1 /*
2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "ci/ciValueKlass.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/c2/barrierSetC2.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/idealKit.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/narrowptrnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/valuetypenode.hpp"
47 #include "runtime/deoptimization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49
50 //----------------------------GraphKit-----------------------------------------
51 // Main utility constructor.
52 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
53 : Phase(Phase::Parser),
54 _env(C->env()),
55 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()),
56 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
57 {
58 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
59 _exceptions = jvms->map()->next_exception();
60 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
61 set_jvms(jvms);
62 #ifdef ASSERT
63 if (_gvn.is_IterGVN() != NULL) {
64 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
65 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
66 _worklist_size = _gvn.C->for_igvn()->size();
67 }
68 #endif
69 }
70
71 // Private constructor for parser.
72 GraphKit::GraphKit()
73 : Phase(Phase::Parser),
74 _env(C->env()),
75 _gvn(*C->initial_gvn()),
76 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
77 {
78 _exceptions = NULL;
79 set_map(NULL);
80 debug_only(_sp = -99);
81 debug_only(set_bci(-99));
82 }
83
84
85
86 //---------------------------clean_stack---------------------------------------
87 // Clear away rubbish from the stack area of the JVM state.
88 // This destroys any arguments that may be waiting on the stack.
89 void GraphKit::clean_stack(int from_sp) {
90 SafePointNode* map = this->map();
91 JVMState* jvms = this->jvms();
92 int stk_size = jvms->stk_size();
93 int stkoff = jvms->stkoff();
94 Node* top = this->top();
95 for (int i = from_sp; i < stk_size; i++) {
96 if (map->in(stkoff + i) != top) {
97 map->set_req(stkoff + i, top);
98 }
99 }
100 }
101
102
103 //--------------------------------sync_jvms-----------------------------------
104 // Make sure our current jvms agrees with our parse state.
105 JVMState* GraphKit::sync_jvms() const {
106 JVMState* jvms = this->jvms();
107 jvms->set_bci(bci()); // Record the new bci in the JVMState
108 jvms->set_sp(sp()); // Record the new sp in the JVMState
109 assert(jvms_in_sync(), "jvms is now in sync");
110 return jvms;
111 }
112
113 //--------------------------------sync_jvms_for_reexecute---------------------
114 // Make sure our current jvms agrees with our parse state. This version
115 // uses the reexecute_sp for reexecuting bytecodes.
116 JVMState* GraphKit::sync_jvms_for_reexecute() {
117 JVMState* jvms = this->jvms();
118 jvms->set_bci(bci()); // Record the new bci in the JVMState
119 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
120 return jvms;
121 }
122
123 #ifdef ASSERT
124 bool GraphKit::jvms_in_sync() const {
125 Parse* parse = is_Parse();
126 if (parse == NULL) {
127 if (bci() != jvms()->bci()) return false;
128 if (sp() != (int)jvms()->sp()) return false;
129 return true;
130 }
131 if (jvms()->method() != parse->method()) return false;
132 if (jvms()->bci() != parse->bci()) return false;
133 int jvms_sp = jvms()->sp();
134 if (jvms_sp != parse->sp()) return false;
135 int jvms_depth = jvms()->depth();
136 if (jvms_depth != parse->depth()) return false;
137 return true;
138 }
139
140 // Local helper checks for special internal merge points
141 // used to accumulate and merge exception states.
142 // They are marked by the region's in(0) edge being the map itself.
143 // Such merge points must never "escape" into the parser at large,
144 // until they have been handed to gvn.transform.
145 static bool is_hidden_merge(Node* reg) {
146 if (reg == NULL) return false;
147 if (reg->is_Phi()) {
148 reg = reg->in(0);
149 if (reg == NULL) return false;
150 }
151 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
152 }
153
154 void GraphKit::verify_map() const {
155 if (map() == NULL) return; // null map is OK
156 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
157 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
158 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
159 }
160
161 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
162 assert(ex_map->next_exception() == NULL, "not already part of a chain");
163 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
164 }
165 #endif
166
167 //---------------------------stop_and_kill_map---------------------------------
168 // Set _map to NULL, signalling a stop to further bytecode execution.
169 // First smash the current map's control to a constant, to mark it dead.
170 void GraphKit::stop_and_kill_map() {
171 SafePointNode* dead_map = stop();
172 if (dead_map != NULL) {
173 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
174 assert(dead_map->is_killed(), "must be so marked");
175 }
176 }
177
178
179 //--------------------------------stopped--------------------------------------
180 // Tell if _map is NULL, or control is top.
181 bool GraphKit::stopped() {
182 if (map() == NULL) return true;
183 else if (control() == top()) return true;
184 else return false;
185 }
186
187
188 //-----------------------------has_ex_handler----------------------------------
189 // Tell if this method or any caller method has exception handlers.
190 bool GraphKit::has_ex_handler() {
191 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
192 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
193 return true;
194 }
195 }
196 return false;
197 }
198
199 //------------------------------save_ex_oop------------------------------------
200 // Save an exception without blowing stack contents or other JVM state.
201 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
202 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
203 ex_map->add_req(ex_oop);
204 debug_only(verify_exception_state(ex_map));
205 }
206
207 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
208 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
209 Node* ex_oop = ex_map->in(ex_map->req()-1);
210 if (clear_it) ex_map->del_req(ex_map->req()-1);
211 return ex_oop;
212 }
213
214 //-----------------------------saved_ex_oop------------------------------------
215 // Recover a saved exception from its map.
216 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
217 return common_saved_ex_oop(ex_map, false);
218 }
219
220 //--------------------------clear_saved_ex_oop---------------------------------
221 // Erase a previously saved exception from its map.
222 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
223 return common_saved_ex_oop(ex_map, true);
224 }
225
226 #ifdef ASSERT
227 //---------------------------has_saved_ex_oop----------------------------------
228 // Erase a previously saved exception from its map.
229 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
230 return ex_map->req() == ex_map->jvms()->endoff()+1;
231 }
232 #endif
233
234 //-------------------------make_exception_state--------------------------------
235 // Turn the current JVM state into an exception state, appending the ex_oop.
236 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
237 sync_jvms();
238 SafePointNode* ex_map = stop(); // do not manipulate this map any more
239 set_saved_ex_oop(ex_map, ex_oop);
240 return ex_map;
241 }
242
243
244 //--------------------------add_exception_state--------------------------------
245 // Add an exception to my list of exceptions.
246 void GraphKit::add_exception_state(SafePointNode* ex_map) {
247 if (ex_map == NULL || ex_map->control() == top()) {
248 return;
249 }
250 #ifdef ASSERT
251 verify_exception_state(ex_map);
252 if (has_exceptions()) {
253 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
254 }
255 #endif
256
257 // If there is already an exception of exactly this type, merge with it.
258 // In particular, null-checks and other low-level exceptions common up here.
259 Node* ex_oop = saved_ex_oop(ex_map);
260 const Type* ex_type = _gvn.type(ex_oop);
261 if (ex_oop == top()) {
262 // No action needed.
263 return;
264 }
265 assert(ex_type->isa_instptr(), "exception must be an instance");
266 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
267 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
268 // We check sp also because call bytecodes can generate exceptions
269 // both before and after arguments are popped!
270 if (ex_type2 == ex_type
271 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
272 combine_exception_states(ex_map, e2);
273 return;
274 }
275 }
276
277 // No pre-existing exception of the same type. Chain it on the list.
278 push_exception_state(ex_map);
279 }
280
281 //-----------------------add_exception_states_from-----------------------------
282 void GraphKit::add_exception_states_from(JVMState* jvms) {
283 SafePointNode* ex_map = jvms->map()->next_exception();
284 if (ex_map != NULL) {
285 jvms->map()->set_next_exception(NULL);
286 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
287 next_map = ex_map->next_exception();
288 ex_map->set_next_exception(NULL);
289 add_exception_state(ex_map);
290 }
291 }
292 }
293
294 //-----------------------transfer_exceptions_into_jvms-------------------------
295 JVMState* GraphKit::transfer_exceptions_into_jvms() {
296 if (map() == NULL) {
297 // We need a JVMS to carry the exceptions, but the map has gone away.
298 // Create a scratch JVMS, cloned from any of the exception states...
299 if (has_exceptions()) {
300 _map = _exceptions;
301 _map = clone_map();
302 _map->set_next_exception(NULL);
303 clear_saved_ex_oop(_map);
304 debug_only(verify_map());
305 } else {
306 // ...or created from scratch
307 JVMState* jvms = new (C) JVMState(_method, NULL);
308 jvms->set_bci(_bci);
309 jvms->set_sp(_sp);
310 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
311 set_jvms(jvms);
312 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
313 set_all_memory(top());
314 while (map()->req() < jvms->endoff()) map()->add_req(top());
315 }
316 // (This is a kludge, in case you didn't notice.)
317 set_control(top());
318 }
319 JVMState* jvms = sync_jvms();
320 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
321 jvms->map()->set_next_exception(_exceptions);
322 _exceptions = NULL; // done with this set of exceptions
323 return jvms;
324 }
325
326 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
327 assert(is_hidden_merge(dstphi), "must be a special merge node");
328 assert(is_hidden_merge(srcphi), "must be a special merge node");
329 uint limit = srcphi->req();
330 for (uint i = PhiNode::Input; i < limit; i++) {
331 dstphi->add_req(srcphi->in(i));
332 }
333 }
334 static inline void add_one_req(Node* dstphi, Node* src) {
335 assert(is_hidden_merge(dstphi), "must be a special merge node");
336 assert(!is_hidden_merge(src), "must not be a special merge node");
337 dstphi->add_req(src);
338 }
339
340 //-----------------------combine_exception_states------------------------------
341 // This helper function combines exception states by building phis on a
342 // specially marked state-merging region. These regions and phis are
343 // untransformed, and can build up gradually. The region is marked by
344 // having a control input of its exception map, rather than NULL. Such
345 // regions do not appear except in this function, and in use_exception_state.
346 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
347 if (failing()) return; // dying anyway...
348 JVMState* ex_jvms = ex_map->_jvms;
349 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
350 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
351 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
352 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
353 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
354 assert(ex_map->req() == phi_map->req(), "matching maps");
355 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
356 Node* hidden_merge_mark = root();
357 Node* region = phi_map->control();
358 MergeMemNode* phi_mem = phi_map->merged_memory();
359 MergeMemNode* ex_mem = ex_map->merged_memory();
360 if (region->in(0) != hidden_merge_mark) {
361 // The control input is not (yet) a specially-marked region in phi_map.
362 // Make it so, and build some phis.
363 region = new RegionNode(2);
364 _gvn.set_type(region, Type::CONTROL);
365 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
366 region->init_req(1, phi_map->control());
367 phi_map->set_control(region);
368 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
369 record_for_igvn(io_phi);
370 _gvn.set_type(io_phi, Type::ABIO);
371 phi_map->set_i_o(io_phi);
372 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
373 Node* m = mms.memory();
374 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
375 record_for_igvn(m_phi);
376 _gvn.set_type(m_phi, Type::MEMORY);
377 mms.set_memory(m_phi);
378 }
379 }
380
381 // Either or both of phi_map and ex_map might already be converted into phis.
382 Node* ex_control = ex_map->control();
383 // if there is special marking on ex_map also, we add multiple edges from src
384 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
385 // how wide was the destination phi_map, originally?
386 uint orig_width = region->req();
387
388 if (add_multiple) {
389 add_n_reqs(region, ex_control);
390 add_n_reqs(phi_map->i_o(), ex_map->i_o());
391 } else {
392 // ex_map has no merges, so we just add single edges everywhere
393 add_one_req(region, ex_control);
394 add_one_req(phi_map->i_o(), ex_map->i_o());
395 }
396 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
397 if (mms.is_empty()) {
398 // get a copy of the base memory, and patch some inputs into it
399 const TypePtr* adr_type = mms.adr_type(C);
400 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
401 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
402 mms.set_memory(phi);
403 // Prepare to append interesting stuff onto the newly sliced phi:
404 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
405 }
406 // Append stuff from ex_map:
407 if (add_multiple) {
408 add_n_reqs(mms.memory(), mms.memory2());
409 } else {
410 add_one_req(mms.memory(), mms.memory2());
411 }
412 }
413 uint limit = ex_map->req();
414 for (uint i = TypeFunc::Parms; i < limit; i++) {
415 // Skip everything in the JVMS after tos. (The ex_oop follows.)
416 if (i == tos) i = ex_jvms->monoff();
417 Node* src = ex_map->in(i);
418 Node* dst = phi_map->in(i);
419 if (src != dst) {
420 PhiNode* phi;
421 if (dst->in(0) != region) {
422 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
423 record_for_igvn(phi);
424 _gvn.set_type(phi, phi->type());
425 phi_map->set_req(i, dst);
426 // Prepare to append interesting stuff onto the new phi:
427 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
428 } else {
429 assert(dst->is_Phi(), "nobody else uses a hidden region");
430 phi = dst->as_Phi();
431 }
432 if (add_multiple && src->in(0) == ex_control) {
433 // Both are phis.
434 add_n_reqs(dst, src);
435 } else {
436 while (dst->req() < region->req()) add_one_req(dst, src);
437 }
438 const Type* srctype = _gvn.type(src);
439 if (phi->type() != srctype) {
440 const Type* dsttype = phi->type()->meet_speculative(srctype);
441 if (phi->type() != dsttype) {
442 phi->set_type(dsttype);
443 _gvn.set_type(phi, dsttype);
444 }
445 }
446 }
447 }
448 phi_map->merge_replaced_nodes_with(ex_map);
449 }
450
451 //--------------------------use_exception_state--------------------------------
452 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
453 if (failing()) { stop(); return top(); }
454 Node* region = phi_map->control();
455 Node* hidden_merge_mark = root();
456 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
457 Node* ex_oop = clear_saved_ex_oop(phi_map);
458 if (region->in(0) == hidden_merge_mark) {
459 // Special marking for internal ex-states. Process the phis now.
460 region->set_req(0, region); // now it's an ordinary region
461 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
462 // Note: Setting the jvms also sets the bci and sp.
463 set_control(_gvn.transform(region));
464 uint tos = jvms()->stkoff() + sp();
465 for (uint i = 1; i < tos; i++) {
466 Node* x = phi_map->in(i);
467 if (x->in(0) == region) {
468 assert(x->is_Phi(), "expected a special phi");
469 phi_map->set_req(i, _gvn.transform(x));
470 }
471 }
472 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
473 Node* x = mms.memory();
474 if (x->in(0) == region) {
475 assert(x->is_Phi(), "nobody else uses a hidden region");
476 mms.set_memory(_gvn.transform(x));
477 }
478 }
479 if (ex_oop->in(0) == region) {
480 assert(ex_oop->is_Phi(), "expected a special phi");
481 ex_oop = _gvn.transform(ex_oop);
482 }
483 } else {
484 set_jvms(phi_map->jvms());
485 }
486
487 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
488 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
489 return ex_oop;
490 }
491
492 //---------------------------------java_bc-------------------------------------
493 Bytecodes::Code GraphKit::java_bc() const {
494 ciMethod* method = this->method();
495 int bci = this->bci();
496 if (method != NULL && bci != InvocationEntryBci)
497 return method->java_code_at_bci(bci);
498 else
499 return Bytecodes::_illegal;
500 }
501
502 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
503 bool must_throw) {
504 // if the exception capability is set, then we will generate code
505 // to check the JavaThread.should_post_on_exceptions flag to see
506 // if we actually need to report exception events (for this
507 // thread). If we don't need to report exception events, we will
508 // take the normal fast path provided by add_exception_events. If
509 // exception event reporting is enabled for this thread, we will
510 // take the uncommon_trap in the BuildCutout below.
511
512 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
513 Node* jthread = _gvn.transform(new ThreadLocalNode());
514 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
515 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
516
517 // Test the should_post_on_exceptions_flag vs. 0
518 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
519 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
520
521 // Branch to slow_path if should_post_on_exceptions_flag was true
522 { BuildCutout unless(this, tst, PROB_MAX);
523 // Do not try anything fancy if we're notifying the VM on every throw.
524 // Cf. case Bytecodes::_athrow in parse2.cpp.
525 uncommon_trap(reason, Deoptimization::Action_none,
526 (ciKlass*)NULL, (char*)NULL, must_throw);
527 }
528
529 }
530
531 //------------------------------builtin_throw----------------------------------
532 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
533 bool must_throw = true;
534
535 if (env()->jvmti_can_post_on_exceptions()) {
536 // check if we must post exception events, take uncommon trap if so
537 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
538 // here if should_post_on_exceptions is false
539 // continue on with the normal codegen
540 }
541
542 // If this particular condition has not yet happened at this
543 // bytecode, then use the uncommon trap mechanism, and allow for
544 // a future recompilation if several traps occur here.
545 // If the throw is hot, try to use a more complicated inline mechanism
546 // which keeps execution inside the compiled code.
547 bool treat_throw_as_hot = false;
548 ciMethodData* md = method()->method_data();
549
550 if (ProfileTraps) {
551 if (too_many_traps(reason)) {
552 treat_throw_as_hot = true;
553 }
554 // (If there is no MDO at all, assume it is early in
555 // execution, and that any deopts are part of the
556 // startup transient, and don't need to be remembered.)
557
558 // Also, if there is a local exception handler, treat all throws
559 // as hot if there has been at least one in this method.
560 if (C->trap_count(reason) != 0
561 && method()->method_data()->trap_count(reason) != 0
562 && has_ex_handler()) {
563 treat_throw_as_hot = true;
564 }
565 }
566
567 // If this throw happens frequently, an uncommon trap might cause
568 // a performance pothole. If there is a local exception handler,
569 // and if this particular bytecode appears to be deoptimizing often,
570 // let us handle the throw inline, with a preconstructed instance.
571 // Note: If the deopt count has blown up, the uncommon trap
572 // runtime is going to flush this nmethod, not matter what.
573 if (treat_throw_as_hot
574 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
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 if (java_bc() == Bytecodes::_aastore) {
594 ex_obj = env()->ArrayStoreException_instance();
595 } else {
596 ex_obj = env()->ClassCastException_instance();
597 }
598 break;
599 default:
600 break;
601 }
602 if (failing()) { stop(); return; } // exception allocation might fail
603 if (ex_obj != NULL) {
604 // Cheat with a preallocated exception object.
605 if (C->log() != NULL)
606 C->log()->elem("hot_throw preallocated='1' reason='%s'",
607 Deoptimization::trap_reason_name(reason));
608 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
609 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
610
611 // Clear the detail message of the preallocated exception object.
