1 /*
2 * Copyright (c) 2001, 2018, 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 "ci/ciUtilities.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/g1/g1BarrierSet.hpp"
29 #include "gc/g1/g1CardTable.hpp"
30 #include "gc/g1/heapRegion.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/cardTable.hpp"
33 #include "gc/shared/cardTableBarrierSet.hpp"
34 #include "gc/shared/collectedHeap.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "opto/addnode.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/intrinsicnode.hpp"
43 #include "opto/locknode.hpp"
44 #include "opto/machnode.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/parse.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "runtime/deoptimization.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #if INCLUDE_ALL_GCS
52 #include "gc/g1/g1ThreadLocalData.hpp"
53 #endif // INCLUDE_ALL_GCS
54
55 //----------------------------GraphKit-----------------------------------------
56 // Main utility constructor.
57 GraphKit::GraphKit(JVMState* jvms)
58 : Phase(Phase::Parser),
59 _env(C->env()),
60 _gvn(*C->initial_gvn())
61 {
62 _exceptions = jvms->map()->next_exception();
63 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
64 set_jvms(jvms);
65 }
66
67 // Private constructor for parser.
68 GraphKit::GraphKit()
69 : Phase(Phase::Parser),
70 _env(C->env()),
71 _gvn(*C->initial_gvn())
72 {
73 _exceptions = NULL;
74 set_map(NULL);
75 debug_only(_sp = -99);
76 debug_only(set_bci(-99));
77 }
78
79
80
81 //---------------------------clean_stack---------------------------------------
82 // Clear away rubbish from the stack area of the JVM state.
83 // This destroys any arguments that may be waiting on the stack.
84 void GraphKit::clean_stack(int from_sp) {
85 SafePointNode* map = this->map();
86 JVMState* jvms = this->jvms();
87 int stk_size = jvms->stk_size();
88 int stkoff = jvms->stkoff();
89 Node* top = this->top();
90 for (int i = from_sp; i < stk_size; i++) {
91 if (map->in(stkoff + i) != top) {
92 map->set_req(stkoff + i, top);
93 }
94 }
95 }
96
97
98 //--------------------------------sync_jvms-----------------------------------
99 // Make sure our current jvms agrees with our parse state.
100 JVMState* GraphKit::sync_jvms() const {
101 JVMState* jvms = this->jvms();
102 jvms->set_bci(bci()); // Record the new bci in the JVMState
103 jvms->set_sp(sp()); // Record the new sp in the JVMState
104 assert(jvms_in_sync(), "jvms is now in sync");
105 return jvms;
106 }
107
108 //--------------------------------sync_jvms_for_reexecute---------------------
109 // Make sure our current jvms agrees with our parse state. This version
110 // uses the reexecute_sp for reexecuting bytecodes.
111 JVMState* GraphKit::sync_jvms_for_reexecute() {
112 JVMState* jvms = this->jvms();
113 jvms->set_bci(bci()); // Record the new bci in the JVMState
114 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
115 return jvms;
116 }
117
118 #ifdef ASSERT
119 bool GraphKit::jvms_in_sync() const {
120 Parse* parse = is_Parse();
121 if (parse == NULL) {
122 if (bci() != jvms()->bci()) return false;
123 if (sp() != (int)jvms()->sp()) return false;
124 return true;
125 }
126 if (jvms()->method() != parse->method()) return false;
127 if (jvms()->bci() != parse->bci()) return false;
128 int jvms_sp = jvms()->sp();
129 if (jvms_sp != parse->sp()) return false;
130 int jvms_depth = jvms()->depth();
131 if (jvms_depth != parse->depth()) return false;
132 return true;
133 }
134
135 // Local helper checks for special internal merge points
136 // used to accumulate and merge exception states.
137 // They are marked by the region's in(0) edge being the map itself.
138 // Such merge points must never "escape" into the parser at large,
139 // until they have been handed to gvn.transform.
140 static bool is_hidden_merge(Node* reg) {
141 if (reg == NULL) return false;
142 if (reg->is_Phi()) {
143 reg = reg->in(0);
144 if (reg == NULL) return false;
145 }
146 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
147 }
148
149 void GraphKit::verify_map() const {
150 if (map() == NULL) return; // null map is OK
151 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
152 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
153 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
154 }
155
156 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
157 assert(ex_map->next_exception() == NULL, "not already part of a chain");
158 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
159 }
160 #endif
161
162 //---------------------------stop_and_kill_map---------------------------------
163 // Set _map to NULL, signalling a stop to further bytecode execution.
164 // First smash the current map's control to a constant, to mark it dead.
165 void GraphKit::stop_and_kill_map() {
166 SafePointNode* dead_map = stop();
167 if (dead_map != NULL) {
168 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
169 assert(dead_map->is_killed(), "must be so marked");
170 }
171 }
172
173
174 //--------------------------------stopped--------------------------------------
175 // Tell if _map is NULL, or control is top.
176 bool GraphKit::stopped() {
177 if (map() == NULL) return true;
178 else if (control() == top()) return true;
179 else return false;
180 }
181
182
183 //-----------------------------has_ex_handler----------------------------------
184 // Tell if this method or any caller method has exception handlers.
185 bool GraphKit::has_ex_handler() {
186 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
187 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
188 return true;
189 }
190 }
191 return false;
192 }
193
194 //------------------------------save_ex_oop------------------------------------
195 // Save an exception without blowing stack contents or other JVM state.
196 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
197 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
198 ex_map->add_req(ex_oop);
199 debug_only(verify_exception_state(ex_map));
200 }
201
202 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
203 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
204 Node* ex_oop = ex_map->in(ex_map->req()-1);
205 if (clear_it) ex_map->del_req(ex_map->req()-1);
206 return ex_oop;
207 }
208
209 //-----------------------------saved_ex_oop------------------------------------
210 // Recover a saved exception from its map.
211 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
212 return common_saved_ex_oop(ex_map, false);
213 }
214
215 //--------------------------clear_saved_ex_oop---------------------------------
216 // Erase a previously saved exception from its map.
217 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
218 return common_saved_ex_oop(ex_map, true);
219 }
220
221 #ifdef ASSERT
222 //---------------------------has_saved_ex_oop----------------------------------
223 // Erase a previously saved exception from its map.
224 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
225 return ex_map->req() == ex_map->jvms()->endoff()+1;
226 }
227 #endif
228
229 //-------------------------make_exception_state--------------------------------
230 // Turn the current JVM state into an exception state, appending the ex_oop.
231 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
232 sync_jvms();
233 SafePointNode* ex_map = stop(); // do not manipulate this map any more
234 set_saved_ex_oop(ex_map, ex_oop);
235 return ex_map;
236 }
237
238
239 //--------------------------add_exception_state--------------------------------
240 // Add an exception to my list of exceptions.
241 void GraphKit::add_exception_state(SafePointNode* ex_map) {
242 if (ex_map == NULL || ex_map->control() == top()) {
243 return;
244 }
245 #ifdef ASSERT
246 verify_exception_state(ex_map);
247 if (has_exceptions()) {
248 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
249 }
250 #endif
251
252 // If there is already an exception of exactly this type, merge with it.
253 // In particular, null-checks and other low-level exceptions common up here.
254 Node* ex_oop = saved_ex_oop(ex_map);
255 const Type* ex_type = _gvn.type(ex_oop);
256 if (ex_oop == top()) {
257 // No action needed.
258 return;
259 }
260 assert(ex_type->isa_instptr(), "exception must be an instance");
261 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
262 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
263 // We check sp also because call bytecodes can generate exceptions
264 // both before and after arguments are popped!
265 if (ex_type2 == ex_type
266 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
267 combine_exception_states(ex_map, e2);
268 return;
269 }
270 }
271
272 // No pre-existing exception of the same type. Chain it on the list.
273 push_exception_state(ex_map);
274 }
275
276 //-----------------------add_exception_states_from-----------------------------
277 void GraphKit::add_exception_states_from(JVMState* jvms) {
278 SafePointNode* ex_map = jvms->map()->next_exception();
279 if (ex_map != NULL) {
280 jvms->map()->set_next_exception(NULL);
281 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
282 next_map = ex_map->next_exception();
283 ex_map->set_next_exception(NULL);
284 add_exception_state(ex_map);
285 }
286 }
287 }
288
289 //-----------------------transfer_exceptions_into_jvms-------------------------
290 JVMState* GraphKit::transfer_exceptions_into_jvms() {
291 if (map() == NULL) {
292 // We need a JVMS to carry the exceptions, but the map has gone away.
293 // Create a scratch JVMS, cloned from any of the exception states...
294 if (has_exceptions()) {
295 _map = _exceptions;
296 _map = clone_map();
297 _map->set_next_exception(NULL);
298 clear_saved_ex_oop(_map);
299 debug_only(verify_map());
300 } else {
301 // ...or created from scratch
302 JVMState* jvms = new (C) JVMState(_method, NULL);
303 jvms->set_bci(_bci);
304 jvms->set_sp(_sp);
305 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
306 set_jvms(jvms);
307 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
308 set_all_memory(top());
309 while (map()->req() < jvms->endoff()) map()->add_req(top());
310 }
311 // (This is a kludge, in case you didn't notice.)
312 set_control(top());
313 }
314 JVMState* jvms = sync_jvms();
315 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
316 jvms->map()->set_next_exception(_exceptions);
317 _exceptions = NULL; // done with this set of exceptions
318 return jvms;
319 }
320
321 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
322 assert(is_hidden_merge(dstphi), "must be a special merge node");
323 assert(is_hidden_merge(srcphi), "must be a special merge node");
324 uint limit = srcphi->req();
325 for (uint i = PhiNode::Input; i < limit; i++) {
326 dstphi->add_req(srcphi->in(i));
327 }
328 }
329 static inline void add_one_req(Node* dstphi, Node* src) {
330 assert(is_hidden_merge(dstphi), "must be a special merge node");
331 assert(!is_hidden_merge(src), "must not be a special merge node");
332 dstphi->add_req(src);
333 }
334
335 //-----------------------combine_exception_states------------------------------
336 // This helper function combines exception states by building phis on a
337 // specially marked state-merging region. These regions and phis are
338 // untransformed, and can build up gradually. The region is marked by
339 // having a control input of its exception map, rather than NULL. Such
340 // regions do not appear except in this function, and in use_exception_state.
341 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
342 if (failing()) return; // dying anyway...
343 JVMState* ex_jvms = ex_map->_jvms;
344 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
345 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
346 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
347 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
348 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
349 assert(ex_map->req() == phi_map->req(), "matching maps");
350 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
351 Node* hidden_merge_mark = root();
352 Node* region = phi_map->control();
353 MergeMemNode* phi_mem = phi_map->merged_memory();
354 MergeMemNode* ex_mem = ex_map->merged_memory();
355 if (region->in(0) != hidden_merge_mark) {
356 // The control input is not (yet) a specially-marked region in phi_map.
357 // Make it so, and build some phis.
358 region = new RegionNode(2);
359 _gvn.set_type(region, Type::CONTROL);
360 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
361 region->init_req(1, phi_map->control());
362 phi_map->set_control(region);
363 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
364 record_for_igvn(io_phi);
365 _gvn.set_type(io_phi, Type::ABIO);
366 phi_map->set_i_o(io_phi);
367 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
368 Node* m = mms.memory();
369 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
370 record_for_igvn(m_phi);
371 _gvn.set_type(m_phi, Type::MEMORY);
372 mms.set_memory(m_phi);
373 }
374 }
375
376 // Either or both of phi_map and ex_map might already be converted into phis.
377 Node* ex_control = ex_map->control();
378 // if there is special marking on ex_map also, we add multiple edges from src
379 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
380 // how wide was the destination phi_map, originally?
381 uint orig_width = region->req();
382
383 if (add_multiple) {
384 add_n_reqs(region, ex_control);
385 add_n_reqs(phi_map->i_o(), ex_map->i_o());
386 } else {
387 // ex_map has no merges, so we just add single edges everywhere
388 add_one_req(region, ex_control);
389 add_one_req(phi_map->i_o(), ex_map->i_o());
390 }
391 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
392 if (mms.is_empty()) {
393 // get a copy of the base memory, and patch some inputs into it
394 const TypePtr* adr_type = mms.adr_type(C);
395 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
396 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
397 mms.set_memory(phi);
398 // Prepare to append interesting stuff onto the newly sliced phi:
399 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
400 }
401 // Append stuff from ex_map:
402 if (add_multiple) {
403 add_n_reqs(mms.memory(), mms.memory2());
404 } else {
405 add_one_req(mms.memory(), mms.memory2());
406 }
407 }
408 uint limit = ex_map->req();
409 for (uint i = TypeFunc::Parms; i < limit; i++) {
410 // Skip everything in the JVMS after tos. (The ex_oop follows.)
411 if (i == tos) i = ex_jvms->monoff();
412 Node* src = ex_map->in(i);
413 Node* dst = phi_map->in(i);
414 if (src != dst) {
415 PhiNode* phi;
416 if (dst->in(0) != region) {
417 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
418 record_for_igvn(phi);
419 _gvn.set_type(phi, phi->type());
420 phi_map->set_req(i, dst);
421 // Prepare to append interesting stuff onto the new phi:
422 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
423 } else {
424 assert(dst->is_Phi(), "nobody else uses a hidden region");
425 phi = dst->as_Phi();
426 }
427 if (add_multiple && src->in(0) == ex_control) {
428 // Both are phis.
429 add_n_reqs(dst, src);
430 } else {
431 while (dst->req() < region->req()) add_one_req(dst, src);
432 }
433 const Type* srctype = _gvn.type(src);
434 if (phi->type() != srctype) {
435 const Type* dsttype = phi->type()->meet_speculative(srctype);
436 if (phi->type() != dsttype) {
437 phi->set_type(dsttype);
438 _gvn.set_type(phi, dsttype);
439 }
440 }
441 }
442 }
443 phi_map->merge_replaced_nodes_with(ex_map);
444 }
445
446 //--------------------------use_exception_state--------------------------------
447 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
448 if (failing()) { stop(); return top(); }
449 Node* region = phi_map->control();
450 Node* hidden_merge_mark = root();
451 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
452 Node* ex_oop = clear_saved_ex_oop(phi_map);
453 if (region->in(0) == hidden_merge_mark) {
454 // Special marking for internal ex-states. Process the phis now.
455 region->set_req(0, region); // now it's an ordinary region
456 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
457 // Note: Setting the jvms also sets the bci and sp.
458 set_control(_gvn.transform(region));
459 uint tos = jvms()->stkoff() + sp();
460 for (uint i = 1; i < tos; i++) {
461 Node* x = phi_map->in(i);
462 if (x->in(0) == region) {
463 assert(x->is_Phi(), "expected a special phi");
464 phi_map->set_req(i, _gvn.transform(x));
465 }
466 }
467 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
468 Node* x = mms.memory();
469 if (x->in(0) == region) {
470 assert(x->is_Phi(), "nobody else uses a hidden region");
471 mms.set_memory(_gvn.transform(x));
472 }
473 }
474 if (ex_oop->in(0) == region) {
475 assert(ex_oop->is_Phi(), "expected a special phi");
476 ex_oop = _gvn.transform(ex_oop);
477 }
478 } else {
479 set_jvms(phi_map->jvms());
480 }
481
482 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
483 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
484 return ex_oop;
485 }
486
487 //---------------------------------java_bc-------------------------------------
488 Bytecodes::Code GraphKit::java_bc() const {
489 ciMethod* method = this->method();
490 int bci = this->bci();
491 if (method != NULL && bci != InvocationEntryBci)
492 return method->java_code_at_bci(bci);
493 else
494 return Bytecodes::_illegal;
495 }
496
497 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
498 bool must_throw) {
499 // if the exception capability is set, then we will generate code
500 // to check the JavaThread.should_post_on_exceptions flag to see
501 // if we actually need to report exception events (for this
502 // thread). If we don't need to report exception events, we will
503 // take the normal fast path provided by add_exception_events. If
504 // exception event reporting is enabled for this thread, we will
505 // take the uncommon_trap in the BuildCutout below.
506
507 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
508 Node* jthread = _gvn.transform(new ThreadLocalNode());
509 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
510 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
511
512 // Test the should_post_on_exceptions_flag vs. 0
513 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
514 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
515
516 // Branch to slow_path if should_post_on_exceptions_flag was true
517 { BuildCutout unless(this, tst, PROB_MAX);
518 // Do not try anything fancy if we're notifying the VM on every throw.
519 // Cf. case Bytecodes::_athrow in parse2.cpp.
520 uncommon_trap(reason, Deoptimization::Action_none,
521 (ciKlass*)NULL, (char*)NULL, must_throw);
522 }
523
524 }
525
526 //------------------------------builtin_throw----------------------------------
527 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
528 bool must_throw = true;
529
530 if (env()->jvmti_can_post_on_exceptions()) {
531 // check if we must post exception events, take uncommon trap if so
532 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
533 // here if should_post_on_exceptions is false
534 // continue on with the normal codegen
535 }
536
537 // If this particular condition has not yet happened at this
538 // bytecode, then use the uncommon trap mechanism, and allow for
539 // a future recompilation if several traps occur here.
540 // If the throw is hot, try to use a more complicated inline mechanism
541 // which keeps execution inside the compiled code.
542 bool treat_throw_as_hot = false;
543 ciMethodData* md = method()->method_data();
544
545 if (ProfileTraps) {
546 if (too_many_traps(reason)) {
547 treat_throw_as_hot = true;
548 }
549 // (If there is no MDO at all, assume it is early in
550 // execution, and that any deopts are part of the
551 // startup transient, and don't need to be remembered.)
552
553 // Also, if there is a local exception handler, treat all throws
554 // as hot if there has been at least one in this method.
555 if (C->trap_count(reason) != 0
556 && method()->method_data()->trap_count(reason) != 0
557 && has_ex_handler()) {
558 treat_throw_as_hot = true;
559 }
560 }
561
562 // If this throw happens frequently, an uncommon trap might cause
563 // a performance pothole. If there is a local exception handler,
564 // and if this particular bytecode appears to be deoptimizing often,
565 // let us handle the throw inline, with a preconstructed instance.
566 // Note: If the deopt count has blown up, the uncommon trap
567 // runtime is going to flush this nmethod, not matter what.
568 if (treat_throw_as_hot
569 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
570 // If the throw is local, we use a pre-existing instance and
571 // punt on the backtrace. This would lead to a missing backtrace
572 // (a repeat of 4292742) if the backtrace object is ever asked
573 // for its backtrace.
574 // Fixing this remaining case of 4292742 requires some flavor of
575 // escape analysis. Leave that for the future.
