial, or is poisoned by a breakpoint.
795 return true;
796 assert(len == jvms->loc_size(), "live map consistent with locals map");
797 for (int local = 0; local < len; local++) {
798 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
799 if (PrintMiscellaneous && (Verbose || WizardMode)) {
800 tty->print_cr("Zombie local %d: ", local);
801 jvms->dump();
802 }
803 return false;
804 }
805 }
806 }
807 return true;
808 }
809
810 #endif //ASSERT
811
812 // Helper function for enforcing certain bytecodes to reexecute if
813 // deoptimization happens
814 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
815 ciMethod* cur_method = jvms->method();
816 int cur_bci = jvms->bci();
817 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
818 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
819 return Interpreter::bytecode_should_reexecute(code) ||
820 is_anewarray && code == Bytecodes::_multianewarray;
821 // Reexecute _multianewarray bytecode which was replaced with
822 // sequence of [a]newarray. See Parse::do_multianewarray().
823 //
824 // Note: interpreter should not have it set since this optimization
825 // is limited by dimensions and guarded by flag so in some cases
826 // multianewarray() runtime calls will be generated and
827 // the bytecode should not be reexecutes (stack will not be reset).
828 } else
829 return false;
830 }
831
832 // Helper function for adding JVMState and debug information to node
833 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
834 // Add the safepoint edges to the call (or other safepoint).
835
836 // Make sure dead locals are set to top. This
837 // should help register allocation time and cut down on the size
838 // of the deoptimization information.
839 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
840
841 // Walk the inline list to fill in the correct set of JVMState's
842 // Also fill in the associated edges for each JVMState.
843
844 // If the bytecode needs to be reexecuted we need to put
845 // the arguments back on the stack.
846 const bool should_reexecute = jvms()->should_reexecute();
847 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
848
849 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
850 // undefined if the bci is different. This is normal for Parse but it
851 // should not happen for LibraryCallKit because only one bci is processed.
852 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
853 "in LibraryCallKit the reexecute bit should not change");
854
855 // If we are guaranteed to throw, we can prune everything but the
856 // input to the current bytecode.
857 bool can_prune_locals = false;
858 uint stack_slots_not_pruned = 0;
859 int inputs = 0, depth = 0;
860 if (must_throw) {
861 assert(method() == youngest_jvms->method(), "sanity");
862 if (compute_stack_effects(inputs, depth)) {
863 can_prune_locals = true;
864 stack_slots_not_pruned = inputs;
865 }
866 }
867
868 if (env()->should_retain_local_variables()) {
869 // At any safepoint, this method can get breakpointed, which would
870 // then require an immediate deoptimization.
871 can_prune_locals = false; // do not prune locals
872 stack_slots_not_pruned = 0;
873 }
874
875 // do not scribble on the input jvms
876 JVMState* out_jvms = youngest_jvms->clone_deep(C);
877 call->set_jvms(out_jvms); // Start jvms list for call node
878
879 // For a known set of bytecodes, the interpreter should reexecute them if
880 // deoptimization happens. We set the reexecute state for them here
881 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
882 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
883 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
884 }
885
886 // Presize the call:
887 DEBUG_ONLY(uint non_debug_edges = call->req());
888 call->add_req_batch(top(), youngest_jvms->debug_depth());
889 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
890
891 // Set up edges so that the call looks like this:
892 // Call [state:] ctl io mem fptr retadr
893 // [parms:] parm0 ... parmN
894 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
895 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
896 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
897 // Note that caller debug info precedes callee debug info.
898
899 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
900 uint debug_ptr = call->req();
901
902 // Loop over the map input edges associated with jvms, add them
903 // to the call node, & reset all offsets to match call node array.
904 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
905 uint debug_end = debug_ptr;
906 uint debug_start = debug_ptr - in_jvms->debug_size();
907 debug_ptr = debug_start; // back up the ptr
908
909 uint p = debug_start; // walks forward in [debug_start, debug_end)
910 uint j, k, l;
911 SafePointNode* in_map = in_jvms->map();
912 out_jvms->set_map(call);
913
914 if (can_prune_locals) {
915 assert(in_jvms->method() == out_jvms->method(), "sanity");
916 // If the current throw can reach an exception handler in this JVMS,
917 // then we must keep everything live that can reach that handler.
918 // As a quick and dirty approximation, we look for any handlers at all.
919 if (in_jvms->method()->has_exception_handlers()) {
920 can_prune_locals = false;
921 }
922 }
923
924 // Add the Locals
925 k = in_jvms->locoff();
926 l = in_jvms->loc_size();
927 out_jvms->set_locoff(p);
928 if (!can_prune_locals) {
929 for (j = 0; j < l; j++)
930 call->set_req(p++, in_map->in(k+j));
931 } else {
932 p += l; // already set to top above by add_req_batch
933 }
934
935 // Add the Expression Stack
936 k = in_jvms->stkoff();
937 l = in_jvms->sp();
938 out_jvms->set_stkoff(p);
939 if (!can_prune_locals) {
940 for (j = 0; j < l; j++)
941 call->set_req(p++, in_map->in(k+j));
942 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
943 // Divide stack into {S0,...,S1}, where S0 is set to top.
944 uint s1 = stack_slots_not_pruned;
945 stack_slots_not_pruned = 0; // for next iteration
946 if (s1 > l) s1 = l;
947 uint s0 = l - s1;
948 p += s0; // skip the tops preinstalled by add_req_batch
949 for (j = s0; j < l; j++)
950 call->set_req(p++, in_map->in(k+j));
951 } else {
952 p += l; // already set to top above by add_req_batch
953 }
954
955 // Add the Monitors
956 k = in_jvms->monoff();
957 l = in_jvms->mon_size();
958 out_jvms->set_monoff(p);
959 for (j = 0; j < l; j++)
960 call->set_req(p++, in_map->in(k+j));
961
962 // Copy any scalar object fields.
963 k = in_jvms->scloff();
964 l = in_jvms->scl_size();
965 out_jvms->set_scloff(p);
966 for (j = 0; j < l; j++)
967 call->set_req(p++, in_map->in(k+j));
968
969 // Finish the new jvms.
970 out_jvms->set_endoff(p);
971
972 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
973 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
974 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
975 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
976 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
977 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
978
979 // Update the two tail pointers in parallel.
980 out_jvms = out_jvms->caller();
981 in_jvms = in_jvms->caller();
982 }
983
984 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
985
986 // Test the correctness of JVMState::debug_xxx accessors:
987 assert(call->jvms()->debug_start() == non_debug_edges, "");
988 assert(call->jvms()->debug_end() == call->req(), "");
989 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
990 }
991
992 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
993 Bytecodes::Code code = java_bc();
994 if (code == Bytecodes::_wide) {
995 code = method()->java_code_at_bci(bci() + 1);
996 }
997
998 BasicType rtype = T_ILLEGAL;
999 int rsize = 0;
1000
1001 if (code != Bytecodes::_illegal) {
1002 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1003 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1004 if (rtype < T_CONFLICT)
1005 rsize = type2size[rtype];
1006 }
1007
1008 switch (code) {
1009 case Bytecodes::_illegal:
1010 return false;
1011
1012 case Bytecodes::_ldc:
1013 case Bytecodes::_ldc_w:
1014 case Bytecodes::_ldc2_w:
1015 inputs = 0;
1016 break;
1017
1018 case Bytecodes::_dup: inputs = 1; break;
1019 case Bytecodes::_dup_x1: inputs = 2; break;
1020 case Bytecodes::_dup_x2: inputs = 3; break;
1021 case Bytecodes::_dup2: inputs = 2; break;
1022 case Bytecodes::_dup2_x1: inputs = 3; break;
1023 case Bytecodes::_dup2_x2: inputs = 4; break;
1024 case Bytecodes::_swap: inputs = 2; break;
1025 case Bytecodes::_arraylength: inputs = 1; break;
1026
1027 case Bytecodes::_getstatic:
1028 case Bytecodes::_putstatic:
1029 case Bytecodes::_getfield:
1030 case Bytecodes::_putfield:
1031 {
1032 bool ignored_will_link;
1033 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1034 int size = field->type()->size();
1035 bool is_get = (depth >= 0), is_static = (depth & 1);
1036 inputs = (is_static ? 0 : 1);
1037 if (is_get) {
1038 depth = size - inputs;
1039 } else {
1040 inputs += size; // putxxx pops the value from the stack
1041 depth = - inputs;
1042 }
1043 }
1044 break;
1045
1046 case Bytecodes::_invokevirtual:
1047 case Bytecodes::_invokespecial:
1048 case Bytecodes::_invokestatic:
1049 case Bytecodes::_invokedynamic:
1050 case Bytecodes::_invokeinterface:
1051 {
1052 int size = method()->get_stack_effect_at_invoke(bci(), code, inputs);
1053 depth = size;
1054 }
1055 break;
1056
1057 case Bytecodes::_multianewarray:
1058 {
1059 ciBytecodeStream iter(method());
1060 iter.reset_to_bci(bci());
1061 iter.next();
1062 inputs = iter.get_dimensions();
1063 assert(rsize == 1, "");
1064 depth = rsize - inputs;
1065 }
1066 break;
1067
1068 case Bytecodes::_ireturn:
1069 case Bytecodes::_lreturn:
1070 case Bytecodes::_freturn:
1071 case Bytecodes::_dreturn:
1072 case Bytecodes::_areturn:
1073 assert(rsize = -depth, "");
1074 inputs = rsize;
1075 break;
1076
1077 case Bytecodes::_jsr:
1078 case Bytecodes::_jsr_w:
1079 inputs = 0;
1080 depth = 1; // S.B. depth=1, not zero
1081 break;
1082
1083 default:
1084 // bytecode produces a typed result
1085 inputs = rsize - depth;
1086 assert(inputs >= 0, "");
1087 break;
1088 }
1089
1090 #ifdef ASSERT
1091 // spot check
1092 int outputs = depth + inputs;
1093 assert(outputs >= 0, "sanity");
1094 switch (code) {
1095 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1096 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1097 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1098 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1099 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1100 }
1101 #endif //ASSERT
1102
1103 return true;
1104 }
1105
1106
1107
1108 //------------------------------basic_plus_adr---------------------------------
1109 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1110 // short-circuit a common case
1111 if (offset == intcon(0)) return ptr;
1112 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1113 }
1114
1115 Node* GraphKit::ConvI2L(Node* offset) {
1116 // short-circuit a common case
1117 jint offset_con = find_int_con(offset, Type::OffsetBot);
1118 if (offset_con != Type::OffsetBot) {
1119 return longcon((jlong) offset_con);
1120 }
1121 return _gvn.transform( new (C) ConvI2LNode(offset));
1122 }
1123 Node* GraphKit::ConvL2I(Node* offset) {
1124 // short-circuit a common case
1125 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1126 if (offset_con != (jlong)Type::OffsetBot) {
1127 return intcon((int) offset_con);
1128 }
1129 return _gvn.transform( new (C) ConvL2INode(offset));
1130 }
1131
1132 //-------------------------load_object_klass-----------------------------------
1133 Node* GraphKit::load_object_klass(Node* obj) {
1134 // Special-case a fresh allocation to avoid building nodes:
1135 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1136 if (akls != NULL) return akls;
1137 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1138 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1139 }
1140
1141 //-------------------------load_array_length-----------------------------------
1142 Node* GraphKit::load_array_length(Node* array) {
1143 // Special-case a fresh allocation to avoid building nodes:
1144 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1145 Node *alen;
1146 if (alloc == NULL) {
1147 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1148 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1149 } else {
1150 alen = alloc->Ideal_length();
1151 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1152 if (ccast != alen) {
1153 alen = _gvn.transform(ccast);
1154 }
1155 }
1156 return alen;
1157 }
1158
1159 //------------------------------do_null_check----------------------------------
1160 // Helper function to do a NULL pointer check. Returned value is
1161 // the incoming address with NULL casted away. You are allowed to use the
1162 // not-null value only if you are control dependent on the test.
1163 extern int explicit_null_checks_inserted,
1164 explicit_null_checks_elided;
1165 Node* GraphKit::null_check_common(Node* value, BasicType type,
1166 // optional arguments for variations:
1167 bool assert_null,
1168 Node* *null_control) {
1169 assert(!assert_null || null_control == NULL, "not both at once");
1170 if (stopped()) return top();
1171 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1172 // For some performance testing, we may wish to suppress null checking.
1173 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1174 return value;
1175 }
1176 explicit_null_checks_inserted++;
1177
1178 // Construct NULL check
1179 Node *chk = NULL;
1180 switch(type) {
1181 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1182 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1183 case T_ARRAY : // fall through
1184 type = T_OBJECT; // simplify further tests
1185 case T_OBJECT : {
1186 const Type *t = _gvn.type( value );
1187
1188 const TypeOopPtr* tp = t->isa_oopptr();
1189 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1190 // Only for do_null_check, not any of its siblings:
1191 && !assert_null && null_control == NULL) {
1192 // Usually, any field access or invocation on an unloaded oop type
1193 // will simply fail to link, since the statically linked class is
1194 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1195 // the static class is loaded but the sharper oop type is not.
1196 // Rather than checking for this obscure case in lots of places,
1197 // we simply observe that a null check on an unloaded class
1198 // will always be followed by a nonsense operation, so we
1199 // can just issue the uncommon trap here.
1200 // Our access to the unloaded class will only be correct
1201 // after it has been loaded and initialized, which requires
1202 // a trip through the interpreter.
1203 #ifndef PRODUCT
1204 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1205 #endif
1206 uncommon_trap(Deoptimization::Reason_unloaded,
1207 Deoptimization::Action_reinterpret,
1208 tp->klass(), "!loaded");
1209 return top();
1210 }
1211
1212 if (assert_null) {
1213 // See if the type is contained in NULL_PTR.
