1 /*
2 * Copyright (c) 2001, 2017, 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 "ci/ciValueKlass.hpp"
28 #include "gc/g1/g1SATBCardTableModRefBS.hpp"
29 #include "gc/g1/heapRegion.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/cardTableModRefBS.hpp"
32 #include "gc/shared/collectedHeap.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/valuetypenode.hpp"
47 #include "runtime/deoptimization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49
50 //----------------------------GraphKit-----------------------------------------
51 // Main utility constructor.
52 GraphKit::GraphKit(JVMState* jvms)
53 : Phase(Phase::Parser),
54 _env(C->env()),
55 _gvn(*C->initial_gvn())
56 {
57 _exceptions = jvms->map()->next_exception();
58 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
59 set_jvms(jvms);
60 }
61
62 // Private constructor for parser.
63 GraphKit::GraphKit()
64 : Phase(Phase::Parser),
65 _env(C->env()),
66 _gvn(*C->initial_gvn())
67 {
68 _exceptions = NULL;
69 set_map(NULL);
70 debug_only(_sp = -99);
71 debug_only(set_bci(-99));
72 }
73
74
75
76 //---------------------------clean_stack---------------------------------------
77 // Clear away rubbish from the stack area of the JVM state.
78 // This destroys any arguments that may be waiting on the stack.
79 void GraphKit::clean_stack(int from_sp) {
80 SafePointNode* map = this->map();
81 JVMState* jvms = this->jvms();
82 int stk_size = jvms->stk_size();
83 int stkoff = jvms->stkoff();
84 Node* top = this->top();
85 for (int i = from_sp; i < stk_size; i++) {
86 if (map->in(stkoff + i) != top) {
87 map->set_req(stkoff + i, top);
88 }
89 }
90 }
91
92
93 //--------------------------------sync_jvms-----------------------------------
94 // Make sure our current jvms agrees with our parse state.
95 JVMState* GraphKit::sync_jvms() const {
96 JVMState* jvms = this->jvms();
97 jvms->set_bci(bci()); // Record the new bci in the JVMState
98 jvms->set_sp(sp()); // Record the new sp in the JVMState
99 assert(jvms_in_sync(), "jvms is now in sync");
100 return jvms;
101 }
102
103 //--------------------------------sync_jvms_for_reexecute---------------------
104 // Make sure our current jvms agrees with our parse state. This version
105 // uses the reexecute_sp for reexecuting bytecodes.
106 JVMState* GraphKit::sync_jvms_for_reexecute() {
107 JVMState* jvms = this->jvms();
108 jvms->set_bci(bci()); // Record the new bci in the JVMState
109 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
110 return jvms;
111 }
112
113 #ifdef ASSERT
114 bool GraphKit::jvms_in_sync() const {
115 Parse* parse = is_Parse();
116 if (parse == NULL) {
117 if (bci() != jvms()->bci()) return false;
118 if (sp() != (int)jvms()->sp()) return false;
119 return true;
120 }
121 if (jvms()->method() != parse->method()) return false;
122 if (jvms()->bci() != parse->bci()) return false;
123 int jvms_sp = jvms()->sp();
124 if (jvms_sp != parse->sp()) return false;
125 int jvms_depth = jvms()->depth();
126 if (jvms_depth != parse->depth()) return false;
127 return true;
128 }
129
130 // Local helper checks for special internal merge points
131 // used to accumulate and merge exception states.
132 // They are marked by the region's in(0) edge being the map itself.
133 // Such merge points must never "escape" into the parser at large,
134 // until they have been handed to gvn.transform.
135 static bool is_hidden_merge(Node* reg) {
136 if (reg == NULL) return false;
137 if (reg->is_Phi()) {
138 reg = reg->in(0);
139 if (reg == NULL) return false;
140 }
141 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
142 }
143
144 void GraphKit::verify_map() const {
145 if (map() == NULL) return; // null map is OK
146 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
147 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
148 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
149 }
150
151 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
152 assert(ex_map->next_exception() == NULL, "not already part of a chain");
153 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
154 }
155 #endif
156
157 //---------------------------stop_and_kill_map---------------------------------
158 // Set _map to NULL, signalling a stop to further bytecode execution.
159 // First smash the current map's control to a constant, to mark it dead.
160 void GraphKit::stop_and_kill_map() {
161 SafePointNode* dead_map = stop();
162 if (dead_map != NULL) {
163 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
164 assert(dead_map->is_killed(), "must be so marked");
165 }
166 }
167
168
169 //--------------------------------stopped--------------------------------------
170 // Tell if _map is NULL, or control is top.
171 bool GraphKit::stopped() {
172 if (map() == NULL) return true;
173 else if (control() == top()) return true;
174 else return false;
175 }
176
177
178 //-----------------------------has_ex_handler----------------------------------
179 // Tell if this method or any caller method has exception handlers.
180 bool GraphKit::has_ex_handler() {
181 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
182 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
183 return true;
184 }
185 }
186 return false;
187 }
188
189 //------------------------------save_ex_oop------------------------------------
190 // Save an exception without blowing stack contents or other JVM state.
191 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
192 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
193 ex_map->add_req(ex_oop);
194 debug_only(verify_exception_state(ex_map));
195 }
196
197 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
198 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
199 Node* ex_oop = ex_map->in(ex_map->req()-1);
200 if (clear_it) ex_map->del_req(ex_map->req()-1);
201 return ex_oop;
202 }
203
204 //-----------------------------saved_ex_oop------------------------------------
205 // Recover a saved exception from its map.
206 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
207 return common_saved_ex_oop(ex_map, false);
208 }
209
210 //--------------------------clear_saved_ex_oop---------------------------------
211 // Erase a previously saved exception from its map.
212 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
213 return common_saved_ex_oop(ex_map, true);
214 }
215
216 #ifdef ASSERT
217 //---------------------------has_saved_ex_oop----------------------------------
218 // Erase a previously saved exception from its map.
219 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
220 return ex_map->req() == ex_map->jvms()->endoff()+1;
221 }
222 #endif
223
224 //-------------------------make_exception_state--------------------------------
225 // Turn the current JVM state into an exception state, appending the ex_oop.
226 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
227 sync_jvms();
228 SafePointNode* ex_map = stop(); // do not manipulate this map any more
229 set_saved_ex_oop(ex_map, ex_oop);
230 return ex_map;
231 }
232
233
234 //--------------------------add_exception_state--------------------------------
235 // Add an exception to my list of exceptions.
236 void GraphKit::add_exception_state(SafePointNode* ex_map) {
237 if (ex_map == NULL || ex_map->control() == top()) {
238 return;
239 }
240 #ifdef ASSERT
241 verify_exception_state(ex_map);
242 if (has_exceptions()) {
243 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
244 }
245 #endif
246
247 // If there is already an exception of exactly this type, merge with it.
248 // In particular, null-checks and other low-level exceptions common up here.
249 Node* ex_oop = saved_ex_oop(ex_map);
250 const Type* ex_type = _gvn.type(ex_oop);
251 if (ex_oop == top()) {
252 // No action needed.
253 return;
254 }
255 assert(ex_type->isa_instptr(), "exception must be an instance");
256 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
257 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
258 // We check sp also because call bytecodes can generate exceptions
259 // both before and after arguments are popped!
260 if (ex_type2 == ex_type
261 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
262 combine_exception_states(ex_map, e2);
263 return;
264 }
265 }
266
267 // No pre-existing exception of the same type. Chain it on the list.
268 push_exception_state(ex_map);
269 }
270
271 //-----------------------add_exception_states_from-----------------------------
272 void GraphKit::add_exception_states_from(JVMState* jvms) {
273 SafePointNode* ex_map = jvms->map()->next_exception();
274 if (ex_map != NULL) {
275 jvms->map()->set_next_exception(NULL);
276 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
277 next_map = ex_map->next_exception();
278 ex_map->set_next_exception(NULL);
279 add_exception_state(ex_map);
280 }
281 }
282 }
283
284 //-----------------------transfer_exceptions_into_jvms-------------------------
285 JVMState* GraphKit::transfer_exceptions_into_jvms() {
286 if (map() == NULL) {
287 // We need a JVMS to carry the exceptions, but the map has gone away.
288 // Create a scratch JVMS, cloned from any of the exception states...
289 if (has_exceptions()) {
290 _map = _exceptions;
291 _map = clone_map();
292 _map->set_next_exception(NULL);
293 clear_saved_ex_oop(_map);
294 debug_only(verify_map());
295 } else {
296 // ...or created from scratch
297 JVMState* jvms = new (C) JVMState(_method, NULL);
298 jvms->set_bci(_bci);
299 jvms->set_sp(_sp);
300 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
301 set_jvms(jvms);
302 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
303 set_all_memory(top());
304 while (map()->req() < jvms->endoff()) map()->add_req(top());
305 }
306 // (This is a kludge, in case you didn't notice.)
307 set_control(top());
308 }
309 JVMState* jvms = sync_jvms();
310 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
311 jvms->map()->set_next_exception(_exceptions);
312 _exceptions = NULL; // done with this set of exceptions
313 return jvms;
314 }
315
316 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
317 assert(is_hidden_merge(dstphi), "must be a special merge node");
318 assert(is_hidden_merge(srcphi), "must be a special merge node");
319 uint limit = srcphi->req();
320 for (uint i = PhiNode::Input; i < limit; i++) {
321 dstphi->add_req(srcphi->in(i));
322 }
323 }
324 static inline void add_one_req(Node* dstphi, Node* src) {
325 assert(is_hidden_merge(dstphi), "must be a special merge node");
326 assert(!is_hidden_merge(src), "must not be a special merge node");
327 dstphi->add_req(src);
328 }
329
330 //-----------------------combine_exception_states------------------------------
331 // This helper function combines exception states by building phis on a
332 // specially marked state-merging region. These regions and phis are
333 // untransformed, and can build up gradually. The region is marked by
334 // having a control input of its exception map, rather than NULL. Such
335 // regions do not appear except in this function, and in use_exception_state.
336 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
337 if (failing()) return; // dying anyway...
338 JVMState* ex_jvms = ex_map->_jvms;
339 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
340 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
341 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
342 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
343 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
344 assert(ex_map->req() == phi_map->req(), "matching maps");
345 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
346 Node* hidden_merge_mark = root();
347 Node* region = phi_map->control();
348 MergeMemNode* phi_mem = phi_map->merged_memory();
349 MergeMemNode* ex_mem = ex_map->merged_memory();
350 if (region->in(0) != hidden_merge_mark) {
351 // The control input is not (yet) a specially-marked region in phi_map.
352 // Make it so, and build some phis.
353 region = new RegionNode(2);
354 _gvn.set_type(region, Type::CONTROL);
355 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
356 region->init_req(1, phi_map->control());
357 phi_map->set_control(region);
358 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
359 record_for_igvn(io_phi);
360 _gvn.set_type(io_phi, Type::ABIO);
361 phi_map->set_i_o(io_phi);
362 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
363 Node* m = mms.memory();
364 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
365 record_for_igvn(m_phi);
366 _gvn.set_type(m_phi, Type::MEMORY);
367 mms.set_memory(m_phi);
368 }
369 }
370
371 // Either or both of phi_map and ex_map might already be converted into phis.
372 Node* ex_control = ex_map->control();
373 // if there is special marking on ex_map also, we add multiple edges from src
374 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
375 // how wide was the destination phi_map, originally?
376 uint orig_width = region->req();
377
378 if (add_multiple) {
379 add_n_reqs(region, ex_control);
380 add_n_reqs(phi_map->i_o(), ex_map->i_o());
381 } else {
382 // ex_map has no merges, so we just add single edges everywhere
383 add_one_req(region, ex_control);
384 add_one_req(phi_map->i_o(), ex_map->i_o());
385 }
386 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
387 if (mms.is_empty()) {
388 // get a copy of the base memory, and patch some inputs into it
389 const TypePtr* adr_type = mms.adr_type(C);
390 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
391 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
392 mms.set_memory(phi);
393 // Prepare to append interesting stuff onto the newly sliced phi:
394 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
395 }
396 // Append stuff from ex_map:
397 if (add_multiple) {
398 add_n_reqs(mms.memory(), mms.memory2());
399 } else {
400 add_one_req(mms.memory(), mms.memory2());
401 }
402 }
403 uint limit = ex_map->req();
404 for (uint i = TypeFunc::Parms; i < limit; i++) {
405 // Skip everything in the JVMS after tos. (The ex_oop follows.)
406 if (i == tos) i = ex_jvms->monoff();
407 Node* src = ex_map->in(i);
408 Node* dst = phi_map->in(i);
409 if (src != dst) {
410 PhiNode* phi;
411 if (dst->in(0) != region) {
412 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
413 record_for_igvn(phi);
414 _gvn.set_type(phi, phi->type());
415 phi_map->set_req(i, dst);
416 // Prepare to append interesting stuff onto the new phi:
417 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
418 } else {
419 assert(dst->is_Phi(), "nobody else uses a hidden region");
420 phi = dst->as_Phi();
421 }
422 if (add_multiple && src->in(0) == ex_control) {
423 // Both are phis.
424 add_n_reqs(dst, src);
425 } else {
426 while (dst->req() < region->req()) add_one_req(dst, src);
427 }
428 const Type* srctype = _gvn.type(src);
429 if (phi->type() != srctype) {
430 const Type* dsttype = phi->type()->meet_speculative(srctype);
431 if (phi->type() != dsttype) {
432 phi->set_type(dsttype);
433 _gvn.set_type(phi, dsttype);
434 }
435 }
436 }
437 }
438 phi_map->merge_replaced_nodes_with(ex_map);
439 }
440
441 //--------------------------use_exception_state--------------------------------
442 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
443 if (failing()) { stop(); return top(); }
444 Node* region = phi_map->control();
445 Node* hidden_merge_mark = root();
446 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
447 Node* ex_oop = clear_saved_ex_oop(phi_map);
448 if (region->in(0) == hidden_merge_mark) {
449 // Special marking for internal ex-states. Process the phis now.
450 region->set_req(0, region); // now it's an ordinary region
451 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
452 // Note: Setting the jvms also sets the bci and sp.
453 set_control(_gvn.transform(region));
454 uint tos = jvms()->stkoff() + sp();
455 for (uint i = 1; i < tos; i++) {
456 Node* x = phi_map->in(i);
457 if (x->in(0) == region) {
458 assert(x->is_Phi(), "expected a special phi");
459 phi_map->set_req(i, _gvn.transform(x));
460 }
461 }
462 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
463 Node* x = mms.memory();
464 if (x->in(0) == region) {
465 assert(x->is_Phi(), "nobody else uses a hidden region");
466 mms.set_memory(_gvn.transform(x));
467 }
468 }
469 if (ex_oop->in(0) == region) {
470 assert(ex_oop->is_Phi(), "expected a special phi");
471 ex_oop = _gvn.transform(ex_oop);
472 }
473 } else {
474 set_jvms(phi_map->jvms());
475 }
476
477 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
478 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
479 return ex_oop;
480 }
481
482 //---------------------------------java_bc-------------------------------------
483 Bytecodes::Code GraphKit::java_bc() const {
484 ciMethod* method = this->method();
485 int bci = this->bci();
486 if (method != NULL && bci != InvocationEntryBci)
487 return method->java_code_at_bci(bci);
488 else
489 return Bytecodes::_illegal;
490 }
491
492 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
493 bool must_throw) {
494 // if the exception capability is set, then we will generate code
495 // to check the JavaThread.should_post_on_exceptions flag to see
496 // if we actually need to report exception events (for this
497 // thread). If we don't need to report exception events, we will
498 // take the normal fast path provided by add_exception_events. If
499 // exception event reporting is enabled for this thread, we will
500 // take the uncommon_trap in the BuildCutout below.
501
502 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
503 Node* jthread = _gvn.transform(new ThreadLocalNode());
504 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
505 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
506
507 // Test the should_post_on_exceptions_flag vs. 0
508 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
509 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
510
511 // Branch to slow_path if should_post_on_exceptions_flag was true
512 { BuildCutout unless(this, tst, PROB_MAX);
513 // Do not try anything fancy if we're notifying the VM on every throw.
514 // Cf. case Bytecodes::_athrow in parse2.cpp.
515 uncommon_trap(reason, Deoptimization::Action_none,
516 (ciKlass*)NULL, (char*)NULL, must_throw);
517 }
518
519 }
520
521 //------------------------------builtin_throw----------------------------------
522 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
523 bool must_throw = true;
524
525 if (env()->jvmti_can_post_on_exceptions()) {
526 // check if we must post exception events, take uncommon trap if so
527 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
528 // here if should_post_on_exceptions is false
529 // continue on with the normal codegen
530 }
531
532 // If this particular condition has not yet happened at this
533 // bytecode, then use the uncommon trap mechanism, and allow for
534 // a future recompilation if several traps occur here.
535 // If the throw is hot, try to use a more complicated inline mechanism
536 // which keeps execution inside the compiled code.
537 bool treat_throw_as_hot = false;
538 ciMethodData* md = method()->method_data();
539
540 if (ProfileTraps) {
541 if (too_many_traps(reason)) {
542 treat_throw_as_hot = true;
543 }
544 // (If there is no MDO at all, assume it is early in
545 // execution, and that any deopts are part of the
546 // startup transient, and don't need to be remembered.)
547
548 // Also, if there is a local exception handler, treat all throws
549 // as hot if there has been at least one in this method.
550 if (C->trap_count(reason) != 0
551 && method()->method_data()->trap_count(reason) != 0
552 && has_ex_handler()) {
553 treat_throw_as_hot = true;
554 }
555 }
556
557 // If this throw happens frequently, an uncommon trap might cause
558 // a performance pothole. If there is a local exception handler,
559 // and if this particular bytecode appears to be deoptimizing often,
560 // let us handle the throw inline, with a preconstructed instance.
561 // Note: If the deopt count has blown up, the uncommon trap
562 // runtime is going to flush this nmethod, not matter what.
563 if (treat_throw_as_hot
564 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
565 // If the throw is local, we use a pre-existing instance and
566 // punt on the backtrace. This would lead to a missing backtrace
567 // (a repeat of 4292742) if the backtrace object is ever asked
568 // for its backtrace.
569 // Fixing this remaining case of 4292742 requires some flavor of
570 // escape analysis. Leave that for the future.
571 ciInstance* ex_obj = NULL;
572 switch (reason) {
573 case Deoptimization::Reason_null_check:
574 ex_obj = env()->NullPointerException_instance();
575 break;
576 case Deoptimization::Reason_div0_check:
577 ex_obj = env()->ArithmeticException_instance();
578 break;
579 case Deoptimization::Reason_range_check:
580 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
581 break;
582 case Deoptimization::Reason_class_check:
583 if (java_bc() == Bytecodes::_aastore || java_bc() == Bytecodes::_vastore) {
584 ex_obj = env()->ArrayStoreException_instance();
585 } else {
586 ex_obj = env()->ClassCastException_instance();
587 }
588 break;
589 default:
590 break;
591 }
592 if (failing()) { stop(); return; } // exception allocation might fail
593 if (ex_obj != NULL) {
594 // Cheat with a preallocated exception object.
