1 /*
2 * Copyright (c) 2001, 2022, 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/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciUtilities.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "ci/ciObjArray.hpp"
31 #include "asm/register.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.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/inlinetypenode.hpp"
43 #include "opto/intrinsicnode.hpp"
44 #include "opto/locknode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "runtime/deoptimization.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "utilities/bitMap.inline.hpp"
55 #include "utilities/powerOfTwo.hpp"
56 #include "utilities/growableArray.hpp"
57
58 //----------------------------GraphKit-----------------------------------------
59 // Main utility constructor.
60 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
61 : Phase(Phase::Parser),
62 _env(C->env()),
63 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()),
64 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
65 {
66 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
67 _exceptions = jvms->map()->next_exception();
68 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
69 set_jvms(jvms);
70 #ifdef ASSERT
71 if (_gvn.is_IterGVN() != NULL) {
72 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
73 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
74 _worklist_size = _gvn.C->for_igvn()->size();
75 }
76 #endif
77 }
78
79 // Private constructor for parser.
80 GraphKit::GraphKit()
81 : Phase(Phase::Parser),
82 _env(C->env()),
83 _gvn(*C->initial_gvn()),
84 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
85 {
86 _exceptions = NULL;
87 set_map(NULL);
88 debug_only(_sp = -99);
89 debug_only(set_bci(-99));
90 }
91
92
93
94 //---------------------------clean_stack---------------------------------------
95 // Clear away rubbish from the stack area of the JVM state.
96 // This destroys any arguments that may be waiting on the stack.
97 void GraphKit::clean_stack(int from_sp) {
98 SafePointNode* map = this->map();
99 JVMState* jvms = this->jvms();
100 int stk_size = jvms->stk_size();
101 int stkoff = jvms->stkoff();
102 Node* top = this->top();
103 for (int i = from_sp; i < stk_size; i++) {
104 if (map->in(stkoff + i) != top) {
105 map->set_req(stkoff + i, top);
106 }
107 }
108 }
109
110
111 //--------------------------------sync_jvms-----------------------------------
112 // Make sure our current jvms agrees with our parse state.
113 JVMState* GraphKit::sync_jvms() const {
114 JVMState* jvms = this->jvms();
115 jvms->set_bci(bci()); // Record the new bci in the JVMState
116 jvms->set_sp(sp()); // Record the new sp in the JVMState
117 assert(jvms_in_sync(), "jvms is now in sync");
118 return jvms;
119 }
120
121 //--------------------------------sync_jvms_for_reexecute---------------------
122 // Make sure our current jvms agrees with our parse state. This version
123 // uses the reexecute_sp for reexecuting bytecodes.
124 JVMState* GraphKit::sync_jvms_for_reexecute() {
125 JVMState* jvms = this->jvms();
126 jvms->set_bci(bci()); // Record the new bci in the JVMState
127 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
128 return jvms;
129 }
130
131 #ifdef ASSERT
132 bool GraphKit::jvms_in_sync() const {
133 Parse* parse = is_Parse();
134 if (parse == NULL) {
135 if (bci() != jvms()->bci()) return false;
136 if (sp() != (int)jvms()->sp()) return false;
137 return true;
138 }
139 if (jvms()->method() != parse->method()) return false;
140 if (jvms()->bci() != parse->bci()) return false;
141 int jvms_sp = jvms()->sp();
142 if (jvms_sp != parse->sp()) return false;
143 int jvms_depth = jvms()->depth();
144 if (jvms_depth != parse->depth()) return false;
145 return true;
146 }
147
148 // Local helper checks for special internal merge points
149 // used to accumulate and merge exception states.
150 // They are marked by the region's in(0) edge being the map itself.
151 // Such merge points must never "escape" into the parser at large,
152 // until they have been handed to gvn.transform.
153 static bool is_hidden_merge(Node* reg) {
154 if (reg == NULL) return false;
155 if (reg->is_Phi()) {
156 reg = reg->in(0);
157 if (reg == NULL) return false;
158 }
159 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
160 }
161
162 void GraphKit::verify_map() const {
163 if (map() == NULL) return; // null map is OK
164 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
165 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
166 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
167 }
168
169 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
170 assert(ex_map->next_exception() == NULL, "not already part of a chain");
171 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
172 }
173 #endif
174
175 //---------------------------stop_and_kill_map---------------------------------
176 // Set _map to NULL, signalling a stop to further bytecode execution.
177 // First smash the current map's control to a constant, to mark it dead.
178 void GraphKit::stop_and_kill_map() {
179 SafePointNode* dead_map = stop();
180 if (dead_map != NULL) {
181 dead_map->disconnect_inputs(C); // Mark the map as killed.
182 assert(dead_map->is_killed(), "must be so marked");
183 }
184 }
185
186
187 //--------------------------------stopped--------------------------------------
188 // Tell if _map is NULL, or control is top.
189 bool GraphKit::stopped() {
190 if (map() == NULL) return true;
191 else if (control() == top()) return true;
192 else return false;
193 }
194
195
196 //-----------------------------has_ex_handler----------------------------------
197 // Tell if this method or any caller method has exception handlers.
198 bool GraphKit::has_ex_handler() {
199 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
200 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
201 return true;
202 }
203 }
204 return false;
205 }
206
207 //------------------------------save_ex_oop------------------------------------
208 // Save an exception without blowing stack contents or other JVM state.
209 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
210 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
211 ex_map->add_req(ex_oop);
212 debug_only(verify_exception_state(ex_map));
213 }
214
215 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
216 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
217 Node* ex_oop = ex_map->in(ex_map->req()-1);
218 if (clear_it) ex_map->del_req(ex_map->req()-1);
219 return ex_oop;
220 }
221
222 //-----------------------------saved_ex_oop------------------------------------
223 // Recover a saved exception from its map.
224 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
225 return common_saved_ex_oop(ex_map, false);
226 }
227
228 //--------------------------clear_saved_ex_oop---------------------------------
229 // Erase a previously saved exception from its map.
230 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
231 return common_saved_ex_oop(ex_map, true);
232 }
233
234 #ifdef ASSERT
235 //---------------------------has_saved_ex_oop----------------------------------
236 // Erase a previously saved exception from its map.
237 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
238 return ex_map->req() == ex_map->jvms()->endoff()+1;
239 }
240 #endif
241
242 //-------------------------make_exception_state--------------------------------
243 // Turn the current JVM state into an exception state, appending the ex_oop.
244 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
245 sync_jvms();
246 SafePointNode* ex_map = stop(); // do not manipulate this map any more
247 set_saved_ex_oop(ex_map, ex_oop);
248 return ex_map;
249 }
250
251
252 //--------------------------add_exception_state--------------------------------
253 // Add an exception to my list of exceptions.
254 void GraphKit::add_exception_state(SafePointNode* ex_map) {
255 if (ex_map == NULL || ex_map->control() == top()) {
256 return;
257 }
258 #ifdef ASSERT
259 verify_exception_state(ex_map);
260 if (has_exceptions()) {
261 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
262 }
263 #endif
264
265 // If there is already an exception of exactly this type, merge with it.
266 // In particular, null-checks and other low-level exceptions common up here.
267 Node* ex_oop = saved_ex_oop(ex_map);
268 const Type* ex_type = _gvn.type(ex_oop);
269 if (ex_oop == top()) {
270 // No action needed.
271 return;
272 }
273 assert(ex_type->isa_instptr(), "exception must be an instance");
274 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
275 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
276 // We check sp also because call bytecodes can generate exceptions
277 // both before and after arguments are popped!
278 if (ex_type2 == ex_type
279 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
280 combine_exception_states(ex_map, e2);
281 return;
282 }
283 }
284
285 // No pre-existing exception of the same type. Chain it on the list.
286 push_exception_state(ex_map);
287 }
288
289 //-----------------------add_exception_states_from-----------------------------
290 void GraphKit::add_exception_states_from(JVMState* jvms) {
291 SafePointNode* ex_map = jvms->map()->next_exception();
292 if (ex_map != NULL) {
293 jvms->map()->set_next_exception(NULL);
294 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
295 next_map = ex_map->next_exception();
296 ex_map->set_next_exception(NULL);
297 add_exception_state(ex_map);
298 }
299 }
300 }
301
302 //-----------------------transfer_exceptions_into_jvms-------------------------
303 JVMState* GraphKit::transfer_exceptions_into_jvms() {
304 if (map() == NULL) {
305 // We need a JVMS to carry the exceptions, but the map has gone away.
306 // Create a scratch JVMS, cloned from any of the exception states...
307 if (has_exceptions()) {
308 _map = _exceptions;
309 _map = clone_map();
310 _map->set_next_exception(NULL);
311 clear_saved_ex_oop(_map);
312 debug_only(verify_map());
313 } else {
314 // ...or created from scratch
315 JVMState* jvms = new (C) JVMState(_method, NULL);
316 jvms->set_bci(_bci);
317 jvms->set_sp(_sp);
318 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
319 set_jvms(jvms);
320 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
321 set_all_memory(top());
322 while (map()->req() < jvms->endoff()) map()->add_req(top());
323 }
324 // (This is a kludge, in case you didn't notice.)
325 set_control(top());
326 }
327 JVMState* jvms = sync_jvms();
328 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
329 jvms->map()->set_next_exception(_exceptions);
330 _exceptions = NULL; // done with this set of exceptions
331 return jvms;
332 }
333
334 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
335 assert(is_hidden_merge(dstphi), "must be a special merge node");
336 assert(is_hidden_merge(srcphi), "must be a special merge node");
337 uint limit = srcphi->req();
338 for (uint i = PhiNode::Input; i < limit; i++) {
339 dstphi->add_req(srcphi->in(i));
340 }
341 }
342 static inline void add_one_req(Node* dstphi, Node* src) {
343 assert(is_hidden_merge(dstphi), "must be a special merge node");
344 assert(!is_hidden_merge(src), "must not be a special merge node");
345 dstphi->add_req(src);
346 }
347
348 //-----------------------combine_exception_states------------------------------
349 // This helper function combines exception states by building phis on a
350 // specially marked state-merging region. These regions and phis are
351 // untransformed, and can build up gradually. The region is marked by
352 // having a control input of its exception map, rather than NULL. Such
353 // regions do not appear except in this function, and in use_exception_state.
354 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
355 if (failing()) return; // dying anyway...
356 JVMState* ex_jvms = ex_map->_jvms;
357 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
358 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
359 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
360 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
361 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
362 assert(ex_map->req() == phi_map->req(), "matching maps");
363 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
364 Node* hidden_merge_mark = root();
365 Node* region = phi_map->control();
366 MergeMemNode* phi_mem = phi_map->merged_memory();
367 MergeMemNode* ex_mem = ex_map->merged_memory();
368 if (region->in(0) != hidden_merge_mark) {
369 // The control input is not (yet) a specially-marked region in phi_map.
370 // Make it so, and build some phis.
371 region = new RegionNode(2);
372 _gvn.set_type(region, Type::CONTROL);
373 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
374 region->init_req(1, phi_map->control());
375 phi_map->set_control(region);
376 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
377 record_for_igvn(io_phi);
378 _gvn.set_type(io_phi, Type::ABIO);
379 phi_map->set_i_o(io_phi);
380 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
381 Node* m = mms.memory();
382 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
383 record_for_igvn(m_phi);
384 _gvn.set_type(m_phi, Type::MEMORY);
385 mms.set_memory(m_phi);
386 }
387 }
388
389 // Either or both of phi_map and ex_map might already be converted into phis.
390 Node* ex_control = ex_map->control();
391 // if there is special marking on ex_map also, we add multiple edges from src
392 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
393 // how wide was the destination phi_map, originally?
394 uint orig_width = region->req();
395
396 if (add_multiple) {
397 add_n_reqs(region, ex_control);
398 add_n_reqs(phi_map->i_o(), ex_map->i_o());
399 } else {
400 // ex_map has no merges, so we just add single edges everywhere
401 add_one_req(region, ex_control);
402 add_one_req(phi_map->i_o(), ex_map->i_o());
403 }
404 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
405 if (mms.is_empty()) {
406 // get a copy of the base memory, and patch some inputs into it
407 const TypePtr* adr_type = mms.adr_type(C);
408 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
409 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
410 mms.set_memory(phi);
411 // Prepare to append interesting stuff onto the newly sliced phi:
412 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
413 }
414 // Append stuff from ex_map:
415 if (add_multiple) {
416 add_n_reqs(mms.memory(), mms.memory2());
417 } else {
418 add_one_req(mms.memory(), mms.memory2());
419 }
420 }
421 uint limit = ex_map->req();
422 for (uint i = TypeFunc::Parms; i < limit; i++) {
423 // Skip everything in the JVMS after tos. (The ex_oop follows.)
424 if (i == tos) i = ex_jvms->monoff();
425 Node* src = ex_map->in(i);
426 Node* dst = phi_map->in(i);
427 if (src != dst) {
428 PhiNode* phi;
429 if (dst->in(0) != region) {
430 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
431 record_for_igvn(phi);
432 _gvn.set_type(phi, phi->type());
433 phi_map->set_req(i, dst);
434 // Prepare to append interesting stuff onto the new phi:
435 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
436 } else {
437 assert(dst->is_Phi(), "nobody else uses a hidden region");
438 phi = dst->as_Phi();
439 }
440 if (add_multiple && src->in(0) == ex_control) {
441 // Both are phis.
442 add_n_reqs(dst, src);
443 } else {
444 while (dst->req() < region->req()) add_one_req(dst, src);
445 }
446 const Type* srctype = _gvn.type(src);
447 if (phi->type() != srctype) {
448 const Type* dsttype = phi->type()->meet_speculative(srctype);
449 if (phi->type() != dsttype) {
450 phi->set_type(dsttype);
451 _gvn.set_type(phi, dsttype);
452 }
453 }
454 }
455 }
456 phi_map->merge_replaced_nodes_with(ex_map);
457 }
458
459 //--------------------------use_exception_state--------------------------------
460 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
461 if (failing()) { stop(); return top(); }
462 Node* region = phi_map->control();
463 Node* hidden_merge_mark = root();
464 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
465 Node* ex_oop = clear_saved_ex_oop(phi_map);
466 if (region->in(0) == hidden_merge_mark) {
467 // Special marking for internal ex-states. Process the phis now.
468 region->set_req(0, region); // now it's an ordinary region
469 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
470 // Note: Setting the jvms also sets the bci and sp.
471 set_control(_gvn.transform(region));
472 uint tos = jvms()->stkoff() + sp();
473 for (uint i = 1; i < tos; i++) {
474 Node* x = phi_map->in(i);
475 if (x->in(0) == region) {
476 assert(x->is_Phi(), "expected a special phi");
477 phi_map->set_req(i, _gvn.transform(x));
478 }
479 }
480 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
481 Node* x = mms.memory();
482 if (x->in(0) == region) {
483 assert(x->is_Phi(), "nobody else uses a hidden region");
484 mms.set_memory(_gvn.transform(x));
485 }
486 }
487 if (ex_oop->in(0) == region) {
488 assert(ex_oop->is_Phi(), "expected a special phi");
489 ex_oop = _gvn.transform(ex_oop);
490 }
491 } else {
492 set_jvms(phi_map->jvms());
493 }
494
495 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
496 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
497 return ex_oop;
498 }
499
500 //---------------------------------java_bc-------------------------------------
501 Bytecodes::Code GraphKit::java_bc() const {
502 ciMethod* method = this->method();
503 int bci = this->bci();
504 if (method != NULL && bci != InvocationEntryBci)
505 return method->java_code_at_bci(bci);
506 else
507 return Bytecodes::_illegal;
508 }
509
510 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
511 bool must_throw) {
512 // if the exception capability is set, then we will generate code
513 // to check the JavaThread.should_post_on_exceptions flag to see
514 // if we actually need to report exception events (for this
515 // thread). If we don't need to report exception events, we will
516 // take the normal fast path provided by add_exception_events. If
517 // exception event reporting is enabled for this thread, we will
518 // take the uncommon_trap in the BuildCutout below.
519
520 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
521 Node* jthread = _gvn.transform(new ThreadLocalNode());
522 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
523 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
524
525 // Test the should_post_on_exceptions_flag vs. 0
526 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
527 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
528
529 // Branch to slow_path if should_post_on_exceptions_flag was true
530 { BuildCutout unless(this, tst, PROB_MAX);
531 // Do not try anything fancy if we're notifying the VM on every throw.
532 // Cf. case Bytecodes::_athrow in parse2.cpp.
533 uncommon_trap(reason, Deoptimization::Action_none,
534 (ciKlass*)NULL, (char*)NULL, must_throw);
535 }
536
537 }
538
539 //------------------------------builtin_throw----------------------------------
540 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) {
541 bool must_throw = true;
542
543 // If this particular condition has not yet happened at this
544 // bytecode, then use the uncommon trap mechanism, and allow for
545 // a future recompilation if several traps occur here.
546 // If the throw is hot, try to use a more complicated inline mechanism
547 // which keeps execution inside the compiled code.
548 bool treat_throw_as_hot = false;
549 ciMethodData* md = method()->method_data();
550
551 if (ProfileTraps) {
552 if (too_many_traps(reason)) {
553 treat_throw_as_hot = true;
554 }
555 // (If there is no MDO at all, assume it is early in
556 // execution, and that any deopts are part of the
557 // startup transient, and don't need to be remembered.)
558
559 // Also, if there is a local exception handler, treat all throws
560 // as hot if there has been at least one in this method.
561 if (C->trap_count(reason) != 0
562 && method()->method_data()->trap_count(reason) != 0
563 && has_ex_handler()) {
564 treat_throw_as_hot = true;
565 }
566 }
567
568 // If this throw happens frequently, an uncommon trap might cause
569 // a performance pothole. If there is a local exception handler,
570 // and if this particular bytecode appears to be deoptimizing often,
571 // let us handle the throw inline, with a preconstructed instance.
572 // Note: If the deopt count has blown up, the uncommon trap
573 // runtime is going to flush this nmethod, not matter what.
574 if (treat_throw_as_hot && method()->can_omit_stack_trace()) {
575 // If the throw is local, we use a pre-existing instance and
576 // punt on the backtrace. This would lead to a missing backtrace
577 // (a repeat of 4292742) if the backtrace object is ever asked
578 // for its backtrace.
579 // Fixing this remaining case of 4292742 requires some flavor of
580 // escape analysis. Leave that for the future.
581 ciInstance* ex_obj = NULL;
582 switch (reason) {
583 case Deoptimization::Reason_null_check:
584 ex_obj = env()->NullPointerException_instance();
585 break;
586 case Deoptimization::Reason_div0_check:
587 ex_obj = env()->ArithmeticException_instance();
588 break;
589 case Deoptimization::Reason_range_check:
590 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
591 break;
592 case Deoptimization::Reason_class_check:
593 ex_obj = env()->ClassCastException_instance();
594 break;
595 case Deoptimization::Reason_array_check:
596 ex_obj = env()->ArrayStoreException_instance();
597 break;
598 default:
599 break;
600 }
601 if (failing()) { stop(); return; } // exception allocation might fail
602 if (ex_obj != NULL) {
603 if (env()->jvmti_can_post_on_exceptions()) {
604 // check if we must post exception events, take uncommon trap if so
605 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
606 // here if should_post_on_exceptions is false
607 // continue on with the normal codegen
608 }
609
610 // Cheat with a preallocated exception object.