612 // Weblogic sometimes mutates the detail message of exceptions
613 // using reflection.
614 int offset = java_lang_Throwable::get_detailMessage_offset();
615 const TypePtr* adr_typ = ex_con->add_offset(offset);
616
617 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
618 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
619 Node *store = access_store_at(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
620
621 add_exception_state(make_exception_state(ex_node));
622 return;
623 }
624 }
625
626 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
627 // It won't be much cheaper than bailing to the interp., since we'll
628 // have to pass up all the debug-info, and the runtime will have to
629 // create the stack trace.
630
631 // Usual case: Bail to interpreter.
632 // Reserve the right to recompile if we haven't seen anything yet.
633
634 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
635 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
636 if (treat_throw_as_hot
637 && (method()->method_data()->trap_recompiled_at(bci(), m)
638 || C->too_many_traps(reason))) {
639 // We cannot afford to take more traps here. Suffer in the interpreter.
640 if (C->log() != NULL)
641 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
642 Deoptimization::trap_reason_name(reason),
643 C->trap_count(reason));
644 action = Deoptimization::Action_none;
645 }
646
647 // "must_throw" prunes the JVM state to include only the stack, if there
648 // are no local exception handlers. This should cut down on register
649 // allocation time and code size, by drastically reducing the number
650 // of in-edges on the call to the uncommon trap.
651
652 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
653 }
654
655
656 //----------------------------PreserveJVMState---------------------------------
657 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
658 debug_only(kit->verify_map());
659 _kit = kit;
660 _map = kit->map(); // preserve the map
661 _sp = kit->sp();
662 kit->set_map(clone_map ? kit->clone_map() : NULL);
663 #ifdef ASSERT
664 _bci = kit->bci();
665 Parse* parser = kit->is_Parse();
666 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
667 _block = block;
668 #endif
669 }
670 PreserveJVMState::~PreserveJVMState() {
671 GraphKit* kit = _kit;
672 #ifdef ASSERT
673 assert(kit->bci() == _bci, "bci must not shift");
674 Parse* parser = kit->is_Parse();
675 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
676 assert(block == _block, "block must not shift");
677 #endif
678 kit->set_map(_map);
679 kit->set_sp(_sp);
680 }
681
682
683 //-----------------------------BuildCutout-------------------------------------
684 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
685 : PreserveJVMState(kit)
686 {
687 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
688 SafePointNode* outer_map = _map; // preserved map is caller's
689 SafePointNode* inner_map = kit->map();
690 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
691 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
692 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
693 }
694 BuildCutout::~BuildCutout() {
695 GraphKit* kit = _kit;
696 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
697 }
698
699 //---------------------------PreserveReexecuteState----------------------------
700 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
701 assert(!kit->stopped(), "must call stopped() before");
702 _kit = kit;
703 _sp = kit->sp();
704 _reexecute = kit->jvms()->_reexecute;
705 }
706 PreserveReexecuteState::~PreserveReexecuteState() {
707 if (_kit->stopped()) return;
708 _kit->jvms()->_reexecute = _reexecute;
709 _kit->set_sp(_sp);
710 }
711
712 //------------------------------clone_map--------------------------------------
713 // Implementation of PreserveJVMState
714 //
715 // Only clone_map(...) here. If this function is only used in the
716 // PreserveJVMState class we may want to get rid of this extra
717 // function eventually and do it all there.
718
719 SafePointNode* GraphKit::clone_map() {
720 if (map() == NULL) return NULL;
721
722 // Clone the memory edge first
723 Node* mem = MergeMemNode::make(map()->memory());
724 gvn().set_type_bottom(mem);
725
726 SafePointNode *clonemap = (SafePointNode*)map()->clone();
727 JVMState* jvms = this->jvms();
728 JVMState* clonejvms = jvms->clone_shallow(C);
729 clonemap->set_memory(mem);
730 clonemap->set_jvms(clonejvms);
731 clonejvms->set_map(clonemap);
732 record_for_igvn(clonemap);
733 gvn().set_type_bottom(clonemap);
734 return clonemap;
735 }
736
737
738 //-----------------------------set_map_clone-----------------------------------
739 void GraphKit::set_map_clone(SafePointNode* m) {
740 _map = m;
741 _map = clone_map();
742 _map->set_next_exception(NULL);
743 debug_only(verify_map());
744 }
745
746
747 //----------------------------kill_dead_locals---------------------------------
748 // Detect any locals which are known to be dead, and force them to top.
749 void GraphKit::kill_dead_locals() {
750 // Consult the liveness information for the locals. If any
751 // of them are unused, then they can be replaced by top(). This
752 // should help register allocation time and cut down on the size
753 // of the deoptimization information.
754
755 // This call is made from many of the bytecode handling
756 // subroutines called from the Big Switch in do_one_bytecode.
757 // Every bytecode which might include a slow path is responsible
758 // for killing its dead locals. The more consistent we
759 // are about killing deads, the fewer useless phis will be
760 // constructed for them at various merge points.
761
762 // bci can be -1 (InvocationEntryBci). We return the entry
763 // liveness for the method.
764
765 if (method() == NULL || method()->code_size() == 0) {
766 // We are building a graph for a call to a native method.
767 // All locals are live.
768 return;
769 }
770
771 ResourceMark rm;
772
773 // Consult the liveness information for the locals. If any
774 // of them are unused, then they can be replaced by top(). This
775 // should help register allocation time and cut down on the size
776 // of the deoptimization information.
777 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
778
779 int len = (int)live_locals.size();
780 assert(len <= jvms()->loc_size(), "too many live locals");
781 for (int local = 0; local < len; local++) {
782 if (!live_locals.at(local)) {
783 set_local(local, top());
784 }
785 }
786 }
787
788 #ifdef ASSERT
789 //-------------------------dead_locals_are_killed------------------------------
790 // Return true if all dead locals are set to top in the map.
791 // Used to assert "clean" debug info at various points.
792 bool GraphKit::dead_locals_are_killed() {
793 if (method() == NULL || method()->code_size() == 0) {
794 // No locals need to be dead, so all is as it should be.
795 return true;
796 }
797
798 // Make sure somebody called kill_dead_locals upstream.
799 ResourceMark rm;
800 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
801 if (jvms->loc_size() == 0) continue; // no locals to consult
802 SafePointNode* map = jvms->map();
803 ciMethod* method = jvms->method();
804 int bci = jvms->bci();
805 if (jvms == this->jvms()) {
806 bci = this->bci(); // it might not yet be synched
807 }
808 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
809 int len = (int)live_locals.size();
810 if (!live_locals.is_valid() || len == 0)
811 // This method is trivial, or is poisoned by a breakpoint.
812 return true;
813 assert(len == jvms->loc_size(), "live map consistent with locals map");
814 for (int local = 0; local < len; local++) {
815 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
816 if (PrintMiscellaneous && (Verbose || WizardMode)) {
817 tty->print_cr("Zombie local %d: ", local);
818 jvms->dump();
819 }
820 return false;
821 }
822 }
823 }
824 return true;
825 }
826
827 #endif //ASSERT
828
829 // Helper function for enforcing certain bytecodes to reexecute if
830 // deoptimization happens
831 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
832 ciMethod* cur_method = jvms->method();
833 int cur_bci = jvms->bci();
834 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
835 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
836 return Interpreter::bytecode_should_reexecute(code) ||
837 (is_anewarray && (code == Bytecodes::_multianewarray));
838 // Reexecute _multianewarray bytecode which was replaced with
839 // sequence of [a]newarray. See Parse::do_multianewarray().
840 //
841 // Note: interpreter should not have it set since this optimization
842 // is limited by dimensions and guarded by flag so in some cases
843 // multianewarray() runtime calls will be generated and
844 // the bytecode should not be reexecutes (stack will not be reset).
845 } else {
846 return false;
847 }
848 }
849
850 // Helper function for adding JVMState and debug information to node
851 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
852 // Add the safepoint edges to the call (or other safepoint).
853
854 // Make sure dead locals are set to top. This
855 // should help register allocation time and cut down on the size
856 // of the deoptimization information.
857 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
858
859 // Walk the inline list to fill in the correct set of JVMState's
860 // Also fill in the associated edges for each JVMState.
861
862 // If the bytecode needs to be reexecuted we need to put
863 // the arguments back on the stack.
864 const bool should_reexecute = jvms()->should_reexecute();
865 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
866
867 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
868 // undefined if the bci is different. This is normal for Parse but it
869 // should not happen for LibraryCallKit because only one bci is processed.
870 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
871 "in LibraryCallKit the reexecute bit should not change");
872
873 // If we are guaranteed to throw, we can prune everything but the
874 // input to the current bytecode.
875 bool can_prune_locals = false;
876 uint stack_slots_not_pruned = 0;
877 int inputs = 0, depth = 0;
878 if (must_throw) {
879 assert(method() == youngest_jvms->method(), "sanity");
880 if (compute_stack_effects(inputs, depth)) {
881 can_prune_locals = true;
882 stack_slots_not_pruned = inputs;
883 }
884 }
885
886 if (env()->should_retain_local_variables()) {
887 // At any safepoint, this method can get breakpointed, which would
888 // then require an immediate deoptimization.
889 can_prune_locals = false; // do not prune locals
890 stack_slots_not_pruned = 0;
891 }
892
893 // do not scribble on the input jvms
894 JVMState* out_jvms = youngest_jvms->clone_deep(C);
895 call->set_jvms(out_jvms); // Start jvms list for call node
896
897 // For a known set of bytecodes, the interpreter should reexecute them if
898 // deoptimization happens. We set the reexecute state for them here
899 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
900 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
901 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
902 }
903
904 // Presize the call:
905 DEBUG_ONLY(uint non_debug_edges = call->req());
906 call->add_req_batch(top(), youngest_jvms->debug_depth());
907 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
908
909 // Set up edges so that the call looks like this:
910 // Call [state:] ctl io mem fptr retadr
911 // [parms:] parm0 ... parmN
912 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
913 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
914 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
915 // Note that caller debug info precedes callee debug info.
916
917 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
918 uint debug_ptr = call->req();
919
920 // Loop over the map input edges associated with jvms, add them
921 // to the call node, & reset all offsets to match call node array.
922 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
923 uint debug_end = debug_ptr;
924 uint debug_start = debug_ptr - in_jvms->debug_size();
925 debug_ptr = debug_start; // back up the ptr
926
927 uint p = debug_start; // walks forward in [debug_start, debug_end)
928 uint j, k, l;
929 SafePointNode* in_map = in_jvms->map();
930 out_jvms->set_map(call);
931
932 if (can_prune_locals) {
933 assert(in_jvms->method() == out_jvms->method(), "sanity");
934 // If the current throw can reach an exception handler in this JVMS,
935 // then we must keep everything live that can reach that handler.
936 // As a quick and dirty approximation, we look for any handlers at all.
937 if (in_jvms->method()->has_exception_handlers()) {
938 can_prune_locals = false;
939 }
940 }
941
942 // Add the Locals
943 k = in_jvms->locoff();
944 l = in_jvms->loc_size();
945 out_jvms->set_locoff(p);
946 if (!can_prune_locals) {
947 for (j = 0; j < l; j++)
948 call->set_req(p++, in_map->in(k+j));
949 } else {
950 p += l; // already set to top above by add_req_batch
951 }
952
953 // Add the Expression Stack
954 k = in_jvms->stkoff();
955 l = in_jvms->sp();
956 out_jvms->set_stkoff(p);
957 if (!can_prune_locals) {
958 for (j = 0; j < l; j++)
959 call->set_req(p++, in_map->in(k+j));
960 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
961 // Divide stack into {S0,...,S1}, where S0 is set to top.
962 uint s1 = stack_slots_not_pruned;
963 stack_slots_not_pruned = 0; // for next iteration
964 if (s1 > l) s1 = l;
965 uint s0 = l - s1;
966 p += s0; // skip the tops preinstalled by add_req_batch
967 for (j = s0; j < l; j++)
968 call->set_req(p++, in_map->in(k+j));
969 } else {
970 p += l; // already set to top above by add_req_batch
971 }
972
973 // Add the Monitors
974 k = in_jvms->monoff();
975 l = in_jvms->mon_size();
976 out_jvms->set_monoff(p);
977 for (j = 0; j < l; j++)
978 call->set_req(p++, in_map->in(k+j));
979
980 // Copy any scalar object fields.
981 k = in_jvms->scloff();
982 l = in_jvms->scl_size();
983 out_jvms->set_scloff(p);
984 for (j = 0; j < l; j++)
985 call->set_req(p++, in_map->in(k+j));
986
987 // Finish the new jvms.
988 out_jvms->set_endoff(p);
989
990 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
991 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
992 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
993 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
994 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
995 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
996
997 // Update the two tail pointers in parallel.
998 out_jvms = out_jvms->caller();
999 in_jvms = in_jvms->caller();
1000 }
1001
1002 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1003
1004 // Test the correctness of JVMState::debug_xxx accessors:
1005 assert(call->jvms()->debug_start() == non_debug_edges, "");
1006 assert(call->jvms()->debug_end() == call->req(), "");
1007 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1008 }
1009
1010 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1011 Bytecodes::Code code = java_bc();
1012 if (code == Bytecodes::_wide) {
1013 code = method()->java_code_at_bci(bci() + 1);
1014 }
1015
1016 BasicType rtype = T_ILLEGAL;
1017 int rsize = 0;
1018
1019 if (code != Bytecodes::_illegal) {
1020 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1021 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1022 if (rtype < T_CONFLICT)
1023 rsize = type2size[rtype];
1024 }
1025
1026 switch (code) {
1027 case Bytecodes::_illegal:
1028 return false;
1029
1030 case Bytecodes::_ldc:
1031 case Bytecodes::_ldc_w:
1032 case Bytecodes::_ldc2_w:
1033 inputs = 0;
1034 break;
1035
1036 case Bytecodes::_dup: inputs = 1; break;
1037 case Bytecodes::_dup_x1: inputs = 2; break;
1038 case Bytecodes::_dup_x2: inputs = 3; break;
1039 case Bytecodes::_dup2: inputs = 2; break;
1040 case Bytecodes::_dup2_x1: inputs = 3; break;
1041 case Bytecodes::_dup2_x2: inputs = 4; break;
1042 case Bytecodes::_swap: inputs = 2; break;
1043 case Bytecodes::_arraylength: inputs = 1; break;
1044
1045 case Bytecodes::_getstatic:
1046 case Bytecodes::_putstatic:
1047 case Bytecodes::_getfield:
1048 case Bytecodes::_putfield:
1049 {
1050 bool ignored_will_link;
1051 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1052 int size = field->type()->size();
1053 bool is_get = (depth >= 0), is_static = (depth & 1);
1054 inputs = (is_static ? 0 : 1);
1055 if (is_get) {
1056 depth = size - inputs;
1057 } else {
1058 inputs += size; // putxxx pops the value from the stack
1059 depth = - inputs;
1060 }
1061 }
1062 break;
1063
1064 case Bytecodes::_invokevirtual:
1065 case Bytecodes::_invokespecial:
1066 case Bytecodes::_invokestatic:
1067 case Bytecodes::_invokedynamic:
1068 case Bytecodes::_invokeinterface:
1069 {
1070 bool ignored_will_link;
1071 ciSignature* declared_signature = NULL;
1072 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1073 assert(declared_signature != NULL, "cannot be null");
1074 inputs = declared_signature->arg_size_for_bc(code);
1075 int size = declared_signature->return_type()->size();
1076 depth = size - inputs;
1077 }
1078 break;
1079
1080 case Bytecodes::_multianewarray:
1081 {
1082 ciBytecodeStream iter(method());
1083 iter.reset_to_bci(bci());
1084 iter.next();
1085 inputs = iter.get_dimensions();
1086 assert(rsize == 1, "");
1087 depth = rsize - inputs;
1088 }
1089 break;
1090
1091 case Bytecodes::_withfield: {
1092 bool ignored_will_link;
1093 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1094 int size = field->type()->size();
1095 inputs = size+1;
1096 depth = rsize - inputs;
1097 break;
1098 }
1099
1100 case Bytecodes::_ireturn:
1101 case Bytecodes::_lreturn:
1102 case Bytecodes::_freturn:
1103 case Bytecodes::_dreturn:
1104 case Bytecodes::_areturn:
1105 assert(rsize == -depth, "");
1106 inputs = rsize;
1107 break;
1108
1109 case Bytecodes::_jsr:
1110 case Bytecodes::_jsr_w:
1111 inputs = 0;
1112 depth = 1; // S.B. depth=1, not zero
1113 break;
1114
1115 default:
1116 // bytecode produces a typed result
1117 inputs = rsize - depth;
1118 assert(inputs >= 0, "");
1119 break;
1120 }
1121
1122 #ifdef ASSERT
1123 // spot check
1124 int outputs = depth + inputs;
1125 assert(outputs >= 0, "sanity");
1126 switch (code) {
1127 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1128 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1129 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1130 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1131 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1132 default: break;
1133 }
1134 #endif //ASSERT
1135
1136 return true;
1137 }
1138
1139
1140
1141 //------------------------------basic_plus_adr---------------------------------
1142 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1143 // short-circuit a common case
1144 if (offset == intcon(0)) return ptr;
1145 return _gvn.transform( new AddPNode(base, ptr, offset) );
1146 }
1147
1148 Node* GraphKit::ConvI2L(Node* offset) {
1149 // short-circuit a common case
1150 jint offset_con = find_int_con(offset, Type::OffsetBot);
1151 if (offset_con != Type::OffsetBot) {
1152 return longcon((jlong) offset_con);
1153 }
1154 return _gvn.transform( new ConvI2LNode(offset));
1155 }
1156
1157 Node* GraphKit::ConvI2UL(Node* offset) {
1158 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1159 if (offset_con != (juint) Type::OffsetBot) {
1160 return longcon((julong) offset_con);
1161 }
1162 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1163 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1164 return _gvn.transform( new AndLNode(conv, mask) );
1165 }
1166
1167 Node* GraphKit::ConvL2I(Node* offset) {
1168 // short-circuit a common case
1169 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1170 if (offset_con != (jlong)Type::OffsetBot) {
1171 return intcon((int) offset_con);
1172 }
1173 return _gvn.transform( new ConvL2INode(offset));
1174 }
1175
1176 //-------------------------load_object_klass-----------------------------------
1177 Node* GraphKit::load_object_klass(Node* obj) {
1178 // Special-case a fresh allocation to avoid building nodes:
1179 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1180 if (akls != NULL) return akls;
1181 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1182 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1183 }
1184
1185 //-------------------------load_array_length-----------------------------------
1186 Node* GraphKit::load_array_length(Node* array) {
1187 // Special-case a fresh allocation to avoid building nodes:
1188 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1189 Node *alen;
1190 if (alloc == NULL) {
1191 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1192 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1193 } else {
1194 alen = alloc->Ideal_length();
1195 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1196 if (ccast != alen) {
1197 alen = _gvn.transform(ccast);
1198 }
1199 }
1200 return alen;
1201 }
1202
1203 //------------------------------do_null_check----------------------------------
1204 // Helper function to do a NULL pointer check. Returned value is
1205 // the incoming address with NULL casted away. You are allowed to use the
1206 // not-null value only if you are control dependent on the test.