595 if (C->log() != NULL)
596 C->log()->elem("hot_throw preallocated='1' reason='%s'",
597 Deoptimization::trap_reason_name(reason));
598 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
599 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
600
601 // Clear the detail message of the preallocated exception object.
602 // Weblogic sometimes mutates the detail message of exceptions
603 // using reflection.
604 int offset = java_lang_Throwable::get_detailMessage_offset();
605 const TypePtr* adr_typ = ex_con->add_offset(offset);
606
607 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
608 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
609 // Conservatively release stores of object references.
610 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
611
612 add_exception_state(make_exception_state(ex_node));
613 return;
614 }
615 }
616
617 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
618 // It won't be much cheaper than bailing to the interp., since we'll
619 // have to pass up all the debug-info, and the runtime will have to
620 // create the stack trace.
621
622 // Usual case: Bail to interpreter.
623 // Reserve the right to recompile if we haven't seen anything yet.
624
625 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
626 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
627 if (treat_throw_as_hot
628 && (method()->method_data()->trap_recompiled_at(bci(), m)
629 || C->too_many_traps(reason))) {
630 // We cannot afford to take more traps here. Suffer in the interpreter.
631 if (C->log() != NULL)
632 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
633 Deoptimization::trap_reason_name(reason),
634 C->trap_count(reason));
635 action = Deoptimization::Action_none;
636 }
637
638 // "must_throw" prunes the JVM state to include only the stack, if there
639 // are no local exception handlers. This should cut down on register
640 // allocation time and code size, by drastically reducing the number
641 // of in-edges on the call to the uncommon trap.
642
643 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
644 }
645
646
647 //----------------------------PreserveJVMState---------------------------------
648 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
649 debug_only(kit->verify_map());
650 _kit = kit;
651 _map = kit->map(); // preserve the map
652 _sp = kit->sp();
653 kit->set_map(clone_map ? kit->clone_map() : NULL);
654 #ifdef ASSERT
655 _bci = kit->bci();
656 Parse* parser = kit->is_Parse();
657 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
658 _block = block;
659 #endif
660 }
661 PreserveJVMState::~PreserveJVMState() {
662 GraphKit* kit = _kit;
663 #ifdef ASSERT
664 assert(kit->bci() == _bci, "bci must not shift");
665 Parse* parser = kit->is_Parse();
666 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
667 assert(block == _block, "block must not shift");
668 #endif
669 kit->set_map(_map);
670 kit->set_sp(_sp);
671 }
672
673
674 //-----------------------------BuildCutout-------------------------------------
675 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
676 : PreserveJVMState(kit)
677 {
678 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
679 SafePointNode* outer_map = _map; // preserved map is caller's
680 SafePointNode* inner_map = kit->map();
681 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
682 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
683 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
684 }
685 BuildCutout::~BuildCutout() {
686 GraphKit* kit = _kit;
687 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
688 }
689
690 //---------------------------PreserveReexecuteState----------------------------
691 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
692 assert(!kit->stopped(), "must call stopped() before");
693 _kit = kit;
694 _sp = kit->sp();
695 _reexecute = kit->jvms()->_reexecute;
696 }
697 PreserveReexecuteState::~PreserveReexecuteState() {
698 if (_kit->stopped()) return;
699 _kit->jvms()->_reexecute = _reexecute;
700 _kit->set_sp(_sp);
701 }
702
703 //------------------------------clone_map--------------------------------------
704 // Implementation of PreserveJVMState
705 //
706 // Only clone_map(...) here. If this function is only used in the
707 // PreserveJVMState class we may want to get rid of this extra
708 // function eventually and do it all there.
709
710 SafePointNode* GraphKit::clone_map() {
711 if (map() == NULL) return NULL;
712
713 // Clone the memory edge first
714 Node* mem = MergeMemNode::make(map()->memory());
715 gvn().set_type_bottom(mem);
716
717 SafePointNode *clonemap = (SafePointNode*)map()->clone();
718 JVMState* jvms = this->jvms();
719 JVMState* clonejvms = jvms->clone_shallow(C);
720 clonemap->set_memory(mem);
721 clonemap->set_jvms(clonejvms);
722 clonejvms->set_map(clonemap);
723 record_for_igvn(clonemap);
724 gvn().set_type_bottom(clonemap);
725 return clonemap;
726 }
727
728
729 //-----------------------------set_map_clone-----------------------------------
730 void GraphKit::set_map_clone(SafePointNode* m) {
731 _map = m;
732 _map = clone_map();
733 _map->set_next_exception(NULL);
734 debug_only(verify_map());
735 }
736
737
738 //----------------------------kill_dead_locals---------------------------------
739 // Detect any locals which are known to be dead, and force them to top.
740 void GraphKit::kill_dead_locals() {
741 // Consult the liveness information for the locals. If any
742 // of them are unused, then they can be replaced by top(). This
743 // should help register allocation time and cut down on the size
744 // of the deoptimization information.
745
746 // This call is made from many of the bytecode handling
747 // subroutines called from the Big Switch in do_one_bytecode.
748 // Every bytecode which might include a slow path is responsible
749 // for killing its dead locals. The more consistent we
750 // are about killing deads, the fewer useless phis will be
751 // constructed for them at various merge points.
752
753 // bci can be -1 (InvocationEntryBci). We return the entry
754 // liveness for the method.
755
756 if (method() == NULL || method()->code_size() == 0) {
757 // We are building a graph for a call to a native method.
758 // All locals are live.
759 return;
760 }
761
762 ResourceMark rm;
763
764 // Consult the liveness information for the locals. If any
765 // of them are unused, then they can be replaced by top(). This
766 // should help register allocation time and cut down on the size
767 // of the deoptimization information.
768 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
769
770 int len = (int)live_locals.size();
771 assert(len <= jvms()->loc_size(), "too many live locals");
772 for (int local = 0; local < len; local++) {
773 if (!live_locals.at(local)) {
774 set_local(local, top());
775 }
776 }
777 }
778
779 #ifdef ASSERT
780 //-------------------------dead_locals_are_killed------------------------------
781 // Return true if all dead locals are set to top in the map.
782 // Used to assert "clean" debug info at various points.
783 bool GraphKit::dead_locals_are_killed() {
784 if (method() == NULL || method()->code_size() == 0) {
785 // No locals need to be dead, so all is as it should be.
786 return true;
787 }
788
789 // Make sure somebody called kill_dead_locals upstream.
790 ResourceMark rm;
791 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
792 if (jvms->loc_size() == 0) continue; // no locals to consult
793 SafePointNode* map = jvms->map();
794 ciMethod* method = jvms->method();
795 int bci = jvms->bci();
796 if (jvms == this->jvms()) {
797 bci = this->bci(); // it might not yet be synched
798 }
799 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
800 int len = (int)live_locals.size();
801 if (!live_locals.is_valid() || len == 0)
802 // This method is trivial, or is poisoned by a breakpoint.
803 return true;
804 assert(len == jvms->loc_size(), "live map consistent with locals map");
805 for (int local = 0; local < len; local++) {
806 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
807 if (PrintMiscellaneous && (Verbose || WizardMode)) {
808 tty->print_cr("Zombie local %d: ", local);
809 jvms->dump();
810 }
811 return false;
812 }
813 }
814 }
815 return true;
816 }
817
818 #endif //ASSERT
819
820 // Helper function for adding JVMState and debug information to node
821 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
822 // Add the safepoint edges to the call (or other safepoint).
823
824 // Make sure dead locals are set to top. This
825 // should help register allocation time and cut down on the size
826 // of the deoptimization information.
827 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
828
829 // Walk the inline list to fill in the correct set of JVMState's
830 // Also fill in the associated edges for each JVMState.
831
832 // If the bytecode needs to be reexecuted we need to put
833 // the arguments back on the stack.
834 const bool should_reexecute = jvms()->should_reexecute();
835 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
836
837 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
838 // undefined if the bci is different. This is normal for Parse but it
839 // should not happen for LibraryCallKit because only one bci is processed.
840 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
841 "in LibraryCallKit the reexecute bit should not change");
842
843 // If we are guaranteed to throw, we can prune everything but the
844 // input to the current bytecode.
845 bool can_prune_locals = false;
846 uint stack_slots_not_pruned = 0;
847 int inputs = 0, depth = 0;
848 if (must_throw) {
849 assert(method() == youngest_jvms->method(), "sanity");
850 if (compute_stack_effects(inputs, depth)) {
851 can_prune_locals = true;
852 stack_slots_not_pruned = inputs;
853 }
854 }
855
856 if (env()->should_retain_local_variables()) {
857 // At any safepoint, this method can get breakpointed, which would
858 // then require an immediate deoptimization.
859 can_prune_locals = false; // do not prune locals
860 stack_slots_not_pruned = 0;
861 }
862
863 C->add_safepoint_edges(call, youngest_jvms, can_prune_locals, stack_slots_not_pruned);
864 }
865
866 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
867 Bytecodes::Code code = java_bc();
868 if (code == Bytecodes::_wide) {
869 code = method()->java_code_at_bci(bci() + 1);
870 }
871
872 BasicType rtype = T_ILLEGAL;
873 int rsize = 0;
874
875 if (code != Bytecodes::_illegal) {
876 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
877 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
878 if (rtype < T_CONFLICT)
879 rsize = type2size[rtype];
880 }
881
882 switch (code) {
883 case Bytecodes::_illegal:
884 return false;
885
886 case Bytecodes::_ldc:
887 case Bytecodes::_ldc_w:
888 case Bytecodes::_ldc2_w:
889 inputs = 0;
890 break;
891
892 case Bytecodes::_dup: inputs = 1; break;
893 case Bytecodes::_dup_x1: inputs = 2; break;
894 case Bytecodes::_dup_x2: inputs = 3; break;
895 case Bytecodes::_dup2: inputs = 2; break;
896 case Bytecodes::_dup2_x1: inputs = 3; break;
897 case Bytecodes::_dup2_x2: inputs = 4; break;
898 case Bytecodes::_swap: inputs = 2; break;
899 case Bytecodes::_arraylength: inputs = 1; break;
900
901 case Bytecodes::_getstatic:
902 case Bytecodes::_putstatic:
903 case Bytecodes::_getfield:
904 case Bytecodes::_putfield:
905 {
906 bool ignored_will_link;
907 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
908 int size = field->type()->size();
909 bool is_get = (depth >= 0), is_static = (depth & 1);
910 inputs = (is_static ? 0 : 1);
911 if (is_get) {
912 depth = size - inputs;
913 } else {
914 inputs += size; // putxxx pops the value from the stack
915 depth = - inputs;
916 }
917 }
918 break;
919
920 case Bytecodes::_invokevirtual:
921 case Bytecodes::_invokespecial:
922 case Bytecodes::_invokestatic:
923 case Bytecodes::_invokedynamic:
924 case Bytecodes::_invokeinterface:
925 {
926 bool ignored_will_link;
927 ciSignature* declared_signature = NULL;
928 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
929 assert(declared_signature != NULL, "cannot be null");
930 inputs = declared_signature->arg_size_for_bc(code);
931 int size = declared_signature->return_type()->size();
932 depth = size - inputs;
933 }
934 break;
935
936 case Bytecodes::_multianewarray:
937 {
938 ciBytecodeStream iter(method());
939 iter.reset_to_bci(bci());
940 iter.next();
941 inputs = iter.get_dimensions();
942 assert(rsize == 1, "");
943 depth = rsize - inputs;
944 }
945 break;
946
947 case Bytecodes::_vwithfield: {
948 bool ignored_will_link;
949 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
950 int size = field->type()->size();
951 inputs = size+1;
952 depth = rsize - inputs;
953 break;
954 }
955
956 case Bytecodes::_ireturn:
957 case Bytecodes::_lreturn:
958 case Bytecodes::_freturn:
959 case Bytecodes::_dreturn:
960 case Bytecodes::_areturn:
961 case Bytecodes::_vreturn:
962 assert(rsize == -depth, "");
963 inputs = rsize;
964 break;
965
966 case Bytecodes::_jsr:
967 case Bytecodes::_jsr_w:
968 inputs = 0;
969 depth = 1; // S.B. depth=1, not zero
970 break;
971
972 default:
973 // bytecode produces a typed result
974 inputs = rsize - depth;
975 assert(inputs >= 0, "");
976 break;
977 }
978
979 #ifdef ASSERT
980 // spot check
981 int outputs = depth + inputs;
982 assert(outputs >= 0, "sanity");
983 switch (code) {
984 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
985 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
986 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
987 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
988 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
989 default: break;
990 }
991 #endif //ASSERT
992
993 return true;
994 }
995
996
997
998 //------------------------------basic_plus_adr---------------------------------
999 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1000 // short-circuit a common case
1001 if (offset == intcon(0)) return ptr;
1002 return _gvn.transform( new AddPNode(base, ptr, offset) );
1003 }
1004
1005 Node* GraphKit::ConvI2L(Node* offset) {
1006 // short-circuit a common case
1007 jint offset_con = find_int_con(offset, Type::OffsetBot);
1008 if (offset_con != Type::OffsetBot) {
1009 return longcon((jlong) offset_con);
1010 }
1011 return _gvn.transform( new ConvI2LNode(offset));
1012 }
1013
1014 Node* GraphKit::ConvI2UL(Node* offset) {
1015 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1016 if (offset_con != (juint) Type::OffsetBot) {
1017 return longcon((julong) offset_con);
1018 }
1019 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1020 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1021 return _gvn.transform( new AndLNode(conv, mask) );
1022 }
1023
1024 Node* GraphKit::ConvL2I(Node* offset) {
1025 // short-circuit a common case
1026 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1027 if (offset_con != (jlong)Type::OffsetBot) {
1028 return intcon((int) offset_con);
1029 }
1030 return _gvn.transform( new ConvL2INode(offset));
1031 }
1032
1033 //-------------------------load_object_klass-----------------------------------
1034 Node* GraphKit::load_object_klass(Node* obj) {
1035 // Special-case a fresh allocation to avoid building nodes:
1036 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1037 if (akls != NULL) return akls;
1038 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1039 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1040 }
1041
1042 //-------------------------load_array_length-----------------------------------
1043 Node* GraphKit::load_array_length(Node* array) {
1044 // Special-case a fresh allocation to avoid building nodes:
1045 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1046 Node *alen;
1047 if (alloc == NULL) {
1048 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1049 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1050 } else {
1051 alen = alloc->Ideal_length();
1052 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1053 if (ccast != alen) {
1054 alen = _gvn.transform(ccast);
1055 }
1056 }
1057 return alen;
1058 }
1059
1060 //------------------------------do_null_check----------------------------------
1061 // Helper function to do a NULL pointer check. Returned value is
1062 // the incoming address with NULL casted away. You are allowed to use the
1063 // not-null value only if you are control dependent on the test.
1064 #ifndef PRODUCT
1065 extern int explicit_null_checks_inserted,
1066 explicit_null_checks_elided;
1067 #endif
1068 Node* GraphKit::null_check_common(Node* value, BasicType type,
1069 // optional arguments for variations:
1070 bool assert_null,
1071 Node* *null_control,
1072 bool speculative) {
1073 assert(!assert_null || null_control == NULL, "not both at once");
1074 if (stopped()) return top();
1075 NOT_PRODUCT(explicit_null_checks_inserted++);
1076
1077 // Construct NULL check
1078 Node *chk = NULL;
1079 switch(type) {
1080 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1081 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1082 case T_VALUETYPE : // fall through
1083 case T_ARRAY : // fall through
1084 type = T_OBJECT; // simplify further tests
1085 case T_OBJECT : {
1086 const Type *t = _gvn.type( value );
1087
1088 const TypeOopPtr* tp = t->isa_oopptr();
1089 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1090 // Only for do_null_check, not any of its siblings:
1091 && !assert_null && null_control == NULL) {
1092 // Usually, any field access or invocation on an unloaded oop type
1093 // will simply fail to link, since the statically linked class is
1094 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1095 // the static class is loaded but the sharper oop type is not.
1096 // Rather than checking for this obscure case in lots of places,
1097 // we simply observe that a null check on an unloaded class
1098 // will always be followed by a nonsense operation, so we
1099 // can just issue the uncommon trap here.
1100 // Our access to the unloaded class will only be correct
1101 // after it has been loaded and initialized, which requires
1102 // a trip through the interpreter.
1103 #ifndef PRODUCT
1104 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1105 #endif
1106 uncommon_trap(Deoptimization::Reason_unloaded,
1107 Deoptimization::Action_reinterpret,
1108 tp->klass(), "!loaded");
1109 return top();
1110 }
1111
1112 if (assert_null) {
1113 // See if the type is contained in NULL_PTR.
1114 // If so, then the value is already null.
1115 if (t->higher_equal(TypePtr::NULL_PTR)) {
1116 NOT_PRODUCT(explicit_null_checks_elided++);
1117 return value; // Elided null assert quickly!
1118 }
1119 } else {
1120 // See if mixing in the NULL pointer changes type.
1121 // If so, then the NULL pointer was not allowed in the original
1122 // type. In other words, "value" was not-null.
1123 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1124 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1125 NOT_PRODUCT(explicit_null_checks_elided++);
1126 return value; // Elided null check quickly!
1127 }
1128 }
1129 chk = new CmpPNode( value, null() );
1130 break;
1131 }
1132
1133 default:
1134 fatal("unexpected type: %s", type2name(type));
1135 }
1136 assert(chk != NULL, "sanity check");
1137 chk = _gvn.transform(chk);
1138
1139 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1140 BoolNode *btst = new BoolNode( chk, btest);
1141 Node *tst = _gvn.transform( btst );
1142
1143 //-----------
1144 // if peephole optimizations occurred, a prior test existed.
1145 // If a prior test existed, maybe it dominates as we can avoid this test.
1146 if (tst != btst && type == T_OBJECT) {
1147 // At this point we want to scan up the CFG to see if we can
1148 // find an identical test (and so avoid this test altogether).
1149 Node *cfg = control();
1150 int depth = 0;
1151 while( depth < 16 ) { // Limit search depth for speed
1152 if( cfg->Opcode() == Op_IfTrue &&
1153 cfg->in(0)->in(1) == tst ) {
1154 // Found prior test. Use "cast_not_null" to construct an identical
1155 // CastPP (and hence hash to) as already exists for the prior test.
1156 // Return that casted value.
1157 if (assert_null) {
1158 replace_in_map(value, null());
1159 return null(); // do not issue the redundant test
1160 }
1161 Node *oldcontrol = control();
1162 set_control(cfg);
1163 Node *res = cast_not_null(value);
1164 set_control(oldcontrol);
1165 NOT_PRODUCT(explicit_null_checks_elided++);
1166 return res;
1167 }
1168 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1169 if (cfg == NULL) break; // Quit at region nodes
1170 depth++;
1171 }
1172 }
1173
1174 //-----------
1175 // Branch to failure if null
1176 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1177 Deoptimization::DeoptReason reason;
1178 if (assert_null) {
1179 reason = Deoptimization::reason_null_assert(speculative);
1180 } else if (type == T_OBJECT) {
1181 reason = Deoptimization::reason_null_check(speculative);
1182 } else {
1183 reason = Deoptimization::Reason_div0_check;
1184 }
1185 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1186 // ciMethodData::has_trap_at will return a conservative -1 if any
1187 // must-be-null assertion has failed. This could cause performance
1188 // problems for a method after its first do_null_assert failure.
