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/opaquenode.hpp"
42 #include "opto/parse.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/valuetypenode.hpp"
46 #include "runtime/deoptimization.hpp"
47 #include "runtime/sharedRuntime.hpp"
48
49 //----------------------------GraphKit-----------------------------------------
50 // Main utility constructor.
51 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
52 : Phase(Phase::Parser),
53 _env(C->env()),
54 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()),
55 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
56 {
57 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
58 _exceptions = jvms->map()->next_exception();
59 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
60 set_jvms(jvms);
61 #ifdef ASSERT
62 if (_gvn.is_IterGVN() != NULL) {
63 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
64 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
65 _worklist_size = _gvn.C->for_igvn()->size();
66 }
67 #endif
68 }
69
70 // Private constructor for parser.
71 GraphKit::GraphKit()
72 : Phase(Phase::Parser),
73 _env(C->env()),
74 _gvn(*C->initial_gvn()),
75 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
76 {
77 _exceptions = NULL;
78 set_map(NULL);
79 debug_only(_sp = -99);
80 debug_only(set_bci(-99));
81 }
82
83
84
85 //---------------------------clean_stack---------------------------------------
86 // Clear away rubbish from the stack area of the JVM state.
87 // This destroys any arguments that may be waiting on the stack.
88 void GraphKit::clean_stack(int from_sp) {
89 SafePointNode* map = this->map();
90 JVMState* jvms = this->jvms();
91 int stk_size = jvms->stk_size();
92 int stkoff = jvms->stkoff();
93 Node* top = this->top();
94 for (int i = from_sp; i < stk_size; i++) {
95 if (map->in(stkoff + i) != top) {
96 map->set_req(stkoff + i, top);
97 }
98 }
99 }
100
101
102 //--------------------------------sync_jvms-----------------------------------
103 // Make sure our current jvms agrees with our parse state.
104 JVMState* GraphKit::sync_jvms() const {
105 JVMState* jvms = this->jvms();
106 jvms->set_bci(bci()); // Record the new bci in the JVMState
107 jvms->set_sp(sp()); // Record the new sp in the JVMState
108 assert(jvms_in_sync(), "jvms is now in sync");
109 return jvms;
110 }
111
112 //--------------------------------sync_jvms_for_reexecute---------------------
113 // Make sure our current jvms agrees with our parse state. This version
114 // uses the reexecute_sp for reexecuting bytecodes.
115 JVMState* GraphKit::sync_jvms_for_reexecute() {
116 JVMState* jvms = this->jvms();
117 jvms->set_bci(bci()); // Record the new bci in the JVMState
118 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
119 return jvms;
120 }
121
122 #ifdef ASSERT
123 bool GraphKit::jvms_in_sync() const {
124 Parse* parse = is_Parse();
125 if (parse == NULL) {
126 if (bci() != jvms()->bci()) return false;
127 if (sp() != (int)jvms()->sp()) return false;
128 return true;
129 }
130 if (jvms()->method() != parse->method()) return false;
131 if (jvms()->bci() != parse->bci()) return false;
132 int jvms_sp = jvms()->sp();
133 if (jvms_sp != parse->sp()) return false;
134 int jvms_depth = jvms()->depth();
135 if (jvms_depth != parse->depth()) return false;
136 return true;
137 }
138
139 // Local helper checks for special internal merge points
140 // used to accumulate and merge exception states.
141 // They are marked by the region's in(0) edge being the map itself.
142 // Such merge points must never "escape" into the parser at large,
143 // until they have been handed to gvn.transform.
144 static bool is_hidden_merge(Node* reg) {
145 if (reg == NULL) return false;
146 if (reg->is_Phi()) {
147 reg = reg->in(0);
148 if (reg == NULL) return false;
149 }
150 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
151 }
152
153 void GraphKit::verify_map() const {
154 if (map() == NULL) return; // null map is OK
155 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
156 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
157 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
158 }
159
160 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
161 assert(ex_map->next_exception() == NULL, "not already part of a chain");
162 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
163 }
164 #endif
165
166 //---------------------------stop_and_kill_map---------------------------------
167 // Set _map to NULL, signalling a stop to further bytecode execution.
168 // First smash the current map's control to a constant, to mark it dead.
169 void GraphKit::stop_and_kill_map() {
170 SafePointNode* dead_map = stop();
171 if (dead_map != NULL) {
172 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
173 assert(dead_map->is_killed(), "must be so marked");
174 }
175 }
176
177
178 //--------------------------------stopped--------------------------------------
179 // Tell if _map is NULL, or control is top.
180 bool GraphKit::stopped() {
181 if (map() == NULL) return true;
182 else if (control() == top()) return true;
183 else return false;
184 }
185
186
187 //-----------------------------has_ex_handler----------------------------------
188 // Tell if this method or any caller method has exception handlers.
189 bool GraphKit::has_ex_handler() {
190 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
191 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
192 return true;
193 }
194 }
195 return false;
196 }
197
198 //------------------------------save_ex_oop------------------------------------
199 // Save an exception without blowing stack contents or other JVM state.
200 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
201 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
202 ex_map->add_req(ex_oop);
203 debug_only(verify_exception_state(ex_map));
204 }
205
206 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
207 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
208 Node* ex_oop = ex_map->in(ex_map->req()-1);
209 if (clear_it) ex_map->del_req(ex_map->req()-1);
210 return ex_oop;
211 }
212
213 //-----------------------------saved_ex_oop------------------------------------
214 // Recover a saved exception from its map.
215 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
216 return common_saved_ex_oop(ex_map, false);
217 }
218
219 //--------------------------clear_saved_ex_oop---------------------------------
220 // Erase a previously saved exception from its map.
221 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
222 return common_saved_ex_oop(ex_map, true);
223 }
224
225 #ifdef ASSERT
226 //---------------------------has_saved_ex_oop----------------------------------
227 // Erase a previously saved exception from its map.
228 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
229 return ex_map->req() == ex_map->jvms()->endoff()+1;
230 }
231 #endif
232
233 //-------------------------make_exception_state--------------------------------
234 // Turn the current JVM state into an exception state, appending the ex_oop.
235 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
236 sync_jvms();
237 SafePointNode* ex_map = stop(); // do not manipulate this map any more
238 set_saved_ex_oop(ex_map, ex_oop);
239 return ex_map;
240 }
241
242
243 //--------------------------add_exception_state--------------------------------
244 // Add an exception to my list of exceptions.
245 void GraphKit::add_exception_state(SafePointNode* ex_map) {
246 if (ex_map == NULL || ex_map->control() == top()) {
247 return;
248 }
249 #ifdef ASSERT
250 verify_exception_state(ex_map);
251 if (has_exceptions()) {
252 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
253 }
254 #endif
255
256 // If there is already an exception of exactly this type, merge with it.
257 // In particular, null-checks and other low-level exceptions common up here.
258 Node* ex_oop = saved_ex_oop(ex_map);
259 const Type* ex_type = _gvn.type(ex_oop);
260 if (ex_oop == top()) {
261 // No action needed.
262 return;
263 }
264 assert(ex_type->isa_instptr(), "exception must be an instance");
265 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
266 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
267 // We check sp also because call bytecodes can generate exceptions
268 // both before and after arguments are popped!
269 if (ex_type2 == ex_type
270 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
271 combine_exception_states(ex_map, e2);
272 return;
273 }
274 }
275
276 // No pre-existing exception of the same type. Chain it on the list.
277 push_exception_state(ex_map);
278 }
279
280 //-----------------------add_exception_states_from-----------------------------
281 void GraphKit::add_exception_states_from(JVMState* jvms) {
282 SafePointNode* ex_map = jvms->map()->next_exception();
283 if (ex_map != NULL) {
284 jvms->map()->set_next_exception(NULL);
285 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
286 next_map = ex_map->next_exception();
287 ex_map->set_next_exception(NULL);
288 add_exception_state(ex_map);
289 }
290 }
291 }
292
293 //-----------------------transfer_exceptions_into_jvms-------------------------
294 JVMState* GraphKit::transfer_exceptions_into_jvms() {
295 if (map() == NULL) {
296 // We need a JVMS to carry the exceptions, but the map has gone away.
297 // Create a scratch JVMS, cloned from any of the exception states...
298 if (has_exceptions()) {
299 _map = _exceptions;
300 _map = clone_map();
301 _map->set_next_exception(NULL);
302 clear_saved_ex_oop(_map);
303 debug_only(verify_map());
304 } else {
305 // ...or created from scratch
306 JVMState* jvms = new (C) JVMState(_method, NULL);
307 jvms->set_bci(_bci);
308 jvms->set_sp(_sp);
309 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
310 set_jvms(jvms);
311 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
312 set_all_memory(top());
313 while (map()->req() < jvms->endoff()) map()->add_req(top());
314 }
315 // (This is a kludge, in case you didn't notice.)
316 set_control(top());
317 }
318 JVMState* jvms = sync_jvms();
319 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
320 jvms->map()->set_next_exception(_exceptions);
321 _exceptions = NULL; // done with this set of exceptions
322 return jvms;
323 }
324
325 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
326 assert(is_hidden_merge(dstphi), "must be a special merge node");
327 assert(is_hidden_merge(srcphi), "must be a special merge node");
328 uint limit = srcphi->req();
329 for (uint i = PhiNode::Input; i < limit; i++) {
330 dstphi->add_req(srcphi->in(i));
331 }
332 }
333 static inline void add_one_req(Node* dstphi, Node* src) {
334 assert(is_hidden_merge(dstphi), "must be a special merge node");
335 assert(!is_hidden_merge(src), "must not be a special merge node");
336 dstphi->add_req(src);
337 }
338
339 //-----------------------combine_exception_states------------------------------
340 // This helper function combines exception states by building phis on a
341 // specially marked state-merging region. These regions and phis are
342 // untransformed, and can build up gradually. The region is marked by
343 // having a control input of its exception map, rather than NULL. Such
344 // regions do not appear except in this function, and in use_exception_state.
345 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
346 if (failing()) return; // dying anyway...
347 JVMState* ex_jvms = ex_map->_jvms;
348 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
349 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
350 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
351 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
352 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
353 assert(ex_map->req() == phi_map->req(), "matching maps");
354 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
355 Node* hidden_merge_mark = root();
356 Node* region = phi_map->control();
357 MergeMemNode* phi_mem = phi_map->merged_memory();
358 MergeMemNode* ex_mem = ex_map->merged_memory();
359 if (region->in(0) != hidden_merge_mark) {
360 // The control input is not (yet) a specially-marked region in phi_map.
361 // Make it so, and build some phis.
362 region = new RegionNode(2);
363 _gvn.set_type(region, Type::CONTROL);
364 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
365 region->init_req(1, phi_map->control());
366 phi_map->set_control(region);
367 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
368 record_for_igvn(io_phi);
369 _gvn.set_type(io_phi, Type::ABIO);
370 phi_map->set_i_o(io_phi);
371 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
372 Node* m = mms.memory();
373 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
374 record_for_igvn(m_phi);
375 _gvn.set_type(m_phi, Type::MEMORY);
376 mms.set_memory(m_phi);
377 }
378 }
379
380 // Either or both of phi_map and ex_map might already be converted into phis.
381 Node* ex_control = ex_map->control();
382 // if there is special marking on ex_map also, we add multiple edges from src
383 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
384 // how wide was the destination phi_map, originally?
385 uint orig_width = region->req();
386
387 if (add_multiple) {
388 add_n_reqs(region, ex_control);
389 add_n_reqs(phi_map->i_o(), ex_map->i_o());
390 } else {
391 // ex_map has no merges, so we just add single edges everywhere
392 add_one_req(region, ex_control);
393 add_one_req(phi_map->i_o(), ex_map->i_o());
394 }
395 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
396 if (mms.is_empty()) {
397 // get a copy of the base memory, and patch some inputs into it
398 const TypePtr* adr_type = mms.adr_type(C);
399 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
400 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
401 mms.set_memory(phi);
402 // Prepare to append interesting stuff onto the newly sliced phi:
403 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
404 }
405 // Append stuff from ex_map:
406 if (add_multiple) {
407 add_n_reqs(mms.memory(), mms.memory2());
408 } else {
409 add_one_req(mms.memory(), mms.memory2());
410 }
411 }
412 uint limit = ex_map->req();
413 for (uint i = TypeFunc::Parms; i < limit; i++) {
414 // Skip everything in the JVMS after tos. (The ex_oop follows.)
415 if (i == tos) i = ex_jvms->monoff();
416 Node* src = ex_map->in(i);
417 Node* dst = phi_map->in(i);
418 if (src != dst) {
419 PhiNode* phi;
420 if (dst->in(0) != region) {
421 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
422 record_for_igvn(phi);
423 _gvn.set_type(phi, phi->type());
424 phi_map->set_req(i, dst);
425 // Prepare to append interesting stuff onto the new phi:
426 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
427 } else {
428 assert(dst->is_Phi(), "nobody else uses a hidden region");
429 phi = dst->as_Phi();
430 }
431 if (add_multiple && src->in(0) == ex_control) {
432 // Both are phis.
433 add_n_reqs(dst, src);
434 } else {
435 while (dst->req() < region->req()) add_one_req(dst, src);
436 }
437 const Type* srctype = _gvn.type(src);
438 if (phi->type() != srctype) {
439 const Type* dsttype = phi->type()->meet_speculative(srctype);
440 if (phi->type() != dsttype) {
441 phi->set_type(dsttype);
442 _gvn.set_type(phi, dsttype);
443 }
444 }
445 }
446 }
447 phi_map->merge_replaced_nodes_with(ex_map);
448 }
449
450 //--------------------------use_exception_state--------------------------------
451 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
452 if (failing()) { stop(); return top(); }
453 Node* region = phi_map->control();
454 Node* hidden_merge_mark = root();
455 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
456 Node* ex_oop = clear_saved_ex_oop(phi_map);
457 if (region->in(0) == hidden_merge_mark) {
458 // Special marking for internal ex-states. Process the phis now.
459 region->set_req(0, region); // now it's an ordinary region
460 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
461 // Note: Setting the jvms also sets the bci and sp.
