1 /*
2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
28 #include "gc_implementation/g1/heapRegion.hpp"
29 #include "gc_interface/collectedHeap.hpp"
30 #include "memory/barrierSet.hpp"
31 #include "memory/cardTableModRefBS.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/graphKit.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "runtime/deoptimization.hpp"
41 #include "runtime/sharedRuntime.hpp"
42
43 //----------------------------GraphKit-----------------------------------------
44 // Main utility constructor.
45 GraphKit::GraphKit(JVMState* jvms)
46 : Phase(Phase::Parser),
47 _env(C->env()),
48 _gvn(*C->initial_gvn())
49 {
50 _exceptions = jvms->map()->next_exception();
51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
52 set_jvms(jvms);
53 }
54
55 // Private constructor for parser.
56 GraphKit::GraphKit()
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn())
60 {
61 _exceptions = NULL;
62 set_map(NULL);
63 debug_only(_sp = -99);
64 debug_only(set_bci(-99));
65 }
66
67
68
69 //---------------------------clean_stack---------------------------------------
70 // Clear away rubbish from the stack area of the JVM state.
71 // This destroys any arguments that may be waiting on the stack.
72 void GraphKit::clean_stack(int from_sp) {
73 SafePointNode* map = this->map();
74 JVMState* jvms = this->jvms();
75 int stk_size = jvms->stk_size();
76 int stkoff = jvms->stkoff();
77 Node* top = this->top();
78 for (int i = from_sp; i < stk_size; i++) {
79 if (map->in(stkoff + i) != top) {
80 map->set_req(stkoff + i, top);
81 }
82 }
83 }
84
85
86 //--------------------------------sync_jvms-----------------------------------
87 // Make sure our current jvms agrees with our parse state.
88 JVMState* GraphKit::sync_jvms() const {
89 JVMState* jvms = this->jvms();
90 jvms->set_bci(bci()); // Record the new bci in the JVMState
91 jvms->set_sp(sp()); // Record the new sp in the JVMState
92 assert(jvms_in_sync(), "jvms is now in sync");
93 return jvms;
94 }
95
96 //--------------------------------sync_jvms_for_reexecute---------------------
97 // Make sure our current jvms agrees with our parse state. This version
98 // uses the reexecute_sp for reexecuting bytecodes.
99 JVMState* GraphKit::sync_jvms_for_reexecute() {
100 JVMState* jvms = this->jvms();
101 jvms->set_bci(bci()); // Record the new bci in the JVMState
102 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
103 return jvms;
104 }
105
106 #ifdef ASSERT
107 bool GraphKit::jvms_in_sync() const {
108 Parse* parse = is_Parse();
109 if (parse == NULL) {
110 if (bci() != jvms()->bci()) return false;
111 if (sp() != (int)jvms()->sp()) return false;
112 return true;
113 }
114 if (jvms()->method() != parse->method()) return false;
115 if (jvms()->bci() != parse->bci()) return false;
116 int jvms_sp = jvms()->sp();
117 if (jvms_sp != parse->sp()) return false;
118 int jvms_depth = jvms()->depth();
119 if (jvms_depth != parse->depth()) return false;
120 return true;
121 }
122
123 // Local helper checks for special internal merge points
124 // used to accumulate and merge exception states.
125 // They are marked by the region's in(0) edge being the map itself.
126 // Such merge points must never "escape" into the parser at large,
127 // until they have been handed to gvn.transform.
128 static bool is_hidden_merge(Node* reg) {
129 if (reg == NULL) return false;
130 if (reg->is_Phi()) {
131 reg = reg->in(0);
132 if (reg == NULL) return false;
133 }
134 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
135 }
136
137 void GraphKit::verify_map() const {
138 if (map() == NULL) return; // null map is OK
139 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
140 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
141 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
142 }
143
144 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
145 assert(ex_map->next_exception() == NULL, "not already part of a chain");
146 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
147 }
148 #endif
149
150 //---------------------------stop_and_kill_map---------------------------------
151 // Set _map to NULL, signalling a stop to further bytecode execution.
152 // First smash the current map's control to a constant, to mark it dead.
153 void GraphKit::stop_and_kill_map() {
154 SafePointNode* dead_map = stop();
155 if (dead_map != NULL) {
156 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
157 assert(dead_map->is_killed(), "must be so marked");
158 }
159 }
160
161
162 //--------------------------------stopped--------------------------------------
163 // Tell if _map is NULL, or control is top.
164 bool GraphKit::stopped() {
165 if (map() == NULL) return true;
166 else if (control() == top()) return true;
167 else return false;
168 }
169
170
171 //-----------------------------has_ex_handler----------------------------------
172 // Tell if this method or any caller method has exception handlers.
173 bool GraphKit::has_ex_handler() {
174 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
175 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
176 return true;
177 }
178 }
179 return false;
180 }
181
182 //------------------------------save_ex_oop------------------------------------
183 // Save an exception without blowing stack contents or other JVM state.
184 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
185 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
186 ex_map->add_req(ex_oop);
187 debug_only(verify_exception_state(ex_map));
188 }
189
190 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
191 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
192 Node* ex_oop = ex_map->in(ex_map->req()-1);
193 if (clear_it) ex_map->del_req(ex_map->req()-1);
194 return ex_oop;
195 }
196
197 //-----------------------------saved_ex_oop------------------------------------
198 // Recover a saved exception from its map.
199 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
200 return common_saved_ex_oop(ex_map, false);
201 }
202
203 //--------------------------clear_saved_ex_oop---------------------------------
204 // Erase a previously saved exception from its map.
205 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
206 return common_saved_ex_oop(ex_map, true);
207 }
208
209 #ifdef ASSERT
210 //---------------------------has_saved_ex_oop----------------------------------
211 // Erase a previously saved exception from its map.
212 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
213 return ex_map->req() == ex_map->jvms()->endoff()+1;
214 }
215 #endif
216
217 //-------------------------make_exception_state--------------------------------
218 // Turn the current JVM state into an exception state, appending the ex_oop.
219 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
220 sync_jvms();
221 SafePointNode* ex_map = stop(); // do not manipulate this map any more
222 set_saved_ex_oop(ex_map, ex_oop);
223 return ex_map;
224 }
225
226
227 //--------------------------add_exception_state--------------------------------
228 // Add an exception to my list of exceptions.
229 void GraphKit::add_exception_state(SafePointNode* ex_map) {
230 if (ex_map == NULL || ex_map->control() == top()) {
231 return;
232 }
233 #ifdef ASSERT
234 verify_exception_state(ex_map);
235 if (has_exceptions()) {
236 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
237 }
238 #endif
239
240 // If there is already an exception of exactly this type, merge with it.
241 // In particular, null-checks and other low-level exceptions common up here.
242 Node* ex_oop = saved_ex_oop(ex_map);
243 const Type* ex_type = _gvn.type(ex_oop);
244 if (ex_oop == top()) {
245 // No action needed.
246 return;
247 }
248 assert(ex_type->isa_instptr(), "exception must be an instance");
249 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
250 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
251 // We check sp also because call bytecodes can generate exceptions
252 // both before and after arguments are popped!
253 if (ex_type2 == ex_type
254 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
255 combine_exception_states(ex_map, e2);
256 return;
257 }
258 }
259
260 // No pre-existing exception of the same type. Chain it on the list.
261 push_exception_state(ex_map);
262 }
263
264 //-----------------------add_exception_states_from-----------------------------
265 void GraphKit::add_exception_states_from(JVMState* jvms) {
266 SafePointNode* ex_map = jvms->map()->next_exception();
267 if (ex_map != NULL) {
268 jvms->map()->set_next_exception(NULL);
269 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
270 next_map = ex_map->next_exception();
271 ex_map->set_next_exception(NULL);
272 add_exception_state(ex_map);
273 }
274 }
275 }
276
277 //-----------------------transfer_exceptions_into_jvms-------------------------
278 JVMState* GraphKit::transfer_exceptions_into_jvms() {
279 if (map() == NULL) {
280 // We need a JVMS to carry the exceptions, but the map has gone away.
281 // Create a scratch JVMS, cloned from any of the exception states...
282 if (has_exceptions()) {
283 _map = _exceptions;
284 _map = clone_map();
285 _map->set_next_exception(NULL);
286 clear_saved_ex_oop(_map);
287 debug_only(verify_map());
288 } else {
289 // ...or created from scratch
290 JVMState* jvms = new (C) JVMState(_method, NULL);
291 jvms->set_bci(_bci);
292 jvms->set_sp(_sp);
293 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms));
294 set_jvms(jvms);
295 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
296 set_all_memory(top());
297 while (map()->req() < jvms->endoff()) map()->add_req(top());
298 }
299 // (This is a kludge, in case you didn't notice.)
300 set_control(top());
301 }
302 JVMState* jvms = sync_jvms();
303 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
304 jvms->map()->set_next_exception(_exceptions);
305 _exceptions = NULL; // done with this set of exceptions
306 return jvms;
307 }
308
309 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
310 assert(is_hidden_merge(dstphi), "must be a special merge node");
311 assert(is_hidden_merge(srcphi), "must be a special merge node");
312 uint limit = srcphi->req();
313 for (uint i = PhiNode::Input; i < limit; i++) {
314 dstphi->add_req(srcphi->in(i));
315 }
316 }
317 static inline void add_one_req(Node* dstphi, Node* src) {
318 assert(is_hidden_merge(dstphi), "must be a special merge node");
319 assert(!is_hidden_merge(src), "must not be a special merge node");
320 dstphi->add_req(src);
321 }
322
323 //-----------------------combine_exception_states------------------------------
324 // This helper function combines exception states by building phis on a
325 // specially marked state-merging region. These regions and phis are
326 // untransformed, and can build up gradually. The region is marked by
327 // having a control input of its exception map, rather than NULL. Such
328 // regions do not appear except in this function, and in use_exception_state.
329 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
330 if (failing()) return; // dying anyway...
331 JVMState* ex_jvms = ex_map->_jvms;
332 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
333 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
334 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
335 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
336 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
337 assert(ex_map->req() == phi_map->req(), "matching maps");
338 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
339 Node* hidden_merge_mark = root();
340 Node* region = phi_map->control();
341 MergeMemNode* phi_mem = phi_map->merged_memory();
342 MergeMemNode* ex_mem = ex_map->merged_memory();
343 if (region->in(0) != hidden_merge_mark) {
344 // The control input is not (yet) a specially-marked region in phi_map.
345 // Make it so, and build some phis.
346 region = new (C) RegionNode(2);
347 _gvn.set_type(region, Type::CONTROL);
348 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
349 region->init_req(1, phi_map->control());
350 phi_map->set_control(region);
351 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
352 record_for_igvn(io_phi);
353 _gvn.set_type(io_phi, Type::ABIO);
354 phi_map->set_i_o(io_phi);
355 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
356 Node* m = mms.memory();
357 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
358 record_for_igvn(m_phi);
359 _gvn.set_type(m_phi, Type::MEMORY);
360 mms.set_memory(m_phi);
361 }
362 }
363
364 // Either or both of phi_map and ex_map might already be converted into phis.
365 Node* ex_control = ex_map->control();
366 // if there is special marking on ex_map also, we add multiple edges from src
367 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
368 // how wide was the destination phi_map, originally?
369 uint orig_width = region->req();
370
371 if (add_multiple) {
372 add_n_reqs(region, ex_control);
373 add_n_reqs(phi_map->i_o(), ex_map->i_o());
374 } else {
375 // ex_map has no merges, so we just add single edges everywhere
376 add_one_req(region, ex_control);
377 add_one_req(phi_map->i_o(), ex_map->i_o());
378 }
379 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
380 if (mms.is_empty()) {
381 // get a copy of the base memory, and patch some inputs into it
382 const TypePtr* adr_type = mms.adr_type(C);
383 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
384 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
385 mms.set_memory(phi);
386 // Prepare to append interesting stuff onto the newly sliced phi:
387 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
388 }
389 // Append stuff from ex_map:
390 if (add_multiple) {
391 add_n_reqs(mms.memory(), mms.memory2());
392 } else {
393 add_one_req(mms.memory(), mms.memory2());
394 }
395 }
396 uint limit = ex_map->req();
397 for (uint i = TypeFunc::Parms; i < limit; i++) {
398 // Skip everything in the JVMS after tos. (The ex_oop follows.)
399 if (i == tos) i = ex_jvms->monoff();
400 Node* src = ex_map->in(i);
401 Node* dst = phi_map->in(i);
402 if (src != dst) {
403 PhiNode* phi;
404 if (dst->in(0) != region) {
405 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
406 record_for_igvn(phi);
407 _gvn.set_type(phi, phi->type());
408 phi_map->set_req(i, dst);
409 // Prepare to append interesting stuff onto the new phi:
410 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
411 } else {
412 assert(dst->is_Phi(), "nobody else uses a hidden region");
413 phi = dst->as_Phi();
414 }
415 if (add_multiple && src->in(0) == ex_control) {
416 // Both are phis.
417 add_n_reqs(dst, src);
418 } else {
419 while (dst->req() < region->req()) add_one_req(dst, src);
420 }
421 const Type* srctype = _gvn.type(src);
422 if (phi->type() != srctype) {
423 const Type* dsttype = phi->type()->meet_speculative(srctype);
424 if (phi->type() != dsttype) {
425 phi->set_type(dsttype);
426 _gvn.set_type(phi, dsttype);
427 }
428 }
429 }
430 }
431 }
432
433 //--------------------------use_exception_state--------------------------------
434 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
435 if (failing()) { stop(); return top(); }
436 Node* region = phi_map->control();
437 Node* hidden_merge_mark = root();
438 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
439 Node* ex_oop = clear_saved_ex_oop(phi_map);
440 if (region->in(0) == hidden_merge_mark) {
441 // Special marking for internal ex-states. Process the phis now.
442 region->set_req(0, region); // now it's an ordinary region
443 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
444 // Note: Setting the jvms also sets the bci and sp.
445 set_control(_gvn.transform(region));
446 uint tos = jvms()->stkoff() + sp();
447 for (uint i = 1; i < tos; i++) {
448 Node* x = phi_map->in(i);
449 if (x->in(0) == region) {
450 assert(x->is_Phi(), "expected a special phi");
451 phi_map->set_req(i, _gvn.transform(x));
452 }
453 }
454 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
455 Node* x = mms.memory();
456 if (x->in(0) == region) {
457 assert(x->is_Phi(), "nobody else uses a hidden region");
458 mms.set_memory(_gvn.transform(x));
459 }
460 }
461 if (ex_oop->in(0) == region) {
462 assert(ex_oop->is_Phi(), "expected a special phi");
463 ex_oop = _gvn.transform(ex_oop);
464 }
465 } else {
466 set_jvms(phi_map->jvms());
467 }
468
469 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
470 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
471 return ex_oop;
472 }
473
474 //---------------------------------java_bc-------------------------------------
475 Bytecodes::Code GraphKit::java_bc() const {
476 ciMethod* method = this->method();
477 int bci = this->bci();
478 if (method != NULL && bci != InvocationEntryBci)
479 return method->java_code_at_bci(bci);
480 else
481 return Bytecodes::_illegal;
482 }
483
484 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
485 bool must_throw) {
486 // if the exception capability is set, then we will generate code
487 // to check the JavaThread.should_post_on_exceptions flag to see
488 // if we actually need to report exception events (for this
489 // thread). If we don't need to report exception events, we will
490 // take the normal fast path provided by add_exception_events. If
491 // exception event reporting is enabled for this thread, we will
492 // take the uncommon_trap in the BuildCutout below.
493
494 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
495 Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
496 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
497 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
498
499 // Test the should_post_on_exceptions_flag vs. 0
500 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
501 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
502
503 // Branch to slow_path if should_post_on_exceptions_flag was true
504 { BuildCutout unless(this, tst, PROB_MAX);
505 // Do not try anything fancy if we're notifying the VM on every throw.
506 // Cf. case Bytecodes::_athrow in parse2.cpp.
507 uncommon_trap(reason, Deoptimization::Action_none,
508 (ciKlass*)NULL, (char*)NULL, must_throw);
509 }
510
511 }
512
513 //------------------------------builtin_throw----------------------------------
514 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
515 bool must_throw = true;
516
517 if (env()->jvmti_can_post_on_exceptions()) {
518 // check if we must post exception events, take uncommon trap if so
519 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
520 // here if should_post_on_exceptions is false
521 // continue on with the normal codegen
522 }
523
524 // If this particular condition has not yet happened at this
525 // bytecode, then use the uncommon trap mechanism, and allow for
526 // a future recompilation if several traps occur here.
527 // If the throw is hot, try to use a more complicated inline mechanism
528 // which keeps execution inside the compiled code.
529 bool treat_throw_as_hot = false;
530 ciMethodData* md = method()->method_data();
531
532 if (ProfileTraps) {
533 if (too_many_traps(reason)) {
534 treat_throw_as_hot = true;
535 }
536 // (If there is no MDO at all, assume it is early in
537 // execution, and that any deopts are part of the
538 // startup transient, and don't need to be remembered.)