1189 // Consider using 'Reason_class_check' instead?
1190
1191 // To cause an implicit null check, we set the not-null probability
1192 // to the maximum (PROB_MAX). For an explicit check the probability
1193 // is set to a smaller value.
1194 if (null_control != NULL || too_many_traps(reason)) {
1195 // probability is less likely
1196 ok_prob = PROB_LIKELY_MAG(3);
1197 } else if (!assert_null &&
1198 (ImplicitNullCheckThreshold > 0) &&
1199 method() != NULL &&
1200 (method()->method_data()->trap_count(reason)
1201 >= (uint)ImplicitNullCheckThreshold)) {
1202 ok_prob = PROB_LIKELY_MAG(3);
1203 }
1204
1205 if (null_control != NULL) {
1206 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1207 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1208 set_control( _gvn.transform( new IfTrueNode(iff)));
1209 #ifndef PRODUCT
1210 if (null_true == top()) {
1211 explicit_null_checks_elided++;
1212 }
1213 #endif
1214 (*null_control) = null_true;
1215 } else {
1216 BuildCutout unless(this, tst, ok_prob);
1217 // Check for optimizer eliding test at parse time
1218 if (stopped()) {
1219 // Failure not possible; do not bother making uncommon trap.
1220 NOT_PRODUCT(explicit_null_checks_elided++);
1221 } else if (assert_null) {
1222 uncommon_trap(reason,
1223 Deoptimization::Action_make_not_entrant,
1224 NULL, "assert_null");
1225 } else {
1226 replace_in_map(value, zerocon(type));
1227 builtin_throw(reason);
1228 }
1229 }
1230
1231 // Must throw exception, fall-thru not possible?
1232 if (stopped()) {
1233 return top(); // No result
1234 }
1235
1236 if (assert_null) {
1237 // Cast obj to null on this path.
1238 replace_in_map(value, zerocon(type));
1239 return zerocon(type);
1240 }
1241
1242 // Cast obj to not-null on this path, if there is no null_control.
1243 // (If there is a null_control, a non-null value may come back to haunt us.)
1244 if (type == T_OBJECT) {
1245 Node* cast = cast_not_null(value, false);
1246 if (null_control == NULL || (*null_control) == top())
1247 replace_in_map(value, cast);
1248 value = cast;
1249 }
1250
1251 return value;
1252 }
1253
1254
1255 //------------------------------cast_not_null----------------------------------
1256 // Cast obj to not-null on this path
1257 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1258 const Type *t = _gvn.type(obj);
1259 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1260 // Object is already not-null?
1261 if( t == t_not_null ) return obj;
1262
1263 Node *cast = new CastPPNode(obj,t_not_null);
1264 cast->init_req(0, control());
1265 cast = _gvn.transform( cast );
1266
1267 // Scan for instances of 'obj' in the current JVM mapping.
1268 // These instances are known to be not-null after the test.
1269 if (do_replace_in_map)
1270 replace_in_map(obj, cast);
1271
1272 return cast; // Return casted value
1273 }
1274
1275 // Sometimes in intrinsics, we implicitly know an object is not null
1276 // (there's no actual null check) so we can cast it to not null. In
1277 // the course of optimizations, the input to the cast can become null.
1278 // In that case that data path will die and we need the control path
1279 // to become dead as well to keep the graph consistent. So we have to
1280 // add a check for null for which one branch can't be taken. It uses
1281 // an Opaque4 node that will cause the check to be removed after loop
1282 // opts so the test goes away and the compiled code doesn't execute a
1283 // useless check.
1284 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1285 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1286 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1287 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1288 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1289 _gvn.set_type(iff, iff->Value(&_gvn));
1290 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1291 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1292 Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1293 C->root()->add_req(halt);
1294 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1295 set_control(if_t);
1296 return cast_not_null(value, do_replace_in_map);
1297 }
1298
1299
1300 //--------------------------replace_in_map-------------------------------------
1301 void GraphKit::replace_in_map(Node* old, Node* neww) {
1302 if (old == neww) {
1303 return;
1304 }
1305
1306 map()->replace_edge(old, neww);
1307
1308 // Note: This operation potentially replaces any edge
1309 // on the map. This includes locals, stack, and monitors
1310 // of the current (innermost) JVM state.
1311
1312 // don't let inconsistent types from profiling escape this
1313 // method
1314
1315 const Type* told = _gvn.type(old);
1316 const Type* tnew = _gvn.type(neww);
1317
1318 if (!tnew->higher_equal(told)) {
1319 return;
1320 }
1321
1322 map()->record_replaced_node(old, neww);
1323 }
1324
1325
1326 //=============================================================================
1327 //--------------------------------memory---------------------------------------
1328 Node* GraphKit::memory(uint alias_idx) {
1329 MergeMemNode* mem = merged_memory();
1330 Node* p = mem->memory_at(alias_idx);
1331 _gvn.set_type(p, Type::MEMORY); // must be mapped
1332 return p;
1333 }
1334
1335 //-----------------------------reset_memory------------------------------------
1336 Node* GraphKit::reset_memory() {
1337 Node* mem = map()->memory();
1338 // do not use this node for any more parsing!
1339 debug_only( map()->set_memory((Node*)NULL) );
1340 return _gvn.transform( mem );
1341 }
1342
1343 //------------------------------set_all_memory---------------------------------
1344 void GraphKit::set_all_memory(Node* newmem) {
1345 Node* mergemem = MergeMemNode::make(newmem);
1346 gvn().set_type_bottom(mergemem);
1347 map()->set_memory(mergemem);
1348 }
1349
1350 //------------------------------set_all_memory_call----------------------------
1351 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1352 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1353 set_all_memory(newmem);
1354 }
1355
1356 //=============================================================================
1357 //
1358 // parser factory methods for MemNodes
1359 //
1360 // These are layered on top of the factory methods in LoadNode and StoreNode,
1361 // and integrate with the parser's memory state and _gvn engine.
1362 //
1363
1364 // factory methods in "int adr_idx"
1365 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1366 int adr_idx,
1367 MemNode::MemOrd mo,
1368 LoadNode::ControlDependency control_dependency,
1369 bool require_atomic_access,
1370 bool unaligned,
1371 bool mismatched) {
1372 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1373 const TypePtr* adr_type = NULL; // debug-mode-only argument
1374 debug_only(adr_type = C->get_adr_type(adr_idx));
1375 Node* mem = memory(adr_idx);
1376 Node* ld;
1377 if (require_atomic_access && bt == T_LONG) {
1378 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1379 } else if (require_atomic_access && bt == T_DOUBLE) {
1380 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1381 } else {
1382 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1383 }
1384 ld = _gvn.transform(ld);
1385 if (bt == T_VALUETYPE) {
1386 // Load non-flattened value type from memory. Add a null check and let the
1387 // interpreter take care of initializing the field to the default value type.
1388 Node* null_ctl = top();
1389 ld = null_check_common(ld, bt, false, &null_ctl, false);
1390 if (null_ctl != top()) {
1391 assert(!adr_type->isa_aryptr(), "value type array must be initialized");
1392 PreserveJVMState pjvms(this);
1393 set_control(null_ctl);
1394 uncommon_trap(Deoptimization::reason_null_check(false), Deoptimization::Action_maybe_recompile,
1395 t->is_valuetypeptr()->value_type()->value_klass(), "uninitialized non-flattened value type");
1396 }
1397 ld = ValueTypeNode::make(gvn(), map()->memory(), ld);
1398 } else if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1399 // Improve graph before escape analysis and boxing elimination.
1400 record_for_igvn(ld);
1401 }
1402 return ld;
1403 }
1404
1405 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1406 int adr_idx,
1407 MemNode::MemOrd mo,
1408 bool require_atomic_access,
1409 bool unaligned,
1410 bool mismatched) {
1411 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1412 const TypePtr* adr_type = NULL;
1413 debug_only(adr_type = C->get_adr_type(adr_idx));
1414 Node *mem = memory(adr_idx);
1415 Node* st;
1416 if (require_atomic_access && bt == T_LONG) {
1417 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1418 } else if (require_atomic_access && bt == T_DOUBLE) {
1419 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1420 } else {
1421 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1422 }
1423 if (unaligned) {
1424 st->as_Store()->set_unaligned_access();
1425 }
1426 if (mismatched) {
1427 st->as_Store()->set_mismatched_access();
1428 }
1429 st = _gvn.transform(st);
1430 set_memory(st, adr_idx);
1431 // Back-to-back stores can only remove intermediate store with DU info
1432 // so push on worklist for optimizer.
1433 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1434 record_for_igvn(st);
1435
1436 return st;
1437 }
1438
1439
1440 void GraphKit::pre_barrier(bool do_load,
1441 Node* ctl,
1442 Node* obj,
1443 Node* adr,
1444 uint adr_idx,
1445 Node* val,
1446 const TypeOopPtr* val_type,
1447 Node* pre_val,
1448 BasicType bt) {
1449
1450 BarrierSet* bs = Universe::heap()->barrier_set();
1451 set_control(ctl);
1452 switch (bs->kind()) {
1453 case BarrierSet::G1SATBCTLogging:
1454 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1455 break;
1456
1457 case BarrierSet::CardTableForRS:
1458 case BarrierSet::CardTableExtension:
1459 case BarrierSet::ModRef:
1460 break;
1461
1462 default :
1463 ShouldNotReachHere();
1464
1465 }
1466 }
1467
1468 bool GraphKit::can_move_pre_barrier() const {
1469 BarrierSet* bs = Universe::heap()->barrier_set();
1470 switch (bs->kind()) {
1471 case BarrierSet::G1SATBCTLogging:
1472 return true; // Can move it if no safepoint
1473
1474 case BarrierSet::CardTableForRS:
1475 case BarrierSet::CardTableExtension:
1476 case BarrierSet::ModRef:
1477 return true; // There is no pre-barrier
1478
1479 default :
1480 ShouldNotReachHere();
1481 }
1482 return false;
1483 }
1484
1485 void GraphKit::post_barrier(Node* ctl,
1486 Node* store,
1487 Node* obj,
1488 Node* adr,
1489 uint adr_idx,
1490 Node* val,
1491 BasicType bt,
1492 bool use_precise) {
1493 BarrierSet* bs = Universe::heap()->barrier_set();
1494 set_control(ctl);
1495 switch (bs->kind()) {
1496 case BarrierSet::G1SATBCTLogging:
1497 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1498 break;
1499
1500 case BarrierSet::CardTableForRS:
1501 case BarrierSet::CardTableExtension:
1502 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1503 break;
1504
1505 case BarrierSet::ModRef:
1506 break;
1507
1508 default :
1509 ShouldNotReachHere();
1510
1511 }
1512 }
1513
1514 Node* GraphKit::store_oop(Node* ctl,
1515 Node* obj,
1516 Node* adr,
1517 const TypePtr* adr_type,
1518 Node* val,
1519 const TypeOopPtr* val_type,
1520 BasicType bt,
1521 bool use_precise,
1522 MemNode::MemOrd mo,
1523 bool mismatched) {
1524 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1525 // could be delayed during Parse (for example, in adjust_map_after_if()).
1526 // Execute transformation here to avoid barrier generation in such case.
1527 if (_gvn.type(val) == TypePtr::NULL_PTR)
1528 val = _gvn.makecon(TypePtr::NULL_PTR);
1529
1530 set_control(ctl);
1531 if (stopped()) return top(); // Dead path ?
1532
1533 assert(bt == T_OBJECT || bt == T_VALUETYPE, "sanity");
1534 assert(val != NULL, "not dead path");
1535 uint adr_idx = C->get_alias_index(adr_type);
1536 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1537
1538 if (bt == T_VALUETYPE) {
1539 // Allocate value type and store oop
1540 val = val->as_ValueType()->allocate(this);
1541 }
1542
1543 pre_barrier(true /* do_load */,
1544 control(), obj, adr, adr_idx, val, val_type,
1545 NULL /* pre_val */,
1546 bt);
1547
1548 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1549 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1550 return store;
1551 }
1552
1553 // Could be an array or object we don't know at compile time (unsafe ref.)
1554 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1555 Node* obj, // containing obj
1556 Node* adr, // actual adress to store val at
1557 const TypePtr* adr_type,
1558 Node* val,
1559 BasicType bt,
1560 MemNode::MemOrd mo,
1561 bool mismatched) {
1562 Compile::AliasType* at = C->alias_type(adr_type);
1563 const TypeOopPtr* val_type = NULL;
1564 if (adr_type->isa_instptr()) {
1565 if (at->field() != NULL) {
1566 // known field. This code is a copy of the do_put_xxx logic.
1567 ciField* field = at->field();
1568 if (!field->type()->is_loaded()) {
1569 val_type = TypeInstPtr::BOTTOM;
1570 } else {
1571 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1572 }
1573 }
1574 } else if (adr_type->isa_aryptr()) {
1575 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1576 }
1577 if (val_type == NULL) {
1578 val_type = TypeInstPtr::BOTTOM;
1579 }
1580 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1581 }
1582
1583
1584 //-------------------------array_element_address-------------------------
1585 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1586 const TypeInt* sizetype, Node* ctrl) {
1587 uint shift = exact_log2(type2aelembytes(elembt));
1588 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1589 if (arytype_klass->is_value_array_klass()) {
1590 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass();
1591 shift = vak->log2_element_size();
1592 }
1593 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1594
1595 // short-circuit a common case (saves lots of confusing waste motion)
1596 jint idx_con = find_int_con(idx, -1);
1597 if (idx_con >= 0) {
1598 intptr_t offset = header + ((intptr_t)idx_con << shift);
1599 return basic_plus_adr(ary, offset);
1600 }
1601
1602 // must be correct type for alignment purposes
1603 Node* base = basic_plus_adr(ary, header);
1604 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1605 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1606 return basic_plus_adr(ary, base, scale);
1607 }
1608
1609 //-------------------------load_array_element-------------------------
1610 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1611 const Type* elemtype = arytype->elem();
1612 BasicType elembt = elemtype->array_element_basic_type();
1613 assert(elembt != T_VALUETYPE, "value types are not supported by this method");
1614 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1615 if (elembt == T_NARROWOOP) {
1616 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1617 }
1618 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1619 return ld;
1620 }
1621
1622 //-------------------------set_arguments_for_java_call-------------------------
1623 // Arguments (pre-popped from the stack) are taken from the JVMS.
1624 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1625 // Add the call arguments:
1626 const TypeTuple* domain = call->tf()->domain_sig();
1627 uint nargs = domain->cnt();
1628 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1629 Node* arg = argument(i-TypeFunc::Parms);
1630 if (ValueTypePassFieldsAsArgs) {
1631 if (arg->is_ValueType()) {
1632 ValueTypeNode* vt = arg->as_ValueType();
1633 if (domain->field_at(i)->is_valuetypeptr()->klass() != C->env()->___Value_klass()) {
1634 // We don't pass value type arguments by reference but instead
1635 // pass each field of the value type
1636 idx += vt->pass_fields(call, idx, *this);
1637 // If a value type argument is passed as fields, attach the Method* to the call site
1638 // to be able to access the extended signature later via attached_method_before_pc().
1639 // For example, see CompiledMethod::preserve_callee_argument_oops().
1640 call->set_override_symbolic_info(true);
1641 } else {
1642 arg = arg->as_ValueType()->allocate(this);
1643 call->init_req(idx, arg);
1644 idx++;
1645 }
1646 } else {
1647 call->init_req(idx, arg);
1648 idx++;
1649 }
1650 } else {
1651 if (arg->is_ValueType()) {
1652 // Pass value type argument via oop to callee
1653 arg = arg->as_ValueType()->allocate(this);
1654 }
1655 call->init_req(i, arg);
1656 }
1657 }
1658 }
1659
1660 //---------------------------set_edges_for_java_call---------------------------
1661 // Connect a newly created call into the current JVMS.
1662 // A return value node (if any) is returned from set_edges_for_java_call.
1663 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1664
1665 // Add the predefined inputs:
1666 call->init_req( TypeFunc::Control, control() );
1667 call->init_req( TypeFunc::I_O , i_o() );
1668 call->init_req( TypeFunc::Memory , reset_memory() );
1669 call->init_req( TypeFunc::FramePtr, frameptr() );
1670 call->init_req( TypeFunc::ReturnAdr, top() );
1671
1672 add_safepoint_edges(call, must_throw);
1673
1674 Node* xcall = _gvn.transform(call);
1675
1676 if (xcall == top()) {
1677 set_control(top());
1678 return;
1679 }
1680 assert(xcall == call, "call identity is stable");
1681
1682 // Re-use the current map to produce the result.
1683
1684 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1685 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1686 set_all_memory_call(xcall, separate_io_proj);
1687
1688 //return xcall; // no need, caller already has it
1689 }
1690
1691 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1692 if (stopped()) return top(); // maybe the call folded up?
1693
1694 // Capture the return value, if any.
1695 Node* ret;
1696 if (call->method() == NULL ||
1697 call->method()->return_type()->basic_type() == T_VOID)
1698 ret = top();
1699 else {
1700 if (!call->tf()->returns_value_type_as_fields()) {
1701 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1702 } else {
1703 // Return of multiple values (value type fields): we create a
1704 // ValueType node, each field is a projection from the call.
1705 const TypeTuple *range_sig = call->tf()->range_sig();
1706 const Type* t = range_sig->field_at(TypeFunc::Parms);
1707 assert(t->isa_valuetypeptr(), "only value types for multiple return values");
1708 ciValueKlass* vk = t->is_valuetypeptr()->value_type()->value_klass();
1709 ret = ValueTypeNode::make(_gvn, call, vk, TypeFunc::Parms+1, false);
1710 }
1711 }
1712
1713 // Note: Since any out-of-line call can produce an exception,
1714 // we always insert an I_O projection from the call into the result.
1715
1716 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1717
1718 if (separate_io_proj) {
1719 // The caller requested separate projections be used by the fall
1720 // through and exceptional paths, so replace the projections for
1721 // the fall through path.
1722 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1723 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1724 }
1725 return ret;
1726 }
1727
1728 //--------------------set_predefined_input_for_runtime_call--------------------
1729 // Reading and setting the memory state is way conservative here.
1730 // The real problem is that I am not doing real Type analysis on memory,
1731 // so I cannot distinguish card mark stores from other stores. Across a GC
1732 // point the Store Barrier and the card mark memory has to agree. I cannot
1733 // have a card mark store and its barrier split across the GC point from
1734 // either above or below. Here I get that to happen by reading ALL of memory.