1207 #ifndef PRODUCT
1208 extern int explicit_null_checks_inserted,
1209 explicit_null_checks_elided;
1210 #endif
1211 Node* GraphKit::null_check_common(Node* value, BasicType type,
1212 // optional arguments for variations:
1213 bool assert_null,
1214 Node* *null_control,
1215 bool speculative) {
1216 assert(!assert_null || null_control == NULL, "not both at once");
1217 if (stopped()) return top();
1218 NOT_PRODUCT(explicit_null_checks_inserted++);
1219
1220 // Construct NULL check
1221 Node *chk = NULL;
1222 switch(type) {
1223 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1224 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1225 case T_VALUETYPE : // fall through
1226 case T_ARRAY : // fall through
1227 type = T_OBJECT; // simplify further tests
1228 case T_OBJECT : {
1229 const Type *t = _gvn.type( value );
1230
1231 const TypeOopPtr* tp = t->isa_oopptr();
1232 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1233 // Only for do_null_check, not any of its siblings:
1234 && !assert_null && null_control == NULL) {
1235 // Usually, any field access or invocation on an unloaded oop type
1236 // will simply fail to link, since the statically linked class is
1237 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1238 // the static class is loaded but the sharper oop type is not.
1239 // Rather than checking for this obscure case in lots of places,
1240 // we simply observe that a null check on an unloaded class
1241 // will always be followed by a nonsense operation, so we
1242 // can just issue the uncommon trap here.
1243 // Our access to the unloaded class will only be correct
1244 // after it has been loaded and initialized, which requires
1245 // a trip through the interpreter.
1246 #ifndef PRODUCT
1247 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1248 #endif
1249 uncommon_trap(Deoptimization::Reason_unloaded,
1250 Deoptimization::Action_reinterpret,
1251 tp->klass(), "!loaded");
1252 return top();
1253 }
1254
1255 if (assert_null) {
1256 // See if the type is contained in NULL_PTR.
1257 // If so, then the value is already null.
1258 if (t->higher_equal(TypePtr::NULL_PTR)) {
1259 NOT_PRODUCT(explicit_null_checks_elided++);
1260 return value; // Elided null assert quickly!
1261 }
1262 } else {
1263 // See if mixing in the NULL pointer changes type.
1264 // If so, then the NULL pointer was not allowed in the original
1265 // type. In other words, "value" was not-null.
1266 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1267 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1268 NOT_PRODUCT(explicit_null_checks_elided++);
1269 return value; // Elided null check quickly!
1270 }
1271 }
1272 chk = new CmpPNode( value, null() );
1273 break;
1274 }
1275
1276 default:
1277 fatal("unexpected type: %s", type2name(type));
1278 }
1279 assert(chk != NULL, "sanity check");
1280 chk = _gvn.transform(chk);
1281
1282 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1283 BoolNode *btst = new BoolNode( chk, btest);
1284 Node *tst = _gvn.transform( btst );
1285
1286 //-----------
1287 // if peephole optimizations occurred, a prior test existed.
1288 // If a prior test existed, maybe it dominates as we can avoid this test.
1289 if (tst != btst && type == T_OBJECT) {
1290 // At this point we want to scan up the CFG to see if we can
1291 // find an identical test (and so avoid this test altogether).
1292 Node *cfg = control();
1293 int depth = 0;
1294 while( depth < 16 ) { // Limit search depth for speed
1295 if( cfg->Opcode() == Op_IfTrue &&
1296 cfg->in(0)->in(1) == tst ) {
1297 // Found prior test. Use "cast_not_null" to construct an identical
1298 // CastPP (and hence hash to) as already exists for the prior test.
1299 // Return that casted value.
1300 if (assert_null) {
1301 replace_in_map(value, null());
1302 return null(); // do not issue the redundant test
1303 }
1304 Node *oldcontrol = control();
1305 set_control(cfg);
1306 Node *res = cast_not_null(value);
1307 set_control(oldcontrol);
1308 NOT_PRODUCT(explicit_null_checks_elided++);
1309 return res;
1310 }
1311 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1312 if (cfg == NULL) break; // Quit at region nodes
1313 depth++;
1314 }
1315 }
1316
1317 //-----------
1318 // Branch to failure if null
1319 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1320 Deoptimization::DeoptReason reason;
1321 if (assert_null) {
1322 reason = Deoptimization::reason_null_assert(speculative);
1323 } else if (type == T_OBJECT) {
1324 reason = Deoptimization::reason_null_check(speculative);
1325 } else {
1326 reason = Deoptimization::Reason_div0_check;
1327 }
1328 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1329 // ciMethodData::has_trap_at will return a conservative -1 if any
1330 // must-be-null assertion has failed. This could cause performance
1331 // problems for a method after its first do_null_assert failure.
1332 // Consider using 'Reason_class_check' instead?
1333
1334 // To cause an implicit null check, we set the not-null probability
1335 // to the maximum (PROB_MAX). For an explicit check the probability
1336 // is set to a smaller value.
1337 if (null_control != NULL || too_many_traps(reason)) {
1338 // probability is less likely
1339 ok_prob = PROB_LIKELY_MAG(3);
1340 } else if (!assert_null &&
1341 (ImplicitNullCheckThreshold > 0) &&
1342 method() != NULL &&
1343 (method()->method_data()->trap_count(reason)
1344 >= (uint)ImplicitNullCheckThreshold)) {
1345 ok_prob = PROB_LIKELY_MAG(3);
1346 }
1347
1348 if (null_control != NULL) {
1349 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1350 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1351 set_control( _gvn.transform( new IfTrueNode(iff)));
1352 #ifndef PRODUCT
1353 if (null_true == top()) {
1354 explicit_null_checks_elided++;
1355 }
1356 #endif
1357 (*null_control) = null_true;
1358 } else {
1359 BuildCutout unless(this, tst, ok_prob);
1360 // Check for optimizer eliding test at parse time
1361 if (stopped()) {
1362 // Failure not possible; do not bother making uncommon trap.
1363 NOT_PRODUCT(explicit_null_checks_elided++);
1364 } else if (assert_null) {
1365 uncommon_trap(reason,
1366 Deoptimization::Action_make_not_entrant,
1367 NULL, "assert_null");
1368 } else {
1369 replace_in_map(value, zerocon(type));
1370 builtin_throw(reason);
1371 }
1372 }
1373
1374 // Must throw exception, fall-thru not possible?
1375 if (stopped()) {
1376 return top(); // No result
1377 }
1378
1379 if (assert_null) {
1380 // Cast obj to null on this path.
1381 replace_in_map(value, zerocon(type));
1382 return zerocon(type);
1383 }
1384
1385 // Cast obj to not-null on this path, if there is no null_control.
1386 // (If there is a null_control, a non-null value may come back to haunt us.)
1387 if (type == T_OBJECT) {
1388 Node* cast = cast_not_null(value, false);
1389 if (null_control == NULL || (*null_control) == top())
1390 replace_in_map(value, cast);
1391 value = cast;
1392 }
1393
1394 return value;
1395 }
1396
1397
1398 //------------------------------cast_not_null----------------------------------
1399 // Cast obj to not-null on this path
1400 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1401 assert(!obj->is_ValueType(), "should not cast value type");
1402 const Type *t = _gvn.type(obj);
1403 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1404 // Object is already not-null?
1405 if( t == t_not_null ) return obj;
1406
1407 Node *cast = new CastPPNode(obj,t_not_null);
1408 cast->init_req(0, control());
1409 cast = _gvn.transform( cast );
1410
1411 // Scan for instances of 'obj' in the current JVM mapping.
1412 // These instances are known to be not-null after the test.
1413 if (do_replace_in_map)
1414 replace_in_map(obj, cast);
1415
1416 return cast; // Return casted value
1417 }
1418
1419 // Sometimes in intrinsics, we implicitly know an object is not null
1420 // (there's no actual null check) so we can cast it to not null. In
1421 // the course of optimizations, the input to the cast can become null.
1422 // In that case that data path will die and we need the control path
1423 // to become dead as well to keep the graph consistent. So we have to
1424 // add a check for null for which one branch can't be taken. It uses
1425 // an Opaque4 node that will cause the check to be removed after loop
1426 // opts so the test goes away and the compiled code doesn't execute a
1427 // useless check.
1428 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1429 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1430 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1431 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1432 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1433 _gvn.set_type(iff, iff->Value(&_gvn));
1434 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1435 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1436 Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1437 C->root()->add_req(halt);
1438 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1439 set_control(if_t);
1440 return cast_not_null(value, do_replace_in_map);
1441 }
1442
1443
1444 //--------------------------replace_in_map-------------------------------------
1445 void GraphKit::replace_in_map(Node* old, Node* neww) {
1446 if (old == neww) {
1447 return;
1448 }
1449
1450 map()->replace_edge(old, neww);
1451
1452 // Note: This operation potentially replaces any edge
1453 // on the map. This includes locals, stack, and monitors
1454 // of the current (innermost) JVM state.
1455
1456 // don't let inconsistent types from profiling escape this
1457 // method
1458
1459 const Type* told = _gvn.type(old);
1460 const Type* tnew = _gvn.type(neww);
1461
1462 if (!tnew->higher_equal(told)) {
1463 return;
1464 }
1465
1466 map()->record_replaced_node(old, neww);
1467 }
1468
1469
1470 //=============================================================================
1471 //--------------------------------memory---------------------------------------
1472 Node* GraphKit::memory(uint alias_idx) {
1473 MergeMemNode* mem = merged_memory();
1474 Node* p = mem->memory_at(alias_idx);
1475 _gvn.set_type(p, Type::MEMORY); // must be mapped
1476 return p;
1477 }
1478
1479 //-----------------------------reset_memory------------------------------------
1480 Node* GraphKit::reset_memory() {
1481 Node* mem = map()->memory();
1482 // do not use this node for any more parsing!
1483 debug_only( map()->set_memory((Node*)NULL) );
1484 return _gvn.transform( mem );
1485 }
1486
1487 //------------------------------set_all_memory---------------------------------
1488 void GraphKit::set_all_memory(Node* newmem) {
1489 Node* mergemem = MergeMemNode::make(newmem);
1490 gvn().set_type_bottom(mergemem);
1491 map()->set_memory(mergemem);
1492 }
1493
1494 //------------------------------set_all_memory_call----------------------------
1495 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1496 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1497 set_all_memory(newmem);
1498 }
1499
1500 //=============================================================================
1501 //
1502 // parser factory methods for MemNodes
1503 //
1504 // These are layered on top of the factory methods in LoadNode and StoreNode,
1505 // and integrate with the parser's memory state and _gvn engine.
1506 //
1507
1508 // factory methods in "int adr_idx"
1509 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1510 int adr_idx,
1511 MemNode::MemOrd mo,
1512 LoadNode::ControlDependency control_dependency,
1513 bool require_atomic_access,
1514 bool unaligned,
1515 bool mismatched) {
1516 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1517 const TypePtr* adr_type = NULL; // debug-mode-only argument
1518 debug_only(adr_type = C->get_adr_type(adr_idx));
1519 Node* mem = memory(adr_idx);
1520 Node* ld;
1521 if (require_atomic_access && bt == T_LONG) {
1522 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1523 } else if (require_atomic_access && bt == T_DOUBLE) {
1524 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1525 } else {
1526 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1527 }
1528 ld = _gvn.transform(ld);
1529
1530 if (((bt == T_OBJECT || bt == T_VALUETYPE || bt == T_VALUETYPEPTR) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1531 // Improve graph before escape analysis and boxing elimination.
1532 record_for_igvn(ld);
1533 }
1534 return ld;
1535 }
1536
1537 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1538 int adr_idx,
1539 MemNode::MemOrd mo,
1540 bool require_atomic_access,
1541 bool unaligned,
1542 bool mismatched) {
1543 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1544 const TypePtr* adr_type = NULL;
1545 debug_only(adr_type = C->get_adr_type(adr_idx));
1546 Node *mem = memory(adr_idx);
1547 Node* st;
1548 if (require_atomic_access && bt == T_LONG) {
1549 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1550 } else if (require_atomic_access && bt == T_DOUBLE) {
1551 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1552 } else {
1553 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1554 }
1555 if (unaligned) {
1556 st->as_Store()->set_unaligned_access();
1557 }
1558 if (mismatched) {
1559 st->as_Store()->set_mismatched_access();
1560 }
1561 st = _gvn.transform(st);
1562 set_memory(st, adr_idx);
1563 // Back-to-back stores can only remove intermediate store with DU info
1564 // so push on worklist for optimizer.
1565 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1566 record_for_igvn(st);
1567
1568 return st;
1569 }
1570
1571 Node* GraphKit::access_store_at(Node* ctl,
1572 Node* obj,
1573 Node* adr,
1574 const TypePtr* adr_type,
1575 Node* val,
1576 const Type* val_type,
1577 BasicType bt,
1578 DecoratorSet decorators,
1579 bool deoptimize_on_exception) {
1580 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1581 // could be delayed during Parse (for example, in adjust_map_after_if()).
1582 // Execute transformation here to avoid barrier generation in such case.
1583 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1584 val = _gvn.makecon(TypePtr::NULL_PTR);
1585 }
1586
1587 set_control(ctl);
1588 if (stopped()) {
1589 return top(); // Dead path ?
1590 }
1591
1592 assert(val != NULL, "not dead path");
1593 if (val->is_ValueType()) {
1594 // Allocate value type and get oop
1595 val = val->as_ValueType()->allocate(this, deoptimize_on_exception)->get_oop();
1596 }
1597
1598 C2AccessValuePtr addr(adr, adr_type);
1599 C2AccessValue value(val, val_type);
1600 C2Access access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1601 if (access.is_raw()) {
1602 return _barrier_set->BarrierSetC2::store_at(access, value);
1603 } else {
1604 return _barrier_set->store_at(access, value);
1605 }
1606 }
1607
1608 Node* GraphKit::access_load_at(Node* obj, // containing obj
1609 Node* adr, // actual adress to store val at
1610 const TypePtr* adr_type,
1611 const Type* val_type,
1612 BasicType bt,
1613 DecoratorSet decorators) {
1614 if (stopped()) {
1615 return top(); // Dead path ?
1616 }
1617
1618 C2AccessValuePtr addr(adr, adr_type);
1619 C2Access access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1620 if (access.is_raw()) {
1621 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1622 } else {
1623 return _barrier_set->load_at(access, val_type);
1624 }
1625 }
1626
1627 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* ctl,
1628 Node* obj,
1629 Node* adr,
1630 const TypePtr* adr_type,
1631 int alias_idx,
1632 Node* expected_val,
1633 Node* new_val,
1634 const Type* value_type,
1635 BasicType bt,
1636 DecoratorSet decorators) {
1637 set_control(ctl);
1638 C2AccessValuePtr addr(adr, adr_type);
1639 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1640 bt, obj, addr, alias_idx);
1641 if (access.is_raw()) {
1642 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1643 } else {
1644 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1645 }
1646 }
1647
1648 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* ctl,
1649 Node* obj,
1650 Node* adr,
1651 const TypePtr* adr_type,
1652 int alias_idx,
1653 Node* expected_val,
1654 Node* new_val,
1655 const Type* value_type,
1656 BasicType bt,
1657 DecoratorSet decorators) {
1658 set_control(ctl);
1659 C2AccessValuePtr addr(adr, adr_type);
1660 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1661 bt, obj, addr, alias_idx);
1662 if (access.is_raw()) {
1663 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1664 } else {
1665 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1666 }
1667 }
1668
1669 Node* GraphKit::access_atomic_xchg_at(Node* ctl,
1670 Node* obj,
1671 Node* adr,
1672 const TypePtr* adr_type,
1673 int alias_idx,
1674 Node* new_val,
1675 const Type* value_type,
1676 BasicType bt,
1677 DecoratorSet decorators) {
1678 set_control(ctl);
1679 C2AccessValuePtr addr(adr, adr_type);
1680 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1681 bt, obj, addr, alias_idx);
1682 if (access.is_raw()) {
1683 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1684 } else {
1685 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1686 }
1687 }
1688
1689 Node* GraphKit::access_atomic_add_at(Node* ctl,
1690 Node* obj,
1691 Node* adr,
1692 const TypePtr* adr_type,
1693 int alias_idx,
1694 Node* new_val,
1695 const Type* value_type,
1696 BasicType bt,
1697 DecoratorSet decorators) {
1698 set_control(ctl);
1699 C2AccessValuePtr addr(adr, adr_type);
1700 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1701 if (access.is_raw()) {
1702 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1703 } else {
1704 return _barrier_set->atomic_add_at(access, new_val, value_type);
1705 }
1706 }
1707
1708 void GraphKit::access_clone(Node* ctl, Node* src, Node* dst, Node* size, bool is_array) {
1709 set_control(ctl);
1710 return _barrier_set->clone(this, src, dst, size, is_array);
1711 }
1712
1713 //-------------------------array_element_address-------------------------
1714 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1715 const TypeInt* sizetype, Node* ctrl) {
1716 uint shift = exact_log2(type2aelembytes(elembt));
1717 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1718 if (arytype_klass->is_value_array_klass()) {
1719 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass();
1720 shift = vak->log2_element_size();
1721 }
1722 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1723
1724 // short-circuit a common case (saves lots of confusing waste motion)
1725 jint idx_con = find_int_con(idx, -1);
1726 if (idx_con >= 0) {
1727 intptr_t offset = header + ((intptr_t)idx_con << shift);
1728 return basic_plus_adr(ary, offset);
1729 }
1730
1731 // must be correct type for alignment purposes
1732 Node* base = basic_plus_adr(ary, header);
1733 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1734 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1735 return basic_plus_adr(ary, base, scale);
1736 }
1737
1738 //-------------------------load_array_element-------------------------
1739 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1740 const Type* elemtype = arytype->elem();
1741 BasicType elembt = elemtype->array_element_basic_type();
1742 assert(elembt != T_VALUETYPE, "value types are not supported by this method");
1743 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1744 if (elembt == T_NARROWOOP) {
1745 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1746 }
1747 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1748 return ld;
1749 }
1750
1751 //-------------------------set_arguments_for_java_call-------------------------
1752 // Arguments (pre-popped from the stack) are taken from the JVMS.