576 ciInstance* ex_obj = NULL;
577 switch (reason) {
578 case Deoptimization::Reason_null_check:
579 ex_obj = env()->NullPointerException_instance();
580 break;
581 case Deoptimization::Reason_div0_check:
582 ex_obj = env()->ArithmeticException_instance();
583 break;
584 case Deoptimization::Reason_range_check:
585 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
586 break;
587 case Deoptimization::Reason_class_check:
588 if (java_bc() == Bytecodes::_aastore) {
589 ex_obj = env()->ArrayStoreException_instance();
590 } else {
591 ex_obj = env()->ClassCastException_instance();
592 }
593 break;
594 default:
595 break;
596 }
597 if (failing()) { stop(); return; } // exception allocation might fail
598 if (ex_obj != NULL) {
599 // Cheat with a preallocated exception object.
600 if (C->log() != NULL)
601 C->log()->elem("hot_throw preallocated='1' reason='%s'",
602 Deoptimization::trap_reason_name(reason));
603 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
604 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
605
606 // Clear the detail message of the preallocated exception object.
607 // Weblogic sometimes mutates the detail message of exceptions
608 // using reflection.
609 int offset = java_lang_Throwable::get_detailMessage_offset();
610 const TypePtr* adr_typ = ex_con->add_offset(offset);
611
612 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
613 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
614 // Conservatively release stores of object references.
615 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
616
617 add_exception_state(make_exception_state(ex_node));
618 return;
619 }
620 }
621
622 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
623 // It won't be much cheaper than bailing to the interp., since we'll
624 // have to pass up all the debug-info, and the runtime will have to
625 // create the stack trace.
626
627 // Usual case: Bail to interpreter.
628 // Reserve the right to recompile if we haven't seen anything yet.
629
630 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
631 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
632 if (treat_throw_as_hot
633 && (method()->method_data()->trap_recompiled_at(bci(), m)
634 || C->too_many_traps(reason))) {
635 // We cannot afford to take more traps here. Suffer in the interpreter.
636 if (C->log() != NULL)
637 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
638 Deoptimization::trap_reason_name(reason),
639 C->trap_count(reason));
640 action = Deoptimization::Action_none;
641 }
642
643 // "must_throw" prunes the JVM state to include only the stack, if there
644 // are no local exception handlers. This should cut down on register
645 // allocation time and code size, by drastically reducing the number
646 // of in-edges on the call to the uncommon trap.
647
648 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
649 }
650
651
652 //----------------------------PreserveJVMState---------------------------------
653 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
654 debug_only(kit->verify_map());
655 _kit = kit;
656 _map = kit->map(); // preserve the map
657 _sp = kit->sp();
658 kit->set_map(clone_map ? kit->clone_map() : NULL);
659 #ifdef ASSERT
660 _bci = kit->bci();
661 Parse* parser = kit->is_Parse();
662 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
663 _block = block;
664 #endif
665 }
666 PreserveJVMState::~PreserveJVMState() {
667 GraphKit* kit = _kit;
668 #ifdef ASSERT
669 assert(kit->bci() == _bci, "bci must not shift");
670 Parse* parser = kit->is_Parse();
671 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
672 assert(block == _block, "block must not shift");
673 #endif
674 kit->set_map(_map);
675 kit->set_sp(_sp);
676 }
677
678
679 //-----------------------------BuildCutout-------------------------------------
680 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
681 : PreserveJVMState(kit)
682 {
683 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
684 SafePointNode* outer_map = _map; // preserved map is caller's
685 SafePointNode* inner_map = kit->map();
686 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
687 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
688 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
689 }
690 BuildCutout::~BuildCutout() {
691 GraphKit* kit = _kit;
692 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
693 }
694
695 //---------------------------PreserveReexecuteState----------------------------
696 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
697 assert(!kit->stopped(), "must call stopped() before");
698 _kit = kit;
699 _sp = kit->sp();
700 _reexecute = kit->jvms()->_reexecute;
701 }
702 PreserveReexecuteState::~PreserveReexecuteState() {
703 if (_kit->stopped()) return;
704 _kit->jvms()->_reexecute = _reexecute;
705 _kit->set_sp(_sp);
706 }
707
708 //------------------------------clone_map--------------------------------------
709 // Implementation of PreserveJVMState
710 //
711 // Only clone_map(...) here. If this function is only used in the
712 // PreserveJVMState class we may want to get rid of this extra
713 // function eventually and do it all there.
714
715 SafePointNode* GraphKit::clone_map() {
716 if (map() == NULL) return NULL;
717
718 // Clone the memory edge first
719 Node* mem = MergeMemNode::make(map()->memory());
720 gvn().set_type_bottom(mem);
721
722 SafePointNode *clonemap = (SafePointNode*)map()->clone();
723 JVMState* jvms = this->jvms();
724 JVMState* clonejvms = jvms->clone_shallow(C);
725 clonemap->set_memory(mem);
726 clonemap->set_jvms(clonejvms);
727 clonejvms->set_map(clonemap);
728 record_for_igvn(clonemap);
729 gvn().set_type_bottom(clonemap);
730 return clonemap;
731 }
732
733
734 //-----------------------------set_map_clone-----------------------------------
735 void GraphKit::set_map_clone(SafePointNode* m) {
736 _map = m;
737 _map = clone_map();
738 _map->set_next_exception(NULL);
739 debug_only(verify_map());
740 }
741
742
743 //----------------------------kill_dead_locals---------------------------------
744 // Detect any locals which are known to be dead, and force them to top.
745 void GraphKit::kill_dead_locals() {
746 // Consult the liveness information for the locals. If any
747 // of them are unused, then they can be replaced by top(). This
748 // should help register allocation time and cut down on the size
749 // of the deoptimization information.
750
751 // This call is made from many of the bytecode handling
752 // subroutines called from the Big Switch in do_one_bytecode.
753 // Every bytecode which might include a slow path is responsible
754 // for killing its dead locals. The more consistent we
755 // are about killing deads, the fewer useless phis will be
756 // constructed for them at various merge points.
757
758 // bci can be -1 (InvocationEntryBci). We return the entry
759 // liveness for the method.
760
761 if (method() == NULL || method()->code_size() == 0) {
762 // We are building a graph for a call to a native method.
763 // All locals are live.
764 return;
765 }
766
767 ResourceMark rm;
768
769 // Consult the liveness information for the locals. If any
770 // of them are unused, then they can be replaced by top(). This
771 // should help register allocation time and cut down on the size
772 // of the deoptimization information.
773 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
774
775 int len = (int)live_locals.size();
776 assert(len <= jvms()->loc_size(), "too many live locals");
777 for (int local = 0; local < len; local++) {
778 if (!live_locals.at(local)) {
779 set_local(local, top());
780 }
781 }
782 }
783
784 #ifdef ASSERT
785 //-------------------------dead_locals_are_killed------------------------------
786 // Return true if all dead locals are set to top in the map.
787 // Used to assert "clean" debug info at various points.
788 bool GraphKit::dead_locals_are_killed() {
789 if (method() == NULL || method()->code_size() == 0) {
790 // No locals need to be dead, so all is as it should be.
791 return true;
792 }
793
794 // Make sure somebody called kill_dead_locals upstream.
795 ResourceMark rm;
796 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
797 if (jvms->loc_size() == 0) continue; // no locals to consult
798 SafePointNode* map = jvms->map();
799 ciMethod* method = jvms->method();
800 int bci = jvms->bci();
801 if (jvms == this->jvms()) {
802 bci = this->bci(); // it might not yet be synched
803 }
804 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
805 int len = (int)live_locals.size();
806 if (!live_locals.is_valid() || len == 0)
807 // This method is trivial, or is poisoned by a breakpoint.
808 return true;
809 assert(len == jvms->loc_size(), "live map consistent with locals map");
810 for (int local = 0; local < len; local++) {
811 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
812 if (PrintMiscellaneous && (Verbose || WizardMode)) {
813 tty->print_cr("Zombie local %d: ", local);
814 jvms->dump();
815 }
816 return false;
817 }
818 }
819 }
820 return true;
821 }
822
823 #endif //ASSERT
824
825 // Helper function for enforcing certain bytecodes to reexecute if
826 // deoptimization happens
827 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
828 ciMethod* cur_method = jvms->method();
829 int cur_bci = jvms->bci();
830 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
831 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
832 return Interpreter::bytecode_should_reexecute(code) ||
833 (is_anewarray && code == Bytecodes::_multianewarray);
834 // Reexecute _multianewarray bytecode which was replaced with
835 // sequence of [a]newarray. See Parse::do_multianewarray().
836 //
837 // Note: interpreter should not have it set since this optimization
838 // is limited by dimensions and guarded by flag so in some cases
839 // multianewarray() runtime calls will be generated and
840 // the bytecode should not be reexecutes (stack will not be reset).
841 } else
842 return false;
843 }
844
845 // Helper function for adding JVMState and debug information to node
846 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
847 // Add the safepoint edges to the call (or other safepoint).
848
849 // Make sure dead locals are set to top. This
850 // should help register allocation time and cut down on the size
851 // of the deoptimization information.
852 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
853
854 // Walk the inline list to fill in the correct set of JVMState's
855 // Also fill in the associated edges for each JVMState.
856
857 // If the bytecode needs to be reexecuted we need to put
858 // the arguments back on the stack.
859 const bool should_reexecute = jvms()->should_reexecute();
860 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
861
862 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
863 // undefined if the bci is different. This is normal for Parse but it
864 // should not happen for LibraryCallKit because only one bci is processed.
865 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
866 "in LibraryCallKit the reexecute bit should not change");
867
868 // If we are guaranteed to throw, we can prune everything but the
869 // input to the current bytecode.
870 bool can_prune_locals = false;
871 uint stack_slots_not_pruned = 0;
872 int inputs = 0, depth = 0;
873 if (must_throw) {
874 assert(method() == youngest_jvms->method(), "sanity");
875 if (compute_stack_effects(inputs, depth)) {
876 can_prune_locals = true;
877 stack_slots_not_pruned = inputs;
878 }
879 }
880
881 if (env()->should_retain_local_variables()) {
882 // At any safepoint, this method can get breakpointed, which would
883 // then require an immediate deoptimization.
884 can_prune_locals = false; // do not prune locals
885 stack_slots_not_pruned = 0;
886 }
887
888 // do not scribble on the input jvms
889 JVMState* out_jvms = youngest_jvms->clone_deep(C);
890 call->set_jvms(out_jvms); // Start jvms list for call node
891
892 // For a known set of bytecodes, the interpreter should reexecute them if
893 // deoptimization happens. We set the reexecute state for them here
894 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
895 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
896 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
897 }
898
899 // Presize the call:
900 DEBUG_ONLY(uint non_debug_edges = call->req());
901 call->add_req_batch(top(), youngest_jvms->debug_depth());
902 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
903
904 // Set up edges so that the call looks like this:
905 // Call [state:] ctl io mem fptr retadr
906 // [parms:] parm0 ... parmN
907 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
908 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
909 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
910 // Note that caller debug info precedes callee debug info.
911
912 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
913 uint debug_ptr = call->req();
914
915 // Loop over the map input edges associated with jvms, add them
916 // to the call node, & reset all offsets to match call node array.
917 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
918 uint debug_end = debug_ptr;
919 uint debug_start = debug_ptr - in_jvms->debug_size();
920 debug_ptr = debug_start; // back up the ptr
921
922 uint p = debug_start; // walks forward in [debug_start, debug_end)
923 uint j, k, l;
924 SafePointNode* in_map = in_jvms->map();
925 out_jvms->set_map(call);
926
927 if (can_prune_locals) {
928 assert(in_jvms->method() == out_jvms->method(), "sanity");
929 // If the current throw can reach an exception handler in this JVMS,
930 // then we must keep everything live that can reach that handler.
931 // As a quick and dirty approximation, we look for any handlers at all.
932 if (in_jvms->method()->has_exception_handlers()) {
933 can_prune_locals = false;
934 }
935 }
936
937 // Add the Locals
938 k = in_jvms->locoff();
939 l = in_jvms->loc_size();
940 out_jvms->set_locoff(p);
941 if (!can_prune_locals) {
942 for (j = 0; j < l; j++)
943 call->set_req(p++, in_map->in(k+j));
944 } else {
945 p += l; // already set to top above by add_req_batch
946 }
947
948 // Add the Expression Stack
949 k = in_jvms->stkoff();
950 l = in_jvms->sp();
951 out_jvms->set_stkoff(p);
952 if (!can_prune_locals) {
953 for (j = 0; j < l; j++)
954 call->set_req(p++, in_map->in(k+j));
955 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
956 // Divide stack into {S0,...,S1}, where S0 is set to top.
957 uint s1 = stack_slots_not_pruned;
958 stack_slots_not_pruned = 0; // for next iteration
959 if (s1 > l) s1 = l;
960 uint s0 = l - s1;
961 p += s0; // skip the tops preinstalled by add_req_batch
962 for (j = s0; j < l; j++)
963 call->set_req(p++, in_map->in(k+j));
964 } else {
965 p += l; // already set to top above by add_req_batch
966 }
967
968 // Add the Monitors
969 k = in_jvms->monoff();
970 l = in_jvms->mon_size();
971 out_jvms->set_monoff(p);
972 for (j = 0; j < l; j++)
973 call->set_req(p++, in_map->in(k+j));
974
975 // Copy any scalar object fields.
976 k = in_jvms->scloff();
977 l = in_jvms->scl_size();
978 out_jvms->set_scloff(p);
979 for (j = 0; j < l; j++)
980 call->set_req(p++, in_map->in(k+j));
981
982 // Finish the new jvms.
983 out_jvms->set_endoff(p);
984
985 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
986 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
987 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
988 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
989 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
990 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
991
992 // Update the two tail pointers in parallel.
993 out_jvms = out_jvms->caller();
994 in_jvms = in_jvms->caller();
995 }
996
997 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
998
999 // Test the correctness of JVMState::debug_xxx accessors:
1000 assert(call->jvms()->debug_start() == non_debug_edges, "");
1001 assert(call->jvms()->debug_end() == call->req(), "");
1002 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1003 }
1004
1005 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1006 Bytecodes::Code code = java_bc();
1007 if (code == Bytecodes::_wide) {
1008 code = method()->java_code_at_bci(bci() + 1);
1009 }
1010
1011 BasicType rtype = T_ILLEGAL;
1012 int rsize = 0;
1013
1014 if (code != Bytecodes::_illegal) {
1015 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1016 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1017 if (rtype < T_CONFLICT)
1018 rsize = type2size[rtype];
1019 }
1020
1021 switch (code) {
1022 case Bytecodes::_illegal:
1023 return false;
1024
1025 case Bytecodes::_ldc:
1026 case Bytecodes::_ldc_w:
1027 case Bytecodes::_ldc2_w:
1028 inputs = 0;
1029 break;
1030
1031 case Bytecodes::_dup: inputs = 1; break;
1032 case Bytecodes::_dup_x1: inputs = 2; break;
1033 case Bytecodes::_dup_x2: inputs = 3; break;
1034 case Bytecodes::_dup2: inputs = 2; break;
1035 case Bytecodes::_dup2_x1: inputs = 3; break;
1036 case Bytecodes::_dup2_x2: inputs = 4; break;
1037 case Bytecodes::_swap: inputs = 2; break;
1038 case Bytecodes::_arraylength: inputs = 1; break;
1039
1040 case Bytecodes::_getstatic:
1041 case Bytecodes::_putstatic:
1042 case Bytecodes::_getfield:
1043 case Bytecodes::_putfield:
1044 {
1045 bool ignored_will_link;
1046 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1047 int size = field->type()->size();
1048 bool is_get = (depth >= 0), is_static = (depth & 1);
1049 inputs = (is_static ? 0 : 1);
1050 if (is_get) {
1051 depth = size - inputs;
1052 } else {
1053 inputs += size; // putxxx pops the value from the stack
1054 depth = - inputs;
1055 }
1056 }
1057 break;
1058
1059 case Bytecodes::_invokevirtual:
1060 case Bytecodes::_invokespecial:
1061 case Bytecodes::_invokestatic:
1062 case Bytecodes::_invokedynamic:
1063 case Bytecodes::_invokeinterface:
1064 {
1065 bool ignored_will_link;
1066 ciSignature* declared_signature = NULL;
1067 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1068 assert(declared_signature != NULL, "cannot be null");
1069 inputs = declared_signature->arg_size_for_bc(code);
1070 int size = declared_signature->return_type()->size();
1071 depth = size - inputs;
1072 }
1073 break;
1074
1075 case Bytecodes::_multianewarray:
1076 {
1077 ciBytecodeStream iter(method());
1078 iter.reset_to_bci(bci());
1079 iter.next();
1080 inputs = iter.get_dimensions();
1081 assert(rsize == 1, "");
1082 depth = rsize - inputs;
1083 }
1084 break;
1085
1086 case Bytecodes::_ireturn:
1087 case Bytecodes::_lreturn:
1088 case Bytecodes::_freturn:
1089 case Bytecodes::_dreturn:
1090 case Bytecodes::_areturn:
1091 assert(rsize == -depth, "");
1092 inputs = rsize;
1093 break;
1094
1095 case Bytecodes::_jsr:
1096 case Bytecodes::_jsr_w:
1097 inputs = 0;
1098 depth = 1; // S.B. depth=1, not zero
1099 break;
1100
1101 default:
1102 // bytecode produces a typed result
1103 inputs = rsize - depth;
1104 assert(inputs >= 0, "");
1105 break;
1106 }
1107
1108 #ifdef ASSERT
1109 // spot check
1110 int outputs = depth + inputs;
1111 assert(outputs >= 0, "sanity");
1112 switch (code) {
1113 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1114 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1115 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1116 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1117 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1118 default: break;
1119 }
1120 #endif //ASSERT
1121
1122 return true;
1123 }
1124
1125
1126
1127 //------------------------------basic_plus_adr---------------------------------
1128 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1129 // short-circuit a common case
1130 if (offset == intcon(0)) return ptr;
1131 return _gvn.transform( new AddPNode(base, ptr, offset) );
1132 }
1133
1134 Node* GraphKit::ConvI2L(Node* offset) {
1135 // short-circuit a common case
1136 jint offset_con = find_int_con(offset, Type::OffsetBot);
1137 if (offset_con != Type::OffsetBot) {
1138 return longcon((jlong) offset_con);
1139 }
1140 return _gvn.transform( new ConvI2LNode(offset));
1141 }
1142
1143 Node* GraphKit::ConvI2UL(Node* offset) {
1144 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1145 if (offset_con != (juint) Type::OffsetBot) {
1146 return longcon((julong) offset_con);
1147 }
1148 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1149 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1150 return _gvn.transform( new AndLNode(conv, mask) );
1151 }
1152
1153 Node* GraphKit::ConvL2I(Node* offset) {
1154 // short-circuit a common case
1155 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1156 if (offset_con != (jlong)Type::OffsetBot) {
1157 return intcon((int) offset_con);
1158 }
1159 return _gvn.transform( new ConvL2INode(offset));
1160 }
1161
1162 //-------------------------load_object_klass-----------------------------------
1163 Node* GraphKit::load_object_klass(Node* obj) {
1164 // Special-case a fresh allocation to avoid building nodes:
1165 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1166 if (akls != NULL) return akls;
1167 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1168 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1169 }
1170
1171 //-------------------------load_array_length-----------------------------------
1172 Node* GraphKit::load_array_length(Node* array) {
1173 // Special-case a fresh allocation to avoid building nodes:
1174 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1175 Node *alen;
1176 if (alloc == NULL) {
1177 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1178 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1179 } else {
1180 alen = alloc->Ideal_length();
1181 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1182 if (ccast != alen) {
1183 alen = _gvn.transform(ccast);
1184 }
1185 }
1186 return alen;
1187 }
1188
1189 //------------------------------do_null_check----------------------------------
1190 // Helper function to do a NULL pointer check. Returned value is
1191 // the incoming address with NULL casted away. You are allowed to use the
1192 // not-null value only if you are control dependent on the test.