539
540 // Also, if there is a local exception handler, treat all throws
541 // as hot if there has been at least one in this method.
542 if (C->trap_count(reason) != 0
543 && method()->method_data()->trap_count(reason) != 0
544 && has_ex_handler()) {
545 treat_throw_as_hot = true;
546 }
547 }
548
549 // If this throw happens frequently, an uncommon trap might cause
550 // a performance pothole. If there is a local exception handler,
551 // and if this particular bytecode appears to be deoptimizing often,
552 // let us handle the throw inline, with a preconstructed instance.
553 // Note: If the deopt count has blown up, the uncommon trap
554 // runtime is going to flush this nmethod, not matter what.
555 if (treat_throw_as_hot
556 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
557 // If the throw is local, we use a pre-existing instance and
558 // punt on the backtrace. This would lead to a missing backtrace
559 // (a repeat of 4292742) if the backtrace object is ever asked
560 // for its backtrace.
561 // Fixing this remaining case of 4292742 requires some flavor of
562 // escape analysis. Leave that for the future.
563 ciInstance* ex_obj = NULL;
564 switch (reason) {
565 case Deoptimization::Reason_null_check:
566 ex_obj = env()->NullPointerException_instance();
567 break;
568 case Deoptimization::Reason_div0_check:
569 ex_obj = env()->ArithmeticException_instance();
570 break;
571 case Deoptimization::Reason_range_check:
572 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
573 break;
574 case Deoptimization::Reason_class_check:
575 if (java_bc() == Bytecodes::_aastore) {
576 ex_obj = env()->ArrayStoreException_instance();
577 } else {
578 ex_obj = env()->ClassCastException_instance();
579 }
580 break;
581 }
582 if (failing()) { stop(); return; } // exception allocation might fail
583 if (ex_obj != NULL) {
584 // Cheat with a preallocated exception object.
585 if (C->log() != NULL)
586 C->log()->elem("hot_throw preallocated='1' reason='%s'",
587 Deoptimization::trap_reason_name(reason));
588 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
589 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
590
591 // Clear the detail message of the preallocated exception object.
592 // Weblogic sometimes mutates the detail message of exceptions
593 // using reflection.
594 int offset = java_lang_Throwable::get_detailMessage_offset();
595 const TypePtr* adr_typ = ex_con->add_offset(offset);
596
597 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
598 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
599 // Conservatively release stores of object references.
600 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
601
602 add_exception_state(make_exception_state(ex_node));
603 return;
604 }
605 }
606
607 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
608 // It won't be much cheaper than bailing to the interp., since we'll
609 // have to pass up all the debug-info, and the runtime will have to
610 // create the stack trace.
611
612 // Usual case: Bail to interpreter.
613 // Reserve the right to recompile if we haven't seen anything yet.
614
615 assert(!Deoptimization::reason_is_speculate(reason), "unsupported");
616 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
617 if (treat_throw_as_hot
618 && (method()->method_data()->trap_recompiled_at(bci(), NULL)
619 || C->too_many_traps(reason))) {
620 // We cannot afford to take more traps here. Suffer in the interpreter.
621 if (C->log() != NULL)
622 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
623 Deoptimization::trap_reason_name(reason),
624 C->trap_count(reason));
625 action = Deoptimization::Action_none;
626 }
627
628 // "must_throw" prunes the JVM state to include only the stack, if there
629 // are no local exception handlers. This should cut down on register
630 // allocation time and code size, by drastically reducing the number
631 // of in-edges on the call to the uncommon trap.
632
633 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
634 }
635
636
637 //----------------------------PreserveJVMState---------------------------------
638 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
639 debug_only(kit->verify_map());
640 _kit = kit;
641 _map = kit->map(); // preserve the map
642 _sp = kit->sp();
643 kit->set_map(clone_map ? kit->clone_map() : NULL);
644 Compile::current()->inc_preserve_jvm_state();
645 #ifdef ASSERT
646 _bci = kit->bci();
647 Parse* parser = kit->is_Parse();
648 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
649 _block = block;
650 #endif
651 }
652 PreserveJVMState::~PreserveJVMState() {
653 GraphKit* kit = _kit;
654 #ifdef ASSERT
655 assert(kit->bci() == _bci, "bci must not shift");
656 Parse* parser = kit->is_Parse();
657 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
658 assert(block == _block, "block must not shift");
659 #endif
660 kit->set_map(_map);
661 kit->set_sp(_sp);
662 Compile::current()->dec_preserve_jvm_state();
663 }
664
665
666 //-----------------------------BuildCutout-------------------------------------
667 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
668 : PreserveJVMState(kit)
669 {
670 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
671 SafePointNode* outer_map = _map; // preserved map is caller's
672 SafePointNode* inner_map = kit->map();
673 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
674 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
675 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
676 }
677 BuildCutout::~BuildCutout() {
678 GraphKit* kit = _kit;
679 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
680 }
681
682 //---------------------------PreserveReexecuteState----------------------------
683 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
684 assert(!kit->stopped(), "must call stopped() before");
685 _kit = kit;
686 _sp = kit->sp();
687 _reexecute = kit->jvms()->_reexecute;
688 }
689 PreserveReexecuteState::~PreserveReexecuteState() {
690 if (_kit->stopped()) return;
691 _kit->jvms()->_reexecute = _reexecute;
692 _kit->set_sp(_sp);
693 }
694
695 //------------------------------clone_map--------------------------------------
696 // Implementation of PreserveJVMState
697 //
698 // Only clone_map(...) here. If this function is only used in the
699 // PreserveJVMState class we may want to get rid of this extra
700 // function eventually and do it all there.
701
702 SafePointNode* GraphKit::clone_map() {
703 if (map() == NULL) return NULL;
704
705 // Clone the memory edge first
706 Node* mem = MergeMemNode::make(C, map()->memory());
707 gvn().set_type_bottom(mem);
708
709 SafePointNode *clonemap = (SafePointNode*)map()->clone();
710 JVMState* jvms = this->jvms();
711 JVMState* clonejvms = jvms->clone_shallow(C);
712 clonemap->set_memory(mem);
713 clonemap->set_jvms(clonejvms);
714 clonejvms->set_map(clonemap);
715 record_for_igvn(clonemap);
716 gvn().set_type_bottom(clonemap);
717 return clonemap;
718 }
719
720
721 //-----------------------------set_map_clone-----------------------------------
722 void GraphKit::set_map_clone(SafePointNode* m) {
723 _map = m;
724 _map = clone_map();
725 _map->set_next_exception(NULL);
726 debug_only(verify_map());
727 }
728
729
730 //----------------------------kill_dead_locals---------------------------------
731 // Detect any locals which are known to be dead, and force them to top.
732 void GraphKit::kill_dead_locals() {
733 // Consult the liveness information for the locals. If any
734 // of them are unused, then they can be replaced by top(). This
735 // should help register allocation time and cut down on the size
736 // of the deoptimization information.
737
738 // This call is made from many of the bytecode handling
739 // subroutines called from the Big Switch in do_one_bytecode.
740 // Every bytecode which might include a slow path is responsible
741 // for killing its dead locals. The more consistent we
742 // are about killing deads, the fewer useless phis will be
743 // constructed for them at various merge points.
744
745 // bci can be -1 (InvocationEntryBci). We return the entry
746 // liveness for the method.
747
748 if (method() == NULL || method()->code_size() == 0) {
749 // We are building a graph for a call to a native method.
750 // All locals are live.
751 return;
752 }
753
754 ResourceMark rm;
755
756 // Consult the liveness information for the locals. If any
757 // of them are unused, then they can be replaced by top(). This
758 // should help register allocation time and cut down on the size
759 // of the deoptimization information.
760 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
761
762 int len = (int)live_locals.size();
763 assert(len <= jvms()->loc_size(), "too many live locals");
764 for (int local = 0; local < len; local++) {
765 if (!live_locals.at(local)) {
766 set_local(local, top());
767 }
768 }
769 }
770
771 #ifdef ASSERT
772 //-------------------------dead_locals_are_killed------------------------------
773 // Return true if all dead locals are set to top in the map.
774 // Used to assert "clean" debug info at various points.
775 bool GraphKit::dead_locals_are_killed() {
776 if (method() == NULL || method()->code_size() == 0) {
777 // No locals need to be dead, so all is as it should be.
778 return true;
779 }
780
781 // Make sure somebody called kill_dead_locals upstream.
782 ResourceMark rm;
783 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
784 if (jvms->loc_size() == 0) continue; // no locals to consult
785 SafePointNode* map = jvms->map();
786 ciMethod* method = jvms->method();
787 int bci = jvms->bci();
788 if (jvms == this->jvms()) {
789 bci = this->bci(); // it might not yet be synched
790 }
791 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
792 int len = (int)live_locals.size();
793 if (!live_locals.is_valid() || len == 0)
794 // This method is trivial, or is poisoned by a breakpoint.
795 return true;
796 assert(len == jvms->loc_size(), "live map consistent with locals map");
797 for (int local = 0; local < len; local++) {
798 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
799 if (PrintMiscellaneous && (Verbose || WizardMode)) {
800 tty->print_cr("Zombie local %d: ", local);
801 jvms->dump();
802 }
803 return false;
804 }
805 }
806 }
807 return true;
808 }
809
810 #endif //ASSERT
811
812 // Helper function for enforcing certain bytecodes to reexecute if
813 // deoptimization happens
814 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
815 ciMethod* cur_method = jvms->method();
816 int cur_bci = jvms->bci();
817 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
818 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
819 return Interpreter::bytecode_should_reexecute(code) ||
820 is_anewarray && code == Bytecodes::_multianewarray;
821 // Reexecute _multianewarray bytecode which was replaced with
822 // sequence of [a]newarray. See Parse::do_multianewarray().
823 //
824 // Note: interpreter should not have it set since this optimization
825 // is limited by dimensions and guarded by flag so in some cases
826 // multianewarray() runtime calls will be generated and
827 // the bytecode should not be reexecutes (stack will not be reset).
828 } else
829 return false;
830 }
831
832 // Helper function for adding JVMState and debug information to node
833 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
834 // Add the safepoint edges to the call (or other safepoint).
835
836 // Make sure dead locals are set to top. This
837 // should help register allocation time and cut down on the size
838 // of the deoptimization information.
839 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
840
841 // Walk the inline list to fill in the correct set of JVMState's
842 // Also fill in the associated edges for each JVMState.
843
844 // If the bytecode needs to be reexecuted we need to put
845 // the arguments back on the stack.
846 const bool should_reexecute = jvms()->should_reexecute();
847 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
848
849 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
850 // undefined if the bci is different. This is normal for Parse but it
851 // should not happen for LibraryCallKit because only one bci is processed.
852 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
853 "in LibraryCallKit the reexecute bit should not change");
854
855 // If we are guaranteed to throw, we can prune everything but the
856 // input to the current bytecode.
857 bool can_prune_locals = false;
858 uint stack_slots_not_pruned = 0;
859 int inputs = 0, depth = 0;
860 if (must_throw) {
861 assert(method() == youngest_jvms->method(), "sanity");
862 if (compute_stack_effects(inputs, depth)) {
863 can_prune_locals = true;
864 stack_slots_not_pruned = inputs;
865 }
866 }
867
868 if (env()->should_retain_local_variables()) {
869 // At any safepoint, this method can get breakpointed, which would
870 // then require an immediate deoptimization.
871 can_prune_locals = false; // do not prune locals
872 stack_slots_not_pruned = 0;
873 }
874
875 // do not scribble on the input jvms
876 JVMState* out_jvms = youngest_jvms->clone_deep(C);
877 call->set_jvms(out_jvms); // Start jvms list for call node
878
879 // For a known set of bytecodes, the interpreter should reexecute them if
880 // deoptimization happens. We set the reexecute state for them here
881 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
882 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
883 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
884 }
885
886 // Presize the call:
887 DEBUG_ONLY(uint non_debug_edges = call->req());
888 call->add_req_batch(top(), youngest_jvms->debug_depth());
889 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
890
891 // Set up edges so that the call looks like this:
892 // Call [state:] ctl io mem fptr retadr
893 // [parms:] parm0 ... parmN
894 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
895 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
896 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
897 // Note that caller debug info precedes callee debug info.
898
899 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
900 uint debug_ptr = call->req();
901
902 // Loop over the map input edges associated with jvms, add them
903 // to the call node, & reset all offsets to match call node array.
904 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
905 uint debug_end = debug_ptr;
906 uint debug_start = debug_ptr - in_jvms->debug_size();
907 debug_ptr = debug_start; // back up the ptr
908
909 uint p = debug_start; // walks forward in [debug_start, debug_end)
910 uint j, k, l;
911 SafePointNode* in_map = in_jvms->map();
912 out_jvms->set_map(call);
913
914 if (can_prune_locals) {
915 assert(in_jvms->method() == out_jvms->method(), "sanity");
916 // If the current throw can reach an exception handler in this JVMS,
917 // then we must keep everything live that can reach that handler.
918 // As a quick and dirty approximation, we look for any handlers at all.
919 if (in_jvms->method()->has_exception_handlers()) {
920 can_prune_locals = false;
921 }
922 }
923
924 // Add the Locals
925 k = in_jvms->locoff();
926 l = in_jvms->loc_size();
927 out_jvms->set_locoff(p);
928 if (!can_prune_locals) {
929 for (j = 0; j < l; j++)
930 call->set_req(p++, in_map->in(k+j));
931 } else {
932 p += l; // already set to top above by add_req_batch
933 }
934
935 // Add the Expression Stack
936 k = in_jvms->stkoff();
937 l = in_jvms->sp();
938 out_jvms->set_stkoff(p);
939 if (!can_prune_locals) {
940 for (j = 0; j < l; j++)
941 call->set_req(p++, in_map->in(k+j));
942 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
943 // Divide stack into {S0,...,S1}, where S0 is set to top.
944 uint s1 = stack_slots_not_pruned;
945 stack_slots_not_pruned = 0; // for next iteration
946 if (s1 > l) s1 = l;
947 uint s0 = l - s1;
948 p += s0; // skip the tops preinstalled by add_req_batch
949 for (j = s0; j < l; j++)
950 call->set_req(p++, in_map->in(k+j));
951 } else {
952 p += l; // already set to top above by add_req_batch
953 }
954
955 // Add the Monitors
956 k = in_jvms->monoff();
957 l = in_jvms->mon_size();
958 out_jvms->set_monoff(p);
959 for (j = 0; j < l; j++)
960 call->set_req(p++, in_map->in(k+j));
961
962 // Copy any scalar object fields.
963 k = in_jvms->scloff();
964 l = in_jvms->scl_size();
965 out_jvms->set_scloff(p);
966 for (j = 0; j < l; j++)
967 call->set_req(p++, in_map->in(k+j));
968
969 // Finish the new jvms.
970 out_jvms->set_endoff(p);
971
972 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
973 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
974 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
975 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
976 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
977 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
978
979 // Update the two tail pointers in parallel.