611 if (C->log() != NULL)
612 C->log()->elem("hot_throw preallocated='1' reason='%s'",
613 Deoptimization::trap_reason_name(reason));
614 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
615 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
616
617 // Clear the detail message of the preallocated exception object.
618 // Weblogic sometimes mutates the detail message of exceptions
619 // using reflection.
620 int offset = java_lang_Throwable::get_detailMessage_offset();
621 const TypePtr* adr_typ = ex_con->add_offset(offset);
622
623 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
624 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
625 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
626
627 if (!method()->has_exception_handlers()) {
628 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway.
629 // This prevents issues with exception handling and late inlining.
630 set_sp(0);
631 clean_stack(0);
632 }
633
634 add_exception_state(make_exception_state(ex_node));
635 return;
636 }
637 }
638
639 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
640 // It won't be much cheaper than bailing to the interp., since we'll
641 // have to pass up all the debug-info, and the runtime will have to
642 // create the stack trace.
643
644 // Usual case: Bail to interpreter.
645 // Reserve the right to recompile if we haven't seen anything yet.
646
647 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
648 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
649 if (treat_throw_as_hot
650 && (method()->method_data()->trap_recompiled_at(bci(), m)
651 || C->too_many_traps(reason))) {
652 // We cannot afford to take more traps here. Suffer in the interpreter.
653 if (C->log() != NULL)
654 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
655 Deoptimization::trap_reason_name(reason),
656 C->trap_count(reason));
657 action = Deoptimization::Action_none;
658 }
659
660 // "must_throw" prunes the JVM state to include only the stack, if there
661 // are no local exception handlers. This should cut down on register
662 // allocation time and code size, by drastically reducing the number
663 // of in-edges on the call to the uncommon trap.
664
665 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
666 }
667
668
669 //----------------------------PreserveJVMState---------------------------------
670 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
671 debug_only(kit->verify_map());
672 _kit = kit;
673 _map = kit->map(); // preserve the map
674 _sp = kit->sp();
675 kit->set_map(clone_map ? kit->clone_map() : NULL);
676 #ifdef ASSERT
677 _bci = kit->bci();
678 Parse* parser = kit->is_Parse();
679 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
680 _block = block;
681 #endif
682 }
683 PreserveJVMState::~PreserveJVMState() {
684 GraphKit* kit = _kit;
685 #ifdef ASSERT
686 assert(kit->bci() == _bci, "bci must not shift");
687 Parse* parser = kit->is_Parse();
688 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
689 assert(block == _block, "block must not shift");
690 #endif
691 kit->set_map(_map);
692 kit->set_sp(_sp);
693 }
694
695
696 //-----------------------------BuildCutout-------------------------------------
697 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
698 : PreserveJVMState(kit)
699 {
700 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
701 SafePointNode* outer_map = _map; // preserved map is caller's
702 SafePointNode* inner_map = kit->map();
703 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
704 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
705 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
706 }
707 BuildCutout::~BuildCutout() {
708 GraphKit* kit = _kit;
709 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
710 }
711
712 //---------------------------PreserveReexecuteState----------------------------
713 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
714 assert(!kit->stopped(), "must call stopped() before");
715 _kit = kit;
716 _sp = kit->sp();
717 _reexecute = kit->jvms()->_reexecute;
718 }
719 PreserveReexecuteState::~PreserveReexecuteState() {
720 if (_kit->stopped()) return;
721 _kit->jvms()->_reexecute = _reexecute;
722 _kit->set_sp(_sp);
723 }
724
725 //------------------------------clone_map--------------------------------------
726 // Implementation of PreserveJVMState
727 //
728 // Only clone_map(...) here. If this function is only used in the
729 // PreserveJVMState class we may want to get rid of this extra
730 // function eventually and do it all there.
731
732 SafePointNode* GraphKit::clone_map() {
733 if (map() == NULL) return NULL;
734
735 // Clone the memory edge first
736 Node* mem = MergeMemNode::make(map()->memory());
737 gvn().set_type_bottom(mem);
738
739 SafePointNode *clonemap = (SafePointNode*)map()->clone();
740 JVMState* jvms = this->jvms();
741 JVMState* clonejvms = jvms->clone_shallow(C);
742 clonemap->set_memory(mem);
743 clonemap->set_jvms(clonejvms);
744 clonejvms->set_map(clonemap);
745 record_for_igvn(clonemap);
746 gvn().set_type_bottom(clonemap);
747 return clonemap;
748 }
749
750
751 //-----------------------------set_map_clone-----------------------------------
752 void GraphKit::set_map_clone(SafePointNode* m) {
753 _map = m;
754 _map = clone_map();
755 _map->set_next_exception(NULL);
756 debug_only(verify_map());
757 }
758
759
760 //----------------------------kill_dead_locals---------------------------------
761 // Detect any locals which are known to be dead, and force them to top.
762 void GraphKit::kill_dead_locals() {
763 // Consult the liveness information for the locals. If any
764 // of them are unused, then they can be replaced by top(). This
765 // should help register allocation time and cut down on the size
766 // of the deoptimization information.
767
768 // This call is made from many of the bytecode handling
769 // subroutines called from the Big Switch in do_one_bytecode.
770 // Every bytecode which might include a slow path is responsible
771 // for killing its dead locals. The more consistent we
772 // are about killing deads, the fewer useless phis will be
773 // constructed for them at various merge points.
774
775 // bci can be -1 (InvocationEntryBci). We return the entry
776 // liveness for the method.
777
778 if (method() == NULL || method()->code_size() == 0) {
779 // We are building a graph for a call to a native method.
780 // All locals are live.
781 return;
782 }
783
784 ResourceMark rm;
785
786 // Consult the liveness information for the locals. If any
787 // of them are unused, then they can be replaced by top(). This
788 // should help register allocation time and cut down on the size
789 // of the deoptimization information.
790 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
791
792 int len = (int)live_locals.size();
793 assert(len <= jvms()->loc_size(), "too many live locals");
794 for (int local = 0; local < len; local++) {
795 if (!live_locals.at(local)) {
796 set_local(local, top());
797 }
798 }
799 }
800
801 #ifdef ASSERT
802 //-------------------------dead_locals_are_killed------------------------------
803 // Return true if all dead locals are set to top in the map.
804 // Used to assert "clean" debug info at various points.
805 bool GraphKit::dead_locals_are_killed() {
806 if (method() == NULL || method()->code_size() == 0) {
807 // No locals need to be dead, so all is as it should be.
808 return true;
809 }
810
811 // Make sure somebody called kill_dead_locals upstream.
812 ResourceMark rm;
813 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
814 if (jvms->loc_size() == 0) continue; // no locals to consult
815 SafePointNode* map = jvms->map();
816 ciMethod* method = jvms->method();
817 int bci = jvms->bci();
818 if (jvms == this->jvms()) {
819 bci = this->bci(); // it might not yet be synched
820 }
821 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
822 int len = (int)live_locals.size();
823 if (!live_locals.is_valid() || len == 0)
824 // This method is trivial, or is poisoned by a breakpoint.
825 return true;
826 assert(len == jvms->loc_size(), "live map consistent with locals map");
827 for (int local = 0; local < len; local++) {
828 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
829 if (PrintMiscellaneous && (Verbose || WizardMode)) {
830 tty->print_cr("Zombie local %d: ", local);
831 jvms->dump();
832 }
833 return false;
834 }
835 }
836 }
837 return true;
838 }
839
840 #endif //ASSERT
841
842 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
843 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
844 ciMethod* cur_method = jvms->method();
845 int cur_bci = jvms->bci();
846 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
847 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
848 return Interpreter::bytecode_should_reexecute(code) ||
849 (is_anewarray && (code == Bytecodes::_multianewarray));
850 // Reexecute _multianewarray bytecode which was replaced with
851 // sequence of [a]newarray. See Parse::do_multianewarray().
852 //
853 // Note: interpreter should not have it set since this optimization
854 // is limited by dimensions and guarded by flag so in some cases
855 // multianewarray() runtime calls will be generated and
856 // the bytecode should not be reexecutes (stack will not be reset).
857 } else {
858 return false;
859 }
860 }
861
862 // Helper function for adding JVMState and debug information to node
863 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
864 // Add the safepoint edges to the call (or other safepoint).
865
866 // Make sure dead locals are set to top. This
867 // should help register allocation time and cut down on the size
868 // of the deoptimization information.
869 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
870
871 // Walk the inline list to fill in the correct set of JVMState's
872 // Also fill in the associated edges for each JVMState.
873
874 // If the bytecode needs to be reexecuted we need to put
875 // the arguments back on the stack.
876 const bool should_reexecute = jvms()->should_reexecute();
877 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
878
879 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
880 // undefined if the bci is different. This is normal for Parse but it
881 // should not happen for LibraryCallKit because only one bci is processed.
882 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
883 "in LibraryCallKit the reexecute bit should not change");
884
885 // If we are guaranteed to throw, we can prune everything but the
886 // input to the current bytecode.
887 bool can_prune_locals = false;
888 uint stack_slots_not_pruned = 0;
889 int inputs = 0, depth = 0;
890 if (must_throw) {
891 assert(method() == youngest_jvms->method(), "sanity");
892 if (compute_stack_effects(inputs, depth)) {
893 can_prune_locals = true;
894 stack_slots_not_pruned = inputs;
895 }
896 }
897
898 if (env()->should_retain_local_variables()) {
899 // At any safepoint, this method can get breakpointed, which would
900 // then require an immediate deoptimization.
901 can_prune_locals = false; // do not prune locals
902 stack_slots_not_pruned = 0;
903 }
904
905 // do not scribble on the input jvms
906 JVMState* out_jvms = youngest_jvms->clone_deep(C);
907 call->set_jvms(out_jvms); // Start jvms list for call node
908
909 // For a known set of bytecodes, the interpreter should reexecute them if
910 // deoptimization happens. We set the reexecute state for them here
911 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
912 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
913 #ifdef ASSERT
914 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
915 assert(method() == youngest_jvms->method(), "sanity");
916 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
917 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
918 #endif // ASSERT
919 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
920 }
921
922 // Presize the call:
923 DEBUG_ONLY(uint non_debug_edges = call->req());
924 call->add_req_batch(top(), youngest_jvms->debug_depth());
925 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
926
927 // Set up edges so that the call looks like this:
928 // Call [state:] ctl io mem fptr retadr
929 // [parms:] parm0 ... parmN
930 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
931 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
932 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
933 // Note that caller debug info precedes callee debug info.
934
935 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
936 uint debug_ptr = call->req();
937
938 // Loop over the map input edges associated with jvms, add them
939 // to the call node, & reset all offsets to match call node array.
940 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
941 uint debug_end = debug_ptr;
942 uint debug_start = debug_ptr - in_jvms->debug_size();
943 debug_ptr = debug_start; // back up the ptr
944
945 uint p = debug_start; // walks forward in [debug_start, debug_end)
946 uint j, k, l;
947 SafePointNode* in_map = in_jvms->map();
948 out_jvms->set_map(call);
949
950 if (can_prune_locals) {
951 assert(in_jvms->method() == out_jvms->method(), "sanity");
952 // If the current throw can reach an exception handler in this JVMS,
953 // then we must keep everything live that can reach that handler.
954 // As a quick and dirty approximation, we look for any handlers at all.
955 if (in_jvms->method()->has_exception_handlers()) {
956 can_prune_locals = false;
957 }
958 }
959
960 // Add the Locals
961 k = in_jvms->locoff();
962 l = in_jvms->loc_size();
963 out_jvms->set_locoff(p);
964 if (!can_prune_locals) {
965 for (j = 0; j < l; j++)
966 call->set_req(p++, in_map->in(k+j));
967 } else {
968 p += l; // already set to top above by add_req_batch
969 }
970
971 // Add the Expression Stack
972 k = in_jvms->stkoff();
973 l = in_jvms->sp();
974 out_jvms->set_stkoff(p);
975 if (!can_prune_locals) {
976 for (j = 0; j < l; j++)
977 call->set_req(p++, in_map->in(k+j));
978 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
979 // Divide stack into {S0,...,S1}, where S0 is set to top.
980 uint s1 = stack_slots_not_pruned;
981 stack_slots_not_pruned = 0; // for next iteration
982 if (s1 > l) s1 = l;
983 uint s0 = l - s1;
984 p += s0; // skip the tops preinstalled by add_req_batch
985 for (j = s0; j < l; j++)
986 call->set_req(p++, in_map->in(k+j));
987 } else {
988 p += l; // already set to top above by add_req_batch
989 }
990
991 // Add the Monitors
992 k = in_jvms->monoff();
993 l = in_jvms->mon_size();
994 out_jvms->set_monoff(p);
995 for (j = 0; j < l; j++)
996 call->set_req(p++, in_map->in(k+j));
997
998 // Copy any scalar object fields.
999 k = in_jvms->scloff();
1000 l = in_jvms->scl_size();
1001 out_jvms->set_scloff(p);
1002 for (j = 0; j < l; j++)
1003 call->set_req(p++, in_map->in(k+j));
1004
1005 // Finish the new jvms.
1006 out_jvms->set_endoff(p);
1007
1008 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1009 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1010 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1011 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1012 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1013 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1014
1015 // Update the two tail pointers in parallel.
1016 out_jvms = out_jvms->caller();
1017 in_jvms = in_jvms->caller();
1018 }
1019
1020 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1021
1022 // Test the correctness of JVMState::debug_xxx accessors:
1023 assert(call->jvms()->debug_start() == non_debug_edges, "");
1024 assert(call->jvms()->debug_end() == call->req(), "");
1025 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1026 }
1027
1028 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1029 Bytecodes::Code code = java_bc();
1030 if (code == Bytecodes::_wide) {
1031 code = method()->java_code_at_bci(bci() + 1);
1032 }
1033
1034 BasicType rtype = T_ILLEGAL;
1035 int rsize = 0;
1036
1037 if (code != Bytecodes::_illegal) {
1038 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1039 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1040 if (rtype < T_CONFLICT)
1041 rsize = type2size[rtype];
1042 }
1043
1044 switch (code) {
1045 case Bytecodes::_illegal:
1046 return false;
1047
1048 case Bytecodes::_ldc:
1049 case Bytecodes::_ldc_w:
1050 case Bytecodes::_ldc2_w:
1051 inputs = 0;
1052 break;
1053
1054 case Bytecodes::_dup: inputs = 1; break;
1055 case Bytecodes::_dup_x1: inputs = 2; break;
1056 case Bytecodes::_dup_x2: inputs = 3; break;
1057 case Bytecodes::_dup2: inputs = 2; break;
1058 case Bytecodes::_dup2_x1: inputs = 3; break;
1059 case Bytecodes::_dup2_x2: inputs = 4; break;
1060 case Bytecodes::_swap: inputs = 2; break;
1061 case Bytecodes::_arraylength: inputs = 1; break;
1062
1063 case Bytecodes::_getstatic:
1064 case Bytecodes::_putstatic:
1065 case Bytecodes::_getfield:
1066 case Bytecodes::_putfield:
1067 {
1068 bool ignored_will_link;
1069 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1070 int size = field->type()->size();
1071 bool is_get = (depth >= 0), is_static = (depth & 1);
1072 inputs = (is_static ? 0 : 1);
1073 if (is_get) {
1074 depth = size - inputs;
1075 } else {
1076 inputs += size; // putxxx pops the value from the stack
1077 depth = - inputs;
1078 }
1079 }
1080 break;
1081
1082 case Bytecodes::_invokevirtual:
1083 case Bytecodes::_invokespecial:
1084 case Bytecodes::_invokestatic:
1085 case Bytecodes::_invokedynamic:
1086 case Bytecodes::_invokeinterface:
1087 {
1088 bool ignored_will_link;
1089 ciSignature* declared_signature = NULL;
1090 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1091 assert(declared_signature != NULL, "cannot be null");
1092 inputs = declared_signature->arg_size_for_bc(code);
1093 int size = declared_signature->return_type()->size();
1094 depth = size - inputs;
1095 }
1096 break;
1097
1098 case Bytecodes::_multianewarray:
1099 {
1100 ciBytecodeStream iter(method());
1101 iter.reset_to_bci(bci());
1102 iter.next();
1103 inputs = iter.get_dimensions();
1104 assert(rsize == 1, "");
1105 depth = rsize - inputs;
1106 }
1107 break;
1108
1109 case Bytecodes::_withfield: {
1110 bool ignored_will_link;
1111 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1112 int size = field->type()->size();
1113 inputs = size+1;
1114 depth = rsize - inputs;
1115 break;
1116 }
1117
1118 case Bytecodes::_ireturn:
1119 case Bytecodes::_lreturn:
1120 case Bytecodes::_freturn:
1121 case Bytecodes::_dreturn:
1122 case Bytecodes::_areturn:
1123 assert(rsize == -depth, "");
1124 inputs = rsize;
1125 break;
1126
1127 case Bytecodes::_jsr:
1128 case Bytecodes::_jsr_w:
1129 inputs = 0;
1130 depth = 1; // S.B. depth=1, not zero
1131 break;
1132
1133 default:
1134 // bytecode produces a typed result
1135 inputs = rsize - depth;
1136 assert(inputs >= 0, "");
1137 break;
1138 }
1139
1140 #ifdef ASSERT
1141 // spot check
1142 int outputs = depth + inputs;
1143 assert(outputs >= 0, "sanity");
1144 switch (code) {
1145 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1146 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1147 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1148 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1149 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1150 default: break;
1151 }
1152 #endif //ASSERT
1153
1154 return true;
1155 }
1156
1157
1158
1159 //------------------------------basic_plus_adr---------------------------------
1160 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1161 // short-circuit a common case
1162 if (offset == intcon(0)) return ptr;
1163 return _gvn.transform( new AddPNode(base, ptr, offset) );
1164 }
1165
1166 Node* GraphKit::ConvI2L(Node* offset) {
1167 // short-circuit a common case
1168 jint offset_con = find_int_con(offset, Type::OffsetBot);
1169 if (offset_con != Type::OffsetBot) {
1170 return longcon((jlong) offset_con);
1171 }
1172 return _gvn.transform( new ConvI2LNode(offset));
1173 }
1174
1175 Node* GraphKit::ConvI2UL(Node* offset) {
1176 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1177 if (offset_con != (juint) Type::OffsetBot) {
1178 return longcon((julong) offset_con);
1179 }
1180 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1181 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1182 return _gvn.transform( new AndLNode(conv, mask) );
1183 }
1184
1185 Node* GraphKit::ConvL2I(Node* offset) {
1186 // short-circuit a common case
1187 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1188 if (offset_con != (jlong)Type::OffsetBot) {
1189 return intcon((int) offset_con);
1190 }
1191 return _gvn.transform( new ConvL2INode(offset));
1192 }
1193
1194 //-------------------------load_object_klass-----------------------------------
1195 Node* GraphKit::load_object_klass(Node* obj) {
1196 // Special-case a fresh allocation to avoid building nodes:
1197 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1198 if (akls != NULL) return akls;
1199 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1200 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1201 }
1202
1203 //-------------------------load_array_length-----------------------------------
1204 Node* GraphKit::load_array_length(Node* array) {
1205 // Special-case a fresh allocation to avoid building nodes:
1206 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1207 Node *alen;
1208 if (alloc == NULL) {
1209 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1210 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1211 } else {
1212 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1213 }
1214 return alen;
1215 }
1216
1217 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1218 const TypeOopPtr* oop_type,
1219 bool replace_length_in_map) {
1220 Node* length = alloc->Ideal_length();
1221 if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1222 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1223 if (ccast != length) {
1224 // do not transform ccast here, it might convert to top node for
1225 // negative array length and break assumptions in parsing stage.