462 set_control(_gvn.transform(region));
463 uint tos = jvms()->stkoff() + sp();
464 for (uint i = 1; i < tos; i++) {
465 Node* x = phi_map->in(i);
466 if (x->in(0) == region) {
467 assert(x->is_Phi(), "expected a special phi");
468 phi_map->set_req(i, _gvn.transform(x));
469 }
470 }
471 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
472 Node* x = mms.memory();
473 if (x->in(0) == region) {
474 assert(x->is_Phi(), "nobody else uses a hidden region");
475 mms.set_memory(_gvn.transform(x));
476 }
477 }
478 if (ex_oop->in(0) == region) {
479 assert(ex_oop->is_Phi(), "expected a special phi");
480 ex_oop = _gvn.transform(ex_oop);
481 }
482 } else {
483 set_jvms(phi_map->jvms());
484 }
485
486 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
487 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
488 return ex_oop;
489 }
490
491 //---------------------------------java_bc-------------------------------------
492 Bytecodes::Code GraphKit::java_bc() const {
493 ciMethod* method = this->method();
494 int bci = this->bci();
495 if (method != NULL && bci != InvocationEntryBci)
496 return method->java_code_at_bci(bci);
497 else
498 return Bytecodes::_illegal;
499 }
500
501 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
502 bool must_throw) {
503 // if the exception capability is set, then we will generate code
504 // to check the JavaThread.should_post_on_exceptions flag to see
505 // if we actually need to report exception events (for this
506 // thread). If we don't need to report exception events, we will
507 // take the normal fast path provided by add_exception_events. If
508 // exception event reporting is enabled for this thread, we will
509 // take the uncommon_trap in the BuildCutout below.
510
511 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
512 Node* jthread = _gvn.transform(new ThreadLocalNode());
513 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
514 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
515
516 // Test the should_post_on_exceptions_flag vs. 0
517 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
518 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
519
520 // Branch to slow_path if should_post_on_exceptions_flag was true
521 { BuildCutout unless(this, tst, PROB_MAX);
522 // Do not try anything fancy if we're notifying the VM on every throw.
523 // Cf. case Bytecodes::_athrow in parse2.cpp.
524 uncommon_trap(reason, Deoptimization::Action_none,
525 (ciKlass*)NULL, (char*)NULL, must_throw);
526 }
527
528 }
529
530 //------------------------------builtin_throw----------------------------------
531 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
532 bool must_throw = true;
533
534 if (env()->jvmti_can_post_on_exceptions()) {
535 // check if we must post exception events, take uncommon trap if so
536 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
537 // here if should_post_on_exceptions is false
538 // continue on with the normal codegen
539 }
540
541 // If this particular condition has not yet happened at this
542 // bytecode, then use the uncommon trap mechanism, and allow for
543 // a future recompilation if several traps occur here.
544 // If the throw is hot, try to use a more complicated inline mechanism
545 // which keeps execution inside the compiled code.
546 bool treat_throw_as_hot = false;
547 ciMethodData* md = method()->method_data();
548
549 if (ProfileTraps) {
550 if (too_many_traps(reason)) {
551 treat_throw_as_hot = true;
552 }
553 // (If there is no MDO at all, assume it is early in
554 // execution, and that any deopts are part of the
555 // startup transient, and don't need to be remembered.)
556
557 // Also, if there is a local exception handler, treat all throws
558 // as hot if there has been at least one in this method.
559 if (C->trap_count(reason) != 0
560 && method()->method_data()->trap_count(reason) != 0
561 && has_ex_handler()) {
562 treat_throw_as_hot = true;
563 }
564 }
565
566 // If this throw happens frequently, an uncommon trap might cause
567 // a performance pothole. If there is a local exception handler,
568 // and if this particular bytecode appears to be deoptimizing often,
569 // let us handle the throw inline, with a preconstructed instance.
570 // Note: If the deopt count has blown up, the uncommon trap
571 // runtime is going to flush this nmethod, not matter what.
572 if (treat_throw_as_hot
573 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
574 // If the throw is local, we use a pre-existing instance and
575 // punt on the backtrace. This would lead to a missing backtrace
576 // (a repeat of 4292742) if the backtrace object is ever asked
577 // for its backtrace.
578 // Fixing this remaining case of 4292742 requires some flavor of
579 // escape analysis. Leave that for the future.
580 ciInstance* ex_obj = NULL;
581 switch (reason) {
582 case Deoptimization::Reason_null_check:
583 ex_obj = env()->NullPointerException_instance();
584 break;
585 case Deoptimization::Reason_div0_check:
586 ex_obj = env()->ArithmeticException_instance();
587 break;
588 case Deoptimization::Reason_range_check:
589 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
590 break;
591 case Deoptimization::Reason_class_check:
592 if (java_bc() == Bytecodes::_aastore) {
593 ex_obj = env()->ArrayStoreException_instance();
594 } else {
595 ex_obj = env()->ClassCastException_instance();
596 }
597 break;
598 default:
599 break;
600 }
601 if (failing()) { stop(); return; } // exception allocation might fail
602 if (ex_obj != NULL) {
603 // Cheat with a preallocated exception object.
604 if (C->log() != NULL)
605 C->log()->elem("hot_throw preallocated='1' reason='%s'",
606 Deoptimization::trap_reason_name(reason));
607 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
608 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
609
610 // Clear the detail message of the preallocated exception object.
611 // Weblogic sometimes mutates the detail message of exceptions
612 // using reflection.
613 int offset = java_lang_Throwable::get_detailMessage_offset();
614 const TypePtr* adr_typ = ex_con->add_offset(offset);
615
616 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
617 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
618 Node *store = access_store_at(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
619
620 add_exception_state(make_exception_state(ex_node));
621 return;
622 }
623 }
624
625 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
626 // It won't be much cheaper than bailing to the interp., since we'll
627 // have to pass up all the debug-info, and the runtime will have to
628 // create the stack trace.
629
630 // Usual case: Bail to interpreter.
631 // Reserve the right to recompile if we haven't seen anything yet.
632
633 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
634 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
635 if (treat_throw_as_hot
636 && (method()->method_data()->trap_recompiled_at(bci(), m)
637 || C->too_many_traps(reason))) {
638 // We cannot afford to take more traps here. Suffer in the interpreter.
639 if (C->log() != NULL)
640 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
641 Deoptimization::trap_reason_name(reason),
642 C->trap_count(reason));
643 action = Deoptimization::Action_none;
644 }
645
646 // "must_throw" prunes the JVM state to include only the stack, if there
647 // are no local exception handlers. This should cut down on register
648 // allocation time and code size, by drastically reducing the number
649 // of in-edges on the call to the uncommon trap.
650
651 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
652 }
653
654
655 //----------------------------PreserveJVMState---------------------------------
656 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
657 debug_only(kit->verify_map());
658 _kit = kit;
659 _map = kit->map(); // preserve the map
660 _sp = kit->sp();
661 kit->set_map(clone_map ? kit->clone_map() : NULL);
662 #ifdef ASSERT
663 _bci = kit->bci();
664 Parse* parser = kit->is_Parse();
665 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
666 _block = block;
667 #endif
668 }
669 PreserveJVMState::~PreserveJVMState() {
670 GraphKit* kit = _kit;
671 #ifdef ASSERT
672 assert(kit->bci() == _bci, "bci must not shift");
673 Parse* parser = kit->is_Parse();
674 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
675 assert(block == _block, "block must not shift");
676 #endif
677 kit->set_map(_map);
678 kit->set_sp(_sp);
679 }
680
681
682 //-----------------------------BuildCutout-------------------------------------
683 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
684 : PreserveJVMState(kit)
685 {
686 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
687 SafePointNode* outer_map = _map; // preserved map is caller's
688 SafePointNode* inner_map = kit->map();
689 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
690 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
691 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
692 }
693 BuildCutout::~BuildCutout() {
694 GraphKit* kit = _kit;
695 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
696 }
697
698 //---------------------------PreserveReexecuteState----------------------------
699 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
700 assert(!kit->stopped(), "must call stopped() before");
701 _kit = kit;
702 _sp = kit->sp();
703 _reexecute = kit->jvms()->_reexecute;
704 }
705 PreserveReexecuteState::~PreserveReexecuteState() {
706 if (_kit->stopped()) return;
707 _kit->jvms()->_reexecute = _reexecute;
708 _kit->set_sp(_sp);
709 }
710
711 //------------------------------clone_map--------------------------------------
712 // Implementation of PreserveJVMState
713 //
714 // Only clone_map(...) here. If this function is only used in the
715 // PreserveJVMState class we may want to get rid of this extra
716 // function eventually and do it all there.
717
718 SafePointNode* GraphKit::clone_map() {
719 if (map() == NULL) return NULL;
720
721 // Clone the memory edge first
722 Node* mem = MergeMemNode::make(map()->memory());
723 gvn().set_type_bottom(mem);
724
725 SafePointNode *clonemap = (SafePointNode*)map()->clone();
726 JVMState* jvms = this->jvms();
727 JVMState* clonejvms = jvms->clone_shallow(C);
728 clonemap->set_memory(mem);
729 clonemap->set_jvms(clonejvms);
730 clonejvms->set_map(clonemap);
731 record_for_igvn(clonemap);
732 gvn().set_type_bottom(clonemap);
733 return clonemap;
734 }
735
736
737 //-----------------------------set_map_clone-----------------------------------
738 void GraphKit::set_map_clone(SafePointNode* m) {
739 _map = m;
740 _map = clone_map();
741 _map->set_next_exception(NULL);
742 debug_only(verify_map());
743 }
744
745
746 //----------------------------kill_dead_locals---------------------------------
747 // Detect any locals which are known to be dead, and force them to top.
748 void GraphKit::kill_dead_locals() {
749 // Consult the liveness information for the locals. If any
750 // of them are unused, then they can be replaced by top(). This
751 // should help register allocation time and cut down on the size
752 // of the deoptimization information.
753
754 // This call is made from many of the bytecode handling
755 // subroutines called from the Big Switch in do_one_bytecode.
756 // Every bytecode which might include a slow path is responsible
757 // for killing its dead locals. The more consistent we
758 // are about killing deads, the fewer useless phis will be
759 // constructed for them at various merge points.
760
761 // bci can be -1 (InvocationEntryBci). We return the entry
762 // liveness for the method.
763
764 if (method() == NULL || method()->code_size() == 0) {
765 // We are building a graph for a call to a native method.
766 // All locals are live.
767 return;
768 }
769
770 ResourceMark rm;
771
772 // Consult the liveness information for the locals. If any
773 // of them are unused, then they can be replaced by top(). This
774 // should help register allocation time and cut down on the size
775 // of the deoptimization information.
776 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
777
778 int len = (int)live_locals.size();
779 assert(len <= jvms()->loc_size(), "too many live locals");
780 for (int local = 0; local < len; local++) {
781 if (!live_locals.at(local)) {
782 set_local(local, top());
783 }
784 }
785 }
786
787 #ifdef ASSERT
788 //-------------------------dead_locals_are_killed------------------------------
789 // Return true if all dead locals are set to top in the map.
790 // Used to assert "clean" debug info at various points.
791 bool GraphKit::dead_locals_are_killed() {
792 if (method() == NULL || method()->code_size() == 0) {
793 // No locals need to be dead, so all is as it should be.
794 return true;
795 }
796
797 // Make sure somebody called kill_dead_locals upstream.
798 ResourceMark rm;
799 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
800 if (jvms->loc_size() == 0) continue; // no locals to consult
801 SafePointNode* map = jvms->map();
802 ciMethod* method = jvms->method();
803 int bci = jvms->bci();
804 if (jvms == this->jvms()) {
805 bci = this->bci(); // it might not yet be synched
806 }
807 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
808 int len = (int)live_locals.size();
809 if (!live_locals.is_valid() || len == 0)
810 // This method is trivial, or is poisoned by a breakpoint.
811 return true;
812 assert(len == jvms->loc_size(), "live map consistent with locals map");
813 for (int local = 0; local < len; local++) {
814 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
815 if (PrintMiscellaneous && (Verbose || WizardMode)) {
816 tty->print_cr("Zombie local %d: ", local);
817 jvms->dump();
818 }
819 return false;
820 }
821 }
822 }
823 return true;
824 }
825
826 #endif //ASSERT
827
828 // Helper function for enforcing certain bytecodes to reexecute if
829 // deoptimization happens
830 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
831 ciMethod* cur_method = jvms->method();
832 int cur_bci = jvms->bci();
833 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
834 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
835 return Interpreter::bytecode_should_reexecute(code) ||
836 (is_anewarray && (code == Bytecodes::_multianewarray));
837 // Reexecute _multianewarray bytecode which was replaced with
838 // sequence of [a]newarray. See Parse::do_multianewarray().
839 //
840 // Note: interpreter should not have it set since this optimization
841 // is limited by dimensions and guarded by flag so in some cases
842 // multianewarray() runtime calls will be generated and
843 // the bytecode should not be reexecutes (stack will not be reset).
844 } else {
845 return false;
846 }
847 }
848
849 // Helper function for adding JVMState and debug information to node
850 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
851 // Add the safepoint edges to the call (or other safepoint).
852
853 // Make sure dead locals are set to top. This
854 // should help register allocation time and cut down on the size
855 // of the deoptimization information.
856 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
857
858 // Walk the inline list to fill in the correct set of JVMState's
859 // Also fill in the associated edges for each JVMState.
860
861 // If the bytecode needs to be reexecuted we need to put
862 // the arguments back on the stack.
863 const bool should_reexecute = jvms()->should_reexecute();
864 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
865
866 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
867 // undefined if the bci is different. This is normal for Parse but it
868 // should not happen for LibraryCallKit because only one bci is processed.
869 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
870 "in LibraryCallKit the reexecute bit should not change");
871
872 // If we are guaranteed to throw, we can prune everything but the
873 // input to the current bytecode.
874 bool can_prune_locals = false;
875 uint stack_slots_not_pruned = 0;
876 int inputs = 0, depth = 0;
877 if (must_throw) {
878 assert(method() == youngest_jvms->method(), "sanity");
879 if (compute_stack_effects(inputs, depth)) {
880 can_prune_locals = true;
881 stack_slots_not_pruned = inputs;
882 }
883 }
884
885 if (env()->should_retain_local_variables()) {
886 // At any safepoint, this method can get breakpointed, which would
887 // then require an immediate deoptimization.
888 can_prune_locals = false; // do not prune locals
889 stack_slots_not_pruned = 0;
890 }
891
892 // do not scribble on the input jvms
893 JVMState* out_jvms = youngest_jvms->clone_deep(C);
894 call->set_jvms(out_jvms); // Start jvms list for call node
895
896 // For a known set of bytecodes, the interpreter should reexecute them if
897 // deoptimization happens. We set the reexecute state for them here
898 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
899 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
900 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
901 }
902
903 // Presize the call:
904 DEBUG_ONLY(uint non_debug_edges = call->req());
905 call->add_req_batch(top(), youngest_jvms->debug_depth());
906 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
907
908 // Set up edges so that the call looks like this:
909 // Call [state:] ctl io mem fptr retadr
910 // [parms:] parm0 ... parmN
911 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
912 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
913 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
914 // Note that caller debug info precedes callee debug info.
915
916 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
917 uint debug_ptr = call->req();
918
919 // Loop over the map input edges associated with jvms, add them
920 // to the call node, & reset all offsets to match call node array.