1193 #ifndef PRODUCT
1194 extern int explicit_null_checks_inserted,
1195 explicit_null_checks_elided;
1196 #endif
1197 Node* GraphKit::null_check_common(Node* value, BasicType type,
1198 // optional arguments for variations:
1199 bool assert_null,
1200 Node* *null_control,
1201 bool speculative) {
1202 assert(!assert_null || null_control == NULL, "not both at once");
1203 if (stopped()) return top();
1204 NOT_PRODUCT(explicit_null_checks_inserted++);
1205
1206 // Construct NULL check
1207 Node *chk = NULL;
1208 switch(type) {
1209 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1210 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1211 case T_ARRAY : // fall through
1212 type = T_OBJECT; // simplify further tests
1213 case T_OBJECT : {
1214 const Type *t = _gvn.type( value );
1215
1216 const TypeOopPtr* tp = t->isa_oopptr();
1217 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1218 // Only for do_null_check, not any of its siblings:
1219 && !assert_null && null_control == NULL) {
1220 // Usually, any field access or invocation on an unloaded oop type
1221 // will simply fail to link, since the statically linked class is
1222 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1223 // the static class is loaded but the sharper oop type is not.
1224 // Rather than checking for this obscure case in lots of places,
1225 // we simply observe that a null check on an unloaded class
1226 // will always be followed by a nonsense operation, so we
1227 // can just issue the uncommon trap here.
1228 // Our access to the unloaded class will only be correct
1229 // after it has been loaded and initialized, which requires
1230 // a trip through the interpreter.
1231 #ifndef PRODUCT
1232 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1233 #endif
1234 uncommon_trap(Deoptimization::Reason_unloaded,
1235 Deoptimization::Action_reinterpret,
1236 tp->klass(), "!loaded");
1237 return top();
1238 }
1239
1240 if (assert_null) {
1241 // See if the type is contained in NULL_PTR.
1242 // If so, then the value is already null.
1243 if (t->higher_equal(TypePtr::NULL_PTR)) {
1244 NOT_PRODUCT(explicit_null_checks_elided++);
1245 return value; // Elided null assert quickly!
1246 }
1247 } else {
1248 // See if mixing in the NULL pointer changes type.
1249 // If so, then the NULL pointer was not allowed in the original
1250 // type. In other words, "value" was not-null.
1251 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1252 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1253 NOT_PRODUCT(explicit_null_checks_elided++);
1254 return value; // Elided null check quickly!
1255 }
1256 }
1257 chk = new CmpPNode( value, null() );
1258 break;
1259 }
1260
1261 default:
1262 fatal("unexpected type: %s", type2name(type));
1263 }
1264 assert(chk != NULL, "sanity check");
1265 chk = _gvn.transform(chk);
1266
1267 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1268 BoolNode *btst = new BoolNode( chk, btest);
1269 Node *tst = _gvn.transform( btst );
1270
1271 //-----------
1272 // if peephole optimizations occurred, a prior test existed.
1273 // If a prior test existed, maybe it dominates as we can avoid this test.
1274 if (tst != btst && type == T_OBJECT) {
1275 // At this point we want to scan up the CFG to see if we can
1276 // find an identical test (and so avoid this test altogether).
1277 Node *cfg = control();
1278 int depth = 0;
1279 while( depth < 16 ) { // Limit search depth for speed
1280 if( cfg->Opcode() == Op_IfTrue &&
1281 cfg->in(0)->in(1) == tst ) {
1282 // Found prior test. Use "cast_not_null" to construct an identical
1283 // CastPP (and hence hash to) as already exists for the prior test.
1284 // Return that casted value.
1285 if (assert_null) {
1286 replace_in_map(value, null());
1287 return null(); // do not issue the redundant test
1288 }
1289 Node *oldcontrol = control();
1290 set_control(cfg);
1291 Node *res = cast_not_null(value);
1292 set_control(oldcontrol);
1293 NOT_PRODUCT(explicit_null_checks_elided++);
1294 return res;
1295 }
1296 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1297 if (cfg == NULL) break; // Quit at region nodes
1298 depth++;
1299 }
1300 }
1301
1302 //-----------
1303 // Branch to failure if null
1304 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1305 Deoptimization::DeoptReason reason;
1306 if (assert_null) {
1307 reason = Deoptimization::reason_null_assert(speculative);
1308 } else if (type == T_OBJECT) {
1309 reason = Deoptimization::reason_null_check(speculative);
1310 } else {
1311 reason = Deoptimization::Reason_div0_check;
1312 }
1313 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1314 // ciMethodData::has_trap_at will return a conservative -1 if any
1315 // must-be-null assertion has failed. This could cause performance
1316 // problems for a method after its first do_null_assert failure.
1317 // Consider using 'Reason_class_check' instead?
1318
1319 // To cause an implicit null check, we set the not-null probability
1320 // to the maximum (PROB_MAX). For an explicit check the probability
1321 // is set to a smaller value.
1322 if (null_control != NULL || too_many_traps(reason)) {
1323 // probability is less likely
1324 ok_prob = PROB_LIKELY_MAG(3);
1325 } else if (!assert_null &&
1326 (ImplicitNullCheckThreshold > 0) &&
1327 method() != NULL &&
1328 (method()->method_data()->trap_count(reason)
1329 >= (uint)ImplicitNullCheckThreshold)) {
1330 ok_prob = PROB_LIKELY_MAG(3);
1331 }
1332
1333 if (null_control != NULL) {
1334 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1335 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1336 set_control( _gvn.transform( new IfTrueNode(iff)));
1337 #ifndef PRODUCT
1338 if (null_true == top()) {
1339 explicit_null_checks_elided++;
1340 }
1341 #endif
1342 (*null_control) = null_true;
1343 } else {
1344 BuildCutout unless(this, tst, ok_prob);
1345 // Check for optimizer eliding test at parse time
1346 if (stopped()) {
1347 // Failure not possible; do not bother making uncommon trap.
1348 NOT_PRODUCT(explicit_null_checks_elided++);
1349 } else if (assert_null) {
1350 uncommon_trap(reason,
1351 Deoptimization::Action_make_not_entrant,
1352 NULL, "assert_null");
1353 } else {
1354 replace_in_map(value, zerocon(type));
1355 builtin_throw(reason);
1356 }
1357 }
1358
1359 // Must throw exception, fall-thru not possible?
1360 if (stopped()) {
1361 return top(); // No result
1362 }
1363
1364 if (assert_null) {
1365 // Cast obj to null on this path.
1366 replace_in_map(value, zerocon(type));
1367 return zerocon(type);
1368 }
1369
1370 // Cast obj to not-null on this path, if there is no null_control.
1371 // (If there is a null_control, a non-null value may come back to haunt us.)
1372 if (type == T_OBJECT) {
1373 Node* cast = cast_not_null(value, false);
1374 if (null_control == NULL || (*null_control) == top())
1375 replace_in_map(value, cast);
1376 value = cast;
1377 }
1378
1379 return value;
1380 }
1381
1382
1383 //------------------------------cast_not_null----------------------------------
1384 // Cast obj to not-null on this path
1385 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1386 const Type *t = _gvn.type(obj);
1387 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1388 // Object is already not-null?
1389 if( t == t_not_null ) return obj;
1390
1391 Node *cast = new CastPPNode(obj,t_not_null);
1392 cast->init_req(0, control());
1393 cast = _gvn.transform( cast );
1394
1395 // Scan for instances of 'obj' in the current JVM mapping.
1396 // These instances are known to be not-null after the test.
1397 if (do_replace_in_map)
1398 replace_in_map(obj, cast);
1399
1400 return cast; // Return casted value
1401 }
1402
1403 // Sometimes in intrinsics, we implicitly know an object is not null
1404 // (there's no actual null check) so we can cast it to not null. In
1405 // the course of optimizations, the input to the cast can become null.
1406 // In that case that data path will die and we need the control path
1407 // to become dead as well to keep the graph consistent. So we have to
1408 // add a check for null for which one branch can't be taken. It uses
1409 // an Opaque4 node that will cause the check to be removed after loop
1410 // opts so the test goes away and the compiled code doesn't execute a
1411 // useless check.
1412 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1413 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1414 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1415 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1416 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1417 _gvn.set_type(iff, iff->Value(&_gvn));
1418 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1419 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1420 Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1421 C->root()->add_req(halt);
1422 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1423 set_control(if_t);
1424 return cast_not_null(value, do_replace_in_map);
1425 }
1426
1427
1428 //--------------------------replace_in_map-------------------------------------
1429 void GraphKit::replace_in_map(Node* old, Node* neww) {
1430 if (old == neww) {
1431 return;
1432 }
1433
1434 map()->replace_edge(old, neww);
1435
1436 // Note: This operation potentially replaces any edge
1437 // on the map. This includes locals, stack, and monitors
1438 // of the current (innermost) JVM state.
1439
1440 // don't let inconsistent types from profiling escape this
1441 // method
1442
1443 const Type* told = _gvn.type(old);
1444 const Type* tnew = _gvn.type(neww);
1445
1446 if (!tnew->higher_equal(told)) {
1447 return;
1448 }
1449
1450 map()->record_replaced_node(old, neww);
1451 }
1452
1453
1454 //=============================================================================
1455 //--------------------------------memory---------------------------------------
1456 Node* GraphKit::memory(uint alias_idx) {
1457 MergeMemNode* mem = merged_memory();
1458 Node* p = mem->memory_at(alias_idx);
1459 _gvn.set_type(p, Type::MEMORY); // must be mapped
1460 return p;
1461 }
1462
1463 //-----------------------------reset_memory------------------------------------
1464 Node* GraphKit::reset_memory() {
1465 Node* mem = map()->memory();
1466 // do not use this node for any more parsing!
1467 debug_only( map()->set_memory((Node*)NULL) );
1468 return _gvn.transform( mem );
1469 }
1470
1471 //------------------------------set_all_memory---------------------------------
1472 void GraphKit::set_all_memory(Node* newmem) {
1473 Node* mergemem = MergeMemNode::make(newmem);
1474 gvn().set_type_bottom(mergemem);
1475 map()->set_memory(mergemem);
1476 }
1477
1478 //------------------------------set_all_memory_call----------------------------
1479 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1480 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1481 set_all_memory(newmem);
1482 }
1483
1484 //=============================================================================
1485 //
1486 // parser factory methods for MemNodes
1487 //
1488 // These are layered on top of the factory methods in LoadNode and StoreNode,
1489 // and integrate with the parser's memory state and _gvn engine.
1490 //
1491
1492 // factory methods in "int adr_idx"
1493 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1494 int adr_idx,
1495 MemNode::MemOrd mo,
1496 LoadNode::ControlDependency control_dependency,
1497 bool require_atomic_access,
1498 bool unaligned,
1499 bool mismatched) {
1500 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1501 const TypePtr* adr_type = NULL; // debug-mode-only argument
1502 debug_only(adr_type = C->get_adr_type(adr_idx));
1503 Node* mem = memory(adr_idx);
1504 Node* ld;
1505 if (require_atomic_access && bt == T_LONG) {
1506 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1507 } else if (require_atomic_access && bt == T_DOUBLE) {
1508 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1509 } else {
1510 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1511 }
1512 ld = _gvn.transform(ld);
1513 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1514 // Improve graph before escape analysis and boxing elimination.
1515 record_for_igvn(ld);
1516 }
1517 return ld;
1518 }
1519
1520 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1521 int adr_idx,
1522 MemNode::MemOrd mo,
1523 bool require_atomic_access,
1524 bool unaligned,
1525 bool mismatched) {
1526 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1527 const TypePtr* adr_type = NULL;
1528 debug_only(adr_type = C->get_adr_type(adr_idx));
1529 Node *mem = memory(adr_idx);
1530 Node* st;
1531 if (require_atomic_access && bt == T_LONG) {
1532 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1533 } else if (require_atomic_access && bt == T_DOUBLE) {
1534 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1535 } else {
1536 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1537 }
1538 if (unaligned) {
1539 st->as_Store()->set_unaligned_access();
1540 }
1541 if (mismatched) {
1542 st->as_Store()->set_mismatched_access();
1543 }
1544 st = _gvn.transform(st);
1545 set_memory(st, adr_idx);
1546 // Back-to-back stores can only remove intermediate store with DU info
1547 // so push on worklist for optimizer.
1548 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1549 record_for_igvn(st);
1550
1551 return st;
1552 }
1553
1554
1555 void GraphKit::pre_barrier(bool do_load,
1556 Node* ctl,
1557 Node* obj,
1558 Node* adr,
1559 uint adr_idx,
1560 Node* val,
1561 const TypeOopPtr* val_type,
1562 Node* pre_val,
1563 BasicType bt) {
1564
1565 BarrierSet* bs = BarrierSet::barrier_set();
1566 set_control(ctl);
1567 switch (bs->kind()) {
1568 case BarrierSet::G1BarrierSet:
1569 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1570 break;
1571
1572 case BarrierSet::CardTableBarrierSet:
1573 case BarrierSet::Epsilon:
1574 break;
1575
1576 default :
1577 ShouldNotReachHere();
1578
1579 }
1580 }
1581
1582 bool GraphKit::can_move_pre_barrier() const {
1583 BarrierSet* bs = BarrierSet::barrier_set();
1584 switch (bs->kind()) {
1585 case BarrierSet::G1BarrierSet:
1586 return true; // Can move it if no safepoint
1587
1588 case BarrierSet::CardTableBarrierSet:
1589 case BarrierSet::Epsilon:
1590 return true; // There is no pre-barrier
1591
1592 default :
1593 ShouldNotReachHere();
1594 }
1595 return false;
1596 }
1597
1598 void GraphKit::post_barrier(Node* ctl,
1599 Node* store,
1600 Node* obj,
1601 Node* adr,
1602 uint adr_idx,
1603 Node* val,
1604 BasicType bt,
1605 bool use_precise) {
1606 BarrierSet* bs = BarrierSet::barrier_set();
1607 set_control(ctl);
1608 switch (bs->kind()) {
1609 case BarrierSet::G1BarrierSet:
1610 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1611 break;
1612
1613 case BarrierSet::CardTableBarrierSet:
1614 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1615 break;
1616
1617 case BarrierSet::Epsilon:
1618 break;
1619
1620 default :
1621 ShouldNotReachHere();
1622
1623 }
1624 }
1625
1626 Node* GraphKit::store_oop(Node* ctl,
1627 Node* obj,
1628 Node* adr,
1629 const TypePtr* adr_type,
1630 Node* val,
1631 const TypeOopPtr* val_type,
1632 BasicType bt,
1633 bool use_precise,
1634 MemNode::MemOrd mo,
1635 bool mismatched) {
1636 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1637 // could be delayed during Parse (for example, in adjust_map_after_if()).
1638 // Execute transformation here to avoid barrier generation in such case.
1639 if (_gvn.type(val) == TypePtr::NULL_PTR)
1640 val = _gvn.makecon(TypePtr::NULL_PTR);
1641
1642 set_control(ctl);
1643 if (stopped()) return top(); // Dead path ?
1644
1645 assert(bt == T_OBJECT, "sanity");
1646 assert(val != NULL, "not dead path");
1647 uint adr_idx = C->get_alias_index(adr_type);
1648 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1649
1650 pre_barrier(true /* do_load */,
1651 control(), obj, adr, adr_idx, val, val_type,
1652 NULL /* pre_val */,
1653 bt);
1654
1655 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1656 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1657 return store;
1658 }
1659
1660 // Could be an array or object we don't know at compile time (unsafe ref.)
1661 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1662 Node* obj, // containing obj
1663 Node* adr, // actual adress to store val at
1664 const TypePtr* adr_type,
1665 Node* val,
1666 BasicType bt,
1667 MemNode::MemOrd mo,
1668 bool mismatched) {
1669 Compile::AliasType* at = C->alias_type(adr_type);
1670 const TypeOopPtr* val_type = NULL;
1671 if (adr_type->isa_instptr()) {
1672 if (at->field() != NULL) {
1673 // known field. This code is a copy of the do_put_xxx logic.
1674 ciField* field = at->field();
1675 if (!field->type()->is_loaded()) {
1676 val_type = TypeInstPtr::BOTTOM;
1677 } else {
1678 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1679 }
1680 }
1681 } else if (adr_type->isa_aryptr()) {
1682 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1683 }
1684 if (val_type == NULL) {
1685 val_type = TypeInstPtr::BOTTOM;
1686 }
1687 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1688 }
1689
1690
1691 //-------------------------array_element_address-------------------------
1692 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1693 const TypeInt* sizetype, Node* ctrl) {
1694 uint shift = exact_log2(type2aelembytes(elembt));
1695 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1696
1697 // short-circuit a common case (saves lots of confusing waste motion)
1698 jint idx_con = find_int_con(idx, -1);
1699 if (idx_con >= 0) {
1700 intptr_t offset = header + ((intptr_t)idx_con << shift);
1701 return basic_plus_adr(ary, offset);
1702 }
1703
1704 // must be correct type for alignment purposes
1705 Node* base = basic_plus_adr(ary, header);
1706 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1707 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1708 return basic_plus_adr(ary, base, scale);
1709 }
1710
1711 //-------------------------load_array_element-------------------------
1712 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1713 const Type* elemtype = arytype->elem();
1714 BasicType elembt = elemtype->array_element_basic_type();
1715 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1716 if (elembt == T_NARROWOOP) {
1717 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1718 }
1719 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1720 return ld;
1721 }
1722
1723 //-------------------------set_arguments_for_java_call-------------------------
1724 // Arguments (pre-popped from the stack) are taken from the JVMS.