1226 _gvn.set_type_bottom(ccast);
1227 record_for_igvn(ccast);
1228 if (replace_length_in_map) {
1229 replace_in_map(length, ccast);
1230 }
1231 return ccast;
1232 }
1233 }
1234 return length;
1235 }
1236
1237 //------------------------------do_null_check----------------------------------
1238 // Helper function to do a NULL pointer check. Returned value is
1239 // the incoming address with NULL casted away. You are allowed to use the
1240 // not-null value only if you are control dependent on the test.
1241 #ifndef PRODUCT
1242 extern int explicit_null_checks_inserted,
1243 explicit_null_checks_elided;
1244 #endif
1245 Node* GraphKit::null_check_common(Node* value, BasicType type,
1246 // optional arguments for variations:
1247 bool assert_null,
1248 Node* *null_control,
1249 bool speculative,
1250 bool is_init_check) {
1251 assert(!assert_null || null_control == NULL, "not both at once");
1252 if (stopped()) return top();
1253 NOT_PRODUCT(explicit_null_checks_inserted++);
1254
1255 if (value->is_InlineType()) {
1256 // Null checking a scalarized but nullable inline type. Check the IsInit
1257 // input instead of the oop input to avoid keeping buffer allocations alive.
1258 InlineTypeNode* vtptr = value->as_InlineType();
1259 while (vtptr->get_oop()->is_InlineType()) {
1260 vtptr = vtptr->get_oop()->as_InlineType();
1261 }
1262 null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true);
1263 if (stopped()) {
1264 return top();
1265 }
1266 if (assert_null) {
1267 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1268 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1269 // replace_in_map(value, vtptr);
1270 // return vtptr;
1271 return null();
1272 }
1273 bool do_replace_in_map = (null_control == NULL || (*null_control) == top());
1274 return cast_not_null(value, do_replace_in_map);
1275 }
1276
1277 // Construct NULL check
1278 Node *chk = NULL;
1279 switch(type) {
1280 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1281 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1282 case T_PRIMITIVE_OBJECT : // fall through
1283 case T_ARRAY : // fall through
1284 type = T_OBJECT; // simplify further tests
1285 case T_OBJECT : {
1286 const Type *t = _gvn.type( value );
1287
1288 const TypeOopPtr* tp = t->isa_oopptr();
1289 if (tp != NULL && !tp->is_loaded()
1290 // Only for do_null_check, not any of its siblings:
1291 && !assert_null && null_control == NULL) {
1292 // Usually, any field access or invocation on an unloaded oop type
1293 // will simply fail to link, since the statically linked class is
1294 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1295 // the static class is loaded but the sharper oop type is not.
1296 // Rather than checking for this obscure case in lots of places,
1297 // we simply observe that a null check on an unloaded class
1298 // will always be followed by a nonsense operation, so we
1299 // can just issue the uncommon trap here.
1300 // Our access to the unloaded class will only be correct
1301 // after it has been loaded and initialized, which requires
1302 // a trip through the interpreter.
1303 ciKlass* klass = tp->unloaded_klass();
1304 #ifndef PRODUCT
1305 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1306 #endif
1307 uncommon_trap(Deoptimization::Reason_unloaded,
1308 Deoptimization::Action_reinterpret,
1309 klass, "!loaded");
1310 return top();
1311 }
1312
1313 if (assert_null) {
1314 // See if the type is contained in NULL_PTR.
1315 // If so, then the value is already null.
1316 if (t->higher_equal(TypePtr::NULL_PTR)) {
1317 NOT_PRODUCT(explicit_null_checks_elided++);
1318 return value; // Elided null assert quickly!
1319 }
1320 } else {
1321 // See if mixing in the NULL pointer changes type.
1322 // If so, then the NULL pointer was not allowed in the original
1323 // type. In other words, "value" was not-null.
1324 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1325 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1326 NOT_PRODUCT(explicit_null_checks_elided++);
1327 return value; // Elided null check quickly!
1328 }
1329 }
1330 chk = new CmpPNode( value, null() );
1331 break;
1332 }
1333
1334 default:
1335 fatal("unexpected type: %s", type2name(type));
1336 }
1337 assert(chk != NULL, "sanity check");
1338 chk = _gvn.transform(chk);
1339
1340 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1341 BoolNode *btst = new BoolNode( chk, btest);
1342 Node *tst = _gvn.transform( btst );
1343
1344 //-----------
1345 // if peephole optimizations occurred, a prior test existed.
1346 // If a prior test existed, maybe it dominates as we can avoid this test.
1347 if (tst != btst && type == T_OBJECT) {
1348 // At this point we want to scan up the CFG to see if we can
1349 // find an identical test (and so avoid this test altogether).
1350 Node *cfg = control();
1351 int depth = 0;
1352 while( depth < 16 ) { // Limit search depth for speed
1353 if( cfg->Opcode() == Op_IfTrue &&
1354 cfg->in(0)->in(1) == tst ) {
1355 // Found prior test. Use "cast_not_null" to construct an identical
1356 // CastPP (and hence hash to) as already exists for the prior test.
1357 // Return that casted value.
1358 if (assert_null) {
1359 replace_in_map(value, null());
1360 return null(); // do not issue the redundant test
1361 }
1362 Node *oldcontrol = control();
1363 set_control(cfg);
1364 Node *res = cast_not_null(value);
1365 set_control(oldcontrol);
1366 NOT_PRODUCT(explicit_null_checks_elided++);
1367 return res;
1368 }
1369 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1370 if (cfg == NULL) break; // Quit at region nodes
1371 depth++;
1372 }
1373 }
1374
1375 //-----------
1376 // Branch to failure if null
1377 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1378 Deoptimization::DeoptReason reason;
1379 if (assert_null) {
1380 reason = Deoptimization::reason_null_assert(speculative);
1381 } else if (type == T_OBJECT || is_init_check) {
1382 reason = Deoptimization::reason_null_check(speculative);
1383 } else {
1384 reason = Deoptimization::Reason_div0_check;
1385 }
1386 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1387 // ciMethodData::has_trap_at will return a conservative -1 if any
1388 // must-be-null assertion has failed. This could cause performance
1389 // problems for a method after its first do_null_assert failure.
1390 // Consider using 'Reason_class_check' instead?
1391
1392 // To cause an implicit null check, we set the not-null probability
1393 // to the maximum (PROB_MAX). For an explicit check the probability
1394 // is set to a smaller value.
1395 if (null_control != NULL || too_many_traps(reason)) {
1396 // probability is less likely
1397 ok_prob = PROB_LIKELY_MAG(3);
1398 } else if (!assert_null &&
1399 (ImplicitNullCheckThreshold > 0) &&
1400 method() != NULL &&
1401 (method()->method_data()->trap_count(reason)
1402 >= (uint)ImplicitNullCheckThreshold)) {
1403 ok_prob = PROB_LIKELY_MAG(3);
1404 }
1405
1406 if (null_control != NULL) {
1407 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1408 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1409 set_control( _gvn.transform( new IfTrueNode(iff)));
1410 #ifndef PRODUCT
1411 if (null_true == top()) {
1412 explicit_null_checks_elided++;
1413 }
1414 #endif
1415 (*null_control) = null_true;
1416 } else {
1417 BuildCutout unless(this, tst, ok_prob);
1418 // Check for optimizer eliding test at parse time
1419 if (stopped()) {
1420 // Failure not possible; do not bother making uncommon trap.
1421 NOT_PRODUCT(explicit_null_checks_elided++);
1422 } else if (assert_null) {
1423 uncommon_trap(reason,
1424 Deoptimization::Action_make_not_entrant,
1425 NULL, "assert_null");
1426 } else {
1427 replace_in_map(value, zerocon(type));
1428 builtin_throw(reason);
1429 }
1430 }
1431
1432 // Must throw exception, fall-thru not possible?
1433 if (stopped()) {
1434 return top(); // No result
1435 }
1436
1437 if (assert_null) {
1438 // Cast obj to null on this path.
1439 replace_in_map(value, zerocon(type));
1440 return zerocon(type);
1441 }
1442
1443 // Cast obj to not-null on this path, if there is no null_control.
1444 // (If there is a null_control, a non-null value may come back to haunt us.)
1445 if (type == T_OBJECT) {
1446 Node* cast = cast_not_null(value, false);
1447 if (null_control == NULL || (*null_control) == top())
1448 replace_in_map(value, cast);
1449 value = cast;
1450 }
1451
1452 return value;
1453 }
1454
1455 //------------------------------cast_not_null----------------------------------
1456 // Cast obj to not-null on this path
1457 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1458 if (obj->is_InlineType()) {
1459 Node* vt = obj->clone();
1460 vt->as_InlineType()->set_is_init(_gvn);
1461 vt = _gvn.transform(vt);
1462 if (do_replace_in_map) {
1463 replace_in_map(obj, vt);
1464 }
1465 return vt;
1466 }
1467 const Type *t = _gvn.type(obj);
1468 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1469 // Object is already not-null?
1470 if( t == t_not_null ) return obj;
1471
1472 Node *cast = new CastPPNode(obj,t_not_null);
1473 cast->init_req(0, control());
1474 cast = _gvn.transform( cast );
1475
1476 // Scan for instances of 'obj' in the current JVM mapping.
1477 // These instances are known to be not-null after the test.
1478 if (do_replace_in_map)
1479 replace_in_map(obj, cast);
1480
1481 return cast; // Return casted value
1482 }
1483
1484 // Sometimes in intrinsics, we implicitly know an object is not null
1485 // (there's no actual null check) so we can cast it to not null. In
1486 // the course of optimizations, the input to the cast can become null.
1487 // In that case that data path will die and we need the control path
1488 // to become dead as well to keep the graph consistent. So we have to
1489 // add a check for null for which one branch can't be taken. It uses
1490 // an Opaque4 node that will cause the check to be removed after loop
1491 // opts so the test goes away and the compiled code doesn't execute a
1492 // useless check.
1493 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1494 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1495 return value;
1496 }
1497 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1498 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1499 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1500 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1501 _gvn.set_type(iff, iff->Value(&_gvn));
1502 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1503 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1504 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1505 C->root()->add_req(halt);
1506 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1507 set_control(if_t);
1508 return cast_not_null(value, do_replace_in_map);
1509 }
1510
1511
1512 //--------------------------replace_in_map-------------------------------------
1513 void GraphKit::replace_in_map(Node* old, Node* neww) {
1514 if (old == neww) {
1515 return;
1516 }
1517
1518 map()->replace_edge(old, neww);
1519
1520 // Note: This operation potentially replaces any edge
1521 // on the map. This includes locals, stack, and monitors
1522 // of the current (innermost) JVM state.
1523
1524 // don't let inconsistent types from profiling escape this
1525 // method
1526
1527 const Type* told = _gvn.type(old);
1528 const Type* tnew = _gvn.type(neww);
1529
1530 if (!tnew->higher_equal(told)) {
1531 return;
1532 }
1533
1534 map()->record_replaced_node(old, neww);
1535 }
1536
1537
1538 //=============================================================================
1539 //--------------------------------memory---------------------------------------
1540 Node* GraphKit::memory(uint alias_idx) {
1541 MergeMemNode* mem = merged_memory();
1542 Node* p = mem->memory_at(alias_idx);
1543 assert(p != mem->empty_memory(), "empty");
1544 _gvn.set_type(p, Type::MEMORY); // must be mapped
1545 return p;
1546 }
1547
1548 //-----------------------------reset_memory------------------------------------
1549 Node* GraphKit::reset_memory() {
1550 Node* mem = map()->memory();
1551 // do not use this node for any more parsing!
1552 debug_only( map()->set_memory((Node*)NULL) );
1553 return _gvn.transform( mem );
1554 }
1555
1556 //------------------------------set_all_memory---------------------------------
1557 void GraphKit::set_all_memory(Node* newmem) {
1558 Node* mergemem = MergeMemNode::make(newmem);
1559 gvn().set_type_bottom(mergemem);
1560 map()->set_memory(mergemem);
1561 }
1562
1563 //------------------------------set_all_memory_call----------------------------
1564 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1565 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1566 set_all_memory(newmem);
1567 }
1568
1569 //=============================================================================
1570 //
1571 // parser factory methods for MemNodes
1572 //
1573 // These are layered on top of the factory methods in LoadNode and StoreNode,
1574 // and integrate with the parser's memory state and _gvn engine.
1575 //
1576
1577 // factory methods in "int adr_idx"
1578 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1579 int adr_idx,
1580 MemNode::MemOrd mo,
1581 LoadNode::ControlDependency control_dependency,
1582 bool require_atomic_access,
1583 bool unaligned,
1584 bool mismatched,
1585 bool unsafe,
1586 uint8_t barrier_data) {
1587 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1588 const TypePtr* adr_type = NULL; // debug-mode-only argument
1589 debug_only(adr_type = C->get_adr_type(adr_idx));
1590 Node* mem = memory(adr_idx);
1591 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1592 ld = _gvn.transform(ld);
1593
1594 if (((bt == T_OBJECT || bt == T_PRIMITIVE_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1595 // Improve graph before escape analysis and boxing elimination.
1596 record_for_igvn(ld);
1597 }
1598 return ld;
1599 }
1600
1601 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1602 int adr_idx,
1603 MemNode::MemOrd mo,
1604 bool require_atomic_access,
1605 bool unaligned,
1606 bool mismatched,
1607 bool unsafe) {
1608 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1609 const TypePtr* adr_type = NULL;
1610 debug_only(adr_type = C->get_adr_type(adr_idx));
1611 Node *mem = memory(adr_idx);
1612 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1613 if (unaligned) {
1614 st->as_Store()->set_unaligned_access();
1615 }
1616 if (mismatched) {
1617 st->as_Store()->set_mismatched_access();
1618 }
1619 if (unsafe) {
1620 st->as_Store()->set_unsafe_access();
1621 }
1622 st = _gvn.transform(st);
1623 set_memory(st, adr_idx);
1624 // Back-to-back stores can only remove intermediate store with DU info
1625 // so push on worklist for optimizer.
1626 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1627 record_for_igvn(st);
1628
1629 return st;
1630 }
1631
1632 Node* GraphKit::access_store_at(Node* obj,
1633 Node* adr,
1634 const TypePtr* adr_type,
1635 Node* val,
1636 const Type* val_type,
1637 BasicType bt,
1638 DecoratorSet decorators,
1639 bool safe_for_replace) {
1640 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1641 // could be delayed during Parse (for example, in adjust_map_after_if()).
1642 // Execute transformation here to avoid barrier generation in such case.
1643 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1644 val = _gvn.makecon(TypePtr::NULL_PTR);
1645 }
1646
1647 if (stopped()) {
1648 return top(); // Dead path ?
1649 }
1650
1651 assert(val != NULL, "not dead path");
1652 if (val->is_InlineType()) {
1653 // Store to non-flattened field. Buffer the inline type and make sure
1654 // the store is re-executed if the allocation triggers deoptimization.
1655 PreserveReexecuteState preexecs(this);
1656 jvms()->set_should_reexecute(true);
1657 val = val->as_InlineType()->buffer(this, safe_for_replace);
1658 }
1659
1660 C2AccessValuePtr addr(adr, adr_type);
1661 C2AccessValue value(val, val_type);
1662 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1663 if (access.is_raw()) {
1664 return _barrier_set->BarrierSetC2::store_at(access, value);
1665 } else {
1666 return _barrier_set->store_at(access, value);
1667 }
1668 }
1669
1670 Node* GraphKit::access_load_at(Node* obj, // containing obj
1671 Node* adr, // actual address to store val at
1672 const TypePtr* adr_type,
1673 const Type* val_type,
1674 BasicType bt,
1675 DecoratorSet decorators,
1676 Node* ctl) {
1677 if (stopped()) {
1678 return top(); // Dead path ?
1679 }
1680
1681 C2AccessValuePtr addr(adr, adr_type);
1682 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1683 if (access.is_raw()) {
1684 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1685 } else {
1686 return _barrier_set->load_at(access, val_type);
1687 }
1688 }
1689
1690 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1691 const Type* val_type,
1692 BasicType bt,
1693 DecoratorSet decorators) {
1694 if (stopped()) {
1695 return top(); // Dead path ?
1696 }
1697
1698 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1699 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1700 if (access.is_raw()) {
1701 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1702 } else {
1703 return _barrier_set->load_at(access, val_type);
1704 }
1705 }
1706
1707 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1708 Node* adr,
1709 const TypePtr* adr_type,
1710 int alias_idx,
1711 Node* expected_val,
1712 Node* new_val,
1713 const Type* value_type,
1714 BasicType bt,
1715 DecoratorSet decorators) {
1716 C2AccessValuePtr addr(adr, adr_type);
1717 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1718 bt, obj, addr, alias_idx);
1719 if (access.is_raw()) {
1720 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1721 } else {
1722 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1723 }
1724 }
1725
1726 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1727 Node* adr,
1728 const TypePtr* adr_type,
1729 int alias_idx,
1730 Node* expected_val,
1731 Node* new_val,
1732 const Type* value_type,
1733 BasicType bt,
1734 DecoratorSet decorators) {
1735 C2AccessValuePtr addr(adr, adr_type);
1736 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1737 bt, obj, addr, alias_idx);
1738 if (access.is_raw()) {
1739 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1740 } else {
1741 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1742 }
1743 }
1744
1745 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1746 Node* adr,
1747 const TypePtr* adr_type,
1748 int alias_idx,
1749 Node* new_val,
1750 const Type* value_type,
1751 BasicType bt,
1752 DecoratorSet decorators) {
1753 C2AccessValuePtr addr(adr, adr_type);
1754 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1755 bt, obj, addr, alias_idx);
1756 if (access.is_raw()) {
1757 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1758 } else {
1759 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1760 }
1761 }
1762
1763 Node* GraphKit::access_atomic_add_at(Node* obj,
1764 Node* adr,
1765 const TypePtr* adr_type,
1766 int alias_idx,
1767 Node* new_val,
1768 const Type* value_type,
1769 BasicType bt,
1770 DecoratorSet decorators) {
1771 C2AccessValuePtr addr(adr, adr_type);
1772 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1773 if (access.is_raw()) {
1774 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1775 } else {
1776 return _barrier_set->atomic_add_at(access, new_val, value_type);
1777 }
1778 }
1779
1780 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1781 return _barrier_set->clone(this, src, dst, size, is_array);
1782 }
1783
1784 //-------------------------array_element_address-------------------------
1785 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1786 const TypeInt* sizetype, Node* ctrl) {
1787 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1788 uint shift = arytype->is_flat() ? arytype->flat_log_elem_size() : exact_log2(type2aelembytes(elembt));
1789 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1790
1791 // short-circuit a common case (saves lots of confusing waste motion)
1792 jint idx_con = find_int_con(idx, -1);
1793 if (idx_con >= 0) {
1794 intptr_t offset = header + ((intptr_t)idx_con << shift);
1795 return basic_plus_adr(ary, offset);
1796 }
1797
1798 // must be correct type for alignment purposes
1799 Node* base = basic_plus_adr(ary, header);
1800 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1801 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1802 return basic_plus_adr(ary, base, scale);
1803 }
1804
1805 //-------------------------load_array_element-------------------------
1806 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1807 const Type* elemtype = arytype->elem();
1808 BasicType elembt = elemtype->array_element_basic_type();
1809 assert(elembt != T_PRIMITIVE_OBJECT, "inline types are not supported by this method");
1810 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1811 if (elembt == T_NARROWOOP) {
1812 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1813 }
1814 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1815 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1816 return ld;
1817 }
1818
1819 //-------------------------set_arguments_for_java_call-------------------------
1820 // Arguments (pre-popped from the stack) are taken from the JVMS.