921 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
922 uint debug_end = debug_ptr;
923 uint debug_start = debug_ptr - in_jvms->debug_size();
924 debug_ptr = debug_start; // back up the ptr
925
926 uint p = debug_start; // walks forward in [debug_start, debug_end)
927 uint j, k, l;
928 SafePointNode* in_map = in_jvms->map();
929 out_jvms->set_map(call);
930
931 if (can_prune_locals) {
932 assert(in_jvms->method() == out_jvms->method(), "sanity");
933 // If the current throw can reach an exception handler in this JVMS,
934 // then we must keep everything live that can reach that handler.
935 // As a quick and dirty approximation, we look for any handlers at all.
936 if (in_jvms->method()->has_exception_handlers()) {
937 can_prune_locals = false;
938 }
939 }
940
941 // Add the Locals
942 k = in_jvms->locoff();
943 l = in_jvms->loc_size();
944 out_jvms->set_locoff(p);
945 if (!can_prune_locals) {
946 for (j = 0; j < l; j++)
947 call->set_req(p++, in_map->in(k+j));
948 } else {
949 p += l; // already set to top above by add_req_batch
950 }
951
952 // Add the Expression Stack
953 k = in_jvms->stkoff();
954 l = in_jvms->sp();
955 out_jvms->set_stkoff(p);
956 if (!can_prune_locals) {
957 for (j = 0; j < l; j++)
958 call->set_req(p++, in_map->in(k+j));
959 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
960 // Divide stack into {S0,...,S1}, where S0 is set to top.
961 uint s1 = stack_slots_not_pruned;
962 stack_slots_not_pruned = 0; // for next iteration
963 if (s1 > l) s1 = l;
964 uint s0 = l - s1;
965 p += s0; // skip the tops preinstalled by add_req_batch
966 for (j = s0; j < l; j++)
967 call->set_req(p++, in_map->in(k+j));
968 } else {
969 p += l; // already set to top above by add_req_batch
970 }
971
972 // Add the Monitors
973 k = in_jvms->monoff();
974 l = in_jvms->mon_size();
975 out_jvms->set_monoff(p);
976 for (j = 0; j < l; j++)
977 call->set_req(p++, in_map->in(k+j));
978
979 // Copy any scalar object fields.
980 k = in_jvms->scloff();
981 l = in_jvms->scl_size();
982 out_jvms->set_scloff(p);
983 for (j = 0; j < l; j++)
984 call->set_req(p++, in_map->in(k+j));
985
986 // Finish the new jvms.
987 out_jvms->set_endoff(p);
988
989 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
990 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
991 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
992 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
993 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
994 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
995
996 // Update the two tail pointers in parallel.
997 out_jvms = out_jvms->caller();
998 in_jvms = in_jvms->caller();
999 }
1000
1001 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1002
1003 // Test the correctness of JVMState::debug_xxx accessors:
1004 assert(call->jvms()->debug_start() == non_debug_edges, "");
1005 assert(call->jvms()->debug_end() == call->req(), "");
1006 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1007 }
1008
1009 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1010 Bytecodes::Code code = java_bc();
1011 if (code == Bytecodes::_wide) {
1012 code = method()->java_code_at_bci(bci() + 1);
1013 }
1014
1015 BasicType rtype = T_ILLEGAL;
1016 int rsize = 0;
1017
1018 if (code != Bytecodes::_illegal) {
1019 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1020 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1021 if (rtype < T_CONFLICT)
1022 rsize = type2size[rtype];
1023 }
1024
1025 switch (code) {
1026 case Bytecodes::_illegal:
1027 return false;
1028
1029 case Bytecodes::_ldc:
1030 case Bytecodes::_ldc_w:
1031 case Bytecodes::_ldc2_w:
1032 inputs = 0;
1033 break;
1034
1035 case Bytecodes::_dup: inputs = 1; break;
1036 case Bytecodes::_dup_x1: inputs = 2; break;
1037 case Bytecodes::_dup_x2: inputs = 3; break;
1038 case Bytecodes::_dup2: inputs = 2; break;
1039 case Bytecodes::_dup2_x1: inputs = 3; break;
1040 case Bytecodes::_dup2_x2: inputs = 4; break;
1041 case Bytecodes::_swap: inputs = 2; break;
1042 case Bytecodes::_arraylength: inputs = 1; break;
1043
1044 case Bytecodes::_getstatic:
1045 case Bytecodes::_putstatic:
1046 case Bytecodes::_getfield:
1047 case Bytecodes::_putfield:
1048 {
1049 bool ignored_will_link;
1050 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1051 int size = field->type()->size();
1052 bool is_get = (depth >= 0), is_static = (depth & 1);
1053 inputs = (is_static ? 0 : 1);
1054 if (is_get) {
1055 depth = size - inputs;
1056 } else {
1057 inputs += size; // putxxx pops the value from the stack
1058 depth = - inputs;
1059 }
1060 }
1061 break;
1062
1063 case Bytecodes::_invokevirtual:
1064 case Bytecodes::_invokespecial:
1065 case Bytecodes::_invokestatic:
1066 case Bytecodes::_invokedynamic:
1067 case Bytecodes::_invokeinterface:
1068 {
1069 bool ignored_will_link;
1070 ciSignature* declared_signature = NULL;
1071 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1072 assert(declared_signature != NULL, "cannot be null");
1073 inputs = declared_signature->arg_size_for_bc(code);
1074 int size = declared_signature->return_type()->size();
1075 depth = size - inputs;
1076 }
1077 break;
1078
1079 case Bytecodes::_multianewarray:
1080 {
1081 ciBytecodeStream iter(method());
1082 iter.reset_to_bci(bci());
1083 iter.next();
1084 inputs = iter.get_dimensions();
1085 assert(rsize == 1, "");
1086 depth = rsize - inputs;
1087 }
1088 break;
1089
1090 case Bytecodes::_withfield: {
1091 bool ignored_will_link;
1092 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1093 int size = field->type()->size();
1094 inputs = size+1;
1095 depth = rsize - inputs;
1096 break;
1097 }
1098
1099 case Bytecodes::_ireturn:
1100 case Bytecodes::_lreturn:
1101 case Bytecodes::_freturn:
1102 case Bytecodes::_dreturn:
1103 case Bytecodes::_areturn:
1104 assert(rsize == -depth, "");
1105 inputs = rsize;
1106 break;
1107
1108 case Bytecodes::_jsr:
1109 case Bytecodes::_jsr_w:
1110 inputs = 0;
1111 depth = 1; // S.B. depth=1, not zero
1112 break;
1113
1114 default:
1115 // bytecode produces a typed result
1116 inputs = rsize - depth;
1117 assert(inputs >= 0, "");
1118 break;
1119 }
1120
1121 #ifdef ASSERT
1122 // spot check
1123 int outputs = depth + inputs;
1124 assert(outputs >= 0, "sanity");
1125 switch (code) {
1126 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1127 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1128 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1129 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1130 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1131 default: break;
1132 }
1133 #endif //ASSERT
1134
1135 return true;
1136 }
1137
1138
1139
1140 //------------------------------basic_plus_adr---------------------------------
1141 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1142 // short-circuit a common case
1143 if (offset == intcon(0)) return ptr;
1144 return _gvn.transform( new AddPNode(base, ptr, offset) );
1145 }
1146
1147 Node* GraphKit::ConvI2L(Node* offset) {
1148 // short-circuit a common case
1149 jint offset_con = find_int_con(offset, Type::OffsetBot);
1150 if (offset_con != Type::OffsetBot) {
1151 return longcon((jlong) offset_con);
1152 }
1153 return _gvn.transform( new ConvI2LNode(offset));
1154 }
1155
1156 Node* GraphKit::ConvI2UL(Node* offset) {
1157 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1158 if (offset_con != (juint) Type::OffsetBot) {
1159 return longcon((julong) offset_con);
1160 }
1161 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1162 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1163 return _gvn.transform( new AndLNode(conv, mask) );
1164 }
1165
1166 Node* GraphKit::ConvL2I(Node* offset) {
1167 // short-circuit a common case
1168 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1169 if (offset_con != (jlong)Type::OffsetBot) {
1170 return intcon((int) offset_con);
1171 }
1172 return _gvn.transform( new ConvL2INode(offset));
1173 }
1174
1175 //-------------------------load_object_klass-----------------------------------
1176 Node* GraphKit::load_object_klass(Node* obj) {
1177 // Special-case a fresh allocation to avoid building nodes:
1178 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1179 if (akls != NULL) return akls;
1180 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1181 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1182 }
1183
1184 //-------------------------load_array_length-----------------------------------
1185 Node* GraphKit::load_array_length(Node* array) {
1186 // Special-case a fresh allocation to avoid building nodes:
1187 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1188 Node *alen;
1189 if (alloc == NULL) {
1190 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1191 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1192 } else {
1193 alen = alloc->Ideal_length();
1194 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1195 if (ccast != alen) {
1196 alen = _gvn.transform(ccast);
1197 }
1198 }
1199 return alen;
1200 }
1201
1202 //------------------------------do_null_check----------------------------------
1203 // Helper function to do a NULL pointer check. Returned value is
1204 // the incoming address with NULL casted away. You are allowed to use the
1205 // not-null value only if you are control dependent on the test.
1206 #ifndef PRODUCT
1207 extern int explicit_null_checks_inserted,
1208 explicit_null_checks_elided;
1209 #endif
1210 Node* GraphKit::null_check_common(Node* value, BasicType type,
1211 // optional arguments for variations:
1212 bool assert_null,
1213 Node* *null_control,
1214 bool speculative) {
1215 assert(!assert_null || null_control == NULL, "not both at once");
1216 if (stopped()) return top();
1217 NOT_PRODUCT(explicit_null_checks_inserted++);
1218
1219 // Construct NULL check
1220 Node *chk = NULL;
1221 switch(type) {
1222 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1223 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1224 case T_VALUETYPE : // fall through
1225 case T_ARRAY : // fall through
1226 type = T_OBJECT; // simplify further tests
1227 case T_OBJECT : {
1228 const Type *t = _gvn.type( value );
1229
1230 const TypeOopPtr* tp = t->isa_oopptr();
1231 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1232 // Only for do_null_check, not any of its siblings:
1233 && !assert_null && null_control == NULL) {
1234 // Usually, any field access or invocation on an unloaded oop type
1235 // will simply fail to link, since the statically linked class is
1236 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1237 // the static class is loaded but the sharper oop type is not.
1238 // Rather than checking for this obscure case in lots of places,
1239 // we simply observe that a null check on an unloaded class
1240 // will always be followed by a nonsense operation, so we
1241 // can just issue the uncommon trap here.
1242 // Our access to the unloaded class will only be correct
1243 // after it has been loaded and initialized, which requires
1244 // a trip through the interpreter.
1245 #ifndef PRODUCT
1246 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1247 #endif
1248 uncommon_trap(Deoptimization::Reason_unloaded,
1249 Deoptimization::Action_reinterpret,
1250 tp->klass(), "!loaded");
1251 return top();
1252 }
1253
1254 if (assert_null) {
1255 // See if the type is contained in NULL_PTR.
1256 // If so, then the value is already null.
1257 if (t->higher_equal(TypePtr::NULL_PTR)) {
1258 NOT_PRODUCT(explicit_null_checks_elided++);
1259 return value; // Elided null assert quickly!
1260 }
1261 } else {
1262 // See if mixing in the NULL pointer changes type.
1263 // If so, then the NULL pointer was not allowed in the original
1264 // type. In other words, "value" was not-null.
1265 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1266 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1267 NOT_PRODUCT(explicit_null_checks_elided++);
1268 return value; // Elided null check quickly!
1269 }
1270 }
1271 chk = new CmpPNode( value, null() );
1272 break;
1273 }
1274
1275 default:
1276 fatal("unexpected type: %s", type2name(type));
1277 }
1278 assert(chk != NULL, "sanity check");
1279 chk = _gvn.transform(chk);
1280
1281 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1282 BoolNode *btst = new BoolNode( chk, btest);
1283 Node *tst = _gvn.transform( btst );
1284
1285 //-----------
1286 // if peephole optimizations occurred, a prior test existed.
1287 // If a prior test existed, maybe it dominates as we can avoid this test.
1288 if (tst != btst && type == T_OBJECT) {
1289 // At this point we want to scan up the CFG to see if we can
1290 // find an identical test (and so avoid this test altogether).
1291 Node *cfg = control();
1292 int depth = 0;
1293 while( depth < 16 ) { // Limit search depth for speed
1294 if( cfg->Opcode() == Op_IfTrue &&
1295 cfg->in(0)->in(1) == tst ) {
1296 // Found prior test. Use "cast_not_null" to construct an identical
1297 // CastPP (and hence hash to) as already exists for the prior test.
1298 // Return that casted value.
1299 if (assert_null) {
1300 replace_in_map(value, null());
1301 return null(); // do not issue the redundant test
1302 }
1303 Node *oldcontrol = control();
1304 set_control(cfg);
1305 Node *res = cast_not_null(value);
1306 set_control(oldcontrol);
1307 NOT_PRODUCT(explicit_null_checks_elided++);
1308 return res;
1309 }
1310 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1311 if (cfg == NULL) break; // Quit at region nodes
1312 depth++;
1313 }
1314 }
1315
1316 //-----------
1317 // Branch to failure if null
1318 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1319 Deoptimization::DeoptReason reason;
1320 if (assert_null) {
1321 reason = Deoptimization::reason_null_assert(speculative);
1322 } else if (type == T_OBJECT) {
1323 reason = Deoptimization::reason_null_check(speculative);
1324 } else {
1325 reason = Deoptimization::Reason_div0_check;
1326 }
1327 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1328 // ciMethodData::has_trap_at will return a conservative -1 if any
1329 // must-be-null assertion has failed. This could cause performance
1330 // problems for a method after its first do_null_assert failure.
1331 // Consider using 'Reason_class_check' instead?
1332
1333 // To cause an implicit null check, we set the not-null probability
1334 // to the maximum (PROB_MAX). For an explicit check the probability
1335 // is set to a smaller value.
1336 if (null_control != NULL || too_many_traps(reason)) {
1337 // probability is less likely
1338 ok_prob = PROB_LIKELY_MAG(3);
1339 } else if (!assert_null &&
1340 (ImplicitNullCheckThreshold > 0) &&
1341 method() != NULL &&
1342 (method()->method_data()->trap_count(reason)
1343 >= (uint)ImplicitNullCheckThreshold)) {
1344 ok_prob = PROB_LIKELY_MAG(3);
1345 }
1346
1347 if (null_control != NULL) {
1348 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1349 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1350 set_control( _gvn.transform( new IfTrueNode(iff)));
1351 #ifndef PRODUCT
1352 if (null_true == top()) {
1353 explicit_null_checks_elided++;
1354 }
1355 #endif
1356 (*null_control) = null_true;
1357 } else {
1358 BuildCutout unless(this, tst, ok_prob);
1359 // Check for optimizer eliding test at parse time
1360 if (stopped()) {
1361 // Failure not possible; do not bother making uncommon trap.