1725 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1726 // Add the call arguments:
1727 uint nargs = call->method()->arg_size();
1728 for (uint i = 0; i < nargs; i++) {
1729 Node* arg = argument(i);
1730 call->init_req(i + TypeFunc::Parms, arg);
1731 }
1732 }
1733
1734 //---------------------------set_edges_for_java_call---------------------------
1735 // Connect a newly created call into the current JVMS.
1736 // A return value node (if any) is returned from set_edges_for_java_call.
1737 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1738
1739 // Add the predefined inputs:
1740 call->init_req( TypeFunc::Control, control() );
1741 call->init_req( TypeFunc::I_O , i_o() );
1742 call->init_req( TypeFunc::Memory , reset_memory() );
1743 call->init_req( TypeFunc::FramePtr, frameptr() );
1744 call->init_req( TypeFunc::ReturnAdr, top() );
1745
1746 add_safepoint_edges(call, must_throw);
1747
1748 Node* xcall = _gvn.transform(call);
1749
1750 if (xcall == top()) {
1751 set_control(top());
1752 return;
1753 }
1754 assert(xcall == call, "call identity is stable");
1755
1756 // Re-use the current map to produce the result.
1757
1758 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1759 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1760 set_all_memory_call(xcall, separate_io_proj);
1761
1762 //return xcall; // no need, caller already has it
1763 }
1764
1765 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1766 if (stopped()) return top(); // maybe the call folded up?
1767
1768 // Capture the return value, if any.
1769 Node* ret;
1770 if (call->method() == NULL ||
1771 call->method()->return_type()->basic_type() == T_VOID)
1772 ret = top();
1773 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1774
1775 // Note: Since any out-of-line call can produce an exception,
1776 // we always insert an I_O projection from the call into the result.
1777
1778 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1779
1780 if (separate_io_proj) {
1781 // The caller requested separate projections be used by the fall
1782 // through and exceptional paths, so replace the projections for
1783 // the fall through path.
1784 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1785 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1786 }
1787 return ret;
1788 }
1789
1790 //--------------------set_predefined_input_for_runtime_call--------------------
1791 // Reading and setting the memory state is way conservative here.
1792 // The real problem is that I am not doing real Type analysis on memory,
1793 // so I cannot distinguish card mark stores from other stores. Across a GC
1794 // point the Store Barrier and the card mark memory has to agree. I cannot
1795 // have a card mark store and its barrier split across the GC point from
1796 // either above or below. Here I get that to happen by reading ALL of memory.
1797 // A better answer would be to separate out card marks from other memory.
1798 // For now, return the input memory state, so that it can be reused
1799 // after the call, if this call has restricted memory effects.
1800 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1801 // Set fixed predefined input arguments
1802 Node* memory = reset_memory();
1803 call->init_req( TypeFunc::Control, control() );
1804 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1805 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1806 call->init_req( TypeFunc::FramePtr, frameptr() );
1807 call->init_req( TypeFunc::ReturnAdr, top() );
1808 return memory;
1809 }
1810
1811 //-------------------set_predefined_output_for_runtime_call--------------------
1812 // Set control and memory (not i_o) from the call.
1813 // If keep_mem is not NULL, use it for the output state,
1814 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1815 // If hook_mem is NULL, this call produces no memory effects at all.
1816 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1817 // then only that memory slice is taken from the call.
1818 // In the last case, we must put an appropriate memory barrier before
1819 // the call, so as to create the correct anti-dependencies on loads
1820 // preceding the call.
1821 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1822 Node* keep_mem,
1823 const TypePtr* hook_mem) {
1824 // no i/o
1825 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1826 if (keep_mem) {
1827 // First clone the existing memory state
1828 set_all_memory(keep_mem);
1829 if (hook_mem != NULL) {
1830 // Make memory for the call
1831 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1832 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1833 // We also use hook_mem to extract specific effects from arraycopy stubs.
1834 set_memory(mem, hook_mem);
1835 }
1836 // ...else the call has NO memory effects.
1837
1838 // Make sure the call advertises its memory effects precisely.
1839 // This lets us build accurate anti-dependences in gcm.cpp.
1840 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1841 "call node must be constructed correctly");
1842 } else {
1843 assert(hook_mem == NULL, "");
1844 // This is not a "slow path" call; all memory comes from the call.
1845 set_all_memory_call(call);
1846 }
1847 }
1848
1849
1850 // Replace the call with the current state of the kit.
1851 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1852 JVMState* ejvms = NULL;
1853 if (has_exceptions()) {
1854 ejvms = transfer_exceptions_into_jvms();
1855 }
1856
1857 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1858 ReplacedNodes replaced_nodes_exception;
1859 Node* ex_ctl = top();
1860
1861 SafePointNode* final_state = stop();
1862
1863 // Find all the needed outputs of this call
1864 CallProjections callprojs;
1865 call->extract_projections(&callprojs, true);
1866
1867 Node* init_mem = call->in(TypeFunc::Memory);
1868 Node* final_mem = final_state->in(TypeFunc::Memory);
1869 Node* final_ctl = final_state->in(TypeFunc::Control);
1870 Node* final_io = final_state->in(TypeFunc::I_O);
1871
1872 // Replace all the old call edges with the edges from the inlining result
1873 if (callprojs.fallthrough_catchproj != NULL) {
1874 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1875 }
1876 if (callprojs.fallthrough_memproj != NULL) {
1877 if (final_mem->is_MergeMem()) {
1878 // Parser's exits MergeMem was not transformed but may be optimized
1879 final_mem = _gvn.transform(final_mem);
1880 }
1881 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1882 }
1883 if (callprojs.fallthrough_ioproj != NULL) {
1884 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1885 }
1886
1887 // Replace the result with the new result if it exists and is used
1888 if (callprojs.resproj != NULL && result != NULL) {
1889 C->gvn_replace_by(callprojs.resproj, result);
1890 }
1891
1892 if (ejvms == NULL) {
1893 // No exception edges to simply kill off those paths
1894 if (callprojs.catchall_catchproj != NULL) {
1895 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1896 }
1897 if (callprojs.catchall_memproj != NULL) {
1898 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1899 }
1900 if (callprojs.catchall_ioproj != NULL) {
1901 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1902 }
1903 // Replace the old exception object with top
1904 if (callprojs.exobj != NULL) {
1905 C->gvn_replace_by(callprojs.exobj, C->top());
1906 }
1907 } else {
1908 GraphKit ekit(ejvms);
1909
1910 // Load my combined exception state into the kit, with all phis transformed:
1911 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1912 replaced_nodes_exception = ex_map->replaced_nodes();
1913
1914 Node* ex_oop = ekit.use_exception_state(ex_map);
1915
1916 if (callprojs.catchall_catchproj != NULL) {
1917 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1918 ex_ctl = ekit.control();
1919 }
1920 if (callprojs.catchall_memproj != NULL) {
1921 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1922 }
1923 if (callprojs.catchall_ioproj != NULL) {
1924 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1925 }
1926
1927 // Replace the old exception object with the newly created one
1928 if (callprojs.exobj != NULL) {
1929 C->gvn_replace_by(callprojs.exobj, ex_oop);
1930 }
1931 }
1932
1933 // Disconnect the call from the graph
1934 call->disconnect_inputs(NULL, C);
1935 C->gvn_replace_by(call, C->top());
1936
1937 // Clean up any MergeMems that feed other MergeMems since the
1938 // optimizer doesn't like that.
1939 if (final_mem->is_MergeMem()) {
1940 Node_List wl;
1941 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1942 Node* m = i.get();
1943 if (m->is_MergeMem() && !wl.contains(m)) {
1944 wl.push(m);
1945 }
1946 }
1947 while (wl.size() > 0) {
1948 _gvn.transform(wl.pop());
1949 }
1950 }
1951
1952 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1953 replaced_nodes.apply(C, final_ctl);
1954 }
1955 if (!ex_ctl->is_top() && do_replaced_nodes) {
1956 replaced_nodes_exception.apply(C, ex_ctl);
1957 }
1958 }
1959
1960
1961 //------------------------------increment_counter------------------------------
1962 // for statistics: increment a VM counter by 1
1963
1964 void GraphKit::increment_counter(address counter_addr) {
1965 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1966 increment_counter(adr1);
1967 }
1968
1969 void GraphKit::increment_counter(Node* counter_addr) {
1970 int adr_type = Compile::AliasIdxRaw;
1971 Node* ctrl = control();
1972 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1973 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1974 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1975 }
1976
1977
1978 //------------------------------uncommon_trap----------------------------------
1979 // Bail out to the interpreter in mid-method. Implemented by calling the
1980 // uncommon_trap blob. This helper function inserts a runtime call with the
1981 // right debug info.
1982 void GraphKit::uncommon_trap(int trap_request,
1983 ciKlass* klass, const char* comment,
1984 bool must_throw,
1985 bool keep_exact_action) {
1986 if (failing()) stop();
1987 if (stopped()) return; // trap reachable?
1988
1989 // Note: If ProfileTraps is true, and if a deopt. actually
1990 // occurs here, the runtime will make sure an MDO exists. There is
1991 // no need to call method()->ensure_method_data() at this point.
1992
1993 // Set the stack pointer to the right value for reexecution:
1994 set_sp(reexecute_sp());
1995
1996 #ifdef ASSERT
1997 if (!must_throw) {
1998 // Make sure the stack has at least enough depth to execute
1999 // the current bytecode.
2000 int inputs, ignored_depth;
2001 if (compute_stack_effects(inputs, ignored_depth)) {
2002 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2003 Bytecodes::name(java_bc()), sp(), inputs);
2004 }
2005 }
2006 #endif
2007
2008 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2009 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2010
2011 switch (action) {
2012 case Deoptimization::Action_maybe_recompile:
2013 case Deoptimization::Action_reinterpret:
2014 // Temporary fix for 6529811 to allow virtual calls to be sure they
2015 // get the chance to go from mono->bi->mega
2016 if (!keep_exact_action &&
2017 Deoptimization::trap_request_index(trap_request) < 0 &&
2018 too_many_recompiles(reason)) {
2019 // This BCI is causing too many recompilations.
2020 if (C->log() != NULL) {
2021 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2022 Deoptimization::trap_reason_name(reason),
2023 Deoptimization::trap_action_name(action));
2024 }
2025 action = Deoptimization::Action_none;
2026 trap_request = Deoptimization::make_trap_request(reason, action);
2027 } else {
2028 C->set_trap_can_recompile(true);
2029 }
2030 break;
2031 case Deoptimization::Action_make_not_entrant:
2032 C->set_trap_can_recompile(true);
2033 break;
2034 case Deoptimization::Action_none:
2035 case Deoptimization::Action_make_not_compilable:
2036 break;
2037 default:
2038 #ifdef ASSERT
2039 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2040 #endif
2041 break;
2042 }
2043
2044 if (TraceOptoParse) {
2045 char buf[100];
2046 tty->print_cr("Uncommon trap %s at bci:%d",
2047 Deoptimization::format_trap_request(buf, sizeof(buf),
2048 trap_request), bci());
2049 }
2050
2051 CompileLog* log = C->log();
2052 if (log != NULL) {
2053 int kid = (klass == NULL)? -1: log->identify(klass);
2054 log->begin_elem("uncommon_trap bci='%d'", bci());
2055 char buf[100];
2056 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2057 trap_request));
2058 if (kid >= 0) log->print(" klass='%d'", kid);
2059 if (comment != NULL) log->print(" comment='%s'", comment);
2060 log->end_elem();
2061 }
2062
2063 // Make sure any guarding test views this path as very unlikely
2064 Node *i0 = control()->in(0);
2065 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2066 IfNode *iff = i0->as_If();
2067 float f = iff->_prob; // Get prob
2068 if (control()->Opcode() == Op_IfTrue) {
2069 if (f > PROB_UNLIKELY_MAG(4))
2070 iff->_prob = PROB_MIN;
2071 } else {
2072 if (f < PROB_LIKELY_MAG(4))
2073 iff->_prob = PROB_MAX;
2074 }
2075 }
2076
2077 // Clear out dead values from the debug info.
2078 kill_dead_locals();
2079
2080 // Now insert the uncommon trap subroutine call
2081 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2082 const TypePtr* no_memory_effects = NULL;
2083 // Pass the index of the class to be loaded
2084 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2085 (must_throw ? RC_MUST_THROW : 0),
2086 OptoRuntime::uncommon_trap_Type(),
2087 call_addr, "uncommon_trap", no_memory_effects,
2088 intcon(trap_request));
2089 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2090 "must extract request correctly from the graph");
2091 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2092
2093 call->set_req(TypeFunc::ReturnAdr, returnadr());
2094 // The debug info is the only real input to this call.
2095
2096 // Halt-and-catch fire here. The above call should never return!
2097 HaltNode* halt = new HaltNode(control(), frameptr());
2098 _gvn.set_type_bottom(halt);
2099 root()->add_req(halt);
2100
2101 stop_and_kill_map();
2102 }
2103
2104
2105 //--------------------------just_allocated_object------------------------------
2106 // Report the object that was just allocated.
2107 // It must be the case that there are no intervening safepoints.
2108 // We use this to determine if an object is so "fresh" that
2109 // it does not require card marks.
2110 Node* GraphKit::just_allocated_object(Node* current_control) {
2111 if (C->recent_alloc_ctl() == current_control)
2112 return C->recent_alloc_obj();
2113 return NULL;
2114 }
2115
2116
2117 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2118 // (Note: TypeFunc::make has a cache that makes this fast.)
2119 const TypeFunc* tf = TypeFunc::make(dest_method);
2120 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2121 for (int j = 0; j < nargs; j++) {
2122 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2123 if( targ->basic_type() == T_DOUBLE ) {
2124 // If any parameters are doubles, they must be rounded before
2125 // the call, dstore_rounding does gvn.transform
2126 Node *arg = argument(j);
2127 arg = dstore_rounding(arg);
2128 set_argument(j, arg);
2129 }
2130 }
2131 }
2132
2133 /**
2134 * Record profiling data exact_kls for Node n with the type system so
2135 * that it can propagate it (speculation)
2136 *
2137 * @param n node that the type applies to
2138 * @param exact_kls type from profiling
2139 * @param maybe_null did profiling see null?
2140 *
2141 * @return node with improved type
2142 */
2143 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2144 const Type* current_type = _gvn.type(n);
2145 assert(UseTypeSpeculation, "type speculation must be on");
2146
2147 const TypePtr* speculative = current_type->speculative();
2148
2149 // Should the klass from the profile be recorded in the speculative type?
2150 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2151 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2152 const TypeOopPtr* xtype = tklass->as_instance_type();
2153 assert(xtype->klass_is_exact(), "Should be exact");
2154 // Any reason to believe n is not null (from this profiling or a previous one)?
2155 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2156 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2157 // record the new speculative type's depth
2158 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2159 speculative = speculative->with_inline_depth(jvms()->depth());
2160 } else if (current_type->would_improve_ptr(ptr_kind)) {
2161 // Profiling report that null was never seen so we can change the
2162 // speculative type to non null ptr.
2163 if (ptr_kind == ProfileAlwaysNull) {
2164 speculative = TypePtr::NULL_PTR;
2165 } else {
2166 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2167 const TypePtr* ptr = TypePtr::NOTNULL;
2168 if (speculative != NULL) {
2169 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2170 } else {
2171 speculative = ptr;
2172 }
2173 }
2174 }
2175
2176 if (speculative != current_type->speculative()) {
2177 // Build a type with a speculative type (what we think we know
2178 // about the type but will need a guard when we use it)
2179 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2180 // We're changing the type, we need a new CheckCast node to carry
2181 // the new type. The new type depends on the control: what
2182 // profiling tells us is only valid from here as far as we can
2183 // tell.
2184 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2185 cast = _gvn.transform(cast);
2186 replace_in_map(n, cast);
2187 n = cast;
2188 }
2189
2190 return n;
2191 }
2192
2193 /**
2194 * Record profiling data from receiver profiling at an invoke with the
2195 * type system so that it can propagate it (speculation)
2196 *
2197 * @param n receiver node
2198 *
2199 * @return node with improved type
2200 */
2201 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2202 if (!UseTypeSpeculation) {
2203 return n;
2204 }
2205 ciKlass* exact_kls = profile_has_unique_klass();
2206 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2207 if ((java_bc() == Bytecodes::_checkcast ||
2208 java_bc() == Bytecodes::_instanceof ||
2209 java_bc() == Bytecodes::_aastore) &&
2210 method()->method_data()->is_mature()) {
2211 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2212 if (data != NULL) {
2213 if (!data->as_BitData()->null_seen()) {
2214 ptr_kind = ProfileNeverNull;
2215 } else {
2216 assert(data->is_ReceiverTypeData(), "bad profile data type");
2217 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2218 uint i = 0;
2219 for (; i < call->row_limit(); i++) {
2220 ciKlass* receiver = call->receiver(i);
2221 if (receiver != NULL) {
2222 break;
2223 }
2224 }
2225 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2226 }
2227 }
2228 }
2229 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2230 }
2231
2232 /**
2233 * Record profiling data from argument profiling at an invoke with the
2234 * type system so that it can propagate it (speculation)
2235 *
2236 * @param dest_method target method for the call
2237 * @param bc what invoke bytecode is this?