1821 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1822 PreserveReexecuteState preexecs(this);
1823 if (EnableValhalla) {
1824 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1825 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1826 jvms()->set_should_reexecute(true);
1827 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1828 inc_sp(arg_size);
1829 }
1830 // Add the call arguments
1831 const TypeTuple* domain = call->tf()->domain_sig();
1832 uint nargs = domain->cnt();
1833 int arg_num = 0;
1834 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1835 Node* arg = argument(i-TypeFunc::Parms);
1836 const Type* t = domain->field_at(i);
1837 if (t->is_inlinetypeptr() && call->method()->is_scalarized_arg(arg_num)) {
1838 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1839 if (!arg->is_InlineType()) {
1840 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
1841 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass(), t->inline_klass()->is_null_free());
1842 }
1843 InlineTypeNode* vt = arg->as_InlineType();
1844 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1845 // If an inline type argument is passed as fields, attach the Method* to the call site
1846 // to be able to access the extended signature later via attached_method_before_pc().
1847 // For example, see CompiledMethod::preserve_callee_argument_oops().
1848 call->set_override_symbolic_info(true);
1849 arg_num++;
1850 continue;
1851 } else if (arg->is_InlineType()) {
1852 // Pass inline type argument via oop to callee
1853 arg = arg->as_InlineType()->buffer(this);
1854 if (!is_late_inline) {
1855 arg = arg->as_InlineType()->get_oop();
1856 }
1857 }
1858 if (t != Type::HALF) {
1859 arg_num++;
1860 }
1861 call->init_req(idx++, arg);
1862 }
1863 }
1864
1865 //---------------------------set_edges_for_java_call---------------------------
1866 // Connect a newly created call into the current JVMS.
1867 // A return value node (if any) is returned from set_edges_for_java_call.
1868 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1869
1870 // Add the predefined inputs:
1871 call->init_req( TypeFunc::Control, control() );
1872 call->init_req( TypeFunc::I_O , i_o() );
1873 call->init_req( TypeFunc::Memory , reset_memory() );
1874 call->init_req( TypeFunc::FramePtr, frameptr() );
1875 call->init_req( TypeFunc::ReturnAdr, top() );
1876
1877 add_safepoint_edges(call, must_throw);
1878
1879 Node* xcall = _gvn.transform(call);
1880
1881 if (xcall == top()) {
1882 set_control(top());
1883 return;
1884 }
1885 assert(xcall == call, "call identity is stable");
1886
1887 // Re-use the current map to produce the result.
1888
1889 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1890 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1891 set_all_memory_call(xcall, separate_io_proj);
1892
1893 //return xcall; // no need, caller already has it
1894 }
1895
1896 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1897 if (stopped()) return top(); // maybe the call folded up?
1898
1899 // Note: Since any out-of-line call can produce an exception,
1900 // we always insert an I_O projection from the call into the result.
1901
1902 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1903
1904 if (separate_io_proj) {
1905 // The caller requested separate projections be used by the fall
1906 // through and exceptional paths, so replace the projections for
1907 // the fall through path.
1908 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1909 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1910 }
1911
1912 // Capture the return value, if any.
1913 Node* ret;
1914 if (call->method() == NULL || call->method()->return_type()->basic_type() == T_VOID) {
1915 ret = top();
1916 } else if (call->tf()->returns_inline_type_as_fields()) {
1917 // Return of multiple values (inline type fields): we create a
1918 // InlineType node, each field is a projection from the call.
1919 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
1920 uint base_input = TypeFunc::Parms;
1921 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, call->method()->signature()->returns_null_free_inline_type());
1922 } else {
1923 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1924 }
1925
1926 return ret;
1927 }
1928
1929 //--------------------set_predefined_input_for_runtime_call--------------------
1930 // Reading and setting the memory state is way conservative here.
1931 // The real problem is that I am not doing real Type analysis on memory,
1932 // so I cannot distinguish card mark stores from other stores. Across a GC
1933 // point the Store Barrier and the card mark memory has to agree. I cannot
1934 // have a card mark store and its barrier split across the GC point from
1935 // either above or below. Here I get that to happen by reading ALL of memory.
1936 // A better answer would be to separate out card marks from other memory.
1937 // For now, return the input memory state, so that it can be reused
1938 // after the call, if this call has restricted memory effects.
1939 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1940 // Set fixed predefined input arguments
1941 Node* memory = reset_memory();
1942 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1943 call->init_req( TypeFunc::Control, control() );
1944 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1945 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1946 call->init_req( TypeFunc::FramePtr, frameptr() );
1947 call->init_req( TypeFunc::ReturnAdr, top() );
1948 return memory;
1949 }
1950
1951 //-------------------set_predefined_output_for_runtime_call--------------------
1952 // Set control and memory (not i_o) from the call.
1953 // If keep_mem is not NULL, use it for the output state,
1954 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1955 // If hook_mem is NULL, this call produces no memory effects at all.
1956 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1957 // then only that memory slice is taken from the call.
1958 // In the last case, we must put an appropriate memory barrier before
1959 // the call, so as to create the correct anti-dependencies on loads
1960 // preceding the call.
1961 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1962 Node* keep_mem,
1963 const TypePtr* hook_mem) {
1964 // no i/o
1965 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1966 if (keep_mem) {
1967 // First clone the existing memory state
1968 set_all_memory(keep_mem);
1969 if (hook_mem != NULL) {
1970 // Make memory for the call
1971 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1972 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1973 // We also use hook_mem to extract specific effects from arraycopy stubs.
1974 set_memory(mem, hook_mem);
1975 }
1976 // ...else the call has NO memory effects.
1977
1978 // Make sure the call advertises its memory effects precisely.
1979 // This lets us build accurate anti-dependences in gcm.cpp.
1980 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1981 "call node must be constructed correctly");
1982 } else {
1983 assert(hook_mem == NULL, "");
1984 // This is not a "slow path" call; all memory comes from the call.
1985 set_all_memory_call(call);
1986 }
1987 }
1988
1989 // Keep track of MergeMems feeding into other MergeMems
1990 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
1991 if (!mem->is_MergeMem()) {
1992 return;
1993 }
1994 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
1995 Node* use = i.get();
1996 if (use->is_MergeMem()) {
1997 wl.push(use);
1998 }
1999 }
2000 }
2001
2002 // Replace the call with the current state of the kit.
2003 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
2004 JVMState* ejvms = NULL;
2005 if (has_exceptions()) {
2006 ejvms = transfer_exceptions_into_jvms();
2007 }
2008
2009 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2010 ReplacedNodes replaced_nodes_exception;
2011 Node* ex_ctl = top();
2012
2013 SafePointNode* final_state = stop();
2014
2015 // Find all the needed outputs of this call
2016 CallProjections* callprojs = call->extract_projections(true);
2017
2018 Unique_Node_List wl;
2019 Node* init_mem = call->in(TypeFunc::Memory);
2020 Node* final_mem = final_state->in(TypeFunc::Memory);
2021 Node* final_ctl = final_state->in(TypeFunc::Control);
2022 Node* final_io = final_state->in(TypeFunc::I_O);
2023
2024 // Replace all the old call edges with the edges from the inlining result
2025 if (callprojs->fallthrough_catchproj != NULL) {
2026 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2027 }
2028 if (callprojs->fallthrough_memproj != NULL) {
2029 if (final_mem->is_MergeMem()) {
2030 // Parser's exits MergeMem was not transformed but may be optimized
2031 final_mem = _gvn.transform(final_mem);
2032 }
2033 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2034 add_mergemem_users_to_worklist(wl, final_mem);
2035 }
2036 if (callprojs->fallthrough_ioproj != NULL) {
2037 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2038 }
2039
2040 // Replace the result with the new result if it exists and is used
2041 if (callprojs->resproj[0] != NULL && result != NULL) {
2042 // If the inlined code is dead, the result projections for an inline type returned as
2043 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2044 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2045 "unexpected number of results");
2046 C->gvn_replace_by(callprojs->resproj[0], result);
2047 }
2048
2049 if (ejvms == NULL) {
2050 // No exception edges to simply kill off those paths
2051 if (callprojs->catchall_catchproj != NULL) {
2052 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2053 }
2054 if (callprojs->catchall_memproj != NULL) {
2055 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2056 }
2057 if (callprojs->catchall_ioproj != NULL) {
2058 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2059 }
2060 // Replace the old exception object with top
2061 if (callprojs->exobj != NULL) {
2062 C->gvn_replace_by(callprojs->exobj, C->top());
2063 }
2064 } else {
2065 GraphKit ekit(ejvms);
2066
2067 // Load my combined exception state into the kit, with all phis transformed:
2068 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2069 replaced_nodes_exception = ex_map->replaced_nodes();
2070
2071 Node* ex_oop = ekit.use_exception_state(ex_map);
2072
2073 if (callprojs->catchall_catchproj != NULL) {
2074 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2075 ex_ctl = ekit.control();
2076 }
2077 if (callprojs->catchall_memproj != NULL) {
2078 Node* ex_mem = ekit.reset_memory();
2079 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2080 add_mergemem_users_to_worklist(wl, ex_mem);
2081 }
2082 if (callprojs->catchall_ioproj != NULL) {
2083 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2084 }
2085
2086 // Replace the old exception object with the newly created one
2087 if (callprojs->exobj != NULL) {
2088 C->gvn_replace_by(callprojs->exobj, ex_oop);
2089 }
2090 }
2091
2092 // Disconnect the call from the graph
2093 call->disconnect_inputs(C);
2094 C->gvn_replace_by(call, C->top());
2095
2096 // Clean up any MergeMems that feed other MergeMems since the
2097 // optimizer doesn't like that.
2098 while (wl.size() > 0) {
2099 _gvn.transform(wl.pop());
2100 }
2101
2102 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2103 replaced_nodes.apply(C, final_ctl);
2104 }
2105 if (!ex_ctl->is_top() && do_replaced_nodes) {
2106 replaced_nodes_exception.apply(C, ex_ctl);
2107 }
2108 }
2109
2110
2111 //------------------------------increment_counter------------------------------
2112 // for statistics: increment a VM counter by 1
2113
2114 void GraphKit::increment_counter(address counter_addr) {
2115 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2116 increment_counter(adr1);
2117 }
2118
2119 void GraphKit::increment_counter(Node* counter_addr) {
2120 int adr_type = Compile::AliasIdxRaw;
2121 Node* ctrl = control();
2122 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered);
2123 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2124 store_to_memory(ctrl, counter_addr, incr, T_LONG, adr_type, MemNode::unordered);
2125 }
2126
2127
2128 //------------------------------uncommon_trap----------------------------------
2129 // Bail out to the interpreter in mid-method. Implemented by calling the
2130 // uncommon_trap blob. This helper function inserts a runtime call with the
2131 // right debug info.
2132 Node* GraphKit::uncommon_trap(int trap_request,
2133 ciKlass* klass, const char* comment,
2134 bool must_throw,
2135 bool keep_exact_action) {
2136 if (failing()) stop();
2137 if (stopped()) return NULL; // trap reachable?
2138
2139 // Note: If ProfileTraps is true, and if a deopt. actually
2140 // occurs here, the runtime will make sure an MDO exists. There is
2141 // no need to call method()->ensure_method_data() at this point.
2142
2143 // Set the stack pointer to the right value for reexecution:
2144 set_sp(reexecute_sp());
2145
2146 #ifdef ASSERT
2147 if (!must_throw) {
2148 // Make sure the stack has at least enough depth to execute
2149 // the current bytecode.
2150 int inputs, ignored_depth;
2151 if (compute_stack_effects(inputs, ignored_depth)) {
2152 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2153 Bytecodes::name(java_bc()), sp(), inputs);
2154 }
2155 }
2156 #endif
2157
2158 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2159 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2160
2161 switch (action) {
2162 case Deoptimization::Action_maybe_recompile:
2163 case Deoptimization::Action_reinterpret:
2164 // Temporary fix for 6529811 to allow virtual calls to be sure they
2165 // get the chance to go from mono->bi->mega
2166 if (!keep_exact_action &&
2167 Deoptimization::trap_request_index(trap_request) < 0 &&
2168 too_many_recompiles(reason)) {
2169 // This BCI is causing too many recompilations.
2170 if (C->log() != NULL) {
2171 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2172 Deoptimization::trap_reason_name(reason),
2173 Deoptimization::trap_action_name(action));
2174 }
2175 action = Deoptimization::Action_none;
2176 trap_request = Deoptimization::make_trap_request(reason, action);
2177 } else {
2178 C->set_trap_can_recompile(true);
2179 }
2180 break;
2181 case Deoptimization::Action_make_not_entrant:
2182 C->set_trap_can_recompile(true);
2183 break;
2184 case Deoptimization::Action_none:
2185 case Deoptimization::Action_make_not_compilable:
2186 break;
2187 default:
2188 #ifdef ASSERT
2189 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2190 #endif
2191 break;
2192 }
2193
2194 if (TraceOptoParse) {
2195 char buf[100];
2196 tty->print_cr("Uncommon trap %s at bci:%d",
2197 Deoptimization::format_trap_request(buf, sizeof(buf),
2198 trap_request), bci());
2199 }
2200
2201 CompileLog* log = C->log();
2202 if (log != NULL) {
2203 int kid = (klass == NULL)? -1: log->identify(klass);
2204 log->begin_elem("uncommon_trap bci='%d'", bci());
2205 char buf[100];
2206 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2207 trap_request));
2208 if (kid >= 0) log->print(" klass='%d'", kid);
2209 if (comment != NULL) log->print(" comment='%s'", comment);
2210 log->end_elem();
2211 }
2212
2213 // Make sure any guarding test views this path as very unlikely
2214 Node *i0 = control()->in(0);
2215 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2216 IfNode *iff = i0->as_If();
2217 float f = iff->_prob; // Get prob
2218 if (control()->Opcode() == Op_IfTrue) {
2219 if (f > PROB_UNLIKELY_MAG(4))
2220 iff->_prob = PROB_MIN;
2221 } else {
2222 if (f < PROB_LIKELY_MAG(4))
2223 iff->_prob = PROB_MAX;
2224 }
2225 }
2226
2227 // Clear out dead values from the debug info.
2228 kill_dead_locals();
2229
2230 // Now insert the uncommon trap subroutine call
2231 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2232 const TypePtr* no_memory_effects = NULL;
2233 // Pass the index of the class to be loaded
2234 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2235 (must_throw ? RC_MUST_THROW : 0),
2236 OptoRuntime::uncommon_trap_Type(),
2237 call_addr, "uncommon_trap", no_memory_effects,
2238 intcon(trap_request));
2239 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2240 "must extract request correctly from the graph");
2241 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2242
2243 call->set_req(TypeFunc::ReturnAdr, returnadr());
2244 // The debug info is the only real input to this call.
2245
2246 // Halt-and-catch fire here. The above call should never return!
2247 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2248 PRODUCT_ONLY(COMMA /*reachable*/false));
2249 _gvn.set_type_bottom(halt);
2250 root()->add_req(halt);
2251
2252 stop_and_kill_map();
2253 return call;
2254 }
2255
2256
2257 //--------------------------just_allocated_object------------------------------
2258 // Report the object that was just allocated.
2259 // It must be the case that there are no intervening safepoints.
2260 // We use this to determine if an object is so "fresh" that
2261 // it does not require card marks.
2262 Node* GraphKit::just_allocated_object(Node* current_control) {
2263 Node* ctrl = current_control;
2264 // Object::<init> is invoked after allocation, most of invoke nodes
2265 // will be reduced, but a region node is kept in parse time, we check
2266 // the pattern and skip the region node if it degraded to a copy.
2267 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 &&
2268 ctrl->as_Region()->is_copy()) {
2269 ctrl = ctrl->as_Region()->is_copy();
2270 }
2271 if (C->recent_alloc_ctl() == ctrl) {
2272 return C->recent_alloc_obj();
2273 }
2274 return NULL;
2275 }
2276
2277
2278 /**
2279 * Record profiling data exact_kls for Node n with the type system so
2280 * that it can propagate it (speculation)
2281 *
2282 * @param n node that the type applies to
2283 * @param exact_kls type from profiling
2284 * @param maybe_null did profiling see null?
2285 *
2286 * @return node with improved type
2287 */
2288 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2289 const Type* current_type = _gvn.type(n);
2290 assert(UseTypeSpeculation, "type speculation must be on");
2291
2292 const TypePtr* speculative = current_type->speculative();
2293
2294 // Should the klass from the profile be recorded in the speculative type?
2295 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2296 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2297 const TypeOopPtr* xtype = tklass->as_instance_type();
2298 assert(xtype->klass_is_exact(), "Should be exact");
2299 // Any reason to believe n is not null (from this profiling or a previous one)?
2300 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2301 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2302 // record the new speculative type's depth
2303 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2304 speculative = speculative->with_inline_depth(jvms()->depth());
2305 } else if (current_type->would_improve_ptr(ptr_kind)) {
2306 // Profiling report that null was never seen so we can change the
2307 // speculative type to non null ptr.
2308 if (ptr_kind == ProfileAlwaysNull) {
2309 speculative = TypePtr::NULL_PTR;
2310 } else {
2311 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2312 const TypePtr* ptr = TypePtr::NOTNULL;
2313 if (speculative != NULL) {
2314 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2315 } else {
2316 speculative = ptr;
2317 }
2318 }
2319 }
2320
2321 if (speculative != current_type->speculative()) {
2322 // Build a type with a speculative type (what we think we know
2323 // about the type but will need a guard when we use it)
2324 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2325 // We're changing the type, we need a new CheckCast node to carry
2326 // the new type. The new type depends on the control: what
2327 // profiling tells us is only valid from here as far as we can
2328 // tell.
2329 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2330 cast = _gvn.transform(cast);
2331 replace_in_map(n, cast);
2332 n = cast;
2333 }
2334
2335 return n;
2336 }
2337
2338 /**
2339 * Record profiling data from receiver profiling at an invoke with the
2340 * type system so that it can propagate it (speculation)
2341 *
2342 * @param n receiver node
2343 *
2344 * @return node with improved type
2345 */
2346 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2347 if (!UseTypeSpeculation) {
2348 return n;
2349 }
2350 ciKlass* exact_kls = profile_has_unique_klass();
2351 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2352 if ((java_bc() == Bytecodes::_checkcast ||
2353 java_bc() == Bytecodes::_instanceof ||
2354 java_bc() == Bytecodes::_aastore) &&
2355 method()->method_data()->is_mature()) {
2356 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2357 if (data != NULL) {
2358 if (java_bc() == Bytecodes::_aastore) {
2359 ciKlass* array_type = NULL;
2360 ciKlass* element_type = NULL;
2361 ProfilePtrKind element_ptr = ProfileMaybeNull;
2362 bool flat_array = true;
2363 bool null_free_array = true;
2364 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2365 exact_kls = element_type;
2366 ptr_kind = element_ptr;
2367 } else {
2368 if (!data->as_BitData()->null_seen()) {
2369 ptr_kind = ProfileNeverNull;
2370 } else {
2371 assert(data->is_ReceiverTypeData(), "bad profile data type");
2372 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2373 uint i = 0;
2374 for (; i < call->row_limit(); i++) {
2375 ciKlass* receiver = call->receiver(i);
2376 if (receiver != NULL) {
2377 break;
2378 }
2379 }
2380 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2381 }
2382 }
2383 }
2384 }
2385 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2386 }
2387
2388 /**
2389 * Record profiling data from argument profiling at an invoke with the
2390 * type system so that it can propagate it (speculation)
2391 *
2392 * @param dest_method target method for the call
2393 * @param bc what invoke bytecode is this?