1753 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1754 // Add the call arguments:
1755 const TypeTuple* domain = call->tf()->domain_sig();
1756 uint nargs = domain->cnt();
1757 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1758 Node* arg = argument(i-TypeFunc::Parms);
1759 const Type* t = domain->field_at(i);
1760 if (arg->is_ValueType()) {
1761 assert(t->is_oopptr()->can_be_value_type(), "wrong argument type");
1762 ValueTypeNode* vt = arg->as_ValueType();
1763 if (ValueTypePassFieldsAsArgs) {
1764 // We don't pass value type arguments by reference but instead
1765 // pass each field of the value type
1766 idx += vt->pass_fields(call, idx, *this);
1767 // If a value type argument is passed as fields, attach the Method* to the call site
1768 // to be able to access the extended signature later via attached_method_before_pc().
1769 // For example, see CompiledMethod::preserve_callee_argument_oops().
1770 call->set_override_symbolic_info(true);
1771 continue;
1772 } else {
1773 // Pass value type argument via oop to callee
1774 arg = vt->allocate(this)->get_oop();
1775 }
1776 } else if (t->is_valuetypeptr()) {
1777 ciMethod* declared_method = method()->get_method_at_bci(bci());
1778 if (arg->is_ValueTypePtr()) {
1779 // We are calling an Object method with a value type receiver
1780 ValueTypePtrNode* vt = arg->isa_ValueTypePtr();
1781 assert(i == TypeFunc::Parms && !declared_method->is_static(), "argument must be receiver");
1782 assert(vt->is_allocated(&gvn()), "argument must be allocated");
1783 arg = vt->get_oop();
1784 } else {
1785 // Constant null passed for a value type argument
1786 assert(arg->bottom_type()->remove_speculative() == TypePtr::NULL_PTR, "Anything other than null?");
1787 int arg_size = declared_method->signature()->arg_size_for_bc(java_bc());
1788 inc_sp(arg_size); // restore arguments
1789 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none);
1790 return;
1791 }
1792 }
1793 call->init_req(idx, arg);
1794 idx++;
1795 }
1796 }
1797
1798 //---------------------------set_edges_for_java_call---------------------------
1799 // Connect a newly created call into the current JVMS.
1800 // A return value node (if any) is returned from set_edges_for_java_call.
1801 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1802
1803 // Add the predefined inputs:
1804 call->init_req( TypeFunc::Control, control() );
1805 call->init_req( TypeFunc::I_O , i_o() );
1806 call->init_req( TypeFunc::Memory , reset_memory() );
1807 call->init_req( TypeFunc::FramePtr, frameptr() );
1808 call->init_req( TypeFunc::ReturnAdr, top() );
1809
1810 add_safepoint_edges(call, must_throw);
1811
1812 Node* xcall = _gvn.transform(call);
1813
1814 if (xcall == top()) {
1815 set_control(top());
1816 return;
1817 }
1818 assert(xcall == call, "call identity is stable");
1819
1820 // Re-use the current map to produce the result.
1821
1822 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1823 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1824 set_all_memory_call(xcall, separate_io_proj);
1825
1826 //return xcall; // no need, caller already has it
1827 }
1828
1829 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1830 if (stopped()) return top(); // maybe the call folded up?
1831
1832 // Note: Since any out-of-line call can produce an exception,
1833 // we always insert an I_O projection from the call into the result.
1834
1835 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1836
1837 if (separate_io_proj) {
1838 // The caller requested separate projections be used by the fall
1839 // through and exceptional paths, so replace the projections for
1840 // the fall through path.
1841 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1842 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1843 }
1844
1845 // Capture the return value, if any.
1846 Node* ret;
1847 if (call->method() == NULL ||
1848 call->method()->return_type()->basic_type() == T_VOID) {
1849 ret = top();
1850 } else {
1851 if (!call->tf()->returns_value_type_as_fields()) {
1852 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1853 } else {
1854 // Return of multiple values (value type fields): we create a
1855 // ValueType node, each field is a projection from the call.
1856 const TypeTuple* range_sig = call->tf()->range_sig();
1857 const Type* t = range_sig->field_at(TypeFunc::Parms);
1858 assert(t->is_valuetypeptr(), "only value types for multiple return values");
1859 ciValueKlass* vk = t->value_klass();
1860 ret = ValueTypeNode::make_from_multi(this, call, vk, TypeFunc::Parms+1, false);
1861 }
1862 }
1863
1864 return ret;
1865 }
1866
1867 //--------------------set_predefined_input_for_runtime_call--------------------
1868 // Reading and setting the memory state is way conservative here.
1869 // The real problem is that I am not doing real Type analysis on memory,
1870 // so I cannot distinguish card mark stores from other stores. Across a GC
1871 // point the Store Barrier and the card mark memory has to agree. I cannot
1872 // have a card mark store and its barrier split across the GC point from
1873 // either above or below. Here I get that to happen by reading ALL of memory.
1874 // A better answer would be to separate out card marks from other memory.
1875 // For now, return the input memory state, so that it can be reused
1876 // after the call, if this call has restricted memory effects.
1877 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1878 // Set fixed predefined input arguments
1879 Node* memory = reset_memory();
1880 call->init_req( TypeFunc::Control, control() );
1881 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1882 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1883 call->init_req( TypeFunc::FramePtr, frameptr() );
1884 call->init_req( TypeFunc::ReturnAdr, top() );
1885 return memory;
1886 }
1887
1888 //-------------------set_predefined_output_for_runtime_call--------------------
1889 // Set control and memory (not i_o) from the call.
1890 // If keep_mem is not NULL, use it for the output state,
1891 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1892 // If hook_mem is NULL, this call produces no memory effects at all.
1893 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1894 // then only that memory slice is taken from the call.
1895 // In the last case, we must put an appropriate memory barrier before
1896 // the call, so as to create the correct anti-dependencies on loads
1897 // preceding the call.
1898 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1899 Node* keep_mem,
1900 const TypePtr* hook_mem) {
1901 // no i/o
1902 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1903 if (keep_mem) {
1904 // First clone the existing memory state
1905 set_all_memory(keep_mem);
1906 if (hook_mem != NULL) {
1907 // Make memory for the call
1908 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1909 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1910 // We also use hook_mem to extract specific effects from arraycopy stubs.
1911 set_memory(mem, hook_mem);
1912 }
1913 // ...else the call has NO memory effects.
1914
1915 // Make sure the call advertises its memory effects precisely.
1916 // This lets us build accurate anti-dependences in gcm.cpp.
1917 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1918 "call node must be constructed correctly");
1919 } else {
1920 assert(hook_mem == NULL, "");
1921 // This is not a "slow path" call; all memory comes from the call.
1922 set_all_memory_call(call);
1923 }
1924 }
1925
1926
1927 // Replace the call with the current state of the kit.
1928 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1929 JVMState* ejvms = NULL;
1930 if (has_exceptions()) {
1931 ejvms = transfer_exceptions_into_jvms();
1932 }
1933
1934 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1935 ReplacedNodes replaced_nodes_exception;
1936 Node* ex_ctl = top();
1937
1938 SafePointNode* final_state = stop();
1939
1940 // Find all the needed outputs of this call
1941 CallProjections* callprojs = call->extract_projections(true);
1942
1943 Node* init_mem = call->in(TypeFunc::Memory);
1944 Node* final_mem = final_state->in(TypeFunc::Memory);
1945 Node* final_ctl = final_state->in(TypeFunc::Control);
1946 Node* final_io = final_state->in(TypeFunc::I_O);
1947
1948 // Replace all the old call edges with the edges from the inlining result
1949 if (callprojs->fallthrough_catchproj != NULL) {
1950 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
1951 }
1952 if (callprojs->fallthrough_memproj != NULL) {
1953 if (final_mem->is_MergeMem()) {
1954 // Parser's exits MergeMem was not transformed but may be optimized
1955 final_mem = _gvn.transform(final_mem);
1956 }
1957 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
1958 }
1959 if (callprojs->fallthrough_ioproj != NULL) {
1960 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
1961 }
1962
1963 // Replace the result with the new result if it exists and is used
1964 if (callprojs->resproj[0] != NULL && result != NULL) {
1965 assert(callprojs->nb_resproj == 1, "unexpected number of results");
1966 C->gvn_replace_by(callprojs->resproj[0], result);
1967 }
1968
1969 if (ejvms == NULL) {
1970 // No exception edges to simply kill off those paths
1971 if (callprojs->catchall_catchproj != NULL) {
1972 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
1973 }
1974 if (callprojs->catchall_memproj != NULL) {
1975 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
1976 }
1977 if (callprojs->catchall_ioproj != NULL) {
1978 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
1979 }
1980 // Replace the old exception object with top
1981 if (callprojs->exobj != NULL) {
1982 C->gvn_replace_by(callprojs->exobj, C->top());
1983 }
1984 } else {
1985 GraphKit ekit(ejvms);
1986
1987 // Load my combined exception state into the kit, with all phis transformed:
1988 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1989 replaced_nodes_exception = ex_map->replaced_nodes();
1990
1991 Node* ex_oop = ekit.use_exception_state(ex_map);
1992
1993 if (callprojs->catchall_catchproj != NULL) {
1994 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
1995 ex_ctl = ekit.control();
1996 }
1997 if (callprojs->catchall_memproj != NULL) {
1998 C->gvn_replace_by(callprojs->catchall_memproj, ekit.reset_memory());
1999 }
2000 if (callprojs->catchall_ioproj != NULL) {
2001 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2002 }
2003
2004 // Replace the old exception object with the newly created one
2005 if (callprojs->exobj != NULL) {
2006 C->gvn_replace_by(callprojs->exobj, ex_oop);
2007 }
2008 }
2009
2010 // Disconnect the call from the graph
2011 call->disconnect_inputs(NULL, C);
2012 C->gvn_replace_by(call, C->top());
2013
2014 // Clean up any MergeMems that feed other MergeMems since the
2015 // optimizer doesn't like that.
2016 if (final_mem->is_MergeMem()) {
2017 Node_List wl;
2018 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
2019 Node* m = i.get();
2020 if (m->is_MergeMem() && !wl.contains(m)) {
2021 wl.push(m);
2022 }
2023 }
2024 while (wl.size() > 0) {
2025 _gvn.transform(wl.pop());
2026 }
2027 }
2028
2029 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2030 replaced_nodes.apply(C, final_ctl);
2031 }
2032 if (!ex_ctl->is_top() && do_replaced_nodes) {
2033 replaced_nodes_exception.apply(C, ex_ctl);
2034 }
2035 }
2036
2037
2038 //------------------------------increment_counter------------------------------
2039 // for statistics: increment a VM counter by 1
2040
2041 void GraphKit::increment_counter(address counter_addr) {
2042 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2043 increment_counter(adr1);
2044 }
2045
2046 void GraphKit::increment_counter(Node* counter_addr) {
2047 int adr_type = Compile::AliasIdxRaw;
2048 Node* ctrl = control();
2049 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2050 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
2051 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
2052 }
2053
2054
2055 //------------------------------uncommon_trap----------------------------------
2056 // Bail out to the interpreter in mid-method. Implemented by calling the
2057 // uncommon_trap blob. This helper function inserts a runtime call with the
2058 // right debug info.
2059 void GraphKit::uncommon_trap(int trap_request,
2060 ciKlass* klass, const char* comment,
2061 bool must_throw,
2062 bool keep_exact_action) {
2063 if (failing()) stop();
2064 if (stopped()) return; // trap reachable?
2065
2066 // Note: If ProfileTraps is true, and if a deopt. actually
2067 // occurs here, the runtime will make sure an MDO exists. There is
2068 // no need to call method()->ensure_method_data() at this point.
2069
2070 // Set the stack pointer to the right value for reexecution:
2071 set_sp(reexecute_sp());
2072
2073 #ifdef ASSERT
2074 if (!must_throw) {
2075 // Make sure the stack has at least enough depth to execute
2076 // the current bytecode.
2077 int inputs, ignored_depth;
2078 if (compute_stack_effects(inputs, ignored_depth)) {
2079 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2080 Bytecodes::name(java_bc()), sp(), inputs);
2081 }
2082 }
2083 #endif
2084
2085 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2086 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2087
2088 switch (action) {
2089 case Deoptimization::Action_maybe_recompile:
2090 case Deoptimization::Action_reinterpret:
2091 // Temporary fix for 6529811 to allow virtual calls to be sure they
2092 // get the chance to go from mono->bi->mega
2093 if (!keep_exact_action &&
2094 Deoptimization::trap_request_index(trap_request) < 0 &&
2095 too_many_recompiles(reason)) {
2096 // This BCI is causing too many recompilations.
2097 if (C->log() != NULL) {
2098 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2099 Deoptimization::trap_reason_name(reason),
2100 Deoptimization::trap_action_name(action));
2101 }
2102 action = Deoptimization::Action_none;
2103 trap_request = Deoptimization::make_trap_request(reason, action);
2104 } else {
2105 C->set_trap_can_recompile(true);
2106 }
2107 break;
2108 case Deoptimization::Action_make_not_entrant:
2109 C->set_trap_can_recompile(true);
2110 break;
2111 case Deoptimization::Action_none:
2112 case Deoptimization::Action_make_not_compilable:
2113 break;
2114 default:
2115 #ifdef ASSERT
2116 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2117 #endif
2118 break;
2119 }
2120
2121 if (TraceOptoParse) {
2122 char buf[100];
2123 tty->print_cr("Uncommon trap %s at bci:%d",
2124 Deoptimization::format_trap_request(buf, sizeof(buf),
2125 trap_request), bci());
2126 }
2127
2128 CompileLog* log = C->log();
2129 if (log != NULL) {
2130 int kid = (klass == NULL)? -1: log->identify(klass);
2131 log->begin_elem("uncommon_trap bci='%d'", bci());
2132 char buf[100];
2133 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2134 trap_request));
2135 if (kid >= 0) log->print(" klass='%d'", kid);
2136 if (comment != NULL) log->print(" comment='%s'", comment);
2137 log->end_elem();
2138 }
2139
2140 // Make sure any guarding test views this path as very unlikely
2141 Node *i0 = control()->in(0);
2142 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2143 IfNode *iff = i0->as_If();
2144 float f = iff->_prob; // Get prob
2145 if (control()->Opcode() == Op_IfTrue) {
2146 if (f > PROB_UNLIKELY_MAG(4))
2147 iff->_prob = PROB_MIN;
2148 } else {
2149 if (f < PROB_LIKELY_MAG(4))
2150 iff->_prob = PROB_MAX;
2151 }
2152 }
2153
2154 // Clear out dead values from the debug info.
2155 kill_dead_locals();
2156
2157 // Now insert the uncommon trap subroutine call
2158 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2159 const TypePtr* no_memory_effects = NULL;
2160 // Pass the index of the class to be loaded
2161 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2162 (must_throw ? RC_MUST_THROW : 0),
2163 OptoRuntime::uncommon_trap_Type(),
2164 call_addr, "uncommon_trap", no_memory_effects,
2165 intcon(trap_request));
2166 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2167 "must extract request correctly from the graph");
2168 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2169
2170 call->set_req(TypeFunc::ReturnAdr, returnadr());
2171 // The debug info is the only real input to this call.
2172
2173 // Halt-and-catch fire here. The above call should never return!
2174 HaltNode* halt = new HaltNode(control(), frameptr());
2175 _gvn.set_type_bottom(halt);
2176 root()->add_req(halt);
2177
2178 stop_and_kill_map();
2179 }
2180
2181
2182 //--------------------------just_allocated_object------------------------------
2183 // Report the object that was just allocated.
2184 // It must be the case that there are no intervening safepoints.
2185 // We use this to determine if an object is so "fresh" that
2186 // it does not require card marks.
2187 Node* GraphKit::just_allocated_object(Node* current_control) {
2188 if (C->recent_alloc_ctl() == current_control)
2189 return C->recent_alloc_obj();
2190 return NULL;
2191 }
2192
2193
2194 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2195 // (Note: TypeFunc::make has a cache that makes this fast.)
2196 const TypeFunc* tf = TypeFunc::make(dest_method);
2197 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2198 for (int j = 0; j < nargs; j++) {
2199 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2200 if( targ->basic_type() == T_DOUBLE ) {
2201 // If any parameters are doubles, they must be rounded before
2202 // the call, dstore_rounding does gvn.transform
2203 Node *arg = argument(j);
2204 arg = dstore_rounding(arg);
2205 set_argument(j, arg);
2206 }
2207 }
2208 }
2209
2210 /**
2211 * Record profiling data exact_kls for Node n with the type system so
2212 * that it can propagate it (speculation)
2213 *
2214 * @param n node that the type applies to
2215 * @param exact_kls type from profiling
2216 * @param maybe_null did profiling see null?
2217 *
2218 * @return node with improved type
2219 */
2220 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2221 const Type* current_type = _gvn.type(n);
2222 assert(UseTypeSpeculation, "type speculation must be on");
2223
2224 const TypePtr* speculative = current_type->speculative();
2225
2226 // Should the klass from the profile be recorded in the speculative type?
2227 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2228 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2229 const TypeOopPtr* xtype = tklass->as_instance_type();
2230 assert(xtype->klass_is_exact(), "Should be exact");
2231 // Any reason to believe n is not null (from this profiling or a previous one)?
2232 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2233 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2234 // record the new speculative type's depth
2235 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2236 speculative = speculative->with_inline_depth(jvms()->depth());
2237 } else if (current_type->would_improve_ptr(ptr_kind)) {
2238 // Profiling report that null was never seen so we can change the
2239 // speculative type to non null ptr.