1214 // If so, then the value is already null.
1215 if (t->higher_equal(TypePtr::NULL_PTR)) {
1216 explicit_null_checks_elided++;
1217 return value; // Elided null assert quickly!
1218 }
1219 } else {
1220 // See if mixing in the NULL pointer changes type.
1221 // If so, then the NULL pointer was not allowed in the original
1222 // type. In other words, "value" was not-null.
1223 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1224 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1225 explicit_null_checks_elided++;
1226 return value; // Elided null check quickly!
1227 }
1228 }
1229 chk = new (C) CmpPNode( value, null() );
1230 break;
1231 }
1232
1233 default:
1234 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1235 }
1236 assert(chk != NULL, "sanity check");
1237 chk = _gvn.transform(chk);
1238
1239 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1240 BoolNode *btst = new (C) BoolNode( chk, btest);
1241 Node *tst = _gvn.transform( btst );
1242
1243 //-----------
1244 // if peephole optimizations occurred, a prior test existed.
1245 // If a prior test existed, maybe it dominates as we can avoid this test.
1246 if (tst != btst && type == T_OBJECT) {
1247 // At this point we want to scan up the CFG to see if we can
1248 // find an identical test (and so avoid this test altogether).
1249 Node *cfg = control();
1250 int depth = 0;
1251 while( depth < 16 ) { // Limit search depth for speed
1252 if( cfg->Opcode() == Op_IfTrue &&
1253 cfg->in(0)->in(1) == tst ) {
1254 // Found prior test. Use "cast_not_null" to construct an identical
1255 // CastPP (and hence hash to) as already exists for the prior test.
1256 // Return that casted value.
1257 if (assert_null) {
1258 replace_in_map(value, null());
1259 return null(); // do not issue the redundant test
1260 }
1261 Node *oldcontrol = control();
1262 set_control(cfg);
1263 Node *res = cast_not_null(value);
1264 set_control(oldcontrol);
1265 explicit_null_checks_elided++;
1266 return res;
1267 }
1268 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1269 if (cfg == NULL) break; // Quit at region nodes
1270 depth++;
1271 }
1272 }
1273
1274 //-----------
1275 // Branch to failure if null
1276 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1277 Deoptimization::DeoptReason reason;
1278 if (assert_null)
1279 reason = Deoptimization::Reason_null_assert;
1280 else if (type == T_OBJECT)
1281 reason = Deoptimization::Reason_null_check;
1282 else
1283 reason = Deoptimization::Reason_div0_check;
1284
1285 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1286 // ciMethodData::has_trap_at will return a conservative -1 if any
1287 // must-be-null assertion has failed. This could cause performance
1288 // problems for a method after its first do_null_assert failure.
1289 // Consider using 'Reason_class_check' instead?
1290
1291 // To cause an implicit null check, we set the not-null probability
1292 // to the maximum (PROB_MAX). For an explicit check the probability
1293 // is set to a smaller value.
1294 if (null_control != NULL || too_many_traps(reason)) {
1295 // probability is less likely
1296 ok_prob = PROB_LIKELY_MAG(3);
1297 } else if (!assert_null &&
1298 (ImplicitNullCheckThreshold > 0) &&
1299 method() != NULL &&
1300 (method()->method_data()->trap_count(reason)
1301 >= (uint)ImplicitNullCheckThreshold)) {
1302 ok_prob = PROB_LIKELY_MAG(3);
1303 }
1304
1305 if (null_control != NULL) {
1306 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1307 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1308 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1309 if (null_true == top())
1310 explicit_null_checks_elided++;
1311 (*null_control) = null_true;
1312 } else {
1313 BuildCutout unless(this, tst, ok_prob);
1314 // Check for optimizer eliding test at parse time
1315 if (stopped()) {
1316 // Failure not possible; do not bother making uncommon trap.
1317 explicit_null_checks_elided++;
1318 } else if (assert_null) {
1319 uncommon_trap(reason,
1320 Deoptimization::Action_make_not_entrant,
1321 NULL, "assert_null");
1322 } else {
1323 replace_in_map(value, zerocon(type));
1324 builtin_throw(reason);
1325 }
1326 }
1327
1328 // Must throw exception, fall-thru not possible?
1329 if (stopped()) {
1330 return top(); // No result
1331 }
1332
1333 if (assert_null) {
1334 // Cast obj to null on this path.
1335 replace_in_map(value, zerocon(type));
1336 return zerocon(type);
1337 }
1338
1339 // Cast obj to not-null on this path, if there is no null_control.
1340 // (If there is a null_control, a non-null value may come back to haunt us.)
1341 if (type == T_OBJECT) {
1342 Node* cast = cast_not_null(value, false);
1343 if (null_control == NULL || (*null_control) == top())
1344 replace_in_map(value, cast);
1345 value = cast;
1346 }
1347
1348 return value;
1349 }
1350
1351
1352 //------------------------------cast_not_null----------------------------------
1353 // Cast obj to not-null on this path
1354 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1355 const Type *t = _gvn.type(obj);
1356 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1357 // Object is already not-null?
1358 if( t == t_not_null ) return obj;
1359
1360 Node *cast = new (C) CastPPNode(obj,t_not_null);
1361 cast->init_req(0, control());
1362 cast = _gvn.transform( cast );
1363
1364 // Scan for instances of 'obj' in the current JVM mapping.
1365 // These instances are known to be not-null after the test.
1366 if (do_replace_in_map)
1367 replace_in_map(obj, cast);
1368
1369 return cast; // Return casted value
1370 }
1371
1372
1373 //--------------------------replace_in_map-------------------------------------
1374 void GraphKit::replace_in_map(Node* old, Node* neww) {
1375 if (old == neww) {
1376 return;
1377 }
1378
1379 map()->replace_edge(old, neww);
1380
1381 // Note: This operation potentially replaces any edge
1382 // on the map. This includes locals, stack, and monitors
1383 // of the current (innermost) JVM state.
1384
1385 if (!ReplaceInParentMaps) {
1386 return;
1387 }
1388
1389 // PreserveJVMState doesn't do a deep copy so we can't modify
1390 // parents
1391 if (Compile::current()->has_preserve_jvm_state()) {
1392 return;
1393 }
1394
1395 Parse* parser = is_Parse();
1396 bool progress = true;
1397 Node* ctrl = map()->in(0);
1398 // Follow the chain of parsers and see whether the update can be
1399 // done in the map of callers. We can do the replace for a caller if
1400 // the current control post dominates the control of a caller.
1401 while (parser != NULL && parser->caller() != NULL && progress) {
1402 progress = false;
1403 Node* parent_map = parser->caller()->map();
1404 assert(parser->exits().map()->jvms()->depth() == parser->caller()->depth(), "map mismatch");
1405
1406 Node* parent_ctrl = parent_map->in(0);
1407
1408 while (parent_ctrl->is_Region()) {
1409 Node* n = parent_ctrl->as_Region()->is_copy();
1410 if (n == NULL) {
1411 break;
1412 }
1413 parent_ctrl = n;
1414 }
1415
1416 for (;;) {
1417 if (ctrl == parent_ctrl) {
1418 // update the map of the exits which is the one that will be
1419 // used when compilation resume after inlining
1420 parser->exits().map()->replace_edge(old, neww);
1421 progress = true;
1422 break;
1423 }
1424 if (ctrl->is_Proj() && ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
1425 ctrl = ctrl->in(0)->in(0);
1426 } else if (ctrl->is_Region()) {
1427 Node* n = ctrl->as_Region()->is_copy();
1428 if (n == NULL) {
1429 break;
1430 }
1431 ctrl = n;
1432 } else {
1433 break;
1434 }
1435 }
1436
1437 parser = parser->parent_parser();
1438 }
1439 }
1440
1441
1442 //=============================================================================
1443 //--------------------------------memory---------------------------------------
1444 Node* GraphKit::memory(uint alias_idx) {
1445 MergeMemNode* mem = merged_memory();
1446 Node* p = mem->memory_at(alias_idx);
1447 _gvn.set_type(p, Type::MEMORY); // must be mapped
1448 return p;
1449 }
1450
1451 //-----------------------------reset_memory------------------------------------
1452 Node* GraphKit::reset_memory() {
1453 Node* mem = map()->memory();
1454 // do not use this node for any more parsing!
1455 debug_only( map()->set_memory((Node*)NULL) );
1456 return _gvn.transform( mem );
1457 }
1458
1459 //------------------------------set_all_memory---------------------------------
1460 void GraphKit::set_all_memory(Node* newmem) {
1461 Node* mergemem = MergeMemNode::make(C, newmem);
1462 gvn().set_type_bottom(mergemem);
1463 map()->set_memory(mergemem);
1464 }
1465
1466 //------------------------------set_all_memory_call----------------------------
1467 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1468 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1469 set_all_memory(newmem);
1470 }
1471
1472 //=============================================================================
1473 //
1474 // parser factory methods for MemNodes
1475 //
1476 // These are layered on top of the factory methods in LoadNode and StoreNode,
1477 // and integrate with the parser's memory state and _gvn engine.
1478 //
1479
1480 // factory methods in "int adr_idx"
1481 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1482 int adr_idx,
1483 MemNode::MemOrd mo, bool require_atomic_access) {
1484 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1485 const TypePtr* adr_type = NULL; // debug-mode-only argument
1486 debug_only(adr_type = C->get_adr_type(adr_idx));
1487 Node* mem = memory(adr_idx);
1488 Node* ld;
1489 if (require_atomic_access && bt == T_LONG) {
1490 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo);
1491 } else {
1492 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo);
1493 }
1494 ld = _gvn.transform(ld);
1495 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1496 // Improve graph before escape analysis and boxing elimination.
1497 record_for_igvn(ld);
1498 }
1499 return ld;
1500 }
1501
1502 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1503 int adr_idx,
1504 MemNode::MemOrd mo,
1505 bool require_atomic_access) {
1506 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1507 const TypePtr* adr_type = NULL;
1508 debug_only(adr_type = C->get_adr_type(adr_idx));
1509 Node *mem = memory(adr_idx);
1510 Node* st;
1511 if (require_atomic_access && bt == T_LONG) {
1512 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1513 } else {
1514 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1515 }
1516 st = _gvn.transform(st);
1517 set_memory(st, adr_idx);
1518 // Back-to-back stores can only remove intermediate store with DU info
1519 // so push on worklist for optimizer.
1520 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1521 record_for_igvn(st);
1522
1523 return st;
1524 }
1525
1526
1527 void GraphKit::pre_barrier(bool do_load,
1528 Node* ctl,
1529 Node* obj,
1530 Node* adr,
1531 uint adr_idx,
1532 Node* val,
1533 const TypeOopPtr* val_type,
1534 Node* pre_val,
1535 BasicType bt) {
1536
1537 BarrierSet* bs = Universe::heap()->barrier_set();
1538 set_control(ctl);
1539 switch (bs->kind()) {
1540 case BarrierSet::G1SATBCT:
1541 case BarrierSet::G1SATBCTLogging:
1542 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1543 break;
1544
1545 case BarrierSet::CardTableModRef:
1546 case BarrierSet::CardTableExtension:
1547 case BarrierSet::ModRef:
1548 break;
1549
1550 case BarrierSet::Other:
1551 default :
1552 ShouldNotReachHere();
1553
1554 }
1555 }
1556
1557 bool GraphKit::can_move_pre_barrier() const {
1558 BarrierSet* bs = Universe::heap()->barrier_set();
1559 switch (bs->kind()) {
1560 case BarrierSet::G1SATBCT:
1561 case BarrierSet::G1SATBCTLogging:
1562 return true; // Can move it if no safepoint
1563
1564 case BarrierSet::CardTableModRef:
1565 case BarrierSet::CardTableExtension:
1566 case BarrierSet::ModRef:
1567 return true; // There is no pre-barrier
1568
1569 case BarrierSet::Other:
1570 default :
1571 ShouldNotReachHere();
1572 }
1573 return false;
1574 }
1575
1576 void GraphKit::post_barrier(Node* ctl,
1577 Node* store,
1578 Node* obj,
1579 Node* adr,
1580 uint adr_idx,
1581 Node* val,
1582 BasicType bt,
1583 bool use_precise) {
1584 BarrierSet* bs = Universe::heap()->barrier_set();
1585 set_control(ctl);
1586 switch (bs->kind()) {
1587 case BarrierSet::G1SATBCT:
1588 case BarrierSet::G1SATBCTLogging:
1589 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1590 break;
1591
1592 case BarrierSet::CardTableModRef:
1593 case BarrierSet::CardTableExtension:
1594 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1595 break;
1596
1597 case BarrierSet::ModRef:
1598 break;
1599
1600 case BarrierSet::Other:
1601 default :
1602 ShouldNotReachHere();
1603
1604 }
1605 }
1606
1607 Node* GraphKit::store_oop(Node* ctl,
1608 Node* obj,
1609 Node* adr,
1610 const TypePtr* adr_type,
1611 Node* val,
1612 const TypeOopPtr* val_type,
1613 BasicType bt,
1614 bool use_precise,
1615 MemNode::MemOrd mo) {
1616 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1617 // could be delayed during Parse (for example, in adjust_map_after_if()).
1618 // Execute transformation here to avoid barrier generation in such case.
1619 if (_gvn.type(val) == TypePtr::NULL_PTR)
1620 val = _gvn.makecon(TypePtr::NULL_PTR);
1621
1622 set_control(ctl);
1623 if (stopped()) return top(); // Dead path ?
1624
1625 assert(bt == T_OBJECT, "sanity");
1626 assert(val != NULL, "not dead path");
1627 uint adr_idx = C->get_alias_index(adr_type);
1628 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1629
1630 pre_barrier(true /* do_load */,
1631 control(), obj, adr, adr_idx, val, val_type,
1632 NULL /* pre_val */,
1633 bt);
1634
1635 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo);
1636 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1637 return store;
1638 }
1639
1640 // Could be an array or object we don't know at compile time (unsafe ref.)