2394 */
2395 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2396 if (!UseTypeSpeculation) {
2397 return;
2398 }
2399 const TypeFunc* tf = TypeFunc::make(dest_method);
2400 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2401 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2402 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2403 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2404 if (is_reference_type(targ->basic_type())) {
2405 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2406 ciKlass* better_type = NULL;
2407 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2408 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2409 }
2410 i++;
2411 }
2412 }
2413 }
2414
2415 /**
2416 * Record profiling data from parameter profiling at an invoke with
2417 * the type system so that it can propagate it (speculation)
2418 */
2419 void GraphKit::record_profiled_parameters_for_speculation() {
2420 if (!UseTypeSpeculation) {
2421 return;
2422 }
2423 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2424 if (_gvn.type(local(i))->isa_oopptr()) {
2425 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2426 ciKlass* better_type = NULL;
2427 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2428 record_profile_for_speculation(local(i), better_type, ptr_kind);
2429 }
2430 j++;
2431 }
2432 }
2433 }
2434
2435 /**
2436 * Record profiling data from return value profiling at an invoke with
2437 * the type system so that it can propagate it (speculation)
2438 */
2439 void GraphKit::record_profiled_return_for_speculation() {
2440 if (!UseTypeSpeculation) {
2441 return;
2442 }
2443 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2444 ciKlass* better_type = NULL;
2445 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2446 // If profiling reports a single type for the return value,
2447 // feed it to the type system so it can propagate it as a
2448 // speculative type
2449 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2450 }
2451 }
2452
2453 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2454 if (Matcher::strict_fp_requires_explicit_rounding) {
2455 // (Note: TypeFunc::make has a cache that makes this fast.)
2456 const TypeFunc* tf = TypeFunc::make(dest_method);
2457 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2458 for (int j = 0; j < nargs; j++) {
2459 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2460 if (targ->basic_type() == T_DOUBLE) {
2461 // If any parameters are doubles, they must be rounded before
2462 // the call, dprecision_rounding does gvn.transform
2463 Node *arg = argument(j);
2464 arg = dprecision_rounding(arg);
2465 set_argument(j, arg);
2466 }
2467 }
2468 }
2469 }
2470
2471 // rounding for strict float precision conformance
2472 Node* GraphKit::precision_rounding(Node* n) {
2473 if (Matcher::strict_fp_requires_explicit_rounding) {
2474 #ifdef IA32
2475 if (UseSSE == 0) {
2476 return _gvn.transform(new RoundFloatNode(0, n));
2477 }
2478 #else
2479 Unimplemented();
2480 #endif // IA32
2481 }
2482 return n;
2483 }
2484
2485 // rounding for strict double precision conformance
2486 Node* GraphKit::dprecision_rounding(Node *n) {
2487 if (Matcher::strict_fp_requires_explicit_rounding) {
2488 #ifdef IA32
2489 if (UseSSE < 2) {
2490 return _gvn.transform(new RoundDoubleNode(0, n));
2491 }
2492 #else
2493 Unimplemented();
2494 #endif // IA32
2495 }
2496 return n;
2497 }
2498
2499 //=============================================================================
2500 // Generate a fast path/slow path idiom. Graph looks like:
2501 // [foo] indicates that 'foo' is a parameter
2502 //
2503 // [in] NULL
2504 // \ /
2505 // CmpP
2506 // Bool ne
2507 // If
2508 // / \
2509 // True False-<2>
2510 // / |
2511 // / cast_not_null
2512 // Load | | ^
2513 // [fast_test] | |
2514 // gvn to opt_test | |
2515 // / \ | <1>
2516 // True False |
2517 // | \\ |
2518 // [slow_call] \[fast_result]
2519 // Ctl Val \ \
2520 // | \ \
2521 // Catch <1> \ \
2522 // / \ ^ \ \
2523 // Ex No_Ex | \ \
2524 // | \ \ | \ <2> \
2525 // ... \ [slow_res] | | \ [null_result]
2526 // \ \--+--+--- | |
2527 // \ | / \ | /
2528 // --------Region Phi
2529 //
2530 //=============================================================================
2531 // Code is structured as a series of driver functions all called 'do_XXX' that
2532 // call a set of helper functions. Helper functions first, then drivers.
2533
2534 //------------------------------null_check_oop---------------------------------
2535 // Null check oop. Set null-path control into Region in slot 3.
2536 // Make a cast-not-nullness use the other not-null control. Return cast.
2537 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2538 bool never_see_null,
2539 bool safe_for_replace,
2540 bool speculative) {
2541 // Initial NULL check taken path
2542 (*null_control) = top();
2543 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2544
2545 // Generate uncommon_trap:
2546 if (never_see_null && (*null_control) != top()) {
2547 // If we see an unexpected null at a check-cast we record it and force a
2548 // recompile; the offending check-cast will be compiled to handle NULLs.
2549 // If we see more than one offending BCI, then all checkcasts in the
2550 // method will be compiled to handle NULLs.
2551 PreserveJVMState pjvms(this);
2552 set_control(*null_control);
2553 replace_in_map(value, null());
2554 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2555 uncommon_trap(reason,
2556 Deoptimization::Action_make_not_entrant);
2557 (*null_control) = top(); // NULL path is dead
2558 }
2559 if ((*null_control) == top() && safe_for_replace) {
2560 replace_in_map(value, cast);
2561 }
2562
2563 // Cast away null-ness on the result
2564 return cast;
2565 }
2566
2567 //------------------------------opt_iff----------------------------------------
2568 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2569 // Return slow-path control.
2570 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2571 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2572
2573 // Fast path taken; set region slot 2
2574 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2575 region->init_req(2,fast_taken); // Capture fast-control
2576
2577 // Fast path not-taken, i.e. slow path
2578 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2579 return slow_taken;
2580 }
2581
2582 //-----------------------------make_runtime_call-------------------------------
2583 Node* GraphKit::make_runtime_call(int flags,
2584 const TypeFunc* call_type, address call_addr,
2585 const char* call_name,
2586 const TypePtr* adr_type,
2587 // The following parms are all optional.
2588 // The first NULL ends the list.
2589 Node* parm0, Node* parm1,
2590 Node* parm2, Node* parm3,
2591 Node* parm4, Node* parm5,
2592 Node* parm6, Node* parm7) {
2593 assert(call_addr != NULL, "must not call NULL targets");
2594
2595 // Slow-path call
2596 bool is_leaf = !(flags & RC_NO_LEAF);
2597 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2598 if (call_name == NULL) {
2599 assert(!is_leaf, "must supply name for leaf");
2600 call_name = OptoRuntime::stub_name(call_addr);
2601 }
2602 CallNode* call;
2603 if (!is_leaf) {
2604 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2605 } else if (flags & RC_NO_FP) {
2606 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2607 } else if (flags & RC_VECTOR){
2608 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2609 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2610 } else {
2611 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2612 }
2613
2614 // The following is similar to set_edges_for_java_call,
2615 // except that the memory effects of the call are restricted to AliasIdxRaw.
2616
2617 // Slow path call has no side-effects, uses few values
2618 bool wide_in = !(flags & RC_NARROW_MEM);
2619 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2620
2621 Node* prev_mem = NULL;
2622 if (wide_in) {
2623 prev_mem = set_predefined_input_for_runtime_call(call);
2624 } else {
2625 assert(!wide_out, "narrow in => narrow out");
2626 Node* narrow_mem = memory(adr_type);
2627 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2628 }
2629
2630 // Hook each parm in order. Stop looking at the first NULL.
2631 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2632 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2633 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2634 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2635 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2636 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2637 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2638 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2639 /* close each nested if ===> */ } } } } } } } }
2640 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2641
2642 if (!is_leaf) {
2643 // Non-leaves can block and take safepoints:
2644 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2645 }
2646 // Non-leaves can throw exceptions:
2647 if (has_io) {
2648 call->set_req(TypeFunc::I_O, i_o());
2649 }
2650
2651 if (flags & RC_UNCOMMON) {
2652 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2653 // (An "if" probability corresponds roughly to an unconditional count.
2654 // Sort of.)
2655 call->set_cnt(PROB_UNLIKELY_MAG(4));
2656 }
2657
2658 Node* c = _gvn.transform(call);
2659 assert(c == call, "cannot disappear");
2660
2661 if (wide_out) {
2662 // Slow path call has full side-effects.
2663 set_predefined_output_for_runtime_call(call);
2664 } else {
2665 // Slow path call has few side-effects, and/or sets few values.
2666 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2667 }
2668
2669 if (has_io) {
2670 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2671 }
2672 return call;
2673
2674 }
2675
2676 // i2b
2677 Node* GraphKit::sign_extend_byte(Node* in) {
2678 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2679 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2680 }
2681
2682 // i2s
2683 Node* GraphKit::sign_extend_short(Node* in) {
2684 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2685 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2686 }
2687
2688
2689 //------------------------------merge_memory-----------------------------------
2690 // Merge memory from one path into the current memory state.
2691 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2692 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2693 Node* old_slice = mms.force_memory();
2694 Node* new_slice = mms.memory2();
2695 if (old_slice != new_slice) {
2696 PhiNode* phi;
2697 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2698 if (mms.is_empty()) {
2699 // clone base memory Phi's inputs for this memory slice
2700 assert(old_slice == mms.base_memory(), "sanity");
2701 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2702 _gvn.set_type(phi, Type::MEMORY);
2703 for (uint i = 1; i < phi->req(); i++) {
2704 phi->init_req(i, old_slice->in(i));
2705 }
2706 } else {
2707 phi = old_slice->as_Phi(); // Phi was generated already
2708 }
2709 } else {
2710 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2711 _gvn.set_type(phi, Type::MEMORY);
2712 }
2713 phi->set_req(new_path, new_slice);
2714 mms.set_memory(phi);
2715 }
2716 }
2717 }
2718
2719 //------------------------------make_slow_call_ex------------------------------
2720 // Make the exception handler hookups for the slow call
2721 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2722 if (stopped()) return;
2723
2724 // Make a catch node with just two handlers: fall-through and catch-all
2725 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2726 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2727 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2728 _gvn.set_type_bottom(norm);
2729 C->record_for_igvn(norm);
2730 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2731
2732 { PreserveJVMState pjvms(this);
2733 set_control(excp);
2734 set_i_o(i_o);
2735
2736 if (excp != top()) {
2737 if (deoptimize) {
2738 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2739 uncommon_trap(Deoptimization::Reason_unhandled,
2740 Deoptimization::Action_none);
2741 } else {
2742 // Create an exception state also.
2743 // Use an exact type if the caller has a specific exception.
2744 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2745 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2746 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2747 }
2748 }
2749 }
2750
2751 // Get the no-exception control from the CatchNode.
2752 set_control(norm);
2753 }
2754
2755 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2756 Node* cmp = NULL;
2757 switch(bt) {
2758 case T_INT: cmp = new CmpINode(in1, in2); break;
2759 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2760 default: fatal("unexpected comparison type %s", type2name(bt));
2761 }
2762 gvn.transform(cmp);
2763 Node* bol = gvn.transform(new BoolNode(cmp, test));
2764 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2765 gvn.transform(iff);
2766 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2767 return iff;
2768 }
2769
2770 //-------------------------------gen_subtype_check-----------------------------
2771 // Generate a subtyping check. Takes as input the subtype and supertype.
2772 // Returns 2 values: sets the default control() to the true path and returns
2773 // the false path. Only reads invariant memory; sets no (visible) memory.
2774 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2775 // but that's not exposed to the optimizer. This call also doesn't take in an
2776 // Object; if you wish to check an Object you need to load the Object's class
2777 // prior to coming here.
2778 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn) {
2779 Compile* C = gvn.C;
2780 if ((*ctrl)->is_top()) {
2781 return C->top();
2782 }
2783
2784 // Fast check for identical types, perhaps identical constants.
2785 // The types can even be identical non-constants, in cases
2786 // involving Array.newInstance, Object.clone, etc.
2787 if (subklass == superklass)
2788 return C->top(); // false path is dead; no test needed.
2789
2790 if (gvn.type(superklass)->singleton()) {
2791 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2792 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2793
2794 // In the common case of an exact superklass, try to fold up the
2795 // test before generating code. You may ask, why not just generate
2796 // the code and then let it fold up? The answer is that the generated
2797 // code will necessarily include null checks, which do not always
2798 // completely fold away. If they are also needless, then they turn
2799 // into a performance loss. Example:
2800 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2801 // Here, the type of 'fa' is often exact, so the store check
2802 // of fa[1]=x will fold up, without testing the nullness of x.
2803 switch (C->static_subtype_check(superk, subk)) {
2804 case Compile::SSC_always_false:
2805 {
2806 Node* always_fail = *ctrl;
2807 *ctrl = gvn.C->top();
2808 return always_fail;
2809 }
2810 case Compile::SSC_always_true:
2811 return C->top();
2812 case Compile::SSC_easy_test:
2813 {
2814 // Just do a direct pointer compare and be done.
2815 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2816 *ctrl = gvn.transform(new IfTrueNode(iff));
2817 return gvn.transform(new IfFalseNode(iff));
2818 }
2819 case Compile::SSC_full_test:
2820 break;
2821 default:
2822 ShouldNotReachHere();
2823 }
2824 }
2825
2826 // %%% Possible further optimization: Even if the superklass is not exact,
2827 // if the subklass is the unique subtype of the superklass, the check
2828 // will always succeed. We could leave a dependency behind to ensure this.
2829
2830 // First load the super-klass's check-offset
2831 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2832 Node* m = C->immutable_memory();
2833 Node *chk_off = gvn.transform(new LoadINode(NULL, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2834 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2835 bool might_be_cache = (gvn.find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2836
2837 // Load from the sub-klass's super-class display list, or a 1-word cache of
2838 // the secondary superclass list, or a failing value with a sentinel offset
2839 // if the super-klass is an interface or exceptionally deep in the Java
2840 // hierarchy and we have to scan the secondary superclass list the hard way.
2841 // Worst-case type is a little odd: NULL is allowed as a result (usually
2842 // klass loads can never produce a NULL).
2843 Node *chk_off_X = chk_off;
2844 #ifdef _LP64
2845 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2846 #endif
2847 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2848 // For some types like interfaces the following loadKlass is from a 1-word
2849 // cache which is mutable so can't use immutable memory. Other
2850 // types load from the super-class display table which is immutable.
2851 Node *kmem = C->immutable_memory();
2852 // secondary_super_cache is not immutable but can be treated as such because:
2853 // - no ideal node writes to it in a way that could cause an
2854 // incorrect/missed optimization of the following Load.
2855 // - it's a cache so, worse case, not reading the latest value
2856 // wouldn't cause incorrect execution
2857 if (might_be_cache && mem != NULL) {
2858 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
2859 }
2860 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, NULL, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
2861
2862 // Compile speed common case: ARE a subtype and we canNOT fail
2863 if( superklass == nkls )
2864 return C->top(); // false path is dead; no test needed.
2865
2866 // See if we get an immediate positive hit. Happens roughly 83% of the
2867 // time. Test to see if the value loaded just previously from the subklass
2868 // is exactly the superklass.
2869 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2870 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
2871 *ctrl = gvn.transform(new IfFalseNode(iff1));
2872
2873 // Compile speed common case: Check for being deterministic right now. If
2874 // chk_off is a constant and not equal to cacheoff then we are NOT a
2875 // subklass. In this case we need exactly the 1 test above and we can
2876 // return those results immediately.
2877 if (!might_be_cache) {
2878 Node* not_subtype_ctrl = *ctrl;
2879 *ctrl = iftrue1; // We need exactly the 1 test above
2880 return not_subtype_ctrl;
2881 }
2882
2883 // Gather the various success & failures here
2884 RegionNode *r_ok_subtype = new RegionNode(4);
2885 gvn.record_for_igvn(r_ok_subtype);
2886 RegionNode *r_not_subtype = new RegionNode(3);
2887 gvn.record_for_igvn(r_not_subtype);
2888
2889 r_ok_subtype->init_req(1, iftrue1);
2890
2891 // Check for immediate negative hit. Happens roughly 11% of the time (which
2892 // is roughly 63% of the remaining cases). Test to see if the loaded
2893 // check-offset points into the subklass display list or the 1-element
2894 // cache. If it points to the display (and NOT the cache) and the display
2895 // missed then it's not a subtype.
2896 Node *cacheoff = gvn.intcon(cacheoff_con);
2897 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2898 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2899 *ctrl = gvn.transform(new IfFalseNode(iff2));
2900
2901 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2902 // No performance impact (too rare) but allows sharing of secondary arrays
2903 // which has some footprint reduction.
2904 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2905 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2906 *ctrl = gvn.transform(new IfFalseNode(iff3));
2907
2908 // -- Roads not taken here: --
2909 // We could also have chosen to perform the self-check at the beginning
2910 // of this code sequence, as the assembler does. This would not pay off
2911 // the same way, since the optimizer, unlike the assembler, can perform
2912 // static type analysis to fold away many successful self-checks.
2913 // Non-foldable self checks work better here in second position, because
2914 // the initial primary superclass check subsumes a self-check for most
2915 // types. An exception would be a secondary type like array-of-interface,
2916 // which does not appear in its own primary supertype display.
2917 // Finally, we could have chosen to move the self-check into the
2918 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2919 // dependent manner. But it is worthwhile to have the check here,
2920 // where it can be perhaps be optimized. The cost in code space is
2921 // small (register compare, branch).
2922
2923 // Now do a linear scan of the secondary super-klass array. Again, no real
2924 // performance impact (too rare) but it's gotta be done.
2925 // Since the code is rarely used, there is no penalty for moving it
2926 // out of line, and it can only improve I-cache density.
2927 // The decision to inline or out-of-line this final check is platform
2928 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2929 Node* psc = gvn.transform(
2930 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2931
2932 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2933 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2934 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2935
2936 // Return false path; set default control to true path.