980 out_jvms = out_jvms->caller();
981 in_jvms = in_jvms->caller();
982 }
983
984 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
985
986 // Test the correctness of JVMState::debug_xxx accessors:
987 assert(call->jvms()->debug_start() == non_debug_edges, "");
988 assert(call->jvms()->debug_end() == call->req(), "");
989 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
990 }
991
992 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
993 Bytecodes::Code code = java_bc();
994 if (code == Bytecodes::_wide) {
995 code = method()->java_code_at_bci(bci() + 1);
996 }
997
998 BasicType rtype = T_ILLEGAL;
999 int rsize = 0;
1000
1001 if (code != Bytecodes::_illegal) {
1002 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1003 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1004 if (rtype < T_CONFLICT)
1005 rsize = type2size[rtype];
1006 }
1007
1008 switch (code) {
1009 case Bytecodes::_illegal:
1010 return false;
1011
1012 case Bytecodes::_ldc:
1013 case Bytecodes::_ldc_w:
1014 case Bytecodes::_ldc2_w:
1015 inputs = 0;
1016 break;
1017
1018 case Bytecodes::_dup: inputs = 1; break;
1019 case Bytecodes::_dup_x1: inputs = 2; break;
1020 case Bytecodes::_dup_x2: inputs = 3; break;
1021 case Bytecodes::_dup2: inputs = 2; break;
1022 case Bytecodes::_dup2_x1: inputs = 3; break;
1023 case Bytecodes::_dup2_x2: inputs = 4; break;
1024 case Bytecodes::_swap: inputs = 2; break;
1025 case Bytecodes::_arraylength: inputs = 1; break;
1026
1027 case Bytecodes::_getstatic:
1028 case Bytecodes::_putstatic:
1029 case Bytecodes::_getfield:
1030 case Bytecodes::_putfield:
1031 {
1032 bool ignored_will_link;
1033 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1034 int size = field->type()->size();
1035 bool is_get = (depth >= 0), is_static = (depth & 1);
1036 inputs = (is_static ? 0 : 1);
1037 if (is_get) {
1038 depth = size - inputs;
1039 } else {
1040 inputs += size; // putxxx pops the value from the stack
1041 depth = - inputs;
1042 }
1043 }
1044 break;
1045
1046 case Bytecodes::_invokevirtual:
1047 case Bytecodes::_invokespecial:
1048 case Bytecodes::_invokestatic:
1049 case Bytecodes::_invokedynamic:
1050 case Bytecodes::_invokeinterface:
1051 {
1052 bool ignored_will_link;
1053 ciSignature* declared_signature = NULL;
1054 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1055 assert(declared_signature != NULL, "cannot be null");
1056 inputs = declared_signature->arg_size_for_bc(code);
1057 int size = declared_signature->return_type()->size();
1058 depth = size - inputs;
1059 }
1060 break;
1061
1062 case Bytecodes::_multianewarray:
1063 {
1064 ciBytecodeStream iter(method());
1065 iter.reset_to_bci(bci());
1066 iter.next();
1067 inputs = iter.get_dimensions();
1068 assert(rsize == 1, "");
1069 depth = rsize - inputs;
1070 }
1071 break;
1072
1073 case Bytecodes::_ireturn:
1074 case Bytecodes::_lreturn:
1075 case Bytecodes::_freturn:
1076 case Bytecodes::_dreturn:
1077 case Bytecodes::_areturn:
1078 assert(rsize = -depth, "");
1079 inputs = rsize;
1080 break;
1081
1082 case Bytecodes::_jsr:
1083 case Bytecodes::_jsr_w:
1084 inputs = 0;
1085 depth = 1; // S.B. depth=1, not zero
1086 break;
1087
1088 default:
1089 // bytecode produces a typed result
1090 inputs = rsize - depth;
1091 assert(inputs >= 0, "");
1092 break;
1093 }
1094
1095 #ifdef ASSERT
1096 // spot check
1097 int outputs = depth + inputs;
1098 assert(outputs >= 0, "sanity");
1099 switch (code) {
1100 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1101 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1102 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1103 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1104 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1105 }
1106 #endif //ASSERT
1107
1108 return true;
1109 }
1110
1111
1112
1113 //------------------------------basic_plus_adr---------------------------------
1114 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1115 // short-circuit a common case
1116 if (offset == intcon(0)) return ptr;
1117 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1118 }
1119
1120 Node* GraphKit::ConvI2L(Node* offset) {
1121 // short-circuit a common case
1122 jint offset_con = find_int_con(offset, Type::OffsetBot);
1123 if (offset_con != Type::OffsetBot) {
1124 return longcon((jlong) offset_con);
1125 }
1126 return _gvn.transform( new (C) ConvI2LNode(offset));
1127 }
1128 Node* GraphKit::ConvL2I(Node* offset) {
1129 // short-circuit a common case
1130 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1131 if (offset_con != (jlong)Type::OffsetBot) {
1132 return intcon((int) offset_con);
1133 }
1134 return _gvn.transform( new (C) ConvL2INode(offset));
1135 }
1136
1137 //-------------------------load_object_klass-----------------------------------
1138 Node* GraphKit::load_object_klass(Node* obj) {
1139 // Special-case a fresh allocation to avoid building nodes:
1140 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1141 if (akls != NULL) return akls;
1142 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1143 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1144 }
1145
1146 //-------------------------load_array_length-----------------------------------
1147 Node* GraphKit::load_array_length(Node* array) {
1148 // Special-case a fresh allocation to avoid building nodes:
1149 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1150 Node *alen;
1151 if (alloc == NULL) {
1152 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1153 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1154 } else {
1155 alen = alloc->Ideal_length();
1156 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1157 if (ccast != alen) {
1158 alen = _gvn.transform(ccast);
1159 }
1160 }
1161 return alen;
1162 }
1163
1164 //------------------------------do_null_check----------------------------------
1165 // Helper function to do a NULL pointer check. Returned value is
1166 // the incoming address with NULL casted away. You are allowed to use the
1167 // not-null value only if you are control dependent on the test.
1168 extern int explicit_null_checks_inserted,
1169 explicit_null_checks_elided;
1170 Node* GraphKit::null_check_common(Node* value, BasicType type,
1171 // optional arguments for variations:
1172 bool assert_null,
1173 Node* *null_control) {
1174 assert(!assert_null || null_control == NULL, "not both at once");
1175 if (stopped()) return top();
1176 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1177 // For some performance testing, we may wish to suppress null checking.
1178 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1179 return value;
1180 }
1181 explicit_null_checks_inserted++;
1182
1183 // Construct NULL check
1184 Node *chk = NULL;
1185 switch(type) {
1186 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1187 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1188 case T_ARRAY : // fall through
1189 type = T_OBJECT; // simplify further tests
1190 case T_OBJECT : {
1191 const Type *t = _gvn.type( value );
1192
1193 const TypeOopPtr* tp = t->isa_oopptr();
1194 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1195 // Only for do_null_check, not any of its siblings:
1196 && !assert_null && null_control == NULL) {
1197 // Usually, any field access or invocation on an unloaded oop type
1198 // will simply fail to link, since the statically linked class is
1199 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1200 // the static class is loaded but the sharper oop type is not.
1201 // Rather than checking for this obscure case in lots of places,
1202 // we simply observe that a null check on an unloaded class
1203 // will always be followed by a nonsense operation, so we
1204 // can just issue the uncommon trap here.
1205 // Our access to the unloaded class will only be correct
1206 // after it has been loaded and initialized, which requires
1207 // a trip through the interpreter.
1208 #ifndef PRODUCT
1209 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1210 #endif
1211 uncommon_trap(Deoptimization::Reason_unloaded,
1212 Deoptimization::Action_reinterpret,
1213 tp->klass(), "!loaded");
1214 return top();
1215 }
1216
1217 if (assert_null) {
1218 // See if the type is contained in NULL_PTR.
1219 // If so, then the value is already null.
1220 if (t->higher_equal(TypePtr::NULL_PTR)) {
1221 explicit_null_checks_elided++;
1222 return value; // Elided null assert quickly!
1223 }
1224 } else {
1225 // See if mixing in the NULL pointer changes type.
1226 // If so, then the NULL pointer was not allowed in the original
1227 // type. In other words, "value" was not-null.
1228 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1229 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1230 explicit_null_checks_elided++;
1231 return value; // Elided null check quickly!
1232 }
1233 }
1234 chk = new (C) CmpPNode( value, null() );
1235 break;
1236 }
1237
1238 default:
1239 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1240 }
1241 assert(chk != NULL, "sanity check");
1242 chk = _gvn.transform(chk);
1243
1244 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1245 BoolNode *btst = new (C) BoolNode( chk, btest);
1246 Node *tst = _gvn.transform( btst );
1247
1248 //-----------
1249 // if peephole optimizations occurred, a prior test existed.
1250 // If a prior test existed, maybe it dominates as we can avoid this test.
1251 if (tst != btst && type == T_OBJECT) {
1252 // At this point we want to scan up the CFG to see if we can
1253 // find an identical test (and so avoid this test altogether).
1254 Node *cfg = control();
1255 int depth = 0;
1256 while( depth < 16 ) { // Limit search depth for speed
1257 if( cfg->Opcode() == Op_IfTrue &&
1258 cfg->in(0)->in(1) == tst ) {
1259 // Found prior test. Use "cast_not_null" to construct an identical
1260 // CastPP (and hence hash to) as already exists for the prior test.
1261 // Return that casted value.
1262 if (assert_null) {
1263 replace_in_map(value, null());
1264 return null(); // do not issue the redundant test
1265 }
1266 Node *oldcontrol = control();
1267 set_control(cfg);
1268 Node *res = cast_not_null(value);
1269 set_control(oldcontrol);
1270 explicit_null_checks_elided++;
1271 return res;
1272 }
1273 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1274 if (cfg == NULL) break; // Quit at region nodes
1275 depth++;
1276 }
1277 }
1278
1279 //-----------
1280 // Branch to failure if null
1281 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1282 Deoptimization::DeoptReason reason;
1283 if (assert_null)
1284 reason = Deoptimization::Reason_null_assert;
1285 else if (type == T_OBJECT)
1286 reason = Deoptimization::Reason_null_check;
1287 else
1288 reason = Deoptimization::Reason_div0_check;
1289
1290 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1291 // ciMethodData::has_trap_at will return a conservative -1 if any
1292 // must-be-null assertion has failed. This could cause performance
1293 // problems for a method after its first do_null_assert failure.
1294 // Consider using 'Reason_class_check' instead?
1295
1296 // To cause an implicit null check, we set the not-null probability
1297 // to the maximum (PROB_MAX). For an explicit check the probability
1298 // is set to a smaller value.
1299 if (null_control != NULL || too_many_traps(reason)) {
1300 // probability is less likely
1301 ok_prob = PROB_LIKELY_MAG(3);
1302 } else if (!assert_null &&
1303 (ImplicitNullCheckThreshold > 0) &&
1304 method() != NULL &&
1305 (method()->method_data()->trap_count(reason)
1306 >= (uint)ImplicitNullCheckThreshold)) {
1307 ok_prob = PROB_LIKELY_MAG(3);
1308 }
1309
1310 if (null_control != NULL) {
1311 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1312 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1313 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1314 if (null_true == top())
1315 explicit_null_checks_elided++;
1316 (*null_control) = null_true;
1317 } else {
1318 BuildCutout unless(this, tst, ok_prob);
1319 // Check for optimizer eliding test at parse time
1320 if (stopped()) {
1321 // Failure not possible; do not bother making uncommon trap.
1322 explicit_null_checks_elided++;
1323 } else if (assert_null) {
1324 uncommon_trap(reason,
1325 Deoptimization::Action_make_not_entrant,
1326 NULL, "assert_null");
1327 } else {
1328 replace_in_map(value, zerocon(type));
1329 builtin_throw(reason);
1330 }
1331 }
1332
1333 // Must throw exception, fall-thru not possible?
1334 if (stopped()) {
1335 return top(); // No result
1336 }
1337
1338 if (assert_null) {
1339 // Cast obj to null on this path.
1340 replace_in_map(value, zerocon(type));
1341 return zerocon(type);
1342 }
1343
1344 // Cast obj to not-null on this path, if there is no null_control.
1345 // (If there is a null_control, a non-null value may come back to haunt us.)
1346 if (type == T_OBJECT) {
1347 Node* cast = cast_not_null(value, false);
1348 if (null_control == NULL || (*null_control) == top())
1349 replace_in_map(value, cast);
1350 value = cast;
1351 }
1352
1353 return value;
1354 }
1355
1356
1357 //------------------------------cast_not_null----------------------------------
1358 // Cast obj to not-null on this path
1359 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1360 const Type *t = _gvn.type(obj);
1361 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1362 // Object is already not-null?
1363 if( t == t_not_null ) return obj;
1364
1365 Node *cast = new (C) CastPPNode(obj,t_not_null);
1366 cast->init_req(0, control());
1367 cast = _gvn.transform( cast );
1368
1369 // Scan for instances of 'obj' in the current JVM mapping.
1370 // These instances are known to be not-null after the test.
1371 if (do_replace_in_map)
1372 replace_in_map(obj, cast);
1373
1374 return cast; // Return casted value
1375 }
1376
1377
1378 //--------------------------replace_in_map-------------------------------------
1379 void GraphKit::replace_in_map(Node* old, Node* neww) {
1380 if (old == neww) {
1381 return;
1382 }
1383
1384 map()->replace_edge(old, neww);
1385
1386 // Note: This operation potentially replaces any edge
1387 // on the map. This includes locals, stack, and monitors
1388 // of the current (innermost) JVM state.
1389
1390 if (!ReplaceInParentMaps) {
1391 return;
1392 }
1393
1394 // PreserveJVMState doesn't do a deep copy so we can't modify
1395 // parents
1396 if (Compile::current()->has_preserve_jvm_state()) {
1397 return;
1398 }
1399
1400 Parse* parser = is_Parse();
1401 bool progress = true;
1402 Node* ctrl = map()->in(0);
1403 // Follow the chain of parsers and see whether the update can be
1404 // done in the map of callers. We can do the replace for a caller if
1405 // the current control post dominates the control of a caller.
1406 while (parser != NULL && parser->caller() != NULL && progress) {
1407 progress = false;
1408 Node* parent_map = parser->caller()->map();
1409 assert(parser->exits().map()->jvms()->depth() == parser->caller()->depth(), "map mismatch");
1410
1411 Node* parent_ctrl = parent_map->in(0);
1412
1413 while (parent_ctrl->is_Region()) {
1414 Node* n = parent_ctrl->as_Region()->is_copy();
1415 if (n == NULL) {
1416 break;
1417 }
1418 parent_ctrl = n;
1419 }
1420
1421 for (;;) {
1422 if (ctrl == parent_ctrl) {
1423 // update the map of the exits which is the one that will be
1424 // used when compilation resume after inlining
1425 parser->exits().map()->replace_edge(old, neww);
1426 progress = true;
1427 break;
1428 }
1429 if (ctrl->is_Proj() && ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
1430 ctrl = ctrl->in(0)->in(0);
1431 } else if (ctrl->is_Region()) {
1432 Node* n = ctrl->as_Region()->is_copy();
1433 if (n == NULL) {
1434 break;
1435 }
1436 ctrl = n;
1437 } else {
1438 break;
1439 }
1440 }
1441
1442 parser = parser->parent_parser();
1443 }
1444 }
1445
1446
1447 //=============================================================================
1448 //--------------------------------memory---------------------------------------
1449 Node* GraphKit::memory(uint alias_idx) {
1450 MergeMemNode* mem = merged_memory();
1451 Node* p = mem->memory_at(alias_idx);
1452 _gvn.set_type(p, Type::MEMORY); // must be mapped
1453 return p;
1454 }
1455
1456 //-----------------------------reset_memory------------------------------------
1457 Node* GraphKit::reset_memory() {
1458 Node* mem = map()->memory();
1459 // do not use this node for any more parsing!
1460 debug_only( map()->set_memory((Node*)NULL) );
1461 return _gvn.transform( mem );
1462 }
1463
1464 //------------------------------set_all_memory---------------------------------
1465 void GraphKit::set_all_memory(Node* newmem) {
1466 Node* mergemem = MergeMemNode::make(C, newmem);
1467 gvn().set_type_bottom(mergemem);
1468 map()->set_memory(mergemem);
1469 }
1470
1471 //------------------------------set_all_memory_call----------------------------
1472 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1473 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1474 set_all_memory(newmem);
1475 }
1476
1477 //=============================================================================
1478 //
1479 // parser factory methods for MemNodes
1480 //
1481 // These are layered on top of the factory methods in LoadNode and StoreNode,
1482 // and integrate with the parser's memory state and _gvn engine.
1483 //
1484
1485 // factory methods in "int adr_idx"
1486 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1487 int adr_idx,
1488 MemNode::MemOrd mo, bool require_atomic_access) {
1489 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1490 const TypePtr* adr_type = NULL; // debug-mode-only argument
1491 debug_only(adr_type = C->get_adr_type(adr_idx));
1492 Node* mem = memory(adr_idx);
1493 Node* ld;
1494 if (require_atomic_access && bt == T_LONG) {
1495 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo);
1496 } else {
1497 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo);
1498 }
1499 ld = _gvn.transform(ld);
1500 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1501 // Improve graph before escape analysis and boxing elimination.
1502 record_for_igvn(ld);
1503 }
1504 return ld;
1505 }
1506
1507 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1508 int adr_idx,
1509 MemNode::MemOrd mo,
1510 bool require_atomic_access) {
1511 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1512 const TypePtr* adr_type = NULL;
1513 debug_only(adr_type = C->get_adr_type(adr_idx));
1514 Node *mem = memory(adr_idx);
1515 Node* st;
1516 if (require_atomic_access && bt == T_LONG) {
1517 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1518 } else {
1519 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1520 }
1521 st = _gvn.transform(st);
1522 set_memory(st, adr_idx);
1523 // Back-to-back stores can only remove intermediate store with DU info
1524 // so push on worklist for optimizer.