1641 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1642 Node* obj, // containing obj
1643 Node* adr, // actual adress to store val at
1644 const TypePtr* adr_type,
1645 Node* val,
1646 BasicType bt,
1647 MemNode::MemOrd mo) {
1648 Compile::AliasType* at = C->alias_type(adr_type);
1649 const TypeOopPtr* val_type = NULL;
1650 if (adr_type->isa_instptr()) {
1651 if (at->field() != NULL) {
1652 // known field. This code is a copy of the do_put_xxx logic.
1653 ciField* field = at->field();
1654 if (!field->type()->is_loaded()) {
1655 val_type = TypeInstPtr::BOTTOM;
1656 } else {
1657 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1658 }
1659 }
1660 } else if (adr_type->isa_aryptr()) {
1661 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1662 }
1663 if (val_type == NULL) {
1664 val_type = TypeInstPtr::BOTTOM;
1665 }
1666 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo);
1667 }
1668
1669
1670 //-------------------------array_element_address-------------------------
1671 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1672 const TypeInt* sizetype) {
1673 uint shift = exact_log2(type2aelembytes(elembt));
1674 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1675
1676 // short-circuit a common case (saves lots of confusing waste motion)
1677 jint idx_con = find_int_con(idx, -1);
1678 if (idx_con >= 0) {
1679 intptr_t offset = header + ((intptr_t)idx_con << shift);
1680 return basic_plus_adr(ary, offset);
1681 }
1682
1683 // must be correct type for alignment purposes
1684 Node* base = basic_plus_adr(ary, header);
1685 #ifdef _LP64
1686 // The scaled index operand to AddP must be a clean 64-bit value.
1687 // Java allows a 32-bit int to be incremented to a negative
1688 // value, which appears in a 64-bit register as a large
1689 // positive number. Using that large positive number as an
1690 // operand in pointer arithmetic has bad consequences.
1691 // On the other hand, 32-bit overflow is rare, and the possibility
1692 // can often be excluded, if we annotate the ConvI2L node with
1693 // a type assertion that its value is known to be a small positive
1694 // number. (The prior range check has ensured this.)
1695 // This assertion is used by ConvI2LNode::Ideal.
1696 int index_max = max_jint - 1; // array size is max_jint, index is one less
1697 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1698 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1699 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) );
1700 #endif
1701 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1702 return basic_plus_adr(ary, base, scale);
1703 }
1704
1705 //-------------------------load_array_element-------------------------
1706 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1707 const Type* elemtype = arytype->elem();
1708 BasicType elembt = elemtype->array_element_basic_type();
1709 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1710 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1711 return ld;
1712 }
1713
1714 //-------------------------set_arguments_for_java_call-------------------------
1715 // Arguments (pre-popped from the stack) are taken from the JVMS.
1716 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1717 // Add the call arguments:
1718 uint nargs = call->method()->arg_size();
1719 for (uint i = 0; i < nargs; i++) {
1720 Node* arg = argument(i);
1721 call->init_req(i + TypeFunc::Parms, arg);
1722 }
1723 }
1724
1725 //---------------------------set_edges_for_java_call---------------------------
1726 // Connect a newly created call into the current JVMS.
1727 // A return value node (if any) is returned from set_edges_for_java_call.
1728 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1729
1730 // Add the predefined inputs:
1731 call->init_req( TypeFunc::Control, control() );
1732 call->init_req( TypeFunc::I_O , i_o() );
1733 call->init_req( TypeFunc::Memory , reset_memory() );
1734 call->init_req( TypeFunc::FramePtr, frameptr() );
1735 call->init_req( TypeFunc::ReturnAdr, top() );
1736
1737 add_safepoint_edges(call, must_throw);
1738
1739 Node* xcall = _gvn.transform(call);
1740
1741 if (xcall == top()) {
1742 set_control(top());
1743 return;
1744 }
1745 assert(xcall == call, "call identity is stable");
1746
1747 // Re-use the current map to produce the result.
1748
1749 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1750 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1751 set_all_memory_call(xcall, separate_io_proj);
1752
1753 //return xcall; // no need, caller already has it
1754 }
1755
1756 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1757 if (stopped()) return top(); // maybe the call folded up?
1758
1759 // Capture the return value, if any.
1760 Node* ret;
1761 if (call->method() == NULL ||
1762 call->method()->return_type()->basic_type() == T_VOID)
1763 ret = top();
1764 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1765
1766 // Note: Since any out-of-line call can produce an exception,
1767 // we always insert an I_O projection from the call into the result.
1768
1769 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1770
1771 if (separate_io_proj) {
1772 // The caller requested separate projections be used by the fall
1773 // through and exceptional paths, so replace the projections for
1774 // the fall through path.
1775 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1776 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1777 }
1778 return ret;
1779 }
1780
1781 //--------------------set_predefined_input_for_runtime_call--------------------
1782 // Reading and setting the memory state is way conservative here.
1783 // The real problem is that I am not doing real Type analysis on memory,
1784 // so I cannot distinguish card mark stores from other stores. Across a GC
1785 // point the Store Barrier and the card mark memory has to agree. I cannot
1786 // have a card mark store and its barrier split across the GC point from
1787 // either above or below. Here I get that to happen by reading ALL of memory.
1788 // A better answer would be to separate out card marks from other memory.
1789 // For now, return the input memory state, so that it can be reused
1790 // after the call, if this call has restricted memory effects.
1791 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1792 // Set fixed predefined input arguments
1793 Node* memory = reset_memory();
1794 call->init_req( TypeFunc::Control, control() );
1795 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1796 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1797 call->init_req( TypeFunc::FramePtr, frameptr() );
1798 call->init_req( TypeFunc::ReturnAdr, top() );
1799 return memory;
1800 }
1801
1802 //-------------------set_predefined_output_for_runtime_call--------------------
1803 // Set control and memory (not i_o) from the call.
1804 // If keep_mem is not NULL, use it for the output state,
1805 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1806 // If hook_mem is NULL, this call produces no memory effects at all.
1807 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1808 // then only that memory slice is taken from the call.
1809 // In the last case, we must put an appropriate memory barrier before
1810 // the call, so as to create the correct anti-dependencies on loads
1811 // preceding the call.
1812 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1813 Node* keep_mem,
1814 const TypePtr* hook_mem) {
1815 // no i/o
1816 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1817 if (keep_mem) {
1818 // First clone the existing memory state
1819 set_all_memory(keep_mem);
1820 if (hook_mem != NULL) {
1821 // Make memory for the call
1822 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1823 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1824 // We also use hook_mem to extract specific effects from arraycopy stubs.
1825 set_memory(mem, hook_mem);
1826 }
1827 // ...else the call has NO memory effects.
1828
1829 // Make sure the call advertises its memory effects precisely.
1830 // This lets us build accurate anti-dependences in gcm.cpp.
1831 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1832 "call node must be constructed correctly");
1833 } else {
1834 assert(hook_mem == NULL, "");
1835 // This is not a "slow path" call; all memory comes from the call.
1836 set_all_memory_call(call);
1837 }
1838 }
1839
1840
1841 // Replace the call with the current state of the kit.
1842 void GraphKit::replace_call(CallNode* call, Node* result) {
1843 JVMState* ejvms = NULL;
1844 if (has_exceptions()) {
1845 ejvms = transfer_exceptions_into_jvms();
1846 }
1847
1848 SafePointNode* final_state = stop();
1849
1850 // Find all the needed outputs of this call
1851 CallProjections callprojs;
1852 call->extract_projections(&callprojs, true);
1853
1854 Node* init_mem = call->in(TypeFunc::Memory);
1855 Node* final_mem = final_state->in(TypeFunc::Memory);
1856 Node* final_ctl = final_state->in(TypeFunc::Control);
1857 Node* final_io = final_state->in(TypeFunc::I_O);
1858
1859 // Replace all the old call edges with the edges from the inlining result
1860 if (callprojs.fallthrough_catchproj != NULL) {
1861 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1862 }
1863 if (callprojs.fallthrough_memproj != NULL) {
1864 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1865 }
1866 if (callprojs.fallthrough_ioproj != NULL) {
1867 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1868 }
1869
1870 // Replace the result with the new result if it exists and is used
1871 if (callprojs.resproj != NULL && result != NULL) {
1872 C->gvn_replace_by(callprojs.resproj, result);
1873 }
1874
1875 if (ejvms == NULL) {
1876 // No exception edges to simply kill off those paths
1877 if (callprojs.catchall_catchproj != NULL) {
1878 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1879 }
1880 if (callprojs.catchall_memproj != NULL) {
1881 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1882 }
1883 if (callprojs.catchall_ioproj != NULL) {
1884 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1885 }
1886 // Replace the old exception object with top
1887 if (callprojs.exobj != NULL) {
1888 C->gvn_replace_by(callprojs.exobj, C->top());
1889 }
1890 } else {
1891 GraphKit ekit(ejvms);
1892
1893 // Load my combined exception state into the kit, with all phis transformed:
1894 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1895
1896 Node* ex_oop = ekit.use_exception_state(ex_map);
1897 if (callprojs.catchall_catchproj != NULL) {
1898 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1899 }
1900 if (callprojs.catchall_memproj != NULL) {
1901 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1902 }
1903 if (callprojs.catchall_ioproj != NULL) {
1904 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1905 }
1906
1907 // Replace the old exception object with the newly created one
1908 if (callprojs.exobj != NULL) {
1909 C->gvn_replace_by(callprojs.exobj, ex_oop);
1910 }
1911 }
1912
1913 // Disconnect the call from the graph
1914 call->disconnect_inputs(NULL, C);
1915 C->gvn_replace_by(call, C->top());
1916
1917 // Clean up any MergeMems that feed other MergeMems since the
1918 // optimizer doesn't like that.
1919 if (final_mem->is_MergeMem()) {
1920 Node_List wl;
1921 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1922 Node* m = i.get();
1923 if (m->is_MergeMem() && !wl.contains(m)) {
1924 wl.push(m);
1925 }
1926 }
1927 while (wl.size() > 0) {
1928 _gvn.transform(wl.pop());
1929 }
1930 }
1931 }
1932
1933
1934 //------------------------------increment_counter------------------------------
1935 // for statistics: increment a VM counter by 1
1936
1937 void GraphKit::increment_counter(address counter_addr) {
1938 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1939 increment_counter(adr1);
1940 }
1941
1942 void GraphKit::increment_counter(Node* counter_addr) {
1943 int adr_type = Compile::AliasIdxRaw;
1944 Node* ctrl = control();
1945 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1946 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1947 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1948 }
1949
1950
1951 //------------------------------uncommon_trap----------------------------------
1952 // Bail out to the interpreter in mid-method. Implemented by calling the
1953 // uncommon_trap blob. This helper function inserts a runtime call with the
1954 // right debug info.
1955 void GraphKit::uncommon_trap(int trap_request,
1956 ciKlass* klass, const char* comment,
1957 bool must_throw,
1958 bool keep_exact_action) {
1959 if (failing()) stop();
1960 if (stopped()) return; // trap reachable?
1961
1962 // Note: If ProfileTraps is true, and if a deopt. actually
1963 // occurs here, the runtime will make sure an MDO exists. There is
1964 // no need to call method()->ensure_method_data() at this point.
1965
1966 // Set the stack pointer to the right value for reexecution:
1967 set_sp(reexecute_sp());
1968
1969 #ifdef ASSERT
1970 if (!must_throw) {
1971 // Make sure the stack has at least enough depth to execute
1972 // the current bytecode.
1973 int inputs, ignored_depth;
1974 if (compute_stack_effects(inputs, ignored_depth)) {
1975 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1976 Bytecodes::name(java_bc()), sp(), inputs));
1977 }
1978 }
1979 #endif
1980
1981 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1982 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1983
1984 switch (action) {
1985 case Deoptimization::Action_maybe_recompile:
1986 case Deoptimization::Action_reinterpret:
1987 // Temporary fix for 6529811 to allow virtual calls to be sure they
1988 // get the chance to go from mono->bi->mega
1989 if (!keep_exact_action &&
1990 Deoptimization::trap_request_index(trap_request) < 0 &&
1991 too_many_recompiles(reason)) {
1992 // This BCI is causing too many recompilations.
1993 action = Deoptimization::Action_none;
1994 trap_request = Deoptimization::make_trap_request(reason, action);
1995 } else {
1996 C->set_trap_can_recompile(true);
1997 }
1998 break;
1999 case Deoptimization::Action_make_not_entrant:
2000 C->set_trap_can_recompile(true);
2001 break;
2002 #ifdef ASSERT
2003 case Deoptimization::Action_none:
2004 case Deoptimization::Action_make_not_compilable:
2005 break;
2006 default:
2007 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
2008 break;
2009 #endif
2010 }
2011
2012 if (TraceOptoParse) {
2013 char buf[100];
2014 tty->print_cr("Uncommon trap %s at bci:%d",
2015 Deoptimization::format_trap_request(buf, sizeof(buf),
2016 trap_request), bci());
2017 }
2018
2019 CompileLog* log = C->log();
2020 if (log != NULL) {
2021 int kid = (klass == NULL)? -1: log->identify(klass);
2022 log->begin_elem("uncommon_trap bci='%d'", bci());
2023 char buf[100];
2024 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2025 trap_request));
2026 if (kid >= 0) log->print(" klass='%d'", kid);
2027 if (comment != NULL) log->print(" comment='%s'", comment);
2028 log->end_elem();
2029 }
2030
2031 // Make sure any guarding test views this path as very unlikely
2032 Node *i0 = control()->in(0);
2033 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2034 IfNode *iff = i0->as_If();
2035 float f = iff->_prob; // Get prob
2036 if (control()->Opcode() == Op_IfTrue) {
2037 if (f > PROB_UNLIKELY_MAG(4))
2038 iff->_prob = PROB_MIN;
2039 } else {
2040 if (f < PROB_LIKELY_MAG(4))
2041 iff->_prob = PROB_MAX;
2042 }
2043 }
2044
2045 // Clear out dead values from the debug info.