1362 NOT_PRODUCT(explicit_null_checks_elided++);
1363 } else if (assert_null) {
1364 uncommon_trap(reason,
1365 Deoptimization::Action_make_not_entrant,
1366 NULL, "assert_null");
1367 } else {
1368 replace_in_map(value, zerocon(type));
1369 builtin_throw(reason);
1370 }
1371 }
1372
1373 // Must throw exception, fall-thru not possible?
1374 if (stopped()) {
1375 return top(); // No result
1376 }
1377
1378 if (assert_null) {
1379 // Cast obj to null on this path.
1380 replace_in_map(value, zerocon(type));
1381 return zerocon(type);
1382 }
1383
1384 // Cast obj to not-null on this path, if there is no null_control.
1385 // (If there is a null_control, a non-null value may come back to haunt us.)
1386 if (type == T_OBJECT) {
1387 Node* cast = cast_not_null(value, false);
1388 if (null_control == NULL || (*null_control) == top())
1389 replace_in_map(value, cast);
1390 value = cast;
1391 }
1392
1393 return value;
1394 }
1395
1396
1397 //------------------------------cast_not_null----------------------------------
1398 // Cast obj to not-null on this path
1399 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1400 const Type *t = _gvn.type(obj);
1401 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1402 // Object is already not-null?
1403 if( t == t_not_null ) return obj;
1404
1405 Node *cast = new CastPPNode(obj,t_not_null);
1406 cast->init_req(0, control());
1407 cast = _gvn.transform( cast );
1408
1409 // Scan for instances of 'obj' in the current JVM mapping.
1410 // These instances are known to be not-null after the test.
1411 if (do_replace_in_map)
1412 replace_in_map(obj, cast);
1413
1414 return cast; // Return casted value
1415 }
1416
1417 // Sometimes in intrinsics, we implicitly know an object is not null
1418 // (there's no actual null check) so we can cast it to not null. In
1419 // the course of optimizations, the input to the cast can become null.
1420 // In that case that data path will die and we need the control path
1421 // to become dead as well to keep the graph consistent. So we have to
1422 // add a check for null for which one branch can't be taken. It uses
1423 // an Opaque4 node that will cause the check to be removed after loop
1424 // opts so the test goes away and the compiled code doesn't execute a
1425 // useless check.
1426 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1427 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1428 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1429 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1430 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1431 _gvn.set_type(iff, iff->Value(&_gvn));
1432 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1433 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1434 Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1435 C->root()->add_req(halt);
1436 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1437 set_control(if_t);
1438 return cast_not_null(value, do_replace_in_map);
1439 }
1440
1441
1442 //--------------------------replace_in_map-------------------------------------
1443 void GraphKit::replace_in_map(Node* old, Node* neww) {
1444 if (old == neww) {
1445 return;
1446 }
1447
1448 map()->replace_edge(old, neww);
1449
1450 // Note: This operation potentially replaces any edge
1451 // on the map. This includes locals, stack, and monitors
1452 // of the current (innermost) JVM state.
1453
1454 // don't let inconsistent types from profiling escape this
1455 // method
1456
1457 const Type* told = _gvn.type(old);
1458 const Type* tnew = _gvn.type(neww);
1459
1460 if (!tnew->higher_equal(told)) {
1461 return;
1462 }
1463
1464 map()->record_replaced_node(old, neww);
1465 }
1466
1467
1468 //=============================================================================
1469 //--------------------------------memory---------------------------------------
1470 Node* GraphKit::memory(uint alias_idx) {
1471 MergeMemNode* mem = merged_memory();
1472 Node* p = mem->memory_at(alias_idx);
1473 _gvn.set_type(p, Type::MEMORY); // must be mapped
1474 return p;
1475 }
1476
1477 //-----------------------------reset_memory------------------------------------
1478 Node* GraphKit::reset_memory() {
1479 Node* mem = map()->memory();
1480 // do not use this node for any more parsing!
1481 debug_only( map()->set_memory((Node*)NULL) );
1482 return _gvn.transform( mem );
1483 }
1484
1485 //------------------------------set_all_memory---------------------------------
1486 void GraphKit::set_all_memory(Node* newmem) {
1487 Node* mergemem = MergeMemNode::make(newmem);
1488 gvn().set_type_bottom(mergemem);
1489 map()->set_memory(mergemem);
1490 }
1491
1492 //------------------------------set_all_memory_call----------------------------
1493 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1494 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1495 set_all_memory(newmem);
1496 }
1497
1498 //=============================================================================
1499 //
1500 // parser factory methods for MemNodes
1501 //
1502 // These are layered on top of the factory methods in LoadNode and StoreNode,
1503 // and integrate with the parser's memory state and _gvn engine.
1504 //
1505
1506 // factory methods in "int adr_idx"
1507 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1508 int adr_idx,
1509 MemNode::MemOrd mo,
1510 LoadNode::ControlDependency control_dependency,
1511 bool require_atomic_access,
1512 bool unaligned,
1513 bool mismatched) {
1514 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1515 const TypePtr* adr_type = NULL; // debug-mode-only argument
1516 debug_only(adr_type = C->get_adr_type(adr_idx));
1517 Node* mem = memory(adr_idx);
1518 Node* ld;
1519 if (require_atomic_access && bt == T_LONG) {
1520 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1521 } else if (require_atomic_access && bt == T_DOUBLE) {
1522 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1523 } else {
1524 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1525 }
1526 ld = _gvn.transform(ld);
1527
1528 if (((bt == T_OBJECT || bt == T_VALUETYPE || bt == T_VALUETYPEPTR) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1529 // Improve graph before escape analysis and boxing elimination.
1530 record_for_igvn(ld);
1531 }
1532 return ld;
1533 }
1534
1535 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1536 int adr_idx,
1537 MemNode::MemOrd mo,
1538 bool require_atomic_access,
1539 bool unaligned,
1540 bool mismatched) {
1541 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1542 const TypePtr* adr_type = NULL;
1543 debug_only(adr_type = C->get_adr_type(adr_idx));
1544 Node *mem = memory(adr_idx);
1545 Node* st;
1546 if (require_atomic_access && bt == T_LONG) {
1547 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1548 } else if (require_atomic_access && bt == T_DOUBLE) {
1549 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1550 } else {
1551 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1552 }
1553 if (unaligned) {
1554 st->as_Store()->set_unaligned_access();
1555 }
1556 if (mismatched) {
1557 st->as_Store()->set_mismatched_access();
1558 }
1559 st = _gvn.transform(st);
1560 set_memory(st, adr_idx);
1561 // Back-to-back stores can only remove intermediate store with DU info
1562 // so push on worklist for optimizer.
1563 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1564 record_for_igvn(st);
1565
1566 return st;
1567 }
1568
1569 Node* GraphKit::access_store_at(Node* ctl,
1570 Node* obj,
1571 Node* adr,
1572 const TypePtr* adr_type,
1573 Node* val,
1574 const Type* val_type,
1575 BasicType bt,
1576 DecoratorSet decorators,
1577 bool deoptimize_on_exception) {
1578 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1579 // could be delayed during Parse (for example, in adjust_map_after_if()).
1580 // Execute transformation here to avoid barrier generation in such case.
1581 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1582 val = _gvn.makecon(TypePtr::NULL_PTR);
1583 }
1584
1585 set_control(ctl);
1586 if (stopped()) {
1587 return top(); // Dead path ?
1588 }
1589
1590 assert(val != NULL, "not dead path");
1591 if (val->is_ValueType()) {
1592 // Allocate value type and get oop
1593 val = val->as_ValueType()->allocate(this, deoptimize_on_exception)->get_oop();
1594 }
1595
1596 C2AccessValuePtr addr(adr, adr_type);
1597 C2AccessValue value(val, val_type);
1598 C2Access access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1599 if (access.is_raw()) {
1600 return _barrier_set->BarrierSetC2::store_at(access, value);
1601 } else {
1602 return _barrier_set->store_at(access, value);
1603 }
1604 }
1605
1606 Node* GraphKit::access_load_at(Node* obj, // containing obj
1607 Node* adr, // actual adress to store val at
1608 const TypePtr* adr_type,
1609 const Type* val_type,
1610 BasicType bt,
1611 DecoratorSet decorators) {
1612 if (stopped()) {
1613 return top(); // Dead path ?
1614 }
1615
1616 C2AccessValuePtr addr(adr, adr_type);
1617 C2Access access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1618 if (access.is_raw()) {
1619 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1620 } else {
1621 return _barrier_set->load_at(access, val_type);
1622 }
1623 }
1624
1625 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* ctl,
1626 Node* obj,
1627 Node* adr,
1628 const TypePtr* adr_type,
1629 int alias_idx,
1630 Node* expected_val,
1631 Node* new_val,
1632 const Type* value_type,
1633 BasicType bt,
1634 DecoratorSet decorators) {
1635 set_control(ctl);
1636 C2AccessValuePtr addr(adr, adr_type);
1637 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1638 bt, obj, addr, alias_idx);
1639 if (access.is_raw()) {
1640 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1641 } else {
1642 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1643 }
1644 }
1645
1646 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* ctl,
1647 Node* obj,
1648 Node* adr,
1649 const TypePtr* adr_type,
1650 int alias_idx,
1651 Node* expected_val,
1652 Node* new_val,
1653 const Type* value_type,
1654 BasicType bt,
1655 DecoratorSet decorators) {
1656 set_control(ctl);
1657 C2AccessValuePtr addr(adr, adr_type);
1658 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1659 bt, obj, addr, alias_idx);
1660 if (access.is_raw()) {
1661 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1662 } else {
1663 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1664 }
1665 }
1666
1667 Node* GraphKit::access_atomic_xchg_at(Node* ctl,
1668 Node* obj,
1669 Node* adr,
1670 const TypePtr* adr_type,
1671 int alias_idx,
1672 Node* new_val,
1673 const Type* value_type,
1674 BasicType bt,
1675 DecoratorSet decorators) {
1676 set_control(ctl);
1677 C2AccessValuePtr addr(adr, adr_type);
1678 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1679 bt, obj, addr, alias_idx);
1680 if (access.is_raw()) {
1681 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1682 } else {
1683 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1684 }
1685 }
1686
1687 Node* GraphKit::access_atomic_add_at(Node* ctl,
1688 Node* obj,
1689 Node* adr,
1690 const TypePtr* adr_type,
1691 int alias_idx,
1692 Node* new_val,
1693 const Type* value_type,
1694 BasicType bt,
1695 DecoratorSet decorators) {
1696 set_control(ctl);
1697 C2AccessValuePtr addr(adr, adr_type);
1698 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1699 if (access.is_raw()) {
1700 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1701 } else {
1702 return _barrier_set->atomic_add_at(access, new_val, value_type);
1703 }
1704 }
1705
1706 void GraphKit::access_clone(Node* ctl, Node* src, Node* dst, Node* size, bool is_array) {
1707 set_control(ctl);
1708 return _barrier_set->clone(this, src, dst, size, is_array);
1709 }
1710
1711 //-------------------------array_element_address-------------------------
1712 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1713 const TypeInt* sizetype, Node* ctrl) {
1714 uint shift = exact_log2(type2aelembytes(elembt));
1715 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1716 if (arytype_klass->is_value_array_klass()) {
1717 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass();
1718 shift = vak->log2_element_size();
1719 }
1720 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1721
1722 // short-circuit a common case (saves lots of confusing waste motion)
1723 jint idx_con = find_int_con(idx, -1);
1724 if (idx_con >= 0) {
1725 intptr_t offset = header + ((intptr_t)idx_con << shift);
1726 return basic_plus_adr(ary, offset);
1727 }
1728
1729 // must be correct type for alignment purposes
1730 Node* base = basic_plus_adr(ary, header);
1731 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1732 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1733 return basic_plus_adr(ary, base, scale);
1734 }
1735
1736 //-------------------------load_array_element-------------------------
1737 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1738 const Type* elemtype = arytype->elem();
1739 BasicType elembt = elemtype->array_element_basic_type();
1740 assert(elembt != T_VALUETYPE, "value types are not supported by this method");
1741 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1742 if (elembt == T_NARROWOOP) {
1743 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1744 }
1745 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1746 return ld;
1747 }
1748
1749 //-------------------------set_arguments_for_java_call-------------------------
1750 // Arguments (pre-popped from the stack) are taken from the JVMS.
1751 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1752 // Add the call arguments:
1753 const TypeTuple* domain = call->tf()->domain_sig();
1754 uint nargs = domain->cnt();
1755 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1756 Node* arg = argument(i-TypeFunc::Parms);
1757 const Type* t = domain->field_at(i);
1758 if (arg->is_ValueType()) {
1759 assert(t->is_oopptr()->can_be_value_type(), "wrong argument type");
1760 ValueTypeNode* vt = arg->as_ValueType();
1761 if (ValueTypePassFieldsAsArgs) {
1762 // We don't pass value type arguments by reference but instead
1763 // pass each field of the value type
1764 idx += vt->pass_fields(call, idx, *this);
1765 // If a value type argument is passed as fields, attach the Method* to the call site
1766 // to be able to access the extended signature later via attached_method_before_pc().
1767 // For example, see CompiledMethod::preserve_callee_argument_oops().
1768 call->set_override_symbolic_info(true);
1769 continue;
1770 } else {
1771 // Pass value type argument via oop to callee
1772 arg = vt->allocate(this)->get_oop();
1773 }
1774 } else if (t->is_valuetypeptr()) {
1775 // Constant null passed for a value type argument
1776 assert(arg->bottom_type()->remove_speculative() == TypePtr::NULL_PTR, "Anything other than null?");
1777 ciMethod* declared_method = method()->get_method_at_bci(bci());
1778 int arg_size = declared_method->signature()->arg_size_for_bc(java_bc());
1779 inc_sp(arg_size); // restore arguments
1780 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none);
1781 return;
1782 }
1783 call->init_req(idx, arg);
1784 idx++;
1785 }
1786 }
1787
1788 //---------------------------set_edges_for_java_call---------------------------
1789 // Connect a newly created call into the current JVMS.
1790 // A return value node (if any) is returned from set_edges_for_java_call.
1791 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1792
1793 // Add the predefined inputs:
1794 call->init_req( TypeFunc::Control, control() );
1795 call->init_req( TypeFunc::I_O , i_o() );
1796 call->init_req( TypeFunc::Memory , reset_memory() );
1797 call->init_req( TypeFunc::FramePtr, frameptr() );
1798 call->init_req( TypeFunc::ReturnAdr, top() );
1799
1800 add_safepoint_edges(call, must_throw);
1801
1802 Node* xcall = _gvn.transform(call);
1803
1804 if (xcall == top()) {
1805 set_control(top());
1806 return;
1807 }
1808 assert(xcall == call, "call identity is stable");
1809
1810 // Re-use the current map to produce the result.