1735 // A better answer would be to separate out card marks from other memory.
1736 // For now, return the input memory state, so that it can be reused
1737 // after the call, if this call has restricted memory effects.
1738 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1739 // Set fixed predefined input arguments
1740 Node* memory = reset_memory();
1741 call->init_req( TypeFunc::Control, control() );
1742 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1743 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1744 call->init_req( TypeFunc::FramePtr, frameptr() );
1745 call->init_req( TypeFunc::ReturnAdr, top() );
1746 return memory;
1747 }
1748
1749 //-------------------set_predefined_output_for_runtime_call--------------------
1750 // Set control and memory (not i_o) from the call.
1751 // If keep_mem is not NULL, use it for the output state,
1752 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1753 // If hook_mem is NULL, this call produces no memory effects at all.
1754 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1755 // then only that memory slice is taken from the call.
1756 // In the last case, we must put an appropriate memory barrier before
1757 // the call, so as to create the correct anti-dependencies on loads
1758 // preceding the call.
1759 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1760 Node* keep_mem,
1761 const TypePtr* hook_mem) {
1762 // no i/o
1763 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1764 if (keep_mem) {
1765 // First clone the existing memory state
1766 set_all_memory(keep_mem);
1767 if (hook_mem != NULL) {
1768 // Make memory for the call
1769 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1770 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1771 // We also use hook_mem to extract specific effects from arraycopy stubs.
1772 set_memory(mem, hook_mem);
1773 }
1774 // ...else the call has NO memory effects.
1775
1776 // Make sure the call advertises its memory effects precisely.
1777 // This lets us build accurate anti-dependences in gcm.cpp.
1778 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1779 "call node must be constructed correctly");
1780 } else {
1781 assert(hook_mem == NULL, "");
1782 // This is not a "slow path" call; all memory comes from the call.
1783 set_all_memory_call(call);
1784 }
1785 }
1786
1787
1788 // Replace the call with the current state of the kit.
1789 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1790 JVMState* ejvms = NULL;
1791 if (has_exceptions()) {
1792 ejvms = transfer_exceptions_into_jvms();
1793 }
1794
1795 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1796 ReplacedNodes replaced_nodes_exception;
1797 Node* ex_ctl = top();
1798
1799 SafePointNode* final_state = stop();
1800
1801 // Find all the needed outputs of this call
1802 CallProjections callprojs;
1803 call->extract_projections(&callprojs, true);
1804
1805 Node* init_mem = call->in(TypeFunc::Memory);
1806 Node* final_mem = final_state->in(TypeFunc::Memory);
1807 Node* final_ctl = final_state->in(TypeFunc::Control);
1808 Node* final_io = final_state->in(TypeFunc::I_O);
1809
1810 // Replace all the old call edges with the edges from the inlining result
1811 if (callprojs.fallthrough_catchproj != NULL) {
1812 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1813 }
1814 if (callprojs.fallthrough_memproj != NULL) {
1815 if (final_mem->is_MergeMem()) {
1816 // Parser's exits MergeMem was not transformed but may be optimized
1817 final_mem = _gvn.transform(final_mem);
1818 }
1819 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1820 }
1821 if (callprojs.fallthrough_ioproj != NULL) {
1822 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1823 }
1824
1825 // Replace the result with the new result if it exists and is used
1826 if (callprojs.resproj != NULL && result != NULL) {
1827 C->gvn_replace_by(callprojs.resproj, result);
1828 }
1829
1830 if (ejvms == NULL) {
1831 // No exception edges to simply kill off those paths
1832 if (callprojs.catchall_catchproj != NULL) {
1833 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1834 }
1835 if (callprojs.catchall_memproj != NULL) {
1836 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1837 }
1838 if (callprojs.catchall_ioproj != NULL) {
1839 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1840 }
1841 // Replace the old exception object with top
1842 if (callprojs.exobj != NULL) {
1843 C->gvn_replace_by(callprojs.exobj, C->top());
1844 }
1845 } else {
1846 GraphKit ekit(ejvms);
1847
1848 // Load my combined exception state into the kit, with all phis transformed:
1849 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1850 replaced_nodes_exception = ex_map->replaced_nodes();
1851
1852 Node* ex_oop = ekit.use_exception_state(ex_map);
1853
1854 if (callprojs.catchall_catchproj != NULL) {
1855 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1856 ex_ctl = ekit.control();
1857 }
1858 if (callprojs.catchall_memproj != NULL) {
1859 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1860 }
1861 if (callprojs.catchall_ioproj != NULL) {
1862 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1863 }
1864
1865 // Replace the old exception object with the newly created one
1866 if (callprojs.exobj != NULL) {
1867 C->gvn_replace_by(callprojs.exobj, ex_oop);
1868 }
1869 }
1870
1871 // Disconnect the call from the graph
1872 call->disconnect_inputs(NULL, C);
1873 C->gvn_replace_by(call, C->top());
1874
1875 // Clean up any MergeMems that feed other MergeMems since the
1876 // optimizer doesn't like that.
1877 if (final_mem->is_MergeMem()) {
1878 Node_List wl;
1879 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1880 Node* m = i.get();
1881 if (m->is_MergeMem() && !wl.contains(m)) {
1882 wl.push(m);
1883 }
1884 }
1885 while (wl.size() > 0) {
1886 _gvn.transform(wl.pop());
1887 }
1888 }
1889
1890 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1891 replaced_nodes.apply(C, final_ctl);
1892 }
1893 if (!ex_ctl->is_top() && do_replaced_nodes) {
1894 replaced_nodes_exception.apply(C, ex_ctl);
1895 }
1896 }
1897
1898
1899 //------------------------------increment_counter------------------------------
1900 // for statistics: increment a VM counter by 1
1901
1902 void GraphKit::increment_counter(address counter_addr) {
1903 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1904 increment_counter(adr1);
1905 }
1906
1907 void GraphKit::increment_counter(Node* counter_addr) {
1908 int adr_type = Compile::AliasIdxRaw;
1909 Node* ctrl = control();
1910 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1911 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1912 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1913 }
1914
1915
1916 //------------------------------uncommon_trap----------------------------------
1917 // Bail out to the interpreter in mid-method. Implemented by calling the
1918 // uncommon_trap blob. This helper function inserts a runtime call with the
1919 // right debug info.
1920 void GraphKit::uncommon_trap(int trap_request,
1921 ciKlass* klass, const char* comment,
1922 bool must_throw,
1923 bool keep_exact_action) {
1924 if (failing()) stop();
1925 if (stopped()) return; // trap reachable?
1926
1927 // Note: If ProfileTraps is true, and if a deopt. actually
1928 // occurs here, the runtime will make sure an MDO exists. There is
1929 // no need to call method()->ensure_method_data() at this point.
1930
1931 // Set the stack pointer to the right value for reexecution:
1932 set_sp(reexecute_sp());
1933
1934 #ifdef ASSERT
1935 if (!must_throw) {
1936 // Make sure the stack has at least enough depth to execute
1937 // the current bytecode.
1938 int inputs, ignored_depth;
1939 if (compute_stack_effects(inputs, ignored_depth)) {
1940 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1941 Bytecodes::name(java_bc()), sp(), inputs);
1942 }
1943 }
1944 #endif
1945
1946 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1947 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1948
1949 switch (action) {
1950 case Deoptimization::Action_maybe_recompile:
1951 case Deoptimization::Action_reinterpret:
1952 // Temporary fix for 6529811 to allow virtual calls to be sure they
1953 // get the chance to go from mono->bi->mega
1954 if (!keep_exact_action &&
1955 Deoptimization::trap_request_index(trap_request) < 0 &&
1956 too_many_recompiles(reason)) {
1957 // This BCI is causing too many recompilations.
1958 if (C->log() != NULL) {
1959 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
1960 Deoptimization::trap_reason_name(reason),
1961 Deoptimization::trap_action_name(action));
1962 }
1963 action = Deoptimization::Action_none;
1964 trap_request = Deoptimization::make_trap_request(reason, action);
1965 } else {
1966 C->set_trap_can_recompile(true);
1967 }
1968 break;
1969 case Deoptimization::Action_make_not_entrant:
1970 C->set_trap_can_recompile(true);
1971 break;
1972 case Deoptimization::Action_none:
1973 case Deoptimization::Action_make_not_compilable:
1974 break;
1975 default:
1976 #ifdef ASSERT
1977 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
1978 #endif
1979 break;
1980 }
1981
1982 if (TraceOptoParse) {
1983 char buf[100];
1984 tty->print_cr("Uncommon trap %s at bci:%d",
1985 Deoptimization::format_trap_request(buf, sizeof(buf),
1986 trap_request), bci());
1987 }
1988
1989 CompileLog* log = C->log();
1990 if (log != NULL) {
1991 int kid = (klass == NULL)? -1: log->identify(klass);
1992 log->begin_elem("uncommon_trap bci='%d'", bci());
1993 char buf[100];
1994 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1995 trap_request));
1996 if (kid >= 0) log->print(" klass='%d'", kid);
1997 if (comment != NULL) log->print(" comment='%s'", comment);
1998 log->end_elem();
1999 }
2000
2001 // Make sure any guarding test views this path as very unlikely
2002 Node *i0 = control()->in(0);
2003 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2004 IfNode *iff = i0->as_If();
2005 float f = iff->_prob; // Get prob
2006 if (control()->Opcode() == Op_IfTrue) {
2007 if (f > PROB_UNLIKELY_MAG(4))
2008 iff->_prob = PROB_MIN;
2009 } else {
2010 if (f < PROB_LIKELY_MAG(4))
2011 iff->_prob = PROB_MAX;
2012 }
2013 }
2014
2015 // Clear out dead values from the debug info.
2016 kill_dead_locals();
2017
2018 // Now insert the uncommon trap subroutine call
2019 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2020 const TypePtr* no_memory_effects = NULL;
2021 // Pass the index of the class to be loaded
2022 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2023 (must_throw ? RC_MUST_THROW : 0),
2024 OptoRuntime::uncommon_trap_Type(),
2025 call_addr, "uncommon_trap", no_memory_effects,
2026 intcon(trap_request));
2027 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2028 "must extract request correctly from the graph");
2029 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2030
2031 call->set_req(TypeFunc::ReturnAdr, returnadr());
2032 // The debug info is the only real input to this call.
2033
2034 // Halt-and-catch fire here. The above call should never return!
2035 HaltNode* halt = new HaltNode(control(), frameptr());
2036 _gvn.set_type_bottom(halt);
2037 root()->add_req(halt);
2038
2039 stop_and_kill_map();
2040 }
2041
2042
2043 //--------------------------just_allocated_object------------------------------
2044 // Report the object that was just allocated.
2045 // It must be the case that there are no intervening safepoints.
2046 // We use this to determine if an object is so "fresh" that
2047 // it does not require card marks.
2048 Node* GraphKit::just_allocated_object(Node* current_control) {
2049 if (C->recent_alloc_ctl() == current_control)
2050 return C->recent_alloc_obj();
2051 return NULL;
2052 }
2053
2054
2055 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2056 // (Note: TypeFunc::make has a cache that makes this fast.)
2057 const TypeFunc* tf = TypeFunc::make(dest_method);
2058 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2059 for (int j = 0; j < nargs; j++) {
2060 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2061 if( targ->basic_type() == T_DOUBLE ) {
2062 // If any parameters are doubles, they must be rounded before
2063 // the call, dstore_rounding does gvn.transform
2064 Node *arg = argument(j);
2065 arg = dstore_rounding(arg);
2066 set_argument(j, arg);
2067 }
2068 }
2069 }
2070
2071 /**
2072 * Record profiling data exact_kls for Node n with the type system so
2073 * that it can propagate it (speculation)
2074 *
2075 * @param n node that the type applies to
2076 * @param exact_kls type from profiling
2077 * @param maybe_null did profiling see null?
2078 *
2079 * @return node with improved type
2080 */
2081 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2082 const Type* current_type = _gvn.type(n);
2083 assert(UseTypeSpeculation, "type speculation must be on");
2084
2085 const TypePtr* speculative = current_type->speculative();
2086
2087 // Should the klass from the profile be recorded in the speculative type?
2088 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2089 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2090 const TypeOopPtr* xtype = tklass->as_instance_type();
2091 assert(xtype->klass_is_exact(), "Should be exact");
2092 // Any reason to believe n is not null (from this profiling or a previous one)?
2093 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2094 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2095 // record the new speculative type's depth
2096 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2097 speculative = speculative->with_inline_depth(jvms()->depth());
2098 } else if (current_type->would_improve_ptr(ptr_kind)) {
2099 // Profiling report that null was never seen so we can change the
2100 // speculative type to non null ptr.
2101 if (ptr_kind == ProfileAlwaysNull) {
2102 speculative = TypePtr::NULL_PTR;
2103 } else {
2104 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2105 const TypePtr* ptr = TypePtr::NOTNULL;
2106 if (speculative != NULL) {
2107 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2108 } else {
2109 speculative = ptr;
2110 }
2111 }
2112 }
2113
2114 if (speculative != current_type->speculative()) {
2115 // Build a type with a speculative type (what we think we know
2116 // about the type but will need a guard when we use it)
2117 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2118 // We're changing the type, we need a new CheckCast node to carry
2119 // the new type. The new type depends on the control: what
2120 // profiling tells us is only valid from here as far as we can
2121 // tell.
2122 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2123 cast = _gvn.transform(cast);
2124 replace_in_map(n, cast);
2125 n = cast;
2126 }
2127
2128 return n;
2129 }
2130
2131 /**
2132 * Record profiling data from receiver profiling at an invoke with the
2133 * type system so that it can propagate it (speculation)
2134 *
2135 * @param n receiver node
2136 *
2137 * @return node with improved type
2138 */
2139 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2140 if (!UseTypeSpeculation) {
2141 return n;
2142 }
2143 ciKlass* exact_kls = profile_has_unique_klass();
2144 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2145 if ((java_bc() == Bytecodes::_checkcast ||
2146 java_bc() == Bytecodes::_instanceof ||
2147 java_bc() == Bytecodes::_aastore) &&
2148 method()->method_data()->is_mature()) {
2149 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2150 if (data != NULL) {
2151 if (!data->as_BitData()->null_seen()) {
2152 ptr_kind = ProfileNeverNull;
2153 } else {
2154 assert(data->is_ReceiverTypeData(), "bad profile data type");
2155 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2156 uint i = 0;
2157 for (; i < call->row_limit(); i++) {
2158 ciKlass* receiver = call->receiver(i);
2159 if (receiver != NULL) {
2160 break;
2161 }
2162 }
2163 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2164 }
2165 }
2166 }
2167 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2168 }
2169
2170 /**
2171 * Record profiling data from argument profiling at an invoke with the
2172 * type system so that it can propagate it (speculation)
2173 *
2174 * @param dest_method target method for the call
2175 * @param bc what invoke bytecode is this?
2176 */
2177 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2178 if (!UseTypeSpeculation) {
2179 return;
2180 }
2181 const TypeFunc* tf = TypeFunc::make(dest_method);
2182 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2183 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2184 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2185 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2186 if (targ->isa_oopptr() && !targ->isa_valuetypeptr()) {
2187 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2188 ciKlass* better_type = NULL;
2189 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2190 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2191 }
2192 i++;
2193 }
2194 }
2195 }
2196
2197 /**
2198 * Record profiling data from parameter profiling at an invoke with
2199 * the type system so that it can propagate it (speculation)
2200 */
2201 void GraphKit::record_profiled_parameters_for_speculation() {
2202 if (!UseTypeSpeculation) {
2203 return;
2204 }
2205 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2206 if (_gvn.type(local(i))->isa_oopptr()) {
2207 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2208 ciKlass* better_type = NULL;
2209 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2210 record_profile_for_speculation(local(i), better_type, ptr_kind);
2211 }
2212 j++;
2213 }
2214 }
2215 }
2216
2217 /**
2218 * Record profiling data from return value profiling at an invoke with
2219 * the type system so that it can propagate it (speculation)
2220 */
2221 void GraphKit::record_profiled_return_for_speculation() {
2222 if (!UseTypeSpeculation) {
2223 return;
2224 }
2225 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2226 ciKlass* better_type = NULL;
2227 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2228 // If profiling reports a single type for the return value,
2229 // feed it to the type system so it can propagate it as a
2230 // speculative type
2231 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2232 }
2233 }
2234
2235 void GraphKit::round_double_result(ciMethod* dest_method) {
2236 // A non-strict method may return a double value which has an extended
2237 // exponent, but this must not be visible in a caller which is 'strict'
2238 // If a strict caller invokes a non-strict callee, round a double result
2239
2240 BasicType result_type = dest_method->return_type()->basic_type();
2241 assert( method() != NULL, "must have caller context");
2242 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2243 // Destination method's return value is on top of stack
2244 // dstore_rounding() does gvn.transform
2245 Node *result = pop_pair();
2246 result = dstore_rounding(result);
2247 push_pair(result);
2248 }
2249 }
2250
2251 // rounding for strict float precision conformance
2252 Node* GraphKit::precision_rounding(Node* n) {
2253 return UseStrictFP && _method->flags().is_strict()
2254 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2255 ? _gvn.transform( new RoundFloatNode(0, n) )
2256 : n;
2257 }
2258
2259 // rounding for strict double precision conformance
2260 Node* GraphKit::dprecision_rounding(Node *n) {
2261 return UseStrictFP && _method->flags().is_strict()
2262 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2263 ? _gvn.transform( new RoundDoubleNode(0, n) )
2264 : n;
2265 }
2266
2267 // rounding for non-strict double stores
2268 Node* GraphKit::dstore_rounding(Node* n) {
2269 return Matcher::strict_fp_requires_explicit_rounding
2270 && UseSSE <= 1
2271 ? _gvn.transform( new RoundDoubleNode(0, n) )
2272 : n;
2273 }
2274
2275 //=============================================================================
2276 // Generate a fast path/slow path idiom. Graph looks like:
2277 // [foo] indicates that 'foo' is a parameter
2278 //
2279 // [in] NULL
2280 // \ /
2281 // CmpP
2282 // Bool ne
2283 // If
2284 // / \
2285 // True False-<2>
2286 // / |
2287 // / cast_not_null
2288 // Load | | ^
2289 // [fast_test] | |
2290 // gvn to opt_test | |
2291 // / \ | <1>
2292 // True False |
2293 // | \\ |
2294 // [slow_call] \[fast_result]
2295 // Ctl Val \ \
2296 // | \ \
2297 // Catch <1> \ \
2298 // / \ ^ \ \
2299 // Ex No_Ex | \ \
2300 // | \ \ | \ <2> \
2301 // ... \ [slow_res] | | \ [null_result]
2302 // \ \--+--+--- | |
2303 // \ | / \ | /
2304 // --------Region Phi
2305 //
2306 //=============================================================================
2307 // Code is structured as a series of driver functions all called 'do_XXX' that
2308 // call a set of helper functions. Helper functions first, then drivers.