2240 if (ptr_kind == ProfileAlwaysNull) {
2241 speculative = TypePtr::NULL_PTR;
2242 } else {
2243 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2244 const TypePtr* ptr = TypePtr::NOTNULL;
2245 if (speculative != NULL) {
2246 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2247 } else {
2248 speculative = ptr;
2249 }
2250 }
2251 }
2252
2253 if (speculative != current_type->speculative()) {
2254 // Build a type with a speculative type (what we think we know
2255 // about the type but will need a guard when we use it)
2256 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2257 // We're changing the type, we need a new CheckCast node to carry
2258 // the new type. The new type depends on the control: what
2259 // profiling tells us is only valid from here as far as we can
2260 // tell.
2261 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2262 cast = _gvn.transform(cast);
2263 replace_in_map(n, cast);
2264 n = cast;
2265 }
2266
2267 return n;
2268 }
2269
2270 /**
2271 * Record profiling data from receiver profiling at an invoke with the
2272 * type system so that it can propagate it (speculation)
2273 *
2274 * @param n receiver node
2275 *
2276 * @return node with improved type
2277 */
2278 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2279 if (!UseTypeSpeculation) {
2280 return n;
2281 }
2282 ciKlass* exact_kls = profile_has_unique_klass();
2283 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2284 if ((java_bc() == Bytecodes::_checkcast ||
2285 java_bc() == Bytecodes::_instanceof ||
2286 java_bc() == Bytecodes::_aastore) &&
2287 method()->method_data()->is_mature()) {
2288 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2289 if (data != NULL) {
2290 if (!data->as_BitData()->null_seen()) {
2291 ptr_kind = ProfileNeverNull;
2292 } else {
2293 assert(data->is_ReceiverTypeData(), "bad profile data type");
2294 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2295 uint i = 0;
2296 for (; i < call->row_limit(); i++) {
2297 ciKlass* receiver = call->receiver(i);
2298 if (receiver != NULL) {
2299 break;
2300 }
2301 }
2302 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2303 }
2304 }
2305 }
2306 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2307 }
2308
2309 /**
2310 * Record profiling data from argument profiling at an invoke with the
2311 * type system so that it can propagate it (speculation)
2312 *
2313 * @param dest_method target method for the call
2314 * @param bc what invoke bytecode is this?
2315 */
2316 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2317 if (!UseTypeSpeculation) {
2318 return;
2319 }
2320 const TypeFunc* tf = TypeFunc::make(dest_method);
2321 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2322 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2323 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2324 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2325 if (targ->isa_oopptr()) {
2326 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2327 ciKlass* better_type = NULL;
2328 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2329 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2330 }
2331 i++;
2332 }
2333 }
2334 }
2335
2336 /**
2337 * Record profiling data from parameter profiling at an invoke with
2338 * the type system so that it can propagate it (speculation)
2339 */
2340 void GraphKit::record_profiled_parameters_for_speculation() {
2341 if (!UseTypeSpeculation) {
2342 return;
2343 }
2344 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2345 if (_gvn.type(local(i))->isa_oopptr()) {
2346 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2347 ciKlass* better_type = NULL;
2348 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2349 record_profile_for_speculation(local(i), better_type, ptr_kind);
2350 }
2351 j++;
2352 }
2353 }
2354 }
2355
2356 /**
2357 * Record profiling data from return value profiling at an invoke with
2358 * the type system so that it can propagate it (speculation)
2359 */
2360 void GraphKit::record_profiled_return_for_speculation() {
2361 if (!UseTypeSpeculation) {
2362 return;
2363 }
2364 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2365 ciKlass* better_type = NULL;
2366 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2367 // If profiling reports a single type for the return value,
2368 // feed it to the type system so it can propagate it as a
2369 // speculative type
2370 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2371 }
2372 }
2373
2374 void GraphKit::round_double_result(ciMethod* dest_method) {
2375 // A non-strict method may return a double value which has an extended
2376 // exponent, but this must not be visible in a caller which is 'strict'
2377 // If a strict caller invokes a non-strict callee, round a double result
2378
2379 BasicType result_type = dest_method->return_type()->basic_type();
2380 assert( method() != NULL, "must have caller context");
2381 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2382 // Destination method's return value is on top of stack
2383 // dstore_rounding() does gvn.transform
2384 Node *result = pop_pair();
2385 result = dstore_rounding(result);
2386 push_pair(result);
2387 }
2388 }
2389
2390 // rounding for strict float precision conformance
2391 Node* GraphKit::precision_rounding(Node* n) {
2392 return UseStrictFP && _method->flags().is_strict()
2393 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2394 ? _gvn.transform( new RoundFloatNode(0, n) )
2395 : n;
2396 }
2397
2398 // rounding for strict double precision conformance
2399 Node* GraphKit::dprecision_rounding(Node *n) {
2400 return UseStrictFP && _method->flags().is_strict()
2401 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2402 ? _gvn.transform( new RoundDoubleNode(0, n) )
2403 : n;
2404 }
2405
2406 // rounding for non-strict double stores
2407 Node* GraphKit::dstore_rounding(Node* n) {
2408 return Matcher::strict_fp_requires_explicit_rounding
2409 && UseSSE <= 1
2410 ? _gvn.transform( new RoundDoubleNode(0, n) )
2411 : n;
2412 }
2413
2414 //=============================================================================
2415 // Generate a fast path/slow path idiom. Graph looks like:
2416 // [foo] indicates that 'foo' is a parameter
2417 //
2418 // [in] NULL
2419 // \ /
2420 // CmpP
2421 // Bool ne
2422 // If
2423 // / \
2424 // True False-<2>
2425 // / |
2426 // / cast_not_null
2427 // Load | | ^
2428 // [fast_test] | |
2429 // gvn to opt_test | |
2430 // / \ | <1>
2431 // True False |
2432 // | \\ |
2433 // [slow_call] \[fast_result]
2434 // Ctl Val \ \
2435 // | \ \
2436 // Catch <1> \ \
2437 // / \ ^ \ \
2438 // Ex No_Ex | \ \
2439 // | \ \ | \ <2> \
2440 // ... \ [slow_res] | | \ [null_result]
2441 // \ \--+--+--- | |
2442 // \ | / \ | /
2443 // --------Region Phi
2444 //
2445 //=============================================================================
2446 // Code is structured as a series of driver functions all called 'do_XXX' that
2447 // call a set of helper functions. Helper functions first, then drivers.
2448
2449 //------------------------------null_check_oop---------------------------------
2450 // Null check oop. Set null-path control into Region in slot 3.
2451 // Make a cast-not-nullness use the other not-null control. Return cast.
2452 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2453 bool never_see_null,
2454 bool safe_for_replace,
2455 bool speculative) {
2456 // Initial NULL check taken path
2457 (*null_control) = top();
2458 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2459
2460 // Generate uncommon_trap:
2461 if (never_see_null && (*null_control) != top()) {
2462 // If we see an unexpected null at a check-cast we record it and force a
2463 // recompile; the offending check-cast will be compiled to handle NULLs.
2464 // If we see more than one offending BCI, then all checkcasts in the
2465 // method will be compiled to handle NULLs.
2466 PreserveJVMState pjvms(this);
2467 set_control(*null_control);
2468 replace_in_map(value, null());
2469 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2470 uncommon_trap(reason,
2471 Deoptimization::Action_make_not_entrant);
2472 (*null_control) = top(); // NULL path is dead
2473 }
2474 if ((*null_control) == top() && safe_for_replace) {
2475 replace_in_map(value, cast);
2476 }
2477
2478 // Cast away null-ness on the result
2479 return cast;
2480 }
2481
2482 //------------------------------opt_iff----------------------------------------
2483 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2484 // Return slow-path control.
2485 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2486 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2487
2488 // Fast path taken; set region slot 2
2489 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2490 region->init_req(2,fast_taken); // Capture fast-control
2491
2492 // Fast path not-taken, i.e. slow path
2493 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2494 return slow_taken;
2495 }
2496
2497 //-----------------------------make_runtime_call-------------------------------
2498 Node* GraphKit::make_runtime_call(int flags,
2499 const TypeFunc* call_type, address call_addr,
2500 const char* call_name,
2501 const TypePtr* adr_type,
2502 // The following parms are all optional.
2503 // The first NULL ends the list.
2504 Node* parm0, Node* parm1,
2505 Node* parm2, Node* parm3,
2506 Node* parm4, Node* parm5,
2507 Node* parm6, Node* parm7) {
2508 // Slow-path call
2509 bool is_leaf = !(flags & RC_NO_LEAF);
2510 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2511 if (call_name == NULL) {
2512 assert(!is_leaf, "must supply name for leaf");
2513 call_name = OptoRuntime::stub_name(call_addr);
2514 }
2515 CallNode* call;
2516 if (!is_leaf) {
2517 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2518 bci(), adr_type);
2519 } else if (flags & RC_NO_FP) {
2520 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2521 } else {
2522 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2523 }
2524
2525 // The following is similar to set_edges_for_java_call,
2526 // except that the memory effects of the call are restricted to AliasIdxRaw.
2527
2528 // Slow path call has no side-effects, uses few values
2529 bool wide_in = !(flags & RC_NARROW_MEM);
2530 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2531
2532 Node* prev_mem = NULL;
2533 if (wide_in) {
2534 prev_mem = set_predefined_input_for_runtime_call(call);
2535 } else {
2536 assert(!wide_out, "narrow in => narrow out");
2537 Node* narrow_mem = memory(adr_type);
2538 prev_mem = reset_memory();
2539 map()->set_memory(narrow_mem);
2540 set_predefined_input_for_runtime_call(call);
2541 }
2542
2543 // Hook each parm in order. Stop looking at the first NULL.
2544 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2545 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2546 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2547 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2548 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2549 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2550 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2551 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2552 /* close each nested if ===> */ } } } } } } } }
2553 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2554
2555 if (!is_leaf) {
2556 // Non-leaves can block and take safepoints:
2557 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2558 }
2559 // Non-leaves can throw exceptions:
2560 if (has_io) {
2561 call->set_req(TypeFunc::I_O, i_o());
2562 }
2563
2564 if (flags & RC_UNCOMMON) {
2565 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2566 // (An "if" probability corresponds roughly to an unconditional count.
2567 // Sort of.)
2568 call->set_cnt(PROB_UNLIKELY_MAG(4));
2569 }
2570
2571 Node* c = _gvn.transform(call);
2572 assert(c == call, "cannot disappear");
2573
2574 if (wide_out) {
2575 // Slow path call has full side-effects.
2576 set_predefined_output_for_runtime_call(call);
2577 } else {
2578 // Slow path call has few side-effects, and/or sets few values.
2579 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2580 }
2581
2582 if (has_io) {
2583 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2584 }
2585 return call;
2586
2587 }
2588
2589 //------------------------------merge_memory-----------------------------------
2590 // Merge memory from one path into the current memory state.
2591 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2592 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2593 Node* old_slice = mms.force_memory();
2594 Node* new_slice = mms.memory2();
2595 if (old_slice != new_slice) {
2596 PhiNode* phi;
2597 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2598 if (mms.is_empty()) {
2599 // clone base memory Phi's inputs for this memory slice
2600 assert(old_slice == mms.base_memory(), "sanity");
2601 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2602 _gvn.set_type(phi, Type::MEMORY);
2603 for (uint i = 1; i < phi->req(); i++) {
2604 phi->init_req(i, old_slice->in(i));
2605 }
2606 } else {
2607 phi = old_slice->as_Phi(); // Phi was generated already
2608 }
2609 } else {
2610 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2611 _gvn.set_type(phi, Type::MEMORY);
2612 }
2613 phi->set_req(new_path, new_slice);
2614 mms.set_memory(phi);
2615 }
2616 }
2617 }
2618
2619 //------------------------------make_slow_call_ex------------------------------
2620 // Make the exception handler hookups for the slow call
2621 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2622 if (stopped()) return;
2623
2624 // Make a catch node with just two handlers: fall-through and catch-all
2625 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2626 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2627 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2628 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2629
2630 { PreserveJVMState pjvms(this);
2631 set_control(excp);
2632 set_i_o(i_o);
2633
2634 if (excp != top()) {
2635 if (deoptimize) {
2636 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2637 uncommon_trap(Deoptimization::Reason_unhandled,
2638 Deoptimization::Action_none);
2639 } else {
2640 // Create an exception state also.
2641 // Use an exact type if the caller has specified a specific exception.
2642 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2643 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2644 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2645 }
2646 }
2647 }
2648
2649 // Get the no-exception control from the CatchNode.
2650 set_control(norm);
2651 }
2652
2653 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2654 Node* cmp = NULL;
2655 switch(bt) {
2656 case T_INT: cmp = new CmpINode(in1, in2); break;
2657 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2658 default: fatal("unexpected comparison type %s", type2name(bt));
2659 }
2660 gvn->transform(cmp);
2661 Node* bol = gvn->transform(new BoolNode(cmp, test));
2662 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2663 gvn->transform(iff);
2664 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2665 return iff;
2666 }
2667
2668
2669 //-------------------------------gen_subtype_check-----------------------------
2670 // Generate a subtyping check. Takes as input the subtype and supertype.
2671 // Returns 2 values: sets the default control() to the true path and returns
2672 // the false path. Only reads invariant memory; sets no (visible) memory.
2673 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2674 // but that's not exposed to the optimizer. This call also doesn't take in an
2675 // Object; if you wish to check an Object you need to load the Object's class
2676 // prior to coming here.
2677 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2678 Compile* C = gvn->C;
2679
2680 if ((*ctrl)->is_top()) {
2681 return C->top();
2682 }
2683
2684 // Fast check for identical types, perhaps identical constants.
2685 // The types can even be identical non-constants, in cases
2686 // involving Array.newInstance, Object.clone, etc.
2687 if (subklass == superklass)
2688 return C->top(); // false path is dead; no test needed.
2689
2690 if (gvn->type(superklass)->singleton()) {
2691 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2692 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2693
2694 // In the common case of an exact superklass, try to fold up the
2695 // test before generating code. You may ask, why not just generate
2696 // the code and then let it fold up? The answer is that the generated
2697 // code will necessarily include null checks, which do not always
2698 // completely fold away. If they are also needless, then they turn
2699 // into a performance loss. Example:
2700 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2701 // Here, the type of 'fa' is often exact, so the store check
2702 // of fa[1]=x will fold up, without testing the nullness of x.
2703 switch (C->static_subtype_check(superk, subk)) {
2704 case Compile::SSC_always_false:
2705 {
2706 Node* always_fail = *ctrl;
2707 *ctrl = gvn->C->top();
2708 return always_fail;
2709 }
2710 case Compile::SSC_always_true:
2711 return C->top();
2712 case Compile::SSC_easy_test:
2713 {
2714 // Just do a direct pointer compare and be done.
2715 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2716 *ctrl = gvn->transform(new IfTrueNode(iff));
2717 return gvn->transform(new IfFalseNode(iff));
2718 }
2719 case Compile::SSC_full_test:
2720 break;
2721 default:
2722 ShouldNotReachHere();
2723 }
2724 }
2725
2726 // %%% Possible further optimization: Even if the superklass is not exact,
2727 // if the subklass is the unique subtype of the superklass, the check
2728 // will always succeed. We could leave a dependency behind to ensure this.
2729
2730 // First load the super-klass's check-offset
2731 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2732 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2733 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2734 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2735 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2736
2737 // Load from the sub-klass's super-class display list, or a 1-word cache of
2738 // the secondary superclass list, or a failing value with a sentinel offset
2739 // if the super-klass is an interface or exceptionally deep in the Java
2740 // hierarchy and we have to scan the secondary superclass list the hard way.
2741 // Worst-case type is a little odd: NULL is allowed as a result (usually
2742 // klass loads can never produce a NULL).
2743 Node *chk_off_X = chk_off;
2744 #ifdef _LP64
2745 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2746 #endif
2747 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2748 // For some types like interfaces the following loadKlass is from a 1-word
2749 // cache which is mutable so can't use immutable memory. Other
2750 // types load from the super-class display table which is immutable.
2751 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2752 Node *kmem = might_be_cache ? m : C->immutable_memory();
2753 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::BOTTOM));
2754
2755 // Compile speed common case: ARE a subtype and we canNOT fail
2756 if( superklass == nkls )
2757 return C->top(); // false path is dead; no test needed.
2758
2759 // See if we get an immediate positive hit. Happens roughly 83% of the
2760 // time. Test to see if the value loaded just previously from the subklass
2761 // is exactly the superklass.
2762 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2763 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2764 *ctrl = gvn->transform(new IfFalseNode(iff1));
2765
2766 // Compile speed common case: Check for being deterministic right now. If
2767 // chk_off is a constant and not equal to cacheoff then we are NOT a
2768 // subklass. In this case we need exactly the 1 test above and we can
2769 // return those results immediately.
2770 if (!might_be_cache) {
2771 Node* not_subtype_ctrl = *ctrl;
2772 *ctrl = iftrue1; // We need exactly the 1 test above
2773 return not_subtype_ctrl;
2774 }
2775
2776 // Gather the various success & failures here
2777 RegionNode *r_ok_subtype = new RegionNode(4);
2778 gvn->record_for_igvn(r_ok_subtype);
2779 RegionNode *r_not_subtype = new RegionNode(3);
2780 gvn->record_for_igvn(r_not_subtype);
2781
2782 r_ok_subtype->init_req(1, iftrue1);
2783
2784 // Check for immediate negative hit. Happens roughly 11% of the time (which
2785 // is roughly 63% of the remaining cases). Test to see if the loaded
2786 // check-offset points into the subklass display list or the 1-element
2787 // cache. If it points to the display (and NOT the cache) and the display
2788 // missed then it's not a subtype.
2789 Node *cacheoff = gvn->intcon(cacheoff_con);
2790 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2791 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2792 *ctrl = gvn->transform(new IfFalseNode(iff2));
2793
2794 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2795 // No performance impact (too rare) but allows sharing of secondary arrays
2796 // which has some footprint reduction.
2797 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2798 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2799 *ctrl = gvn->transform(new IfFalseNode(iff3));
2800
2801 // -- Roads not taken here: --
2802 // We could also have chosen to perform the self-check at the beginning
2803 // of this code sequence, as the assembler does. This would not pay off
2804 // the same way, since the optimizer, unlike the assembler, can perform
2805 // static type analysis to fold away many successful self-checks.
2806 // Non-foldable self checks work better here in second position, because
2807 // the initial primary superclass check subsumes a self-check for most
2808 // types. An exception would be a secondary type like array-of-interface,
2809 // which does not appear in its own primary supertype display.