1811
1812 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1813 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1814 set_all_memory_call(xcall, separate_io_proj);
1815
1816 //return xcall; // no need, caller already has it
1817 }
1818
1819 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1820 if (stopped()) return top(); // maybe the call folded up?
1821
1822 // Note: Since any out-of-line call can produce an exception,
1823 // we always insert an I_O projection from the call into the result.
1824
1825 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1826
1827 if (separate_io_proj) {
1828 // The caller requested separate projections be used by the fall
1829 // through and exceptional paths, so replace the projections for
1830 // the fall through path.
1831 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1832 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1833 }
1834
1835 // Capture the return value, if any.
1836 Node* ret;
1837 if (call->method() == NULL ||
1838 call->method()->return_type()->basic_type() == T_VOID) {
1839 ret = top();
1840 } else {
1841 if (!call->tf()->returns_value_type_as_fields()) {
1842 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1843 } else {
1844 // Return of multiple values (value type fields): we create a
1845 // ValueType node, each field is a projection from the call.
1846 const TypeTuple* range_sig = call->tf()->range_sig();
1847 const Type* t = range_sig->field_at(TypeFunc::Parms);
1848 assert(t->is_valuetypeptr(), "only value types for multiple return values");
1849 ciValueKlass* vk = t->value_klass();
1850 ret = ValueTypeNode::make_from_multi(this, call, vk, TypeFunc::Parms+1, false);
1851 }
1852 }
1853
1854 return ret;
1855 }
1856
1857 //--------------------set_predefined_input_for_runtime_call--------------------
1858 // Reading and setting the memory state is way conservative here.
1859 // The real problem is that I am not doing real Type analysis on memory,
1860 // so I cannot distinguish card mark stores from other stores. Across a GC
1861 // point the Store Barrier and the card mark memory has to agree. I cannot
1862 // have a card mark store and its barrier split across the GC point from
1863 // either above or below. Here I get that to happen by reading ALL of memory.
1864 // A better answer would be to separate out card marks from other memory.
1865 // For now, return the input memory state, so that it can be reused
1866 // after the call, if this call has restricted memory effects.
1867 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1868 // Set fixed predefined input arguments
1869 Node* memory = reset_memory();
1870 call->init_req( TypeFunc::Control, control() );
1871 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1872 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1873 call->init_req( TypeFunc::FramePtr, frameptr() );
1874 call->init_req( TypeFunc::ReturnAdr, top() );
1875 return memory;
1876 }
1877
1878 //-------------------set_predefined_output_for_runtime_call--------------------
1879 // Set control and memory (not i_o) from the call.
1880 // If keep_mem is not NULL, use it for the output state,
1881 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1882 // If hook_mem is NULL, this call produces no memory effects at all.
1883 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1884 // then only that memory slice is taken from the call.
1885 // In the last case, we must put an appropriate memory barrier before
1886 // the call, so as to create the correct anti-dependencies on loads
1887 // preceding the call.
1888 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1889 Node* keep_mem,
1890 const TypePtr* hook_mem) {
1891 // no i/o
1892 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1893 if (keep_mem) {
1894 // First clone the existing memory state
1895 set_all_memory(keep_mem);
1896 if (hook_mem != NULL) {
1897 // Make memory for the call
1898 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1899 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1900 // We also use hook_mem to extract specific effects from arraycopy stubs.
1901 set_memory(mem, hook_mem);
1902 }
1903 // ...else the call has NO memory effects.
1904
1905 // Make sure the call advertises its memory effects precisely.
1906 // This lets us build accurate anti-dependences in gcm.cpp.
1907 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1908 "call node must be constructed correctly");
1909 } else {
1910 assert(hook_mem == NULL, "");
1911 // This is not a "slow path" call; all memory comes from the call.
1912 set_all_memory_call(call);
1913 }
1914 }
1915
1916
1917 // Replace the call with the current state of the kit.
1918 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1919 JVMState* ejvms = NULL;
1920 if (has_exceptions()) {
1921 ejvms = transfer_exceptions_into_jvms();
1922 }
1923
1924 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1925 ReplacedNodes replaced_nodes_exception;
1926 Node* ex_ctl = top();
1927
1928 SafePointNode* final_state = stop();
1929
1930 // Find all the needed outputs of this call
1931 CallProjections* callprojs = call->extract_projections(true);
1932
1933 Node* init_mem = call->in(TypeFunc::Memory);
1934 Node* final_mem = final_state->in(TypeFunc::Memory);
1935 Node* final_ctl = final_state->in(TypeFunc::Control);
1936 Node* final_io = final_state->in(TypeFunc::I_O);
1937
1938 // Replace all the old call edges with the edges from the inlining result
1939 if (callprojs->fallthrough_catchproj != NULL) {
1940 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
1941 }
1942 if (callprojs->fallthrough_memproj != NULL) {
1943 if (final_mem->is_MergeMem()) {
1944 // Parser's exits MergeMem was not transformed but may be optimized
1945 final_mem = _gvn.transform(final_mem);
1946 }
1947 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
1948 }
1949 if (callprojs->fallthrough_ioproj != NULL) {
1950 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
1951 }
1952
1953 // Replace the result with the new result if it exists and is used
1954 if (callprojs->resproj[0] != NULL && result != NULL) {
1955 assert(callprojs->nb_resproj == 1, "unexpected number of results");
1956 C->gvn_replace_by(callprojs->resproj[0], result);
1957 }
1958
1959 if (ejvms == NULL) {
1960 // No exception edges to simply kill off those paths
1961 if (callprojs->catchall_catchproj != NULL) {
1962 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
1963 }
1964 if (callprojs->catchall_memproj != NULL) {
1965 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
1966 }
1967 if (callprojs->catchall_ioproj != NULL) {
1968 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
1969 }
1970 // Replace the old exception object with top
1971 if (callprojs->exobj != NULL) {
1972 C->gvn_replace_by(callprojs->exobj, C->top());
1973 }
1974 } else {
1975 GraphKit ekit(ejvms);
1976
1977 // Load my combined exception state into the kit, with all phis transformed:
1978 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1979 replaced_nodes_exception = ex_map->replaced_nodes();
1980
1981 Node* ex_oop = ekit.use_exception_state(ex_map);
1982
1983 if (callprojs->catchall_catchproj != NULL) {
1984 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
1985 ex_ctl = ekit.control();
1986 }
1987 if (callprojs->catchall_memproj != NULL) {
1988 C->gvn_replace_by(callprojs->catchall_memproj, ekit.reset_memory());
1989 }
1990 if (callprojs->catchall_ioproj != NULL) {
1991 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
1992 }
1993
1994 // Replace the old exception object with the newly created one
1995 if (callprojs->exobj != NULL) {
1996 C->gvn_replace_by(callprojs->exobj, ex_oop);
1997 }
1998 }
1999
2000 // Disconnect the call from the graph
2001 call->disconnect_inputs(NULL, C);
2002 C->gvn_replace_by(call, C->top());
2003
2004 // Clean up any MergeMems that feed other MergeMems since the
2005 // optimizer doesn't like that.
2006 if (final_mem->is_MergeMem()) {
2007 Node_List wl;
2008 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
2009 Node* m = i.get();
2010 if (m->is_MergeMem() && !wl.contains(m)) {
2011 wl.push(m);
2012 }
2013 }
2014 while (wl.size() > 0) {
2015 _gvn.transform(wl.pop());
2016 }
2017 }
2018
2019 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2020 replaced_nodes.apply(C, final_ctl);
2021 }
2022 if (!ex_ctl->is_top() && do_replaced_nodes) {
2023 replaced_nodes_exception.apply(C, ex_ctl);
2024 }
2025 }
2026
2027
2028 //------------------------------increment_counter------------------------------
2029 // for statistics: increment a VM counter by 1
2030
2031 void GraphKit::increment_counter(address counter_addr) {
2032 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2033 increment_counter(adr1);
2034 }
2035
2036 void GraphKit::increment_counter(Node* counter_addr) {
2037 int adr_type = Compile::AliasIdxRaw;
2038 Node* ctrl = control();
2039 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2040 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
2041 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
2042 }
2043
2044
2045 //------------------------------uncommon_trap----------------------------------
2046 // Bail out to the interpreter in mid-method. Implemented by calling the
2047 // uncommon_trap blob. This helper function inserts a runtime call with the
2048 // right debug info.
2049 void GraphKit::uncommon_trap(int trap_request,
2050 ciKlass* klass, const char* comment,
2051 bool must_throw,
2052 bool keep_exact_action) {
2053 if (failing()) stop();
2054 if (stopped()) return; // trap reachable?
2055
2056 // Note: If ProfileTraps is true, and if a deopt. actually
2057 // occurs here, the runtime will make sure an MDO exists. There is
2058 // no need to call method()->ensure_method_data() at this point.
2059
2060 // Set the stack pointer to the right value for reexecution:
2061 set_sp(reexecute_sp());
2062
2063 #ifdef ASSERT
2064 if (!must_throw) {
2065 // Make sure the stack has at least enough depth to execute
2066 // the current bytecode.
2067 int inputs, ignored_depth;
2068 if (compute_stack_effects(inputs, ignored_depth)) {
2069 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2070 Bytecodes::name(java_bc()), sp(), inputs);
2071 }
2072 }
2073 #endif
2074
2075 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2076 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2077
2078 switch (action) {
2079 case Deoptimization::Action_maybe_recompile:
2080 case Deoptimization::Action_reinterpret:
2081 // Temporary fix for 6529811 to allow virtual calls to be sure they
2082 // get the chance to go from mono->bi->mega
2083 if (!keep_exact_action &&
2084 Deoptimization::trap_request_index(trap_request) < 0 &&
2085 too_many_recompiles(reason)) {
2086 // This BCI is causing too many recompilations.
2087 if (C->log() != NULL) {
2088 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2089 Deoptimization::trap_reason_name(reason),
2090 Deoptimization::trap_action_name(action));
2091 }
2092 action = Deoptimization::Action_none;
2093 trap_request = Deoptimization::make_trap_request(reason, action);
2094 } else {
2095 C->set_trap_can_recompile(true);
2096 }
2097 break;
2098 case Deoptimization::Action_make_not_entrant:
2099 C->set_trap_can_recompile(true);
2100 break;
2101 case Deoptimization::Action_none:
2102 case Deoptimization::Action_make_not_compilable:
2103 break;
2104 default:
2105 #ifdef ASSERT
2106 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2107 #endif
2108 break;
2109 }
2110
2111 if (TraceOptoParse) {
2112 char buf[100];
2113 tty->print_cr("Uncommon trap %s at bci:%d",
2114 Deoptimization::format_trap_request(buf, sizeof(buf),
2115 trap_request), bci());
2116 }
2117
2118 CompileLog* log = C->log();
2119 if (log != NULL) {
2120 int kid = (klass == NULL)? -1: log->identify(klass);
2121 log->begin_elem("uncommon_trap bci='%d'", bci());
2122 char buf[100];
2123 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2124 trap_request));
2125 if (kid >= 0) log->print(" klass='%d'", kid);
2126 if (comment != NULL) log->print(" comment='%s'", comment);
2127 log->end_elem();
2128 }
2129
2130 // Make sure any guarding test views this path as very unlikely
2131 Node *i0 = control()->in(0);
2132 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2133 IfNode *iff = i0->as_If();
2134 float f = iff->_prob; // Get prob
2135 if (control()->Opcode() == Op_IfTrue) {
2136 if (f > PROB_UNLIKELY_MAG(4))
2137 iff->_prob = PROB_MIN;
2138 } else {
2139 if (f < PROB_LIKELY_MAG(4))
2140 iff->_prob = PROB_MAX;
2141 }
2142 }
2143
2144 // Clear out dead values from the debug info.
2145 kill_dead_locals();
2146
2147 // Now insert the uncommon trap subroutine call
2148 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2149 const TypePtr* no_memory_effects = NULL;
2150 // Pass the index of the class to be loaded
2151 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2152 (must_throw ? RC_MUST_THROW : 0),
2153 OptoRuntime::uncommon_trap_Type(),
2154 call_addr, "uncommon_trap", no_memory_effects,
2155 intcon(trap_request));
2156 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2157 "must extract request correctly from the graph");
2158 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2159
2160 call->set_req(TypeFunc::ReturnAdr, returnadr());
2161 // The debug info is the only real input to this call.
2162
2163 // Halt-and-catch fire here. The above call should never return!
2164 HaltNode* halt = new HaltNode(control(), frameptr());
2165 _gvn.set_type_bottom(halt);
2166 root()->add_req(halt);
2167
2168 stop_and_kill_map();
2169 }
2170
2171
2172 //--------------------------just_allocated_object------------------------------
2173 // Report the object that was just allocated.
2174 // It must be the case that there are no intervening safepoints.
2175 // We use this to determine if an object is so "fresh" that
2176 // it does not require card marks.
2177 Node* GraphKit::just_allocated_object(Node* current_control) {
2178 if (C->recent_alloc_ctl() == current_control)
2179 return C->recent_alloc_obj();
2180 return NULL;
2181 }
2182
2183
2184 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2185 // (Note: TypeFunc::make has a cache that makes this fast.)
2186 const TypeFunc* tf = TypeFunc::make(dest_method);
2187 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2188 for (int j = 0; j < nargs; j++) {
2189 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2190 if( targ->basic_type() == T_DOUBLE ) {
2191 // If any parameters are doubles, they must be rounded before
2192 // the call, dstore_rounding does gvn.transform
2193 Node *arg = argument(j);
2194 arg = dstore_rounding(arg);
2195 set_argument(j, arg);
2196 }
2197 }
2198 }
2199
2200 /**
2201 * Record profiling data exact_kls for Node n with the type system so
2202 * that it can propagate it (speculation)
2203 *
2204 * @param n node that the type applies to
2205 * @param exact_kls type from profiling
2206 * @param maybe_null did profiling see null?
2207 *
2208 * @return node with improved type
2209 */
2210 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2211 const Type* current_type = _gvn.type(n);
2212 assert(UseTypeSpeculation, "type speculation must be on");
2213
2214 const TypePtr* speculative = current_type->speculative();
2215
2216 // Should the klass from the profile be recorded in the speculative type?
2217 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2218 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2219 const TypeOopPtr* xtype = tklass->as_instance_type();
2220 assert(xtype->klass_is_exact(), "Should be exact");
2221 // Any reason to believe n is not null (from this profiling or a previous one)?