2937 *ctrl = gvn.transform(r_ok_subtype);
2938 return gvn.transform(r_not_subtype);
2939 }
2940
2941 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2942 const Type* sub_t = _gvn.type(obj_or_subklass);
2943 if (sub_t->make_oopptr() != NULL && sub_t->make_oopptr()->is_inlinetypeptr()) {
2944 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
2945 obj_or_subklass = makecon(sub_t);
2946 }
2947 bool expand_subtype_check = C->post_loop_opts_phase() || // macro node expansion is over
2948 ExpandSubTypeCheckAtParseTime; // forced expansion
2949 if (expand_subtype_check) {
2950 MergeMemNode* mem = merged_memory();
2951 Node* ctrl = control();
2952 Node* subklass = obj_or_subklass;
2953 if (!sub_t->isa_klassptr()) {
2954 subklass = load_object_klass(obj_or_subklass);
2955 }
2956 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn);
2957 set_control(ctrl);
2958 return n;
2959 }
2960
2961 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass));
2962 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2963 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2964 set_control(_gvn.transform(new IfTrueNode(iff)));
2965 return _gvn.transform(new IfFalseNode(iff));
2966 }
2967
2968 // Profile-driven exact type check:
2969 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2970 float prob, Node* *casted_receiver) {
2971 assert(!klass->is_interface(), "no exact type check on interfaces");
2972 Node* fail = top();
2973 const Type* rec_t = _gvn.type(receiver);
2974 if (rec_t->is_inlinetypeptr()) {
2975 if (klass->equals(rec_t->inline_klass())) {
2976 (*casted_receiver) = receiver; // Always passes
2977 } else {
2978 (*casted_receiver) = top(); // Always fails
2979 fail = control();
2980 set_control(top());
2981 }
2982 return fail;
2983 }
2984 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2985 Node* recv_klass = load_object_klass(receiver);
2986 fail = type_check(recv_klass, tklass, prob);
2987
2988 if (!stopped()) {
2989 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2990 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2991 assert(recv_xtype->klass_is_exact(), "");
2992
2993 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
2994 // Subsume downstream occurrences of receiver with a cast to
2995 // recv_xtype, since now we know what the type will be.
2996 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2997 Node* res = _gvn.transform(cast);
2998 if (recv_xtype->is_inlinetypeptr()) {
2999 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3000 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3001 }
3002 (*casted_receiver) = res;
3003 // (User must make the replace_in_map call.)
3004 }
3005 }
3006
3007 return fail;
3008 }
3009
3010 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3011 float prob) {
3012 Node* want_klass = makecon(tklass);
3013 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3014 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3015 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3016 set_control(_gvn.transform(new IfTrueNode (iff)));
3017 Node* fail = _gvn.transform(new IfFalseNode(iff));
3018 return fail;
3019 }
3020
3021 //------------------------------subtype_check_receiver-------------------------
3022 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3023 Node** casted_receiver) {
3024 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
3025 Node* want_klass = makecon(tklass);
3026
3027 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3028
3029 // Ignore interface type information until interface types are properly tracked.
3030 if (!stopped() && !klass->is_interface()) {
3031 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3032 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3033 if (receiver_type != NULL && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3034 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3035 if (recv_type->is_inlinetypeptr()) {
3036 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3037 }
3038 (*casted_receiver) = cast;
3039 }
3040 }
3041
3042 return slow_ctl;
3043 }
3044
3045 //------------------------------seems_never_null-------------------------------
3046 // Use null_seen information if it is available from the profile.
3047 // If we see an unexpected null at a type check we record it and force a
3048 // recompile; the offending check will be recompiled to handle NULLs.
3049 // If we see several offending BCIs, then all checks in the
3050 // method will be recompiled.
3051 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3052 speculating = !_gvn.type(obj)->speculative_maybe_null();
3053 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3054 if (UncommonNullCast // Cutout for this technique
3055 && obj != null() // And not the -Xcomp stupid case?
3056 && !too_many_traps(reason)
3057 ) {
3058 if (speculating) {
3059 return true;
3060 }
3061 if (data == NULL)
3062 // Edge case: no mature data. Be optimistic here.
3063 return true;
3064 // If the profile has not seen a null, assume it won't happen.
3065 assert(java_bc() == Bytecodes::_checkcast ||
3066 java_bc() == Bytecodes::_instanceof ||
3067 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3068 if (java_bc() == Bytecodes::_aastore) {
3069 return ((ciArrayLoadStoreData*)data->as_ArrayLoadStoreData())->element()->ptr_kind() == ProfileNeverNull;
3070 }
3071 return !data->as_BitData()->null_seen();
3072 }
3073 speculating = false;
3074 return false;
3075 }
3076
3077 void GraphKit::guard_klass_being_initialized(Node* klass) {
3078 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3079 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3080 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
3081 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3082 T_BYTE, MemNode::unordered);
3083 init_state = _gvn.transform(init_state);
3084
3085 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3086
3087 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3088 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3089
3090 { BuildCutout unless(this, tst, PROB_MAX);
3091 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3092 }
3093 }
3094
3095 void GraphKit::guard_init_thread(Node* klass) {
3096 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3097 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3098
3099 Node* init_thread = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
3100 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3101 T_ADDRESS, MemNode::unordered);
3102 init_thread = _gvn.transform(init_thread);
3103
3104 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3105
3106 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3107 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3108
3109 { BuildCutout unless(this, tst, PROB_MAX);
3110 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3111 }
3112 }
3113
3114 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3115 if (ik->is_being_initialized()) {
3116 if (C->needs_clinit_barrier(ik, context)) {
3117 Node* klass = makecon(TypeKlassPtr::make(ik));
3118 guard_klass_being_initialized(klass);
3119 guard_init_thread(klass);
3120 insert_mem_bar(Op_MemBarCPUOrder);
3121 }
3122 } else if (ik->is_initialized()) {
3123 return; // no barrier needed
3124 } else {
3125 uncommon_trap(Deoptimization::Reason_uninitialized,
3126 Deoptimization::Action_reinterpret,
3127 NULL);
3128 }
3129 }
3130
3131 //------------------------maybe_cast_profiled_receiver-------------------------
3132 // If the profile has seen exactly one type, narrow to exactly that type.
3133 // Subsequent type checks will always fold up.
3134 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3135 const TypeKlassPtr* require_klass,
3136 ciKlass* spec_klass,
3137 bool safe_for_replace) {
3138 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
3139
3140 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
3141
3142 // Make sure we haven't already deoptimized from this tactic.
3143 if (too_many_traps_or_recompiles(reason))
3144 return NULL;
3145
3146 // (No, this isn't a call, but it's enough like a virtual call
3147 // to use the same ciMethod accessor to get the profile info...)
3148 // If we have a speculative type use it instead of profiling (which
3149 // may not help us)
3150 ciKlass* exact_kls = spec_klass;
3151 if (exact_kls == NULL) {
3152 if (java_bc() == Bytecodes::_aastore) {
3153 ciKlass* array_type = NULL;
3154 ciKlass* element_type = NULL;
3155 ProfilePtrKind element_ptr = ProfileMaybeNull;
3156 bool flat_array = true;
3157 bool null_free_array = true;
3158 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3159 exact_kls = element_type;
3160 } else {
3161 exact_kls = profile_has_unique_klass();
3162 }
3163 }
3164 if (exact_kls != NULL) {// no cast failures here
3165 if (require_klass == NULL ||
3166 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls)) == Compile::SSC_always_true) {
3167 // If we narrow the type to match what the type profile sees or
3168 // the speculative type, we can then remove the rest of the
3169 // cast.
3170 // This is a win, even if the exact_kls is very specific,
3171 // because downstream operations, such as method calls,
3172 // will often benefit from the sharper type.
3173 Node* exact_obj = not_null_obj; // will get updated in place...
3174 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3175 &exact_obj);
3176 { PreserveJVMState pjvms(this);
3177 set_control(slow_ctl);
3178 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3179 }
3180 if (safe_for_replace) {
3181 replace_in_map(not_null_obj, exact_obj);
3182 }
3183 return exact_obj;
3184 }
3185 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3186 }
3187
3188 return NULL;
3189 }
3190
3191 /**
3192 * Cast obj to type and emit guard unless we had too many traps here
3193 * already
3194 *
3195 * @param obj node being casted
3196 * @param type type to cast the node to
3197 * @param not_null true if we know node cannot be null
3198 */
3199 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3200 ciKlass* type,
3201 bool not_null) {
3202 if (stopped()) {
3203 return obj;
3204 }
3205
3206 // type == NULL if profiling tells us this object is always null
3207 if (type != NULL) {
3208 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3209 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3210
3211 if (!too_many_traps_or_recompiles(null_reason) &&
3212 !too_many_traps_or_recompiles(class_reason)) {
3213 Node* not_null_obj = NULL;
3214 // not_null is true if we know the object is not null and
3215 // there's no need for a null check
3216 if (!not_null) {
3217 Node* null_ctl = top();
3218 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3219 assert(null_ctl->is_top(), "no null control here");
3220 } else {
3221 not_null_obj = obj;
3222 }
3223
3224 Node* exact_obj = not_null_obj;
3225 ciKlass* exact_kls = type;
3226 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3227 &exact_obj);
3228 {
3229 PreserveJVMState pjvms(this);
3230 set_control(slow_ctl);
3231 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3232 }
3233 replace_in_map(not_null_obj, exact_obj);
3234 obj = exact_obj;
3235 }
3236 } else {
3237 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3238 Node* exact_obj = null_assert(obj);
3239 replace_in_map(obj, exact_obj);
3240 obj = exact_obj;
3241 }
3242 }
3243 return obj;
3244 }
3245
3246 //-------------------------------gen_instanceof--------------------------------
3247 // Generate an instance-of idiom. Used by both the instance-of bytecode
3248 // and the reflective instance-of call.
3249 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3250 kill_dead_locals(); // Benefit all the uncommon traps
3251 assert( !stopped(), "dead parse path should be checked in callers" );
3252 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3253 "must check for not-null not-dead klass in callers");
3254
3255 // Make the merge point
3256 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3257 RegionNode* region = new RegionNode(PATH_LIMIT);
3258 Node* phi = new PhiNode(region, TypeInt::BOOL);
3259 C->set_has_split_ifs(true); // Has chance for split-if optimization
3260
3261 ciProfileData* data = NULL;
3262 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3263 data = method()->method_data()->bci_to_data(bci());
3264 }
3265 bool speculative_not_null = false;
3266 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3267 && seems_never_null(obj, data, speculative_not_null));
3268
3269 // Null check; get casted pointer; set region slot 3
3270 Node* null_ctl = top();
3271 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3272
3273 // If not_null_obj is dead, only null-path is taken
3274 if (stopped()) { // Doing instance-of on a NULL?
3275 set_control(null_ctl);
3276 return intcon(0);
3277 }
3278 region->init_req(_null_path, null_ctl);
3279 phi ->init_req(_null_path, intcon(0)); // Set null path value
3280 if (null_ctl == top()) {
3281 // Do this eagerly, so that pattern matches like is_diamond_phi
3282 // will work even during parsing.
3283 assert(_null_path == PATH_LIMIT-1, "delete last");
3284 region->del_req(_null_path);
3285 phi ->del_req(_null_path);
3286 }
3287
3288 // Do we know the type check always succeed?
3289 bool known_statically = false;
3290 if (_gvn.type(superklass)->singleton()) {
3291 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3292 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3293 if (subk != NULL && subk->is_loaded()) {
3294 int static_res = C->static_subtype_check(superk, subk);
3295 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3296 }
3297 }
3298
3299 if (!known_statically) {
3300 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3301 // We may not have profiling here or it may not help us. If we
3302 // have a speculative type use it to perform an exact cast.
3303 ciKlass* spec_obj_type = obj_type->speculative_type();
3304 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3305 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3306 if (stopped()) { // Profile disagrees with this path.
3307 set_control(null_ctl); // Null is the only remaining possibility.
3308 return intcon(0);
3309 }
3310 if (cast_obj != NULL) {
3311 not_null_obj = cast_obj;
3312 }
3313 }
3314 }
3315
3316 // Generate the subtype check
3317 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3318
3319 // Plug in the success path to the general merge in slot 1.
3320 region->init_req(_obj_path, control());
3321 phi ->init_req(_obj_path, intcon(1));
3322
3323 // Plug in the failing path to the general merge in slot 2.
3324 region->init_req(_fail_path, not_subtype_ctrl);
3325 phi ->init_req(_fail_path, intcon(0));
3326
3327 // Return final merged results
3328 set_control( _gvn.transform(region) );
3329 record_for_igvn(region);
3330
3331 // If we know the type check always succeeds then we don't use the
3332 // profiling data at this bytecode. Don't lose it, feed it to the
3333 // type system as a speculative type.
3334 if (safe_for_replace) {
3335 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3336 replace_in_map(obj, casted_obj);
3337 }
3338
3339 return _gvn.transform(phi);
3340 }
3341
3342 //-------------------------------gen_checkcast---------------------------------
3343 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3344 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3345 // uncommon-trap paths work. Adjust stack after this call.
3346 // If failure_control is supplied and not null, it is filled in with
3347 // the control edge for the cast failure. Otherwise, an appropriate
3348 // uncommon trap or exception is thrown.
3349 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool null_free) {
3350 kill_dead_locals(); // Benefit all the uncommon traps
3351 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr();
3352 const TypeOopPtr* toop = tk->cast_to_exactness(false)->as_instance_type();
3353 bool safe_for_replace = (failure_control == NULL);
3354 assert(!null_free || toop->is_inlinetypeptr(), "must be an inline type pointer");
3355
3356 // Fast cutout: Check the case that the cast is vacuously true.
3357 // This detects the common cases where the test will short-circuit
3358 // away completely. We do this before we perform the null check,
3359 // because if the test is going to turn into zero code, we don't
3360 // want a residual null check left around. (Causes a slowdown,
3361 // for example, in some objArray manipulations, such as a[i]=a[j].)
3362 if (tk->singleton()) {
3363 const TypeKlassPtr* kptr = NULL;
3364 const Type* t = _gvn.type(obj);
3365 if (t->isa_oop_ptr()) {
3366 kptr = t->is_oopptr()->as_klass_type();
3367 } else if (obj->is_InlineType()) {
3368 ciInlineKlass* vk = t->inline_klass();
3369 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0), vk->flatten_array());
3370 }
3371 if (kptr != NULL) {
3372 switch (C->static_subtype_check(tk, kptr)) {
3373 case Compile::SSC_always_true:
3374 // If we know the type check always succeed then we don't use
3375 // the profiling data at this bytecode. Don't lose it, feed it
3376 // to the type system as a speculative type.
3377 obj = record_profiled_receiver_for_speculation(obj);
3378 if (null_free) {
3379 assert(safe_for_replace, "must be");
3380 obj = null_check(obj);
3381 }
3382 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3383 return obj;
3384 case Compile::SSC_always_false:
3385 if (null_free) {
3386 assert(safe_for_replace, "must be");
3387 obj = null_check(obj);
3388 }
3389 // It needs a null check because a null will *pass* the cast check.
3390 if (t->isa_oopptr() != NULL && !t->is_oopptr()->maybe_null()) {
3391 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3392 Deoptimization::DeoptReason reason = is_aastore ?
3393 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3394 builtin_throw(reason);
3395 return top();
3396 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3397 return null_assert(obj);
3398 }
3399 break; // Fall through to full check
3400 default:
3401 break;
3402 }
3403 }
3404 }
3405
3406 ciProfileData* data = NULL;
3407 if (failure_control == NULL) { // use MDO in regular case only
3408 assert(java_bc() == Bytecodes::_aastore ||
3409 java_bc() == Bytecodes::_checkcast,
3410 "interpreter profiles type checks only for these BCs");
3411 if (method()->method_data()->is_mature()) {
3412 data = method()->method_data()->bci_to_data(bci());
3413 }
3414 }
3415
3416 // Make the merge point
3417 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3418 RegionNode* region = new RegionNode(PATH_LIMIT);
3419 Node* phi = new PhiNode(region, toop);
3420 _gvn.set_type(region, Type::CONTROL);
3421 _gvn.set_type(phi, toop);
3422
3423 C->set_has_split_ifs(true); // Has chance for split-if optimization
3424
3425 // Use null-cast information if it is available
3426 bool speculative_not_null = false;
3427 bool never_see_null = ((failure_control == NULL) // regular case only
3428 && seems_never_null(obj, data, speculative_not_null));
3429
3430 if (obj->is_InlineType()) {
3431 // Re-execute if buffering during triggers deoptimization
3432 PreserveReexecuteState preexecs(this);
3433 jvms()->set_should_reexecute(true);
3434 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3435 }
3436
3437 // Null check; get casted pointer; set region slot 3
3438 Node* null_ctl = top();
3439 Node* not_null_obj = NULL;
3440 if (null_free) {
3441 assert(safe_for_replace, "must be");
3442 not_null_obj = null_check(obj);
3443 } else {
3444 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3445 }
3446
3447 // If not_null_obj is dead, only null-path is taken
3448 if (stopped()) { // Doing instance-of on a NULL?
3449 set_control(null_ctl);
3450 if (toop->is_inlinetypeptr()) {
3451 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3452 }
3453 return null();
3454 }
3455 region->init_req(_null_path, null_ctl);
3456 phi ->init_req(_null_path, null()); // Set null path value
3457 if (null_ctl == top()) {
3458 // Do this eagerly, so that pattern matches like is_diamond_phi
3459 // will work even during parsing.
3460 assert(_null_path == PATH_LIMIT-1, "delete last");
3461 region->del_req(_null_path);
3462 phi ->del_req(_null_path);
3463 }
3464
3465 Node* cast_obj = NULL;
3466 if (tk->klass_is_exact()) {
3467 // The following optimization tries to statically cast the speculative type of the object
3468 // (for example obtained during profiling) to the type of the superklass and then do a
3469 // dynamic check that the type of the object is what we expect. To work correctly
3470 // for checkcast and aastore the type of superklass should be exact.
3471 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3472 // We may not have profiling here or it may not help us. If we have
3473 // a speculative type use it to perform an exact cast.
3474 ciKlass* spec_obj_type = obj_type->speculative_type();
3475 if (spec_obj_type != NULL || data != NULL) {
3476 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk, spec_obj_type, safe_for_replace);
3477 if (cast_obj != NULL) {
3478 if (failure_control != NULL) // failure is now impossible
3479 (*failure_control) = top();
3480 // adjust the type of the phi to the exact klass:
3481 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3482 }
3483 }
3484 }
3485
3486 if (cast_obj == NULL) {
3487 // Generate the subtype check
3488 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3489
3490 // Plug in success path into the merge
3491 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3492 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3493 if (failure_control == NULL) {
3494 if (not_subtype_ctrl != top()) { // If failure is possible
3495 PreserveJVMState pjvms(this);
3496 set_control(not_subtype_ctrl);
3497 Node* obj_klass = NULL;
3498 if (not_null_obj->is_InlineType()) {
3499 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
3500 } else {
3501 obj_klass = load_object_klass(not_null_obj);
3502 }
3503 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3504 Deoptimization::DeoptReason reason = is_aastore ?
3505 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3506 builtin_throw(reason);
3507 }
3508 } else {
3509 (*failure_control) = not_subtype_ctrl;
3510 }
3511 }
3512
3513 region->init_req(_obj_path, control());
3514 phi ->init_req(_obj_path, cast_obj);
3515
3516 // A merge of NULL or Casted-NotNull obj
3517 Node* res = _gvn.transform(phi);
3518
3519 // Note I do NOT always 'replace_in_map(obj,result)' here.