1525 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1526 record_for_igvn(st);
1527
1528 return st;
1529 }
1530
1531
1532 void GraphKit::pre_barrier(bool do_load,
1533 Node* ctl,
1534 Node* obj,
1535 Node* adr,
1536 uint adr_idx,
1537 Node* val,
1538 const TypeOopPtr* val_type,
1539 Node* pre_val,
1540 BasicType bt) {
1541
1542 BarrierSet* bs = Universe::heap()->barrier_set();
1543 set_control(ctl);
1544 switch (bs->kind()) {
1545 case BarrierSet::G1SATBCT:
1546 case BarrierSet::G1SATBCTLogging:
1547 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1548 break;
1549
1550 case BarrierSet::CardTableModRef:
1551 case BarrierSet::CardTableExtension:
1552 case BarrierSet::ModRef:
1553 break;
1554
1555 case BarrierSet::Other:
1556 default :
1557 ShouldNotReachHere();
1558
1559 }
1560 }
1561
1562 bool GraphKit::can_move_pre_barrier() const {
1563 BarrierSet* bs = Universe::heap()->barrier_set();
1564 switch (bs->kind()) {
1565 case BarrierSet::G1SATBCT:
1566 case BarrierSet::G1SATBCTLogging:
1567 return true; // Can move it if no safepoint
1568
1569 case BarrierSet::CardTableModRef:
1570 case BarrierSet::CardTableExtension:
1571 case BarrierSet::ModRef:
1572 return true; // There is no pre-barrier
1573
1574 case BarrierSet::Other:
1575 default :
1576 ShouldNotReachHere();
1577 }
1578 return false;
1579 }
1580
1581 void GraphKit::post_barrier(Node* ctl,
1582 Node* store,
1583 Node* obj,
1584 Node* adr,
1585 uint adr_idx,
1586 Node* val,
1587 BasicType bt,
1588 bool use_precise) {
1589 BarrierSet* bs = Universe::heap()->barrier_set();
1590 set_control(ctl);
1591 switch (bs->kind()) {
1592 case BarrierSet::G1SATBCT:
1593 case BarrierSet::G1SATBCTLogging:
1594 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1595 break;
1596
1597 case BarrierSet::CardTableModRef:
1598 case BarrierSet::CardTableExtension:
1599 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1600 break;
1601
1602 case BarrierSet::ModRef:
1603 break;
1604
1605 case BarrierSet::Other:
1606 default :
1607 ShouldNotReachHere();
1608
1609 }
1610 }
1611
1612 Node* GraphKit::store_oop(Node* ctl,
1613 Node* obj,
1614 Node* adr,
1615 const TypePtr* adr_type,
1616 Node* val,
1617 const TypeOopPtr* val_type,
1618 BasicType bt,
1619 bool use_precise,
1620 MemNode::MemOrd mo) {
1621 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1622 // could be delayed during Parse (for example, in adjust_map_after_if()).
1623 // Execute transformation here to avoid barrier generation in such case.
1624 if (_gvn.type(val) == TypePtr::NULL_PTR)
1625 val = _gvn.makecon(TypePtr::NULL_PTR);
1626
1627 set_control(ctl);
1628 if (stopped()) return top(); // Dead path ?
1629
1630 assert(bt == T_OBJECT, "sanity");
1631 assert(val != NULL, "not dead path");
1632 uint adr_idx = C->get_alias_index(adr_type);
1633 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1634
1635 pre_barrier(true /* do_load */,
1636 control(), obj, adr, adr_idx, val, val_type,
1637 NULL /* pre_val */,
1638 bt);
1639
1640 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo);
1641 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1642 return store;
1643 }
1644
1645 // Could be an array or object we don't know at compile time (unsafe ref.)
1646 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1647 Node* obj, // containing obj
1648 Node* adr, // actual adress to store val at
1649 const TypePtr* adr_type,
1650 Node* val,
1651 BasicType bt,
1652 MemNode::MemOrd mo) {
1653 Compile::AliasType* at = C->alias_type(adr_type);
1654 const TypeOopPtr* val_type = NULL;
1655 if (adr_type->isa_instptr()) {
1656 if (at->field() != NULL) {
1657 // known field. This code is a copy of the do_put_xxx logic.
1658 ciField* field = at->field();
1659 if (!field->type()->is_loaded()) {
1660 val_type = TypeInstPtr::BOTTOM;
1661 } else {
1662 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1663 }
1664 }
1665 } else if (adr_type->isa_aryptr()) {
1666 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1667 }
1668 if (val_type == NULL) {
1669 val_type = TypeInstPtr::BOTTOM;
1670 }
1671 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo);
1672 }
1673
1674
1675 //-------------------------array_element_address-------------------------
1676 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1677 const TypeInt* sizetype) {
1678 uint shift = exact_log2(type2aelembytes(elembt));
1679 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1680
1681 // short-circuit a common case (saves lots of confusing waste motion)
1682 jint idx_con = find_int_con(idx, -1);
1683 if (idx_con >= 0) {
1684 intptr_t offset = header + ((intptr_t)idx_con << shift);
1685 return basic_plus_adr(ary, offset);
1686 }
1687
1688 // must be correct type for alignment purposes
1689 Node* base = basic_plus_adr(ary, header);
1690 #ifdef _LP64
1691 // The scaled index operand to AddP must be a clean 64-bit value.
1692 // Java allows a 32-bit int to be incremented to a negative
1693 // value, which appears in a 64-bit register as a large
1694 // positive number. Using that large positive number as an
1695 // operand in pointer arithmetic has bad consequences.
1696 // On the other hand, 32-bit overflow is rare, and the possibility
1697 // can often be excluded, if we annotate the ConvI2L node with
1698 // a type assertion that its value is known to be a small positive
1699 // number. (The prior range check has ensured this.)
1700 // This assertion is used by ConvI2LNode::Ideal.
1701 int index_max = max_jint - 1; // array size is max_jint, index is one less
1702 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1703 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1704 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) );
1705 #endif
1706 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1707 return basic_plus_adr(ary, base, scale);
1708 }
1709
1710 //-------------------------load_array_element-------------------------
1711 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1712 const Type* elemtype = arytype->elem();
1713 BasicType elembt = elemtype->array_element_basic_type();
1714 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1715 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1716 return ld;
1717 }
1718
1719 //-------------------------set_arguments_for_java_call-------------------------
1720 // Arguments (pre-popped from the stack) are taken from the JVMS.
1721 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1722 // Add the call arguments:
1723 uint nargs = call->method()->arg_size();
1724 for (uint i = 0; i < nargs; i++) {
1725 Node* arg = argument(i);
1726 call->init_req(i + TypeFunc::Parms, arg);
1727 }
1728 }
1729
1730 //---------------------------set_edges_for_java_call---------------------------
1731 // Connect a newly created call into the current JVMS.
1732 // A return value node (if any) is returned from set_edges_for_java_call.
1733 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1734
1735 // Add the predefined inputs:
1736 call->init_req( TypeFunc::Control, control() );
1737 call->init_req( TypeFunc::I_O , i_o() );
1738 call->init_req( TypeFunc::Memory , reset_memory() );
1739 call->init_req( TypeFunc::FramePtr, frameptr() );
1740 call->init_req( TypeFunc::ReturnAdr, top() );
1741
1742 add_safepoint_edges(call, must_throw);
1743
1744 Node* xcall = _gvn.transform(call);
1745
1746 if (xcall == top()) {
1747 set_control(top());
1748 return;
1749 }
1750 assert(xcall == call, "call identity is stable");
1751
1752 // Re-use the current map to produce the result.
1753
1754 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1755 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1756 set_all_memory_call(xcall, separate_io_proj);
1757
1758 //return xcall; // no need, caller already has it
1759 }
1760
1761 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1762 if (stopped()) return top(); // maybe the call folded up?
1763
1764 // Capture the return value, if any.
1765 Node* ret;
1766 if (call->method() == NULL ||
1767 call->method()->return_type()->basic_type() == T_VOID)
1768 ret = top();
1769 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1770
1771 // Note: Since any out-of-line call can produce an exception,
1772 // we always insert an I_O projection from the call into the result.
1773
1774 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1775
1776 if (separate_io_proj) {
1777 // The caller requested separate projections be used by the fall
1778 // through and exceptional paths, so replace the projections for
1779 // the fall through path.
1780 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1781 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1782 }
1783 return ret;
1784 }
1785
1786 //--------------------set_predefined_input_for_runtime_call--------------------
1787 // Reading and setting the memory state is way conservative here.
1788 // The real problem is that I am not doing real Type analysis on memory,
1789 // so I cannot distinguish card mark stores from other stores. Across a GC
1790 // point the Store Barrier and the card mark memory has to agree. I cannot
1791 // have a card mark store and its barrier split across the GC point from
1792 // either above or below. Here I get that to happen by reading ALL of memory.
1793 // A better answer would be to separate out card marks from other memory.
1794 // For now, return the input memory state, so that it can be reused
1795 // after the call, if this call has restricted memory effects.
1796 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1797 // Set fixed predefined input arguments
1798 Node* memory = reset_memory();
1799 call->init_req( TypeFunc::Control, control() );
1800 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1801 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1802 call->init_req( TypeFunc::FramePtr, frameptr() );
1803 call->init_req( TypeFunc::ReturnAdr, top() );
1804 return memory;
1805 }
1806
1807 //-------------------set_predefined_output_for_runtime_call--------------------
1808 // Set control and memory (not i_o) from the call.
1809 // If keep_mem is not NULL, use it for the output state,
1810 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1811 // If hook_mem is NULL, this call produces no memory effects at all.
1812 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1813 // then only that memory slice is taken from the call.
1814 // In the last case, we must put an appropriate memory barrier before
1815 // the call, so as to create the correct anti-dependencies on loads
1816 // preceding the call.
1817 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1818 Node* keep_mem,
1819 const TypePtr* hook_mem) {
1820 // no i/o
1821 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1822 if (keep_mem) {
1823 // First clone the existing memory state
1824 set_all_memory(keep_mem);
1825 if (hook_mem != NULL) {
1826 // Make memory for the call
1827 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1828 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1829 // We also use hook_mem to extract specific effects from arraycopy stubs.
1830 set_memory(mem, hook_mem);
1831 }
1832 // ...else the call has NO memory effects.
1833
1834 // Make sure the call advertises its memory effects precisely.
1835 // This lets us build accurate anti-dependences in gcm.cpp.
1836 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1837 "call node must be constructed correctly");
1838 } else {
1839 assert(hook_mem == NULL, "");
1840 // This is not a "slow path" call; all memory comes from the call.
1841 set_all_memory_call(call);
1842 }
1843 }
1844
1845
1846 // Replace the call with the current state of the kit.
1847 void GraphKit::replace_call(CallNode* call, Node* result) {
1848 JVMState* ejvms = NULL;
1849 if (has_exceptions()) {
1850 ejvms = transfer_exceptions_into_jvms();
1851 }
1852
1853 SafePointNode* final_state = stop();
1854
1855 // Find all the needed outputs of this call
1856 CallProjections callprojs;
1857 call->extract_projections(&callprojs, true);
1858
1859 Node* init_mem = call->in(TypeFunc::Memory);
1860 Node* final_mem = final_state->in(TypeFunc::Memory);
1861 Node* final_ctl = final_state->in(TypeFunc::Control);
1862 Node* final_io = final_state->in(TypeFunc::I_O);
1863
1864 // Replace all the old call edges with the edges from the inlining result
1865 if (callprojs.fallthrough_catchproj != NULL) {
1866 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1867 }
1868 if (callprojs.fallthrough_memproj != NULL) {
1869 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1870 }
1871 if (callprojs.fallthrough_ioproj != NULL) {
1872 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1873 }
1874
1875 // Replace the result with the new result if it exists and is used
1876 if (callprojs.resproj != NULL && result != NULL) {
1877 C->gvn_replace_by(callprojs.resproj, result);
1878 }
1879
1880 if (ejvms == NULL) {
1881 // No exception edges to simply kill off those paths
1882 if (callprojs.catchall_catchproj != NULL) {
1883 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1884 }
1885 if (callprojs.catchall_memproj != NULL) {
1886 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1887 }
1888 if (callprojs.catchall_ioproj != NULL) {
1889 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1890 }
1891 // Replace the old exception object with top
1892 if (callprojs.exobj != NULL) {
1893 C->gvn_replace_by(callprojs.exobj, C->top());
1894 }
1895 } else {
1896 GraphKit ekit(ejvms);
1897
1898 // Load my combined exception state into the kit, with all phis transformed:
1899 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1900
1901 Node* ex_oop = ekit.use_exception_state(ex_map);
1902 if (callprojs.catchall_catchproj != NULL) {
1903 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1904 }
1905 if (callprojs.catchall_memproj != NULL) {
1906 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1907 }
1908 if (callprojs.catchall_ioproj != NULL) {
1909 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1910 }
1911
1912 // Replace the old exception object with the newly created one
1913 if (callprojs.exobj != NULL) {
1914 C->gvn_replace_by(callprojs.exobj, ex_oop);
1915 }
1916 }
1917
1918 // Disconnect the call from the graph
1919 call->disconnect_inputs(NULL, C);
1920 C->gvn_replace_by(call, C->top());
1921
1922 // Clean up any MergeMems that feed other MergeMems since the
1923 // optimizer doesn't like that.
1924 if (final_mem->is_MergeMem()) {
1925 Node_List wl;
1926 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1927 Node* m = i.get();
1928 if (m->is_MergeMem() && !wl.contains(m)) {
1929 wl.push(m);
1930 }
1931 }
1932 while (wl.size() > 0) {
1933 _gvn.transform(wl.pop());
1934 }
1935 }
1936 }
1937
1938
1939 //------------------------------increment_counter------------------------------
1940 // for statistics: increment a VM counter by 1
1941
1942 void GraphKit::increment_counter(address counter_addr) {
1943 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1944 increment_counter(adr1);
1945 }
1946
1947 void GraphKit::increment_counter(Node* counter_addr) {
1948 int adr_type = Compile::AliasIdxRaw;
1949 Node* ctrl = control();
1950 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1951 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1952 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1953 }
1954
1955
1956 //------------------------------uncommon_trap----------------------------------
1957 // Bail out to the interpreter in mid-method. Implemented by calling the
1958 // uncommon_trap blob. This helper function inserts a runtime call with the
1959 // right debug info.
1960 void GraphKit::uncommon_trap(int trap_request,
1961 ciKlass* klass, const char* comment,
1962 bool must_throw,
1963 bool keep_exact_action) {
1964 if (failing()) stop();
1965 if (stopped()) return; // trap reachable?
1966
1967 // Note: If ProfileTraps is true, and if a deopt. actually
1968 // occurs here, the runtime will make sure an MDO exists. There is
1969 // no need to call method()->ensure_method_data() at this point.
1970
1971 // Set the stack pointer to the right value for reexecution:
1972 set_sp(reexecute_sp());
1973
1974 #ifdef ASSERT
1975 if (!must_throw) {
1976 // Make sure the stack has at least enough depth to execute
1977 // the current bytecode.
1978 int inputs, ignored_depth;
1979 if (compute_stack_effects(inputs, ignored_depth)) {
1980 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1981 Bytecodes::name(java_bc()), sp(), inputs));
1982 }
1983 }
1984 #endif
1985
1986 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1987 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1988
1989 switch (action) {
1990 case Deoptimization::Action_maybe_recompile:
1991 case Deoptimization::Action_reinterpret:
1992 // Temporary fix for 6529811 to allow virtual calls to be sure they
1993 // get the chance to go from mono->bi->mega
1994 if (!keep_exact_action &&
1995 Deoptimization::trap_request_index(trap_request) < 0 &&
1996 too_many_recompiles(reason)) {
1997 // This BCI is causing too many recompilations.
1998 action = Deoptimization::Action_none;
1999 trap_request = Deoptimization::make_trap_request(reason, action);
2000 } else {
2001 C->set_trap_can_recompile(true);
2002 }
2003 break;
2004 case Deoptimization::Action_make_not_entrant:
2005 C->set_trap_can_recompile(true);
2006 break;
2007 #ifdef ASSERT
2008 case Deoptimization::Action_none:
2009 case Deoptimization::Action_make_not_compilable:
2010 break;
2011 default:
2012 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
2013 break;
2014 #endif
2015 }
2016
2017 if (TraceOptoParse) {
2018 char buf[100];
2019 tty->print_cr("Uncommon trap %s at bci:%d",
2020 Deoptimization::format_trap_request(buf, sizeof(buf),
2021 trap_request), bci());
2022 }
2023
2024 CompileLog* log = C->log();
2025 if (log != NULL) {
2026 int kid = (klass == NULL)? -1: log->identify(klass);
2027 log->begin_elem("uncommon_trap bci='%d'", bci());
2028 char buf[100];
2029 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2030 trap_request));
2031 if (kid >= 0) log->print(" klass='%d'", kid);
2032 if (comment != NULL) log->print(" comment='%s'", comment);
2033 log->end_elem();
2034 }
2035
2036 // Make sure any guarding test views this path as very unlikely
2037 Node *i0 = control()->in(0);
2038 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2039 IfNode *iff = i0->as_If();
2040 float f = iff->_prob; // Get prob
2041 if (control()->Opcode() == Op_IfTrue) {
2042 if (f > PROB_UNLIKELY_MAG(4))
2043 iff->_prob = PROB_MIN;
2044 } else {
2045 if (f < PROB_LIKELY_MAG(4))
2046 iff->_prob = PROB_MAX;
2047 }
2048 }
2049
2050 // Clear out dead values from the debug info.