2046 kill_dead_locals();
2047
2048 // Now insert the uncommon trap subroutine call
2049 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2050 const TypePtr* no_memory_effects = NULL;
2051 // Pass the index of the class to be loaded
2052 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2053 (must_throw ? RC_MUST_THROW : 0),
2054 OptoRuntime::uncommon_trap_Type(),
2055 call_addr, "uncommon_trap", no_memory_effects,
2056 intcon(trap_request));
2057 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2058 "must extract request correctly from the graph");
2059 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2060
2061 call->set_req(TypeFunc::ReturnAdr, returnadr());
2062 // The debug info is the only real input to this call.
2063
2064 // Halt-and-catch fire here. The above call should never return!
2065 HaltNode* halt = new(C) HaltNode(control(), frameptr());
2066 _gvn.set_type_bottom(halt);
2067 root()->add_req(halt);
2068
2069 stop_and_kill_map();
2070 }
2071
2072
2073 //--------------------------just_allocated_object------------------------------
2074 // Report the object that was just allocated.
2075 // It must be the case that there are no intervening safepoints.
2076 // We use this to determine if an object is so "fresh" that
2077 // it does not require card marks.
2078 Node* GraphKit::just_allocated_object(Node* current_control) {
2079 if (C->recent_alloc_ctl() == current_control)
2080 return C->recent_alloc_obj();
2081 return NULL;
2082 }
2083
2084
2085 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2086 // (Note: TypeFunc::make has a cache that makes this fast.)
2087 const TypeFunc* tf = TypeFunc::make(dest_method);
2088 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2089 for (int j = 0; j < nargs; j++) {
2090 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2091 if( targ->basic_type() == T_DOUBLE ) {
2092 // If any parameters are doubles, they must be rounded before
2093 // the call, dstore_rounding does gvn.transform
2094 Node *arg = argument(j);
2095 arg = dstore_rounding(arg);
2096 set_argument(j, arg);
2097 }
2098 }
2099 }
2100
2101 /**
2102 * Record profiling data exact_kls for Node n with the type system so
2103 * that it can propagate it (speculation)
2104 *
2105 * @param n node that the type applies to
2106 * @param exact_kls type from profiling
2107 *
2108 * @return node with improved type
2109 */
2110 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) {
2111 const Type* current_type = _gvn.type(n);
2112 assert(UseTypeSpeculation, "type speculation must be on");
2113
2114 const TypeOopPtr* speculative = current_type->speculative();
2115
2116 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2117 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2118 const TypeOopPtr* xtype = tklass->as_instance_type();
2119 assert(xtype->klass_is_exact(), "Should be exact");
2120 // record the new speculative type's depth
2121 speculative = xtype->with_inline_depth(jvms()->depth());
2122 }
2123
2124 if (speculative != current_type->speculative()) {
2125 // Build a type with a speculative type (what we think we know
2126 // about the type but will need a guard when we use it)
2127 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2128 // We're changing the type, we need a new CheckCast node to carry
2129 // the new type. The new type depends on the control: what
2130 // profiling tells us is only valid from here as far as we can
2131 // tell.
2132 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2133 cast = _gvn.transform(cast);
2134 replace_in_map(n, cast);
2135 n = cast;
2136 }
2137
2138 return n;
2139 }
2140
2141 /**
2142 * Record profiling data from receiver profiling at an invoke with the
2143 * type system so that it can propagate it (speculation)
2144 *
2145 * @param n receiver node
2146 *
2147 * @return node with improved type
2148 */
2149 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2150 if (!UseTypeSpeculation) {
2151 return n;
2152 }
2153 ciKlass* exact_kls = profile_has_unique_klass();
2154 return record_profile_for_speculation(n, exact_kls);
2155 }
2156
2157 /**
2158 * Record profiling data from argument profiling at an invoke with the
2159 * type system so that it can propagate it (speculation)
2160 *
2161 * @param dest_method target method for the call
2162 * @param bc what invoke bytecode is this?
2163 */
2164 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2165 if (!UseTypeSpeculation) {
2166 return;
2167 }
2168 const TypeFunc* tf = TypeFunc::make(dest_method);
2169 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2170 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2171 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2172 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2173 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2174 ciKlass* better_type = method()->argument_profiled_type(bci(), i);
2175 if (better_type != NULL) {
2176 record_profile_for_speculation(argument(j), better_type);
2177 }
2178 i++;
2179 }
2180 }
2181 }
2182
2183 /**
2184 * Record profiling data from parameter profiling at an invoke with
2185 * the type system so that it can propagate it (speculation)
2186 */
2187 void GraphKit::record_profiled_parameters_for_speculation() {
2188 if (!UseTypeSpeculation) {
2189 return;
2190 }
2191 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2192 if (_gvn.type(local(i))->isa_oopptr()) {
2193 ciKlass* better_type = method()->parameter_profiled_type(j);
2194 if (better_type != NULL) {
2195 record_profile_for_speculation(local(i), better_type);
2196 }
2197 j++;
2198 }
2199 }
2200 }
2201
2202 void GraphKit::round_double_result(ciMethod* dest_method) {
2203 // A non-strict method may return a double value which has an extended
2204 // exponent, but this must not be visible in a caller which is 'strict'
2205 // If a strict caller invokes a non-strict callee, round a double result
2206
2207 BasicType result_type = dest_method->return_type()->basic_type();
2208 assert( method() != NULL, "must have caller context");
2209 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2210 // Destination method's return value is on top of stack
2211 // dstore_rounding() does gvn.transform
2212 Node *result = pop_pair();
2213 result = dstore_rounding(result);
2214 push_pair(result);
2215 }
2216 }
2217
2218 // rounding for strict float precision conformance
2219 Node* GraphKit::precision_rounding(Node* n) {
2220 return UseStrictFP && _method->flags().is_strict()
2221 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2222 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2223 : n;
2224 }
2225
2226 // rounding for strict double precision conformance
2227 Node* GraphKit::dprecision_rounding(Node *n) {
2228 return UseStrictFP && _method->flags().is_strict()
2229 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2230 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2231 : n;
2232 }
2233
2234 // rounding for non-strict double stores
2235 Node* GraphKit::dstore_rounding(Node* n) {
2236 return Matcher::strict_fp_requires_explicit_rounding
2237 && UseSSE <= 1
2238 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2239 : n;
2240 }
2241
2242 //=============================================================================
2243 // Generate a fast path/slow path idiom. Graph looks like:
2244 // [foo] indicates that 'foo' is a parameter
2245 //
2246 // [in] NULL
2247 // \ /
2248 // CmpP
2249 // Bool ne
2250 // If
2251 // / \
2252 // True False-<2>
2253 // / |
2254 // / cast_not_null
2255 // Load | | ^
2256 // [fast_test] | |
2257 // gvn to opt_test | |
2258 // / \ | <1>
2259 // True False |
2260 // | \\ |
2261 // [slow_call] \[fast_result]
2262 // Ctl Val \ \
2263 // | \ \
2264 // Catch <1> \ \
2265 // / \ ^ \ \
2266 // Ex No_Ex | \ \
2267 // | \ \ | \ <2> \
2268 // ... \ [slow_res] | | \ [null_result]
2269 // \ \--+--+--- | |
2270 // \ | / \ | /
2271 // --------Region Phi
2272 //
2273 //=============================================================================
2274 // Code is structured as a series of driver functions all called 'do_XXX' that
2275 // call a set of helper functions. Helper functions first, then drivers.
2276
2277 //------------------------------null_check_oop---------------------------------
2278 // Null check oop. Set null-path control into Region in slot 3.
2279 // Make a cast-not-nullness use the other not-null control. Return cast.
2280 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2281 bool never_see_null, bool safe_for_replace) {
2282 // Initial NULL check taken path
2283 (*null_control) = top();
2284 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2285
2286 // Generate uncommon_trap:
2287 if (never_see_null && (*null_control) != top()) {
2288 // If we see an unexpected null at a check-cast we record it and force a
2289 // recompile; the offending check-cast will be compiled to handle NULLs.
2290 // If we see more than one offending BCI, then all checkcasts in the
2291 // method will be compiled to handle NULLs.
2292 PreserveJVMState pjvms(this);
2293 set_control(*null_control);
2294 replace_in_map(value, null());
2295 uncommon_trap(Deoptimization::Reason_null_check,
2296 Deoptimization::Action_make_not_entrant);
2297 (*null_control) = top(); // NULL path is dead
2298 }
2299 if ((*null_control) == top() && safe_for_replace) {
2300 replace_in_map(value, cast);
2301 }
2302
2303 // Cast away null-ness on the result
2304 return cast;
2305 }
2306
2307 //------------------------------opt_iff----------------------------------------
2308 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2309 // Return slow-path control.
2310 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2311 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2312
2313 // Fast path taken; set region slot 2
2314 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2315 region->init_req(2,fast_taken); // Capture fast-control
2316
2317 // Fast path not-taken, i.e. slow path
2318 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2319 return slow_taken;
2320 }
2321
2322 //-----------------------------make_runtime_call-------------------------------
2323 Node* GraphKit::make_runtime_call(int flags,
2324 const TypeFunc* call_type, address call_addr,
2325 const char* call_name,
2326 const TypePtr* adr_type,
2327 // The following parms are all optional.
2328 // The first NULL ends the list.
2329 Node* parm0, Node* parm1,
2330 Node* parm2, Node* parm3,
2331 Node* parm4, Node* parm5,
2332 Node* parm6, Node* parm7) {
2333 // Slow-path call
2334 bool is_leaf = !(flags & RC_NO_LEAF);
2335 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2336 if (call_name == NULL) {
2337 assert(!is_leaf, "must supply name for leaf");
2338 call_name = OptoRuntime::stub_name(call_addr);
2339 }
2340 CallNode* call;
2341 if (!is_leaf) {
2342 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2343 bci(), adr_type);
2344 } else if (flags & RC_NO_FP) {
2345 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2346 } else {
2347 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2348 }
2349
2350 // The following is similar to set_edges_for_java_call,
2351 // except that the memory effects of the call are restricted to AliasIdxRaw.
2352
2353 // Slow path call has no side-effects, uses few values
2354 bool wide_in = !(flags & RC_NARROW_MEM);
2355 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2356
2357 Node* prev_mem = NULL;
2358 if (wide_in) {
2359 prev_mem = set_predefined_input_for_runtime_call(call);
2360 } else {
2361 assert(!wide_out, "narrow in => narrow out");
2362 Node* narrow_mem = memory(adr_type);
2363 prev_mem = reset_memory();
2364 map()->set_memory(narrow_mem);
2365 set_predefined_input_for_runtime_call(call);
2366 }
2367
2368 // Hook each parm in order. Stop looking at the first NULL.
2369 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2370 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2371 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2372 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2373 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2374 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2375 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2376 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2377 /* close each nested if ===> */ } } } } } } } }
2378 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2379
2380 if (!is_leaf) {
2381 // Non-leaves can block and take safepoints:
2382 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2383 }
2384 // Non-leaves can throw exceptions:
2385 if (has_io) {
2386 call->set_req(TypeFunc::I_O, i_o());
2387 }
2388
2389 if (flags & RC_UNCOMMON) {
2390 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2391 // (An "if" probability corresponds roughly to an unconditional count.
2392 // Sort of.)
2393 call->set_cnt(PROB_UNLIKELY_MAG(4));
2394 }
2395
2396 Node* c = _gvn.transform(call);
2397 assert(c == call, "cannot disappear");
2398
2399 if (wide_out) {
2400 // Slow path call has full side-effects.
2401 set_predefined_output_for_runtime_call(call);
2402 } else {
2403 // Slow path call has few side-effects, and/or sets few values.
2404 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2405 }
2406
2407 if (has_io) {
2408 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2409 }
2410 return call;
2411
2412 }
2413
2414 //------------------------------merge_memory-----------------------------------
2415 // Merge memory from one path into the current memory state.
2416 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2417 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2418 Node* old_slice = mms.force_memory();
2419 Node* new_slice = mms.memory2();
2420 if (old_slice != new_slice) {
2421 PhiNode* phi;
2422 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2423 phi = new_slice->as_Phi();
2424 #ifdef ASSERT
2425 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2426 old_slice = old_slice->in(new_path);
2427 // Caller is responsible for ensuring that any pre-existing
2428 // phis are already aware of old memory.
2429 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2430 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2431 #endif
2432 mms.set_memory(phi);
2433 } else {
2434 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2435 _gvn.set_type(phi, Type::MEMORY);
2436 phi->set_req(new_path, new_slice);
2437 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2438 }
2439 }
2440 }
2441 }
2442
2443 //------------------------------make_slow_call_ex------------------------------
2444 // Make the exception handler hookups for the slow call
2445 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2446 if (stopped()) return;
2447
2448 // Make a catch node with just two handlers: fall-through and catch-all
2449 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2450 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2451 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2452 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2453
2454 { PreserveJVMState pjvms(this);
2455 set_control(excp);
2456 set_i_o(i_o);
2457
2458 if (excp != top()) {
2459 // Create an exception state also.
2460 // Use an exact type if the caller has specified a specific exception.
2461 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2462 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2463 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2464 }
2465 }
2466
2467 // Get the no-exception control from the CatchNode.
2468 set_control(norm);
2469 }
2470
2471
2472 //-------------------------------gen_subtype_check-----------------------------
2473 // Generate a subtyping check. Takes as input the subtype and supertype.