2810 // Finally, we could have chosen to move the self-check into the
2811 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2812 // dependent manner. But it is worthwhile to have the check here,
2813 // where it can be perhaps be optimized. The cost in code space is
2814 // small (register compare, branch).
2815
2816 // Now do a linear scan of the secondary super-klass array. Again, no real
2817 // performance impact (too rare) but it's gotta be done.
2818 // Since the code is rarely used, there is no penalty for moving it
2819 // out of line, and it can only improve I-cache density.
2820 // The decision to inline or out-of-line this final check is platform
2821 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2822 Node* psc = gvn->transform(
2823 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2824
2825 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2826 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2827 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2828
2829 // Return false path; set default control to true path.
2830 *ctrl = gvn->transform(r_ok_subtype);
2831 return gvn->transform(r_not_subtype);
2832 }
2833
2834 // Profile-driven exact type check:
2835 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2836 float prob,
2837 Node* *casted_receiver) {
2838 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2839 Node* recv_klass = load_object_klass(receiver);
2840 Node* fail = type_check(recv_klass, tklass, prob);
2841 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2842 assert(recv_xtype->klass_is_exact(), "");
2843
2844 // Subsume downstream occurrences of receiver with a cast to
2845 // recv_xtype, since now we know what the type will be.
2846 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2847 Node* res = _gvn.transform(cast);
2848 if (recv_xtype->is_valuetypeptr()) {
2849 assert(!gvn().type(res)->is_ptr()->maybe_null(), "should never be null");
2850 res = ValueTypeNode::make_from_oop(this, res, recv_xtype->value_klass());
2851 }
2852
2853 (*casted_receiver) = res;
2854 // (User must make the replace_in_map call.)
2855
2856 return fail;
2857 }
2858
2859 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
2860 float prob) {
2861 Node* want_klass = makecon(tklass);
2862 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass));
2863 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2864 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2865 set_control( _gvn.transform( new IfTrueNode (iff)));
2866 Node* fail = _gvn.transform( new IfFalseNode(iff));
2867 return fail;
2868 }
2869
2870
2871 //------------------------------seems_never_null-------------------------------
2872 // Use null_seen information if it is available from the profile.
2873 // If we see an unexpected null at a type check we record it and force a
2874 // recompile; the offending check will be recompiled to handle NULLs.
2875 // If we see several offending BCIs, then all checks in the
2876 // method will be recompiled.
2877 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2878 speculating = !_gvn.type(obj)->speculative_maybe_null();
2879 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2880 if (UncommonNullCast // Cutout for this technique
2881 && obj != null() // And not the -Xcomp stupid case?
2882 && !too_many_traps(reason)
2883 ) {
2884 if (speculating) {
2885 return true;
2886 }
2887 if (data == NULL)
2888 // Edge case: no mature data. Be optimistic here.
2889 return true;
2890 // If the profile has not seen a null, assume it won't happen.
2891 assert(java_bc() == Bytecodes::_checkcast ||
2892 java_bc() == Bytecodes::_instanceof ||
2893 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2894 return !data->as_BitData()->null_seen();
2895 }
2896 speculating = false;
2897 return false;
2898 }
2899
2900 //------------------------maybe_cast_profiled_receiver-------------------------
2901 // If the profile has seen exactly one type, narrow to exactly that type.
2902 // Subsequent type checks will always fold up.
2903 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2904 ciKlass* require_klass,
2905 ciKlass* spec_klass,
2906 bool safe_for_replace) {
2907 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2908
2909 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2910
2911 // Make sure we haven't already deoptimized from this tactic.
2912 if (too_many_traps(reason) || too_many_recompiles(reason))
2913 return NULL;
2914
2915 // (No, this isn't a call, but it's enough like a virtual call
2916 // to use the same ciMethod accessor to get the profile info...)
2917 // If we have a speculative type use it instead of profiling (which
2918 // may not help us)
2919 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2920 if (exact_kls != NULL) {// no cast failures here
2921 if (require_klass == NULL ||
2922 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2923 // If we narrow the type to match what the type profile sees or
2924 // the speculative type, we can then remove the rest of the
2925 // cast.
2926 // This is a win, even if the exact_kls is very specific,
2927 // because downstream operations, such as method calls,
2928 // will often benefit from the sharper type.
2929 Node* exact_obj = not_null_obj; // will get updated in place...
2930 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2931 &exact_obj);
2932 { PreserveJVMState pjvms(this);
2933 set_control(slow_ctl);
2934 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2935 }
2936 if (safe_for_replace) {
2937 replace_in_map(not_null_obj, exact_obj);
2938 }
2939 return exact_obj;
2940 }
2941 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2942 }
2943
2944 return NULL;
2945 }
2946
2947 /**
2948 * Cast obj to type and emit guard unless we had too many traps here
2949 * already
2950 *
2951 * @param obj node being casted
2952 * @param type type to cast the node to
2953 * @param not_null true if we know node cannot be null
2954 */
2955 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2956 ciKlass* type,
2957 bool not_null) {
2958 if (stopped()) {
2959 return obj;
2960 }
2961
2962 // type == NULL if profiling tells us this object is always null
2963 if (type != NULL) {
2964 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2965 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2966
2967 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2968 !too_many_traps(class_reason) &&
2969 !too_many_recompiles(class_reason)) {
2970 Node* not_null_obj = NULL;
2971 // not_null is true if we know the object is not null and
2972 // there's no need for a null check
2973 if (!not_null) {
2974 Node* null_ctl = top();
2975 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2976 assert(null_ctl->is_top(), "no null control here");
2977 } else {
2978 not_null_obj = obj;
2979 }
2980
2981 Node* exact_obj = not_null_obj;
2982 ciKlass* exact_kls = type;
2983 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2984 &exact_obj);
2985 {
2986 PreserveJVMState pjvms(this);
2987 set_control(slow_ctl);
2988 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2989 }
2990 replace_in_map(not_null_obj, exact_obj);
2991 obj = exact_obj;
2992 }
2993 } else {
2994 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2995 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2996 Node* exact_obj = null_assert(obj);
2997 replace_in_map(obj, exact_obj);
2998 obj = exact_obj;
2999 }
3000 }
3001 return obj;
3002 }
3003
3004 //-------------------------------gen_instanceof--------------------------------
3005 // Generate an instance-of idiom. Used by both the instance-of bytecode
3006 // and the reflective instance-of call.
3007 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3008 kill_dead_locals(); // Benefit all the uncommon traps
3009 assert( !stopped(), "dead parse path should be checked in callers" );
3010 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3011 "must check for not-null not-dead klass in callers");
3012
3013 // Make the merge point
3014 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3015 RegionNode* region = new RegionNode(PATH_LIMIT);
3016 Node* phi = new PhiNode(region, TypeInt::BOOL);
3017 C->set_has_split_ifs(true); // Has chance for split-if optimization
3018
3019 ciProfileData* data = NULL;
3020 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3021 data = method()->method_data()->bci_to_data(bci());
3022 }
3023 bool speculative_not_null = false;
3024 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3025 && seems_never_null(obj, data, speculative_not_null));
3026 bool is_value = obj->is_ValueType();
3027
3028 // Null check; get casted pointer; set region slot 3
3029 Node* null_ctl = top();
3030 Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3031
3032 // If not_null_obj is dead, only null-path is taken
3033 if (stopped()) { // Doing instance-of on a NULL?
3034 set_control(null_ctl);
3035 return intcon(0);
3036 }
3037 region->init_req(_null_path, null_ctl);
3038 phi ->init_req(_null_path, intcon(0)); // Set null path value
3039 if (null_ctl == top()) {
3040 // Do this eagerly, so that pattern matches like is_diamond_phi
3041 // will work even during parsing.
3042 assert(_null_path == PATH_LIMIT-1, "delete last");
3043 region->del_req(_null_path);
3044 phi ->del_req(_null_path);
3045 }
3046
3047 // Do we know the type check always succeed?
3048 if (!is_value) {
3049 bool known_statically = false;
3050 if (_gvn.type(superklass)->singleton()) {
3051 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3052 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3053 if (subk != NULL && subk->is_loaded()) {
3054 int static_res = C->static_subtype_check(superk, subk);
3055 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3056 }
3057 }
3058
3059 if (!known_statically) {
3060 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3061 // We may not have profiling here or it may not help us. If we
3062 // have a speculative type use it to perform an exact cast.
3063 ciKlass* spec_obj_type = obj_type->speculative_type();
3064 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3065 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3066 if (stopped()) { // Profile disagrees with this path.
3067 set_control(null_ctl); // Null is the only remaining possibility.
3068 return intcon(0);
3069 }
3070 if (cast_obj != NULL &&
3071 // A value that's sometimes null is not something we can optimize well
3072 !(cast_obj->is_ValueType() && null_ctl != top())) {
3073 not_null_obj = cast_obj;
3074 is_value = not_null_obj->is_ValueType();
3075 }
3076 }
3077 }
3078 }
3079
3080 // Load the object's klass
3081 Node* obj_klass = NULL;
3082 if (is_value) {
3083 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->is_valuetype()->value_klass()));
3084 } else {
3085 obj_klass = load_object_klass(not_null_obj);
3086 }
3087
3088 // Generate the subtype check
3089 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3090
3091 // Plug in the success path to the general merge in slot 1.
3092 region->init_req(_obj_path, control());
3093 phi ->init_req(_obj_path, intcon(1));
3094
3095 // Plug in the failing path to the general merge in slot 2.
3096 region->init_req(_fail_path, not_subtype_ctrl);
3097 phi ->init_req(_fail_path, intcon(0));
3098
3099 // Return final merged results
3100 set_control( _gvn.transform(region) );
3101 record_for_igvn(region);
3102
3103 // If we know the type check always succeeds then we don't use the
3104 // profiling data at this bytecode. Don't lose it, feed it to the
3105 // type system as a speculative type.
3106 if (safe_for_replace && !is_value) {
3107 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3108 replace_in_map(obj, casted_obj);
3109 }
3110
3111 return _gvn.transform(phi);
3112 }
3113
3114 //-------------------------------gen_checkcast---------------------------------
3115 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3116 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3117 // uncommon-trap paths work. Adjust stack after this call.
3118 // If failure_control is supplied and not null, it is filled in with
3119 // the control edge for the cast failure. Otherwise, an appropriate
3120 // uncommon trap or exception is thrown.
3121 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3122 Node* *failure_control) {
3123 kill_dead_locals(); // Benefit all the uncommon traps
3124 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr();
3125 const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass());
3126
3127 bool is_value = obj->is_ValueType();
3128
3129 // Fast cutout: Check the case that the cast is vacuously true.
3130 // This detects the common cases where the test will short-circuit
3131 // away completely. We do this before we perform the null check,
3132 // because if the test is going to turn into zero code, we don't
3133 // want a residual null check left around. (Causes a slowdown,
3134 // for example, in some objArray manipulations, such as a[i]=a[j].)
3135 if (tk->singleton()) {
3136 ciKlass* klass = NULL;
3137 if (is_value) {
3138 klass = _gvn.type(obj)->is_valuetype()->value_klass();
3139 } else {
3140 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3141 if (objtp != NULL) {
3142 klass = objtp->klass();
3143 }
3144 }
3145 if (klass != NULL) {
3146 switch (C->static_subtype_check(tk->klass(), klass)) {
3147 case Compile::SSC_always_true:
3148 // If we know the type check always succeed then we don't use
3149 // the profiling data at this bytecode. Don't lose it, feed it
3150 // to the type system as a speculative type.
3151 if (!is_value) {
3152 obj = record_profiled_receiver_for_speculation(obj);
3153 if (toop->is_valuetypeptr()) {
3154 obj = ValueTypeNode::make_from_oop(this, obj, toop->value_klass(), /* buffer_check */ false, /* null2default */ false);
3155 }
3156 }
3157 return obj;
3158 case Compile::SSC_always_false:
3159 // It needs a null check because a null will *pass* the cast check.
3160 // A non-null value will always produce an exception.
3161 if (is_value) {
3162 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(klass)));
3163 return top();
3164 } else {
3165 return null_assert(obj);
3166 }
3167 }
3168 }
3169 }
3170
3171 ciProfileData* data = NULL;
3172 bool safe_for_replace = false;
3173 if (failure_control == NULL) { // use MDO in regular case only
3174 assert(java_bc() == Bytecodes::_aastore ||
3175 java_bc() == Bytecodes::_checkcast,
3176 "interpreter profiles type checks only for these BCs");
3177 data = method()->method_data()->bci_to_data(bci());
3178 safe_for_replace = true;
3179 }
3180
3181 // Make the merge point
3182 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3183 RegionNode* region = new RegionNode(PATH_LIMIT);
3184 Node* phi = new PhiNode(region, toop);
3185 C->set_has_split_ifs(true); // Has chance for split-if optimization
3186
3187 // Use null-cast information if it is available
3188 bool speculative_not_null = false;
3189 bool never_see_null = ((failure_control == NULL) // regular case only
3190 && seems_never_null(obj, data, speculative_not_null));
3191
3192 // Null check; get casted pointer; set region slot 3
3193 Node* null_ctl = top();
3194 Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3195
3196 // If not_null_obj is dead, only null-path is taken
3197 if (stopped()) { // Doing instance-of on a NULL?
3198 set_control(null_ctl);
3199 return null();
3200 }
3201 region->init_req(_null_path, null_ctl);
3202 phi ->init_req(_null_path, null()); // Set null path value
3203 if (null_ctl == top()) {
3204 // Do this eagerly, so that pattern matches like is_diamond_phi
3205 // will work even during parsing.
3206 assert(_null_path == PATH_LIMIT-1, "delete last");
3207 region->del_req(_null_path);
3208 phi ->del_req(_null_path);
3209 }
3210
3211 Node* cast_obj = NULL;
3212 if (!is_value && tk->klass_is_exact()) {
3213 // The following optimization tries to statically cast the speculative type of the object
3214 // (for example obtained during profiling) to the type of the superklass and then do a
3215 // dynamic check that the type of the object is what we expect. To work correctly
3216 // for checkcast and aastore the type of superklass should be exact.
3217 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3218 // We may not have profiling here or it may not help us. If we have
3219 // a speculative type use it to perform an exact cast.
3220 ciKlass* spec_obj_type = obj_type->speculative_type();
3221 if (spec_obj_type != NULL || data != NULL) {
3222 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3223 if (cast_obj != NULL && cast_obj->is_ValueType()) {
3224 if (null_ctl != top()) {
3225 cast_obj = NULL; // A value that's sometimes null is not something we can optimize well
3226 } else {
3227 return cast_obj;
3228 }
3229 }
3230 if (cast_obj != NULL) {
3231 if (failure_control != NULL) // failure is now impossible
3232 (*failure_control) = top();
3233 // adjust the type of the phi to the exact klass:
3234 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3235 }
3236 }
3237 }
3238
3239 if (cast_obj == NULL) {
3240 // Load the object's klass
3241 Node* obj_klass = NULL;
3242 if (is_value) {
3243 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->is_valuetype()->value_klass()));
3244 } else {
3245 obj_klass = load_object_klass(not_null_obj);
3246 }
3247
3248 // Generate the subtype check
3249 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3250
3251 if (is_value) {
3252 ValueTypeNode* vt = not_null_obj->as_ValueType();
3253 vt = vt->allocate(this, true)->as_ValueType();
3254 not_null_obj = ValueTypePtrNode::make_from_value_type(_gvn, vt);
3255 }
3256
3257 // Plug in success path into the merge
3258 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3259 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3260 if (failure_control == NULL) {
3261 if (not_subtype_ctrl != top()) { // If failure is possible
3262 PreserveJVMState pjvms(this);
3263 set_control(not_subtype_ctrl);
3264 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3265 }
3266 } else {
3267 (*failure_control) = not_subtype_ctrl;
3268 }
3269 }
3270
3271 region->init_req(_obj_path, control());
3272 phi ->init_req(_obj_path, cast_obj);
3273
3274 // A merge of NULL or Casted-NotNull obj
3275 Node* res = _gvn.transform(phi);
3276
3277 // Note I do NOT always 'replace_in_map(obj,result)' here.
3278 // if( tk->klass()->can_be_primary_super() )
3279 // This means that if I successfully store an Object into an array-of-String
3280 // I 'forget' that the Object is really now known to be a String. I have to
3281 // do this because we don't have true union types for interfaces - if I store
3282 // a Baz into an array-of-Interface and then tell the optimizer it's an
3283 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3284 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3285 // replace_in_map( obj, res );
3286
3287 // Return final merged results
3288 set_control( _gvn.transform(region) );
3289 record_for_igvn(region);
3290
3291 if (!is_value) {
3292 res = record_profiled_receiver_for_speculation(res);
3293 if (toop->is_valuetypeptr()) {
3294 res = ValueTypeNode::make_from_oop(this, res, toop->value_klass(), /* buffer_check */ false, /* null2default */ false);
3295 }
3296 }
3297 return res;
3298 }
3299
3300 Node* GraphKit::is_always_locked(Node* obj) {
3301 Node* mark_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3302 Node* mark = make_load(NULL, mark_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3303 Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern);
3304 return _gvn.transform(new AndXNode(mark, value_mask));
3305 }
3306
3307 // Deoptimize if 'obj' is a value type
3308 void GraphKit::gen_value_type_guard(Node* obj, int nargs) {
3309 assert(EnableValhalla, "should only be used if value types are enabled");
3310 Node* bol = NULL;
3311 if (obj->is_ValueTypeBase()) {
3312 bol = intcon(0);
3313 } else {
3314 Node* is_value = is_always_locked(obj);
3315 Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern);
3316 Node* cmp = _gvn.transform(new CmpXNode(is_value, value_mask));
3317 bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3318 }
3319 { BuildCutout unless(this, bol, PROB_MAX);
3320 inc_sp(nargs);
3321 uncommon_trap(Deoptimization::Reason_class_check,
3322 Deoptimization::Action_none);
3323 }
3324 }
3325
3326 // Deoptimize if 'ary' is flattened or if 'obj' is null and 'ary' is a value type array
3327 void GraphKit::gen_value_type_array_guard(Node* ary, Node* obj, Node* elem_klass) {
3328 assert(EnableValhalla, "should only be used if value types are enabled");
3329 if (elem_klass == NULL) {
3330 // Load array element klass
3331 Node* kls = load_object_klass(ary);
3332 Node* k_adr = basic_plus_adr(kls, in_bytes(ArrayKlass::element_klass_offset()));
3333 elem_klass = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
3334 }
3335 // Check if element is a value type
3336 Node* flags_addr = basic_plus_adr(elem_klass, in_bytes(Klass::access_flags_offset()));
3337 Node* flags = make_load(NULL, flags_addr, TypeInt::INT, T_INT, MemNode::unordered);
3338 Node* is_value_elem = _gvn.transform(new AndINode(flags, intcon(JVM_ACC_VALUE)));
3339
3340 const Type* objtype = _gvn.type(obj);
3341 if (objtype == TypePtr::NULL_PTR) {
3342 // Object is always null, check if array is a value type array
3343 Node* cmp = _gvn.transform(new CmpINode(is_value_elem, intcon(0)));
3344 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3345 { BuildCutout unless(this, bol, PROB_MAX);
3346 // TODO just deoptimize for now if we store null to a value type array
3347 uncommon_trap(Deoptimization::Reason_array_check,
3348 Deoptimization::Action_none);
3349 }
3350 } else {
3351 // Check if array is flattened or if we are storing null to a value type array
3352 // TODO can we merge these checks?