2238 */
2239 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2240 if (!UseTypeSpeculation) {
2241 return;
2242 }
2243 const TypeFunc* tf = TypeFunc::make(dest_method);
2244 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2245 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2246 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2247 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2248 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2249 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2250 ciKlass* better_type = NULL;
2251 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2252 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2253 }
2254 i++;
2255 }
2256 }
2257 }
2258
2259 /**
2260 * Record profiling data from parameter profiling at an invoke with
2261 * the type system so that it can propagate it (speculation)
2262 */
2263 void GraphKit::record_profiled_parameters_for_speculation() {
2264 if (!UseTypeSpeculation) {
2265 return;
2266 }
2267 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2268 if (_gvn.type(local(i))->isa_oopptr()) {
2269 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2270 ciKlass* better_type = NULL;
2271 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2272 record_profile_for_speculation(local(i), better_type, ptr_kind);
2273 }
2274 j++;
2275 }
2276 }
2277 }
2278
2279 /**
2280 * Record profiling data from return value profiling at an invoke with
2281 * the type system so that it can propagate it (speculation)
2282 */
2283 void GraphKit::record_profiled_return_for_speculation() {
2284 if (!UseTypeSpeculation) {
2285 return;
2286 }
2287 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2288 ciKlass* better_type = NULL;
2289 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2290 // If profiling reports a single type for the return value,
2291 // feed it to the type system so it can propagate it as a
2292 // speculative type
2293 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2294 }
2295 }
2296
2297 void GraphKit::round_double_result(ciMethod* dest_method) {
2298 // A non-strict method may return a double value which has an extended
2299 // exponent, but this must not be visible in a caller which is 'strict'
2300 // If a strict caller invokes a non-strict callee, round a double result
2301
2302 BasicType result_type = dest_method->return_type()->basic_type();
2303 assert( method() != NULL, "must have caller context");
2304 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2305 // Destination method's return value is on top of stack
2306 // dstore_rounding() does gvn.transform
2307 Node *result = pop_pair();
2308 result = dstore_rounding(result);
2309 push_pair(result);
2310 }
2311 }
2312
2313 // rounding for strict float precision conformance
2314 Node* GraphKit::precision_rounding(Node* n) {
2315 return UseStrictFP && _method->flags().is_strict()
2316 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2317 ? _gvn.transform( new RoundFloatNode(0, n) )
2318 : n;
2319 }
2320
2321 // rounding for strict double precision conformance
2322 Node* GraphKit::dprecision_rounding(Node *n) {
2323 return UseStrictFP && _method->flags().is_strict()
2324 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2325 ? _gvn.transform( new RoundDoubleNode(0, n) )
2326 : n;
2327 }
2328
2329 // rounding for non-strict double stores
2330 Node* GraphKit::dstore_rounding(Node* n) {
2331 return Matcher::strict_fp_requires_explicit_rounding
2332 && UseSSE <= 1
2333 ? _gvn.transform( new RoundDoubleNode(0, n) )
2334 : n;
2335 }
2336
2337 //=============================================================================
2338 // Generate a fast path/slow path idiom. Graph looks like:
2339 // [foo] indicates that 'foo' is a parameter
2340 //
2341 // [in] NULL
2342 // \ /
2343 // CmpP
2344 // Bool ne
2345 // If
2346 // / \
2347 // True False-<2>
2348 // / |
2349 // / cast_not_null
2350 // Load | | ^
2351 // [fast_test] | |
2352 // gvn to opt_test | |
2353 // / \ | <1>
2354 // True False |
2355 // | \\ |
2356 // [slow_call] \[fast_result]
2357 // Ctl Val \ \
2358 // | \ \
2359 // Catch <1> \ \
2360 // / \ ^ \ \
2361 // Ex No_Ex | \ \
2362 // | \ \ | \ <2> \
2363 // ... \ [slow_res] | | \ [null_result]
2364 // \ \--+--+--- | |
2365 // \ | / \ | /
2366 // --------Region Phi
2367 //
2368 //=============================================================================
2369 // Code is structured as a series of driver functions all called 'do_XXX' that
2370 // call a set of helper functions. Helper functions first, then drivers.
2371
2372 //------------------------------null_check_oop---------------------------------
2373 // Null check oop. Set null-path control into Region in slot 3.
2374 // Make a cast-not-nullness use the other not-null control. Return cast.
2375 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2376 bool never_see_null,
2377 bool safe_for_replace,
2378 bool speculative) {
2379 // Initial NULL check taken path
2380 (*null_control) = top();
2381 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2382
2383 // Generate uncommon_trap:
2384 if (never_see_null && (*null_control) != top()) {
2385 // If we see an unexpected null at a check-cast we record it and force a
2386 // recompile; the offending check-cast will be compiled to handle NULLs.
2387 // If we see more than one offending BCI, then all checkcasts in the
2388 // method will be compiled to handle NULLs.
2389 PreserveJVMState pjvms(this);
2390 set_control(*null_control);
2391 replace_in_map(value, null());
2392 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2393 uncommon_trap(reason,
2394 Deoptimization::Action_make_not_entrant);
2395 (*null_control) = top(); // NULL path is dead
2396 }
2397 if ((*null_control) == top() && safe_for_replace) {
2398 replace_in_map(value, cast);
2399 }
2400
2401 // Cast away null-ness on the result
2402 return cast;
2403 }
2404
2405 //------------------------------opt_iff----------------------------------------
2406 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2407 // Return slow-path control.
2408 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2409 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2410
2411 // Fast path taken; set region slot 2
2412 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2413 region->init_req(2,fast_taken); // Capture fast-control
2414
2415 // Fast path not-taken, i.e. slow path
2416 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2417 return slow_taken;
2418 }
2419
2420 //-----------------------------make_runtime_call-------------------------------
2421 Node* GraphKit::make_runtime_call(int flags,
2422 const TypeFunc* call_type, address call_addr,
2423 const char* call_name,
2424 const TypePtr* adr_type,
2425 // The following parms are all optional.
2426 // The first NULL ends the list.
2427 Node* parm0, Node* parm1,
2428 Node* parm2, Node* parm3,
2429 Node* parm4, Node* parm5,
2430 Node* parm6, Node* parm7) {
2431 // Slow-path call
2432 bool is_leaf = !(flags & RC_NO_LEAF);
2433 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2434 if (call_name == NULL) {
2435 assert(!is_leaf, "must supply name for leaf");
2436 call_name = OptoRuntime::stub_name(call_addr);
2437 }
2438 CallNode* call;
2439 if (!is_leaf) {
2440 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2441 bci(), adr_type);
2442 } else if (flags & RC_NO_FP) {
2443 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2444 } else {
2445 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2446 }
2447
2448 // The following is similar to set_edges_for_java_call,
2449 // except that the memory effects of the call are restricted to AliasIdxRaw.
2450
2451 // Slow path call has no side-effects, uses few values
2452 bool wide_in = !(flags & RC_NARROW_MEM);
2453 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2454
2455 Node* prev_mem = NULL;
2456 if (wide_in) {
2457 prev_mem = set_predefined_input_for_runtime_call(call);
2458 } else {
2459 assert(!wide_out, "narrow in => narrow out");
2460 Node* narrow_mem = memory(adr_type);
2461 prev_mem = reset_memory();
2462 map()->set_memory(narrow_mem);
2463 set_predefined_input_for_runtime_call(call);
2464 }
2465
2466 // Hook each parm in order. Stop looking at the first NULL.
2467 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2468 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2469 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2470 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2471 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2472 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2473 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2474 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2475 /* close each nested if ===> */ } } } } } } } }
2476 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2477
2478 if (!is_leaf) {
2479 // Non-leaves can block and take safepoints:
2480 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2481 }
2482 // Non-leaves can throw exceptions:
2483 if (has_io) {
2484 call->set_req(TypeFunc::I_O, i_o());
2485 }
2486
2487 if (flags & RC_UNCOMMON) {
2488 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2489 // (An "if" probability corresponds roughly to an unconditional count.
2490 // Sort of.)
2491 call->set_cnt(PROB_UNLIKELY_MAG(4));
2492 }
2493
2494 Node* c = _gvn.transform(call);
2495 assert(c == call, "cannot disappear");
2496
2497 if (wide_out) {
2498 // Slow path call has full side-effects.
2499 set_predefined_output_for_runtime_call(call);
2500 } else {
2501 // Slow path call has few side-effects, and/or sets few values.
2502 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2503 }
2504
2505 if (has_io) {
2506 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2507 }
2508 return call;
2509
2510 }
2511
2512 //------------------------------merge_memory-----------------------------------
2513 // Merge memory from one path into the current memory state.
2514 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2515 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2516 Node* old_slice = mms.force_memory();
2517 Node* new_slice = mms.memory2();
2518 if (old_slice != new_slice) {
2519 PhiNode* phi;
2520 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2521 if (mms.is_empty()) {
2522 // clone base memory Phi's inputs for this memory slice
2523 assert(old_slice == mms.base_memory(), "sanity");
2524 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2525 _gvn.set_type(phi, Type::MEMORY);
2526 for (uint i = 1; i < phi->req(); i++) {
2527 phi->init_req(i, old_slice->in(i));
2528 }
2529 } else {
2530 phi = old_slice->as_Phi(); // Phi was generated already
2531 }
2532 } else {
2533 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2534 _gvn.set_type(phi, Type::MEMORY);
2535 }
2536 phi->set_req(new_path, new_slice);
2537 mms.set_memory(phi);
2538 }
2539 }
2540 }
2541
2542 //------------------------------make_slow_call_ex------------------------------
2543 // Make the exception handler hookups for the slow call
2544 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2545 if (stopped()) return;
2546
2547 // Make a catch node with just two handlers: fall-through and catch-all
2548 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2549 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2550 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2551 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2552
2553 { PreserveJVMState pjvms(this);
2554 set_control(excp);
2555 set_i_o(i_o);
2556
2557 if (excp != top()) {
2558 if (deoptimize) {
2559 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2560 uncommon_trap(Deoptimization::Reason_unhandled,
2561 Deoptimization::Action_none);
2562 } else {
2563 // Create an exception state also.
2564 // Use an exact type if the caller has specified a specific exception.
2565 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2566 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2567 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2568 }
2569 }
2570 }
2571
2572 // Get the no-exception control from the CatchNode.
2573 set_control(norm);
2574 }
2575
2576 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2577 Node* cmp = NULL;
2578 switch(bt) {
2579 case T_INT: cmp = new CmpINode(in1, in2); break;
2580 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2581 default: fatal("unexpected comparison type %s", type2name(bt));
2582 }
2583 gvn->transform(cmp);
2584 Node* bol = gvn->transform(new BoolNode(cmp, test));
2585 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2586 gvn->transform(iff);
2587 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2588 return iff;
2589 }
2590
2591
2592 //-------------------------------gen_subtype_check-----------------------------
2593 // Generate a subtyping check. Takes as input the subtype and supertype.
2594 // Returns 2 values: sets the default control() to the true path and returns
2595 // the false path. Only reads invariant memory; sets no (visible) memory.
2596 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2597 // but that's not exposed to the optimizer. This call also doesn't take in an
2598 // Object; if you wish to check an Object you need to load the Object's class
2599 // prior to coming here.
2600 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2601 Compile* C = gvn->C;
2602
2603 if ((*ctrl)->is_top()) {
2604 return C->top();
2605 }
2606
2607 // Fast check for identical types, perhaps identical constants.
2608 // The types can even be identical non-constants, in cases
2609 // involving Array.newInstance, Object.clone, etc.
2610 if (subklass == superklass)
2611 return C->top(); // false path is dead; no test needed.
2612
2613 if (gvn->type(superklass)->singleton()) {
2614 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2615 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2616
2617 // In the common case of an exact superklass, try to fold up the
2618 // test before generating code. You may ask, why not just generate
2619 // the code and then let it fold up? The answer is that the generated
2620 // code will necessarily include null checks, which do not always
2621 // completely fold away. If they are also needless, then they turn
2622 // into a performance loss. Example:
2623 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2624 // Here, the type of 'fa' is often exact, so the store check
2625 // of fa[1]=x will fold up, without testing the nullness of x.
2626 switch (C->static_subtype_check(superk, subk)) {
2627 case Compile::SSC_always_false:
2628 {
2629 Node* always_fail = *ctrl;
2630 *ctrl = gvn->C->top();
2631 return always_fail;
2632 }
2633 case Compile::SSC_always_true:
2634 return C->top();
2635 case Compile::SSC_easy_test:
2636 {
2637 // Just do a direct pointer compare and be done.
2638 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2639 *ctrl = gvn->transform(new IfTrueNode(iff));
2640 return gvn->transform(new IfFalseNode(iff));
2641 }
2642 case Compile::SSC_full_test:
2643 break;
2644 default:
2645 ShouldNotReachHere();
2646 }
2647 }
2648
2649 // %%% Possible further optimization: Even if the superklass is not exact,
2650 // if the subklass is the unique subtype of the superklass, the check
2651 // will always succeed. We could leave a dependency behind to ensure this.
2652
2653 // First load the super-klass's check-offset
2654 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2655 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2656 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2657 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2658 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2659
2660 // Load from the sub-klass's super-class display list, or a 1-word cache of
2661 // the secondary superclass list, or a failing value with a sentinel offset
2662 // if the super-klass is an interface or exceptionally deep in the Java
2663 // hierarchy and we have to scan the secondary superclass list the hard way.
2664 // Worst-case type is a little odd: NULL is allowed as a result (usually
2665 // klass loads can never produce a NULL).
2666 Node *chk_off_X = chk_off;
2667 #ifdef _LP64
2668 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2669 #endif
2670 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2671 // For some types like interfaces the following loadKlass is from a 1-word
2672 // cache which is mutable so can't use immutable memory. Other
2673 // types load from the super-class display table which is immutable.
2674 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2675 Node *kmem = might_be_cache ? m : C->immutable_memory();
2676 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2677
2678 // Compile speed common case: ARE a subtype and we canNOT fail
2679 if( superklass == nkls )
2680 return C->top(); // false path is dead; no test needed.
2681
2682 // See if we get an immediate positive hit. Happens roughly 83% of the
2683 // time. Test to see if the value loaded just previously from the subklass
2684 // is exactly the superklass.
2685 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2686 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2687 *ctrl = gvn->transform(new IfFalseNode(iff1));
2688
2689 // Compile speed common case: Check for being deterministic right now. If
2690 // chk_off is a constant and not equal to cacheoff then we are NOT a
2691 // subklass. In this case we need exactly the 1 test above and we can
2692 // return those results immediately.
2693 if (!might_be_cache) {
2694 Node* not_subtype_ctrl = *ctrl;
2695 *ctrl = iftrue1; // We need exactly the 1 test above
2696 return not_subtype_ctrl;
2697 }
2698
2699 // Gather the various success & failures here
2700 RegionNode *r_ok_subtype = new RegionNode(4);
2701 gvn->record_for_igvn(r_ok_subtype);
2702 RegionNode *r_not_subtype = new RegionNode(3);
2703 gvn->record_for_igvn(r_not_subtype);
2704
2705 r_ok_subtype->init_req(1, iftrue1);
2706
2707 // Check for immediate negative hit. Happens roughly 11% of the time (which
2708 // is roughly 63% of the remaining cases). Test to see if the loaded
2709 // check-offset points into the subklass display list or the 1-element
2710 // cache. If it points to the display (and NOT the cache) and the display
2711 // missed then it's not a subtype.
2712 Node *cacheoff = gvn->intcon(cacheoff_con);
2713 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2714 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2715 *ctrl = gvn->transform(new IfFalseNode(iff2));
2716
2717 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2718 // No performance impact (too rare) but allows sharing of secondary arrays
2719 // which has some footprint reduction.
2720 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2721 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2722 *ctrl = gvn->transform(new IfFalseNode(iff3));
2723
2724 // -- Roads not taken here: --
2725 // We could also have chosen to perform the self-check at the beginning
2726 // of this code sequence, as the assembler does. This would not pay off
2727 // the same way, since the optimizer, unlike the assembler, can perform
2728 // static type analysis to fold away many successful self-checks.
2729 // Non-foldable self checks work better here in second position, because
2730 // the initial primary superclass check subsumes a self-check for most
2731 // types. An exception would be a secondary type like array-of-interface,
2732 // which does not appear in its own primary supertype display.
2733 // Finally, we could have chosen to move the self-check into the
2734 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2735 // dependent manner. But it is worthwhile to have the check here,
2736 // where it can be perhaps be optimized. The cost in code space is
2737 // small (register compare, branch).
2738
2739 // Now do a linear scan of the secondary super-klass array. Again, no real
2740 // performance impact (too rare) but it's gotta be done.
2741 // Since the code is rarely used, there is no penalty for moving it
2742 // out of line, and it can only improve I-cache density.
2743 // The decision to inline or out-of-line this final check is platform
2744 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2745 Node* psc = gvn->transform(
2746 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2747
2748 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2749 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2750 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2751
2752 // Return false path; set default control to true path.
2753 *ctrl = gvn->transform(r_ok_subtype);
2754 return gvn->transform(r_not_subtype);
2755 }
2756
2757 // Profile-driven exact type check:
2758 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2759 float prob,
2760 Node* *casted_receiver) {
2761 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2762 Node* recv_klass = load_object_klass(receiver);
2763 Node* want_klass = makecon(tklass);
2764 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2765 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2766 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2767 set_control( _gvn.transform( new IfTrueNode (iff) ));
2768 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2769
2770 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2771 assert(recv_xtype->klass_is_exact(), "");
2772
2773 // Subsume downstream occurrences of receiver with a cast to
2774 // recv_xtype, since now we know what the type will be.
2775 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2776 (*casted_receiver) = _gvn.transform(cast);
2777 // (User must make the replace_in_map call.)
2778
2779 return fail;
2780 }
2781
2782
2783 //------------------------------seems_never_null-------------------------------
2784 // Use null_seen information if it is available from the profile.
2785 // If we see an unexpected null at a type check we record it and force a
2786 // recompile; the offending check will be recompiled to handle NULLs.
2787 // If we see several offending BCIs, then all checks in the
2788 // method will be recompiled.
2789 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2790 speculating = !_gvn.type(obj)->speculative_maybe_null();
2791 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2792 if (UncommonNullCast // Cutout for this technique
2793 && obj != null() // And not the -Xcomp stupid case?
2794 && !too_many_traps(reason)
2795 ) {
2796 if (speculating) {
2797 return true;
2798 }
2799 if (data == NULL)
2800 // Edge case: no mature data. Be optimistic here.
2801 return true;
2802 // If the profile has not seen a null, assume it won't happen.
2803 assert(java_bc() == Bytecodes::_checkcast ||
2804 java_bc() == Bytecodes::_instanceof ||
2805 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2806 return !data->as_BitData()->null_seen();
2807 }
2808 speculating = false;
2809 return false;
2810 }
2811
2812 //------------------------maybe_cast_profiled_receiver-------------------------
2813 // If the profile has seen exactly one type, narrow to exactly that type.
2814 // Subsequent type checks will always fold up.
2815 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2816 ciKlass* require_klass,
2817 ciKlass* spec_klass,
2818 bool safe_for_replace) {
2819 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2820
2821 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2822
2823 // Make sure we haven't already deoptimized from this tactic.
2824 if (too_many_traps(reason) || too_many_recompiles(reason))
2825 return NULL;
2826
2827 // (No, this isn't a call, but it's enough like a virtual call
2828 // to use the same ciMethod accessor to get the profile info...)
2829 // If we have a speculative type use it instead of profiling (which
2830 // may not help us)
2831 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2832 if (exact_kls != NULL) {// no cast failures here
2833 if (require_klass == NULL ||
2834 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2835 // If we narrow the type to match what the type profile sees or
2836 // the speculative type, we can then remove the rest of the
2837 // cast.