2051 kill_dead_locals();
2052
2053 // Now insert the uncommon trap subroutine call
2054 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2055 const TypePtr* no_memory_effects = NULL;
2056 // Pass the index of the class to be loaded
2057 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2058 (must_throw ? RC_MUST_THROW : 0),
2059 OptoRuntime::uncommon_trap_Type(),
2060 call_addr, "uncommon_trap", no_memory_effects,
2061 intcon(trap_request));
2062 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2063 "must extract request correctly from the graph");
2064 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2065
2066 call->set_req(TypeFunc::ReturnAdr, returnadr());
2067 // The debug info is the only real input to this call.
2068
2069 // Halt-and-catch fire here. The above call should never return!
2070 HaltNode* halt = new(C) HaltNode(control(), frameptr());
2071 _gvn.set_type_bottom(halt);
2072 root()->add_req(halt);
2073
2074 stop_and_kill_map();
2075 }
2076
2077
2078 //--------------------------just_allocated_object------------------------------
2079 // Report the object that was just allocated.
2080 // It must be the case that there are no intervening safepoints.
2081 // We use this to determine if an object is so "fresh" that
2082 // it does not require card marks.
2083 Node* GraphKit::just_allocated_object(Node* current_control) {
2084 if (C->recent_alloc_ctl() == current_control)
2085 return C->recent_alloc_obj();
2086 return NULL;
2087 }
2088
2089
2090 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2091 // (Note: TypeFunc::make has a cache that makes this fast.)
2092 const TypeFunc* tf = TypeFunc::make(dest_method);
2093 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2094 for (int j = 0; j < nargs; j++) {
2095 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2096 if( targ->basic_type() == T_DOUBLE ) {
2097 // If any parameters are doubles, they must be rounded before
2098 // the call, dstore_rounding does gvn.transform
2099 Node *arg = argument(j);
2100 arg = dstore_rounding(arg);
2101 set_argument(j, arg);
2102 }
2103 }
2104 }
2105
2106 /**
2107 * Record profiling data exact_kls for Node n with the type system so
2108 * that it can propagate it (speculation)
2109 *
2110 * @param n node that the type applies to
2111 * @param exact_kls type from profiling
2112 *
2113 * @return node with improved type
2114 */
2115 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) {
2116 const Type* current_type = _gvn.type(n);
2117 assert(UseTypeSpeculation, "type speculation must be on");
2118
2119 const TypeOopPtr* speculative = current_type->speculative();
2120
2121 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2122 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2123 const TypeOopPtr* xtype = tklass->as_instance_type();
2124 assert(xtype->klass_is_exact(), "Should be exact");
2125 // record the new speculative type's depth
2126 speculative = xtype->with_inline_depth(jvms()->depth());
2127 }
2128
2129 if (speculative != current_type->speculative()) {
2130 // Build a type with a speculative type (what we think we know
2131 // about the type but will need a guard when we use it)
2132 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2133 // We're changing the type, we need a new CheckCast node to carry
2134 // the new type. The new type depends on the control: what
2135 // profiling tells us is only valid from here as far as we can
2136 // tell.
2137 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2138 cast = _gvn.transform(cast);
2139 replace_in_map(n, cast);
2140 n = cast;
2141 }
2142
2143 return n;
2144 }
2145
2146 /**
2147 * Record profiling data from receiver profiling at an invoke with the
2148 * type system so that it can propagate it (speculation)
2149 *
2150 * @param n receiver node
2151 *
2152 * @return node with improved type
2153 */
2154 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2155 if (!UseTypeSpeculation) {
2156 return n;
2157 }
2158 ciKlass* exact_kls = profile_has_unique_klass();
2159 return record_profile_for_speculation(n, exact_kls);
2160 }
2161
2162 /**
2163 * Record profiling data from argument profiling at an invoke with the
2164 * type system so that it can propagate it (speculation)
2165 *
2166 * @param dest_method target method for the call
2167 * @param bc what invoke bytecode is this?
2168 */
2169 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2170 if (!UseTypeSpeculation) {
2171 return;
2172 }
2173 const TypeFunc* tf = TypeFunc::make(dest_method);
2174 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2175 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2176 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2177 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2178 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2179 ciKlass* better_type = method()->argument_profiled_type(bci(), i);
2180 if (better_type != NULL) {
2181 record_profile_for_speculation(argument(j), better_type);
2182 }
2183 i++;
2184 }
2185 }
2186 }
2187
2188 /**
2189 * Record profiling data from parameter profiling at an invoke with
2190 * the type system so that it can propagate it (speculation)
2191 */
2192 void GraphKit::record_profiled_parameters_for_speculation() {
2193 if (!UseTypeSpeculation) {
2194 return;
2195 }
2196 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2197 if (_gvn.type(local(i))->isa_oopptr()) {
2198 ciKlass* better_type = method()->parameter_profiled_type(j);
2199 if (better_type != NULL) {
2200 record_profile_for_speculation(local(i), better_type);
2201 }
2202 j++;
2203 }
2204 }
2205 }
2206
2207 void GraphKit::round_double_result(ciMethod* dest_method) {
2208 // A non-strict method may return a double value which has an extended
2209 // exponent, but this must not be visible in a caller which is 'strict'
2210 // If a strict caller invokes a non-strict callee, round a double result
2211
2212 BasicType result_type = dest_method->return_type()->basic_type();
2213 assert( method() != NULL, "must have caller context");
2214 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2215 // Destination method's return value is on top of stack
2216 // dstore_rounding() does gvn.transform
2217 Node *result = pop_pair();
2218 result = dstore_rounding(result);
2219 push_pair(result);
2220 }
2221 }
2222
2223 // rounding for strict float precision conformance
2224 Node* GraphKit::precision_rounding(Node* n) {
2225 return UseStrictFP && _method->flags().is_strict()
2226 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2227 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2228 : n;
2229 }
2230
2231 // rounding for strict double precision conformance
2232 Node* GraphKit::dprecision_rounding(Node *n) {
2233 return UseStrictFP && _method->flags().is_strict()
2234 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2235 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2236 : n;
2237 }
2238
2239 // rounding for non-strict double stores
2240 Node* GraphKit::dstore_rounding(Node* n) {
2241 return Matcher::strict_fp_requires_explicit_rounding
2242 && UseSSE <= 1
2243 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2244 : n;
2245 }
2246
2247 //=============================================================================
2248 // Generate a fast path/slow path idiom. Graph looks like:
2249 // [foo] indicates that 'foo' is a parameter
2250 //
2251 // [in] NULL
2252 // \ /
2253 // CmpP
2254 // Bool ne
2255 // If
2256 // / \
2257 // True False-<2>
2258 // / |
2259 // / cast_not_null
2260 // Load | | ^
2261 // [fast_test] | |
2262 // gvn to opt_test | |
2263 // / \ | <1>
2264 // True False |
2265 // | \\ |
2266 // [slow_call] \[fast_result]
2267 // Ctl Val \ \
2268 // | \ \
2269 // Catch <1> \ \
2270 // / \ ^ \ \
2271 // Ex No_Ex | \ \
2272 // | \ \ | \ <2> \
2273 // ... \ [slow_res] | | \ [null_result]
2274 // \ \--+--+--- | |
2275 // \ | / \ | /
2276 // --------Region Phi
2277 //
2278 //=============================================================================
2279 // Code is structured as a series of driver functions all called 'do_XXX' that
2280 // call a set of helper functions. Helper functions first, then drivers.
2281
2282 //------------------------------null_check_oop---------------------------------
2283 // Null check oop. Set null-path control into Region in slot 3.
2284 // Make a cast-not-nullness use the other not-null control. Return cast.
2285 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2286 bool never_see_null, bool safe_for_replace) {
2287 // Initial NULL check taken path
2288 (*null_control) = top();
2289 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2290
2291 // Generate uncommon_trap:
2292 if (never_see_null && (*null_control) != top()) {
2293 // If we see an unexpected null at a check-cast we record it and force a
2294 // recompile; the offending check-cast will be compiled to handle NULLs.
2295 // If we see more than one offending BCI, then all checkcasts in the
2296 // method will be compiled to handle NULLs.
2297 PreserveJVMState pjvms(this);
2298 set_control(*null_control);
2299 replace_in_map(value, null());
2300 uncommon_trap(Deoptimization::Reason_null_check,
2301 Deoptimization::Action_make_not_entrant);
2302 (*null_control) = top(); // NULL path is dead
2303 }
2304 if ((*null_control) == top() && safe_for_replace) {
2305 replace_in_map(value, cast);
2306 }
2307
2308 // Cast away null-ness on the result
2309 return cast;
2310 }
2311
2312 //------------------------------opt_iff----------------------------------------
2313 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2314 // Return slow-path control.
2315 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2316 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2317
2318 // Fast path taken; set region slot 2
2319 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2320 region->init_req(2,fast_taken); // Capture fast-control
2321
2322 // Fast path not-taken, i.e. slow path
2323 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2324 return slow_taken;
2325 }
2326
2327 //-----------------------------make_runtime_call-------------------------------
2328 Node* GraphKit::make_runtime_call(int flags,
2329 const TypeFunc* call_type, address call_addr,
2330 const char* call_name,
2331 const TypePtr* adr_type,
2332 // The following parms are all optional.
2333 // The first NULL ends the list.
2334 Node* parm0, Node* parm1,
2335 Node* parm2, Node* parm3,
2336 Node* parm4, Node* parm5,
2337 Node* parm6, Node* parm7) {
2338 // Slow-path call
2339 bool is_leaf = !(flags & RC_NO_LEAF);
2340 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2341 if (call_name == NULL) {
2342 assert(!is_leaf, "must supply name for leaf");
2343 call_name = OptoRuntime::stub_name(call_addr);
2344 }
2345 CallNode* call;
2346 if (!is_leaf) {
2347 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2348 bci(), adr_type);
2349 } else if (flags & RC_NO_FP) {
2350 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2351 } else {
2352 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2353 }
2354
2355 // The following is similar to set_edges_for_java_call,
2356 // except that the memory effects of the call are restricted to AliasIdxRaw.
2357
2358 // Slow path call has no side-effects, uses few values
2359 bool wide_in = !(flags & RC_NARROW_MEM);
2360 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2361
2362 Node* prev_mem = NULL;
2363 if (wide_in) {
2364 prev_mem = set_predefined_input_for_runtime_call(call);
2365 } else {
2366 assert(!wide_out, "narrow in => narrow out");
2367 Node* narrow_mem = memory(adr_type);
2368 prev_mem = reset_memory();
2369 map()->set_memory(narrow_mem);
2370 set_predefined_input_for_runtime_call(call);
2371 }
2372
2373 // Hook each parm in order. Stop looking at the first NULL.
2374 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2375 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2376 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2377 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2378 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2379 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2380 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2381 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2382 /* close each nested if ===> */ } } } } } } } }
2383 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2384
2385 if (!is_leaf) {
2386 // Non-leaves can block and take safepoints:
2387 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2388 }
2389 // Non-leaves can throw exceptions:
2390 if (has_io) {
2391 call->set_req(TypeFunc::I_O, i_o());
2392 }
2393
2394 if (flags & RC_UNCOMMON) {
2395 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2396 // (An "if" probability corresponds roughly to an unconditional count.
2397 // Sort of.)
2398 call->set_cnt(PROB_UNLIKELY_MAG(4));
2399 }
2400
2401 Node* c = _gvn.transform(call);
2402 assert(c == call, "cannot disappear");
2403
2404 if (wide_out) {
2405 // Slow path call has full side-effects.
2406 set_predefined_output_for_runtime_call(call);
2407 } else {
2408 // Slow path call has few side-effects, and/or sets few values.
2409 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2410 }
2411
2412 if (has_io) {
2413 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2414 }
2415 return call;
2416
2417 }
2418
2419 //------------------------------merge_memory-----------------------------------
2420 // Merge memory from one path into the current memory state.
2421 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2422 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2423 Node* old_slice = mms.force_memory();
2424 Node* new_slice = mms.memory2();
2425 if (old_slice != new_slice) {
2426 PhiNode* phi;
2427 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2428 phi = new_slice->as_Phi();
2429 #ifdef ASSERT
2430 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2431 old_slice = old_slice->in(new_path);
2432 // Caller is responsible for ensuring that any pre-existing
2433 // phis are already aware of old memory.
2434 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2435 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2436 #endif
2437 mms.set_memory(phi);
2438 } else {
2439 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2440 _gvn.set_type(phi, Type::MEMORY);
2441 phi->set_req(new_path, new_slice);
2442 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2443 }
2444 }
2445 }
2446 }
2447
2448 //------------------------------make_slow_call_ex------------------------------
2449 // Make the exception handler hookups for the slow call
2450 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2451 if (stopped()) return;
2452
2453 // Make a catch node with just two handlers: fall-through and catch-all
2454 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2455 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2456 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2457 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2458
2459 { PreserveJVMState pjvms(this);
2460 set_control(excp);
2461 set_i_o(i_o);
2462
2463 if (excp != top()) {
2464 // Create an exception state also.
2465 // Use an exact type if the caller has specified a specific exception.
2466 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2467 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2468 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2469 }
2470 }
2471
2472 // Get the no-exception control from the CatchNode.
2473 set_control(norm);
2474 }
2475
2476
2477 //-------------------------------gen_subtype_check-----------------------------
2478 // Generate a subtyping check. Takes as input the subtype and supertype.
2479 // Returns 2 values: sets the default control() to the true path and returns
2480 // the false path. Only reads invariant memory; sets no (visible) memory.
2481 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2482 // but that's not exposed to the optimizer. This call also doesn't take in an
2483 // Object; if you wish to check an Object you need to load the Object's class
2484 // prior to coming here.
2485 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2486 // Fast check for identical types, perhaps identical constants.
2487 // The types can even be identical non-constants, in cases
2488 // involving Array.newInstance, Object.clone, etc.
2489 if (subklass == superklass)
2490 return top(); // false path is dead; no test needed.
2491
2492 if (_gvn.type(superklass)->singleton()) {
2493 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2494 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2495
2496 // In the common case of an exact superklass, try to fold up the
2497 // test before generating code. You may ask, why not just generate
2498 // the code and then let it fold up? The answer is that the generated
2499 // code will necessarily include null checks, which do not always
2500 // completely fold away. If they are also needless, then they turn
2501 // into a performance loss. Example:
2502 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2503 // Here, the type of 'fa' is often exact, so the store check
2504 // of fa[1]=x will fold up, without testing the nullness of x.
2505 switch (static_subtype_check(superk, subk)) {
2506 case SSC_always_false:
2507 {
2508 Node* always_fail = control();
2509 set_control(top());
2510 return always_fail;
2511 }
2512 case SSC_always_true:
2513 return top();
2514 case SSC_easy_test:
2515 {
2516 // Just do a direct pointer compare and be done.
2517 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2518 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2519 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2520 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2521 return _gvn.transform( new(C) IfFalseNode(iff) );
2522 }
2523 case SSC_full_test:
2524 break;
2525 default:
2526 ShouldNotReachHere();
2527 }
2528 }
2529
2530 // %%% Possible further optimization: Even if the superklass is not exact,
2531 // if the subklass is the unique subtype of the superklass, the check
2532 // will always succeed. We could leave a dependency behind to ensure this.
2533
2534 // First load the super-klass's check-offset
2535 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2536 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(),
2537 TypeInt::INT, MemNode::unordered));
2538 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2539 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2540
2541 // Load from the sub-klass's super-class display list, or a 1-word cache of
2542 // the secondary superclass list, or a failing value with a sentinel offset
2543 // if the super-klass is an interface or exceptionally deep in the Java
2544 // hierarchy and we have to scan the secondary superclass list the hard way.
2545 // Worst-case type is a little odd: NULL is allowed as a result (usually
2546 // klass loads can never produce a NULL).
2547 Node *chk_off_X = ConvI2X(chk_off);
2548 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2549 // For some types like interfaces the following loadKlass is from a 1-word
2550 // cache which is mutable so can't use immutable memory. Other
2551 // types load from the super-class display table which is immutable.
2552 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2553 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2554
2555 // Compile speed common case: ARE a subtype and we canNOT fail
2556 if( superklass == nkls )
2557 return top(); // false path is dead; no test needed.
2558
2559 // See if we get an immediate positive hit. Happens roughly 83% of the
2560 // time. Test to see if the value loaded just previously from the subklass
2561 // is exactly the superklass.
2562 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2563 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2564 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2565 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2566 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2567
2568 // Compile speed common case: Check for being deterministic right now. If
2569 // chk_off is a constant and not equal to cacheoff then we are NOT a
2570 // subklass. In this case we need exactly the 1 test above and we can
2571 // return those results immediately.
2572 if (!might_be_cache) {
2573 Node* not_subtype_ctrl = control();
2574 set_control(iftrue1); // We need exactly the 1 test above
2575 return not_subtype_ctrl;
2576 }
2577
2578 // Gather the various success & failures here
2579 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2580 record_for_igvn(r_ok_subtype);
2581 RegionNode *r_not_subtype = new (C) RegionNode(3);
2582 record_for_igvn(r_not_subtype);
2583
2584 r_ok_subtype->init_req(1, iftrue1);
2585
2586 // Check for immediate negative hit. Happens roughly 11% of the time (which
2587 // is roughly 63% of the remaining cases). Test to see if the loaded
2588 // check-offset points into the subklass display list or the 1-element
2589 // cache. If it points to the display (and NOT the cache) and the display
2590 // missed then it's not a subtype.