2309
2310 //------------------------------null_check_oop---------------------------------
2311 // Null check oop. Set null-path control into Region in slot 3.
2312 // Make a cast-not-nullness use the other not-null control. Return cast.
2313 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2314 bool never_see_null,
2315 bool safe_for_replace,
2316 bool speculative) {
2317 // Initial NULL check taken path
2318 (*null_control) = top();
2319 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2320
2321 // Generate uncommon_trap:
2322 if (never_see_null && (*null_control) != top()) {
2323 // If we see an unexpected null at a check-cast we record it and force a
2324 // recompile; the offending check-cast will be compiled to handle NULLs.
2325 // If we see more than one offending BCI, then all checkcasts in the
2326 // method will be compiled to handle NULLs.
2327 PreserveJVMState pjvms(this);
2328 set_control(*null_control);
2329 replace_in_map(value, null());
2330 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2331 uncommon_trap(reason,
2332 Deoptimization::Action_make_not_entrant);
2333 (*null_control) = top(); // NULL path is dead
2334 }
2335 if ((*null_control) == top() && safe_for_replace) {
2336 replace_in_map(value, cast);
2337 }
2338
2339 // Cast away null-ness on the result
2340 return cast;
2341 }
2342
2343 //------------------------------opt_iff----------------------------------------
2344 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2345 // Return slow-path control.
2346 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2347 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2348
2349 // Fast path taken; set region slot 2
2350 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2351 region->init_req(2,fast_taken); // Capture fast-control
2352
2353 // Fast path not-taken, i.e. slow path
2354 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2355 return slow_taken;
2356 }
2357
2358 //-----------------------------make_runtime_call-------------------------------
2359 Node* GraphKit::make_runtime_call(int flags,
2360 const TypeFunc* call_type, address call_addr,
2361 const char* call_name,
2362 const TypePtr* adr_type,
2363 // The following parms are all optional.
2364 // The first NULL ends the list.
2365 Node* parm0, Node* parm1,
2366 Node* parm2, Node* parm3,
2367 Node* parm4, Node* parm5,
2368 Node* parm6, Node* parm7) {
2369 // Slow-path call
2370 bool is_leaf = !(flags & RC_NO_LEAF);
2371 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2372 if (call_name == NULL) {
2373 assert(!is_leaf, "must supply name for leaf");
2374 call_name = OptoRuntime::stub_name(call_addr);
2375 }
2376 CallNode* call;
2377 if (!is_leaf) {
2378 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2379 bci(), adr_type);
2380 } else if (flags & RC_NO_FP) {
2381 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2382 } else {
2383 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2384 }
2385
2386 // The following is similar to set_edges_for_java_call,
2387 // except that the memory effects of the call are restricted to AliasIdxRaw.
2388
2389 // Slow path call has no side-effects, uses few values
2390 bool wide_in = !(flags & RC_NARROW_MEM);
2391 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2392
2393 Node* prev_mem = NULL;
2394 if (wide_in) {
2395 prev_mem = set_predefined_input_for_runtime_call(call);
2396 } else {
2397 assert(!wide_out, "narrow in => narrow out");
2398 Node* narrow_mem = memory(adr_type);
2399 prev_mem = reset_memory();
2400 map()->set_memory(narrow_mem);
2401 set_predefined_input_for_runtime_call(call);
2402 }
2403
2404 // Hook each parm in order. Stop looking at the first NULL.
2405 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2406 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2407 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2408 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2409 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2410 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2411 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2412 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2413 /* close each nested if ===> */ } } } } } } } }
2414 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2415
2416 if (!is_leaf) {
2417 // Non-leaves can block and take safepoints:
2418 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2419 }
2420 // Non-leaves can throw exceptions:
2421 if (has_io) {
2422 call->set_req(TypeFunc::I_O, i_o());
2423 }
2424
2425 if (flags & RC_UNCOMMON) {
2426 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2427 // (An "if" probability corresponds roughly to an unconditional count.
2428 // Sort of.)
2429 call->set_cnt(PROB_UNLIKELY_MAG(4));
2430 }
2431
2432 Node* c = _gvn.transform(call);
2433 assert(c == call, "cannot disappear");
2434
2435 if (wide_out) {
2436 // Slow path call has full side-effects.
2437 set_predefined_output_for_runtime_call(call);
2438 } else {
2439 // Slow path call has few side-effects, and/or sets few values.
2440 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2441 }
2442
2443 if (has_io) {
2444 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2445 }
2446 return call;
2447
2448 }
2449
2450 //------------------------------merge_memory-----------------------------------
2451 // Merge memory from one path into the current memory state.
2452 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2453 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2454 Node* old_slice = mms.force_memory();
2455 Node* new_slice = mms.memory2();
2456 if (old_slice != new_slice) {
2457 PhiNode* phi;
2458 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2459 if (mms.is_empty()) {
2460 // clone base memory Phi's inputs for this memory slice
2461 assert(old_slice == mms.base_memory(), "sanity");
2462 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2463 _gvn.set_type(phi, Type::MEMORY);
2464 for (uint i = 1; i < phi->req(); i++) {
2465 phi->init_req(i, old_slice->in(i));
2466 }
2467 } else {
2468 phi = old_slice->as_Phi(); // Phi was generated already
2469 }
2470 } else {
2471 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2472 _gvn.set_type(phi, Type::MEMORY);
2473 }
2474 phi->set_req(new_path, new_slice);
2475 mms.set_memory(phi);
2476 }
2477 }
2478 }
2479
2480 //------------------------------make_slow_call_ex------------------------------
2481 // Make the exception handler hookups for the slow call
2482 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2483 if (stopped()) return;
2484
2485 // Make a catch node with just two handlers: fall-through and catch-all
2486 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2487 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2488 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2489 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2490
2491 { PreserveJVMState pjvms(this);
2492 set_control(excp);
2493 set_i_o(i_o);
2494
2495 if (excp != top()) {
2496 if (deoptimize) {
2497 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2498 uncommon_trap(Deoptimization::Reason_unhandled,
2499 Deoptimization::Action_none);
2500 } else {
2501 // Create an exception state also.
2502 // Use an exact type if the caller has specified a specific exception.
2503 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2504 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2505 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2506 }
2507 }
2508 }
2509
2510 // Get the no-exception control from the CatchNode.
2511 set_control(norm);
2512 }
2513
2514 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2515 Node* cmp = NULL;
2516 switch(bt) {
2517 case T_INT: cmp = new CmpINode(in1, in2); break;
2518 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2519 default: fatal("unexpected comparison type %s", type2name(bt));
2520 }
2521 gvn->transform(cmp);
2522 Node* bol = gvn->transform(new BoolNode(cmp, test));
2523 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2524 gvn->transform(iff);
2525 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2526 return iff;
2527 }
2528
2529
2530 //-------------------------------gen_subtype_check-----------------------------
2531 // Generate a subtyping check. Takes as input the subtype and supertype.
2532 // Returns 2 values: sets the default control() to the true path and returns
2533 // the false path. Only reads invariant memory; sets no (visible) memory.
2534 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2535 // but that's not exposed to the optimizer. This call also doesn't take in an
2536 // Object; if you wish to check an Object you need to load the Object's class
2537 // prior to coming here.
2538 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2539 Compile* C = gvn->C;
2540
2541 if ((*ctrl)->is_top()) {
2542 return C->top();
2543 }
2544
2545 // Fast check for identical types, perhaps identical constants.
2546 // The types can even be identical non-constants, in cases
2547 // involving Array.newInstance, Object.clone, etc.
2548 if (subklass == superklass)
2549 return C->top(); // false path is dead; no test needed.
2550
2551 if (gvn->type(superklass)->singleton()) {
2552 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2553 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2554
2555 // In the common case of an exact superklass, try to fold up the
2556 // test before generating code. You may ask, why not just generate
2557 // the code and then let it fold up? The answer is that the generated
2558 // code will necessarily include null checks, which do not always
2559 // completely fold away. If they are also needless, then they turn
2560 // into a performance loss. Example:
2561 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2562 // Here, the type of 'fa' is often exact, so the store check
2563 // of fa[1]=x will fold up, without testing the nullness of x.
2564 switch (C->static_subtype_check(superk, subk)) {
2565 case Compile::SSC_always_false:
2566 {
2567 Node* always_fail = *ctrl;
2568 *ctrl = gvn->C->top();
2569 return always_fail;
2570 }
2571 case Compile::SSC_always_true:
2572 return C->top();
2573 case Compile::SSC_easy_test:
2574 {
2575 // Just do a direct pointer compare and be done.
2576 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2577 *ctrl = gvn->transform(new IfTrueNode(iff));
2578 return gvn->transform(new IfFalseNode(iff));
2579 }
2580 case Compile::SSC_full_test:
2581 break;
2582 default:
2583 ShouldNotReachHere();
2584 }
2585 }
2586
2587 // %%% Possible further optimization: Even if the superklass is not exact,
2588 // if the subklass is the unique subtype of the superklass, the check
2589 // will always succeed. We could leave a dependency behind to ensure this.
2590
2591 // First load the super-klass's check-offset
2592 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2593 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2594 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2595 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2596 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2597
2598 // Load from the sub-klass's super-class display list, or a 1-word cache of
2599 // the secondary superclass list, or a failing value with a sentinel offset
2600 // if the super-klass is an interface or exceptionally deep in the Java
2601 // hierarchy and we have to scan the secondary superclass list the hard way.
2602 // Worst-case type is a little odd: NULL is allowed as a result (usually
2603 // klass loads can never produce a NULL).
2604 Node *chk_off_X = chk_off;
2605 #ifdef _LP64
2606 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2607 #endif
2608 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2609 // For some types like interfaces the following loadKlass is from a 1-word
2610 // cache which is mutable so can't use immutable memory. Other
2611 // types load from the super-class display table which is immutable.
2612 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2613 Node *kmem = might_be_cache ? m : C->immutable_memory();
2614 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::BOTTOM));
2615
2616 // Compile speed common case: ARE a subtype and we canNOT fail
2617 if( superklass == nkls )
2618 return C->top(); // false path is dead; no test needed.
2619
2620 // See if we get an immediate positive hit. Happens roughly 83% of the
2621 // time. Test to see if the value loaded just previously from the subklass
2622 // is exactly the superklass.
2623 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2624 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2625 *ctrl = gvn->transform(new IfFalseNode(iff1));
2626
2627 // Compile speed common case: Check for being deterministic right now. If
2628 // chk_off is a constant and not equal to cacheoff then we are NOT a
2629 // subklass. In this case we need exactly the 1 test above and we can
2630 // return those results immediately.
2631 if (!might_be_cache) {
2632 Node* not_subtype_ctrl = *ctrl;
2633 *ctrl = iftrue1; // We need exactly the 1 test above
2634 return not_subtype_ctrl;
2635 }
2636
2637 // Gather the various success & failures here
2638 RegionNode *r_ok_subtype = new RegionNode(4);
2639 gvn->record_for_igvn(r_ok_subtype);
2640 RegionNode *r_not_subtype = new RegionNode(3);
2641 gvn->record_for_igvn(r_not_subtype);
2642
2643 r_ok_subtype->init_req(1, iftrue1);
2644
2645 // Check for immediate negative hit. Happens roughly 11% of the time (which
2646 // is roughly 63% of the remaining cases). Test to see if the loaded
2647 // check-offset points into the subklass display list or the 1-element
2648 // cache. If it points to the display (and NOT the cache) and the display
2649 // missed then it's not a subtype.
2650 Node *cacheoff = gvn->intcon(cacheoff_con);
2651 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2652 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2653 *ctrl = gvn->transform(new IfFalseNode(iff2));
2654
2655 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2656 // No performance impact (too rare) but allows sharing of secondary arrays
2657 // which has some footprint reduction.
2658 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2659 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2660 *ctrl = gvn->transform(new IfFalseNode(iff3));
2661
2662 // -- Roads not taken here: --
2663 // We could also have chosen to perform the self-check at the beginning
2664 // of this code sequence, as the assembler does. This would not pay off
2665 // the same way, since the optimizer, unlike the assembler, can perform
2666 // static type analysis to fold away many successful self-checks.
2667 // Non-foldable self checks work better here in second position, because
2668 // the initial primary superclass check subsumes a self-check for most
2669 // types. An exception would be a secondary type like array-of-interface,
2670 // which does not appear in its own primary supertype display.
2671 // Finally, we could have chosen to move the self-check into the
2672 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2673 // dependent manner. But it is worthwhile to have the check here,
2674 // where it can be perhaps be optimized. The cost in code space is
2675 // small (register compare, branch).
2676
2677 // Now do a linear scan of the secondary super-klass array. Again, no real
2678 // performance impact (too rare) but it's gotta be done.
2679 // Since the code is rarely used, there is no penalty for moving it
2680 // out of line, and it can only improve I-cache density.
2681 // The decision to inline or out-of-line this final check is platform
2682 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2683 Node* psc = gvn->transform(
2684 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2685
2686 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2687 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2688 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2689
2690 // Return false path; set default control to true path.
2691 *ctrl = gvn->transform(r_ok_subtype);
2692 return gvn->transform(r_not_subtype);
2693 }
2694
2695 // Profile-driven exact type check:
2696 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2697 float prob,
2698 Node* *casted_receiver) {
2699 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2700 Node* recv_klass = load_object_klass(receiver);
2701 Node* fail = type_check(recv_klass, tklass, prob);
2702 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2703 assert(recv_xtype->klass_is_exact(), "");
2704
2705 // Subsume downstream occurrences of receiver with a cast to
2706 // recv_xtype, since now we know what the type will be.
2707 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2708 (*casted_receiver) = _gvn.transform(cast);
2709 // (User must make the replace_in_map call.)
2710
2711 return fail;
2712 }
2713
2714 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
2715 float prob) {
2716 //const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2717 Node* want_klass = makecon(tklass);
2718 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass));
2719 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2720 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2721 set_control( _gvn.transform( new IfTrueNode (iff)));
2722 Node* fail = _gvn.transform( new IfFalseNode(iff));
2723 return fail;
2724 }
2725
2726
2727 //------------------------------seems_never_null-------------------------------
2728 // Use null_seen information if it is available from the profile.
2729 // If we see an unexpected null at a type check we record it and force a
2730 // recompile; the offending check will be recompiled to handle NULLs.
2731 // If we see several offending BCIs, then all checks in the
2732 // method will be recompiled.
2733 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2734 speculating = !_gvn.type(obj)->speculative_maybe_null();
2735 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2736 if (UncommonNullCast // Cutout for this technique
2737 && obj != null() // And not the -Xcomp stupid case?
2738 && !too_many_traps(reason)
2739 ) {
2740 if (speculating) {
2741 return true;
2742 }
2743 if (data == NULL)
2744 // Edge case: no mature data. Be optimistic here.
2745 return true;
2746 // If the profile has not seen a null, assume it won't happen.
2747 assert(java_bc() == Bytecodes::_checkcast ||
2748 java_bc() == Bytecodes::_instanceof ||
2749 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2750 return !data->as_BitData()->null_seen();
2751 }
2752 speculating = false;
2753 return false;
2754 }
2755
2756 //------------------------maybe_cast_profiled_receiver-------------------------
2757 // If the profile has seen exactly one type, narrow to exactly that type.
2758 // Subsequent type checks will always fold up.
2759 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2760 ciKlass* require_klass,
2761 ciKlass* spec_klass,
2762 bool safe_for_replace) {
2763 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2764
2765 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2766
2767 // Make sure we haven't already deoptimized from this tactic.
2768 if (too_many_traps(reason) || too_many_recompiles(reason))
2769 return NULL;
2770
2771 // (No, this isn't a call, but it's enough like a virtual call
2772 // to use the same ciMethod accessor to get the profile info...)
2773 // If we have a speculative type use it instead of profiling (which
2774 // may not help us)
2775 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2776 if (exact_kls != NULL) {// no cast failures here
2777 if (require_klass == NULL ||
2778 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2779 // If we narrow the type to match what the type profile sees or
2780 // the speculative type, we can then remove the rest of the
2781 // cast.
2782 // This is a win, even if the exact_kls is very specific,
2783 // because downstream operations, such as method calls,
2784 // will often benefit from the sharper type.
2785 Node* exact_obj = not_null_obj; // will get updated in place...
2786 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2787 &exact_obj);
2788 { PreserveJVMState pjvms(this);
2789 set_control(slow_ctl);
2790 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2791 }
2792 if (safe_for_replace) {
2793 replace_in_map(not_null_obj, exact_obj);
2794 }
2795 return exact_obj;
2796 }
2797 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2798 }
2799
2800 return NULL;
2801 }
2802
2803 /**
2804 * Cast obj to type and emit guard unless we had too many traps here
2805 * already
2806 *
2807 * @param obj node being casted
2808 * @param type type to cast the node to
2809 * @param not_null true if we know node cannot be null
2810 */
2811 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2812 ciKlass* type,
2813 bool not_null) {
2814 if (stopped()) {
2815 return obj;
2816 }
2817
2818 // type == NULL if profiling tells us this object is always null
2819 if (type != NULL) {
2820 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2821 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2822
2823 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2824 !too_many_traps(class_reason) &&
2825 !too_many_recompiles(class_reason)) {
2826 Node* not_null_obj = NULL;
2827 // not_null is true if we know the object is not null and
2828 // there's no need for a null check
2829 if (!not_null) {
2830 Node* null_ctl = top();
2831 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2832 assert(null_ctl->is_top(), "no null control here");
2833 } else {
2834 not_null_obj = obj;
2835 }
2836
2837 Node* exact_obj = not_null_obj;
2838 ciKlass* exact_kls = type;
2839 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2840 &exact_obj);
2841 {
2842 PreserveJVMState pjvms(this);
2843 set_control(slow_ctl);
2844 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2845 }
2846 replace_in_map(not_null_obj, exact_obj);
2847 obj = exact_obj;
2848 }
2849 } else {
2850 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2851 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2852 Node* exact_obj = null_assert(obj);
2853 replace_in_map(obj, exact_obj);
2854 obj = exact_obj;
2855 }
2856 }
2857 return obj;
2858 }
2859
2860 //-------------------------------gen_instanceof--------------------------------
2861 // Generate an instance-of idiom. Used by both the instance-of bytecode
2862 // and the reflective instance-of call.
2863 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2864 kill_dead_locals(); // Benefit all the uncommon traps
2865 assert( !stopped(), "dead parse path should be checked in callers" );
2866 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2867 "must check for not-null not-dead klass in callers");
2868
2869 // Make the merge point
2870 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2871 RegionNode* region = new RegionNode(PATH_LIMIT);
2872 Node* phi = new PhiNode(region, TypeInt::BOOL);
2873 C->set_has_split_ifs(true); // Has chance for split-if optimization
2874
2875 ciProfileData* data = NULL;
2876 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2877 data = method()->method_data()->bci_to_data(bci());
2878 }
2879 bool speculative_not_null = false;
2880 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2881 && seems_never_null(obj, data, speculative_not_null));
2882
2883 // Null check; get casted pointer; set region slot 3
2884 Node* null_ctl = top();
2885 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2886
2887 // If not_null_obj is dead, only null-path is taken
2888 if (stopped()) { // Doing instance-of on a NULL?