2222 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2223 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2224 // record the new speculative type's depth
2225 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2226 speculative = speculative->with_inline_depth(jvms()->depth());
2227 } else if (current_type->would_improve_ptr(ptr_kind)) {
2228 // Profiling report that null was never seen so we can change the
2229 // speculative type to non null ptr.
2230 if (ptr_kind == ProfileAlwaysNull) {
2231 speculative = TypePtr::NULL_PTR;
2232 } else {
2233 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2234 const TypePtr* ptr = TypePtr::NOTNULL;
2235 if (speculative != NULL) {
2236 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2237 } else {
2238 speculative = ptr;
2239 }
2240 }
2241 }
2242
2243 if (speculative != current_type->speculative()) {
2244 // Build a type with a speculative type (what we think we know
2245 // about the type but will need a guard when we use it)
2246 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2247 // We're changing the type, we need a new CheckCast node to carry
2248 // the new type. The new type depends on the control: what
2249 // profiling tells us is only valid from here as far as we can
2250 // tell.
2251 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2252 cast = _gvn.transform(cast);
2253 replace_in_map(n, cast);
2254 n = cast;
2255 }
2256
2257 return n;
2258 }
2259
2260 /**
2261 * Record profiling data from receiver profiling at an invoke with the
2262 * type system so that it can propagate it (speculation)
2263 *
2264 * @param n receiver node
2265 *
2266 * @return node with improved type
2267 */
2268 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2269 if (!UseTypeSpeculation) {
2270 return n;
2271 }
2272 ciKlass* exact_kls = profile_has_unique_klass();
2273 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2274 if ((java_bc() == Bytecodes::_checkcast ||
2275 java_bc() == Bytecodes::_instanceof ||
2276 java_bc() == Bytecodes::_aastore) &&
2277 method()->method_data()->is_mature()) {
2278 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2279 if (data != NULL) {
2280 if (!data->as_BitData()->null_seen()) {
2281 ptr_kind = ProfileNeverNull;
2282 } else {
2283 assert(data->is_ReceiverTypeData(), "bad profile data type");
2284 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2285 uint i = 0;
2286 for (; i < call->row_limit(); i++) {
2287 ciKlass* receiver = call->receiver(i);
2288 if (receiver != NULL) {
2289 break;
2290 }
2291 }
2292 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2293 }
2294 }
2295 }
2296 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2297 }
2298
2299 /**
2300 * Record profiling data from argument profiling at an invoke with the
2301 * type system so that it can propagate it (speculation)
2302 *
2303 * @param dest_method target method for the call
2304 * @param bc what invoke bytecode is this?
2305 */
2306 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2307 if (!UseTypeSpeculation) {
2308 return;
2309 }
2310 const TypeFunc* tf = TypeFunc::make(dest_method);
2311 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2312 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2313 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2314 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2315 if (targ->isa_oopptr()) {
2316 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2317 ciKlass* better_type = NULL;
2318 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2319 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2320 }
2321 i++;
2322 }
2323 }
2324 }
2325
2326 /**
2327 * Record profiling data from parameter profiling at an invoke with
2328 * the type system so that it can propagate it (speculation)
2329 */
2330 void GraphKit::record_profiled_parameters_for_speculation() {
2331 if (!UseTypeSpeculation) {
2332 return;
2333 }
2334 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2335 if (_gvn.type(local(i))->isa_oopptr()) {
2336 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2337 ciKlass* better_type = NULL;
2338 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2339 record_profile_for_speculation(local(i), better_type, ptr_kind);
2340 }
2341 j++;
2342 }
2343 }
2344 }
2345
2346 /**
2347 * Record profiling data from return value profiling at an invoke with
2348 * the type system so that it can propagate it (speculation)
2349 */
2350 void GraphKit::record_profiled_return_for_speculation() {
2351 if (!UseTypeSpeculation) {
2352 return;
2353 }
2354 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2355 ciKlass* better_type = NULL;
2356 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2357 // If profiling reports a single type for the return value,
2358 // feed it to the type system so it can propagate it as a
2359 // speculative type
2360 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2361 }
2362 }
2363
2364 void GraphKit::round_double_result(ciMethod* dest_method) {
2365 // A non-strict method may return a double value which has an extended
2366 // exponent, but this must not be visible in a caller which is 'strict'
2367 // If a strict caller invokes a non-strict callee, round a double result
2368
2369 BasicType result_type = dest_method->return_type()->basic_type();
2370 assert( method() != NULL, "must have caller context");
2371 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2372 // Destination method's return value is on top of stack
2373 // dstore_rounding() does gvn.transform
2374 Node *result = pop_pair();
2375 result = dstore_rounding(result);
2376 push_pair(result);
2377 }
2378 }
2379
2380 // rounding for strict float precision conformance
2381 Node* GraphKit::precision_rounding(Node* n) {
2382 return UseStrictFP && _method->flags().is_strict()
2383 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2384 ? _gvn.transform( new RoundFloatNode(0, n) )
2385 : n;
2386 }
2387
2388 // rounding for strict double precision conformance
2389 Node* GraphKit::dprecision_rounding(Node *n) {
2390 return UseStrictFP && _method->flags().is_strict()
2391 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2392 ? _gvn.transform( new RoundDoubleNode(0, n) )
2393 : n;
2394 }
2395
2396 // rounding for non-strict double stores
2397 Node* GraphKit::dstore_rounding(Node* n) {
2398 return Matcher::strict_fp_requires_explicit_rounding
2399 && UseSSE <= 1
2400 ? _gvn.transform( new RoundDoubleNode(0, n) )
2401 : n;
2402 }
2403
2404 //=============================================================================
2405 // Generate a fast path/slow path idiom. Graph looks like:
2406 // [foo] indicates that 'foo' is a parameter
2407 //
2408 // [in] NULL
2409 // \ /
2410 // CmpP
2411 // Bool ne
2412 // If
2413 // / \
2414 // True False-<2>
2415 // / |
2416 // / cast_not_null
2417 // Load | | ^
2418 // [fast_test] | |
2419 // gvn to opt_test | |
2420 // / \ | <1>
2421 // True False |
2422 // | \\ |
2423 // [slow_call] \[fast_result]
2424 // Ctl Val \ \
2425 // | \ \
2426 // Catch <1> \ \
2427 // / \ ^ \ \
2428 // Ex No_Ex | \ \
2429 // | \ \ | \ <2> \
2430 // ... \ [slow_res] | | \ [null_result]
2431 // \ \--+--+--- | |
2432 // \ | / \ | /
2433 // --------Region Phi
2434 //
2435 //=============================================================================
2436 // Code is structured as a series of driver functions all called 'do_XXX' that
2437 // call a set of helper functions. Helper functions first, then drivers.
2438
2439 //------------------------------null_check_oop---------------------------------
2440 // Null check oop. Set null-path control into Region in slot 3.
2441 // Make a cast-not-nullness use the other not-null control. Return cast.
2442 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2443 bool never_see_null,
2444 bool safe_for_replace,
2445 bool speculative) {
2446 // Initial NULL check taken path
2447 (*null_control) = top();
2448 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2449
2450 // Generate uncommon_trap:
2451 if (never_see_null && (*null_control) != top()) {
2452 // If we see an unexpected null at a check-cast we record it and force a
2453 // recompile; the offending check-cast will be compiled to handle NULLs.
2454 // If we see more than one offending BCI, then all checkcasts in the
2455 // method will be compiled to handle NULLs.
2456 PreserveJVMState pjvms(this);
2457 set_control(*null_control);
2458 replace_in_map(value, null());
2459 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2460 uncommon_trap(reason,
2461 Deoptimization::Action_make_not_entrant);
2462 (*null_control) = top(); // NULL path is dead
2463 }
2464 if ((*null_control) == top() && safe_for_replace) {
2465 replace_in_map(value, cast);
2466 }
2467
2468 // Cast away null-ness on the result
2469 return cast;
2470 }
2471
2472 //------------------------------opt_iff----------------------------------------
2473 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2474 // Return slow-path control.
2475 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2476 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2477
2478 // Fast path taken; set region slot 2
2479 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2480 region->init_req(2,fast_taken); // Capture fast-control
2481
2482 // Fast path not-taken, i.e. slow path
2483 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2484 return slow_taken;
2485 }
2486
2487 //-----------------------------make_runtime_call-------------------------------
2488 Node* GraphKit::make_runtime_call(int flags,
2489 const TypeFunc* call_type, address call_addr,
2490 const char* call_name,
2491 const TypePtr* adr_type,
2492 // The following parms are all optional.
2493 // The first NULL ends the list.
2494 Node* parm0, Node* parm1,
2495 Node* parm2, Node* parm3,
2496 Node* parm4, Node* parm5,
2497 Node* parm6, Node* parm7) {
2498 // Slow-path call
2499 bool is_leaf = !(flags & RC_NO_LEAF);
2500 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2501 if (call_name == NULL) {
2502 assert(!is_leaf, "must supply name for leaf");
2503 call_name = OptoRuntime::stub_name(call_addr);
2504 }
2505 CallNode* call;
2506 if (!is_leaf) {
2507 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2508 bci(), adr_type);
2509 } else if (flags & RC_NO_FP) {
2510 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2511 } else {
2512 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2513 }
2514
2515 // The following is similar to set_edges_for_java_call,
2516 // except that the memory effects of the call are restricted to AliasIdxRaw.
2517
2518 // Slow path call has no side-effects, uses few values
2519 bool wide_in = !(flags & RC_NARROW_MEM);
2520 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2521
2522 Node* prev_mem = NULL;
2523 if (wide_in) {
2524 prev_mem = set_predefined_input_for_runtime_call(call);
2525 } else {
2526 assert(!wide_out, "narrow in => narrow out");
2527 Node* narrow_mem = memory(adr_type);
2528 prev_mem = reset_memory();
2529 map()->set_memory(narrow_mem);
2530 set_predefined_input_for_runtime_call(call);
2531 }
2532
2533 // Hook each parm in order. Stop looking at the first NULL.
2534 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2535 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2536 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2537 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2538 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2539 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2540 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2541 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2542 /* close each nested if ===> */ } } } } } } } }
2543 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2544
2545 if (!is_leaf) {
2546 // Non-leaves can block and take safepoints:
2547 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2548 }
2549 // Non-leaves can throw exceptions:
2550 if (has_io) {
2551 call->set_req(TypeFunc::I_O, i_o());
2552 }
2553
2554 if (flags & RC_UNCOMMON) {
2555 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2556 // (An "if" probability corresponds roughly to an unconditional count.
2557 // Sort of.)
2558 call->set_cnt(PROB_UNLIKELY_MAG(4));
2559 }
2560
2561 Node* c = _gvn.transform(call);
2562 assert(c == call, "cannot disappear");
2563
2564 if (wide_out) {
2565 // Slow path call has full side-effects.
2566 set_predefined_output_for_runtime_call(call);
2567 } else {
2568 // Slow path call has few side-effects, and/or sets few values.
2569 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2570 }
2571
2572 if (has_io) {
2573 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2574 }
2575 return call;
2576
2577 }
2578
2579 //------------------------------merge_memory-----------------------------------
2580 // Merge memory from one path into the current memory state.
2581 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2582 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2583 Node* old_slice = mms.force_memory();
2584 Node* new_slice = mms.memory2();
2585 if (old_slice != new_slice) {
2586 PhiNode* phi;
2587 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2588 if (mms.is_empty()) {
2589 // clone base memory Phi's inputs for this memory slice
2590 assert(old_slice == mms.base_memory(), "sanity");
2591 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2592 _gvn.set_type(phi, Type::MEMORY);
2593 for (uint i = 1; i < phi->req(); i++) {
2594 phi->init_req(i, old_slice->in(i));
2595 }
2596 } else {
2597 phi = old_slice->as_Phi(); // Phi was generated already
2598 }
2599 } else {
2600 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2601 _gvn.set_type(phi, Type::MEMORY);
2602 }
2603 phi->set_req(new_path, new_slice);
2604 mms.set_memory(phi);
2605 }
2606 }
2607 }
2608
2609 //------------------------------make_slow_call_ex------------------------------
2610 // Make the exception handler hookups for the slow call
2611 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2612 if (stopped()) return;
2613
2614 // Make a catch node with just two handlers: fall-through and catch-all
2615 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2616 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2617 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2618 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2619
2620 { PreserveJVMState pjvms(this);
2621 set_control(excp);
2622 set_i_o(i_o);
2623
2624 if (excp != top()) {
2625 if (deoptimize) {
2626 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2627 uncommon_trap(Deoptimization::Reason_unhandled,
2628 Deoptimization::Action_none);
2629 } else {
2630 // Create an exception state also.
2631 // Use an exact type if the caller has specified a specific exception.
2632 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2633 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2634 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2635 }
2636 }
2637 }
2638
2639 // Get the no-exception control from the CatchNode.
2640 set_control(norm);
2641 }
2642
2643 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2644 Node* cmp = NULL;
2645 switch(bt) {
2646 case T_INT: cmp = new CmpINode(in1, in2); break;
2647 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2648 default: fatal("unexpected comparison type %s", type2name(bt));
2649 }
2650 gvn->transform(cmp);
2651 Node* bol = gvn->transform(new BoolNode(cmp, test));
2652 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2653 gvn->transform(iff);
2654 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2655 return iff;
2656 }
2657
2658
2659 //-------------------------------gen_subtype_check-----------------------------
2660 // Generate a subtyping check. Takes as input the subtype and supertype.
2661 // Returns 2 values: sets the default control() to the true path and returns
2662 // the false path. Only reads invariant memory; sets no (visible) memory.
2663 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2664 // but that's not exposed to the optimizer. This call also doesn't take in an
2665 // Object; if you wish to check an Object you need to load the Object's class
2666 // prior to coming here.
2667 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2668 Compile* C = gvn->C;
2669
2670 if ((*ctrl)->is_top()) {
2671 return C->top();
2672 }
2673
2674 // Fast check for identical types, perhaps identical constants.
2675 // The types can even be identical non-constants, in cases
2676 // involving Array.newInstance, Object.clone, etc.
2677 if (subklass == superklass)
2678 return C->top(); // false path is dead; no test needed.
2679
2680 if (gvn->type(superklass)->singleton()) {
2681 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2682 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2683
2684 // In the common case of an exact superklass, try to fold up the
2685 // test before generating code. You may ask, why not just generate
2686 // the code and then let it fold up? The answer is that the generated
2687 // code will necessarily include null checks, which do not always
2688 // completely fold away. If they are also needless, then they turn
2689 // into a performance loss. Example:
2690 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2691 // Here, the type of 'fa' is often exact, so the store check
2692 // of fa[1]=x will fold up, without testing the nullness of x.