3353 gen_flattened_array_guard(ary);
3354 if (objtype->meet(TypePtr::NULL_PTR) == objtype) {
3355 // Check if (is_value_elem && obj_is_null) <=> (!is_value_elem | !obj_is_null == 0)
3356 // TODO what if we later figure out that obj is never null?
3357 Node* not_value = _gvn.transform(new XorINode(is_value_elem, intcon(JVM_ACC_VALUE)));
3358 not_value = _gvn.transform(new ConvI2LNode(not_value));
3359 Node* not_null = _gvn.transform(new CastP2XNode(NULL, obj));
3360 Node* both = _gvn.transform(new OrLNode(not_null, not_value));
3361 Node* cmp = _gvn.transform(new CmpLNode(both, longcon(0)));
3362 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3363 { BuildCutout unless(this, bol, PROB_MAX);
3364 // TODO just deoptimize for now if we store null to a value type array
3365 uncommon_trap(Deoptimization::Reason_array_check,
3366 Deoptimization::Action_none);
3367 }
3368 }
3369 }
3370 }
3371
3372 // Deoptimize if 'ary' is a flattened value type array
3373 void GraphKit::gen_flattened_array_guard(Node* ary, int nargs) {
3374 assert(EnableValhalla, "should only be used if value types are enabled");
3375 if (ValueArrayFlatten) {
3376 // Cannot statically determine if array is flattened, emit runtime check
3377 Node* kls = load_object_klass(ary);
3378 Node* lhp = basic_plus_adr(kls, in_bytes(Klass::layout_helper_offset()));
3379 Node* layout_val = make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3380 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3381 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(Klass::_lh_array_tag_vt_value)));
3382 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3383
3384 { BuildCutout unless(this, bol, PROB_MAX);
3385 // TODO just deoptimize for now if value type array is flattened
3386 inc_sp(nargs);
3387 uncommon_trap(Deoptimization::Reason_array_check,
3388 Deoptimization::Action_none);
3389 }
3390 }
3391 }
3392
3393 //------------------------------next_monitor-----------------------------------
3394 // What number should be given to the next monitor?
3395 int GraphKit::next_monitor() {
3396 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3397 int next = current + C->sync_stack_slots();
3398 // Keep the toplevel high water mark current:
3399 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3400 return current;
3401 }
3402
3403 //------------------------------insert_mem_bar---------------------------------
3404 // Memory barrier to avoid floating things around
3405 // The membar serves as a pinch point between both control and all memory slices.
3406 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3407 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3408 mb->init_req(TypeFunc::Control, control());
3409 mb->init_req(TypeFunc::Memory, reset_memory());
3410 Node* membar = _gvn.transform(mb);
3411 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3412 set_all_memory_call(membar);
3413 return membar;
3414 }
3415
3416 //-------------------------insert_mem_bar_volatile----------------------------
3417 // Memory barrier to avoid floating things around
3418 // The membar serves as a pinch point between both control and memory(alias_idx).
3419 // If you want to make a pinch point on all memory slices, do not use this
3420 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3421 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3422 // When Parse::do_put_xxx updates a volatile field, it appends a series
3423 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3424 // The first membar is on the same memory slice as the field store opcode.
3425 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3426 // All the other membars (for other volatile slices, including AliasIdxBot,
3427 // which stands for all unknown volatile slices) are control-dependent
3428 // on the first membar. This prevents later volatile loads or stores
3429 // from sliding up past the just-emitted store.
3430
3431 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3432 mb->set_req(TypeFunc::Control,control());
3433 if (alias_idx == Compile::AliasIdxBot) {
3434 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3435 } else {
3436 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3437 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3438 }
3439 Node* membar = _gvn.transform(mb);
3440 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3441 if (alias_idx == Compile::AliasIdxBot) {
3442 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3443 } else {
3444 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3445 }
3446 return membar;
3447 }
3448
3449 //------------------------------shared_lock------------------------------------
3450 // Emit locking code.
3451 FastLockNode* GraphKit::shared_lock(Node* obj) {
3452 // bci is either a monitorenter bc or InvocationEntryBci
3453 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3454 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3455
3456 if( !GenerateSynchronizationCode )
3457 return NULL; // Not locking things?
3458
3459 // We cannot lock on a value type
3460 const TypeOopPtr* objptr = _gvn.type(obj)->make_oopptr();
3461 if (objptr->can_be_value_type()) {
3462 gen_value_type_guard(obj, 1);
3463 }
3464
3465 if (stopped()) // Dead monitor?
3466 return NULL;
3467
3468 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3469
3470 // Box the stack location
3471 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3472 Node* mem = reset_memory();
3473
3474 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3475 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3476 // Create the counters for this fast lock.
3477 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3478 }
3479
3480 // Create the rtm counters for this fast lock if needed.
3481 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3482
3483 // Add monitor to debug info for the slow path. If we block inside the
3484 // slow path and de-opt, we need the monitor hanging around
3485 map()->push_monitor( flock );
3486
3487 const TypeFunc *tf = LockNode::lock_type();
3488 LockNode *lock = new LockNode(C, tf);
3489
3490 lock->init_req( TypeFunc::Control, control() );
3491 lock->init_req( TypeFunc::Memory , mem );
3492 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3493 lock->init_req( TypeFunc::FramePtr, frameptr() );
3494 lock->init_req( TypeFunc::ReturnAdr, top() );
3495
3496 lock->init_req(TypeFunc::Parms + 0, obj);
3497 lock->init_req(TypeFunc::Parms + 1, box);
3498 lock->init_req(TypeFunc::Parms + 2, flock);
3499 add_safepoint_edges(lock);
3500
3501 lock = _gvn.transform( lock )->as_Lock();
3502
3503 // lock has no side-effects, sets few values
3504 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3505
3506 insert_mem_bar(Op_MemBarAcquireLock);
3507
3508 // Add this to the worklist so that the lock can be eliminated
3509 record_for_igvn(lock);
3510
3511 #ifndef PRODUCT
3512 if (PrintLockStatistics) {
3513 // Update the counter for this lock. Don't bother using an atomic
3514 // operation since we don't require absolute accuracy.
3515 lock->create_lock_counter(map()->jvms());
3516 increment_counter(lock->counter()->addr());
3517 }
3518 #endif
3519
3520 return flock;
3521 }
3522
3523
3524 //------------------------------shared_unlock----------------------------------
3525 // Emit unlocking code.
3526 void GraphKit::shared_unlock(Node* box, Node* obj) {
3527 // bci is either a monitorenter bc or InvocationEntryBci
3528 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3529 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3530
3531 if( !GenerateSynchronizationCode )
3532 return;
3533 if (stopped()) { // Dead monitor?
3534 map()->pop_monitor(); // Kill monitor from debug info
3535 return;
3536 }
3537 assert(!obj->is_ValueTypeBase(), "should not unlock on value type");
3538
3539 // Memory barrier to avoid floating things down past the locked region
3540 insert_mem_bar(Op_MemBarReleaseLock);
3541
3542 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3543 UnlockNode *unlock = new UnlockNode(C, tf);
3544 #ifdef ASSERT
3545 unlock->set_dbg_jvms(sync_jvms());
3546 #endif
3547 uint raw_idx = Compile::AliasIdxRaw;
3548 unlock->init_req( TypeFunc::Control, control() );
3549 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3550 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3551 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3552 unlock->init_req( TypeFunc::ReturnAdr, top() );
3553
3554 unlock->init_req(TypeFunc::Parms + 0, obj);
3555 unlock->init_req(TypeFunc::Parms + 1, box);
3556 unlock = _gvn.transform(unlock)->as_Unlock();
3557
3558 Node* mem = reset_memory();
3559
3560 // unlock has no side-effects, sets few values
3561 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3562
3563 // Kill monitor from debug info
3564 map()->pop_monitor( );
3565 }
3566
3567 //-------------------------------get_layout_helper-----------------------------
3568 // If the given klass is a constant or known to be an array,
3569 // fetch the constant layout helper value into constant_value
3570 // and return (Node*)NULL. Otherwise, load the non-constant
3571 // layout helper value, and return the node which represents it.
3572 // This two-faced routine is useful because allocation sites
3573 // almost always feature constant types.
3574 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3575 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3576 if (!StressReflectiveCode && inst_klass != NULL) {
3577 ciKlass* klass = inst_klass->klass();
3578 assert(klass != NULL, "klass should not be NULL");
3579 bool xklass = inst_klass->klass_is_exact();
3580 bool can_be_value_array = false;
3581 if (klass->is_array_klass() && EnableValhalla && ValueArrayFlatten) {
3582 ciKlass* elem = klass->as_array_klass()->element_klass();
3583 can_be_value_array = elem != NULL && (elem->is_java_lang_Object() || elem->is_interface());
3584 }
3585 if (xklass || (klass->is_array_klass() && !can_be_value_array)) {
3586 jint lhelper = klass->layout_helper();
3587 if (lhelper != Klass::_lh_neutral_value) {
3588 constant_value = lhelper;
3589 return (Node*) NULL;
3590 }
3591 }
3592 }
3593 constant_value = Klass::_lh_neutral_value; // put in a known value
3594 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3595 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3596 }
3597
3598 // We just put in an allocate/initialize with a big raw-memory effect.
3599 // Hook selected additional alias categories on the initialization.
3600 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3601 MergeMemNode* init_in_merge,
3602 Node* init_out_raw) {
3603 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3604 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3605
3606 Node* prevmem = kit.memory(alias_idx);
3607 init_in_merge->set_memory_at(alias_idx, prevmem);
3608 kit.set_memory(init_out_raw, alias_idx);
3609 }
3610
3611 //---------------------------set_output_for_allocation-------------------------
3612 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3613 const TypeOopPtr* oop_type,
3614 bool deoptimize_on_exception) {
3615 int rawidx = Compile::AliasIdxRaw;
3616 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3617 add_safepoint_edges(alloc);
3618 Node* allocx = _gvn.transform(alloc);
3619 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3620 // create memory projection for i_o
3621 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3622 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3623
3624 // create a memory projection as for the normal control path
3625 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3626 set_memory(malloc, rawidx);
3627
3628 // a normal slow-call doesn't change i_o, but an allocation does
3629 // we create a separate i_o projection for the normal control path
3630 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3631 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3632
3633 // put in an initialization barrier
3634 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3635 rawoop)->as_Initialize();
3636 assert(alloc->initialization() == init, "2-way macro link must work");
3637 assert(init ->allocation() == alloc, "2-way macro link must work");
3638 {
3639 // Extract memory strands which may participate in the new object's
3640 // initialization, and source them from the new InitializeNode.
3641 // This will allow us to observe initializations when they occur,
3642 // and link them properly (as a group) to the InitializeNode.
3643 assert(init->in(InitializeNode::Memory) == malloc, "");
3644 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3645 init->set_req(InitializeNode::Memory, minit_in);
3646 record_for_igvn(minit_in); // fold it up later, if possible
3647 Node* minit_out = memory(rawidx);
3648 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3649 if (oop_type->isa_aryptr()) {
3650 const TypeAryPtr* arytype = oop_type->is_aryptr();
3651 if (arytype->klass()->is_value_array_klass()) {
3652 ciValueArrayKlass* vak = arytype->klass()->as_value_array_klass();
3653 ciValueKlass* vk = vak->element_klass()->as_value_klass();
3654 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
3655 ciField* field = vk->nonstatic_field_at(i);
3656 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3657 continue; // do not bother to track really large numbers of fields
3658 int off_in_vt = field->offset() - vk->first_field_offset();
3659 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
3660 int fieldidx = C->get_alias_index(adr_type);
3661 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3662 }
3663 } else {
3664 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3665 int elemidx = C->get_alias_index(telemref);
3666 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3667 }
3668 } else if (oop_type->isa_instptr()) {
3669 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3670 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3671 ciField* field = ik->nonstatic_field_at(i);
3672 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3673 continue; // do not bother to track really large numbers of fields
3674 // Find (or create) the alias category for this field:
3675 int fieldidx = C->alias_type(field)->index();
3676 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3677 }
3678 }
3679 }
3680
3681 // Cast raw oop to the real thing...
3682 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3683 javaoop = _gvn.transform(javaoop);
3684 C->set_recent_alloc(control(), javaoop);
3685 assert(just_allocated_object(control()) == javaoop, "just allocated");
3686
3687 #ifdef ASSERT
3688 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3689 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3690 "Ideal_allocation works");
3691 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3692 "Ideal_allocation works");
3693 if (alloc->is_AllocateArray()) {
3694 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3695 "Ideal_allocation works");
3696 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3697 "Ideal_allocation works");
3698 } else {
3699 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3700 }
3701 }
3702 #endif //ASSERT
3703
3704 return javaoop;
3705 }
3706
3707 //---------------------------new_instance--------------------------------------
3708 // This routine takes a klass_node which may be constant (for a static type)
3709 // or may be non-constant (for reflective code). It will work equally well
3710 // for either, and the graph will fold nicely if the optimizer later reduces
3711 // the type to a constant.
3712 // The optional arguments are for specialized use by intrinsics:
3713 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3714 // - If 'return_size_val', report the the total object size to the caller.
3715 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3716 Node* GraphKit::new_instance(Node* klass_node,
3717 Node* extra_slow_test,
3718 Node* *return_size_val,
3719 bool deoptimize_on_exception,
3720 ValueTypeBaseNode* value_node) {
3721 // Compute size in doublewords
3722 // The size is always an integral number of doublewords, represented
3723 // as a positive bytewise size stored in the klass's layout_helper.
3724 // The layout_helper also encodes (in a low bit) the need for a slow path.
3725 jint layout_con = Klass::_lh_neutral_value;
3726 Node* layout_val = get_layout_helper(klass_node, layout_con);
3727 bool layout_is_con = (layout_val == NULL);
3728
3729 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3730 // Generate the initial go-slow test. It's either ALWAYS (return a
3731 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3732 // case) a computed value derived from the layout_helper.
3733 Node* initial_slow_test = NULL;
3734 if (layout_is_con) {
3735 assert(!StressReflectiveCode, "stress mode does not use these paths");
3736 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3737 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3738 } else { // reflective case
3739 // This reflective path is used by Unsafe.allocateInstance.
3740 // (It may be stress-tested by specifying StressReflectiveCode.)
3741 // Basically, we want to get into the VM is there's an illegal argument.
3742 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3743 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3744 if (extra_slow_test != intcon(0)) {
3745 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3746 }
3747 // (Macro-expander will further convert this to a Bool, if necessary.)
3748 }
3749
3750 // Find the size in bytes. This is easy; it's the layout_helper.
3751 // The size value must be valid even if the slow path is taken.
3752 Node* size = NULL;
3753 if (layout_is_con) {
3754 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3755 } else { // reflective case
3756 // This reflective path is used by clone and Unsafe.allocateInstance.
3757 size = ConvI2X(layout_val);
3758
3759 // Clear the low bits to extract layout_helper_size_in_bytes:
3760 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3761 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3762 size = _gvn.transform( new AndXNode(size, mask) );
3763 }
3764 if (return_size_val != NULL) {
3765 (*return_size_val) = size;
3766 }
3767
3768 // This is a precise notnull oop of the klass.
3769 // (Actually, it need not be precise if this is a reflective allocation.)
3770 // It's what we cast the result to.
3771 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3772 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3773 const TypeOopPtr* oop_type = tklass->as_instance_type();
3774
3775 // Now generate allocation code
3776
3777 // The entire memory state is needed for slow path of the allocation
3778 // since GC and deoptimization can happen.
3779 Node *mem = reset_memory();
3780 set_all_memory(mem); // Create new memory state
3781
3782 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3783 control(), mem, i_o(),
3784 size, klass_node,
3785 initial_slow_test, value_node);
3786
3787 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3788 }
3789
3790 //-------------------------------new_array-------------------------------------
3791 // helper for newarray and anewarray
3792 // The 'length' parameter is (obviously) the length of the array.
3793 // See comments on new_instance for the meaning of the other arguments.
3794 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3795 Node* length, // number of array elements
3796 int nargs, // number of arguments to push back for uncommon trap
3797 Node* *return_size_val,
3798 bool deoptimize_on_exception) {
3799 jint layout_con = Klass::_lh_neutral_value;
3800 Node* layout_val = get_layout_helper(klass_node, layout_con);
3801 bool layout_is_con = (layout_val == NULL);
3802
3803 if (!layout_is_con && !StressReflectiveCode &&
3804 !too_many_traps(Deoptimization::Reason_class_check)) {
3805 // This is a reflective array creation site.
3806 // Optimistically assume that it is a subtype of Object[],
3807 // so that we can fold up all the address arithmetic.
3808 layout_con = Klass::array_layout_helper(T_OBJECT);
3809 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3810 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3811 { BuildCutout unless(this, bol_lh, PROB_MAX);
3812 inc_sp(nargs);
3813 uncommon_trap(Deoptimization::Reason_class_check,
3814 Deoptimization::Action_maybe_recompile);
3815 }
3816 layout_val = NULL;
3817 layout_is_con = true;
3818 }
3819
3820 // Generate the initial go-slow test. Make sure we do not overflow
3821 // if length is huge (near 2Gig) or negative! We do not need
3822 // exact double-words here, just a close approximation of needed
3823 // double-words. We can't add any offset or rounding bits, lest we
3824 // take a size -1 of bytes and make it positive. Use an unsigned
3825 // compare, so negative sizes look hugely positive.
3826 int fast_size_limit = FastAllocateSizeLimit;
3827 if (layout_is_con) {
3828 assert(!StressReflectiveCode, "stress mode does not use these paths");
3829 // Increase the size limit if we have exact knowledge of array type.