2591 Node *cacheoff = _gvn.intcon(cacheoff_con);
2592 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2593 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2594 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2595 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2596 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2597
2598 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2599 // No performance impact (too rare) but allows sharing of secondary arrays
2600 // which has some footprint reduction.
2601 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2602 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2603 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2604 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2605 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2606
2607 // -- Roads not taken here: --
2608 // We could also have chosen to perform the self-check at the beginning
2609 // of this code sequence, as the assembler does. This would not pay off
2610 // the same way, since the optimizer, unlike the assembler, can perform
2611 // static type analysis to fold away many successful self-checks.
2612 // Non-foldable self checks work better here in second position, because
2613 // the initial primary superclass check subsumes a self-check for most
2614 // types. An exception would be a secondary type like array-of-interface,
2615 // which does not appear in its own primary supertype display.
2616 // Finally, we could have chosen to move the self-check into the
2617 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2618 // dependent manner. But it is worthwhile to have the check here,
2619 // where it can be perhaps be optimized. The cost in code space is
2620 // small (register compare, branch).
2621
2622 // Now do a linear scan of the secondary super-klass array. Again, no real
2623 // performance impact (too rare) but it's gotta be done.
2624 // Since the code is rarely used, there is no penalty for moving it
2625 // out of line, and it can only improve I-cache density.
2626 // The decision to inline or out-of-line this final check is platform
2627 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2628 Node* psc = _gvn.transform(
2629 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2630
2631 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2632 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2633 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2634 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2635 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2636
2637 // Return false path; set default control to true path.
2638 set_control( _gvn.transform(r_ok_subtype) );
2639 return _gvn.transform(r_not_subtype);
2640 }
2641
2642 //----------------------------static_subtype_check-----------------------------
2643 // Shortcut important common cases when superklass is exact:
2644 // (0) superklass is java.lang.Object (can occur in reflective code)
2645 // (1) subklass is already limited to a subtype of superklass => always ok
2646 // (2) subklass does not overlap with superklass => always fail
2647 // (3) superklass has NO subtypes and we can check with a simple compare.
2648 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2649 if (StressReflectiveCode) {
2650 return SSC_full_test; // Let caller generate the general case.
2651 }
2652
2653 if (superk == env()->Object_klass()) {
2654 return SSC_always_true; // (0) this test cannot fail
2655 }
2656
2657 ciType* superelem = superk;
2658 if (superelem->is_array_klass())
2659 superelem = superelem->as_array_klass()->base_element_type();
2660
2661 if (!subk->is_interface()) { // cannot trust static interface types yet
2662 if (subk->is_subtype_of(superk)) {
2663 return SSC_always_true; // (1) false path dead; no dynamic test needed
2664 }
2665 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2666 !superk->is_subtype_of(subk)) {
2667 return SSC_always_false;
2668 }
2669 }
2670
2671 // If casting to an instance klass, it must have no subtypes
2672 if (superk->is_interface()) {
2673 // Cannot trust interfaces yet.
2674 // %%% S.B. superk->nof_implementors() == 1
2675 } else if (superelem->is_instance_klass()) {
2676 ciInstanceKlass* ik = superelem->as_instance_klass();
2677 if (!ik->has_subklass() && !ik->is_interface()) {
2678 if (!ik->is_final()) {
2679 // Add a dependency if there is a chance of a later subclass.
2680 C->dependencies()->assert_leaf_type(ik);
2681 }
2682 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2683 }
2684 } else {
2685 // A primitive array type has no subtypes.
2686 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2687 }
2688
2689 return SSC_full_test;
2690 }
2691
2692 // Profile-driven exact type check:
2693 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2694 float prob,
2695 Node* *casted_receiver) {
2696 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2697 Node* recv_klass = load_object_klass(receiver);
2698 Node* want_klass = makecon(tklass);
2699 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2700 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2701 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2702 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2703 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2704
2705 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2706 assert(recv_xtype->klass_is_exact(), "");
2707
2708 // Subsume downstream occurrences of receiver with a cast to
2709 // recv_xtype, since now we know what the type will be.
2710 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2711 (*casted_receiver) = _gvn.transform(cast);
2712 // (User must make the replace_in_map call.)
2713
2714 return fail;
2715 }
2716
2717
2718 //------------------------------seems_never_null-------------------------------
2719 // Use null_seen information if it is available from the profile.
2720 // If we see an unexpected null at a type check we record it and force a
2721 // recompile; the offending check will be recompiled to handle NULLs.
2722 // If we see several offending BCIs, then all checks in the
2723 // method will be recompiled.
2724 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2725 if (UncommonNullCast // Cutout for this technique
2726 && obj != null() // And not the -Xcomp stupid case?
2727 && !too_many_traps(Deoptimization::Reason_null_check)
2728 ) {
2729 if (data == NULL)
2730 // Edge case: no mature data. Be optimistic here.
2731 return true;
2732 // If the profile has not seen a null, assume it won't happen.
2733 assert(java_bc() == Bytecodes::_checkcast ||
2734 java_bc() == Bytecodes::_instanceof ||
2735 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2736 return !data->as_BitData()->null_seen();
2737 }
2738 return false;
2739 }
2740
2741 //------------------------maybe_cast_profiled_receiver-------------------------
2742 // If the profile has seen exactly one type, narrow to exactly that type.
2743 // Subsequent type checks will always fold up.
2744 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2745 ciKlass* require_klass,
2746 ciKlass* spec_klass,
2747 bool safe_for_replace) {
2748 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2749
2750 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check;
2751
2752 // Make sure we haven't already deoptimized from this tactic.
2753 if (too_many_traps(reason))
2754 return NULL;
2755
2756 // (No, this isn't a call, but it's enough like a virtual call
2757 // to use the same ciMethod accessor to get the profile info...)
2758 // If we have a speculative type use it instead of profiling (which
2759 // may not help us)
2760 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2761 if (exact_kls != NULL) {// no cast failures here
2762 if (require_klass == NULL ||
2763 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2764 // If we narrow the type to match what the type profile sees or
2765 // the speculative type, we can then remove the rest of the
2766 // cast.
2767 // This is a win, even if the exact_kls is very specific,
2768 // because downstream operations, such as method calls,
2769 // will often benefit from the sharper type.
2770 Node* exact_obj = not_null_obj; // will get updated in place...
2771 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2772 &exact_obj);
2773 { PreserveJVMState pjvms(this);
2774 set_control(slow_ctl);
2775 uncommon_trap(reason,
2776 Deoptimization::Action_maybe_recompile);
2777 }
2778 if (safe_for_replace) {
2779 replace_in_map(not_null_obj, exact_obj);
2780 }
2781 return exact_obj;
2782 }
2783 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2784 }
2785
2786 return NULL;
2787 }
2788
2789 /**
2790 * Cast obj to type and emit guard unless we had too many traps here
2791 * already
2792 *
2793 * @param obj node being casted
2794 * @param type type to cast the node to
2795 * @param not_null true if we know node cannot be null
2796 */
2797 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2798 ciKlass* type,
2799 bool not_null) {
2800 // type == NULL if profiling tells us this object is always null
2801 if (type != NULL) {
2802 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2803 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check;
2804 if (!too_many_traps(null_reason) &&
2805 !too_many_traps(class_reason)) {
2806 Node* not_null_obj = NULL;
2807 // not_null is true if we know the object is not null and
2808 // there's no need for a null check
2809 if (!not_null) {
2810 Node* null_ctl = top();
2811 not_null_obj = null_check_oop(obj, &null_ctl, true, true);
2812 assert(null_ctl->is_top(), "no null control here");
2813 } else {
2814 not_null_obj = obj;
2815 }
2816
2817 Node* exact_obj = not_null_obj;
2818 ciKlass* exact_kls = type;
2819 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2820 &exact_obj);
2821 {
2822 PreserveJVMState pjvms(this);
2823 set_control(slow_ctl);
2824 uncommon_trap(class_reason,
2825 Deoptimization::Action_maybe_recompile);
2826 }
2827 replace_in_map(not_null_obj, exact_obj);
2828 obj = exact_obj;
2829 }
2830 } else {
2831 if (!too_many_traps(Deoptimization::Reason_null_assert)) {
2832 Node* exact_obj = null_assert(obj);
2833 replace_in_map(obj, exact_obj);
2834 obj = exact_obj;
2835 }
2836 }
2837 return obj;
2838 }
2839
2840 //-------------------------------gen_instanceof--------------------------------
2841 // Generate an instance-of idiom. Used by both the instance-of bytecode
2842 // and the reflective instance-of call.
2843 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2844 kill_dead_locals(); // Benefit all the uncommon traps
2845 assert( !stopped(), "dead parse path should be checked in callers" );
2846 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2847 "must check for not-null not-dead klass in callers");
2848
2849 // Make the merge point
2850 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2851 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2852 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2853 C->set_has_split_ifs(true); // Has chance for split-if optimization
2854
2855 ciProfileData* data = NULL;
2856 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2857 data = method()->method_data()->bci_to_data(bci());
2858 }
2859 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2860 && seems_never_null(obj, data));
2861
2862 // Null check; get casted pointer; set region slot 3
2863 Node* null_ctl = top();
2864 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2865
2866 // If not_null_obj is dead, only null-path is taken
2867 if (stopped()) { // Doing instance-of on a NULL?
2868 set_control(null_ctl);
2869 return intcon(0);
2870 }
2871 region->init_req(_null_path, null_ctl);
2872 phi ->init_req(_null_path, intcon(0)); // Set null path value
2873 if (null_ctl == top()) {
2874 // Do this eagerly, so that pattern matches like is_diamond_phi
2875 // will work even during parsing.
2876 assert(_null_path == PATH_LIMIT-1, "delete last");
2877 region->del_req(_null_path);
2878 phi ->del_req(_null_path);
2879 }
2880
2881 // Do we know the type check always succeed?
2882 bool known_statically = false;
2883 if (_gvn.type(superklass)->singleton()) {
2884 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2885 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2886 if (subk != NULL && subk->is_loaded()) {
2887 int static_res = static_subtype_check(superk, subk);
2888 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false);
2889 }
2890 }
2891
2892 if (known_statically && UseTypeSpeculation) {
2893 // If we know the type check always succeeds then we don't use the
2894 // profiling data at this bytecode. Don't lose it, feed it to the
2895 // type system as a speculative type.
2896 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2897 } else {
2898 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2899 // We may not have profiling here or it may not help us. If we
2900 // have a speculative type use it to perform an exact cast.
2901 ciKlass* spec_obj_type = obj_type->speculative_type();
2902 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2903 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2904 if (stopped()) { // Profile disagrees with this path.
2905 set_control(null_ctl); // Null is the only remaining possibility.
2906 return intcon(0);
2907 }
2908 if (cast_obj != NULL) {
2909 not_null_obj = cast_obj;
2910 }
2911 }
2912 }
2913
2914 // Load the object's klass
2915 Node* obj_klass = load_object_klass(not_null_obj);
2916
2917 // Generate the subtype check
2918 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2919
2920 // Plug in the success path to the general merge in slot 1.
2921 region->init_req(_obj_path, control());
2922 phi ->init_req(_obj_path, intcon(1));
2923
2924 // Plug in the failing path to the general merge in slot 2.
2925 region->init_req(_fail_path, not_subtype_ctrl);
2926 phi ->init_req(_fail_path, intcon(0));
2927
2928 // Return final merged results
2929 set_control( _gvn.transform(region) );
2930 record_for_igvn(region);
2931 return _gvn.transform(phi);
2932 }
2933
2934 //-------------------------------gen_checkcast---------------------------------
2935 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2936 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2937 // uncommon-trap paths work. Adjust stack after this call.
2938 // If failure_control is supplied and not null, it is filled in with
2939 // the control edge for the cast failure. Otherwise, an appropriate
2940 // uncommon trap or exception is thrown.
2941 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2942 Node* *failure_control) {
2943 kill_dead_locals(); // Benefit all the uncommon traps
2944 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2945 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2946
2947 // Fast cutout: Check the case that the cast is vacuously true.
2948 // This detects the common cases where the test will short-circuit
2949 // away completely. We do this before we perform the null check,
2950 // because if the test is going to turn into zero code, we don't
2951 // want a residual null check left around. (Causes a slowdown,
2952 // for example, in some objArray manipulations, such as a[i]=a[j].)
2953 if (tk->singleton()) {
2954 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2955 if (objtp != NULL && objtp->klass() != NULL) {
2956 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2957 case SSC_always_true:
2958 // If we know the type check always succeed then we don't use
2959 // the profiling data at this bytecode. Don't lose it, feed it
2960 // to the type system as a speculative type.
2961 return record_profiled_receiver_for_speculation(obj);
2962 case SSC_always_false:
2963 // It needs a null check because a null will *pass* the cast check.
2964 // A non-null value will always produce an exception.
2965 return null_assert(obj);
2966 }
2967 }
2968 }
2969
2970 ciProfileData* data = NULL;
2971 bool safe_for_replace = false;
2972 if (failure_control == NULL) { // use MDO in regular case only
2973 assert(java_bc() == Bytecodes::_aastore ||
2974 java_bc() == Bytecodes::_checkcast,
2975 "interpreter profiles type checks only for these BCs");
2976 data = method()->method_data()->bci_to_data(bci());
2977 safe_for_replace = true;
2978 }
2979
2980 // Make the merge point
2981 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2982 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
2983 Node* phi = new (C) PhiNode(region, toop);
2984 C->set_has_split_ifs(true); // Has chance for split-if optimization
2985
2986 // Use null-cast information if it is available
2987 bool never_see_null = ((failure_control == NULL) // regular case only
2988 && seems_never_null(obj, data));
2989
2990 // Null check; get casted pointer; set region slot 3
2991 Node* null_ctl = top();
2992 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2993
2994 // If not_null_obj is dead, only null-path is taken
2995 if (stopped()) { // Doing instance-of on a NULL?
2996 set_control(null_ctl);
2997 return null();
2998 }
2999 region->init_req(_null_path, null_ctl);
3000 phi ->init_req(_null_path, null()); // Set null path value
3001 if (null_ctl == top()) {
3002 // Do this eagerly, so that pattern matches like is_diamond_phi
3003 // will work even during parsing.
3004 assert(_null_path == PATH_LIMIT-1, "delete last");
3005 region->del_req(_null_path);
3006 phi ->del_req(_null_path);
3007 }
3008
3009 Node* cast_obj = NULL;
3010 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3011 // We may not have profiling here or it may not help us. If we have
3012 // a speculative type use it to perform an exact cast.
3013 ciKlass* spec_obj_type = obj_type->speculative_type();
3014 if (spec_obj_type != NULL ||
3015 (data != NULL &&
3016 // Counter has never been decremented (due to cast failure).
3017 // ...This is a reasonable thing to expect. It is true of
3018 // all casts inserted by javac to implement generic types.
3019 data->as_CounterData()->count() >= 0)) {
3020 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3021 if (cast_obj != NULL) {
3022 if (failure_control != NULL) // failure is now impossible
3023 (*failure_control) = top();
3024 // adjust the type of the phi to the exact klass:
3025 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3026 }
3027 }
3028
3029 if (cast_obj == NULL) {
3030 // Load the object's klass
3031 Node* obj_klass = load_object_klass(not_null_obj);
3032
3033 // Generate the subtype check
3034 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3035
3036 // Plug in success path into the merge
3037 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
3038 not_null_obj, toop));
3039 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3040 if (failure_control == NULL) {
3041 if (not_subtype_ctrl != top()) { // If failure is possible
3042 PreserveJVMState pjvms(this);
3043 set_control(not_subtype_ctrl);
3044 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3045 }
3046 } else {
3047 (*failure_control) = not_subtype_ctrl;
3048 }
3049 }
3050
3051 region->init_req(_obj_path, control());
3052 phi ->init_req(_obj_path, cast_obj);
3053
3054 // A merge of NULL or Casted-NotNull obj
3055 Node* res = _gvn.transform(phi);
3056
3057 // Note I do NOT always 'replace_in_map(obj,result)' here.
3058 // if( tk->klass()->can_be_primary_super() )
3059 // This means that if I successfully store an Object into an array-of-String
3060 // I 'forget' that the Object is really now known to be a String. I have to
3061 // do this because we don't have true union types for interfaces - if I store
3062 // a Baz into an array-of-Interface and then tell the optimizer it's an
3063 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3064 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3065 // replace_in_map( obj, res );
3066
3067 // Return final merged results
3068 set_control( _gvn.transform(region) );
3069 record_for_igvn(region);
3070 return res;
3071 }
3072
3073 //------------------------------next_monitor-----------------------------------
3074 // What number should be given to the next monitor?