2693 switch (C->static_subtype_check(superk, subk)) {
2694 case Compile::SSC_always_false:
2695 {
2696 Node* always_fail = *ctrl;
2697 *ctrl = gvn->C->top();
2698 return always_fail;
2699 }
2700 case Compile::SSC_always_true:
2701 return C->top();
2702 case Compile::SSC_easy_test:
2703 {
2704 // Just do a direct pointer compare and be done.
2705 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2706 *ctrl = gvn->transform(new IfTrueNode(iff));
2707 return gvn->transform(new IfFalseNode(iff));
2708 }
2709 case Compile::SSC_full_test:
2710 break;
2711 default:
2712 ShouldNotReachHere();
2713 }
2714 }
2715
2716 // %%% Possible further optimization: Even if the superklass is not exact,
2717 // if the subklass is the unique subtype of the superklass, the check
2718 // will always succeed. We could leave a dependency behind to ensure this.
2719
2720 // First load the super-klass's check-offset
2721 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2722 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2723 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2724 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2725 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2726
2727 // Load from the sub-klass's super-class display list, or a 1-word cache of
2728 // the secondary superclass list, or a failing value with a sentinel offset
2729 // if the super-klass is an interface or exceptionally deep in the Java
2730 // hierarchy and we have to scan the secondary superclass list the hard way.
2731 // Worst-case type is a little odd: NULL is allowed as a result (usually
2732 // klass loads can never produce a NULL).
2733 Node *chk_off_X = chk_off;
2734 #ifdef _LP64
2735 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2736 #endif
2737 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2738 // For some types like interfaces the following loadKlass is from a 1-word
2739 // cache which is mutable so can't use immutable memory. Other
2740 // types load from the super-class display table which is immutable.
2741 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2742 Node *kmem = might_be_cache ? m : C->immutable_memory();
2743 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::BOTTOM));
2744
2745 // Compile speed common case: ARE a subtype and we canNOT fail
2746 if( superklass == nkls )
2747 return C->top(); // false path is dead; no test needed.
2748
2749 // See if we get an immediate positive hit. Happens roughly 83% of the
2750 // time. Test to see if the value loaded just previously from the subklass
2751 // is exactly the superklass.
2752 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2753 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2754 *ctrl = gvn->transform(new IfFalseNode(iff1));
2755
2756 // Compile speed common case: Check for being deterministic right now. If
2757 // chk_off is a constant and not equal to cacheoff then we are NOT a
2758 // subklass. In this case we need exactly the 1 test above and we can
2759 // return those results immediately.
2760 if (!might_be_cache) {
2761 Node* not_subtype_ctrl = *ctrl;
2762 *ctrl = iftrue1; // We need exactly the 1 test above
2763 return not_subtype_ctrl;
2764 }
2765
2766 // Gather the various success & failures here
2767 RegionNode *r_ok_subtype = new RegionNode(4);
2768 gvn->record_for_igvn(r_ok_subtype);
2769 RegionNode *r_not_subtype = new RegionNode(3);
2770 gvn->record_for_igvn(r_not_subtype);
2771
2772 r_ok_subtype->init_req(1, iftrue1);
2773
2774 // Check for immediate negative hit. Happens roughly 11% of the time (which
2775 // is roughly 63% of the remaining cases). Test to see if the loaded
2776 // check-offset points into the subklass display list or the 1-element
2777 // cache. If it points to the display (and NOT the cache) and the display
2778 // missed then it's not a subtype.
2779 Node *cacheoff = gvn->intcon(cacheoff_con);
2780 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2781 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2782 *ctrl = gvn->transform(new IfFalseNode(iff2));
2783
2784 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2785 // No performance impact (too rare) but allows sharing of secondary arrays
2786 // which has some footprint reduction.
2787 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2788 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2789 *ctrl = gvn->transform(new IfFalseNode(iff3));
2790
2791 // -- Roads not taken here: --
2792 // We could also have chosen to perform the self-check at the beginning
2793 // of this code sequence, as the assembler does. This would not pay off
2794 // the same way, since the optimizer, unlike the assembler, can perform
2795 // static type analysis to fold away many successful self-checks.
2796 // Non-foldable self checks work better here in second position, because
2797 // the initial primary superclass check subsumes a self-check for most
2798 // types. An exception would be a secondary type like array-of-interface,
2799 // which does not appear in its own primary supertype display.
2800 // Finally, we could have chosen to move the self-check into the
2801 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2802 // dependent manner. But it is worthwhile to have the check here,
2803 // where it can be perhaps be optimized. The cost in code space is
2804 // small (register compare, branch).
2805
2806 // Now do a linear scan of the secondary super-klass array. Again, no real
2807 // performance impact (too rare) but it's gotta be done.
2808 // Since the code is rarely used, there is no penalty for moving it
2809 // out of line, and it can only improve I-cache density.
2810 // The decision to inline or out-of-line this final check is platform
2811 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2812 Node* psc = gvn->transform(
2813 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2814
2815 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2816 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2817 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2818
2819 // Return false path; set default control to true path.
2820 *ctrl = gvn->transform(r_ok_subtype);
2821 return gvn->transform(r_not_subtype);
2822 }
2823
2824 // Profile-driven exact type check:
2825 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2826 float prob,
2827 Node* *casted_receiver) {
2828 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2829 Node* recv_klass = load_object_klass(receiver);
2830 Node* fail = type_check(recv_klass, tklass, prob);
2831 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2832 assert(recv_xtype->klass_is_exact(), "");
2833
2834 // Subsume downstream occurrences of receiver with a cast to
2835 // recv_xtype, since now we know what the type will be.
2836 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2837 Node* res = _gvn.transform(cast);
2838 if (recv_xtype->is_valuetypeptr()) {
2839 assert(!gvn().type(res)->is_ptr()->maybe_null(), "should never be null");
2840 res = ValueTypeNode::make_from_oop(this, res, recv_xtype->value_klass());
2841 }
2842
2843 (*casted_receiver) = res;
2844 // (User must make the replace_in_map call.)
2845
2846 return fail;
2847 }
2848
2849 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
2850 float prob) {
2851 Node* want_klass = makecon(tklass);
2852 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass));
2853 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2854 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2855 set_control( _gvn.transform( new IfTrueNode (iff)));
2856 Node* fail = _gvn.transform( new IfFalseNode(iff));
2857 return fail;
2858 }
2859
2860
2861 //------------------------------seems_never_null-------------------------------
2862 // Use null_seen information if it is available from the profile.
2863 // If we see an unexpected null at a type check we record it and force a
2864 // recompile; the offending check will be recompiled to handle NULLs.
2865 // If we see several offending BCIs, then all checks in the
2866 // method will be recompiled.
2867 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2868 speculating = !_gvn.type(obj)->speculative_maybe_null();
2869 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2870 if (UncommonNullCast // Cutout for this technique
2871 && obj != null() // And not the -Xcomp stupid case?
2872 && !too_many_traps(reason)
2873 ) {
2874 if (speculating) {
2875 return true;
2876 }
2877 if (data == NULL)
2878 // Edge case: no mature data. Be optimistic here.
2879 return true;
2880 // If the profile has not seen a null, assume it won't happen.
2881 assert(java_bc() == Bytecodes::_checkcast ||
2882 java_bc() == Bytecodes::_instanceof ||
2883 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2884 return !data->as_BitData()->null_seen();
2885 }
2886 speculating = false;
2887 return false;
2888 }
2889
2890 //------------------------maybe_cast_profiled_receiver-------------------------
2891 // If the profile has seen exactly one type, narrow to exactly that type.
2892 // Subsequent type checks will always fold up.
2893 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2894 ciKlass* require_klass,
2895 ciKlass* spec_klass,
2896 bool safe_for_replace) {
2897 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2898
2899 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2900
2901 // Make sure we haven't already deoptimized from this tactic.
2902 if (too_many_traps(reason) || too_many_recompiles(reason))
2903 return NULL;
2904
2905 // (No, this isn't a call, but it's enough like a virtual call
2906 // to use the same ciMethod accessor to get the profile info...)
2907 // If we have a speculative type use it instead of profiling (which
2908 // may not help us)
2909 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2910 if (exact_kls != NULL) {// no cast failures here
2911 if (require_klass == NULL ||
2912 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2913 // If we narrow the type to match what the type profile sees or
2914 // the speculative type, we can then remove the rest of the
2915 // cast.
2916 // This is a win, even if the exact_kls is very specific,
2917 // because downstream operations, such as method calls,
2918 // will often benefit from the sharper type.
2919 Node* exact_obj = not_null_obj; // will get updated in place...
2920 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2921 &exact_obj);
2922 { PreserveJVMState pjvms(this);
2923 set_control(slow_ctl);
2924 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2925 }
2926 if (safe_for_replace) {
2927 replace_in_map(not_null_obj, exact_obj);
2928 }
2929 return exact_obj;
2930 }
2931 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2932 }
2933
2934 return NULL;
2935 }
2936
2937 /**
2938 * Cast obj to type and emit guard unless we had too many traps here
2939 * already
2940 *
2941 * @param obj node being casted
2942 * @param type type to cast the node to
2943 * @param not_null true if we know node cannot be null
2944 */
2945 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2946 ciKlass* type,
2947 bool not_null) {
2948 if (stopped()) {
2949 return obj;
2950 }
2951
2952 // type == NULL if profiling tells us this object is always null
2953 if (type != NULL) {
2954 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2955 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2956
2957 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2958 !too_many_traps(class_reason) &&
2959 !too_many_recompiles(class_reason)) {
2960 Node* not_null_obj = NULL;
2961 // not_null is true if we know the object is not null and
2962 // there's no need for a null check
2963 if (!not_null) {
2964 Node* null_ctl = top();
2965 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2966 assert(null_ctl->is_top(), "no null control here");
2967 } else {
2968 not_null_obj = obj;
2969 }
2970
2971 Node* exact_obj = not_null_obj;
2972 ciKlass* exact_kls = type;
2973 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2974 &exact_obj);
2975 {
2976 PreserveJVMState pjvms(this);
2977 set_control(slow_ctl);
2978 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2979 }
2980 replace_in_map(not_null_obj, exact_obj);
2981 obj = exact_obj;
2982 }
2983 } else {
2984 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2985 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2986 Node* exact_obj = null_assert(obj);
2987 replace_in_map(obj, exact_obj);
2988 obj = exact_obj;
2989 }
2990 }
2991 return obj;
2992 }
2993
2994 //-------------------------------gen_instanceof--------------------------------
2995 // Generate an instance-of idiom. Used by both the instance-of bytecode
2996 // and the reflective instance-of call.
2997 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2998 kill_dead_locals(); // Benefit all the uncommon traps
2999 assert( !stopped(), "dead parse path should be checked in callers" );
3000 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3001 "must check for not-null not-dead klass in callers");
3002
3003 // Make the merge point
3004 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3005 RegionNode* region = new RegionNode(PATH_LIMIT);
3006 Node* phi = new PhiNode(region, TypeInt::BOOL);
3007 C->set_has_split_ifs(true); // Has chance for split-if optimization
3008
3009 ciProfileData* data = NULL;
3010 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3011 data = method()->method_data()->bci_to_data(bci());
3012 }
3013 bool speculative_not_null = false;
3014 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3015 && seems_never_null(obj, data, speculative_not_null));
3016 bool is_value = obj->is_ValueType();
3017
3018 // Null check; get casted pointer; set region slot 3
3019 Node* null_ctl = top();
3020 Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3021
3022 // If not_null_obj is dead, only null-path is taken
3023 if (stopped()) { // Doing instance-of on a NULL?
3024 set_control(null_ctl);
3025 return intcon(0);
3026 }
3027 region->init_req(_null_path, null_ctl);
3028 phi ->init_req(_null_path, intcon(0)); // Set null path value
3029 if (null_ctl == top()) {
3030 // Do this eagerly, so that pattern matches like is_diamond_phi
3031 // will work even during parsing.
3032 assert(_null_path == PATH_LIMIT-1, "delete last");
3033 region->del_req(_null_path);
3034 phi ->del_req(_null_path);
3035 }
3036
3037 // Do we know the type check always succeed?
3038 if (!is_value) {
3039 bool known_statically = false;
3040 if (_gvn.type(superklass)->singleton()) {
3041 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3042 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3043 if (subk != NULL && subk->is_loaded()) {
3044 int static_res = C->static_subtype_check(superk, subk);
3045 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3046 }
3047 }
3048
3049 if (!known_statically) {
3050 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3051 // We may not have profiling here or it may not help us. If we
3052 // have a speculative type use it to perform an exact cast.
3053 ciKlass* spec_obj_type = obj_type->speculative_type();
3054 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3055 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3056 if (stopped()) { // Profile disagrees with this path.
3057 set_control(null_ctl); // Null is the only remaining possibility.
3058 return intcon(0);
3059 }
3060 if (cast_obj != NULL &&
3061 // A value that's sometimes null is not something we can optimize well
3062 !(cast_obj->is_ValueType() && null_ctl != top())) {
3063 not_null_obj = cast_obj;
3064 is_value = not_null_obj->is_ValueType();
3065 }
3066 }
3067 }
3068 }
3069
3070 // Load the object's klass
3071 Node* obj_klass = NULL;
3072 if (is_value) {
3073 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->is_valuetype()->value_klass()));
3074 } else {
3075 obj_klass = load_object_klass(not_null_obj);
3076 }
3077
3078 // Generate the subtype check
3079 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3080
3081 // Plug in the success path to the general merge in slot 1.
3082 region->init_req(_obj_path, control());
3083 phi ->init_req(_obj_path, intcon(1));
3084
3085 // Plug in the failing path to the general merge in slot 2.
3086 region->init_req(_fail_path, not_subtype_ctrl);
3087 phi ->init_req(_fail_path, intcon(0));
3088
3089 // Return final merged results
3090 set_control( _gvn.transform(region) );
3091 record_for_igvn(region);
3092
3093 // If we know the type check always succeeds then we don't use the
3094 // profiling data at this bytecode. Don't lose it, feed it to the
3095 // type system as a speculative type.
3096 if (safe_for_replace && !is_value) {
3097 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3098 replace_in_map(obj, casted_obj);
3099 }
3100
3101 return _gvn.transform(phi);
3102 }
3103
3104 //-------------------------------gen_checkcast---------------------------------
3105 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3106 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3107 // uncommon-trap paths work. Adjust stack after this call.
3108 // If failure_control is supplied and not null, it is filled in with
3109 // the control edge for the cast failure. Otherwise, an appropriate
3110 // uncommon trap or exception is thrown.