2474 // Returns 2 values: sets the default control() to the true path and returns
2475 // the false path. Only reads invariant memory; sets no (visible) memory.
2476 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2477 // but that's not exposed to the optimizer. This call also doesn't take in an
2478 // Object; if you wish to check an Object you need to load the Object's class
2479 // prior to coming here.
2480 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2481 // Fast check for identical types, perhaps identical constants.
2482 // The types can even be identical non-constants, in cases
2483 // involving Array.newInstance, Object.clone, etc.
2484 if (subklass == superklass)
2485 return top(); // false path is dead; no test needed.
2486
2487 if (_gvn.type(superklass)->singleton()) {
2488 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2489 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2490
2491 // In the common case of an exact superklass, try to fold up the
2492 // test before generating code. You may ask, why not just generate
2493 // the code and then let it fold up? The answer is that the generated
2494 // code will necessarily include null checks, which do not always
2495 // completely fold away. If they are also needless, then they turn
2496 // into a performance loss. Example:
2497 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2498 // Here, the type of 'fa' is often exact, so the store check
2499 // of fa[1]=x will fold up, without testing the nullness of x.
2500 switch (static_subtype_check(superk, subk)) {
2501 case SSC_always_false:
2502 {
2503 Node* always_fail = control();
2504 set_control(top());
2505 return always_fail;
2506 }
2507 case SSC_always_true:
2508 return top();
2509 case SSC_easy_test:
2510 {
2511 // Just do a direct pointer compare and be done.
2512 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2513 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2514 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2515 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2516 return _gvn.transform( new(C) IfFalseNode(iff) );
2517 }
2518 case SSC_full_test:
2519 break;
2520 default:
2521 ShouldNotReachHere();
2522 }
2523 }
2524
2525 // %%% Possible further optimization: Even if the superklass is not exact,
2526 // if the subklass is the unique subtype of the superklass, the check
2527 // will always succeed. We could leave a dependency behind to ensure this.
2528
2529 // First load the super-klass's check-offset
2530 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2531 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(),
2532 TypeInt::INT, MemNode::unordered));
2533 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2534 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2535
2536 // Load from the sub-klass's super-class display list, or a 1-word cache of
2537 // the secondary superclass list, or a failing value with a sentinel offset
2538 // if the super-klass is an interface or exceptionally deep in the Java
2539 // hierarchy and we have to scan the secondary superclass list the hard way.
2540 // Worst-case type is a little odd: NULL is allowed as a result (usually
2541 // klass loads can never produce a NULL).
2542 Node *chk_off_X = ConvI2X(chk_off);
2543 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2544 // For some types like interfaces the following loadKlass is from a 1-word
2545 // cache which is mutable so can't use immutable memory. Other
2546 // types load from the super-class display table which is immutable.
2547 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2548 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2549
2550 // Compile speed common case: ARE a subtype and we canNOT fail
2551 if( superklass == nkls )
2552 return top(); // false path is dead; no test needed.
2553
2554 // See if we get an immediate positive hit. Happens roughly 83% of the
2555 // time. Test to see if the value loaded just previously from the subklass
2556 // is exactly the superklass.
2557 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2558 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2559 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2560 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2561 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2562
2563 // Compile speed common case: Check for being deterministic right now. If
2564 // chk_off is a constant and not equal to cacheoff then we are NOT a
2565 // subklass. In this case we need exactly the 1 test above and we can
2566 // return those results immediately.
2567 if (!might_be_cache) {
2568 Node* not_subtype_ctrl = control();
2569 set_control(iftrue1); // We need exactly the 1 test above
2570 return not_subtype_ctrl;
2571 }
2572
2573 // Gather the various success & failures here
2574 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2575 record_for_igvn(r_ok_subtype);
2576 RegionNode *r_not_subtype = new (C) RegionNode(3);
2577 record_for_igvn(r_not_subtype);
2578
2579 r_ok_subtype->init_req(1, iftrue1);
2580
2581 // Check for immediate negative hit. Happens roughly 11% of the time (which
2582 // is roughly 63% of the remaining cases). Test to see if the loaded
2583 // check-offset points into the subklass display list or the 1-element
2584 // cache. If it points to the display (and NOT the cache) and the display
2585 // missed then it's not a subtype.
2586 Node *cacheoff = _gvn.intcon(cacheoff_con);
2587 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2588 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2589 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2590 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2591 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2592
2593 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2594 // No performance impact (too rare) but allows sharing of secondary arrays
2595 // which has some footprint reduction.
2596 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2597 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2598 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2599 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2600 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2601
2602 // -- Roads not taken here: --
2603 // We could also have chosen to perform the self-check at the beginning
2604 // of this code sequence, as the assembler does. This would not pay off
2605 // the same way, since the optimizer, unlike the assembler, can perform
2606 // static type analysis to fold away many successful self-checks.
2607 // Non-foldable self checks work better here in second position, because
2608 // the initial primary superclass check subsumes a self-check for most
2609 // types. An exception would be a secondary type like array-of-interface,
2610 // which does not appear in its own primary supertype display.
2611 // Finally, we could have chosen to move the self-check into the
2612 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2613 // dependent manner. But it is worthwhile to have the check here,
2614 // where it can be perhaps be optimized. The cost in code space is
2615 // small (register compare, branch).
2616
2617 // Now do a linear scan of the secondary super-klass array. Again, no real
2618 // performance impact (too rare) but it's gotta be done.
2619 // Since the code is rarely used, there is no penalty for moving it
2620 // out of line, and it can only improve I-cache density.
2621 // The decision to inline or out-of-line this final check is platform
2622 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2623 Node* psc = _gvn.transform(
2624 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2625
2626 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2627 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2628 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2629 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2630 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2631
2632 // Return false path; set default control to true path.
2633 set_control( _gvn.transform(r_ok_subtype) );
2634 return _gvn.transform(r_not_subtype);
2635 }
2636
2637 //----------------------------static_subtype_check-----------------------------
2638 // Shortcut important common cases when superklass is exact:
2639 // (0) superklass is java.lang.Object (can occur in reflective code)
2640 // (1) subklass is already limited to a subtype of superklass => always ok
2641 // (2) subklass does not overlap with superklass => always fail
2642 // (3) superklass has NO subtypes and we can check with a simple compare.
2643 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2644 if (StressReflectiveCode) {
2645 return SSC_full_test; // Let caller generate the general case.
2646 }
2647
2648 if (superk == env()->Object_klass()) {
2649 return SSC_always_true; // (0) this test cannot fail
2650 }
2651
2652 ciType* superelem = superk;
2653 if (superelem->is_array_klass())
2654 superelem = superelem->as_array_klass()->base_element_type();
2655
2656 if (!subk->is_interface()) { // cannot trust static interface types yet
2657 if (subk->is_subtype_of(superk)) {
2658 return SSC_always_true; // (1) false path dead; no dynamic test needed
2659 }
2660 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2661 !superk->is_subtype_of(subk)) {
2662 return SSC_always_false;
2663 }
2664 }
2665
2666 // If casting to an instance klass, it must have no subtypes
2667 if (superk->is_interface()) {
2668 // Cannot trust interfaces yet.
2669 // %%% S.B. superk->nof_implementors() == 1
2670 } else if (superelem->is_instance_klass()) {
2671 ciInstanceKlass* ik = superelem->as_instance_klass();
2672 if (!ik->has_subklass() && !ik->is_interface()) {
2673 if (!ik->is_final()) {
2674 // Add a dependency if there is a chance of a later subclass.
2675 C->dependencies()->assert_leaf_type(ik);
2676 }
2677 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2678 }
2679 } else {
2680 // A primitive array type has no subtypes.
2681 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2682 }
2683
2684 return SSC_full_test;
2685 }
2686
2687 // Profile-driven exact type check:
2688 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2689 float prob,
2690 Node* *casted_receiver) {
2691 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2692 Node* recv_klass = load_object_klass(receiver);
2693 Node* want_klass = makecon(tklass);
2694 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2695 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2696 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2697 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2698 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2699
2700 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2701 assert(recv_xtype->klass_is_exact(), "");
2702
2703 // Subsume downstream occurrences of receiver with a cast to
2704 // recv_xtype, since now we know what the type will be.
2705 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2706 (*casted_receiver) = _gvn.transform(cast);
2707 // (User must make the replace_in_map call.)
2708
2709 return fail;
2710 }
2711
2712
2713 //------------------------------seems_never_null-------------------------------
2714 // Use null_seen information if it is available from the profile.
2715 // If we see an unexpected null at a type check we record it and force a
2716 // recompile; the offending check will be recompiled to handle NULLs.
2717 // If we see several offending BCIs, then all checks in the
2718 // method will be recompiled.
2719 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2720 if (UncommonNullCast // Cutout for this technique
2721 && obj != null() // And not the -Xcomp stupid case?
2722 && !too_many_traps(Deoptimization::Reason_null_check)
2723 ) {
2724 if (data == NULL)
2725 // Edge case: no mature data. Be optimistic here.
2726 return true;
2727 // If the profile has not seen a null, assume it won't happen.
2728 assert(java_bc() == Bytecodes::_checkcast ||
2729 java_bc() == Bytecodes::_instanceof ||
2730 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2731 return !data->as_BitData()->null_seen();
2732 }
2733 return false;
2734 }
2735
2736 //------------------------maybe_cast_profiled_receiver-------------------------
2737 // If the profile has seen exactly one type, narrow to exactly that type.
2738 // Subsequent type checks will always fold up.
2739 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2740 ciKlass* require_klass,
2741 ciKlass* spec_klass,
2742 bool safe_for_replace) {
2743 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2744
2745 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check;
2746
2747 // Make sure we haven't already deoptimized from this tactic.
2748 if (too_many_traps(reason))
2749 return NULL;
2750
2751 // (No, this isn't a call, but it's enough like a virtual call
2752 // to use the same ciMethod accessor to get the profile info...)
2753 // If we have a speculative type use it instead of profiling (which
2754 // may not help us)
2755 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2756 if (exact_kls != NULL) {// no cast failures here
2757 if (require_klass == NULL ||
2758 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2759 // If we narrow the type to match what the type profile sees or
2760 // the speculative type, we can then remove the rest of the
2761 // cast.
2762 // This is a win, even if the exact_kls is very specific,
2763 // because downstream operations, such as method calls,
2764 // will often benefit from the sharper type.
2765 Node* exact_obj = not_null_obj; // will get updated in place...
2766 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2767 &exact_obj);
2768 { PreserveJVMState pjvms(this);
2769 set_control(slow_ctl);
2770 uncommon_trap(reason,
2771 Deoptimization::Action_maybe_recompile);
2772 }
2773 if (safe_for_replace) {
2774 replace_in_map(not_null_obj, exact_obj);
2775 }
2776 return exact_obj;
2777 }
2778 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2779 }
2780
2781 return NULL;
2782 }
2783
2784 /**
2785 * Cast obj to type and emit guard unless we had too many traps here
2786 * already
2787 *
2788 * @param obj node being casted
2789 * @param type type to cast the node to
2790 * @param not_null true if we know node cannot be null
2791 */
2792 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2793 ciKlass* type,
2794 bool not_null) {
2795 // type == NULL if profiling tells us this object is always null
2796 if (type != NULL) {
2797 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2798 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check;
2799 if (!too_many_traps(null_reason) &&
2800 !too_many_traps(class_reason)) {
2801 Node* not_null_obj = NULL;
2802 // not_null is true if we know the object is not null and
2803 // there's no need for a null check
2804 if (!not_null) {
2805 Node* null_ctl = top();
2806 not_null_obj = null_check_oop(obj, &null_ctl, true, true);
2807 assert(null_ctl->is_top(), "no null control here");
2808 } else {
2809 not_null_obj = obj;
2810 }
2811
2812 Node* exact_obj = not_null_obj;
2813 ciKlass* exact_kls = type;
2814 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2815 &exact_obj);
2816 {
2817 PreserveJVMState pjvms(this);
2818 set_control(slow_ctl);
2819 uncommon_trap(class_reason,
2820 Deoptimization::Action_maybe_recompile);
2821 }
2822 replace_in_map(not_null_obj, exact_obj);
2823 obj = exact_obj;
2824 }
2825 } else {
2826 if (!too_many_traps(Deoptimization::Reason_null_assert)) {
2827 Node* exact_obj = null_assert(obj);
2828 replace_in_map(obj, exact_obj);
2829 obj = exact_obj;
2830 }
2831 }
2832 return obj;
2833 }
2834
2835 //-------------------------------gen_instanceof--------------------------------
2836 // Generate an instance-of idiom. Used by both the instance-of bytecode
2837 // and the reflective instance-of call.
2838 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2839 kill_dead_locals(); // Benefit all the uncommon traps
2840 assert( !stopped(), "dead parse path should be checked in callers" );
2841 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2842 "must check for not-null not-dead klass in callers");
2843
2844 // Make the merge point
2845 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2846 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2847 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2848 C->set_has_split_ifs(true); // Has chance for split-if optimization
2849
2850 ciProfileData* data = NULL;
2851 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2852 data = method()->method_data()->bci_to_data(bci());
2853 }
2854 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2855 && seems_never_null(obj, data));
2856
2857 // Null check; get casted pointer; set region slot 3
2858 Node* null_ctl = top();
2859 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2860
2861 // If not_null_obj is dead, only null-path is taken
2862 if (stopped()) { // Doing instance-of on a NULL?
2863 set_control(null_ctl);
2864 return intcon(0);
2865 }
2866 region->init_req(_null_path, null_ctl);
2867 phi ->init_req(_null_path, intcon(0)); // Set null path value
2868 if (null_ctl == top()) {
2869 // Do this eagerly, so that pattern matches like is_diamond_phi
2870 // will work even during parsing.