3520 // if( tk->klass()->can_be_primary_super() )
3521 // This means that if I successfully store an Object into an array-of-String
3522 // I 'forget' that the Object is really now known to be a String. I have to
3523 // do this because we don't have true union types for interfaces - if I store
3524 // a Baz into an array-of-Interface and then tell the optimizer it's an
3525 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3526 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3527 // replace_in_map( obj, res );
3528
3529 // Return final merged results
3530 set_control( _gvn.transform(region) );
3531 record_for_igvn(region);
3532
3533 bool not_inline = !toop->can_be_inline_type();
3534 bool not_flattened = !UseFlatArray || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flatten_array());
3535 if (EnableValhalla && not_flattened) {
3536 // Check if obj has been loaded from an array
3537 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3538 Node* array = NULL;
3539 if (obj->isa_Load()) {
3540 Node* address = obj->in(MemNode::Address);
3541 if (address->isa_AddP()) {
3542 array = address->as_AddP()->in(AddPNode::Base);
3543 }
3544 } else if (obj->is_Phi()) {
3545 Node* region = obj->in(0);
3546 // TODO make this more robust (see JDK-8231346)
3547 if (region->req() == 3 && region->in(2) != NULL && region->in(2)->in(0) != NULL) {
3548 IfNode* iff = region->in(2)->in(0)->isa_If();
3549 if (iff != NULL) {
3550 iff->is_flat_array_check(&_gvn, &array);
3551 }
3552 }
3553 }
3554 if (array != NULL) {
3555 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3556 if (ary_t != NULL) {
3557 if (!ary_t->is_not_null_free() && not_inline) {
3558 // Casting array element to a non-inline-type, mark array as not null-free.
3559 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3560 replace_in_map(array, cast);
3561 } else if (!ary_t->is_not_flat()) {
3562 // Casting array element to a non-flattened type, mark array as not flat.
3563 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3564 replace_in_map(array, cast);
3565 }
3566 }
3567 }
3568 }
3569
3570 if (!stopped() && !res->is_InlineType()) {
3571 res = record_profiled_receiver_for_speculation(res);
3572 if (toop->is_inlinetypeptr()) {
3573 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass(), !gvn().type(res)->maybe_null());
3574 res = vt;
3575 if (safe_for_replace) {
3576 replace_in_map(obj, vt);
3577 replace_in_map(not_null_obj, vt);
3578 replace_in_map(res, vt);
3579 }
3580 }
3581 }
3582 return res;
3583 }
3584
3585 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3586 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3587 Node* mark = make_load(NULL, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3588 Node* mask = MakeConX(markWord::inline_type_pattern);
3589 Node* masked = _gvn.transform(new AndXNode(mark, mask));
3590 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3591 return _gvn.transform(new BoolNode(cmp, is_inline ? BoolTest::eq : BoolTest::ne));
3592 }
3593
3594 Node* GraphKit::is_val_mirror(Node* mirror) {
3595 Node* p = basic_plus_adr(mirror, java_lang_Class::secondary_mirror_offset());
3596 Node* secondary_mirror = access_load_at(mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR), T_OBJECT, IN_HEAP);
3597 Node* cmp = _gvn.transform(new CmpPNode(mirror, secondary_mirror));
3598 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3599 }
3600
3601 Node* GraphKit::array_lh_test(Node* klass, jint mask, jint val, bool eq) {
3602 Node* lh_adr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
3603 // Make sure to use immutable memory here to enable hoisting the check out of loops
3604 Node* lh_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lh_adr, lh_adr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3605 Node* masked = _gvn.transform(new AndINode(lh_val, intcon(mask)));
3606 Node* cmp = _gvn.transform(new CmpINode(masked, intcon(val)));
3607 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3608 }
3609
3610 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3611 // We can't use immutable memory here because the mark word is mutable.
3612 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3613 // check is moved out of loops (mainly to enable loop unswitching).
3614 Node* mem = UseArrayMarkWordCheck ? memory(Compile::AliasIdxRaw) : immutable_memory();
3615 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, mem, array_or_klass));
3616 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3617 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3618 }
3619
3620 Node* GraphKit::null_free_array_test(Node* klass, bool null_free) {
3621 return array_lh_test(klass, Klass::_lh_null_free_array_bit_inplace, 0, !null_free);
3622 }
3623
3624 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3625 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3626 RegionNode* region = new RegionNode(3);
3627 Node* null_ctl = top();
3628 null_check_oop(val, &null_ctl);
3629 if (null_ctl != top()) {
3630 PreserveJVMState pjvms(this);
3631 set_control(null_ctl);
3632 {
3633 // Deoptimize if null-free array
3634 BuildCutout unless(this, null_free_array_test(load_object_klass(ary), /* null_free = */ false), PROB_MAX);
3635 inc_sp(nargs);
3636 uncommon_trap(Deoptimization::Reason_null_check,
3637 Deoptimization::Action_none);
3638 }
3639 region->init_req(1, control());
3640 }
3641 region->init_req(2, control());
3642 set_control(_gvn.transform(region));
3643 record_for_igvn(region);
3644 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3645 // Since we were just successfully storing null, the array can't be null free.
3646 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3647 ary_t = ary_t->cast_to_not_null_free();
3648 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3649 if (safe_for_replace) {
3650 replace_in_map(ary, cast);
3651 }
3652 ary = cast;
3653 }
3654 return ary;
3655 }
3656
3657 //------------------------------next_monitor-----------------------------------
3658 // What number should be given to the next monitor?
3659 int GraphKit::next_monitor() {
3660 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3661 int next = current + C->sync_stack_slots();
3662 // Keep the toplevel high water mark current:
3663 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3664 return current;
3665 }
3666
3667 //------------------------------insert_mem_bar---------------------------------
3668 // Memory barrier to avoid floating things around
3669 // The membar serves as a pinch point between both control and all memory slices.
3670 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3671 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3672 mb->init_req(TypeFunc::Control, control());
3673 mb->init_req(TypeFunc::Memory, reset_memory());
3674 Node* membar = _gvn.transform(mb);
3675 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3676 set_all_memory_call(membar);
3677 return membar;
3678 }
3679
3680 //-------------------------insert_mem_bar_volatile----------------------------
3681 // Memory barrier to avoid floating things around
3682 // The membar serves as a pinch point between both control and memory(alias_idx).
3683 // If you want to make a pinch point on all memory slices, do not use this
3684 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3685 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3686 // When Parse::do_put_xxx updates a volatile field, it appends a series
3687 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3688 // The first membar is on the same memory slice as the field store opcode.
3689 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3690 // All the other membars (for other volatile slices, including AliasIdxBot,
3691 // which stands for all unknown volatile slices) are control-dependent
3692 // on the first membar. This prevents later volatile loads or stores
3693 // from sliding up past the just-emitted store.
3694
3695 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3696 mb->set_req(TypeFunc::Control,control());
3697 if (alias_idx == Compile::AliasIdxBot) {
3698 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3699 } else {
3700 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3701 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3702 }
3703 Node* membar = _gvn.transform(mb);
3704 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3705 if (alias_idx == Compile::AliasIdxBot) {
3706 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3707 } else {
3708 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3709 }
3710 return membar;
3711 }
3712
3713 //------------------------------shared_lock------------------------------------
3714 // Emit locking code.
3715 FastLockNode* GraphKit::shared_lock(Node* obj) {
3716 // bci is either a monitorenter bc or InvocationEntryBci
3717 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3718 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3719
3720 if( !GenerateSynchronizationCode )
3721 return NULL; // Not locking things?
3722
3723 if (stopped()) // Dead monitor?
3724 return NULL;
3725
3726 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3727
3728 // Box the stack location
3729 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3730 Node* mem = reset_memory();
3731
3732 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3733
3734 // Create the rtm counters for this fast lock if needed.
3735 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3736
3737 // Add monitor to debug info for the slow path. If we block inside the
3738 // slow path and de-opt, we need the monitor hanging around
3739 map()->push_monitor( flock );
3740
3741 const TypeFunc *tf = LockNode::lock_type();
3742 LockNode *lock = new LockNode(C, tf);
3743
3744 lock->init_req( TypeFunc::Control, control() );
3745 lock->init_req( TypeFunc::Memory , mem );
3746 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3747 lock->init_req( TypeFunc::FramePtr, frameptr() );
3748 lock->init_req( TypeFunc::ReturnAdr, top() );
3749
3750 lock->init_req(TypeFunc::Parms + 0, obj);
3751 lock->init_req(TypeFunc::Parms + 1, box);
3752 lock->init_req(TypeFunc::Parms + 2, flock);
3753 add_safepoint_edges(lock);
3754
3755 lock = _gvn.transform( lock )->as_Lock();
3756
3757 // lock has no side-effects, sets few values
3758 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3759
3760 insert_mem_bar(Op_MemBarAcquireLock);
3761
3762 // Add this to the worklist so that the lock can be eliminated
3763 record_for_igvn(lock);
3764
3765 #ifndef PRODUCT
3766 if (PrintLockStatistics) {
3767 // Update the counter for this lock. Don't bother using an atomic
3768 // operation since we don't require absolute accuracy.
3769 lock->create_lock_counter(map()->jvms());
3770 increment_counter(lock->counter()->addr());
3771 }
3772 #endif
3773
3774 return flock;
3775 }
3776
3777
3778 //------------------------------shared_unlock----------------------------------
3779 // Emit unlocking code.
3780 void GraphKit::shared_unlock(Node* box, Node* obj) {
3781 // bci is either a monitorenter bc or InvocationEntryBci
3782 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3783 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3784
3785 if( !GenerateSynchronizationCode )
3786 return;
3787 if (stopped()) { // Dead monitor?
3788 map()->pop_monitor(); // Kill monitor from debug info
3789 return;
3790 }
3791 assert(!obj->is_InlineType(), "should not unlock on inline type");
3792
3793 // Memory barrier to avoid floating things down past the locked region
3794 insert_mem_bar(Op_MemBarReleaseLock);
3795
3796 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3797 UnlockNode *unlock = new UnlockNode(C, tf);
3798 #ifdef ASSERT
3799 unlock->set_dbg_jvms(sync_jvms());
3800 #endif
3801 uint raw_idx = Compile::AliasIdxRaw;
3802 unlock->init_req( TypeFunc::Control, control() );
3803 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3804 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3805 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3806 unlock->init_req( TypeFunc::ReturnAdr, top() );
3807
3808 unlock->init_req(TypeFunc::Parms + 0, obj);
3809 unlock->init_req(TypeFunc::Parms + 1, box);
3810 unlock = _gvn.transform(unlock)->as_Unlock();
3811
3812 Node* mem = reset_memory();
3813
3814 // unlock has no side-effects, sets few values
3815 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3816
3817 // Kill monitor from debug info
3818 map()->pop_monitor( );
3819 }
3820
3821 //-------------------------------get_layout_helper-----------------------------
3822 // If the given klass is a constant or known to be an array,
3823 // fetch the constant layout helper value into constant_value
3824 // and return (Node*)NULL. Otherwise, load the non-constant
3825 // layout helper value, and return the node which represents it.
3826 // This two-faced routine is useful because allocation sites
3827 // almost always feature constant types.
3828 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3829 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3830 if (!StressReflectiveCode && inst_klass != NULL) {
3831 bool xklass = inst_klass->klass_is_exact();
3832 bool can_be_flattened = false;
3833 const TypeAryPtr* ary_type = inst_klass->as_instance_type()->isa_aryptr();
3834 if (UseFlatArray && !xklass && ary_type != NULL && !ary_type->is_null_free()) {
3835 // The runtime type of [LMyValue might be [QMyValue due to [QMyValue <: [LMyValue. Don't constant fold.
3836 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
3837 can_be_flattened = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->flatten_array());
3838 }
3839 if (!can_be_flattened && (xklass || inst_klass->isa_aryklassptr())) {
3840 jint lhelper;
3841 if (inst_klass->is_flat()) {
3842 lhelper = ary_type->flat_layout_helper();
3843 } else if (inst_klass->isa_aryklassptr()) {
3844 BasicType elem = ary_type->elem()->array_element_basic_type();
3845 if (is_reference_type(elem, true)) {
3846 elem = T_OBJECT;
3847 }
3848 lhelper = Klass::array_layout_helper(elem);
3849 } else {
3850 lhelper = inst_klass->is_instklassptr()->exact_klass()->layout_helper();
3851 }
3852 if (lhelper != Klass::_lh_neutral_value) {
3853 constant_value = lhelper;
3854 return (Node*) NULL;
3855 }
3856 }
3857 }
3858 constant_value = Klass::_lh_neutral_value; // put in a known value
3859 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3860 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3861 }
3862
3863 // We just put in an allocate/initialize with a big raw-memory effect.
3864 // Hook selected additional alias categories on the initialization.
3865 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3866 MergeMemNode* init_in_merge,
3867 Node* init_out_raw) {
3868 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3869 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3870
3871 Node* prevmem = kit.memory(alias_idx);
3872 init_in_merge->set_memory_at(alias_idx, prevmem);
3873 if (init_out_raw != NULL) {
3874 kit.set_memory(init_out_raw, alias_idx);
3875 }
3876 }
3877
3878 //---------------------------set_output_for_allocation-------------------------
3879 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3880 const TypeOopPtr* oop_type,
3881 bool deoptimize_on_exception) {
3882 int rawidx = Compile::AliasIdxRaw;
3883 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3884 add_safepoint_edges(alloc);
3885 Node* allocx = _gvn.transform(alloc);
3886 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3887 // create memory projection for i_o
3888 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3889 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3890
3891 // create a memory projection as for the normal control path
3892 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3893 set_memory(malloc, rawidx);
3894
3895 // a normal slow-call doesn't change i_o, but an allocation does
3896 // we create a separate i_o projection for the normal control path
3897 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3898 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3899
3900 // put in an initialization barrier
3901 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3902 rawoop)->as_Initialize();
3903 assert(alloc->initialization() == init, "2-way macro link must work");
3904 assert(init ->allocation() == alloc, "2-way macro link must work");
3905 {
3906 // Extract memory strands which may participate in the new object's
3907 // initialization, and source them from the new InitializeNode.
3908 // This will allow us to observe initializations when they occur,
3909 // and link them properly (as a group) to the InitializeNode.
3910 assert(init->in(InitializeNode::Memory) == malloc, "");
3911 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3912 init->set_req(InitializeNode::Memory, minit_in);
3913 record_for_igvn(minit_in); // fold it up later, if possible
3914 _gvn.set_type(minit_in, Type::MEMORY);
3915 Node* minit_out = memory(rawidx);
3916 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3917 // Add an edge in the MergeMem for the header fields so an access
3918 // to one of those has correct memory state
3919 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3920 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3921 if (oop_type->isa_aryptr()) {
3922 const TypeAryPtr* arytype = oop_type->is_aryptr();
3923 if (arytype->is_flat()) {
3924 // Initially all flattened array accesses share a single slice
3925 // but that changes after parsing. Prepare the memory graph so
3926 // it can optimize flattened array accesses properly once they
3927 // don't share a single slice.
3928 assert(C->flattened_accesses_share_alias(), "should be set at parse time");
3929 C->set_flattened_accesses_share_alias(false);
3930 ciInlineKlass* vk = arytype->elem()->inline_klass();
3931 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
3932 ciField* field = vk->nonstatic_field_at(i);
3933 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3934 continue; // do not bother to track really large numbers of fields
3935 int off_in_vt = field->offset() - vk->first_field_offset();
3936 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
3937 int fieldidx = C->get_alias_index(adr_type, true);
3938 // Pass NULL for init_out. Having per flat array element field memory edges as uses of the Initialize node
3939 // can result in per flat array field Phis to be created which confuses the logic of
3940 // Compile::adjust_flattened_array_access_aliases().
3941 hook_memory_on_init(*this, fieldidx, minit_in, NULL);
3942 }
3943 C->set_flattened_accesses_share_alias(true);
3944 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
3945 } else {
3946 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3947 int elemidx = C->get_alias_index(telemref);
3948 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3949 }
3950 } else if (oop_type->isa_instptr()) {
3951 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
3952 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3953 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3954 ciField* field = ik->nonstatic_field_at(i);
3955 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3956 continue; // do not bother to track really large numbers of fields
3957 // Find (or create) the alias category for this field:
3958 int fieldidx = C->alias_type(field)->index();
3959 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3960 }
3961 }
3962 }
3963
3964 // Cast raw oop to the real thing...
3965 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3966 javaoop = _gvn.transform(javaoop);
3967 C->set_recent_alloc(control(), javaoop);
3968 assert(just_allocated_object(control()) == javaoop, "just allocated");
3969
3970 #ifdef ASSERT
3971 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3972 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3973 "Ideal_allocation works");
3974 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3975 "Ideal_allocation works");
3976 if (alloc->is_AllocateArray()) {
3977 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3978 "Ideal_allocation works");
3979 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3980 "Ideal_allocation works");
3981 } else {
3982 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3983 }
3984 }
3985 #endif //ASSERT
3986
3987 return javaoop;
3988 }
3989
3990 //---------------------------new_instance--------------------------------------
3991 // This routine takes a klass_node which may be constant (for a static type)
3992 // or may be non-constant (for reflective code). It will work equally well
3993 // for either, and the graph will fold nicely if the optimizer later reduces
3994 // the type to a constant.
3995 // The optional arguments are for specialized use by intrinsics:
3996 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3997 // - If 'return_size_val', report the total object size to the caller.
3998 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3999 Node* GraphKit::new_instance(Node* klass_node,
4000 Node* extra_slow_test,
4001 Node* *return_size_val,
4002 bool deoptimize_on_exception,
4003 InlineTypeNode* inline_type_node) {
4004 // Compute size in doublewords
4005 // The size is always an integral number of doublewords, represented
4006 // as a positive bytewise size stored in the klass's layout_helper.
4007 // The layout_helper also encodes (in a low bit) the need for a slow path.
4008 jint layout_con = Klass::_lh_neutral_value;
4009 Node* layout_val = get_layout_helper(klass_node, layout_con);
4010 bool layout_is_con = (layout_val == NULL);
4011
4012 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
4013 // Generate the initial go-slow test. It's either ALWAYS (return a
4014 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
4015 // case) a computed value derived from the layout_helper.
4016 Node* initial_slow_test = NULL;
4017 if (layout_is_con) {
4018 assert(!StressReflectiveCode, "stress mode does not use these paths");
4019 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4020 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4021 } else { // reflective case
4022 // This reflective path is used by Unsafe.allocateInstance.
4023 // (It may be stress-tested by specifying StressReflectiveCode.)
4024 // Basically, we want to get into the VM is there's an illegal argument.
4025 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4026 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4027 if (extra_slow_test != intcon(0)) {
4028 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4029 }
4030 // (Macro-expander will further convert this to a Bool, if necessary.)
4031 }
4032
4033 // Find the size in bytes. This is easy; it's the layout_helper.
4034 // The size value must be valid even if the slow path is taken.
4035 Node* size = NULL;
4036 if (layout_is_con) {
4037 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4038 } else { // reflective case
4039 // This reflective path is used by clone and Unsafe.allocateInstance.
4040 size = ConvI2X(layout_val);
4041
4042 // Clear the low bits to extract layout_helper_size_in_bytes:
4043 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4044 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4045 size = _gvn.transform( new AndXNode(size, mask) );
4046 }
4047 if (return_size_val != NULL) {
4048 (*return_size_val) = size;
4049 }
4050
4051 // This is a precise notnull oop of the klass.
4052 // (Actually, it need not be precise if this is a reflective allocation.)
4053 // It's what we cast the result to.
4054 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4055 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4056 const TypeOopPtr* oop_type = tklass->as_instance_type();
4057
4058 // Now generate allocation code
4059
4060 // The entire memory state is needed for slow path of the allocation
4061 // since GC and deoptimization can happen.
4062 Node *mem = reset_memory();
4063 set_all_memory(mem); // Create new memory state
4064
4065 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4066 control(), mem, i_o(),
4067 size, klass_node,
4068 initial_slow_test, inline_type_node);
4069
4070 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4071 }
4072
4073 //-------------------------------new_array-------------------------------------
4074 // helper for newarray and anewarray
4075 // The 'length' parameter is (obviously) the length of the array.