3075 int GraphKit::next_monitor() {
3076 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3077 int next = current + C->sync_stack_slots();
3078 // Keep the toplevel high water mark current:
3079 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3080 return current;
3081 }
3082
3083 //------------------------------insert_mem_bar---------------------------------
3084 // Memory barrier to avoid floating things around
3085 // The membar serves as a pinch point between both control and all memory slices.
3086 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3087 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3088 mb->init_req(TypeFunc::Control, control());
3089 mb->init_req(TypeFunc::Memory, reset_memory());
3090 Node* membar = _gvn.transform(mb);
3091 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3092 set_all_memory_call(membar);
3093 return membar;
3094 }
3095
3096 //-------------------------insert_mem_bar_volatile----------------------------
3097 // Memory barrier to avoid floating things around
3098 // The membar serves as a pinch point between both control and memory(alias_idx).
3099 // If you want to make a pinch point on all memory slices, do not use this
3100 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3101 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3102 // When Parse::do_put_xxx updates a volatile field, it appends a series
3103 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3104 // The first membar is on the same memory slice as the field store opcode.
3105 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3106 // All the other membars (for other volatile slices, including AliasIdxBot,
3107 // which stands for all unknown volatile slices) are control-dependent
3108 // on the first membar. This prevents later volatile loads or stores
3109 // from sliding up past the just-emitted store.
3110
3111 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3112 mb->set_req(TypeFunc::Control,control());
3113 if (alias_idx == Compile::AliasIdxBot) {
3114 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3115 } else {
3116 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3117 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3118 }
3119 Node* membar = _gvn.transform(mb);
3120 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3121 if (alias_idx == Compile::AliasIdxBot) {
3122 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
3123 } else {
3124 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
3125 }
3126 return membar;
3127 }
3128
3129 //------------------------------shared_lock------------------------------------
3130 // Emit locking code.
3131 FastLockNode* GraphKit::shared_lock(Node* obj) {
3132 // bci is either a monitorenter bc or InvocationEntryBci
3133 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3134 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3135
3136 if( !GenerateSynchronizationCode )
3137 return NULL; // Not locking things?
3138 if (stopped()) // Dead monitor?
3139 return NULL;
3140
3141 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3142
3143 // Box the stack location
3144 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
3145 Node* mem = reset_memory();
3146
3147 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
3148 if (PrintPreciseBiasedLockingStatistics) {
3149 // Create the counters for this fast lock.
3150 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3151 }
3152 // Add monitor to debug info for the slow path. If we block inside the
3153 // slow path and de-opt, we need the monitor hanging around
3154 map()->push_monitor( flock );
3155
3156 const TypeFunc *tf = LockNode::lock_type();
3157 LockNode *lock = new (C) LockNode(C, tf);
3158
3159 lock->init_req( TypeFunc::Control, control() );
3160 lock->init_req( TypeFunc::Memory , mem );
3161 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3162 lock->init_req( TypeFunc::FramePtr, frameptr() );
3163 lock->init_req( TypeFunc::ReturnAdr, top() );
3164
3165 lock->init_req(TypeFunc::Parms + 0, obj);
3166 lock->init_req(TypeFunc::Parms + 1, box);
3167 lock->init_req(TypeFunc::Parms + 2, flock);
3168 add_safepoint_edges(lock);
3169
3170 lock = _gvn.transform( lock )->as_Lock();
3171
3172 // lock has no side-effects, sets few values
3173 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3174
3175 insert_mem_bar(Op_MemBarAcquireLock);
3176
3177 // Add this to the worklist so that the lock can be eliminated
3178 record_for_igvn(lock);
3179
3180 #ifndef PRODUCT
3181 if (PrintLockStatistics) {
3182 // Update the counter for this lock. Don't bother using an atomic
3183 // operation since we don't require absolute accuracy.
3184 lock->create_lock_counter(map()->jvms());
3185 increment_counter(lock->counter()->addr());
3186 }
3187 #endif
3188
3189 return flock;
3190 }
3191
3192
3193 //------------------------------shared_unlock----------------------------------
3194 // Emit unlocking code.
3195 void GraphKit::shared_unlock(Node* box, Node* obj) {
3196 // bci is either a monitorenter bc or InvocationEntryBci
3197 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3198 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3199
3200 if( !GenerateSynchronizationCode )
3201 return;
3202 if (stopped()) { // Dead monitor?
3203 map()->pop_monitor(); // Kill monitor from debug info
3204 return;
3205 }
3206
3207 // Memory barrier to avoid floating things down past the locked region
3208 insert_mem_bar(Op_MemBarReleaseLock);
3209
3210 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3211 UnlockNode *unlock = new (C) UnlockNode(C, tf);
3212 uint raw_idx = Compile::AliasIdxRaw;
3213 unlock->init_req( TypeFunc::Control, control() );
3214 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3215 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3216 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3217 unlock->init_req( TypeFunc::ReturnAdr, top() );
3218
3219 unlock->init_req(TypeFunc::Parms + 0, obj);
3220 unlock->init_req(TypeFunc::Parms + 1, box);
3221 unlock = _gvn.transform(unlock)->as_Unlock();
3222
3223 Node* mem = reset_memory();
3224
3225 // unlock has no side-effects, sets few values
3226 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3227
3228 // Kill monitor from debug info
3229 map()->pop_monitor( );
3230 }
3231
3232 //-------------------------------get_layout_helper-----------------------------
3233 // If the given klass is a constant or known to be an array,
3234 // fetch the constant layout helper value into constant_value
3235 // and return (Node*)NULL. Otherwise, load the non-constant
3236 // layout helper value, and return the node which represents it.
3237 // This two-faced routine is useful because allocation sites
3238 // almost always feature constant types.
3239 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3240 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3241 if (!StressReflectiveCode && inst_klass != NULL) {
3242 ciKlass* klass = inst_klass->klass();
3243 bool xklass = inst_klass->klass_is_exact();
3244 if (xklass || klass->is_array_klass()) {
3245 jint lhelper = klass->layout_helper();
3246 if (lhelper != Klass::_lh_neutral_value) {
3247 constant_value = lhelper;
3248 return (Node*) NULL;
3249 }
3250 }
3251 }
3252 constant_value = Klass::_lh_neutral_value; // put in a known value
3253 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3254 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3255 }
3256
3257 // We just put in an allocate/initialize with a big raw-memory effect.
3258 // Hook selected additional alias categories on the initialization.
3259 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3260 MergeMemNode* init_in_merge,
3261 Node* init_out_raw) {
3262 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3263 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3264
3265 Node* prevmem = kit.memory(alias_idx);
3266 init_in_merge->set_memory_at(alias_idx, prevmem);
3267 kit.set_memory(init_out_raw, alias_idx);
3268 }
3269
3270 //---------------------------set_output_for_allocation-------------------------
3271 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3272 const TypeOopPtr* oop_type) {
3273 int rawidx = Compile::AliasIdxRaw;
3274 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3275 add_safepoint_edges(alloc);
3276 Node* allocx = _gvn.transform(alloc);
3277 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3278 // create memory projection for i_o
3279 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3280 make_slow_call_ex(allocx, env()->Throwable_klass(), true);
3281
3282 // create a memory projection as for the normal control path
3283 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3284 set_memory(malloc, rawidx);
3285
3286 // a normal slow-call doesn't change i_o, but an allocation does
3287 // we create a separate i_o projection for the normal control path
3288 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3289 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3290
3291 // put in an initialization barrier
3292 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3293 rawoop)->as_Initialize();
3294 assert(alloc->initialization() == init, "2-way macro link must work");
3295 assert(init ->allocation() == alloc, "2-way macro link must work");
3296 {
3297 // Extract memory strands which may participate in the new object's
3298 // initialization, and source them from the new InitializeNode.
3299 // This will allow us to observe initializations when they occur,
3300 // and link them properly (as a group) to the InitializeNode.
3301 assert(init->in(InitializeNode::Memory) == malloc, "");
3302 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3303 init->set_req(InitializeNode::Memory, minit_in);
3304 record_for_igvn(minit_in); // fold it up later, if possible
3305 Node* minit_out = memory(rawidx);
3306 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3307 if (oop_type->isa_aryptr()) {
3308 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3309 int elemidx = C->get_alias_index(telemref);
3310 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3311 } else if (oop_type->isa_instptr()) {
3312 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3313 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3314 ciField* field = ik->nonstatic_field_at(i);
3315 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3316 continue; // do not bother to track really large numbers of fields
3317 // Find (or create) the alias category for this field:
3318 int fieldidx = C->alias_type(field)->index();
3319 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3320 }
3321 }
3322 }
3323
3324 // Cast raw oop to the real thing...
3325 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3326 javaoop = _gvn.transform(javaoop);
3327 C->set_recent_alloc(control(), javaoop);
3328 assert(just_allocated_object(control()) == javaoop, "just allocated");
3329
3330 #ifdef ASSERT
3331 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3332 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3333 "Ideal_allocation works");
3334 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3335 "Ideal_allocation works");
3336 if (alloc->is_AllocateArray()) {
3337 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3338 "Ideal_allocation works");
3339 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3340 "Ideal_allocation works");
3341 } else {
3342 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3343 }
3344 }
3345 #endif //ASSERT
3346
3347 return javaoop;
3348 }
3349
3350 //---------------------------new_instance--------------------------------------
3351 // This routine takes a klass_node which may be constant (for a static type)
3352 // or may be non-constant (for reflective code). It will work equally well
3353 // for either, and the graph will fold nicely if the optimizer later reduces
3354 // the type to a constant.
3355 // The optional arguments are for specialized use by intrinsics:
3356 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3357 // - If 'return_size_val', report the the total object size to the caller.
3358 Node* GraphKit::new_instance(Node* klass_node,
3359 Node* extra_slow_test,
3360 Node* *return_size_val) {
3361 // Compute size in doublewords
3362 // The size is always an integral number of doublewords, represented
3363 // as a positive bytewise size stored in the klass's layout_helper.
3364 // The layout_helper also encodes (in a low bit) the need for a slow path.
3365 jint layout_con = Klass::_lh_neutral_value;
3366 Node* layout_val = get_layout_helper(klass_node, layout_con);
3367 int layout_is_con = (layout_val == NULL);
3368
3369 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3370 // Generate the initial go-slow test. It's either ALWAYS (return a
3371 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3372 // case) a computed value derived from the layout_helper.
3373 Node* initial_slow_test = NULL;
3374 if (layout_is_con) {
3375 assert(!StressReflectiveCode, "stress mode does not use these paths");
3376 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3377 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3378
3379 } else { // reflective case
3380 // This reflective path is used by Unsafe.allocateInstance.
3381 // (It may be stress-tested by specifying StressReflectiveCode.)
3382 // Basically, we want to get into the VM is there's an illegal argument.
3383 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3384 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3385 if (extra_slow_test != intcon(0)) {
3386 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3387 }
3388 // (Macro-expander will further convert this to a Bool, if necessary.)
3389 }
3390
3391 // Find the size in bytes. This is easy; it's the layout_helper.
3392 // The size value must be valid even if the slow path is taken.
3393 Node* size = NULL;
3394 if (layout_is_con) {
3395 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3396 } else { // reflective case
3397 // This reflective path is used by clone and Unsafe.allocateInstance.
3398 size = ConvI2X(layout_val);
3399
3400 // Clear the low bits to extract layout_helper_size_in_bytes:
3401 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3402 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3403 size = _gvn.transform( new (C) AndXNode(size, mask) );
3404 }
3405 if (return_size_val != NULL) {
3406 (*return_size_val) = size;
3407 }
3408
3409 // This is a precise notnull oop of the klass.
3410 // (Actually, it need not be precise if this is a reflective allocation.)
3411 // It's what we cast the result to.
3412 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3413 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3414 const TypeOopPtr* oop_type = tklass->as_instance_type();
3415
3416 // Now generate allocation code
3417
3418 // The entire memory state is needed for slow path of the allocation
3419 // since GC and deoptimization can happened.
3420 Node *mem = reset_memory();
3421 set_all_memory(mem); // Create new memory state
3422
3423 AllocateNode* alloc
3424 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3425 control(), mem, i_o(),
3426 size, klass_node,
3427 initial_slow_test);
3428
3429 return set_output_for_allocation(alloc, oop_type);
3430 }
3431
3432 //-------------------------------new_array-------------------------------------
3433 // helper for both newarray and anewarray
3434 // The 'length' parameter is (obviously) the length of the array.
3435 // See comments on new_instance for the meaning of the other arguments.
3436 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3437 Node* length, // number of array elements
3438 int nargs, // number of arguments to push back for uncommon trap
3439 Node* *return_size_val) {
3440 jint layout_con = Klass::_lh_neutral_value;
3441 Node* layout_val = get_layout_helper(klass_node, layout_con);
3442 int layout_is_con = (layout_val == NULL);
3443
3444 if (!layout_is_con && !StressReflectiveCode &&
3445 !too_many_traps(Deoptimization::Reason_class_check)) {
3446 // This is a reflective array creation site.
3447 // Optimistically assume that it is a subtype of Object[],
3448 // so that we can fold up all the address arithmetic.
3449 layout_con = Klass::array_layout_helper(T_OBJECT);
3450 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3451 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3452 { BuildCutout unless(this, bol_lh, PROB_MAX);
3453 inc_sp(nargs);
3454 uncommon_trap(Deoptimization::Reason_class_check,
3455 Deoptimization::Action_maybe_recompile);
3456 }
3457 layout_val = NULL;
3458 layout_is_con = true;
3459 }
3460
3461 // Generate the initial go-slow test. Make sure we do not overflow
3462 // if length is huge (near 2Gig) or negative! We do not need
3463 // exact double-words here, just a close approximation of needed
3464 // double-words. We can't add any offset or rounding bits, lest we
3465 // take a size -1 of bytes and make it positive. Use an unsigned
3466 // compare, so negative sizes look hugely positive.
3467 int fast_size_limit = FastAllocateSizeLimit;
3468 if (layout_is_con) {
3469 assert(!StressReflectiveCode, "stress mode does not use these paths");
3470 // Increase the size limit if we have exact knowledge of array type.
3471 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3472 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3473 }
3474
3475 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3476 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3477 if (initial_slow_test->is_Bool()) {
3478 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3479 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3480 }
3481
3482 // --- Size Computation ---
3483 // array_size = round_to_heap(array_header + (length << elem_shift));
3484 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3485 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3486 // The rounding mask is strength-reduced, if possible.
3487 int round_mask = MinObjAlignmentInBytes - 1;
3488 Node* header_size = NULL;
3489 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3490 // (T_BYTE has the weakest alignment and size restrictions...)
3491 if (layout_is_con) {
3492 int hsize = Klass::layout_helper_header_size(layout_con);
3493 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3494 BasicType etype = Klass::layout_helper_element_type(layout_con);
3495 if ((round_mask & ~right_n_bits(eshift)) == 0)
3496 round_mask = 0; // strength-reduce it if it goes away completely
3497 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3498 assert(header_size_min <= hsize, "generic minimum is smallest");
3499 header_size_min = hsize;
3500 header_size = intcon(hsize + round_mask);
3501 } else {
3502 Node* hss = intcon(Klass::_lh_header_size_shift);
3503 Node* hsm = intcon(Klass::_lh_header_size_mask);
3504 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3505 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3506 Node* mask = intcon(round_mask);
3507 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3508 }
3509
3510 Node* elem_shift = NULL;
3511 if (layout_is_con) {
3512 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3513 if (eshift != 0)
3514 elem_shift = intcon(eshift);
3515 } else {
3516 // There is no need to mask or shift this value.
3517 // The semantics of LShiftINode include an implicit mask to 0x1F.
3518 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3519 elem_shift = layout_val;
3520 }
3521
3522 // Transition to native address size for all offset calculations:
3523 Node* lengthx = ConvI2X(length);
3524 Node* headerx = ConvI2X(header_size);
3525 #ifdef _LP64
3526 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3527 if (tllen != NULL && tllen->_lo < 0) {
3528 // Add a manual constraint to a positive range. Cf. array_element_address.
3529 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3530 if (size_max > tllen->_hi) size_max = tllen->_hi;
3531 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3532 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon));
3533 }
3534 }
3535 #endif
3536
3537 // Combine header size (plus rounding) and body size. Then round down.
3538 // This computation cannot overflow, because it is used only in two
3539 // places, one where the length is sharply limited, and the other
3540 // after a successful allocation.
3541 Node* abody = lengthx;
3542 if (elem_shift != NULL)
3543 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3544 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3545 if (round_mask != 0) {
3546 Node* mask = MakeConX(~round_mask);
3547 size = _gvn.transform( new(C) AndXNode(size, mask) );
3548 }
3549 // else if round_mask == 0, the size computation is self-rounding
3550
3551 if (return_size_val != NULL) {
3552 // This is the size
3553 (*return_size_val) = size;
3554 }
3555
3556 // Now generate allocation code
3557
3558 // The entire memory state is needed for slow path of the allocation
3559 // since GC and deoptimization can happened.