2889 set_control(null_ctl);
2890 return intcon(0);
2891 }
2892 region->init_req(_null_path, null_ctl);
2893 phi ->init_req(_null_path, intcon(0)); // Set null path value
2894 if (null_ctl == top()) {
2895 // Do this eagerly, so that pattern matches like is_diamond_phi
2896 // will work even during parsing.
2897 assert(_null_path == PATH_LIMIT-1, "delete last");
2898 region->del_req(_null_path);
2899 phi ->del_req(_null_path);
2900 }
2901
2902 // Do we know the type check always succeed?
2903 bool known_statically = false;
2904 if (_gvn.type(superklass)->singleton()) {
2905 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2906 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2907 if (subk != NULL && subk->is_loaded()) {
2908 int static_res = C->static_subtype_check(superk, subk);
2909 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2910 }
2911 }
2912
2913 if (!known_statically) {
2914 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2915 // We may not have profiling here or it may not help us. If we
2916 // have a speculative type use it to perform an exact cast.
2917 ciKlass* spec_obj_type = obj_type->speculative_type();
2918 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2919 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2920 if (stopped()) { // Profile disagrees with this path.
2921 set_control(null_ctl); // Null is the only remaining possibility.
2922 return intcon(0);
2923 }
2924 if (cast_obj != NULL) {
2925 not_null_obj = cast_obj;
2926 }
2927 }
2928 }
2929
2930 // Load the object's klass
2931 Node* obj_klass = load_object_klass(not_null_obj);
2932
2933 // Generate the subtype check
2934 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2935
2936 // Plug in the success path to the general merge in slot 1.
2937 region->init_req(_obj_path, control());
2938 phi ->init_req(_obj_path, intcon(1));
2939
2940 // Plug in the failing path to the general merge in slot 2.
2941 region->init_req(_fail_path, not_subtype_ctrl);
2942 phi ->init_req(_fail_path, intcon(0));
2943
2944 // Return final merged results
2945 set_control( _gvn.transform(region) );
2946 record_for_igvn(region);
2947
2948 // If we know the type check always succeeds then we don't use the
2949 // profiling data at this bytecode. Don't lose it, feed it to the
2950 // type system as a speculative type.
2951 if (safe_for_replace) {
2952 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
2953 replace_in_map(obj, casted_obj);
2954 }
2955
2956 return _gvn.transform(phi);
2957 }
2958
2959 //-------------------------------gen_checkcast---------------------------------
2960 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2961 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2962 // uncommon-trap paths work. Adjust stack after this call.
2963 // If failure_control is supplied and not null, it is filled in with
2964 // the control edge for the cast failure. Otherwise, an appropriate
2965 // uncommon trap or exception is thrown.
2966 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2967 Node* *failure_control) {
2968 kill_dead_locals(); // Benefit all the uncommon traps
2969 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2970 assert(tk->is_loaded(), "must be loaded");
2971 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2972
2973 // Fast cutout: Check the case that the cast is vacuously true.
2974 // This detects the common cases where the test will short-circuit
2975 // away completely. We do this before we perform the null check,
2976 // because if the test is going to turn into zero code, we don't
2977 // want a residual null check left around. (Causes a slowdown,
2978 // for example, in some objArray manipulations, such as a[i]=a[j].)
2979 if (tk->singleton()) {
2980 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2981 if (objtp != NULL && objtp->klass() != NULL) {
2982 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
2983 case Compile::SSC_always_true:
2984 // If we know the type check always succeed then we don't use
2985 // the profiling data at this bytecode. Don't lose it, feed it
2986 // to the type system as a speculative type.
2987 return record_profiled_receiver_for_speculation(obj);
2988 case Compile::SSC_always_false:
2989 // It needs a null check because a null will *pass* the cast check.
2990 // A non-null value will always produce an exception.
2991 return null_assert(obj);
2992 }
2993 }
2994 }
2995
2996 ciProfileData* data = NULL;
2997 bool safe_for_replace = false;
2998 if (failure_control == NULL) { // use MDO in regular case only
2999 assert(java_bc() == Bytecodes::_aastore ||
3000 java_bc() == Bytecodes::_checkcast,
3001 "interpreter profiles type checks only for these BCs");
3002 data = method()->method_data()->bci_to_data(bci());
3003 safe_for_replace = true;
3004 }
3005
3006 // Make the merge point
3007 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3008 RegionNode* region = new RegionNode(PATH_LIMIT);
3009 Node* phi = new PhiNode(region, toop);
3010 C->set_has_split_ifs(true); // Has chance for split-if optimization
3011
3012 // Use null-cast information if it is available
3013 bool speculative_not_null = false;
3014 bool never_see_null = ((failure_control == NULL) // regular case only
3015 && seems_never_null(obj, data, speculative_not_null));
3016
3017 // Null check; get casted pointer; set region slot 3
3018 Node* null_ctl = top();
3019 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3020
3021 // If not_null_obj is dead, only null-path is taken
3022 if (stopped()) { // Doing instance-of on a NULL?
3023 set_control(null_ctl);
3024 return null();
3025 }
3026 region->init_req(_null_path, null_ctl);
3027 phi ->init_req(_null_path, null()); // Set null path value
3028 if (null_ctl == top()) {
3029 // Do this eagerly, so that pattern matches like is_diamond_phi
3030 // will work even during parsing.
3031 assert(_null_path == PATH_LIMIT-1, "delete last");
3032 region->del_req(_null_path);
3033 phi ->del_req(_null_path);
3034 }
3035
3036 Node* cast_obj = NULL;
3037 if (tk->klass_is_exact()) {
3038 // The following optimization tries to statically cast the speculative type of the object
3039 // (for example obtained during profiling) to the type of the superklass and then do a
3040 // dynamic check that the type of the object is what we expect. To work correctly
3041 // for checkcast and aastore the type of superklass should be exact.
3042 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3043 // We may not have profiling here or it may not help us. If we have
3044 // a speculative type use it to perform an exact cast.
3045 ciKlass* spec_obj_type = obj_type->speculative_type();
3046 if (spec_obj_type != NULL || data != NULL) {
3047 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3048 if (cast_obj != NULL) {
3049 if (failure_control != NULL) // failure is now impossible
3050 (*failure_control) = top();
3051 // adjust the type of the phi to the exact klass:
3052 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3053 }
3054 }
3055 }
3056
3057 if (cast_obj == NULL) {
3058 // Load the object's klass
3059 Node* obj_klass = load_object_klass(not_null_obj);
3060
3061 // Generate the subtype check
3062 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3063
3064 // Plug in success path into the merge
3065 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3066 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3067 if (failure_control == NULL) {
3068 if (not_subtype_ctrl != top()) { // If failure is possible
3069 PreserveJVMState pjvms(this);
3070 set_control(not_subtype_ctrl);
3071 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3072 }
3073 } else {
3074 (*failure_control) = not_subtype_ctrl;
3075 }
3076 }
3077
3078 region->init_req(_obj_path, control());
3079 phi ->init_req(_obj_path, cast_obj);
3080
3081 // A merge of NULL or Casted-NotNull obj
3082 Node* res = _gvn.transform(phi);
3083
3084 // Note I do NOT always 'replace_in_map(obj,result)' here.
3085 // if( tk->klass()->can_be_primary_super() )
3086 // This means that if I successfully store an Object into an array-of-String
3087 // I 'forget' that the Object is really now known to be a String. I have to
3088 // do this because we don't have true union types for interfaces - if I store
3089 // a Baz into an array-of-Interface and then tell the optimizer it's an
3090 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3091 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3092 // replace_in_map( obj, res );
3093
3094 // Return final merged results
3095 set_control( _gvn.transform(region) );
3096 record_for_igvn(region);
3097
3098 return record_profiled_receiver_for_speculation(res);
3099 }
3100
3101 //------------------------------next_monitor-----------------------------------
3102 // What number should be given to the next monitor?
3103 int GraphKit::next_monitor() {
3104 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3105 int next = current + C->sync_stack_slots();
3106 // Keep the toplevel high water mark current:
3107 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3108 return current;
3109 }
3110
3111 //------------------------------insert_mem_bar---------------------------------
3112 // Memory barrier to avoid floating things around
3113 // The membar serves as a pinch point between both control and all memory slices.
3114 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3115 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3116 mb->init_req(TypeFunc::Control, control());
3117 mb->init_req(TypeFunc::Memory, reset_memory());
3118 Node* membar = _gvn.transform(mb);
3119 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3120 set_all_memory_call(membar);
3121 return membar;
3122 }
3123
3124 //-------------------------insert_mem_bar_volatile----------------------------
3125 // Memory barrier to avoid floating things around
3126 // The membar serves as a pinch point between both control and memory(alias_idx).
3127 // If you want to make a pinch point on all memory slices, do not use this
3128 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3129 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3130 // When Parse::do_put_xxx updates a volatile field, it appends a series
3131 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3132 // The first membar is on the same memory slice as the field store opcode.
3133 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3134 // All the other membars (for other volatile slices, including AliasIdxBot,
3135 // which stands for all unknown volatile slices) are control-dependent
3136 // on the first membar. This prevents later volatile loads or stores
3137 // from sliding up past the just-emitted store.
3138
3139 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3140 mb->set_req(TypeFunc::Control,control());
3141 if (alias_idx == Compile::AliasIdxBot) {
3142 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3143 } else {
3144 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3145 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3146 }
3147 Node* membar = _gvn.transform(mb);
3148 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3149 if (alias_idx == Compile::AliasIdxBot) {
3150 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3151 } else {
3152 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3153 }
3154 return membar;
3155 }
3156
3157 //------------------------------shared_lock------------------------------------
3158 // Emit locking code.
3159 FastLockNode* GraphKit::shared_lock(Node* obj) {
3160 // bci is either a monitorenter bc or InvocationEntryBci
3161 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3162 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3163
3164 if( !GenerateSynchronizationCode )
3165 return NULL; // Not locking things?
3166 if (stopped()) // Dead monitor?
3167 return NULL;
3168
3169 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3170
3171 // Box the stack location
3172 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3173 Node* mem = reset_memory();
3174
3175 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3176 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3177 // Create the counters for this fast lock.
3178 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3179 }
3180
3181 // Create the rtm counters for this fast lock if needed.
3182 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3183
3184 // Add monitor to debug info for the slow path. If we block inside the
3185 // slow path and de-opt, we need the monitor hanging around
3186 map()->push_monitor( flock );
3187
3188 const TypeFunc *tf = LockNode::lock_type();
3189 LockNode *lock = new LockNode(C, tf);
3190
3191 lock->init_req( TypeFunc::Control, control() );
3192 lock->init_req( TypeFunc::Memory , mem );
3193 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3194 lock->init_req( TypeFunc::FramePtr, frameptr() );
3195 lock->init_req( TypeFunc::ReturnAdr, top() );
3196
3197 lock->init_req(TypeFunc::Parms + 0, obj);
3198 lock->init_req(TypeFunc::Parms + 1, box);
3199 lock->init_req(TypeFunc::Parms + 2, flock);
3200 add_safepoint_edges(lock);
3201
3202 lock = _gvn.transform( lock )->as_Lock();
3203
3204 // lock has no side-effects, sets few values
3205 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3206
3207 insert_mem_bar(Op_MemBarAcquireLock);
3208
3209 // Add this to the worklist so that the lock can be eliminated
3210 record_for_igvn(lock);
3211
3212 #ifndef PRODUCT
3213 if (PrintLockStatistics) {
3214 // Update the counter for this lock. Don't bother using an atomic
3215 // operation since we don't require absolute accuracy.
3216 lock->create_lock_counter(map()->jvms());
3217 increment_counter(lock->counter()->addr());
3218 }
3219 #endif
3220
3221 return flock;
3222 }
3223
3224
3225 //------------------------------shared_unlock----------------------------------
3226 // Emit unlocking code.
3227 void GraphKit::shared_unlock(Node* box, Node* obj) {
3228 // bci is either a monitorenter bc or InvocationEntryBci
3229 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3230 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3231
3232 if( !GenerateSynchronizationCode )
3233 return;
3234 if (stopped()) { // Dead monitor?
3235 map()->pop_monitor(); // Kill monitor from debug info
3236 return;
3237 }
3238
3239 // Memory barrier to avoid floating things down past the locked region
3240 insert_mem_bar(Op_MemBarReleaseLock);
3241
3242 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3243 UnlockNode *unlock = new UnlockNode(C, tf);
3244 #ifdef ASSERT
3245 unlock->set_dbg_jvms(sync_jvms());
3246 #endif
3247 uint raw_idx = Compile::AliasIdxRaw;
3248 unlock->init_req( TypeFunc::Control, control() );
3249 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3250 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3251 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3252 unlock->init_req( TypeFunc::ReturnAdr, top() );
3253
3254 unlock->init_req(TypeFunc::Parms + 0, obj);
3255 unlock->init_req(TypeFunc::Parms + 1, box);
3256 unlock = _gvn.transform(unlock)->as_Unlock();
3257
3258 Node* mem = reset_memory();
3259
3260 // unlock has no side-effects, sets few values
3261 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3262
3263 // Kill monitor from debug info
3264 map()->pop_monitor( );
3265 }
3266
3267 //-------------------------------get_layout_helper-----------------------------
3268 // If the given klass is a constant or known to be an array,
3269 // fetch the constant layout helper value into constant_value
3270 // and return (Node*)NULL. Otherwise, load the non-constant
3271 // layout helper value, and return the node which represents it.
3272 // This two-faced routine is useful because allocation sites
3273 // almost always feature constant types.
3274 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3275 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3276 if (!StressReflectiveCode && inst_klass != NULL) {
3277 ciKlass* klass = inst_klass->klass();
3278 assert(klass != NULL, "klass should not be NULL");
3279 bool xklass = inst_klass->klass_is_exact();
3280 if (xklass || klass->is_array_klass()) {
3281 jint lhelper = klass->layout_helper();
3282 if (lhelper != Klass::_lh_neutral_value) {
3283 constant_value = lhelper;
3284 return (Node*) NULL;
3285 }
3286 }
3287 }
3288 constant_value = Klass::_lh_neutral_value; // put in a known value
3289 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3290 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3291 }
3292
3293 // We just put in an allocate/initialize with a big raw-memory effect.
3294 // Hook selected additional alias categories on the initialization.
3295 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3296 MergeMemNode* init_in_merge,
3297 Node* init_out_raw) {
3298 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3299 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3300
3301 Node* prevmem = kit.memory(alias_idx);
3302 init_in_merge->set_memory_at(alias_idx, prevmem);
3303 kit.set_memory(init_out_raw, alias_idx);
3304 }
3305
3306 //---------------------------set_output_for_allocation-------------------------
3307 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3308 const TypeOopPtr* oop_type,
3309 bool deoptimize_on_exception) {
3310 int rawidx = Compile::AliasIdxRaw;
3311 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3312 add_safepoint_edges(alloc);
3313 Node* allocx = _gvn.transform(alloc);
3314 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3315 // create memory projection for i_o
3316 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3317 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3318
3319 // create a memory projection as for the normal control path
3320 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3321 set_memory(malloc, rawidx);
3322
3323 // a normal slow-call doesn't change i_o, but an allocation does
3324 // we create a separate i_o projection for the normal control path
3325 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3326 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3327
3328 // put in an initialization barrier
3329 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3330 rawoop)->as_Initialize();
3331 assert(alloc->initialization() == init, "2-way macro link must work");
3332 assert(init ->allocation() == alloc, "2-way macro link must work");
3333 {
3334 // Extract memory strands which may participate in the new object's
3335 // initialization, and source them from the new InitializeNode.
3336 // This will allow us to observe initializations when they occur,
3337 // and link them properly (as a group) to the InitializeNode.
3338 assert(init->in(InitializeNode::Memory) == malloc, "");
3339 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3340 init->set_req(InitializeNode::Memory, minit_in);
3341 record_for_igvn(minit_in); // fold it up later, if possible
3342 Node* minit_out = memory(rawidx);
3343 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3344 if (oop_type->isa_aryptr()) {
3345 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3346 int elemidx = C->get_alias_index(telemref);
3347 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3348 } else if (oop_type->isa_instptr() || oop_type->isa_valuetypeptr()) {
3349 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3350 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3351 ciField* field = ik->nonstatic_field_at(i);
3352 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3353 continue; // do not bother to track really large numbers of fields
3354 // Find (or create) the alias category for this field:
3355 int fieldidx = C->alias_type(field)->index();
3356 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3357 }
3358 }
3359 }
3360
3361 // Cast raw oop to the real thing...
3362 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3363 javaoop = _gvn.transform(javaoop);
3364 C->set_recent_alloc(control(), javaoop);
3365 assert(just_allocated_object(control()) == javaoop, "just allocated");
3366
3367 #ifdef ASSERT
3368 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3369 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3370 "Ideal_allocation works");
3371 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3372 "Ideal_allocation works");
3373 if (alloc->is_AllocateArray()) {
3374 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3375 "Ideal_allocation works");
3376 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3377 "Ideal_allocation works");
3378 } else {
3379 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3380 }
3381 }
3382 #endif //ASSERT
3383
3384 return javaoop;
3385 }
3386
3387 //---------------------------new_instance--------------------------------------
3388 // This routine takes a klass_node which may be constant (for a static type)
3389 // or may be non-constant (for reflective code). It will work equally well
3390 // for either, and the graph will fold nicely if the optimizer later reduces
3391 // the type to a constant.
3392 // The optional arguments are for specialized use by intrinsics:
3393 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3394 // - If 'return_size_val', report the the total object size to the caller.
3395 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3396 Node* GraphKit::new_instance(Node* klass_node,
3397 Node* extra_slow_test,
3398 Node* *return_size_val,
3399 bool deoptimize_on_exception,
3400 ValueTypeNode* value_node) {
3401 // Compute size in doublewords
3402 // The size is always an integral number of doublewords, represented
3403 // as a positive bytewise size stored in the klass's layout_helper.
3404 // The layout_helper also encodes (in a low bit) the need for a slow path.
3405 jint layout_con = Klass::_lh_neutral_value;
3406 Node* layout_val = get_layout_helper(klass_node, layout_con);
3407 bool layout_is_con = (layout_val == NULL);
3408
3409 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3410 // Generate the initial go-slow test. It's either ALWAYS (return a
3411 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3412 // case) a computed value derived from the layout_helper.
3413 Node* initial_slow_test = NULL;
3414 if (layout_is_con) {
3415 assert(!StressReflectiveCode, "stress mode does not use these paths");
3416 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3417 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3418 } else { // reflective case
3419 // This reflective path is used by Unsafe.allocateInstance.
3420 // (It may be stress-tested by specifying StressReflectiveCode.)
3421 // Basically, we want to get into the VM is there's an illegal argument.
3422 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3423 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3424 if (extra_slow_test != intcon(0)) {
3425 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3426 }
3427 // (Macro-expander will further convert this to a Bool, if necessary.)