2838 // This is a win, even if the exact_kls is very specific,
2839 // because downstream operations, such as method calls,
2840 // will often benefit from the sharper type.
2841 Node* exact_obj = not_null_obj; // will get updated in place...
2842 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2843 &exact_obj);
2844 { PreserveJVMState pjvms(this);
2845 set_control(slow_ctl);
2846 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2847 }
2848 if (safe_for_replace) {
2849 replace_in_map(not_null_obj, exact_obj);
2850 }
2851 return exact_obj;
2852 }
2853 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2854 }
2855
2856 return NULL;
2857 }
2858
2859 /**
2860 * Cast obj to type and emit guard unless we had too many traps here
2861 * already
2862 *
2863 * @param obj node being casted
2864 * @param type type to cast the node to
2865 * @param not_null true if we know node cannot be null
2866 */
2867 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2868 ciKlass* type,
2869 bool not_null) {
2870 if (stopped()) {
2871 return obj;
2872 }
2873
2874 // type == NULL if profiling tells us this object is always null
2875 if (type != NULL) {
2876 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2877 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2878
2879 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2880 !too_many_traps(class_reason) &&
2881 !too_many_recompiles(class_reason)) {
2882 Node* not_null_obj = NULL;
2883 // not_null is true if we know the object is not null and
2884 // there's no need for a null check
2885 if (!not_null) {
2886 Node* null_ctl = top();
2887 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2888 assert(null_ctl->is_top(), "no null control here");
2889 } else {
2890 not_null_obj = obj;
2891 }
2892
2893 Node* exact_obj = not_null_obj;
2894 ciKlass* exact_kls = type;
2895 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2896 &exact_obj);
2897 {
2898 PreserveJVMState pjvms(this);
2899 set_control(slow_ctl);
2900 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2901 }
2902 replace_in_map(not_null_obj, exact_obj);
2903 obj = exact_obj;
2904 }
2905 } else {
2906 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2907 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2908 Node* exact_obj = null_assert(obj);
2909 replace_in_map(obj, exact_obj);
2910 obj = exact_obj;
2911 }
2912 }
2913 return obj;
2914 }
2915
2916 //-------------------------------gen_instanceof--------------------------------
2917 // Generate an instance-of idiom. Used by both the instance-of bytecode
2918 // and the reflective instance-of call.
2919 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2920 kill_dead_locals(); // Benefit all the uncommon traps
2921 assert( !stopped(), "dead parse path should be checked in callers" );
2922 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2923 "must check for not-null not-dead klass in callers");
2924
2925 // Make the merge point
2926 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2927 RegionNode* region = new RegionNode(PATH_LIMIT);
2928 Node* phi = new PhiNode(region, TypeInt::BOOL);
2929 C->set_has_split_ifs(true); // Has chance for split-if optimization
2930
2931 ciProfileData* data = NULL;
2932 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2933 data = method()->method_data()->bci_to_data(bci());
2934 }
2935 bool speculative_not_null = false;
2936 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2937 && seems_never_null(obj, data, speculative_not_null));
2938
2939 // Null check; get casted pointer; set region slot 3
2940 Node* null_ctl = top();
2941 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2942
2943 // If not_null_obj is dead, only null-path is taken
2944 if (stopped()) { // Doing instance-of on a NULL?
2945 set_control(null_ctl);
2946 return intcon(0);
2947 }
2948 region->init_req(_null_path, null_ctl);
2949 phi ->init_req(_null_path, intcon(0)); // Set null path value
2950 if (null_ctl == top()) {
2951 // Do this eagerly, so that pattern matches like is_diamond_phi
2952 // will work even during parsing.
2953 assert(_null_path == PATH_LIMIT-1, "delete last");
2954 region->del_req(_null_path);
2955 phi ->del_req(_null_path);
2956 }
2957
2958 // Do we know the type check always succeed?
2959 bool known_statically = false;
2960 if (_gvn.type(superklass)->singleton()) {
2961 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2962 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2963 if (subk != NULL && subk->is_loaded()) {
2964 int static_res = C->static_subtype_check(superk, subk);
2965 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2966 }
2967 }
2968
2969 if (!known_statically) {
2970 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2971 // We may not have profiling here or it may not help us. If we
2972 // have a speculative type use it to perform an exact cast.
2973 ciKlass* spec_obj_type = obj_type->speculative_type();
2974 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2975 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2976 if (stopped()) { // Profile disagrees with this path.
2977 set_control(null_ctl); // Null is the only remaining possibility.
2978 return intcon(0);
2979 }
2980 if (cast_obj != NULL) {
2981 not_null_obj = cast_obj;
2982 }
2983 }
2984 }
2985
2986 // Load the object's klass
2987 Node* obj_klass = load_object_klass(not_null_obj);
2988
2989 // Generate the subtype check
2990 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2991
2992 // Plug in the success path to the general merge in slot 1.
2993 region->init_req(_obj_path, control());
2994 phi ->init_req(_obj_path, intcon(1));
2995
2996 // Plug in the failing path to the general merge in slot 2.
2997 region->init_req(_fail_path, not_subtype_ctrl);
2998 phi ->init_req(_fail_path, intcon(0));
2999
3000 // Return final merged results
3001 set_control( _gvn.transform(region) );
3002 record_for_igvn(region);
3003
3004 // If we know the type check always succeeds then we don't use the
3005 // profiling data at this bytecode. Don't lose it, feed it to the
3006 // type system as a speculative type.
3007 if (safe_for_replace) {
3008 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3009 replace_in_map(obj, casted_obj);
3010 }
3011
3012 return _gvn.transform(phi);
3013 }
3014
3015 //-------------------------------gen_checkcast---------------------------------
3016 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3017 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3018 // uncommon-trap paths work. Adjust stack after this call.
3019 // If failure_control is supplied and not null, it is filled in with
3020 // the control edge for the cast failure. Otherwise, an appropriate
3021 // uncommon trap or exception is thrown.
3022 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3023 Node* *failure_control) {
3024 kill_dead_locals(); // Benefit all the uncommon traps
3025 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3026 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3027
3028 // Fast cutout: Check the case that the cast is vacuously true.
3029 // This detects the common cases where the test will short-circuit
3030 // away completely. We do this before we perform the null check,
3031 // because if the test is going to turn into zero code, we don't
3032 // want a residual null check left around. (Causes a slowdown,
3033 // for example, in some objArray manipulations, such as a[i]=a[j].)
3034 if (tk->singleton()) {
3035 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3036 if (objtp != NULL && objtp->klass() != NULL) {
3037 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3038 case Compile::SSC_always_true:
3039 // If we know the type check always succeed then we don't use
3040 // the profiling data at this bytecode. Don't lose it, feed it
3041 // to the type system as a speculative type.
3042 return record_profiled_receiver_for_speculation(obj);
3043 case Compile::SSC_always_false:
3044 // It needs a null check because a null will *pass* the cast check.
3045 // A non-null value will always produce an exception.
3046 return null_assert(obj);
3047 }
3048 }
3049 }
3050
3051 ciProfileData* data = NULL;
3052 bool safe_for_replace = false;
3053 if (failure_control == NULL) { // use MDO in regular case only
3054 assert(java_bc() == Bytecodes::_aastore ||
3055 java_bc() == Bytecodes::_checkcast,
3056 "interpreter profiles type checks only for these BCs");
3057 data = method()->method_data()->bci_to_data(bci());
3058 safe_for_replace = true;
3059 }
3060
3061 // Make the merge point
3062 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3063 RegionNode* region = new RegionNode(PATH_LIMIT);
3064 Node* phi = new PhiNode(region, toop);
3065 C->set_has_split_ifs(true); // Has chance for split-if optimization
3066
3067 // Use null-cast information if it is available
3068 bool speculative_not_null = false;
3069 bool never_see_null = ((failure_control == NULL) // regular case only
3070 && seems_never_null(obj, data, speculative_not_null));
3071
3072 // Null check; get casted pointer; set region slot 3
3073 Node* null_ctl = top();
3074 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3075
3076 // If not_null_obj is dead, only null-path is taken
3077 if (stopped()) { // Doing instance-of on a NULL?
3078 set_control(null_ctl);
3079 return null();
3080 }
3081 region->init_req(_null_path, null_ctl);
3082 phi ->init_req(_null_path, null()); // Set null path value
3083 if (null_ctl == top()) {
3084 // Do this eagerly, so that pattern matches like is_diamond_phi
3085 // will work even during parsing.
3086 assert(_null_path == PATH_LIMIT-1, "delete last");
3087 region->del_req(_null_path);
3088 phi ->del_req(_null_path);
3089 }
3090
3091 Node* cast_obj = NULL;
3092 if (tk->klass_is_exact()) {
3093 // The following optimization tries to statically cast the speculative type of the object
3094 // (for example obtained during profiling) to the type of the superklass and then do a
3095 // dynamic check that the type of the object is what we expect. To work correctly
3096 // for checkcast and aastore the type of superklass should be exact.
3097 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3098 // We may not have profiling here or it may not help us. If we have
3099 // a speculative type use it to perform an exact cast.
3100 ciKlass* spec_obj_type = obj_type->speculative_type();
3101 if (spec_obj_type != NULL || data != NULL) {
3102 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3103 if (cast_obj != NULL) {
3104 if (failure_control != NULL) // failure is now impossible
3105 (*failure_control) = top();
3106 // adjust the type of the phi to the exact klass:
3107 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3108 }
3109 }
3110 }
3111
3112 if (cast_obj == NULL) {
3113 // Load the object's klass
3114 Node* obj_klass = load_object_klass(not_null_obj);
3115
3116 // Generate the subtype check
3117 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3118
3119 // Plug in success path into the merge
3120 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3121 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3122 if (failure_control == NULL) {
3123 if (not_subtype_ctrl != top()) { // If failure is possible
3124 PreserveJVMState pjvms(this);
3125 set_control(not_subtype_ctrl);
3126 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3127 }
3128 } else {
3129 (*failure_control) = not_subtype_ctrl;
3130 }
3131 }
3132
3133 region->init_req(_obj_path, control());
3134 phi ->init_req(_obj_path, cast_obj);
3135
3136 // A merge of NULL or Casted-NotNull obj
3137 Node* res = _gvn.transform(phi);
3138
3139 // Note I do NOT always 'replace_in_map(obj,result)' here.
3140 // if( tk->klass()->can_be_primary_super() )
3141 // This means that if I successfully store an Object into an array-of-String
3142 // I 'forget' that the Object is really now known to be a String. I have to
3143 // do this because we don't have true union types for interfaces - if I store
3144 // a Baz into an array-of-Interface and then tell the optimizer it's an
3145 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3146 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3147 // replace_in_map( obj, res );
3148
3149 // Return final merged results
3150 set_control( _gvn.transform(region) );
3151 record_for_igvn(region);
3152
3153 return record_profiled_receiver_for_speculation(res);
3154 }
3155
3156 //------------------------------next_monitor-----------------------------------
3157 // What number should be given to the next monitor?
3158 int GraphKit::next_monitor() {
3159 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3160 int next = current + C->sync_stack_slots();
3161 // Keep the toplevel high water mark current:
3162 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3163 return current;
3164 }
3165
3166 //------------------------------insert_mem_bar---------------------------------
3167 // Memory barrier to avoid floating things around
3168 // The membar serves as a pinch point between both control and all memory slices.
3169 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3170 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3171 mb->init_req(TypeFunc::Control, control());
3172 mb->init_req(TypeFunc::Memory, reset_memory());
3173 Node* membar = _gvn.transform(mb);
3174 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3175 set_all_memory_call(membar);
3176 return membar;
3177 }
3178
3179 //-------------------------insert_mem_bar_volatile----------------------------
3180 // Memory barrier to avoid floating things around
3181 // The membar serves as a pinch point between both control and memory(alias_idx).
3182 // If you want to make a pinch point on all memory slices, do not use this
3183 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3184 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3185 // When Parse::do_put_xxx updates a volatile field, it appends a series
3186 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3187 // The first membar is on the same memory slice as the field store opcode.
3188 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3189 // All the other membars (for other volatile slices, including AliasIdxBot,
3190 // which stands for all unknown volatile slices) are control-dependent
3191 // on the first membar. This prevents later volatile loads or stores
3192 // from sliding up past the just-emitted store.
3193
3194 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3195 mb->set_req(TypeFunc::Control,control());
3196 if (alias_idx == Compile::AliasIdxBot) {
3197 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3198 } else {
3199 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3200 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3201 }
3202 Node* membar = _gvn.transform(mb);
3203 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3204 if (alias_idx == Compile::AliasIdxBot) {
3205 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3206 } else {
3207 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3208 }
3209 return membar;
3210 }
3211
3212 //------------------------------shared_lock------------------------------------
3213 // Emit locking code.
3214 FastLockNode* GraphKit::shared_lock(Node* obj) {
3215 // bci is either a monitorenter bc or InvocationEntryBci
3216 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3217 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3218
3219 if( !GenerateSynchronizationCode )
3220 return NULL; // Not locking things?
3221 if (stopped()) // Dead monitor?
3222 return NULL;
3223
3224 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3225
3226 // Box the stack location
3227 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3228 Node* mem = reset_memory();
3229
3230 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3231 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3232 // Create the counters for this fast lock.
3233 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3234 }
3235
3236 // Create the rtm counters for this fast lock if needed.
3237 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3238
3239 // Add monitor to debug info for the slow path. If we block inside the
3240 // slow path and de-opt, we need the monitor hanging around
3241 map()->push_monitor( flock );
3242
3243 const TypeFunc *tf = LockNode::lock_type();
3244 LockNode *lock = new LockNode(C, tf);
3245
3246 lock->init_req( TypeFunc::Control, control() );
3247 lock->init_req( TypeFunc::Memory , mem );
3248 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3249 lock->init_req( TypeFunc::FramePtr, frameptr() );
3250 lock->init_req( TypeFunc::ReturnAdr, top() );
3251
3252 lock->init_req(TypeFunc::Parms + 0, obj);
3253 lock->init_req(TypeFunc::Parms + 1, box);
3254 lock->init_req(TypeFunc::Parms + 2, flock);
3255 add_safepoint_edges(lock);
3256
3257 lock = _gvn.transform( lock )->as_Lock();
3258
3259 // lock has no side-effects, sets few values
3260 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3261
3262 insert_mem_bar(Op_MemBarAcquireLock);
3263
3264 // Add this to the worklist so that the lock can be eliminated
3265 record_for_igvn(lock);
3266
3267 #ifndef PRODUCT
3268 if (PrintLockStatistics) {
3269 // Update the counter for this lock. Don't bother using an atomic
3270 // operation since we don't require absolute accuracy.
3271 lock->create_lock_counter(map()->jvms());
3272 increment_counter(lock->counter()->addr());
3273 }
3274 #endif
3275
3276 return flock;
3277 }
3278
3279
3280 //------------------------------shared_unlock----------------------------------
3281 // Emit unlocking code.
3282 void GraphKit::shared_unlock(Node* box, Node* obj) {
3283 // bci is either a monitorenter bc or InvocationEntryBci
3284 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3285 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3286
3287 if( !GenerateSynchronizationCode )
3288 return;
3289 if (stopped()) { // Dead monitor?
3290 map()->pop_monitor(); // Kill monitor from debug info
3291 return;
3292 }
3293
3294 // Memory barrier to avoid floating things down past the locked region
3295 insert_mem_bar(Op_MemBarReleaseLock);
3296
3297 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3298 UnlockNode *unlock = new UnlockNode(C, tf);
3299 #ifdef ASSERT
3300 unlock->set_dbg_jvms(sync_jvms());
3301 #endif
3302 uint raw_idx = Compile::AliasIdxRaw;
3303 unlock->init_req( TypeFunc::Control, control() );
3304 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3305 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3306 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3307 unlock->init_req( TypeFunc::ReturnAdr, top() );
3308
3309 unlock->init_req(TypeFunc::Parms + 0, obj);
3310 unlock->init_req(TypeFunc::Parms + 1, box);
3311 unlock = _gvn.transform(unlock)->as_Unlock();
3312
3313 Node* mem = reset_memory();
3314
3315 // unlock has no side-effects, sets few values
3316 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3317
3318 // Kill monitor from debug info
3319 map()->pop_monitor( );
3320 }
3321
3322 //-------------------------------get_layout_helper-----------------------------
3323 // If the given klass is a constant or known to be an array,
3324 // fetch the constant layout helper value into constant_value
3325 // and return (Node*)NULL. Otherwise, load the non-constant
3326 // layout helper value, and return the node which represents it.
3327 // This two-faced routine is useful because allocation sites
3328 // almost always feature constant types.
3329 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3330 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3331 if (!StressReflectiveCode && inst_klass != NULL) {
3332 ciKlass* klass = inst_klass->klass();
3333 bool xklass = inst_klass->klass_is_exact();
3334 if (xklass || klass->is_array_klass()) {
3335 jint lhelper = klass->layout_helper();
3336 if (lhelper != Klass::_lh_neutral_value) {
3337 constant_value = lhelper;
3338 return (Node*) NULL;
3339 }
3340 }
3341 }
3342 constant_value = Klass::_lh_neutral_value; // put in a known value
3343 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3344 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3345 }
3346
3347 // We just put in an allocate/initialize with a big raw-memory effect.
3348 // Hook selected additional alias categories on the initialization.
3349 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3350 MergeMemNode* init_in_merge,
3351 Node* init_out_raw) {
3352 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3353 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3354
3355 Node* prevmem = kit.memory(alias_idx);
3356 init_in_merge->set_memory_at(alias_idx, prevmem);
3357 kit.set_memory(init_out_raw, alias_idx);
3358 }
3359
3360 //---------------------------set_output_for_allocation-------------------------
3361 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3362 const TypeOopPtr* oop_type,
3363 bool deoptimize_on_exception) {
3364 int rawidx = Compile::AliasIdxRaw;
3365 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3366 add_safepoint_edges(alloc);
3367 Node* allocx = _gvn.transform(alloc);
3368 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3369 // create memory projection for i_o
3370 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3371 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3372
3373 // create a memory projection as for the normal control path
3374 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3375 set_memory(malloc, rawidx);
3376
3377 // a normal slow-call doesn't change i_o, but an allocation does
3378 // we create a separate i_o projection for the normal control path
3379 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3380 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3381
3382 // put in an initialization barrier
3383 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3384 rawoop)->as_Initialize();
3385 assert(alloc->initialization() == init, "2-way macro link must work");
3386 assert(init ->allocation() == alloc, "2-way macro link must work");
3387 {
3388 // Extract memory strands which may participate in the new object's
3389 // initialization, and source them from the new InitializeNode.