3560 Node *mem = reset_memory();
3561 set_all_memory(mem); // Create new memory state
3562
3563 // Create the AllocateArrayNode and its result projections
3564 AllocateArrayNode* alloc
3565 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3566 control(), mem, i_o(),
3567 size, klass_node,
3568 initial_slow_test,
3569 length);
3570
3571 // Cast to correct type. Note that the klass_node may be constant or not,
3572 // and in the latter case the actual array type will be inexact also.
3573 // (This happens via a non-constant argument to inline_native_newArray.)
3574 // In any case, the value of klass_node provides the desired array type.
3575 const TypeInt* length_type = _gvn.find_int_type(length);
3576 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3577 if (ary_type->isa_aryptr() && length_type != NULL) {
3578 // Try to get a better type than POS for the size
3579 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3580 }
3581
3582 Node* javaoop = set_output_for_allocation(alloc, ary_type);
3583
3584 // Cast length on remaining path to be as narrow as possible
3585 if (map()->find_edge(length) >= 0) {
3586 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3587 if (ccast != length) {
3588 _gvn.set_type_bottom(ccast);
3589 record_for_igvn(ccast);
3590 replace_in_map(length, ccast);
3591 }
3592 }
3593
3594 return javaoop;
3595 }
3596
3597 // The following "Ideal_foo" functions are placed here because they recognize
3598 // the graph shapes created by the functions immediately above.
3599
3600 //---------------------------Ideal_allocation----------------------------------
3601 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3602 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3603 if (ptr == NULL) { // reduce dumb test in callers
3604 return NULL;
3605 }
3606 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3607 ptr = ptr->in(1);
3608 if (ptr == NULL) return NULL;
3609 }
3610 // Return NULL for allocations with several casts:
3611 // j.l.reflect.Array.newInstance(jobject, jint)
3612 // Object.clone()
3613 // to keep more precise type from last cast.
3614 if (ptr->is_Proj()) {
3615 Node* allo = ptr->in(0);
3616 if (allo != NULL && allo->is_Allocate()) {
3617 return allo->as_Allocate();
3618 }
3619 }
3620 // Report failure to match.
3621 return NULL;
3622 }
3623
3624 // Fancy version which also strips off an offset (and reports it to caller).
3625 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3626 intptr_t& offset) {
3627 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3628 if (base == NULL) return NULL;
3629 return Ideal_allocation(base, phase);
3630 }
3631
3632 // Trace Initialize <- Proj[Parm] <- Allocate
3633 AllocateNode* InitializeNode::allocation() {
3634 Node* rawoop = in(InitializeNode::RawAddress);
3635 if (rawoop->is_Proj()) {
3636 Node* alloc = rawoop->in(0);
3637 if (alloc->is_Allocate()) {
3638 return alloc->as_Allocate();
3639 }
3640 }
3641 return NULL;
3642 }
3643
3644 // Trace Allocate -> Proj[Parm] -> Initialize
3645 InitializeNode* AllocateNode::initialization() {
3646 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3647 if (rawoop == NULL) return NULL;
3648 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3649 Node* init = rawoop->fast_out(i);
3650 if (init->is_Initialize()) {
3651 assert(init->as_Initialize()->allocation() == this, "2-way link");
3652 return init->as_Initialize();
3653 }
3654 }
3655 return NULL;
3656 }
3657
3658 //----------------------------- loop predicates ---------------------------
3659
3660 //------------------------------add_predicate_impl----------------------------
3661 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3662 // Too many traps seen?
3663 if (too_many_traps(reason)) {
3664 #ifdef ASSERT
3665 if (TraceLoopPredicate) {
3666 int tc = C->trap_count(reason);
3667 tty->print("too many traps=%s tcount=%d in ",
3668 Deoptimization::trap_reason_name(reason), tc);
3669 method()->print(); // which method has too many predicate traps
3670 tty->cr();
3671 }
3672 #endif
3673 // We cannot afford to take more traps here,
3674 // do not generate predicate.
3675 return;
3676 }
3677
3678 Node *cont = _gvn.intcon(1);
3679 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3680 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3681 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3682 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3683 C->add_predicate_opaq(opq);
3684 {
3685 PreserveJVMState pjvms(this);
3686 set_control(iffalse);
3687 inc_sp(nargs);
3688 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3689 }
3690 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3691 set_control(iftrue);
3692 }
3693
3694 //------------------------------add_predicate---------------------------------
3695 void GraphKit::add_predicate(int nargs) {
3696 if (UseLoopPredicate) {
3697 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3698 }
3699 // loop's limit check predicate should be near the loop.
3700 if (LoopLimitCheck) {
3701 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3702 }
3703 }
3704
3705 //----------------------------- store barriers ----------------------------
3706 #define __ ideal.
3707
3708 void GraphKit::sync_kit(IdealKit& ideal) {
3709 set_all_memory(__ merged_memory());
3710 set_i_o(__ i_o());
3711 set_control(__ ctrl());
3712 }
3713
3714 void GraphKit::final_sync(IdealKit& ideal) {
3715 // Final sync IdealKit and graphKit.
3716 sync_kit(ideal);
3717 }
3718
3719 // vanilla/CMS post barrier
3720 // Insert a write-barrier store. This is to let generational GC work; we have
3721 // to flag all oop-stores before the next GC point.
3722 void GraphKit::write_barrier_post(Node* oop_store,
3723 Node* obj,
3724 Node* adr,
3725 uint adr_idx,
3726 Node* val,
3727 bool use_precise) {
3728 // No store check needed if we're storing a NULL or an old object
3729 // (latter case is probably a string constant). The concurrent
3730 // mark sweep garbage collector, however, needs to have all nonNull
3731 // oop updates flagged via card-marks.
3732 if (val != NULL && val->is_Con()) {
3733 // must be either an oop or NULL
3734 const Type* t = val->bottom_type();
3735 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3736 // stores of null never (?) need barriers
3737 return;
3738 }
3739
3740 if (use_ReduceInitialCardMarks()
3741 && obj == just_allocated_object(control())) {
3742 // We can skip marks on a freshly-allocated object in Eden.
3743 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3744 // That routine informs GC to take appropriate compensating steps,
3745 // upon a slow-path allocation, so as to make this card-mark
3746 // elision safe.
3747 return;
3748 }
3749
3750 if (!use_precise) {
3751 // All card marks for a (non-array) instance are in one place:
3752 adr = obj;
3753 }
3754 // (Else it's an array (or unknown), and we want more precise card marks.)
3755 assert(adr != NULL, "");
3756
3757 IdealKit ideal(this, true);
3758
3759 // Convert the pointer to an int prior to doing math on it
3760 Node* cast = __ CastPX(__ ctrl(), adr);
3761
3762 // Divide by card size
3763 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3764 "Only one we handle so far.");
3765 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3766
3767 // Combine card table base and card offset
3768 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3769
3770 // Get the alias_index for raw card-mark memory
3771 int adr_type = Compile::AliasIdxRaw;
3772 Node* zero = __ ConI(0); // Dirty card value
3773 BasicType bt = T_BYTE;
3774
3775 if (UseCondCardMark) {
3776 // The classic GC reference write barrier is typically implemented
3777 // as a store into the global card mark table. Unfortunately
3778 // unconditional stores can result in false sharing and excessive
3779 // coherence traffic as well as false transactional aborts.
3780 // UseCondCardMark enables MP "polite" conditional card mark
3781 // stores. In theory we could relax the load from ctrl() to
3782 // no_ctrl, but that doesn't buy much latitude.
3783 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3784 __ if_then(card_val, BoolTest::ne, zero);
3785 }
3786
3787 // Smash zero into card
3788 if( !UseConcMarkSweepGC ) {
3789 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release);
3790 } else {
3791 // Specialized path for CM store barrier
3792 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3793 }
3794
3795 if (UseCondCardMark) {
3796 __ end_if();
3797 }
3798
3799 // Final sync IdealKit and GraphKit.
3800 final_sync(ideal);
3801 }
3802
3803 // G1 pre/post barriers
3804 void GraphKit::g1_write_barrier_pre(bool do_load,
3805 Node* obj,
3806 Node* adr,
3807 uint alias_idx,
3808 Node* val,
3809 const TypeOopPtr* val_type,
3810 Node* pre_val,
3811 BasicType bt) {
3812
3813 // Some sanity checks
3814 // Note: val is unused in this routine.
3815
3816 if (do_load) {
3817 // We need to generate the load of the previous value
3818 assert(obj != NULL, "must have a base");
3819 assert(adr != NULL, "where are loading from?");
3820 assert(pre_val == NULL, "loaded already?");
3821 assert(val_type != NULL, "need a type");
3822 } else {
3823 // In this case both val_type and alias_idx are unused.
3824 assert(pre_val != NULL, "must be loaded already");
3825 // Nothing to be done if pre_val is null.
3826 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3827 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3828 }
3829 assert(bt == T_OBJECT, "or we shouldn't be here");
3830
3831 IdealKit ideal(this, true);
3832
3833 Node* tls = __ thread(); // ThreadLocalStorage
3834
3835 Node* no_ctrl = NULL;
3836 Node* no_base = __ top();
3837 Node* zero = __ ConI(0);
3838 Node* zeroX = __ ConX(0);
3839
3840 float likely = PROB_LIKELY(0.999);
3841 float unlikely = PROB_UNLIKELY(0.999);
3842
3843 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3844 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3845
3846 // Offsets into the thread
3847 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3848 PtrQueue::byte_offset_of_active());
3849 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3850 PtrQueue::byte_offset_of_index());
3851 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3852 PtrQueue::byte_offset_of_buf());
3853
3854 // Now the actual pointers into the thread
3855 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3856 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3857 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3858
3859 // Now some of the values
3860 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3861
3862 // if (!marking)
3863 __ if_then(marking, BoolTest::ne, zero, unlikely); {
3864 BasicType index_bt = TypeX_X->basic_type();
3865 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3866 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3867
3868 if (do_load) {
3869 // load original value
3870 // alias_idx correct??
3871 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3872 }
3873
3874 // if (pre_val != NULL)
3875 __ if_then(pre_val, BoolTest::ne, null()); {
3876 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3877
3878 // is the queue for this thread full?
3879 __ if_then(index, BoolTest::ne, zeroX, likely); {
3880
3881 // decrement the index
3882 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3883
3884 // Now get the buffer location we will log the previous value into and store it
3885 Node *log_addr = __ AddP(no_base, buffer, next_index);
3886 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
3887 // update the index
3888 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
3889
3890 } __ else_(); {
3891
3892 // logging buffer is full, call the runtime
3893 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3894 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3895 } __ end_if(); // (!index)
3896 } __ end_if(); // (pre_val != NULL)
3897 } __ end_if(); // (!marking)
3898
3899 // Final sync IdealKit and GraphKit.
3900 final_sync(ideal);
3901 }
3902
3903 //
3904 // Update the card table and add card address to the queue
3905 //
3906 void GraphKit::g1_mark_card(IdealKit& ideal,
3907 Node* card_adr,
3908 Node* oop_store,
3909 uint oop_alias_idx,
3910 Node* index,
3911 Node* index_adr,
3912 Node* buffer,
3913 const TypeFunc* tf) {
3914
3915 Node* zero = __ ConI(0);
3916 Node* zeroX = __ ConX(0);
3917 Node* no_base = __ top();
3918 BasicType card_bt = T_BYTE;
3919 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3920 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3921
3922 // Now do the queue work
3923 __ if_then(index, BoolTest::ne, zeroX); {
3924
3925 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3926 Node* log_addr = __ AddP(no_base, buffer, next_index);
3927
3928 // Order, see storeCM.
3929 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
3930 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
3931
3932 } __ else_(); {
3933 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3934 } __ end_if();
3935
3936 }
3937
3938 void GraphKit::g1_write_barrier_post(Node* oop_store,
3939 Node* obj,
3940 Node* adr,
3941 uint alias_idx,
3942 Node* val,
3943 BasicType bt,
3944 bool use_precise) {
3945 // If we are writing a NULL then we need no post barrier
3946
3947 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3948 // Must be NULL
3949 const Type* t = val->bottom_type();
3950 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3951 // No post barrier if writing NULLx
3952 return;
3953 }
3954
3955 if (!use_precise) {
3956 // All card marks for a (non-array) instance are in one place:
3957 adr = obj;
3958 }
3959 // (Else it's an array (or unknown), and we want more precise card marks.)
3960 assert(adr != NULL, "");
3961
3962 IdealKit ideal(this, true);
3963
3964 Node* tls = __ thread(); // ThreadLocalStorage
3965
3966 Node* no_base = __ top();
3967 float likely = PROB_LIKELY(0.999);
3968 float unlikely = PROB_UNLIKELY(0.999);
3969 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
3970 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
3971 Node* zeroX = __ ConX(0);
3972
3973 // Get the alias_index for raw card-mark memory
3974 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3975
3976 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3977
3978 // Offsets into the thread
3979 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3980 PtrQueue::byte_offset_of_index());
3981 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3982 PtrQueue::byte_offset_of_buf());
3983
3984 // Pointers into the thread
3985
3986 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3987 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3988
3989 // Now some values
3990 // Use ctrl to avoid hoisting these values past a safepoint, which could
3991 // potentially reset these fields in the JavaThread.
3992 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
3993 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3994
3995 // Convert the store obj pointer to an int prior to doing math on it
3996 // Must use ctrl to prevent "integerized oop" existing across safepoint
3997 Node* cast = __ CastPX(__ ctrl(), adr);
3998
3999 // Divide pointer by card size
4000 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4001
4002 // Combine card table base and card offset
4003 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4004
4005 // If we know the value being stored does it cross regions?
4006
4007 if (val != NULL) {
4008 // Does the store cause us to cross regions?
4009
4010 // Should be able to do an unsigned compare of region_size instead of
4011 // and extra shift. Do we have an unsigned compare??
4012 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4013 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4014
4015 // if (xor_res == 0) same region so skip
4016 __ if_then(xor_res, BoolTest::ne, zeroX); {
4017
4018 // No barrier if we are storing a NULL
4019 __ if_then(val, BoolTest::ne, null(), unlikely); {
4020
4021 // Ok must mark the card if not already dirty
4022
4023 // load the original value of the card
4024 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4025
4026 __ if_then(card_val, BoolTest::ne, young_card); {
4027 sync_kit(ideal);
4028 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4029 insert_mem_bar(Op_MemBarVolatile, oop_store);
4030 __ sync_kit(this);
4031
4032 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4033 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4034 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4035 } __ end_if();
4036 } __ end_if();
4037 } __ end_if();
4038 } __ end_if();
4039 } else {
4040 // Object.clone() instrinsic uses this path.
4041 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4042 }
4043
4044 // Final sync IdealKit and GraphKit.
4045 final_sync(ideal);
4046 }
4047 #undef __
4048
4049
4050
4051 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
4052 if (java_lang_String::has_offset_field()) {
4053 int offset_offset = java_lang_String::offset_offset_in_bytes();
4054 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4055 false, NULL, 0);
4056 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4057 int offset_field_idx = C->get_alias_index(offset_field_type);
4058 return make_load(ctrl,
4059 basic_plus_adr(str, str, offset_offset),
4060 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered);
4061 } else {
4062 return intcon(0);
4063 }
4064 }
4065
4066 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4067 if (java_lang_String::has_count_field()) {
4068 int count_offset = java_lang_String::count_offset_in_bytes();
4069 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4070 false, NULL, 0);
4071 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4072 int count_field_idx = C->get_alias_index(count_field_type);
4073 return make_load(ctrl,
4074 basic_plus_adr(str, str, count_offset),
4075 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered);
4076 } else {
4077 return load_array_length(load_String_value(ctrl, str));
4078 }
4079 }
4080
4081 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4082 int value_offset = java_lang_String::value_offset_in_bytes();
4083 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4084 false, NULL, 0);
4085 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4086 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4087 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
4088 ciTypeArrayKlass::make(T_CHAR), true, 0);
4089 int value_field_idx = C->get_alias_index(value_field_type);
4090 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4091 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4092 // String.value field is known to be @Stable.
4093 if (UseImplicitStableValues) {
4094 load = cast_array_to_stable(load, value_type);
4095 }
4096 return load;
4097 }
4098
4099 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
4100 int offset_offset = java_lang_String::offset_offset_in_bytes();
4101 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4102 false, NULL, 0);
4103 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4104 int offset_field_idx = C->get_alias_index(offset_field_type);
4105 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
4106 value, T_INT, offset_field_idx, MemNode::unordered);
4107 }
4108
4109 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4110 int value_offset = java_lang_String::value_offset_in_bytes();
4111 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4112 false, NULL, 0);
4113 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4114
4115 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4116 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered);
4117 }
4118
4119 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
4120 int count_offset = java_lang_String::count_offset_in_bytes();
4121 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4122 false, NULL, 0);
4123 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4124 int count_field_idx = C->get_alias_index(count_field_type);
4125 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
4126 value, T_INT, count_field_idx, MemNode::unordered);
4127 }
4128
4129 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4130 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4131 // assumption of CCP analysis.
4132 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
4133 }