2871 assert(_null_path == PATH_LIMIT-1, "delete last");
2872 region->del_req(_null_path);
2873 phi ->del_req(_null_path);
2874 }
2875
2876 // Do we know the type check always succeed?
2877 bool known_statically = false;
2878 if (_gvn.type(superklass)->singleton()) {
2879 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2880 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2881 if (subk != NULL && subk->is_loaded()) {
2882 int static_res = static_subtype_check(superk, subk);
2883 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false);
2884 }
2885 }
2886
2887 if (known_statically && UseTypeSpeculation) {
2888 // If we know the type check always succeeds then we don't use the
2889 // profiling data at this bytecode. Don't lose it, feed it to the
2890 // type system as a speculative type.
2891 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2892 } else {
2893 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2894 // We may not have profiling here or it may not help us. If we
2895 // have a speculative type use it to perform an exact cast.
2896 ciKlass* spec_obj_type = obj_type->speculative_type();
2897 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2898 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2899 if (stopped()) { // Profile disagrees with this path.
2900 set_control(null_ctl); // Null is the only remaining possibility.
2901 return intcon(0);
2902 }
2903 if (cast_obj != NULL) {
2904 not_null_obj = cast_obj;
2905 }
2906 }
2907 }
2908
2909 // Load the object's klass
2910 Node* obj_klass = load_object_klass(not_null_obj);
2911
2912 // Generate the subtype check
2913 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2914
2915 // Plug in the success path to the general merge in slot 1.
2916 region->init_req(_obj_path, control());
2917 phi ->init_req(_obj_path, intcon(1));
2918
2919 // Plug in the failing path to the general merge in slot 2.
2920 region->init_req(_fail_path, not_subtype_ctrl);
2921 phi ->init_req(_fail_path, intcon(0));
2922
2923 // Return final merged results
2924 set_control( _gvn.transform(region) );
2925 record_for_igvn(region);
2926 return _gvn.transform(phi);
2927 }
2928
2929 //-------------------------------gen_checkcast---------------------------------
2930 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2931 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2932 // uncommon-trap paths work. Adjust stack after this call.
2933 // If failure_control is supplied and not null, it is filled in with
2934 // the control edge for the cast failure. Otherwise, an appropriate
2935 // uncommon trap or exception is thrown.
2936 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2937 Node* *failure_control) {
2938 kill_dead_locals(); // Benefit all the uncommon traps
2939 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2940 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2941
2942 // Fast cutout: Check the case that the cast is vacuously true.
2943 // This detects the common cases where the test will short-circuit
2944 // away completely. We do this before we perform the null check,
2945 // because if the test is going to turn into zero code, we don't
2946 // want a residual null check left around. (Causes a slowdown,
2947 // for example, in some objArray manipulations, such as a[i]=a[j].)
2948 if (tk->singleton()) {
2949 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2950 if (objtp != NULL && objtp->klass() != NULL) {
2951 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2952 case SSC_always_true:
2953 // If we know the type check always succeed then we don't use
2954 // the profiling data at this bytecode. Don't lose it, feed it
2955 // to the type system as a speculative type.
2956 return record_profiled_receiver_for_speculation(obj);
2957 case SSC_always_false:
2958 // It needs a null check because a null will *pass* the cast check.
2959 // A non-null value will always produce an exception.
2960 return null_assert(obj);
2961 }
2962 }
2963 }
2964
2965 ciProfileData* data = NULL;
2966 bool safe_for_replace = false;
2967 if (failure_control == NULL) { // use MDO in regular case only
2968 assert(java_bc() == Bytecodes::_aastore ||
2969 java_bc() == Bytecodes::_checkcast,
2970 "interpreter profiles type checks only for these BCs");
2971 data = method()->method_data()->bci_to_data(bci());
2972 safe_for_replace = true;
2973 }
2974
2975 // Make the merge point
2976 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2977 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
2978 Node* phi = new (C) PhiNode(region, toop);
2979 C->set_has_split_ifs(true); // Has chance for split-if optimization
2980
2981 // Use null-cast information if it is available
2982 bool never_see_null = ((failure_control == NULL) // regular case only
2983 && seems_never_null(obj, data));
2984
2985 // Null check; get casted pointer; set region slot 3
2986 Node* null_ctl = top();
2987 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2988
2989 // If not_null_obj is dead, only null-path is taken
2990 if (stopped()) { // Doing instance-of on a NULL?
2991 set_control(null_ctl);
2992 return null();
2993 }
2994 region->init_req(_null_path, null_ctl);
2995 phi ->init_req(_null_path, null()); // Set null path value
2996 if (null_ctl == top()) {
2997 // Do this eagerly, so that pattern matches like is_diamond_phi
2998 // will work even during parsing.
2999 assert(_null_path == PATH_LIMIT-1, "delete last");
3000 region->del_req(_null_path);
3001 phi ->del_req(_null_path);
3002 }
3003
3004 Node* cast_obj = NULL;
3005 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3006 // We may not have profiling here or it may not help us. If we have
3007 // a speculative type use it to perform an exact cast.
3008 ciKlass* spec_obj_type = obj_type->speculative_type();
3009 if (spec_obj_type != NULL ||
3010 (data != NULL &&
3011 // Counter has never been decremented (due to cast failure).
3012 // ...This is a reasonable thing to expect. It is true of
3013 // all casts inserted by javac to implement generic types.
3014 data->as_CounterData()->count() >= 0)) {
3015 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3016 if (cast_obj != NULL) {
3017 if (failure_control != NULL) // failure is now impossible
3018 (*failure_control) = top();
3019 // adjust the type of the phi to the exact klass:
3020 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3021 }
3022 }
3023
3024 if (cast_obj == NULL) {
3025 // Load the object's klass
3026 Node* obj_klass = load_object_klass(not_null_obj);
3027
3028 // Generate the subtype check
3029 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3030
3031 // Plug in success path into the merge
3032 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
3033 not_null_obj, toop));
3034 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3035 if (failure_control == NULL) {
3036 if (not_subtype_ctrl != top()) { // If failure is possible
3037 PreserveJVMState pjvms(this);
3038 set_control(not_subtype_ctrl);
3039 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3040 }
3041 } else {
3042 (*failure_control) = not_subtype_ctrl;
3043 }
3044 }
3045
3046 region->init_req(_obj_path, control());
3047 phi ->init_req(_obj_path, cast_obj);
3048
3049 // A merge of NULL or Casted-NotNull obj
3050 Node* res = _gvn.transform(phi);
3051
3052 // Note I do NOT always 'replace_in_map(obj,result)' here.
3053 // if( tk->klass()->can_be_primary_super() )
3054 // This means that if I successfully store an Object into an array-of-String
3055 // I 'forget' that the Object is really now known to be a String. I have to
3056 // do this because we don't have true union types for interfaces - if I store
3057 // a Baz into an array-of-Interface and then tell the optimizer it's an
3058 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3059 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3060 // replace_in_map( obj, res );
3061
3062 // Return final merged results
3063 set_control( _gvn.transform(region) );
3064 record_for_igvn(region);
3065 return res;
3066 }
3067
3068 //------------------------------next_monitor-----------------------------------
3069 // What number should be given to the next monitor?
3070 int GraphKit::next_monitor() {
3071 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3072 int next = current + C->sync_stack_slots();
3073 // Keep the toplevel high water mark current:
3074 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3075 return current;
3076 }
3077
3078 //------------------------------insert_mem_bar---------------------------------
3079 // Memory barrier to avoid floating things around
3080 // The membar serves as a pinch point between both control and all memory slices.
3081 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3082 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3083 mb->init_req(TypeFunc::Control, control());
3084 mb->init_req(TypeFunc::Memory, reset_memory());
3085 Node* membar = _gvn.transform(mb);
3086 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3087 set_all_memory_call(membar);
3088 return membar;
3089 }
3090
3091 //-------------------------insert_mem_bar_volatile----------------------------
3092 // Memory barrier to avoid floating things around
3093 // The membar serves as a pinch point between both control and memory(alias_idx).
3094 // If you want to make a pinch point on all memory slices, do not use this
3095 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3096 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3097 // When Parse::do_put_xxx updates a volatile field, it appends a series
3098 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3099 // The first membar is on the same memory slice as the field store opcode.
3100 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3101 // All the other membars (for other volatile slices, including AliasIdxBot,
3102 // which stands for all unknown volatile slices) are control-dependent
3103 // on the first membar. This prevents later volatile loads or stores
3104 // from sliding up past the just-emitted store.
3105
3106 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3107 mb->set_req(TypeFunc::Control,control());
3108 if (alias_idx == Compile::AliasIdxBot) {
3109 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3110 } else {
3111 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3112 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3113 }
3114 Node* membar = _gvn.transform(mb);
3115 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3116 if (alias_idx == Compile::AliasIdxBot) {
3117 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
3118 } else {
3119 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
3120 }
3121 return membar;
3122 }
3123
3124 //------------------------------shared_lock------------------------------------
3125 // Emit locking code.
3126 FastLockNode* GraphKit::shared_lock(Node* obj) {
3127 // bci is either a monitorenter bc or InvocationEntryBci
3128 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3129 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3130
3131 if( !GenerateSynchronizationCode )
3132 return NULL; // Not locking things?
3133 if (stopped()) // Dead monitor?
3134 return NULL;
3135
3136 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3137
3138 // Box the stack location
3139 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
3140 Node* mem = reset_memory();
3141
3142 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
3143 if (PrintPreciseBiasedLockingStatistics) {
3144 // Create the counters for this fast lock.
3145 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3146 }
3147 // Add monitor to debug info for the slow path. If we block inside the
3148 // slow path and de-opt, we need the monitor hanging around
3149 map()->push_monitor( flock );
3150
3151 const TypeFunc *tf = LockNode::lock_type();
3152 LockNode *lock = new (C) LockNode(C, tf);
3153
3154 lock->init_req( TypeFunc::Control, control() );
3155 lock->init_req( TypeFunc::Memory , mem );
3156 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3157 lock->init_req( TypeFunc::FramePtr, frameptr() );
3158 lock->init_req( TypeFunc::ReturnAdr, top() );
3159
3160 lock->init_req(TypeFunc::Parms + 0, obj);
3161 lock->init_req(TypeFunc::Parms + 1, box);
3162 lock->init_req(TypeFunc::Parms + 2, flock);
3163 add_safepoint_edges(lock);
3164
3165 lock = _gvn.transform( lock )->as_Lock();
3166
3167 // lock has no side-effects, sets few values
3168 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3169
3170 insert_mem_bar(Op_MemBarAcquireLock);
3171
3172 // Add this to the worklist so that the lock can be eliminated
3173 record_for_igvn(lock);
3174
3175 #ifndef PRODUCT
3176 if (PrintLockStatistics) {
3177 // Update the counter for this lock. Don't bother using an atomic
3178 // operation since we don't require absolute accuracy.
3179 lock->create_lock_counter(map()->jvms());
3180 increment_counter(lock->counter()->addr());
3181 }
3182 #endif
3183
3184 return flock;
3185 }
3186
3187
3188 //------------------------------shared_unlock----------------------------------
3189 // Emit unlocking code.
3190 void GraphKit::shared_unlock(Node* box, Node* obj) {
3191 // bci is either a monitorenter bc or InvocationEntryBci
3192 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3193 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3194
3195 if( !GenerateSynchronizationCode )
3196 return;
3197 if (stopped()) { // Dead monitor?
3198 map()->pop_monitor(); // Kill monitor from debug info
3199 return;
3200 }
3201
3202 // Memory barrier to avoid floating things down past the locked region
3203 insert_mem_bar(Op_MemBarReleaseLock);
3204
3205 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3206 UnlockNode *unlock = new (C) UnlockNode(C, tf);
3207 uint raw_idx = Compile::AliasIdxRaw;
3208 unlock->init_req( TypeFunc::Control, control() );
3209 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3210 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3211 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3212 unlock->init_req( TypeFunc::ReturnAdr, top() );
3213
3214 unlock->init_req(TypeFunc::Parms + 0, obj);
3215 unlock->init_req(TypeFunc::Parms + 1, box);
3216 unlock = _gvn.transform(unlock)->as_Unlock();
3217
3218 Node* mem = reset_memory();
3219
3220 // unlock has no side-effects, sets few values
3221 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3222
3223 // Kill monitor from debug info
3224 map()->pop_monitor( );
3225 }
3226
3227 //-------------------------------get_layout_helper-----------------------------
3228 // If the given klass is a constant or known to be an array,
3229 // fetch the constant layout helper value into constant_value
3230 // and return (Node*)NULL. Otherwise, load the non-constant
3231 // layout helper value, and return the node which represents it.
3232 // This two-faced routine is useful because allocation sites
3233 // almost always feature constant types.
3234 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3235 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3236 if (!StressReflectiveCode && inst_klass != NULL) {
3237 ciKlass* klass = inst_klass->klass();
3238 bool xklass = inst_klass->klass_is_exact();
3239 if (xklass || klass->is_array_klass()) {
3240 jint lhelper = klass->layout_helper();
3241 if (lhelper != Klass::_lh_neutral_value) {
3242 constant_value = lhelper;
3243 return (Node*) NULL;
3244 }
3245 }
3246 }
3247 constant_value = Klass::_lh_neutral_value; // put in a known value
3248 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3249 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3250 }
3251
3252 // We just put in an allocate/initialize with a big raw-memory effect.
3253 // Hook selected additional alias categories on the initialization.