3428 }
3429
3430 // Find the size in bytes. This is easy; it's the layout_helper.
3431 // The size value must be valid even if the slow path is taken.
3432 Node* size = NULL;
3433 if (layout_is_con) {
3434 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3435 } else { // reflective case
3436 // This reflective path is used by clone and Unsafe.allocateInstance.
3437 size = ConvI2X(layout_val);
3438
3439 // Clear the low bits to extract layout_helper_size_in_bytes:
3440 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3441 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3442 size = _gvn.transform( new AndXNode(size, mask) );
3443 }
3444 if (return_size_val != NULL) {
3445 (*return_size_val) = size;
3446 }
3447
3448 // This is a precise notnull oop of the klass.
3449 // (Actually, it need not be precise if this is a reflective allocation.)
3450 // It's what we cast the result to.
3451 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3452 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3453 const TypeOopPtr* oop_type = tklass->as_instance_type();
3454
3455 // Now generate allocation code
3456
3457 // The entire memory state is needed for slow path of the allocation
3458 // since GC and deoptimization can happen.
3459 Node *mem = reset_memory();
3460 set_all_memory(mem); // Create new memory state
3461
3462 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3463 control(), mem, i_o(),
3464 size, klass_node,
3465 initial_slow_test, value_node);
3466
3467 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3468 }
3469
3470 //-------------------------------new_array-------------------------------------
3471 // helper for newarray and anewarray
3472 // The 'length' parameter is (obviously) the length of the array.
3473 // See comments on new_instance for the meaning of the other arguments.
3474 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3475 Node* length, // number of array elements
3476 int nargs, // number of arguments to push back for uncommon trap
3477 Node* *return_size_val,
3478 bool deoptimize_on_exception) {
3479 jint layout_con = Klass::_lh_neutral_value;
3480 Node* layout_val = get_layout_helper(klass_node, layout_con);
3481 bool layout_is_con = (layout_val == NULL);
3482
3483 if (!layout_is_con && !StressReflectiveCode &&
3484 !too_many_traps(Deoptimization::Reason_class_check)) {
3485 // This is a reflective array creation site.
3486 // Optimistically assume that it is a subtype of Object[],
3487 // so that we can fold up all the address arithmetic.
3488 layout_con = Klass::array_layout_helper(T_OBJECT);
3489 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3490 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3491 { BuildCutout unless(this, bol_lh, PROB_MAX);
3492 inc_sp(nargs);
3493 uncommon_trap(Deoptimization::Reason_class_check,
3494 Deoptimization::Action_maybe_recompile);
3495 }
3496 layout_val = NULL;
3497 layout_is_con = true;
3498 }
3499
3500 // Generate the initial go-slow test. Make sure we do not overflow
3501 // if length is huge (near 2Gig) or negative! We do not need
3502 // exact double-words here, just a close approximation of needed
3503 // double-words. We can't add any offset or rounding bits, lest we
3504 // take a size -1 of bytes and make it positive. Use an unsigned
3505 // compare, so negative sizes look hugely positive.
3506 int fast_size_limit = FastAllocateSizeLimit;
3507 if (layout_is_con) {
3508 assert(!StressReflectiveCode, "stress mode does not use these paths");
3509 // Increase the size limit if we have exact knowledge of array type.
3510 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3511 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3512 }
3513
3514 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3515 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3516
3517 // --- Size Computation ---
3518 // array_size = round_to_heap(array_header + (length << elem_shift));
3519 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3520 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3521 // The rounding mask is strength-reduced, if possible.
3522 int round_mask = MinObjAlignmentInBytes - 1;
3523 Node* header_size = NULL;
3524 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3525 // (T_BYTE has the weakest alignment and size restrictions...)
3526 if (layout_is_con) {
3527 int hsize = Klass::layout_helper_header_size(layout_con);
3528 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3529 BasicType etype = Klass::layout_helper_element_type(layout_con);
3530 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con);
3531 if ((round_mask & ~right_n_bits(eshift)) == 0)
3532 round_mask = 0; // strength-reduce it if it goes away completely
3533 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3534 assert(header_size_min <= hsize, "generic minimum is smallest");
3535 header_size_min = hsize;
3536 header_size = intcon(hsize + round_mask);
3537 } else {
3538 Node* hss = intcon(Klass::_lh_header_size_shift);
3539 Node* hsm = intcon(Klass::_lh_header_size_mask);
3540 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3541 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3542 Node* mask = intcon(round_mask);
3543 header_size = _gvn.transform( new AddINode(hsize, mask) );
3544 }
3545
3546 Node* elem_shift = NULL;
3547 if (layout_is_con) {
3548 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3549 if (eshift != 0)
3550 elem_shift = intcon(eshift);
3551 } else {
3552 // There is no need to mask or shift this value.
3553 // The semantics of LShiftINode include an implicit mask to 0x1F.
3554 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3555 elem_shift = layout_val;
3556 }
3557
3558 // Transition to native address size for all offset calculations:
3559 Node* lengthx = ConvI2X(length);
3560 Node* headerx = ConvI2X(header_size);
3561 #ifdef _LP64
3562 { const TypeInt* tilen = _gvn.find_int_type(length);
3563 if (tilen != NULL && tilen->_lo < 0) {
3564 // Add a manual constraint to a positive range. Cf. array_element_address.
3565 jint size_max = fast_size_limit;
3566 if (size_max > tilen->_hi) size_max = tilen->_hi;
3567 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3568
3569 // Only do a narrow I2L conversion if the range check passed.
3570 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3571 _gvn.transform(iff);
3572 RegionNode* region = new RegionNode(3);
3573 _gvn.set_type(region, Type::CONTROL);
3574 lengthx = new PhiNode(region, TypeLong::LONG);
3575 _gvn.set_type(lengthx, TypeLong::LONG);
3576
3577 // Range check passed. Use ConvI2L node with narrow type.
3578 Node* passed = IfFalse(iff);
3579 region->init_req(1, passed);
3580 // Make I2L conversion control dependent to prevent it from
3581 // floating above the range check during loop optimizations.
3582 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3583
3584 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3585 region->init_req(2, IfTrue(iff));
3586 lengthx->init_req(2, ConvI2X(length));
3587
3588 set_control(region);
3589 record_for_igvn(region);
3590 record_for_igvn(lengthx);
3591 }
3592 }
3593 #endif
3594
3595 // Combine header size (plus rounding) and body size. Then round down.
3596 // This computation cannot overflow, because it is used only in two
3597 // places, one where the length is sharply limited, and the other
3598 // after a successful allocation.
3599 Node* abody = lengthx;
3600 if (elem_shift != NULL)
3601 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3602 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3603 if (round_mask != 0) {
3604 Node* mask = MakeConX(~round_mask);
3605 size = _gvn.transform( new AndXNode(size, mask) );
3606 }
3607 // else if round_mask == 0, the size computation is self-rounding
3608
3609 if (return_size_val != NULL) {
3610 // This is the size
3611 (*return_size_val) = size;
3612 }
3613
3614 // Now generate allocation code
3615
3616 // The entire memory state is needed for slow path of the allocation
3617 // since GC and deoptimization can happen.
3618 Node *mem = reset_memory();
3619 set_all_memory(mem); // Create new memory state
3620
3621 if (initial_slow_test->is_Bool()) {
3622 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3623 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3624 }
3625
3626 // Create the AllocateArrayNode and its result projections
3627 AllocateArrayNode* alloc
3628 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3629 control(), mem, i_o(),
3630 size, klass_node,
3631 initial_slow_test,
3632 length);
3633
3634 // Cast to correct type. Note that the klass_node may be constant or not,
3635 // and in the latter case the actual array type will be inexact also.
3636 // (This happens via a non-constant argument to inline_native_newArray.)
3637 // In any case, the value of klass_node provides the desired array type.
3638 const TypeInt* length_type = _gvn.find_int_type(length);
3639 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3640 if (ary_type->isa_aryptr() && length_type != NULL) {
3641 // Try to get a better type than POS for the size
3642 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3643 }
3644
3645 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3646
3647 // Cast length on remaining path to be as narrow as possible
3648 if (map()->find_edge(length) >= 0) {
3649 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3650 if (ccast != length) {
3651 _gvn.set_type_bottom(ccast);
3652 record_for_igvn(ccast);
3653 replace_in_map(length, ccast);
3654 }
3655 }
3656
3657 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
3658 ciKlass* elem_klass = ary_ptr != NULL ? ary_ptr->klass()->as_array_klass()->element_klass() : NULL;
3659 if (elem_klass != NULL && elem_klass->is_valuetype()) {
3660 ciValueKlass* vk = elem_klass->as_value_klass();
3661 if (!vk->flatten_array()) {
3662 // Non-flattened value type arrays need to be initialized with default value type oops
3663 initialize_value_type_array(javaoop, length, elem_klass->as_value_klass(), nargs);
3664 InitializeNode* init = alloc->initialization();
3665 init->set_complete_with_arraycopy();
3666 }
3667 }
3668
3669 return javaoop;
3670 }
3671
3672 void GraphKit::initialize_value_type_array(Node* array, Node* length, ciValueKlass* vk, int nargs) {
3673 // Check for zero length
3674 Node* null_ctl = top();
3675 null_check_common(length, T_INT, false, &null_ctl, false);
3676 if (stopped()) {
3677 set_control(null_ctl); // Always zero
3678 return;
3679 }
3680
3681 // Prepare for merging control and IO
3682 RegionNode* res_ctl = new RegionNode(3);
3683 res_ctl->init_req(1, null_ctl);
3684 gvn().set_type(res_ctl, Type::CONTROL);
3685 record_for_igvn(res_ctl);
3686 Node* res_io = PhiNode::make(res_ctl, i_o(), Type::ABIO);
3687 gvn().set_type(res_io, Type::ABIO);
3688 record_for_igvn(res_io);
3689
3690 // TODO comment
3691 SafePointNode* loop_map = NULL;
3692 {
3693 PreserveJVMState pjvms(this);
3694 // Create default value type and store it to memory
3695 Node* oop = ValueTypeNode::make_default(gvn(), vk);
3696 oop = oop->as_ValueType()->allocate(this);
3697
3698 length = SubI(length, intcon(1));
3699 add_predicate(nargs);
3700 RegionNode* loop = new RegionNode(3);
3701 loop->init_req(1, control());
3702 gvn().set_type(loop, Type::CONTROL);
3703 record_for_igvn(loop);
3704
3705 Node* index = new PhiNode(loop, TypeInt::INT);
3706 index->init_req(1, intcon(0));
3707 gvn().set_type(index, TypeInt::INT);
3708 record_for_igvn(index);
3709
3710 // TODO explain why we need to capture all memory
3711 PhiNode* mem = new PhiNode(loop, Type::MEMORY, TypePtr::BOTTOM);
3712 mem->init_req(1, reset_memory());
3713 gvn().set_type(mem, Type::MEMORY);
3714 record_for_igvn(mem);
3715 set_control(loop);
3716 set_all_memory(mem);
3717 // Initialize array element
3718 Node* adr = array_element_address(array, index, T_OBJECT);
3719 const TypeOopPtr* elemtype = TypeValueTypePtr::make(TypePtr::NotNull, vk);
3720 Node* store = store_oop_to_array(control(), array, adr, TypeAryPtr::OOPS, oop, elemtype, T_OBJECT, MemNode::release);
3721
3722 IfNode* iff = create_and_map_if(control(), Bool(CmpI(index, length), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
3723 loop->init_req(2, IfTrue(iff));
3724 mem->init_req(2, merged_memory());
3725 index->init_req(2, AddI(index, intcon(1)));
3726
3727 res_ctl->init_req(2, IfFalse(iff));
3728 res_io->set_req(2, i_o());
3729 loop_map = stop();
3730 }
3731 // Set merged control, IO and memory
3732 set_control(res_ctl);
3733 set_i_o(res_io);
3734 merge_memory(loop_map->merged_memory(), res_ctl, 2);
3735
3736 // Transform new memory Phis.
3737 for (MergeMemStream mms(merged_memory()); mms.next_non_empty();) {
3738 Node* phi = mms.memory();
3739 if (phi->is_Phi() && phi->in(0) == res_ctl) {
3740 mms.set_memory(gvn().transform(phi));
3741 }
3742 }
3743 }
3744
3745 // The following "Ideal_foo" functions are placed here because they recognize
3746 // the graph shapes created by the functions immediately above.
3747
3748 //---------------------------Ideal_allocation----------------------------------
3749 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3750 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3751 if (ptr == NULL) { // reduce dumb test in callers
3752 return NULL;
3753 }
3754 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3755 ptr = ptr->in(1);
3756 if (ptr == NULL) return NULL;
3757 }
3758 // Return NULL for allocations with several casts:
3759 // j.l.reflect.Array.newInstance(jobject, jint)
3760 // Object.clone()
3761 // to keep more precise type from last cast.
3762 if (ptr->is_Proj()) {
3763 Node* allo = ptr->in(0);
3764 if (allo != NULL && allo->is_Allocate()) {
3765 return allo->as_Allocate();
3766 }
3767 }
3768 // Report failure to match.
3769 return NULL;
3770 }
3771
3772 // Fancy version which also strips off an offset (and reports it to caller).
3773 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3774 intptr_t& offset) {
3775 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3776 if (base == NULL) return NULL;
3777 return Ideal_allocation(base, phase);
3778 }
3779
3780 // Trace Initialize <- Proj[Parm] <- Allocate
3781 AllocateNode* InitializeNode::allocation() {
3782 Node* rawoop = in(InitializeNode::RawAddress);
3783 if (rawoop->is_Proj()) {
3784 Node* alloc = rawoop->in(0);
3785 if (alloc->is_Allocate()) {
3786 return alloc->as_Allocate();
3787 }
3788 }
3789 return NULL;
3790 }
3791
3792 // Trace Allocate -> Proj[Parm] -> Initialize
3793 InitializeNode* AllocateNode::initialization() {
3794 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3795 if (rawoop == NULL) return NULL;
3796 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3797 Node* init = rawoop->fast_out(i);
3798 if (init->is_Initialize()) {
3799 assert(init->as_Initialize()->allocation() == this, "2-way link");
3800 return init->as_Initialize();
3801 }
3802 }
3803 return NULL;
3804 }
3805
3806 //----------------------------- loop predicates ---------------------------
3807
3808 //------------------------------add_predicate_impl----------------------------
3809 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3810 // Too many traps seen?
3811 if (too_many_traps(reason)) {
3812 #ifdef ASSERT
3813 if (TraceLoopPredicate) {
3814 int tc = C->trap_count(reason);
3815 tty->print("too many traps=%s tcount=%d in ",
3816 Deoptimization::trap_reason_name(reason), tc);
3817 method()->print(); // which method has too many predicate traps
3818 tty->cr();
3819 }
3820 #endif
3821 // We cannot afford to take more traps here,
3822 // do not generate predicate.
3823 return;
3824 }
3825
3826 Node *cont = _gvn.intcon(1);
3827 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
3828 Node *bol = _gvn.transform(new Conv2BNode(opq));
3829 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3830 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3831 C->add_predicate_opaq(opq);
3832 {
3833 PreserveJVMState pjvms(this);
3834 set_control(iffalse);
3835 inc_sp(nargs);
3836 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3837 }
3838 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3839 set_control(iftrue);
3840 }
3841
3842 //------------------------------add_predicate---------------------------------
3843 void GraphKit::add_predicate(int nargs) {
3844 if (UseLoopPredicate) {
3845 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3846 }
3847 // loop's limit check predicate should be near the loop.
3848 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3849 }
3850
3851 //----------------------------- store barriers ----------------------------
3852 #define __ ideal.
3853
3854 void GraphKit::sync_kit(IdealKit& ideal) {
3855 set_all_memory(__ merged_memory());
3856 set_i_o(__ i_o());
3857 set_control(__ ctrl());
3858 }
3859
3860 void GraphKit::final_sync(IdealKit& ideal) {
3861 // Final sync IdealKit and graphKit.
3862 sync_kit(ideal);
3863 }
3864
3865 Node* GraphKit::byte_map_base_node() {
3866 // Get base of card map
3867 CardTableModRefBS* ct =
3868 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
3869 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code");
3870 if (ct->byte_map_base != NULL) {
3871 return makecon(TypeRawPtr::make((address)ct->byte_map_base));
3872 } else {
3873 return null();
3874 }
3875 }
3876
3877 // vanilla/CMS post barrier
3878 // Insert a write-barrier store. This is to let generational GC work; we have
3879 // to flag all oop-stores before the next GC point.
3880 void GraphKit::write_barrier_post(Node* oop_store,
3881 Node* obj,
3882 Node* adr,
3883 uint adr_idx,
3884 Node* val,
3885 bool use_precise) {
3886 // No store check needed if we're storing a NULL or an old object
3887 // (latter case is probably a string constant). The concurrent
3888 // mark sweep garbage collector, however, needs to have all nonNull
3889 // oop updates flagged via card-marks.
3890 if (val != NULL && val->is_Con()) {
3891 // must be either an oop or NULL
3892 const Type* t = val->bottom_type();
3893 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3894 // stores of null never (?) need barriers
3895 return;
3896 }
3897
3898 if (use_ReduceInitialCardMarks()
3899 && obj == just_allocated_object(control())) {
3900 // We can skip marks on a freshly-allocated object in Eden.
3901 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3902 // That routine informs GC to take appropriate compensating steps,
3903 // upon a slow-path allocation, so as to make this card-mark
3904 // elision safe.
3905 return;
3906 }
3907
3908 if (!use_precise) {
3909 // All card marks for a (non-array) instance are in one place:
3910 adr = obj;
3911 }
3912 // (Else it's an array (or unknown), and we want more precise card marks.)
3913 assert(adr != NULL, "");
3914
3915 IdealKit ideal(this, true);
3916
3917 // Convert the pointer to an int prior to doing math on it
3918 Node* cast = __ CastPX(__ ctrl(), adr);
3919
3920 // Divide by card size
3921 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef),
3922 "Only one we handle so far.");
3923 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3924
3925 // Combine card table base and card offset
3926 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3927
3928 // Get the alias_index for raw card-mark memory
3929 int adr_type = Compile::AliasIdxRaw;
3930 Node* zero = __ ConI(0); // Dirty card value
3931 BasicType bt = T_BYTE;
3932
3933 if (UseConcMarkSweepGC && UseCondCardMark) {
3934 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
3935 __ sync_kit(this);
3936 }
3937
3938 if (UseCondCardMark) {
3939 // The classic GC reference write barrier is typically implemented
3940 // as a store into the global card mark table. Unfortunately
3941 // unconditional stores can result in false sharing and excessive
3942 // coherence traffic as well as false transactional aborts.
3943 // UseCondCardMark enables MP "polite" conditional card mark
3944 // stores. In theory we could relax the load from ctrl() to
3945 // no_ctrl, but that doesn't buy much latitude.