4076 // See comments on new_instance for the meaning of the other arguments.
4077 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4078 Node* length, // number of array elements
4079 int nargs, // number of arguments to push back for uncommon trap
4080 Node* *return_size_val,
4081 bool deoptimize_on_exception) {
4082 jint layout_con = Klass::_lh_neutral_value;
4083 Node* layout_val = get_layout_helper(klass_node, layout_con);
4084 bool layout_is_con = (layout_val == NULL);
4085
4086 if (!layout_is_con && !StressReflectiveCode &&
4087 !too_many_traps(Deoptimization::Reason_class_check)) {
4088 // This is a reflective array creation site.
4089 // Optimistically assume that it is a subtype of Object[],
4090 // so that we can fold up all the address arithmetic.
4091 layout_con = Klass::array_layout_helper(T_OBJECT);
4092 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4093 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4094 { BuildCutout unless(this, bol_lh, PROB_MAX);
4095 inc_sp(nargs);
4096 uncommon_trap(Deoptimization::Reason_class_check,
4097 Deoptimization::Action_maybe_recompile);
4098 }
4099 layout_val = NULL;
4100 layout_is_con = true;
4101 }
4102
4103 // Generate the initial go-slow test. Make sure we do not overflow
4104 // if length is huge (near 2Gig) or negative! We do not need
4105 // exact double-words here, just a close approximation of needed
4106 // double-words. We can't add any offset or rounding bits, lest we
4107 // take a size -1 of bytes and make it positive. Use an unsigned
4108 // compare, so negative sizes look hugely positive.
4109 int fast_size_limit = FastAllocateSizeLimit;
4110 if (layout_is_con) {
4111 assert(!StressReflectiveCode, "stress mode does not use these paths");
4112 // Increase the size limit if we have exact knowledge of array type.
4113 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4114 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4115 }
4116
4117 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4118 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4119
4120 // --- Size Computation ---
4121 // array_size = round_to_heap(array_header + (length << elem_shift));
4122 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4123 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4124 // The rounding mask is strength-reduced, if possible.
4125 int round_mask = MinObjAlignmentInBytes - 1;
4126 Node* header_size = NULL;
4127 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4128 // (T_BYTE has the weakest alignment and size restrictions...)
4129 if (layout_is_con) {
4130 int hsize = Klass::layout_helper_header_size(layout_con);
4131 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4132 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4133 if ((round_mask & ~right_n_bits(eshift)) == 0)
4134 round_mask = 0; // strength-reduce it if it goes away completely
4135 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4136 assert(header_size_min <= hsize, "generic minimum is smallest");
4137 header_size_min = hsize;
4138 header_size = intcon(hsize + round_mask);
4139 } else {
4140 Node* hss = intcon(Klass::_lh_header_size_shift);
4141 Node* hsm = intcon(Klass::_lh_header_size_mask);
4142 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
4143 hsize = _gvn.transform( new AndINode(hsize, hsm) );
4144 Node* mask = intcon(round_mask);
4145 header_size = _gvn.transform( new AddINode(hsize, mask) );
4146 }
4147
4148 Node* elem_shift = NULL;
4149 if (layout_is_con) {
4150 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4151 if (eshift != 0)
4152 elem_shift = intcon(eshift);
4153 } else {
4154 // There is no need to mask or shift this value.
4155 // The semantics of LShiftINode include an implicit mask to 0x1F.
4156 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4157 elem_shift = layout_val;
4158 }
4159
4160 // Transition to native address size for all offset calculations:
4161 Node* lengthx = ConvI2X(length);
4162 Node* headerx = ConvI2X(header_size);
4163 #ifdef _LP64
4164 { const TypeInt* tilen = _gvn.find_int_type(length);
4165 if (tilen != NULL && tilen->_lo < 0) {
4166 // Add a manual constraint to a positive range. Cf. array_element_address.
4167 jint size_max = fast_size_limit;
4168 if (size_max > tilen->_hi) size_max = tilen->_hi;
4169 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4170
4171 // Only do a narrow I2L conversion if the range check passed.
4172 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4173 _gvn.transform(iff);
4174 RegionNode* region = new RegionNode(3);
4175 _gvn.set_type(region, Type::CONTROL);
4176 lengthx = new PhiNode(region, TypeLong::LONG);
4177 _gvn.set_type(lengthx, TypeLong::LONG);
4178
4179 // Range check passed. Use ConvI2L node with narrow type.
4180 Node* passed = IfFalse(iff);
4181 region->init_req(1, passed);
4182 // Make I2L conversion control dependent to prevent it from
4183 // floating above the range check during loop optimizations.
4184 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4185
4186 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4187 region->init_req(2, IfTrue(iff));
4188 lengthx->init_req(2, ConvI2X(length));
4189
4190 set_control(region);
4191 record_for_igvn(region);
4192 record_for_igvn(lengthx);
4193 }
4194 }
4195 #endif
4196
4197 // Combine header size (plus rounding) and body size. Then round down.
4198 // This computation cannot overflow, because it is used only in two
4199 // places, one where the length is sharply limited, and the other
4200 // after a successful allocation.
4201 Node* abody = lengthx;
4202 if (elem_shift != NULL)
4203 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
4204 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
4205 if (round_mask != 0) {
4206 Node* mask = MakeConX(~round_mask);
4207 size = _gvn.transform( new AndXNode(size, mask) );
4208 }
4209 // else if round_mask == 0, the size computation is self-rounding
4210
4211 if (return_size_val != NULL) {
4212 // This is the size
4213 (*return_size_val) = size;
4214 }
4215
4216 // Now generate allocation code
4217
4218 // The entire memory state is needed for slow path of the allocation
4219 // since GC and deoptimization can happen.
4220 Node *mem = reset_memory();
4221 set_all_memory(mem); // Create new memory state
4222
4223 if (initial_slow_test->is_Bool()) {
4224 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4225 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4226 }
4227
4228 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4229 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4230 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
4231
4232 // Inline type array variants:
4233 // - null-ok: MyValue.ref[] (ciObjArrayKlass "[LMyValue")
4234 // - null-free: MyValue.val[] (ciObjArrayKlass "[QMyValue")
4235 // - null-free, flattened: MyValue.val[] (ciFlatArrayKlass "[QMyValue")
4236 // Check if array is a null-free, non-flattened inline type array
4237 // that needs to be initialized with the default inline type.
4238 Node* default_value = NULL;
4239 Node* raw_default_value = NULL;
4240 if (ary_ptr != NULL && ary_ptr->klass_is_exact()) {
4241 // Array type is known
4242 if (ary_ptr->is_null_free() && !ary_ptr->is_flat()) {
4243 ciInlineKlass* vk = ary_ptr->elem()->make_oopptr()->inline_klass();
4244 default_value = InlineTypeNode::default_oop(gvn(), vk);
4245 }
4246 } else if (ary_type->can_be_inline_array()) {
4247 // Array type is not known, add runtime checks
4248 assert(!ary_klass->klass_is_exact(), "unexpected exact type");
4249 Node* r = new RegionNode(3);
4250 default_value = new PhiNode(r, TypeInstPtr::BOTTOM);
4251
4252 Node* bol = array_lh_test(klass_node, Klass::_lh_array_tag_flat_value_bit_inplace | Klass::_lh_null_free_array_bit_inplace, Klass::_lh_null_free_array_bit_inplace);
4253 IfNode* iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4254
4255 // Null-free, non-flattened inline type array, initialize with the default value
4256 set_control(_gvn.transform(new IfTrueNode(iff)));
4257 Node* p = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset()));
4258 Node* eklass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), p, TypeInstPtr::KLASS));
4259 Node* adr_fixed_block_addr = basic_plus_adr(eklass, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()));
4260 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4261 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(InlineKlass::default_value_offset_offset()));
4262 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered);
4263 Node* elem_mirror = load_mirror_from_klass(eklass);
4264 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset));
4265 Node* val = access_load_at(elem_mirror, default_value_addr, TypeInstPtr::MIRROR, TypeInstPtr::NOTNULL, T_OBJECT, IN_HEAP);
4266 r->init_req(1, control());
4267 default_value->init_req(1, val);
4268
4269 // Otherwise initialize with all zero
4270 r->init_req(2, _gvn.transform(new IfFalseNode(iff)));
4271 default_value->init_req(2, null());
4272
4273 set_control(_gvn.transform(r));
4274 default_value = _gvn.transform(default_value);
4275 }
4276 if (default_value != NULL) {
4277 if (UseCompressedOops) {
4278 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4279 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4280 Node* lower = _gvn.transform(new CastP2XNode(control(), default_value));
4281 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4282 raw_default_value = _gvn.transform(new OrLNode(lower, upper));
4283 } else {
4284 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4285 }
4286 }
4287
4288 Node* valid_length_test = _gvn.intcon(1);
4289 if (ary_type->isa_aryptr()) {
4290 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4291 jint max = TypeAryPtr::max_array_length(bt);
4292 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4293 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4294 }
4295
4296 // Create the AllocateArrayNode and its result projections
4297 AllocateArrayNode* alloc
4298 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4299 control(), mem, i_o(),
4300 size, klass_node,
4301 initial_slow_test,
4302 length, valid_length_test,
4303 default_value, raw_default_value);
4304 // Cast to correct type. Note that the klass_node may be constant or not,
4305 // and in the latter case the actual array type will be inexact also.
4306 // (This happens via a non-constant argument to inline_native_newArray.)
4307 // In any case, the value of klass_node provides the desired array type.
4308 const TypeInt* length_type = _gvn.find_int_type(length);
4309 if (ary_type->isa_aryptr() && length_type != NULL) {
4310 // Try to get a better type than POS for the size
4311 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4312 }
4313
4314 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4315
4316 array_ideal_length(alloc, ary_type, true);
4317 return javaoop;
4318 }
4319
4320 // The following "Ideal_foo" functions are placed here because they recognize
4321 // the graph shapes created by the functions immediately above.
4322
4323 //---------------------------Ideal_allocation----------------------------------
4324 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4325 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
4326 if (ptr == NULL) { // reduce dumb test in callers
4327 return NULL;
4328 }
4329
4330 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4331 ptr = bs->step_over_gc_barrier(ptr);
4332
4333 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4334 ptr = ptr->in(1);
4335 if (ptr == NULL) return NULL;
4336 }
4337 // Return NULL for allocations with several casts:
4338 // j.l.reflect.Array.newInstance(jobject, jint)
4339 // Object.clone()
4340 // to keep more precise type from last cast.
4341 if (ptr->is_Proj()) {
4342 Node* allo = ptr->in(0);
4343 if (allo != NULL && allo->is_Allocate()) {
4344 return allo->as_Allocate();
4345 }
4346 }
4347 // Report failure to match.
4348 return NULL;
4349 }
4350
4351 // Fancy version which also strips off an offset (and reports it to caller).
4352 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
4353 intptr_t& offset) {
4354 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4355 if (base == NULL) return NULL;
4356 return Ideal_allocation(base, phase);
4357 }
4358
4359 // Trace Initialize <- Proj[Parm] <- Allocate
4360 AllocateNode* InitializeNode::allocation() {
4361 Node* rawoop = in(InitializeNode::RawAddress);
4362 if (rawoop->is_Proj()) {
4363 Node* alloc = rawoop->in(0);
4364 if (alloc->is_Allocate()) {
4365 return alloc->as_Allocate();
4366 }
4367 }
4368 return NULL;
4369 }
4370
4371 // Trace Allocate -> Proj[Parm] -> Initialize
4372 InitializeNode* AllocateNode::initialization() {
4373 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4374 if (rawoop == NULL) return NULL;
4375 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4376 Node* init = rawoop->fast_out(i);
4377 if (init->is_Initialize()) {
4378 assert(init->as_Initialize()->allocation() == this, "2-way link");
4379 return init->as_Initialize();
4380 }
4381 }
4382 return NULL;
4383 }
4384
4385 //----------------------------- loop predicates ---------------------------
4386
4387 //------------------------------add_predicate_impl----------------------------
4388 void GraphKit::add_empty_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
4389 // Too many traps seen?
4390 if (too_many_traps(reason)) {
4391 #ifdef ASSERT
4392 if (TraceLoopPredicate) {
4393 int tc = C->trap_count(reason);
4394 tty->print("too many traps=%s tcount=%d in ",
4395 Deoptimization::trap_reason_name(reason), tc);
4396 method()->print(); // which method has too many predicate traps
4397 tty->cr();
4398 }
4399 #endif
4400 // We cannot afford to take more traps here,
4401 // do not generate predicate.
4402 return;
4403 }
4404
4405 Node *cont = _gvn.intcon(1);
4406 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
4407 Node *bol = _gvn.transform(new Conv2BNode(opq));
4408 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4409 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
4410 C->add_predicate_opaq(opq);
4411 {
4412 PreserveJVMState pjvms(this);
4413 set_control(iffalse);
4414 inc_sp(nargs);
4415 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4416 }
4417 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
4418 set_control(iftrue);
4419 }
4420
4421 //------------------------------add_predicate---------------------------------
4422 void GraphKit::add_empty_predicates(int nargs) {
4423 // These loop predicates remain empty. All concrete loop predicates are inserted above the corresponding
4424 // empty loop predicate later by 'PhaseIdealLoop::create_new_if_for_predicate'. All concrete loop predicates of
4425 // a specific kind (normal, profile or limit check) share the same uncommon trap as the empty loop predicate.
4426 if (UseLoopPredicate) {
4427 add_empty_predicate_impl(Deoptimization::Reason_predicate, nargs);
4428 }
4429 if (UseProfiledLoopPredicate) {
4430 add_empty_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
4431 }
4432 // loop's limit check predicate should be near the loop.
4433 add_empty_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
4434 }
4435
4436 void GraphKit::sync_kit(IdealKit& ideal) {
4437 set_all_memory(ideal.merged_memory());
4438 set_i_o(ideal.i_o());
4439 set_control(ideal.ctrl());
4440 }
4441
4442 void GraphKit::final_sync(IdealKit& ideal) {
4443 // Final sync IdealKit and graphKit.
4444 sync_kit(ideal);
4445 }
4446
4447 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4448 Node* len = load_array_length(load_String_value(str, set_ctrl));
4449 Node* coder = load_String_coder(str, set_ctrl);
4450 // Divide length by 2 if coder is UTF16
4451 return _gvn.transform(new RShiftINode(len, coder));
4452 }
4453
4454 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4455 int value_offset = java_lang_String::value_offset();
4456 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4457 false, NULL, Type::Offset(0));
4458 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4459 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4460 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, true, true),
4461 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4462 Node* p = basic_plus_adr(str, str, value_offset);
4463 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4464 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4465 return load;
4466 }
4467
4468 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4469 if (!CompactStrings) {
4470 return intcon(java_lang_String::CODER_UTF16);
4471 }
4472 int coder_offset = java_lang_String::coder_offset();
4473 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4474 false, NULL, Type::Offset(0));
4475 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4476
4477 Node* p = basic_plus_adr(str, str, coder_offset);
4478 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4479 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4480 return load;
4481 }
4482
4483 void GraphKit::store_String_value(Node* str, Node* value) {
4484 int value_offset = java_lang_String::value_offset();
4485 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4486 false, NULL, Type::Offset(0));
4487 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4488
4489 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4490 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4491 }
4492
4493 void GraphKit::store_String_coder(Node* str, Node* value) {
4494 int coder_offset = java_lang_String::coder_offset();
4495 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4496 false, NULL, Type::Offset(0));
4497 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4498
4499 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4500 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4501 }
4502
4503 // Capture src and dst memory state with a MergeMemNode
4504 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4505 if (src_type == dst_type) {
4506 // Types are equal, we don't need a MergeMemNode
4507 return memory(src_type);
4508 }
4509 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4510 record_for_igvn(merge); // fold it up later, if possible
4511 int src_idx = C->get_alias_index(src_type);
4512 int dst_idx = C->get_alias_index(dst_type);
4513 merge->set_memory_at(src_idx, memory(src_idx));
4514 merge->set_memory_at(dst_idx, memory(dst_idx));
4515 return merge;
4516 }
4517
4518 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4519 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4520 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4521 // If input and output memory types differ, capture both states to preserve
4522 // the dependency between preceding and subsequent loads/stores.
4523 // For example, the following program:
4524 // StoreB
4525 // compress_string
4526 // LoadB
4527 // has this memory graph (use->def):
4528 // LoadB -> compress_string -> CharMem
4529 // ... -> StoreB -> ByteMem
4530 // The intrinsic hides the dependency between LoadB and StoreB, causing
4531 // the load to read from memory not containing the result of the StoreB.
4532 // The correct memory graph should look like this:
4533 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4534 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4535 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4536 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4537 set_memory(res_mem, TypeAryPtr::BYTES);
4538 return str;
4539 }
4540
4541 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4542 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4543 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4544 // Capture src and dst memory (see comment in 'compress_string').
4545 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4546 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4547 set_memory(_gvn.transform(str), dst_type);
4548 }
4549
4550 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4551 /**
4552 * int i_char = start;
4553 * for (int i_byte = 0; i_byte < count; i_byte++) {
4554 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4555 * }
4556 */
4557 add_empty_predicates();
4558 C->set_has_loops(true);
4559
4560 RegionNode* head = new RegionNode(3);
4561 head->init_req(1, control());
4562 gvn().set_type(head, Type::CONTROL);
4563 record_for_igvn(head);
4564
4565 Node* i_byte = new PhiNode(head, TypeInt::INT);
4566 i_byte->init_req(1, intcon(0));
4567 gvn().set_type(i_byte, TypeInt::INT);
4568 record_for_igvn(i_byte);
4569
4570 Node* i_char = new PhiNode(head, TypeInt::INT);
4571 i_char->init_req(1, start);
4572 gvn().set_type(i_char, TypeInt::INT);
4573 record_for_igvn(i_char);
4574
4575 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4576 gvn().set_type(mem, Type::MEMORY);
4577 record_for_igvn(mem);
4578 set_control(head);
4579 set_memory(mem, TypeAryPtr::BYTES);
4580 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4581 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4582 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4583 false, false, true /* mismatched */);
4584
4585 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4586 head->init_req(2, IfTrue(iff));
4587 mem->init_req(2, st);
4588 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4589 i_char->init_req(2, AddI(i_char, intcon(2)));
4590
4591 set_control(IfFalse(iff));
4592 set_memory(st, TypeAryPtr::BYTES);
4593 }
4594
4595 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4596 if (!field->is_constant()) {
4597 return NULL; // Field not marked as constant.
4598 }
4599 ciInstance* holder = NULL;
4600 if (!field->is_static()) {
4601 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4602 if (const_oop != NULL && const_oop->is_instance()) {
4603 holder = const_oop->as_instance();
4604 }
4605 }
4606 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4607 /*is_unsigned_load=*/false);
4608 if (con_type != NULL) {
4609 Node* con = makecon(con_type);
4610 if (field->type()->is_inlinetype()) {
4611 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass(), field->is_null_free());
4612 } else if (con_type->is_inlinetypeptr()) {
4613 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass(), field->is_null_free());
4614 }
4615 return con;
4616 }
4617 return NULL;
4618 }
4619
4620 //---------------------------load_mirror_from_klass----------------------------
4621 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4622 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4623 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4624 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4625 // mirror = ((OopHandle)mirror)->resolve();
4626 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4627 }