3254 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3255 MergeMemNode* init_in_merge,
3256 Node* init_out_raw) {
3257 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3258 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3259
3260 Node* prevmem = kit.memory(alias_idx);
3261 init_in_merge->set_memory_at(alias_idx, prevmem);
3262 kit.set_memory(init_out_raw, alias_idx);
3263 }
3264
3265 //---------------------------set_output_for_allocation-------------------------
3266 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3267 const TypeOopPtr* oop_type) {
3268 int rawidx = Compile::AliasIdxRaw;
3269 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3270 add_safepoint_edges(alloc);
3271 Node* allocx = _gvn.transform(alloc);
3272 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3273 // create memory projection for i_o
3274 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3275 make_slow_call_ex(allocx, env()->Throwable_klass(), true);
3276
3277 // create a memory projection as for the normal control path
3278 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3279 set_memory(malloc, rawidx);
3280
3281 // a normal slow-call doesn't change i_o, but an allocation does
3282 // we create a separate i_o projection for the normal control path
3283 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3284 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3285
3286 // put in an initialization barrier
3287 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3288 rawoop)->as_Initialize();
3289 assert(alloc->initialization() == init, "2-way macro link must work");
3290 assert(init ->allocation() == alloc, "2-way macro link must work");
3291 {
3292 // Extract memory strands which may participate in the new object's
3293 // initialization, and source them from the new InitializeNode.
3294 // This will allow us to observe initializations when they occur,
3295 // and link them properly (as a group) to the InitializeNode.
3296 assert(init->in(InitializeNode::Memory) == malloc, "");
3297 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3298 init->set_req(InitializeNode::Memory, minit_in);
3299 record_for_igvn(minit_in); // fold it up later, if possible
3300 Node* minit_out = memory(rawidx);
3301 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3302 if (oop_type->isa_aryptr()) {
3303 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3304 int elemidx = C->get_alias_index(telemref);
3305 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3306 } else if (oop_type->isa_instptr()) {
3307 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3308 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3309 ciField* field = ik->nonstatic_field_at(i);
3310 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3311 continue; // do not bother to track really large numbers of fields
3312 // Find (or create) the alias category for this field:
3313 int fieldidx = C->alias_type(field)->index();
3314 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3315 }
3316 }
3317 }
3318
3319 // Cast raw oop to the real thing...
3320 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3321 javaoop = _gvn.transform(javaoop);
3322 C->set_recent_alloc(control(), javaoop);
3323 assert(just_allocated_object(control()) == javaoop, "just allocated");
3324
3325 #ifdef ASSERT
3326 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3327 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3328 "Ideal_allocation works");
3329 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3330 "Ideal_allocation works");
3331 if (alloc->is_AllocateArray()) {
3332 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3333 "Ideal_allocation works");
3334 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3335 "Ideal_allocation works");
3336 } else {
3337 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3338 }
3339 }
3340 #endif //ASSERT
3341
3342 return javaoop;
3343 }
3344
3345 //---------------------------new_instance--------------------------------------
3346 // This routine takes a klass_node which may be constant (for a static type)
3347 // or may be non-constant (for reflective code). It will work equally well
3348 // for either, and the graph will fold nicely if the optimizer later reduces
3349 // the type to a constant.
3350 // The optional arguments are for specialized use by intrinsics:
3351 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3352 // - If 'return_size_val', report the the total object size to the caller.
3353 Node* GraphKit::new_instance(Node* klass_node,
3354 Node* extra_slow_test,
3355 Node* *return_size_val) {
3356 // Compute size in doublewords
3357 // The size is always an integral number of doublewords, represented
3358 // as a positive bytewise size stored in the klass's layout_helper.
3359 // The layout_helper also encodes (in a low bit) the need for a slow path.
3360 jint layout_con = Klass::_lh_neutral_value;
3361 Node* layout_val = get_layout_helper(klass_node, layout_con);
3362 int layout_is_con = (layout_val == NULL);
3363
3364 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3365 // Generate the initial go-slow test. It's either ALWAYS (return a
3366 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3367 // case) a computed value derived from the layout_helper.
3368 Node* initial_slow_test = NULL;
3369 if (layout_is_con) {
3370 assert(!StressReflectiveCode, "stress mode does not use these paths");
3371 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3372 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3373
3374 } else { // reflective case
3375 // This reflective path is used by Unsafe.allocateInstance.
3376 // (It may be stress-tested by specifying StressReflectiveCode.)
3377 // Basically, we want to get into the VM is there's an illegal argument.
3378 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3379 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3380 if (extra_slow_test != intcon(0)) {
3381 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3382 }
3383 // (Macro-expander will further convert this to a Bool, if necessary.)
3384 }
3385
3386 // Find the size in bytes. This is easy; it's the layout_helper.
3387 // The size value must be valid even if the slow path is taken.
3388 Node* size = NULL;
3389 if (layout_is_con) {
3390 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3391 } else { // reflective case
3392 // This reflective path is used by clone and Unsafe.allocateInstance.
3393 size = ConvI2X(layout_val);
3394
3395 // Clear the low bits to extract layout_helper_size_in_bytes:
3396 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3397 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3398 size = _gvn.transform( new (C) AndXNode(size, mask) );
3399 }
3400 if (return_size_val != NULL) {
3401 (*return_size_val) = size;
3402 }
3403
3404 // This is a precise notnull oop of the klass.
3405 // (Actually, it need not be precise if this is a reflective allocation.)
3406 // It's what we cast the result to.
3407 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3408 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3409 const TypeOopPtr* oop_type = tklass->as_instance_type();
3410
3411 // Now generate allocation code
3412
3413 // The entire memory state is needed for slow path of the allocation
3414 // since GC and deoptimization can happened.
3415 Node *mem = reset_memory();
3416 set_all_memory(mem); // Create new memory state
3417
3418 AllocateNode* alloc
3419 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3420 control(), mem, i_o(),
3421 size, klass_node,
3422 initial_slow_test);
3423
3424 return set_output_for_allocation(alloc, oop_type);
3425 }
3426
3427 //-------------------------------new_array-------------------------------------
3428 // helper for both newarray and anewarray
3429 // The 'length' parameter is (obviously) the length of the array.
3430 // See comments on new_instance for the meaning of the other arguments.
3431 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3432 Node* length, // number of array elements
3433 int nargs, // number of arguments to push back for uncommon trap
3434 Node* *return_size_val) {
3435 jint layout_con = Klass::_lh_neutral_value;
3436 Node* layout_val = get_layout_helper(klass_node, layout_con);
3437 int layout_is_con = (layout_val == NULL);
3438
3439 if (!layout_is_con && !StressReflectiveCode &&
3440 !too_many_traps(Deoptimization::Reason_class_check)) {
3441 // This is a reflective array creation site.
3442 // Optimistically assume that it is a subtype of Object[],
3443 // so that we can fold up all the address arithmetic.
3444 layout_con = Klass::array_layout_helper(T_OBJECT);
3445 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3446 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3447 { BuildCutout unless(this, bol_lh, PROB_MAX);
3448 inc_sp(nargs);
3449 uncommon_trap(Deoptimization::Reason_class_check,
3450 Deoptimization::Action_maybe_recompile);
3451 }
3452 layout_val = NULL;
3453 layout_is_con = true;
3454 }
3455
3456 // Generate the initial go-slow test. Make sure we do not overflow
3457 // if length is huge (near 2Gig) or negative! We do not need
3458 // exact double-words here, just a close approximation of needed
3459 // double-words. We can't add any offset or rounding bits, lest we
3460 // take a size -1 of bytes and make it positive. Use an unsigned
3461 // compare, so negative sizes look hugely positive.
3462 int fast_size_limit = FastAllocateSizeLimit;
3463 if (layout_is_con) {
3464 assert(!StressReflectiveCode, "stress mode does not use these paths");
3465 // Increase the size limit if we have exact knowledge of array type.
3466 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3467 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3468 }
3469
3470 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3471 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3472 if (initial_slow_test->is_Bool()) {
3473 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3474 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3475 }
3476
3477 // --- Size Computation ---
3478 // array_size = round_to_heap(array_header + (length << elem_shift));
3479 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3480 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3481 // The rounding mask is strength-reduced, if possible.
3482 int round_mask = MinObjAlignmentInBytes - 1;
3483 Node* header_size = NULL;
3484 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3485 // (T_BYTE has the weakest alignment and size restrictions...)
3486 if (layout_is_con) {
3487 int hsize = Klass::layout_helper_header_size(layout_con);
3488 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3489 BasicType etype = Klass::layout_helper_element_type(layout_con);
3490 if ((round_mask & ~right_n_bits(eshift)) == 0)
3491 round_mask = 0; // strength-reduce it if it goes away completely
3492 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3493 assert(header_size_min <= hsize, "generic minimum is smallest");
3494 header_size_min = hsize;
3495 header_size = intcon(hsize + round_mask);
3496 } else {
3497 Node* hss = intcon(Klass::_lh_header_size_shift);
3498 Node* hsm = intcon(Klass::_lh_header_size_mask);
3499 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3500 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3501 Node* mask = intcon(round_mask);
3502 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3503 }
3504
3505 Node* elem_shift = NULL;
3506 if (layout_is_con) {
3507 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3508 if (eshift != 0)
3509 elem_shift = intcon(eshift);
3510 } else {
3511 // There is no need to mask or shift this value.
3512 // The semantics of LShiftINode include an implicit mask to 0x1F.
3513 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3514 elem_shift = layout_val;
3515 }
3516
3517 // Transition to native address size for all offset calculations:
3518 Node* lengthx = ConvI2X(length);
3519 Node* headerx = ConvI2X(header_size);
3520 #ifdef _LP64
3521 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3522 if (tllen != NULL && tllen->_lo < 0) {
3523 // Add a manual constraint to a positive range. Cf. array_element_address.
3524 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3525 if (size_max > tllen->_hi) size_max = tllen->_hi;
3526 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3527 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon));
3528 }
3529 }
3530 #endif
3531
3532 // Combine header size (plus rounding) and body size. Then round down.
3533 // This computation cannot overflow, because it is used only in two
3534 // places, one where the length is sharply limited, and the other
3535 // after a successful allocation.
3536 Node* abody = lengthx;
3537 if (elem_shift != NULL)
3538 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3539 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3540 if (round_mask != 0) {
3541 Node* mask = MakeConX(~round_mask);
3542 size = _gvn.transform( new(C) AndXNode(size, mask) );
3543 }
3544 // else if round_mask == 0, the size computation is self-rounding
3545
3546 if (return_size_val != NULL) {
3547 // This is the size
3548 (*return_size_val) = size;
3549 }
3550
3551 // Now generate allocation code
3552
3553 // The entire memory state is needed for slow path of the allocation
3554 // since GC and deoptimization can happened.
3555 Node *mem = reset_memory();
3556 set_all_memory(mem); // Create new memory state
3557
3558 // Create the AllocateArrayNode and its result projections
3559 AllocateArrayNode* alloc
3560 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3561 control(), mem, i_o(),
3562 size, klass_node,
3563 initial_slow_test,
3564 length);
3565
3566 // Cast to correct type. Note that the klass_node may be constant or not,
3567 // and in the latter case the actual array type will be inexact also.
3568 // (This happens via a non-constant argument to inline_native_newArray.)
3569 // In any case, the value of klass_node provides the desired array type.
3570 const TypeInt* length_type = _gvn.find_int_type(length);
3571 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3572 if (ary_type->isa_aryptr() && length_type != NULL) {
3573 // Try to get a better type than POS for the size
3574 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3575 }
3576
3577 Node* javaoop = set_output_for_allocation(alloc, ary_type);
3578
3579 // Cast length on remaining path to be as narrow as possible
3580 if (map()->find_edge(length) >= 0) {
3581 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3582 if (ccast != length) {
3583 _gvn.set_type_bottom(ccast);
3584 record_for_igvn(ccast);
3585 replace_in_map(length, ccast);
3586 }
3587 }
3588
3589 return javaoop;
3590 }
3591
3592 // The following "Ideal_foo" functions are placed here because they recognize
3593 // the graph shapes created by the functions immediately above.
3594
3595 //---------------------------Ideal_allocation----------------------------------
3596 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3597 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3598 if (ptr == NULL) { // reduce dumb test in callers
3599 return NULL;
3600 }
3601 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3602 ptr = ptr->in(1);
3603 if (ptr == NULL) return NULL;
3604 }
3605 // Return NULL for allocations with several casts:
3606 // j.l.reflect.Array.newInstance(jobject, jint)
3607 // Object.clone()
3608 // to keep more precise type from last cast.
3609 if (ptr->is_Proj()) {
3610 Node* allo = ptr->in(0);
3611 if (allo != NULL && allo->is_Allocate()) {
3612 return allo->as_Allocate();
3613 }
3614 }
3615 // Report failure to match.
3616 return NULL;
3617 }
3618
3619 // Fancy version which also strips off an offset (and reports it to caller).
3620 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3621 intptr_t& offset) {
3622 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3623 if (base == NULL) return NULL;
3624 return Ideal_allocation(base, phase);
3625 }
3626
3627 // Trace Initialize <- Proj[Parm] <- Allocate
3628 AllocateNode* InitializeNode::allocation() {
3629 Node* rawoop = in(InitializeNode::RawAddress);
3630 if (rawoop->is_Proj()) {
3631 Node* alloc = rawoop->in(0);
3632 if (alloc->is_Allocate()) {
3633 return alloc->as_Allocate();
3634 }
3635 }
3636 return NULL;
3637 }
3638
3639 // Trace Allocate -> Proj[Parm] -> Initialize
3640 InitializeNode* AllocateNode::initialization() {
3641 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3642 if (rawoop == NULL) return NULL;
3643 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3644 Node* init = rawoop->fast_out(i);
3645 if (init->is_Initialize()) {
3646 assert(init->as_Initialize()->allocation() == this, "2-way link");
3647 return init->as_Initialize();
3648 }
3649 }
3650 return NULL;
3651 }
3652
3653 //----------------------------- loop predicates ---------------------------
3654
3655 //------------------------------add_predicate_impl----------------------------
3656 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3657 // Too many traps seen?