3390 // This will allow us to observe initializations when they occur,
3391 // and link them properly (as a group) to the InitializeNode.
3392 assert(init->in(InitializeNode::Memory) == malloc, "");
3393 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3394 init->set_req(InitializeNode::Memory, minit_in);
3395 record_for_igvn(minit_in); // fold it up later, if possible
3396 Node* minit_out = memory(rawidx);
3397 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3398 if (oop_type->isa_aryptr()) {
3399 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3400 int elemidx = C->get_alias_index(telemref);
3401 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3402 } else if (oop_type->isa_instptr()) {
3403 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3404 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3405 ciField* field = ik->nonstatic_field_at(i);
3406 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3407 continue; // do not bother to track really large numbers of fields
3408 // Find (or create) the alias category for this field:
3409 int fieldidx = C->alias_type(field)->index();
3410 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3411 }
3412 }
3413 }
3414
3415 // Cast raw oop to the real thing...
3416 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3417 javaoop = _gvn.transform(javaoop);
3418 C->set_recent_alloc(control(), javaoop);
3419 assert(just_allocated_object(control()) == javaoop, "just allocated");
3420
3421 #ifdef ASSERT
3422 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3423 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3424 "Ideal_allocation works");
3425 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3426 "Ideal_allocation works");
3427 if (alloc->is_AllocateArray()) {
3428 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3429 "Ideal_allocation works");
3430 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3431 "Ideal_allocation works");
3432 } else {
3433 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3434 }
3435 }
3436 #endif //ASSERT
3437
3438 return javaoop;
3439 }
3440
3441 //---------------------------new_instance--------------------------------------
3442 // This routine takes a klass_node which may be constant (for a static type)
3443 // or may be non-constant (for reflective code). It will work equally well
3444 // for either, and the graph will fold nicely if the optimizer later reduces
3445 // the type to a constant.
3446 // The optional arguments are for specialized use by intrinsics:
3447 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3448 // - If 'return_size_val', report the the total object size to the caller.
3449 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3450 Node* GraphKit::new_instance(Node* klass_node,
3451 Node* extra_slow_test,
3452 Node* *return_size_val,
3453 bool deoptimize_on_exception) {
3454 // Compute size in doublewords
3455 // The size is always an integral number of doublewords, represented
3456 // as a positive bytewise size stored in the klass's layout_helper.
3457 // The layout_helper also encodes (in a low bit) the need for a slow path.
3458 jint layout_con = Klass::_lh_neutral_value;
3459 Node* layout_val = get_layout_helper(klass_node, layout_con);
3460 int layout_is_con = (layout_val == NULL);
3461
3462 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3463 // Generate the initial go-slow test. It's either ALWAYS (return a
3464 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3465 // case) a computed value derived from the layout_helper.
3466 Node* initial_slow_test = NULL;
3467 if (layout_is_con) {
3468 assert(!StressReflectiveCode, "stress mode does not use these paths");
3469 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3470 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3471 } else { // reflective case
3472 // This reflective path is used by Unsafe.allocateInstance.
3473 // (It may be stress-tested by specifying StressReflectiveCode.)
3474 // Basically, we want to get into the VM is there's an illegal argument.
3475 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3476 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3477 if (extra_slow_test != intcon(0)) {
3478 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3479 }
3480 // (Macro-expander will further convert this to a Bool, if necessary.)
3481 }
3482
3483 // Find the size in bytes. This is easy; it's the layout_helper.
3484 // The size value must be valid even if the slow path is taken.
3485 Node* size = NULL;
3486 if (layout_is_con) {
3487 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3488 } else { // reflective case
3489 // This reflective path is used by clone and Unsafe.allocateInstance.
3490 size = ConvI2X(layout_val);
3491
3492 // Clear the low bits to extract layout_helper_size_in_bytes:
3493 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3494 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3495 size = _gvn.transform( new AndXNode(size, mask) );
3496 }
3497 if (return_size_val != NULL) {
3498 (*return_size_val) = size;
3499 }
3500
3501 // This is a precise notnull oop of the klass.
3502 // (Actually, it need not be precise if this is a reflective allocation.)
3503 // It's what we cast the result to.
3504 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3505 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3506 const TypeOopPtr* oop_type = tklass->as_instance_type();
3507
3508 // Now generate allocation code
3509
3510 // The entire memory state is needed for slow path of the allocation
3511 // since GC and deoptimization can happened.
3512 Node *mem = reset_memory();
3513 set_all_memory(mem); // Create new memory state
3514
3515 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3516 control(), mem, i_o(),
3517 size, klass_node,
3518 initial_slow_test);
3519
3520 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3521 }
3522
3523 //-------------------------------new_array-------------------------------------
3524 // helper for both newarray and anewarray
3525 // The 'length' parameter is (obviously) the length of the array.
3526 // See comments on new_instance for the meaning of the other arguments.
3527 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3528 Node* length, // number of array elements
3529 int nargs, // number of arguments to push back for uncommon trap
3530 Node* *return_size_val,
3531 bool deoptimize_on_exception) {
3532 jint layout_con = Klass::_lh_neutral_value;
3533 Node* layout_val = get_layout_helper(klass_node, layout_con);
3534 int layout_is_con = (layout_val == NULL);
3535
3536 if (!layout_is_con && !StressReflectiveCode &&
3537 !too_many_traps(Deoptimization::Reason_class_check)) {
3538 // This is a reflective array creation site.
3539 // Optimistically assume that it is a subtype of Object[],
3540 // so that we can fold up all the address arithmetic.
3541 layout_con = Klass::array_layout_helper(T_OBJECT);
3542 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3543 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3544 { BuildCutout unless(this, bol_lh, PROB_MAX);
3545 inc_sp(nargs);
3546 uncommon_trap(Deoptimization::Reason_class_check,
3547 Deoptimization::Action_maybe_recompile);
3548 }
3549 layout_val = NULL;
3550 layout_is_con = true;
3551 }
3552
3553 // Generate the initial go-slow test. Make sure we do not overflow
3554 // if length is huge (near 2Gig) or negative! We do not need
3555 // exact double-words here, just a close approximation of needed
3556 // double-words. We can't add any offset or rounding bits, lest we
3557 // take a size -1 of bytes and make it positive. Use an unsigned
3558 // compare, so negative sizes look hugely positive.
3559 int fast_size_limit = FastAllocateSizeLimit;
3560 if (layout_is_con) {
3561 assert(!StressReflectiveCode, "stress mode does not use these paths");
3562 // Increase the size limit if we have exact knowledge of array type.
3563 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3564 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3565 }
3566
3567 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3568 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3569
3570 // --- Size Computation ---
3571 // array_size = round_to_heap(array_header + (length << elem_shift));
3572 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3573 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3574 // The rounding mask is strength-reduced, if possible.
3575 int round_mask = MinObjAlignmentInBytes - 1;
3576 Node* header_size = NULL;
3577 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3578 // (T_BYTE has the weakest alignment and size restrictions...)
3579 if (layout_is_con) {
3580 int hsize = Klass::layout_helper_header_size(layout_con);
3581 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3582 BasicType etype = Klass::layout_helper_element_type(layout_con);
3583 if ((round_mask & ~right_n_bits(eshift)) == 0)
3584 round_mask = 0; // strength-reduce it if it goes away completely
3585 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3586 assert(header_size_min <= hsize, "generic minimum is smallest");
3587 header_size_min = hsize;
3588 header_size = intcon(hsize + round_mask);
3589 } else {
3590 Node* hss = intcon(Klass::_lh_header_size_shift);
3591 Node* hsm = intcon(Klass::_lh_header_size_mask);
3592 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3593 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3594 Node* mask = intcon(round_mask);
3595 header_size = _gvn.transform( new AddINode(hsize, mask) );
3596 }
3597
3598 Node* elem_shift = NULL;
3599 if (layout_is_con) {
3600 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3601 if (eshift != 0)
3602 elem_shift = intcon(eshift);
3603 } else {
3604 // There is no need to mask or shift this value.
3605 // The semantics of LShiftINode include an implicit mask to 0x1F.
3606 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3607 elem_shift = layout_val;
3608 }
3609
3610 // Transition to native address size for all offset calculations:
3611 Node* lengthx = ConvI2X(length);
3612 Node* headerx = ConvI2X(header_size);
3613 #ifdef _LP64
3614 { const TypeInt* tilen = _gvn.find_int_type(length);
3615 if (tilen != NULL && tilen->_lo < 0) {
3616 // Add a manual constraint to a positive range. Cf. array_element_address.
3617 jint size_max = fast_size_limit;
3618 if (size_max > tilen->_hi) size_max = tilen->_hi;
3619 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3620
3621 // Only do a narrow I2L conversion if the range check passed.
3622 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3623 _gvn.transform(iff);
3624 RegionNode* region = new RegionNode(3);
3625 _gvn.set_type(region, Type::CONTROL);
3626 lengthx = new PhiNode(region, TypeLong::LONG);
3627 _gvn.set_type(lengthx, TypeLong::LONG);
3628
3629 // Range check passed. Use ConvI2L node with narrow type.
3630 Node* passed = IfFalse(iff);
3631 region->init_req(1, passed);
3632 // Make I2L conversion control dependent to prevent it from
3633 // floating above the range check during loop optimizations.
3634 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3635
3636 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3637 region->init_req(2, IfTrue(iff));
3638 lengthx->init_req(2, ConvI2X(length));
3639
3640 set_control(region);
3641 record_for_igvn(region);
3642 record_for_igvn(lengthx);
3643 }
3644 }
3645 #endif
3646
3647 // Combine header size (plus rounding) and body size. Then round down.
3648 // This computation cannot overflow, because it is used only in two
3649 // places, one where the length is sharply limited, and the other
3650 // after a successful allocation.
3651 Node* abody = lengthx;
3652 if (elem_shift != NULL)
3653 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3654 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3655 if (round_mask != 0) {
3656 Node* mask = MakeConX(~round_mask);
3657 size = _gvn.transform( new AndXNode(size, mask) );
3658 }
3659 // else if round_mask == 0, the size computation is self-rounding
3660
3661 if (return_size_val != NULL) {
3662 // This is the size
3663 (*return_size_val) = size;
3664 }
3665
3666 // Now generate allocation code
3667
3668 // The entire memory state is needed for slow path of the allocation
3669 // since GC and deoptimization can happened.
3670 Node *mem = reset_memory();
3671 set_all_memory(mem); // Create new memory state
3672
3673 if (initial_slow_test->is_Bool()) {
3674 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3675 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3676 }
3677
3678 // Create the AllocateArrayNode and its result projections
3679 AllocateArrayNode* alloc
3680 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3681 control(), mem, i_o(),
3682 size, klass_node,
3683 initial_slow_test,
3684 length);
3685
3686 // Cast to correct type. Note that the klass_node may be constant or not,
3687 // and in the latter case the actual array type will be inexact also.
3688 // (This happens via a non-constant argument to inline_native_newArray.)
3689 // In any case, the value of klass_node provides the desired array type.
3690 const TypeInt* length_type = _gvn.find_int_type(length);
3691 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3692 if (ary_type->isa_aryptr() && length_type != NULL) {
3693 // Try to get a better type than POS for the size
3694 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3695 }
3696
3697 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3698
3699 // Cast length on remaining path to be as narrow as possible
3700 if (map()->find_edge(length) >= 0) {
3701 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3702 if (ccast != length) {
3703 _gvn.set_type_bottom(ccast);
3704 record_for_igvn(ccast);
3705 replace_in_map(length, ccast);
3706 }
3707 }
3708
3709 return javaoop;
3710 }
3711
3712 // The following "Ideal_foo" functions are placed here because they recognize
3713 // the graph shapes created by the functions immediately above.
3714
3715 //---------------------------Ideal_allocation----------------------------------
3716 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3717 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3718 if (ptr == NULL) { // reduce dumb test in callers
3719 return NULL;
3720 }
3721 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3722 ptr = ptr->in(1);
3723 if (ptr == NULL) return NULL;
3724 }
3725 // Return NULL for allocations with several casts:
3726 // j.l.reflect.Array.newInstance(jobject, jint)
3727 // Object.clone()
3728 // to keep more precise type from last cast.
3729 if (ptr->is_Proj()) {
3730 Node* allo = ptr->in(0);
3731 if (allo != NULL && allo->is_Allocate()) {
3732 return allo->as_Allocate();
3733 }
3734 }
3735 // Report failure to match.
3736 return NULL;
3737 }
3738
3739 // Fancy version which also strips off an offset (and reports it to caller).
3740 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3741 intptr_t& offset) {
3742 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3743 if (base == NULL) return NULL;
3744 return Ideal_allocation(base, phase);
3745 }
3746
3747 // Trace Initialize <- Proj[Parm] <- Allocate
3748 AllocateNode* InitializeNode::allocation() {
3749 Node* rawoop = in(InitializeNode::RawAddress);
3750 if (rawoop->is_Proj()) {
3751 Node* alloc = rawoop->in(0);
3752 if (alloc->is_Allocate()) {
3753 return alloc->as_Allocate();
3754 }
3755 }
3756 return NULL;
3757 }
3758
3759 // Trace Allocate -> Proj[Parm] -> Initialize
3760 InitializeNode* AllocateNode::initialization() {
3761 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
3762 if (rawoop == NULL) return NULL;
3763 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3764 Node* init = rawoop->fast_out(i);
3765 if (init->is_Initialize()) {
3766 assert(init->as_Initialize()->allocation() == this, "2-way link");
3767 return init->as_Initialize();
3768 }
3769 }
3770 return NULL;
3771 }
3772
3773 //----------------------------- loop predicates ---------------------------
3774
3775 //------------------------------add_predicate_impl----------------------------
3776 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3777 // Too many traps seen?
3778 if (too_many_traps(reason)) {
3779 #ifdef ASSERT
3780 if (TraceLoopPredicate) {
3781 int tc = C->trap_count(reason);
3782 tty->print("too many traps=%s tcount=%d in ",
3783 Deoptimization::trap_reason_name(reason), tc);
3784 method()->print(); // which method has too many predicate traps
3785 tty->cr();
3786 }
3787 #endif
3788 // We cannot afford to take more traps here,
3789 // do not generate predicate.
3790 return;
3791 }
3792
3793 Node *cont = _gvn.intcon(1);
3794 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
3795 Node *bol = _gvn.transform(new Conv2BNode(opq));
3796 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3797 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3798 C->add_predicate_opaq(opq);
3799 {
3800 PreserveJVMState pjvms(this);
3801 set_control(iffalse);
3802 inc_sp(nargs);
3803 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3804 }
3805 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3806 set_control(iftrue);
3807 }
3808
3809 //------------------------------add_predicate---------------------------------
3810 void GraphKit::add_predicate(int nargs) {
3811 if (UseLoopPredicate) {
3812 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3813 }
3814 // loop's limit check predicate should be near the loop.
3815 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3816 }
3817
3818 //----------------------------- store barriers ----------------------------
3819 #define __ ideal.
3820
3821 bool GraphKit::use_ReduceInitialCardMarks() {
3822 BarrierSet *bs = BarrierSet::barrier_set();
3823 return bs->is_a(BarrierSet::CardTableBarrierSet)
3824 && barrier_set_cast<CardTableBarrierSet>(bs)->can_elide_tlab_store_barriers()
3825 && ReduceInitialCardMarks;
3826 }
3827
3828 void GraphKit::sync_kit(IdealKit& ideal) {
3829 set_all_memory(__ merged_memory());
3830 set_i_o(__ i_o());
3831 set_control(__ ctrl());
3832 }
3833
3834 void GraphKit::final_sync(IdealKit& ideal) {
3835 // Final sync IdealKit and graphKit.
3836 sync_kit(ideal);
3837 }
3838
3839 Node* GraphKit::byte_map_base_node() {
3840 // Get base of card map
3841 jbyte* card_table_base = ci_card_table_address();
3842 if (card_table_base != NULL) {
3843 return makecon(TypeRawPtr::make((address)card_table_base));
3844 } else {
3845 return null();
3846 }
3847 }
3848
3849 // vanilla/CMS post barrier
3850 // Insert a write-barrier store. This is to let generational GC work; we have
3851 // to flag all oop-stores before the next GC point.
3852 void GraphKit::write_barrier_post(Node* oop_store,
3853 Node* obj,
3854 Node* adr,
3855 uint adr_idx,
3856 Node* val,
3857 bool use_precise) {
3858 // No store check needed if we're storing a NULL or an old object
3859 // (latter case is probably a string constant). The concurrent
3860 // mark sweep garbage collector, however, needs to have all nonNull
3861 // oop updates flagged via card-marks.
3862 if (val != NULL && val->is_Con()) {
3863 // must be either an oop or NULL
3864 const Type* t = val->bottom_type();
3865 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3866 // stores of null never (?) need barriers
3867 return;
3868 }
3869
3870 if (use_ReduceInitialCardMarks()
3871 && obj == just_allocated_object(control())) {
3872 // We can skip marks on a freshly-allocated object in Eden.
3873 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
3874 // That routine informs GC to take appropriate compensating steps,
3875 // upon a slow-path allocation, so as to make this card-mark
3876 // elision safe.
3877 return;
3878 }
3879
3880 if (!use_precise) {
3881 // All card marks for a (non-array) instance are in one place:
3882 adr = obj;
3883 }
3884 // (Else it's an array (or unknown), and we want more precise card marks.)
3885 assert(adr != NULL, "");
3886
3887 IdealKit ideal(this, true);
3888
3889 // Convert the pointer to an int prior to doing math on it
3890 Node* cast = __ CastPX(__ ctrl(), adr);
3891
3892 // Divide by card size
3893 assert(BarrierSet::barrier_set()->is_a(BarrierSet::CardTableBarrierSet),
3894 "Only one we handle so far.");
3895 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
3896
3897 // Combine card table base and card offset
3898 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3899
3900 // Get the alias_index for raw card-mark memory
3901 int adr_type = Compile::AliasIdxRaw;
3902 Node* zero = __ ConI(0); // Dirty card value
3903 BasicType bt = T_BYTE;
3904
3905 if (UseConcMarkSweepGC && UseCondCardMark) {
3906 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
3907 __ sync_kit(this);
3908 }
3909
3910 if (UseCondCardMark) {
3911 // The classic GC reference write barrier is typically implemented
3912 // as a store into the global card mark table. Unfortunately
3913 // unconditional stores can result in false sharing and excessive
3914 // coherence traffic as well as false transactional aborts.
3915 // UseCondCardMark enables MP "polite" conditional card mark
3916 // stores. In theory we could relax the load from ctrl() to
3917 // no_ctrl, but that doesn't buy much latitude.