3830 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3831 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
3832 }
3833
3834 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3835 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3836
3837 // --- Size Computation ---
3838 // array_size = round_to_heap(array_header + (length << elem_shift));
3839 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3840 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3841 // The rounding mask is strength-reduced, if possible.
3842 int round_mask = MinObjAlignmentInBytes - 1;
3843 Node* header_size = NULL;
3844 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3845 // (T_BYTE has the weakest alignment and size restrictions...)
3846 if (layout_is_con) {
3847 int hsize = Klass::layout_helper_header_size(layout_con);
3848 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3849 BasicType etype = Klass::layout_helper_element_type(layout_con);
3850 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con);
3851 if ((round_mask & ~right_n_bits(eshift)) == 0)
3852 round_mask = 0; // strength-reduce it if it goes away completely
3853 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3854 assert(header_size_min <= hsize, "generic minimum is smallest");
3855 header_size_min = hsize;
3856 header_size = intcon(hsize + round_mask);
3857 } else {
3858 Node* hss = intcon(Klass::_lh_header_size_shift);
3859 Node* hsm = intcon(Klass::_lh_header_size_mask);
3860 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3861 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3862 Node* mask = intcon(round_mask);
3863 header_size = _gvn.transform( new AddINode(hsize, mask) );
3864 }
3865
3866 Node* elem_shift = NULL;
3867 if (layout_is_con) {
3868 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3869 if (eshift != 0)
3870 elem_shift = intcon(eshift);
3871 } else {
3872 // There is no need to mask or shift this value.
3873 // The semantics of LShiftINode include an implicit mask to 0x1F.
3874 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3875 elem_shift = layout_val;
3876 }
3877
3878 // Transition to native address size for all offset calculations:
3879 Node* lengthx = ConvI2X(length);
3880 Node* headerx = ConvI2X(header_size);
3881 #ifdef _LP64
3882 { const TypeInt* tilen = _gvn.find_int_type(length);
3883 if (tilen != NULL && tilen->_lo < 0) {
3884 // Add a manual constraint to a positive range. Cf. array_element_address.
3885 jint size_max = fast_size_limit;
3886 if (size_max > tilen->_hi) size_max = tilen->_hi;
3887 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3888
3889 // Only do a narrow I2L conversion if the range check passed.
3890 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3891 _gvn.transform(iff);
3892 RegionNode* region = new RegionNode(3);
3893 _gvn.set_type(region, Type::CONTROL);
3894 lengthx = new PhiNode(region, TypeLong::LONG);
3895 _gvn.set_type(lengthx, TypeLong::LONG);
3896
3897 // Range check passed. Use ConvI2L node with narrow type.
3898 Node* passed = IfFalse(iff);
3899 region->init_req(1, passed);
3900 // Make I2L conversion control dependent to prevent it from
3901 // floating above the range check during loop optimizations.
3902 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3903
3904 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3905 region->init_req(2, IfTrue(iff));
3906 lengthx->init_req(2, ConvI2X(length));
3907
3908 set_control(region);
3909 record_for_igvn(region);
3910 record_for_igvn(lengthx);
3911 }
3912 }
3913 #endif
3914
3915 // Combine header size (plus rounding) and body size. Then round down.
3916 // This computation cannot overflow, because it is used only in two
3917 // places, one where the length is sharply limited, and the other
3918 // after a successful allocation.
3919 Node* abody = lengthx;
3920 if (elem_shift != NULL)
3921 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3922 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3923 if (round_mask != 0) {
3924 Node* mask = MakeConX(~round_mask);
3925 size = _gvn.transform( new AndXNode(size, mask) );
3926 }
3927 // else if round_mask == 0, the size computation is self-rounding
3928
3929 if (return_size_val != NULL) {
3930 // This is the size
3931 (*return_size_val) = size;
3932 }
3933
3934 // Now generate allocation code
3935
3936 // The entire memory state is needed for slow path of the allocation
3937 // since GC and deoptimization can happen.
3938 Node *mem = reset_memory();
3939 set_all_memory(mem); // Create new memory state
3940
3941 if (initial_slow_test->is_Bool()) {
3942 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3943 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3944 }
3945
3946 // Create the AllocateArrayNode and its result projections
3947 AllocateArrayNode* alloc
3948 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3949 control(), mem, i_o(),
3950 size, klass_node,
3951 initial_slow_test,
3952 length);
3953
3954 // Cast to correct type. Note that the klass_node may be constant or not,
3955 // and in the latter case the actual array type will be inexact also.
3956 // (This happens via a non-constant argument to inline_native_newArray.)
3957 // In any case, the value of klass_node provides the desired array type.
3958 const TypeInt* length_type = _gvn.find_int_type(length);
3959 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3960 if (ary_type->isa_aryptr() && length_type != NULL) {
3961 // Try to get a better type than POS for the size
3962 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3963 }
3964
3965 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3966
3967 // Cast length on remaining path to be as narrow as possible
3968 if (map()->find_edge(length) >= 0) {
3969 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3970 if (ccast != length) {
3971 _gvn.set_type_bottom(ccast);
3972 record_for_igvn(ccast);
3973 replace_in_map(length, ccast);
3974 }
3975 }
3976
3977 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
3978 ciKlass* elem_klass = ary_ptr != NULL ? ary_ptr->klass()->as_array_klass()->element_klass() : NULL;
3979 if (elem_klass != NULL && elem_klass->is_valuetype()) {
3980 ciValueKlass* vk = elem_klass->as_value_klass();
3981 if (!vk->flatten_array()) {
3982 // Non-flattened value type arrays need to be initialized with default value type oops
3983 initialize_value_type_array(javaoop, length, elem_klass->as_value_klass(), nargs);
3984 // TODO re-enable once JDK-8189802 is fixed
3985 // InitializeNode* init = alloc->initialization();
3986 //init->set_complete_with_arraycopy();
3987 }
3988 } else if (EnableValhalla && (!layout_con || elem_klass == NULL || (elem_klass->is_java_lang_Object() && !ary_type->klass_is_exact()))) {
3989 InitializeNode* init = alloc->initialization();
3990 init->set_unknown_value();
3991 }
3992
3993 return javaoop;
3994 }
3995
3996 void GraphKit::initialize_value_type_array(Node* array, Node* length, ciValueKlass* vk, int nargs) {
3997 // Check for zero length
3998 Node* null_ctl = top();
3999 null_check_common(length, T_INT, false, &null_ctl, false);
4000 if (stopped()) {
4001 set_control(null_ctl); // Always zero
4002 return;
4003 }
4004
4005 RegionNode* res_ctl = new RegionNode(3);
4006 gvn().set_type(res_ctl, Type::CONTROL);
4007 record_for_igvn(res_ctl);
4008
4009 // Length is zero: don't execute initialization loop
4010 res_ctl->init_req(1, null_ctl);
4011 PhiNode* res_io = PhiNode::make(res_ctl, i_o(), Type::ABIO);
4012 PhiNode* res_mem = PhiNode::make(res_ctl, merged_memory(), Type::MEMORY, TypePtr::BOTTOM);
4013 gvn().set_type(res_io, Type::ABIO);
4014 gvn().set_type(res_mem, Type::MEMORY);
4015 record_for_igvn(res_io);
4016 record_for_igvn(res_mem);
4017
4018 // Length is non-zero: execute a loop that initializes the array with the default value type
4019 Node* oop = ValueTypeNode::load_default_oop(gvn(), vk);
4020
4021 add_predicate(nargs);
4022 RegionNode* loop = new RegionNode(3);
4023 loop->init_req(1, control());
4024 PhiNode* index = PhiNode::make(loop, intcon(0), TypeInt::INT);
4025 PhiNode* mem = PhiNode::make(loop, reset_memory(), Type::MEMORY, TypePtr::BOTTOM);
4026
4027 gvn().set_type(loop, Type::CONTROL);
4028 gvn().set_type(index, TypeInt::INT);
4029 gvn().set_type(mem, Type::MEMORY);
4030 record_for_igvn(loop);
4031 record_for_igvn(index);
4032 record_for_igvn(mem);
4033
4034 // Loop body: initialize array element at 'index'
4035 set_control(loop);
4036 set_all_memory(mem);
4037 Node* adr = array_element_address(array, index, T_OBJECT);
4038 const TypeOopPtr* elemtype = TypeInstPtr::make(TypePtr::NotNull, vk);
4039 access_store_at(control(), array, adr, TypeAryPtr::OOPS, oop, elemtype, T_VALUETYPE, MemNode::release);
4040
4041 // Check if we need to execute another loop iteration
4042 length = SubI(length, intcon(1));
4043 IfNode* iff = create_and_map_if(control(), Bool(CmpI(index, length), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4044
4045 // Continue with next iteration
4046 loop->init_req(2, IfTrue(iff));
4047 index->init_req(2, AddI(index, intcon(1)));
4048 mem->init_req(2, merged_memory());
4049
4050 // Exit loop
4051 res_ctl->init_req(2, IfFalse(iff));
4052 res_io->set_req(2, i_o());
4053 res_mem->set_req(2, reset_memory());
4054
4055 // Set merged control, IO and memory
4056 set_control(res_ctl);
4057 set_i_o(res_io);
4058 set_all_memory(res_mem);
4059 }
4060
4061 // The following "Ideal_foo" functions are placed here because they recognize
4062 // the graph shapes created by the functions immediately above.
4063
4064 //---------------------------Ideal_allocation----------------------------------
4065 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4066 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
4067 if (ptr == NULL) { // reduce dumb test in callers
4068 return NULL;
4069 }
4070 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4071 ptr = ptr->in(1);
4072 if (ptr == NULL) return NULL;
4073 }
4074 // Return NULL for allocations with several casts:
4075 // j.l.reflect.Array.newInstance(jobject, jint)
4076 // Object.clone()
4077 // to keep more precise type from last cast.
4078 if (ptr->is_Proj()) {
4079 Node* allo = ptr->in(0);
4080 if (allo != NULL && allo->is_Allocate()) {
4081 return allo->as_Allocate();
4082 }
4083 }
4084 // Report failure to match.
4085 return NULL;
4086 }
4087
4088 // Fancy version which also strips off an offset (and reports it to caller).
4089 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
4090 intptr_t& offset) {
4091 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4092 if (base == NULL) return NULL;
4093 return Ideal_allocation(base, phase);
4094 }
4095
4096 // Trace Initialize <- Proj[Parm] <- Allocate
4097 AllocateNode* InitializeNode::allocation() {
4098 Node* rawoop = in(InitializeNode::RawAddress);
4099 if (rawoop->is_Proj()) {
4100 Node* alloc = rawoop->in(0);
4101 if (alloc->is_Allocate()) {
4102 return alloc->as_Allocate();
4103 }
4104 }
4105 return NULL;
4106 }
4107
4108 // Trace Allocate -> Proj[Parm] -> Initialize
4109 InitializeNode* AllocateNode::initialization() {
4110 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4111 if (rawoop == NULL) return NULL;
4112 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4113 Node* init = rawoop->fast_out(i);
4114 if (init->is_Initialize()) {
4115 assert(init->as_Initialize()->allocation() == this, "2-way link");
4116 return init->as_Initialize();
4117 }
4118 }
4119 return NULL;
4120 }
4121
4122 //----------------------------- loop predicates ---------------------------
4123
4124 //------------------------------add_predicate_impl----------------------------
4125 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
4126 // Too many traps seen?
4127 if (too_many_traps(reason)) {
4128 #ifdef ASSERT
4129 if (TraceLoopPredicate) {
4130 int tc = C->trap_count(reason);
4131 tty->print("too many traps=%s tcount=%d in ",
4132 Deoptimization::trap_reason_name(reason), tc);
4133 method()->print(); // which method has too many predicate traps
4134 tty->cr();
4135 }
4136 #endif
4137 // We cannot afford to take more traps here,
4138 // do not generate predicate.
4139 return;
4140 }
4141
4142 Node *cont = _gvn.intcon(1);
4143 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
4144 Node *bol = _gvn.transform(new Conv2BNode(opq));
4145 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4146 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
4147 C->add_predicate_opaq(opq);
4148 {
4149 PreserveJVMState pjvms(this);
4150 set_control(iffalse);
4151 inc_sp(nargs);
4152 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4153 }
4154 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
4155 set_control(iftrue);
4156 }
4157
4158 //------------------------------add_predicate---------------------------------
4159 void GraphKit::add_predicate(int nargs) {
4160 if (UseLoopPredicate) {
4161 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
4162 }
4163 if (UseProfiledLoopPredicate) {
4164 add_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
4165 }
4166 // loop's limit check predicate should be near the loop.
4167 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
4168 }
4169
4170 void GraphKit::sync_kit(IdealKit& ideal) {
4171 set_all_memory(ideal.merged_memory());
4172 set_i_o(ideal.i_o());
4173 set_control(ideal.ctrl());
4174 }
4175
4176 void GraphKit::final_sync(IdealKit& ideal) {
4177 // Final sync IdealKit and graphKit.
4178 sync_kit(ideal);
4179 }
4180
4181 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4182 Node* len = load_array_length(load_String_value(ctrl, str));
4183 Node* coder = load_String_coder(ctrl, str);
4184 // Divide length by 2 if coder is UTF16
4185 return _gvn.transform(new RShiftINode(len, coder));
4186 }
4187
4188 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4189 int value_offset = java_lang_String::value_offset_in_bytes();
4190 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4191 false, NULL, Type::Offset(0));
4192 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4193 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4194 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4195 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4196 Node* p = basic_plus_adr(str, str, value_offset);
4197 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4198 IN_HEAP | C2_CONTROL_DEPENDENT_LOAD);
4199 // String.value field is known to be @Stable.
4200 if (UseImplicitStableValues) {
4201 load = cast_array_to_stable(load, value_type);
4202 }
4203 return load;
4204 }
4205
4206 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4207 if (!CompactStrings) {
4208 return intcon(java_lang_String::CODER_UTF16);
4209 }
4210 int coder_offset = java_lang_String::coder_offset_in_bytes();
4211 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4212 false, NULL, Type::Offset(0));
4213 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4214 int coder_field_idx = C->get_alias_index(coder_field_type);
4215 return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4216 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4217 }
4218
4219 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4220 int value_offset = java_lang_String::value_offset_in_bytes();
4221 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4222 false, NULL, Type::Offset(0));
4223 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4224 access_store_at(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4225 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP);
4226 }
4227
4228 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4229 int coder_offset = java_lang_String::coder_offset_in_bytes();
4230 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4231 false, NULL, Type::Offset(0));
4232 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4233 int coder_field_idx = C->get_alias_index(coder_field_type);
4234 store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4235 value, T_BYTE, coder_field_idx, MemNode::unordered);
4236 }
4237
4238 // Capture src and dst memory state with a MergeMemNode
4239 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4240 if (src_type == dst_type) {
4241 // Types are equal, we don't need a MergeMemNode
4242 return memory(src_type);
4243 }
4244 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4245 record_for_igvn(merge); // fold it up later, if possible
4246 int src_idx = C->get_alias_index(src_type);
4247 int dst_idx = C->get_alias_index(dst_type);
4248 merge->set_memory_at(src_idx, memory(src_idx));
4249 merge->set_memory_at(dst_idx, memory(dst_idx));
4250 return merge;
4251 }
4252
4253 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4254 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4255 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4256 // If input and output memory types differ, capture both states to preserve
4257 // the dependency between preceding and subsequent loads/stores.
4258 // For example, the following program:
4259 // StoreB
4260 // compress_string
4261 // LoadB
4262 // has this memory graph (use->def):
4263 // LoadB -> compress_string -> CharMem
4264 // ... -> StoreB -> ByteMem
4265 // The intrinsic hides the dependency between LoadB and StoreB, causing
4266 // the load to read from memory not containing the result of the StoreB.
4267 // The correct memory graph should look like this:
4268 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4269 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4270 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4271 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4272 set_memory(res_mem, TypeAryPtr::BYTES);
4273 return str;
4274 }
4275
4276 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4277 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4278 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4279 // Capture src and dst memory (see comment in 'compress_string').
4280 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4281 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4282 set_memory(_gvn.transform(str), dst_type);
4283 }
4284
4285 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4286 /**
4287 * int i_char = start;
4288 * for (int i_byte = 0; i_byte < count; i_byte++) {
4289 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4290 * }
4291 */
4292 add_predicate();
4293 RegionNode* head = new RegionNode(3);
4294 head->init_req(1, control());
4295 gvn().set_type(head, Type::CONTROL);
4296 record_for_igvn(head);
4297
4298 Node* i_byte = new PhiNode(head, TypeInt::INT);
4299 i_byte->init_req(1, intcon(0));
4300 gvn().set_type(i_byte, TypeInt::INT);
4301 record_for_igvn(i_byte);
4302
4303 Node* i_char = new PhiNode(head, TypeInt::INT);
4304 i_char->init_req(1, start);
4305 gvn().set_type(i_char, TypeInt::INT);
4306 record_for_igvn(i_char);
4307
4308 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4309 gvn().set_type(mem, Type::MEMORY);
4310 record_for_igvn(mem);
4311 set_control(head);
4312 set_memory(mem, TypeAryPtr::BYTES);
4313 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4314 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4315 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4316 false, false, true /* mismatched */);
4317
4318 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4319 head->init_req(2, IfTrue(iff));
4320 mem->init_req(2, st);
4321 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4322 i_char->init_req(2, AddI(i_char, intcon(2)));
4323
4324 set_control(IfFalse(iff));
4325 set_memory(st, TypeAryPtr::BYTES);
4326 }
4327
4328 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4329 if (!field->is_constant()) {
4330 return NULL; // Field not marked as constant.
4331 }
4332 ciInstance* holder = NULL;
4333 if (!field->is_static()) {
4334 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4335 if (const_oop != NULL && const_oop->is_instance()) {
4336 holder = const_oop->as_instance();
4337 }
4338 }
4339 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4340 /*is_unsigned_load=*/false);
4341 if (con_type != NULL) {
4342 Node* con = makecon(con_type);
4343 if (field->layout_type() == T_VALUETYPE) {
4344 // Load value type from constant oop
4345 con = ValueTypeNode::make_from_oop(this, con, field->type()->as_value_klass());
4346 }
4347 return con;
4348 }
4349 return NULL;
4350 }
4351
4352 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4353 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4354 // assumption of CCP analysis.
4355 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4356 }