3658 if (too_many_traps(reason)) {
3659 #ifdef ASSERT
3660 if (TraceLoopPredicate) {
3661 int tc = C->trap_count(reason);
3662 tty->print("too many traps=%s tcount=%d in ",
3663 Deoptimization::trap_reason_name(reason), tc);
3664 method()->print(); // which method has too many predicate traps
3665 tty->cr();
3666 }
3667 #endif
3668 // We cannot afford to take more traps here,
3669 // do not generate predicate.
3670 return;
3671 }
3672
3673 Node *cont = _gvn.intcon(1);
3674 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3675 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3676 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3677 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3678 C->add_predicate_opaq(opq);
3679 {
3680 PreserveJVMState pjvms(this);
3681 set_control(iffalse);
3682 inc_sp(nargs);
3683 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3684 }
3685 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3686 set_control(iftrue);
3687 }
3688
3689 //------------------------------add_predicate---------------------------------
3690 void GraphKit::add_predicate(int nargs) {
3691 if (UseLoopPredicate) {
3692 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3693 }
3694 // loop's limit check predicate should be near the loop.
3695 if (LoopLimitCheck) {
3696 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3697 }
3698 }
3699
3700 //----------------------------- store barriers ----------------------------
3701 #define __ ideal.
3702
3703 void GraphKit::sync_kit(IdealKit& ideal) {
3704 set_all_memory(__ merged_memory());
3705 set_i_o(__ i_o());
3706 set_control(__ ctrl());
3707 }
3708
3709 void GraphKit::final_sync(IdealKit& ideal) {
3710 // Final sync IdealKit and graphKit.
3711 sync_kit(ideal);
3712 }
3713
3714 // vanilla/CMS post barrier
3715 // Insert a write-barrier store. This is to let generational GC work; we have
3716 // to flag all oop-stores before the next GC point.
3717 void GraphKit::write_barrier_post(Node* oop_store,
3718 Node* obj,
3719 Node* adr,
3720 uint adr_idx,
3721 Node* val,
3722 bool use_precise) {
3723 // No store check needed if we're storing a NULL or an old object
3724 // (latter case is probably a string constant). The concurrent
3725 // mark sweep garbage collector, however, needs to have all nonNull
3726 // oop updates flagged via card-marks.
3727 if (val != NULL && val->is_Con()) {
3728 // must be either an oop or NULL
3729 const Type* t = val->bottom_type();
3730 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3731 // stores of null never (?) need barriers
3732 return;
3733 }
3734
3735 if (use_ReduceInitialCardMarks()
3736 && obj == just_allocated_object(control())) {
3737 // We can skip marks on a freshly-allocated object in Eden.
3738 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3739 // That routine informs GC to take appropriate compensating steps,
3740 // upon a slow-path allocation, so as to make this card-mark
3741 // elision safe.
3742 return;
3743 }
3744
3745 if (!use_precise) {
3746 // All card marks for a (non-array) instance are in one place:
3747 adr = obj;
3748 }
3749 // (Else it's an array (or unknown), and we want more precise card marks.)
3750 assert(adr != NULL, "");
3751
3752 IdealKit ideal(this, true);
3753
3754 // Convert the pointer to an int prior to doing math on it
3755 Node* cast = __ CastPX(__ ctrl(), adr);
3756
3757 // Divide by card size
3758 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3759 "Only one we handle so far.");
3760 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3761
3762 // Combine card table base and card offset
3763 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3764
3765 // Get the alias_index for raw card-mark memory
3766 int adr_type = Compile::AliasIdxRaw;
3767 Node* zero = __ ConI(0); // Dirty card value
3768 BasicType bt = T_BYTE;
3769
3770 if (UseCondCardMark) {
3771 // The classic GC reference write barrier is typically implemented
3772 // as a store into the global card mark table. Unfortunately
3773 // unconditional stores can result in false sharing and excessive
3774 // coherence traffic as well as false transactional aborts.
3775 // UseCondCardMark enables MP "polite" conditional card mark
3776 // stores. In theory we could relax the load from ctrl() to
3777 // no_ctrl, but that doesn't buy much latitude.
3778 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3779 __ if_then(card_val, BoolTest::ne, zero);
3780 }
3781
3782 // Smash zero into card
3783 if( !UseConcMarkSweepGC ) {
3784 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release);
3785 } else {
3786 // Specialized path for CM store barrier
3787 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3788 }
3789
3790 if (UseCondCardMark) {
3791 __ end_if();
3792 }
3793
3794 // Final sync IdealKit and GraphKit.
3795 final_sync(ideal);
3796 }
3797
3798 // G1 pre/post barriers
3799 void GraphKit::g1_write_barrier_pre(bool do_load,
3800 Node* obj,
3801 Node* adr,
3802 uint alias_idx,
3803 Node* val,
3804 const TypeOopPtr* val_type,
3805 Node* pre_val,
3806 BasicType bt) {
3807
3808 // Some sanity checks
3809 // Note: val is unused in this routine.
3810
3811 if (do_load) {
3812 // We need to generate the load of the previous value
3813 assert(obj != NULL, "must have a base");
3814 assert(adr != NULL, "where are loading from?");
3815 assert(pre_val == NULL, "loaded already?");
3816 assert(val_type != NULL, "need a type");
3817 } else {
3818 // In this case both val_type and alias_idx are unused.
3819 assert(pre_val != NULL, "must be loaded already");
3820 // Nothing to be done if pre_val is null.
3821 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3822 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3823 }
3824 assert(bt == T_OBJECT, "or we shouldn't be here");
3825
3826 IdealKit ideal(this, true);
3827
3828 Node* tls = __ thread(); // ThreadLocalStorage
3829
3830 Node* no_ctrl = NULL;
3831 Node* no_base = __ top();
3832 Node* zero = __ ConI(0);
3833 Node* zeroX = __ ConX(0);
3834
3835 float likely = PROB_LIKELY(0.999);
3836 float unlikely = PROB_UNLIKELY(0.999);
3837
3838 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3839 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3840
3841 // Offsets into the thread
3842 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3843 PtrQueue::byte_offset_of_active());
3844 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3845 PtrQueue::byte_offset_of_index());
3846 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3847 PtrQueue::byte_offset_of_buf());
3848
3849 // Now the actual pointers into the thread
3850 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3851 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3852 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3853
3854 // Now some of the values
3855 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3856
3857 // if (!marking)
3858 __ if_then(marking, BoolTest::ne, zero, unlikely); {
3859 BasicType index_bt = TypeX_X->basic_type();
3860 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3861 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3862
3863 if (do_load) {
3864 // load original value
3865 // alias_idx correct??
3866 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3867 }
3868
3869 // if (pre_val != NULL)
3870 __ if_then(pre_val, BoolTest::ne, null()); {
3871 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3872
3873 // is the queue for this thread full?
3874 __ if_then(index, BoolTest::ne, zeroX, likely); {
3875
3876 // decrement the index
3877 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3878
3879 // Now get the buffer location we will log the previous value into and store it
3880 Node *log_addr = __ AddP(no_base, buffer, next_index);
3881 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
3882 // update the index
3883 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
3884
3885 } __ else_(); {
3886
3887 // logging buffer is full, call the runtime
3888 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3889 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3890 } __ end_if(); // (!index)
3891 } __ end_if(); // (pre_val != NULL)
3892 } __ end_if(); // (!marking)
3893
3894 // Final sync IdealKit and GraphKit.
3895 final_sync(ideal);
3896 }
3897
3898 //
3899 // Update the card table and add card address to the queue
3900 //
3901 void GraphKit::g1_mark_card(IdealKit& ideal,
3902 Node* card_adr,
3903 Node* oop_store,
3904 uint oop_alias_idx,
3905 Node* index,
3906 Node* index_adr,
3907 Node* buffer,
3908 const TypeFunc* tf) {
3909
3910 Node* zero = __ ConI(0);
3911 Node* zeroX = __ ConX(0);
3912 Node* no_base = __ top();
3913 BasicType card_bt = T_BYTE;
3914 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3915 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3916
3917 // Now do the queue work
3918 __ if_then(index, BoolTest::ne, zeroX); {
3919
3920 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3921 Node* log_addr = __ AddP(no_base, buffer, next_index);
3922
3923 // Order, see storeCM.
3924 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
3925 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
3926
3927 } __ else_(); {
3928 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3929 } __ end_if();
3930
3931 }
3932
3933 void GraphKit::g1_write_barrier_post(Node* oop_store,
3934 Node* obj,
3935 Node* adr,
3936 uint alias_idx,
3937 Node* val,
3938 BasicType bt,
3939 bool use_precise) {
3940 // If we are writing a NULL then we need no post barrier
3941
3942 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3943 // Must be NULL
3944 const Type* t = val->bottom_type();
3945 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3946 // No post barrier if writing NULLx
3947 return;
3948 }
3949
3950 if (!use_precise) {
3951 // All card marks for a (non-array) instance are in one place:
3952 adr = obj;
3953 }
3954 // (Else it's an array (or unknown), and we want more precise card marks.)
3955 assert(adr != NULL, "");
3956
3957 IdealKit ideal(this, true);
3958
3959 Node* tls = __ thread(); // ThreadLocalStorage
3960
3961 Node* no_base = __ top();
3962 float likely = PROB_LIKELY(0.999);
3963 float unlikely = PROB_UNLIKELY(0.999);
3964 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
3965 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
3966 Node* zeroX = __ ConX(0);
3967
3968 // Get the alias_index for raw card-mark memory
3969 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3970
3971 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3972
3973 // Offsets into the thread
3974 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3975 PtrQueue::byte_offset_of_index());
3976 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3977 PtrQueue::byte_offset_of_buf());
3978
3979 // Pointers into the thread
3980
3981 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3982 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3983
3984 // Now some values
3985 // Use ctrl to avoid hoisting these values past a safepoint, which could
3986 // potentially reset these fields in the JavaThread.
3987 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
3988 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3989
3990 // Convert the store obj pointer to an int prior to doing math on it
3991 // Must use ctrl to prevent "integerized oop" existing across safepoint
3992 Node* cast = __ CastPX(__ ctrl(), adr);
3993
3994 // Divide pointer by card size
3995 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3996
3997 // Combine card table base and card offset
3998 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3999
4000 // If we know the value being stored does it cross regions?
4001
4002 if (val != NULL) {
4003 // Does the store cause us to cross regions?
4004
4005 // Should be able to do an unsigned compare of region_size instead of
4006 // and extra shift. Do we have an unsigned compare??
4007 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4008 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4009
4010 // if (xor_res == 0) same region so skip
4011 __ if_then(xor_res, BoolTest::ne, zeroX); {
4012
4013 // No barrier if we are storing a NULL
4014 __ if_then(val, BoolTest::ne, null(), unlikely); {
4015
4016 // Ok must mark the card if not already dirty
4017
4018 // load the original value of the card
4019 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4020
4021 __ if_then(card_val, BoolTest::ne, young_card); {
4022 sync_kit(ideal);
4023 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4024 insert_mem_bar(Op_MemBarVolatile, oop_store);
4025 __ sync_kit(this);
4026
4027 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4028 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4029 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4030 } __ end_if();
4031 } __ end_if();
4032 } __ end_if();
4033 } __ end_if();
4034 } else {
4035 // Object.clone() instrinsic uses this path.
4036 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4037 }
4038
4039 // Final sync IdealKit and GraphKit.
4040 final_sync(ideal);
4041 }
4042 #undef __
4043
4044
4045
4046 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
4047 if (java_lang_String::has_offset_field()) {
4048 int offset_offset = java_lang_String::offset_offset_in_bytes();
4049 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4050 false, NULL, 0);
4051 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4052 int offset_field_idx = C->get_alias_index(offset_field_type);
4053 return make_load(ctrl,
4054 basic_plus_adr(str, str, offset_offset),
4055 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered);
4056 } else {
4057 return intcon(0);
4058 }
4059 }
4060
4061 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4062 if (java_lang_String::has_count_field()) {
4063 int count_offset = java_lang_String::count_offset_in_bytes();
4064 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4065 false, NULL, 0);
4066 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4067 int count_field_idx = C->get_alias_index(count_field_type);
4068 return make_load(ctrl,
4069 basic_plus_adr(str, str, count_offset),
4070 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered);
4071 } else {
4072 return load_array_length(load_String_value(ctrl, str));
4073 }
4074 }
4075
4076 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4077 int value_offset = java_lang_String::value_offset_in_bytes();
4078 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4079 false, NULL, 0);
4080 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4081 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4082 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
4083 ciTypeArrayKlass::make(T_CHAR), true, 0);
4084 int value_field_idx = C->get_alias_index(value_field_type);
4085 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4086 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4087 // String.value field is known to be @Stable.
4088 if (UseImplicitStableValues) {
4089 load = cast_array_to_stable(load, value_type);
4090 }
4091 return load;
4092 }
4093
4094 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
4095 int offset_offset = java_lang_String::offset_offset_in_bytes();
4096 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4097 false, NULL, 0);
4098 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4099 int offset_field_idx = C->get_alias_index(offset_field_type);
4100 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
4101 value, T_INT, offset_field_idx, MemNode::unordered);
4102 }
4103
4104 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4105 int value_offset = java_lang_String::value_offset_in_bytes();
4106 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4107 false, NULL, 0);
4108 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4109
4110 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4111 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered);
4112 }
4113
4114 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
4115 int count_offset = java_lang_String::count_offset_in_bytes();
4116 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4117 false, NULL, 0);
4118 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4119 int count_field_idx = C->get_alias_index(count_field_type);
4120 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
4121 value, T_INT, count_field_idx, MemNode::unordered);
4122 }
4123
4124 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4125 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4126 // assumption of CCP analysis.
4127 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
4128 }