3918 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3919 __ if_then(card_val, BoolTest::ne, zero);
3920 }
3921
3922 // Smash zero into card
3923 if( !UseConcMarkSweepGC ) {
3924 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3925 } else {
3926 // Specialized path for CM store barrier
3927 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3928 }
3929
3930 if (UseCondCardMark) {
3931 __ end_if();
3932 }
3933
3934 // Final sync IdealKit and GraphKit.
3935 final_sync(ideal);
3936 }
3937 /*
3938 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
3939 * required by SATB to make sure all objects live at the start of the
3940 * marking are kept alive, all reference updates need to any previous
3941 * reference stored before writing.
3942 *
3943 * If the previous value is NULL there is no need to save the old value.
3944 * References that are NULL are filtered during runtime by the barrier
3945 * code to avoid unnecessary queuing.
3946 *
3947 * However in the case of newly allocated objects it might be possible to
3948 * prove that the reference about to be overwritten is NULL during compile
3949 * time and avoid adding the barrier code completely.
3950 *
3951 * The compiler needs to determine that the object in which a field is about
3952 * to be written is newly allocated, and that no prior store to the same field
3953 * has happened since the allocation.
3954 *
3955 * Returns true if the pre-barrier can be removed
3956 */
3957 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
3958 BasicType bt, uint adr_idx) {
3959 intptr_t offset = 0;
3960 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3961 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
3962
3963 if (offset == Type::OffsetBot) {
3964 return false; // cannot unalias unless there are precise offsets
3965 }
3966
3967 if (alloc == NULL) {
3968 return false; // No allocation found
3969 }
3970
3971 intptr_t size_in_bytes = type2aelembytes(bt);
3972
3973 Node* mem = memory(adr_idx); // start searching here...
3974
3975 for (int cnt = 0; cnt < 50; cnt++) {
3976
3977 if (mem->is_Store()) {
3978
3979 Node* st_adr = mem->in(MemNode::Address);
3980 intptr_t st_offset = 0;
3981 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
3982
3983 if (st_base == NULL) {
3984 break; // inscrutable pointer
3985 }
3986
3987 // Break we have found a store with same base and offset as ours so break
3988 if (st_base == base && st_offset == offset) {
3989 break;
3990 }
3991
3992 if (st_offset != offset && st_offset != Type::OffsetBot) {
3993 const int MAX_STORE = BytesPerLong;
3994 if (st_offset >= offset + size_in_bytes ||
3995 st_offset <= offset - MAX_STORE ||
3996 st_offset <= offset - mem->as_Store()->memory_size()) {
3997 // Success: The offsets are provably independent.
3998 // (You may ask, why not just test st_offset != offset and be done?
3999 // The answer is that stores of different sizes can co-exist
4000 // in the same sequence of RawMem effects. We sometimes initialize
4001 // a whole 'tile' of array elements with a single jint or jlong.)
4002 mem = mem->in(MemNode::Memory);
4003 continue; // advance through independent store memory
4004 }
4005 }
4006
4007 if (st_base != base
4008 && MemNode::detect_ptr_independence(base, alloc, st_base,
4009 AllocateNode::Ideal_allocation(st_base, phase),
4010 phase)) {
4011 // Success: The bases are provably independent.
4012 mem = mem->in(MemNode::Memory);
4013 continue; // advance through independent store memory
4014 }
4015 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4016
4017 InitializeNode* st_init = mem->in(0)->as_Initialize();
4018 AllocateNode* st_alloc = st_init->allocation();
4019
4020 // Make sure that we are looking at the same allocation site.
4021 // The alloc variable is guaranteed to not be null here from earlier check.
4022 if (alloc == st_alloc) {
4023 // Check that the initialization is storing NULL so that no previous store
4024 // has been moved up and directly write a reference
4025 Node* captured_store = st_init->find_captured_store(offset,
4026 type2aelembytes(T_OBJECT),
4027 phase);
4028 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4029 return true;
4030 }
4031 }
4032 }
4033
4034 // Unless there is an explicit 'continue', we must bail out here,
4035 // because 'mem' is an inscrutable memory state (e.g., a call).
4036 break;
4037 }
4038
4039 return false;
4040 }
4041
4042 // G1 pre/post barriers
4043 void GraphKit::g1_write_barrier_pre(bool do_load,
4044 Node* obj,
4045 Node* adr,
4046 uint alias_idx,
4047 Node* val,
4048 const TypeOopPtr* val_type,
4049 Node* pre_val,
4050 BasicType bt) {
4051
4052 // Some sanity checks
4053 // Note: val is unused in this routine.
4054
4055 if (do_load) {
4056 // We need to generate the load of the previous value
4057 assert(obj != NULL, "must have a base");
4058 assert(adr != NULL, "where are loading from?");
4059 assert(pre_val == NULL, "loaded already?");
4060 assert(val_type != NULL, "need a type");
4061
4062 if (use_ReduceInitialCardMarks()
4063 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4064 return;
4065 }
4066
4067 } else {
4068 // In this case both val_type and alias_idx are unused.
4069 assert(pre_val != NULL, "must be loaded already");
4070 // Nothing to be done if pre_val is null.
4071 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4072 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4073 }
4074 assert(bt == T_OBJECT, "or we shouldn't be here");
4075
4076 IdealKit ideal(this, true);
4077
4078 Node* tls = __ thread(); // ThreadLocalStorage
4079
4080 Node* no_ctrl = NULL;
4081 Node* no_base = __ top();
4082 Node* zero = __ ConI(0);
4083 Node* zeroX = __ ConX(0);
4084
4085 float likely = PROB_LIKELY(0.999);
4086 float unlikely = PROB_UNLIKELY(0.999);
4087
4088 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4089 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4090
4091 // Offsets into the thread
4092 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
4093 const int index_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset());
4094 const int buffer_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
4095
4096 // Now the actual pointers into the thread
4097 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4098 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4099 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4100
4101 // Now some of the values
4102 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4103
4104 // if (!marking)
4105 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4106 BasicType index_bt = TypeX_X->basic_type();
4107 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4108 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4109
4110 if (do_load) {
4111 // load original value
4112 // alias_idx correct??
4113 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4114 }
4115
4116 // if (pre_val != NULL)
4117 __ if_then(pre_val, BoolTest::ne, null()); {
4118 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4119
4120 // is the queue for this thread full?
4121 __ if_then(index, BoolTest::ne, zeroX, likely); {
4122
4123 // decrement the index
4124 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4125
4126 // Now get the buffer location we will log the previous value into and store it
4127 Node *log_addr = __ AddP(no_base, buffer, next_index);
4128 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4129 // update the index
4130 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4131
4132 } __ else_(); {
4133
4134 // logging buffer is full, call the runtime
4135 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4136 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4137 } __ end_if(); // (!index)
4138 } __ end_if(); // (pre_val != NULL)
4139 } __ end_if(); // (!marking)
4140
4141 // Final sync IdealKit and GraphKit.
4142 final_sync(ideal);
4143 }
4144
4145 /*
4146 * G1 similar to any GC with a Young Generation requires a way to keep track of
4147 * references from Old Generation to Young Generation to make sure all live
4148 * objects are found. G1 also requires to keep track of object references
4149 * between different regions to enable evacuation of old regions, which is done
4150 * as part of mixed collections. References are tracked in remembered sets and
4151 * is continuously updated as reference are written to with the help of the
4152 * post-barrier.
4153 *
4154 * To reduce the number of updates to the remembered set the post-barrier
4155 * filters updates to fields in objects located in the Young Generation,
4156 * the same region as the reference, when the NULL is being written or
4157 * if the card is already marked as dirty by an earlier write.
4158 *
4159 * Under certain circumstances it is possible to avoid generating the
4160 * post-barrier completely if it is possible during compile time to prove
4161 * the object is newly allocated and that no safepoint exists between the
4162 * allocation and the store.
4163 *
4164 * In the case of slow allocation the allocation code must handle the barrier
4165 * as part of the allocation in the case the allocated object is not located
4166 * in the nursery, this would happen for humongous objects. This is similar to
4167 * how CMS is required to handle this case, see the comments for the method
4168 * CardTableBarrierSet::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier.
4169 * A deferred card mark is required for these objects and handled in the above
4170 * mentioned methods.
4171 *
4172 * Returns true if the post barrier can be removed
4173 */
4174 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4175 Node* adr) {
4176 intptr_t offset = 0;
4177 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4178 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4179
4180 if (offset == Type::OffsetBot) {
4181 return false; // cannot unalias unless there are precise offsets
4182 }
4183
4184 if (alloc == NULL) {
4185 return false; // No allocation found
4186 }
4187
4188 // Start search from Store node
4189 Node* mem = store->in(MemNode::Control);
4190 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4191
4192 InitializeNode* st_init = mem->in(0)->as_Initialize();
4193 AllocateNode* st_alloc = st_init->allocation();
4194
4195 // Make sure we are looking at the same allocation
4196 if (alloc == st_alloc) {
4197 return true;
4198 }
4199 }
4200
4201 return false;
4202 }
4203
4204 //
4205 // Update the card table and add card address to the queue
4206 //
4207 void GraphKit::g1_mark_card(IdealKit& ideal,
4208 Node* card_adr,
4209 Node* oop_store,
4210 uint oop_alias_idx,
4211 Node* index,
4212 Node* index_adr,
4213 Node* buffer,
4214 const TypeFunc* tf) {
4215
4216 Node* zero = __ ConI(0);
4217 Node* zeroX = __ ConX(0);
4218 Node* no_base = __ top();
4219 BasicType card_bt = T_BYTE;
4220 // Smash zero into card. MUST BE ORDERED WRT TO STORE
4221 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4222
4223 // Now do the queue work
4224 __ if_then(index, BoolTest::ne, zeroX); {
4225
4226 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4227 Node* log_addr = __ AddP(no_base, buffer, next_index);
4228
4229 // Order, see storeCM.
4230 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4231 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4232
4233 } __ else_(); {
4234 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4235 } __ end_if();
4236
4237 }
4238
4239 void GraphKit::g1_write_barrier_post(Node* oop_store,
4240 Node* obj,
4241 Node* adr,
4242 uint alias_idx,
4243 Node* val,
4244 BasicType bt,
4245 bool use_precise) {
4246 // If we are writing a NULL then we need no post barrier
4247
4248 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4249 // Must be NULL
4250 const Type* t = val->bottom_type();
4251 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4252 // No post barrier if writing NULLx
4253 return;
4254 }
4255
4256 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4257 // We can skip marks on a freshly-allocated object in Eden.
4258 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
4259 // That routine informs GC to take appropriate compensating steps,
4260 // upon a slow-path allocation, so as to make this card-mark
4261 // elision safe.
4262 return;
4263 }
4264
4265 if (use_ReduceInitialCardMarks()
4266 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4267 return;
4268 }
4269
4270 if (!use_precise) {
4271 // All card marks for a (non-array) instance are in one place:
4272 adr = obj;
4273 }
4274 // (Else it's an array (or unknown), and we want more precise card marks.)
4275 assert(adr != NULL, "");
4276
4277 IdealKit ideal(this, true);
4278
4279 Node* tls = __ thread(); // ThreadLocalStorage
4280
4281 Node* no_base = __ top();
4282 float likely = PROB_LIKELY(0.999);
4283 float unlikely = PROB_UNLIKELY(0.999);
4284 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val());
4285 Node* dirty_card = __ ConI((jint)CardTable::dirty_card_val());
4286 Node* zeroX = __ ConX(0);
4287
4288 // Get the alias_index for raw card-mark memory
4289 const TypePtr* card_type = TypeRawPtr::BOTTOM;
4290
4291 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4292
4293 // Offsets into the thread
4294 const int index_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset());
4295 const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset());
4296
4297 // Pointers into the thread
4298
4299 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4300 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4301
4302 // Now some values
4303 // Use ctrl to avoid hoisting these values past a safepoint, which could
4304 // potentially reset these fields in the JavaThread.
4305 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4306 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4307
4308 // Convert the store obj pointer to an int prior to doing math on it
4309 // Must use ctrl to prevent "integerized oop" existing across safepoint
4310 Node* cast = __ CastPX(__ ctrl(), adr);
4311
4312 // Divide pointer by card size
4313 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
4314
4315 // Combine card table base and card offset
4316 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4317
4318 // If we know the value being stored does it cross regions?
4319
4320 if (val != NULL) {
4321 // Does the store cause us to cross regions?
4322
4323 // Should be able to do an unsigned compare of region_size instead of
4324 // and extra shift. Do we have an unsigned compare??
4325 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4326 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4327
4328 // if (xor_res == 0) same region so skip
4329 __ if_then(xor_res, BoolTest::ne, zeroX); {
4330
4331 // No barrier if we are storing a NULL
4332 __ if_then(val, BoolTest::ne, null(), unlikely); {
4333
4334 // Ok must mark the card if not already dirty
4335
4336 // load the original value of the card
4337 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4338
4339 __ if_then(card_val, BoolTest::ne, young_card); {
4340 sync_kit(ideal);
4341 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4342 insert_mem_bar(Op_MemBarVolatile, oop_store);
4343 __ sync_kit(this);
4344
4345 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4346 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4347 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4348 } __ end_if();
4349 } __ end_if();
4350 } __ end_if();
4351 } __ end_if();
4352 } else {
4353 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
4354 // We don't need a barrier here if the destination is a newly allocated object
4355 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
4356 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
4357 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
4358 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4359 __ if_then(card_val, BoolTest::ne, young_card); {
4360 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4361 } __ end_if();
4362 }
4363
4364 // Final sync IdealKit and GraphKit.
4365 final_sync(ideal);
4366 }
4367 #undef __
4368
4369
4370 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4371 Node* len = load_array_length(load_String_value(ctrl, str));
4372 Node* coder = load_String_coder(ctrl, str);
4373 // Divide length by 2 if coder is UTF16
4374 return _gvn.transform(new RShiftINode(len, coder));
4375 }
4376
4377 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4378 int value_offset = java_lang_String::value_offset_in_bytes();
4379 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4380 false, NULL, 0);
4381 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4382 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4383 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4384 ciTypeArrayKlass::make(T_BYTE), true, 0);
4385 int value_field_idx = C->get_alias_index(value_field_type);
4386 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4387 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4388 // String.value field is known to be @Stable.
4389 if (UseImplicitStableValues) {
4390 load = cast_array_to_stable(load, value_type);
4391 }
4392 return load;
4393 }
4394
4395 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4396 if (!CompactStrings) {
4397 return intcon(java_lang_String::CODER_UTF16);
4398 }
4399 int coder_offset = java_lang_String::coder_offset_in_bytes();
4400 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4401 false, NULL, 0);
4402 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4403 int coder_field_idx = C->get_alias_index(coder_field_type);
4404 return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4405 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4406 }
4407
4408 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4409 int value_offset = java_lang_String::value_offset_in_bytes();
4410 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4411 false, NULL, 0);
4412 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4413 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4414 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4415 }
4416
4417 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4418 int coder_offset = java_lang_String::coder_offset_in_bytes();
4419 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4420 false, NULL, 0);
4421 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4422 int coder_field_idx = C->get_alias_index(coder_field_type);
4423 store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4424 value, T_BYTE, coder_field_idx, MemNode::unordered);
4425 }
4426
4427 // Capture src and dst memory state with a MergeMemNode
4428 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4429 if (src_type == dst_type) {
4430 // Types are equal, we don't need a MergeMemNode
4431 return memory(src_type);
4432 }
4433 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4434 record_for_igvn(merge); // fold it up later, if possible
4435 int src_idx = C->get_alias_index(src_type);
4436 int dst_idx = C->get_alias_index(dst_type);
4437 merge->set_memory_at(src_idx, memory(src_idx));
4438 merge->set_memory_at(dst_idx, memory(dst_idx));
4439 return merge;
4440 }
4441
4442 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4443 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4444 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4445 // If input and output memory types differ, capture both states to preserve
4446 // the dependency between preceding and subsequent loads/stores.
4447 // For example, the following program:
4448 // StoreB
4449 // compress_string
4450 // LoadB
4451 // has this memory graph (use->def):
4452 // LoadB -> compress_string -> CharMem
4453 // ... -> StoreB -> ByteMem
4454 // The intrinsic hides the dependency between LoadB and StoreB, causing
4455 // the load to read from memory not containing the result of the StoreB.
4456 // The correct memory graph should look like this:
4457 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4458 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4459 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4460 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4461 set_memory(res_mem, TypeAryPtr::BYTES);
4462 return str;
4463 }
4464
4465 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4466 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4467 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4468 // Capture src and dst memory (see comment in 'compress_string').
4469 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4470 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4471 set_memory(_gvn.transform(str), dst_type);
4472 }
4473
4474 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4475 /**
4476 * int i_char = start;
4477 * for (int i_byte = 0; i_byte < count; i_byte++) {
4478 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4479 * }
4480 */
4481 add_predicate();
4482 RegionNode* head = new RegionNode(3);
4483 head->init_req(1, control());
4484 gvn().set_type(head, Type::CONTROL);
4485 record_for_igvn(head);
4486
4487 Node* i_byte = new PhiNode(head, TypeInt::INT);
4488 i_byte->init_req(1, intcon(0));
4489 gvn().set_type(i_byte, TypeInt::INT);
4490 record_for_igvn(i_byte);
4491
4492 Node* i_char = new PhiNode(head, TypeInt::INT);
4493 i_char->init_req(1, start);
4494 gvn().set_type(i_char, TypeInt::INT);
4495 record_for_igvn(i_char);
4496
4497 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4498 gvn().set_type(mem, Type::MEMORY);
4499 record_for_igvn(mem);
4500 set_control(head);
4501 set_memory(mem, TypeAryPtr::BYTES);
4502 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4503 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4504 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4505 false, false, true /* mismatched */);
4506
4507 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4508 head->init_req(2, IfTrue(iff));
4509 mem->init_req(2, st);
4510 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4511 i_char->init_req(2, AddI(i_char, intcon(2)));
4512
4513 set_control(IfFalse(iff));
4514 set_memory(st, TypeAryPtr::BYTES);
4515 }
4516
4517 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4518 if (!field->is_constant()) {
4519 return NULL; // Field not marked as constant.
4520 }
4521 ciInstance* holder = NULL;
4522 if (!field->is_static()) {
4523 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4524 if (const_oop != NULL && const_oop->is_instance()) {
4525 holder = const_oop->as_instance();
4526 }
4527 }
4528 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4529 /*is_unsigned_load=*/false);
4530 if (con_type != NULL) {
4531 return makecon(con_type);
4532 }
4533 return NULL;
4534 }
4535
4536 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4537 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4538 // assumption of CCP analysis.
4539 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4540 }