3946 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3947 __ if_then(card_val, BoolTest::ne, zero);
3948 }
3949
3950 // Smash zero into card
3951 if( !UseConcMarkSweepGC ) {
3952 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3953 } else {
3954 // Specialized path for CM store barrier
3955 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3956 }
3957
3958 if (UseCondCardMark) {
3959 __ end_if();
3960 }
3961
3962 // Final sync IdealKit and GraphKit.
3963 final_sync(ideal);
3964 }
3965 /*
3966 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
3967 * required by SATB to make sure all objects live at the start of the
3968 * marking are kept alive, all reference updates need to any previous
3969 * reference stored before writing.
3970 *
3971 * If the previous value is NULL there is no need to save the old value.
3972 * References that are NULL are filtered during runtime by the barrier
3973 * code to avoid unnecessary queuing.
3974 *
3975 * However in the case of newly allocated objects it might be possible to
3976 * prove that the reference about to be overwritten is NULL during compile
3977 * time and avoid adding the barrier code completely.
3978 *
3979 * The compiler needs to determine that the object in which a field is about
3980 * to be written is newly allocated, and that no prior store to the same field
3981 * has happened since the allocation.
3982 *
3983 * Returns true if the pre-barrier can be removed
3984 */
3985 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
3986 BasicType bt, uint adr_idx) {
3987 intptr_t offset = 0;
3988 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3989 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
3990
3991 if (offset == Type::OffsetBot) {
3992 return false; // cannot unalias unless there are precise offsets
3993 }
3994
3995 if (alloc == NULL) {
3996 return false; // No allocation found
3997 }
3998
3999 intptr_t size_in_bytes = type2aelembytes(bt);
4000
4001 Node* mem = memory(adr_idx); // start searching here...
4002
4003 for (int cnt = 0; cnt < 50; cnt++) {
4004
4005 if (mem->is_Store()) {
4006
4007 Node* st_adr = mem->in(MemNode::Address);
4008 intptr_t st_offset = 0;
4009 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
4010
4011 if (st_base == NULL) {
4012 break; // inscrutable pointer
4013 }
4014
4015 // Break we have found a store with same base and offset as ours so break
4016 if (st_base == base && st_offset == offset) {
4017 break;
4018 }
4019
4020 if (st_offset != offset && st_offset != Type::OffsetBot) {
4021 const int MAX_STORE = BytesPerLong;
4022 if (st_offset >= offset + size_in_bytes ||
4023 st_offset <= offset - MAX_STORE ||
4024 st_offset <= offset - mem->as_Store()->memory_size()) {
4025 // Success: The offsets are provably independent.
4026 // (You may ask, why not just test st_offset != offset and be done?
4027 // The answer is that stores of different sizes can co-exist
4028 // in the same sequence of RawMem effects. We sometimes initialize
4029 // a whole 'tile' of array elements with a single jint or jlong.)
4030 mem = mem->in(MemNode::Memory);
4031 continue; // advance through independent store memory
4032 }
4033 }
4034
4035 if (st_base != base
4036 && MemNode::detect_ptr_independence(base, alloc, st_base,
4037 AllocateNode::Ideal_allocation(st_base, phase),
4038 phase)) {
4039 // Success: The bases are provably independent.
4040 mem = mem->in(MemNode::Memory);
4041 continue; // advance through independent store memory
4042 }
4043 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4044
4045 InitializeNode* st_init = mem->in(0)->as_Initialize();
4046 AllocateNode* st_alloc = st_init->allocation();
4047
4048 // Make sure that we are looking at the same allocation site.
4049 // The alloc variable is guaranteed to not be null here from earlier check.
4050 if (alloc == st_alloc) {
4051 // Check that the initialization is storing NULL so that no previous store
4052 // has been moved up and directly write a reference
4053 Node* captured_store = st_init->find_captured_store(offset,
4054 type2aelembytes(T_OBJECT),
4055 phase);
4056 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4057 return true;
4058 }
4059 }
4060 }
4061
4062 // Unless there is an explicit 'continue', we must bail out here,
4063 // because 'mem' is an inscrutable memory state (e.g., a call).
4064 break;
4065 }
4066
4067 return false;
4068 }
4069
4070 // G1 pre/post barriers
4071 void GraphKit::g1_write_barrier_pre(bool do_load,
4072 Node* obj,
4073 Node* adr,
4074 uint alias_idx,
4075 Node* val,
4076 const TypeOopPtr* val_type,
4077 Node* pre_val,
4078 BasicType bt) {
4079
4080 // Some sanity checks
4081 // Note: val is unused in this routine.
4082
4083 if (do_load) {
4084 // We need to generate the load of the previous value
4085 assert(obj != NULL, "must have a base");
4086 assert(adr != NULL, "where are loading from?");
4087 assert(pre_val == NULL, "loaded already?");
4088 assert(val_type != NULL, "need a type");
4089
4090 if (use_ReduceInitialCardMarks()
4091 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4092 return;
4093 }
4094
4095 } else {
4096 // In this case both val_type and alias_idx are unused.
4097 assert(pre_val != NULL, "must be loaded already");
4098 // Nothing to be done if pre_val is null.
4099 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4100 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4101 }
4102 assert(bt == T_OBJECT || bt == T_VALUETYPE, "or we shouldn't be here");
4103
4104 IdealKit ideal(this, true);
4105
4106 Node* tls = __ thread(); // ThreadLocalStorage
4107
4108 Node* no_ctrl = NULL;
4109 Node* no_base = __ top();
4110 Node* zero = __ ConI(0);
4111 Node* zeroX = __ ConX(0);
4112
4113 float likely = PROB_LIKELY(0.999);
4114 float unlikely = PROB_UNLIKELY(0.999);
4115
4116 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4117 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4118
4119 // Offsets into the thread
4120 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
4121 SATBMarkQueue::byte_offset_of_active());
4122 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
4123 SATBMarkQueue::byte_offset_of_index());
4124 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
4125 SATBMarkQueue::byte_offset_of_buf());
4126
4127 // Now the actual pointers into the thread
4128 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4129 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4130 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4131
4132 // Now some of the values
4133 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4134
4135 // if (!marking)
4136 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4137 BasicType index_bt = TypeX_X->basic_type();
4138 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4139 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4140
4141 if (do_load) {
4142 // load original value
4143 // alias_idx correct??
4144 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4145 }
4146
4147 // if (pre_val != NULL)
4148 __ if_then(pre_val, BoolTest::ne, null()); {
4149 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4150
4151 // is the queue for this thread full?
4152 __ if_then(index, BoolTest::ne, zeroX, likely); {
4153
4154 // decrement the index
4155 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4156
4157 // Now get the buffer location we will log the previous value into and store it
4158 Node *log_addr = __ AddP(no_base, buffer, next_index);
4159 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4160 // update the index
4161 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4162
4163 } __ else_(); {
4164
4165 // logging buffer is full, call the runtime
4166 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4167 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4168 } __ end_if(); // (!index)
4169 } __ end_if(); // (pre_val != NULL)
4170 } __ end_if(); // (!marking)
4171
4172 // Final sync IdealKit and GraphKit.
4173 final_sync(ideal);
4174 }
4175
4176 /*
4177 * G1 similar to any GC with a Young Generation requires a way to keep track of
4178 * references from Old Generation to Young Generation to make sure all live
4179 * objects are found. G1 also requires to keep track of object references
4180 * between different regions to enable evacuation of old regions, which is done
4181 * as part of mixed collections. References are tracked in remembered sets and
4182 * is continuously updated as reference are written to with the help of the
4183 * post-barrier.
4184 *
4185 * To reduce the number of updates to the remembered set the post-barrier
4186 * filters updates to fields in objects located in the Young Generation,
4187 * the same region as the reference, when the NULL is being written or
4188 * if the card is already marked as dirty by an earlier write.
4189 *
4190 * Under certain circumstances it is possible to avoid generating the
4191 * post-barrier completely if it is possible during compile time to prove
4192 * the object is newly allocated and that no safepoint exists between the
4193 * allocation and the store.
4194 *
4195 * In the case of slow allocation the allocation code must handle the barrier
4196 * as part of the allocation in the case the allocated object is not located
4197 * in the nursery, this would happen for humongous objects. This is similar to
4198 * how CMS is required to handle this case, see the comments for the method
4199 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier.
4200 * A deferred card mark is required for these objects and handled in the above
4201 * mentioned methods.
4202 *
4203 * Returns true if the post barrier can be removed
4204 */
4205 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4206 Node* adr) {
4207 intptr_t offset = 0;
4208 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4209 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4210
4211 if (offset == Type::OffsetBot) {
4212 return false; // cannot unalias unless there are precise offsets
4213 }
4214
4215 if (alloc == NULL) {
4216 return false; // No allocation found
4217 }
4218
4219 // Start search from Store node
4220 Node* mem = store->in(MemNode::Control);
4221 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4222
4223 InitializeNode* st_init = mem->in(0)->as_Initialize();
4224 AllocateNode* st_alloc = st_init->allocation();
4225
4226 // Make sure we are looking at the same allocation
4227 if (alloc == st_alloc) {
4228 return true;
4229 }
4230 }
4231
4232 return false;
4233 }
4234
4235 //
4236 // Update the card table and add card address to the queue
4237 //
4238 void GraphKit::g1_mark_card(IdealKit& ideal,
4239 Node* card_adr,
4240 Node* oop_store,
4241 uint oop_alias_idx,
4242 Node* index,
4243 Node* index_adr,
4244 Node* buffer,
4245 const TypeFunc* tf) {
4246
4247 Node* zero = __ ConI(0);
4248 Node* zeroX = __ ConX(0);
4249 Node* no_base = __ top();
4250 BasicType card_bt = T_BYTE;
4251 // Smash zero into card. MUST BE ORDERED WRT TO STORE
4252 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4253
4254 // Now do the queue work
4255 __ if_then(index, BoolTest::ne, zeroX); {
4256
4257 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4258 Node* log_addr = __ AddP(no_base, buffer, next_index);
4259
4260 // Order, see storeCM.
4261 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4262 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4263
4264 } __ else_(); {
4265 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4266 } __ end_if();
4267
4268 }
4269
4270 void GraphKit::g1_write_barrier_post(Node* oop_store,
4271 Node* obj,
4272 Node* adr,
4273 uint alias_idx,
4274 Node* val,
4275 BasicType bt,
4276 bool use_precise) {
4277 // If we are writing a NULL then we need no post barrier
4278
4279 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4280 // Must be NULL
4281 const Type* t = val->bottom_type();
4282 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4283 // No post barrier if writing NULLx
4284 return;
4285 }
4286
4287 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4288 // We can skip marks on a freshly-allocated object in Eden.
4289 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
4290 // That routine informs GC to take appropriate compensating steps,
4291 // upon a slow-path allocation, so as to make this card-mark
4292 // elision safe.
4293 return;
4294 }
4295
4296 if (use_ReduceInitialCardMarks()
4297 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4298 return;
4299 }
4300
4301 if (!use_precise) {
4302 // All card marks for a (non-array) instance are in one place:
4303 adr = obj;
4304 }
4305 // (Else it's an array (or unknown), and we want more precise card marks.)
4306 assert(adr != NULL, "");
4307
4308 IdealKit ideal(this, true);
4309
4310 Node* tls = __ thread(); // ThreadLocalStorage
4311
4312 Node* no_base = __ top();
4313 float likely = PROB_LIKELY(0.999);
4314 float unlikely = PROB_UNLIKELY(0.999);
4315 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
4316 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
4317 Node* zeroX = __ ConX(0);
4318
4319 // Get the alias_index for raw card-mark memory
4320 const TypePtr* card_type = TypeRawPtr::BOTTOM;
4321
4322 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4323
4324 // Offsets into the thread
4325 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4326 DirtyCardQueue::byte_offset_of_index());
4327 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4328 DirtyCardQueue::byte_offset_of_buf());
4329
4330 // Pointers into the thread
4331
4332 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4333 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4334
4335 // Now some values
4336 // Use ctrl to avoid hoisting these values past a safepoint, which could
4337 // potentially reset these fields in the JavaThread.
4338 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4339 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4340
4341 // Convert the store obj pointer to an int prior to doing math on it
4342 // Must use ctrl to prevent "integerized oop" existing across safepoint
4343 Node* cast = __ CastPX(__ ctrl(), adr);
4344
4345 // Divide pointer by card size
4346 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4347
4348 // Combine card table base and card offset
4349 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4350
4351 // If we know the value being stored does it cross regions?
4352
4353 if (val != NULL) {
4354 // Does the store cause us to cross regions?
4355
4356 // Should be able to do an unsigned compare of region_size instead of
4357 // and extra shift. Do we have an unsigned compare??
4358 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4359 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4360
4361 // if (xor_res == 0) same region so skip
4362 __ if_then(xor_res, BoolTest::ne, zeroX); {
4363
4364 // No barrier if we are storing a NULL
4365 __ if_then(val, BoolTest::ne, null(), unlikely); {
4366
4367 // Ok must mark the card if not already dirty
4368
4369 // load the original value of the card
4370 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4371
4372 __ if_then(card_val, BoolTest::ne, young_card); {
4373 sync_kit(ideal);
4374 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4375 insert_mem_bar(Op_MemBarVolatile, oop_store);
4376 __ sync_kit(this);
4377
4378 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4379 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4380 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4381 } __ end_if();
4382 } __ end_if();
4383 } __ end_if();
4384 } __ end_if();
4385 } else {
4386 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
4387 // We don't need a barrier here if the destination is a newly allocated object
4388 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
4389 // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()).
4390 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
4391 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4392 __ if_then(card_val, BoolTest::ne, young_card); {
4393 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4394 } __ end_if();
4395 }
4396
4397 // Final sync IdealKit and GraphKit.
4398 final_sync(ideal);
4399 }
4400 #undef __
4401
4402
4403 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4404 Node* len = load_array_length(load_String_value(ctrl, str));
4405 Node* coder = load_String_coder(ctrl, str);
4406 // Divide length by 2 if coder is UTF16
4407 return _gvn.transform(new RShiftINode(len, coder));
4408 }
4409
4410 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4411 int value_offset = java_lang_String::value_offset_in_bytes();
4412 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4413 false, NULL, Type::Offset(0));
4414 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4415 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4416 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4417 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4418 int value_field_idx = C->get_alias_index(value_field_type);
4419 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4420 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4421 // String.value field is known to be @Stable.
4422 if (UseImplicitStableValues) {
4423 load = cast_array_to_stable(load, value_type);
4424 }
4425 return load;
4426 }
4427
4428 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4429 if (!CompactStrings) {
4430 return intcon(java_lang_String::CODER_UTF16);
4431 }
4432 int coder_offset = java_lang_String::coder_offset_in_bytes();
4433 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4434 false, NULL, Type::Offset(0));
4435 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4436 int coder_field_idx = C->get_alias_index(coder_field_type);
4437 return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4438 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4439 }
4440
4441 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4442 int value_offset = java_lang_String::value_offset_in_bytes();
4443 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4444 false, NULL, Type::Offset(0));
4445 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4446 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4447 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4448 }
4449
4450 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4451 int coder_offset = java_lang_String::coder_offset_in_bytes();
4452 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4453 false, NULL, Type::Offset(0));
4454 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4455 int coder_field_idx = C->get_alias_index(coder_field_type);
4456 store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4457 value, T_BYTE, coder_field_idx, MemNode::unordered);
4458 }
4459
4460 // Capture src and dst memory state with a MergeMemNode
4461 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4462 if (src_type == dst_type) {
4463 // Types are equal, we don't need a MergeMemNode
4464 return memory(src_type);
4465 }
4466 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4467 record_for_igvn(merge); // fold it up later, if possible
4468 int src_idx = C->get_alias_index(src_type);
4469 int dst_idx = C->get_alias_index(dst_type);
4470 merge->set_memory_at(src_idx, memory(src_idx));
4471 merge->set_memory_at(dst_idx, memory(dst_idx));
4472 return merge;
4473 }
4474
4475 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4476 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4477 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4478 // If input and output memory types differ, capture both states to preserve
4479 // the dependency between preceding and subsequent loads/stores.
4480 // For example, the following program:
4481 // StoreB
4482 // compress_string
4483 // LoadB
4484 // has this memory graph (use->def):
4485 // LoadB -> compress_string -> CharMem
4486 // ... -> StoreB -> ByteMem
4487 // The intrinsic hides the dependency between LoadB and StoreB, causing
4488 // the load to read from memory not containing the result of the StoreB.
4489 // The correct memory graph should look like this:
4490 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4491 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4492 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4493 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4494 set_memory(res_mem, TypeAryPtr::BYTES);
4495 return str;
4496 }
4497
4498 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4499 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4500 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4501 // Capture src and dst memory (see comment in 'compress_string').
4502 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4503 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4504 set_memory(_gvn.transform(str), dst_type);
4505 }
4506
4507 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4508 /**
4509 * int i_char = start;
4510 * for (int i_byte = 0; i_byte < count; i_byte++) {
4511 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4512 * }
4513 */
4514 add_predicate();
4515 RegionNode* head = new RegionNode(3);
4516 head->init_req(1, control());
4517 gvn().set_type(head, Type::CONTROL);
4518 record_for_igvn(head);
4519
4520 Node* i_byte = new PhiNode(head, TypeInt::INT);
4521 i_byte->init_req(1, intcon(0));
4522 gvn().set_type(i_byte, TypeInt::INT);
4523 record_for_igvn(i_byte);
4524
4525 Node* i_char = new PhiNode(head, TypeInt::INT);
4526 i_char->init_req(1, start);
4527 gvn().set_type(i_char, TypeInt::INT);
4528 record_for_igvn(i_char);
4529
4530 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4531 gvn().set_type(mem, Type::MEMORY);
4532 record_for_igvn(mem);
4533 set_control(head);
4534 set_memory(mem, TypeAryPtr::BYTES);
4535 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4536 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4537 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4538 false, false, true /* mismatched */);
4539
4540 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4541 head->init_req(2, IfTrue(iff));
4542 mem->init_req(2, st);
4543 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4544 i_char->init_req(2, AddI(i_char, intcon(2)));
4545
4546 set_control(IfFalse(iff));
4547 set_memory(st, TypeAryPtr::BYTES);
4548 }
4549
4550 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4551 if (!field->is_constant()) {
4552 return NULL; // Field not marked as constant.
4553 }
4554 ciInstance* holder = NULL;
4555 if (!field->is_static()) {
4556 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4557 if (const_oop != NULL && const_oop->is_instance()) {
4558 holder = const_oop->as_instance();
4559 }
4560 }
4561 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4562 /*is_unsigned_load=*/false);
4563 if (con_type != NULL) {
4564 Node* con = makecon(con_type);
4565 if (field->layout_type() == T_VALUETYPE) {
4566 // Load value type from constant oop
4567 con = ValueTypeNode::make(gvn(), map()->memory(), con);
4568 }
4569 return con;
4570 }
4571 return NULL;
4572 }
4573
4574 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4575 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4576 // assumption of CCP analysis.
4577 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4578 }