3111 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3112 Node* *failure_control) {
3113 kill_dead_locals(); // Benefit all the uncommon traps
3114 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr();
3115 const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass());
3116
3117 // Handle value types
3118 if (obj->is_ValueType()) {
3119 ValueTypeNode* vt = obj->as_ValueType();
3120 if (toop->is_valuetypeptr()) {
3121 ciValueKlass* vk = vt->type()->is_valuetype()->value_klass();
3122 if (toop->klass() == vk) {
3123 return obj;
3124 } else {
3125 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(vk)));
3126 return top();
3127 }
3128 } else {
3129 vt = vt->allocate(this, true)->as_ValueType();
3130 obj = ValueTypePtrNode::make_from_value_type(_gvn, vt);
3131 }
3132 }
3133
3134 // Fast cutout: Check the case that the cast is vacuously true.
3135 // This detects the common cases where the test will short-circuit
3136 // away completely. We do this before we perform the null check,
3137 // because if the test is going to turn into zero code, we don't
3138 // want a residual null check left around. (Causes a slowdown,
3139 // for example, in some objArray manipulations, such as a[i]=a[j].)
3140 if (tk->singleton()) {
3141 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3142 if (objtp != NULL && objtp->klass() != NULL) {
3143 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3144 case Compile::SSC_always_true:
3145 // If we know the type check always succeed then we don't use
3146 // the profiling data at this bytecode. Don't lose it, feed it
3147 // to the type system as a speculative type.
3148 obj = record_profiled_receiver_for_speculation(obj);
3149 if (toop->is_valuetypeptr()) {
3150 obj = ValueTypeNode::make_from_oop(this, obj, toop->value_klass(), /* buffer_check */ false, /* null2default */ false);
3151 }
3152 return obj;
3153 case Compile::SSC_always_false:
3154 // It needs a null check because a null will *pass* the cast check.
3155 // A non-null value will always produce an exception.
3156 return null_assert(obj);
3157 }
3158 }
3159 }
3160
3161 ciProfileData* data = NULL;
3162 bool safe_for_replace = false;
3163 if (failure_control == NULL) { // use MDO in regular case only
3164 assert(java_bc() == Bytecodes::_aastore ||
3165 java_bc() == Bytecodes::_checkcast,
3166 "interpreter profiles type checks only for these BCs");
3167 data = method()->method_data()->bci_to_data(bci());
3168 safe_for_replace = true;
3169 }
3170
3171 // Make the merge point
3172 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3173 RegionNode* region = new RegionNode(PATH_LIMIT);
3174 Node* phi = new PhiNode(region, toop);
3175 C->set_has_split_ifs(true); // Has chance for split-if optimization
3176
3177 // Use null-cast information if it is available
3178 bool speculative_not_null = false;
3179 bool never_see_null = ((failure_control == NULL) // regular case only
3180 && seems_never_null(obj, data, speculative_not_null));
3181
3182 // Null check; get casted pointer; set region slot 3
3183 Node* null_ctl = top();
3184 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3185
3186 // If not_null_obj is dead, only null-path is taken
3187 if (stopped()) { // Doing instance-of on a NULL?
3188 set_control(null_ctl);
3189 return null();
3190 }
3191 region->init_req(_null_path, null_ctl);
3192 phi ->init_req(_null_path, null()); // Set null path value
3193 if (null_ctl == top()) {
3194 // Do this eagerly, so that pattern matches like is_diamond_phi
3195 // will work even during parsing.
3196 assert(_null_path == PATH_LIMIT-1, "delete last");
3197 region->del_req(_null_path);
3198 phi ->del_req(_null_path);
3199 }
3200
3201 Node* cast_obj = NULL;
3202 if (tk->klass_is_exact()) {
3203 // The following optimization tries to statically cast the speculative type of the object
3204 // (for example obtained during profiling) to the type of the superklass and then do a
3205 // dynamic check that the type of the object is what we expect. To work correctly
3206 // for checkcast and aastore the type of superklass should be exact.
3207 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3208 // We may not have profiling here or it may not help us. If we have
3209 // a speculative type use it to perform an exact cast.
3210 ciKlass* spec_obj_type = obj_type->speculative_type();
3211 if (spec_obj_type != NULL || data != NULL) {
3212 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3213 if (cast_obj != NULL && cast_obj->is_ValueType()) {
3214 if (null_ctl != top()) {
3215 cast_obj = NULL; // A value that's sometimes null is not something we can optimize well
3216 } else {
3217 return cast_obj;
3218 }
3219 }
3220 if (cast_obj != NULL) {
3221 if (failure_control != NULL) // failure is now impossible
3222 (*failure_control) = top();
3223 // adjust the type of the phi to the exact klass:
3224 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3225 }
3226 }
3227 }
3228
3229 if (cast_obj == NULL) {
3230 // Load the object's klass
3231 Node* obj_klass = load_object_klass(not_null_obj);
3232
3233 // Generate the subtype check
3234 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3235
3236 // Plug in success path into the merge
3237 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3238 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3239 if (failure_control == NULL) {
3240 if (not_subtype_ctrl != top()) { // If failure is possible
3241 PreserveJVMState pjvms(this);
3242 set_control(not_subtype_ctrl);
3243 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3244 }
3245 } else {
3246 (*failure_control) = not_subtype_ctrl;
3247 }
3248 }
3249
3250 region->init_req(_obj_path, control());
3251 phi ->init_req(_obj_path, cast_obj);
3252
3253 // A merge of NULL or Casted-NotNull obj
3254 Node* res = _gvn.transform(phi);
3255
3256 // Note I do NOT always 'replace_in_map(obj,result)' here.
3257 // if( tk->klass()->can_be_primary_super() )
3258 // This means that if I successfully store an Object into an array-of-String
3259 // I 'forget' that the Object is really now known to be a String. I have to
3260 // do this because we don't have true union types for interfaces - if I store
3261 // a Baz into an array-of-Interface and then tell the optimizer it's an
3262 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3263 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3264 // replace_in_map( obj, res );
3265
3266 // Return final merged results
3267 set_control( _gvn.transform(region) );
3268 record_for_igvn(region);
3269
3270 res = record_profiled_receiver_for_speculation(res);
3271 if (toop->is_valuetypeptr()) {
3272 res = ValueTypeNode::make_from_oop(this, res, toop->value_klass(), /* buffer_check */ false, /* null2default */ false);
3273 }
3274 return res;
3275 }
3276
3277 // Deoptimize if 'ary' is flattened or if 'obj' is null and 'ary' is a value type array
3278 void GraphKit::gen_value_type_array_guard(Node* ary, Node* obj, Node* elem_klass) {
3279 assert(EnableValhalla, "should only be used if value types are enabled");
3280 if (elem_klass == NULL) {
3281 // Load array element klass
3282 Node* kls = load_object_klass(ary);
3283 Node* k_adr = basic_plus_adr(kls, kls, in_bytes(ArrayKlass::element_klass_offset()));
3284 elem_klass = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
3285 }
3286 // Check if element is a value type
3287 Node* flags_addr = basic_plus_adr(elem_klass, in_bytes(Klass::access_flags_offset()));
3288 Node* flags = make_load(NULL, flags_addr, TypeInt::INT, T_INT, MemNode::unordered);
3289 Node* is_value_elem = _gvn.transform(new AndINode(flags, intcon(JVM_ACC_VALUE)));
3290
3291 const Type* objtype = _gvn.type(obj);
3292 if (objtype == TypePtr::NULL_PTR) {
3293 // Object is always null, check if array is a value type array
3294 Node* cmp = _gvn.transform(new CmpINode(is_value_elem, intcon(0)));
3295 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3296 { BuildCutout unless(this, bol, PROB_MAX);
3297 // TODO just deoptimize for now if we store null to a value type array
3298 uncommon_trap(Deoptimization::Reason_array_check,
3299 Deoptimization::Action_none);
3300 }
3301 } else {
3302 // Check if array is flattened or if we are storing null to a value type array
3303 // TODO can we merge these checks?
3304 gen_flattened_array_guard(ary);
3305 if (objtype->meet(TypePtr::NULL_PTR) == objtype) {
3306 // Check if (is_value_elem && obj_is_null) <=> (!is_value_elem | !obj_is_null == 0)
3307 // TODO what if we later figure out that obj is never null?
3308 Node* not_value = _gvn.transform(new XorINode(is_value_elem, intcon(JVM_ACC_VALUE)));
3309 not_value = _gvn.transform(new ConvI2LNode(not_value));
3310 Node* not_null = _gvn.transform(new CastP2XNode(NULL, obj));
3311 Node* both = _gvn.transform(new OrLNode(not_null, not_value));
3312 Node* cmp = _gvn.transform(new CmpLNode(both, longcon(0)));
3313 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3314 { BuildCutout unless(this, bol, PROB_MAX);
3315 // TODO just deoptimize for now if we store null to a value type array
3316 uncommon_trap(Deoptimization::Reason_array_check,
3317 Deoptimization::Action_none);
3318 }
3319 }
3320 }
3321 }
3322
3323 // Deoptimize if 'ary' is a flattened value type array
3324 void GraphKit::gen_flattened_array_guard(Node* ary, int nargs) {
3325 assert(EnableValhalla, "should only be used if value types are enabled");
3326 if (ValueArrayFlatten) {
3327 // Cannot statically determine if array is flattened, emit runtime check
3328 Node* kls = load_object_klass(ary);
3329 Node* lhp = basic_plus_adr(kls, kls, in_bytes(Klass::layout_helper_offset()));
3330 Node* layout_val = make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3331 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3332 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(Klass::_lh_array_tag_vt_value)));
3333 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3334
3335 { BuildCutout unless(this, bol, PROB_MAX);
3336 // TODO just deoptimize for now if value type array is flattened
3337 inc_sp(nargs);
3338 uncommon_trap(Deoptimization::Reason_array_check,
3339 Deoptimization::Action_none);
3340 }
3341 }
3342 }
3343
3344 //------------------------------next_monitor-----------------------------------
3345 // What number should be given to the next monitor?
3346 int GraphKit::next_monitor() {
3347 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3348 int next = current + C->sync_stack_slots();
3349 // Keep the toplevel high water mark current:
3350 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3351 return current;
3352 }
3353
3354 //------------------------------insert_mem_bar---------------------------------
3355 // Memory barrier to avoid floating things around
3356 // The membar serves as a pinch point between both control and all memory slices.
3357 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3358 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3359 mb->init_req(TypeFunc::Control, control());
3360 mb->init_req(TypeFunc::Memory, reset_memory());
3361 Node* membar = _gvn.transform(mb);
3362 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3363 set_all_memory_call(membar);
3364 return membar;
3365 }
3366
3367 //-------------------------insert_mem_bar_volatile----------------------------
3368 // Memory barrier to avoid floating things around
3369 // The membar serves as a pinch point between both control and memory(alias_idx).
3370 // If you want to make a pinch point on all memory slices, do not use this
3371 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3372 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3373 // When Parse::do_put_xxx updates a volatile field, it appends a series
3374 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3375 // The first membar is on the same memory slice as the field store opcode.
3376 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3377 // All the other membars (for other volatile slices, including AliasIdxBot,
3378 // which stands for all unknown volatile slices) are control-dependent
3379 // on the first membar. This prevents later volatile loads or stores
3380 // from sliding up past the just-emitted store.
3381
3382 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3383 mb->set_req(TypeFunc::Control,control());
3384 if (alias_idx == Compile::AliasIdxBot) {
3385 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3386 } else {
3387 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3388 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3389 }
3390 Node* membar = _gvn.transform(mb);
3391 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3392 if (alias_idx == Compile::AliasIdxBot) {
3393 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3394 } else {
3395 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3396 }
3397 return membar;
3398 }
3399
3400 //------------------------------shared_lock------------------------------------
3401 // Emit locking code.
3402 FastLockNode* GraphKit::shared_lock(Node* obj) {
3403 // bci is either a monitorenter bc or InvocationEntryBci
3404 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3405 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3406
3407 if( !GenerateSynchronizationCode )
3408 return NULL; // Not locking things?
3409 if (stopped()) // Dead monitor?
3410 return NULL;
3411
3412 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3413
3414 // Box the stack location
3415 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3416 Node* mem = reset_memory();
3417
3418 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3419 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3420 // Create the counters for this fast lock.
3421 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3422 }
3423
3424 // Create the rtm counters for this fast lock if needed.
3425 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3426
3427 // Add monitor to debug info for the slow path. If we block inside the
3428 // slow path and de-opt, we need the monitor hanging around
3429 map()->push_monitor( flock );
3430
3431 const TypeFunc *tf = LockNode::lock_type();
3432 LockNode *lock = new LockNode(C, tf);
3433
3434 lock->init_req( TypeFunc::Control, control() );
3435 lock->init_req( TypeFunc::Memory , mem );
3436 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3437 lock->init_req( TypeFunc::FramePtr, frameptr() );
3438 lock->init_req( TypeFunc::ReturnAdr, top() );
3439
3440 lock->init_req(TypeFunc::Parms + 0, obj);
3441 lock->init_req(TypeFunc::Parms + 1, box);
3442 lock->init_req(TypeFunc::Parms + 2, flock);
3443 add_safepoint_edges(lock);
3444
3445 lock = _gvn.transform( lock )->as_Lock();
3446
3447 // lock has no side-effects, sets few values
3448 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3449
3450 insert_mem_bar(Op_MemBarAcquireLock);
3451
3452 // Add this to the worklist so that the lock can be eliminated
3453 record_for_igvn(lock);
3454
3455 #ifndef PRODUCT
3456 if (PrintLockStatistics) {
3457 // Update the counter for this lock. Don't bother using an atomic
3458 // operation since we don't require absolute accuracy.
3459 lock->create_lock_counter(map()->jvms());
3460 increment_counter(lock->counter()->addr());
3461 }
3462 #endif
3463
3464 return flock;
3465 }
3466
3467
3468 //------------------------------shared_unlock----------------------------------
3469 // Emit unlocking code.
3470 void GraphKit::shared_unlock(Node* box, Node* obj) {
3471 // bci is either a monitorenter bc or InvocationEntryBci
3472 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3473 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3474
3475 if( !GenerateSynchronizationCode )
3476 return;
3477 if (stopped()) { // Dead monitor?
3478 map()->pop_monitor(); // Kill monitor from debug info
3479 return;
3480 }
3481
3482 // Memory barrier to avoid floating things down past the locked region
3483 insert_mem_bar(Op_MemBarReleaseLock);
3484
3485 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3486 UnlockNode *unlock = new UnlockNode(C, tf);
3487 #ifdef ASSERT
3488 unlock->set_dbg_jvms(sync_jvms());
3489 #endif
3490 uint raw_idx = Compile::AliasIdxRaw;
3491 unlock->init_req( TypeFunc::Control, control() );
3492 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3493 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3494 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3495 unlock->init_req( TypeFunc::ReturnAdr, top() );
3496
3497 unlock->init_req(TypeFunc::Parms + 0, obj);
3498 unlock->init_req(TypeFunc::Parms + 1, box);
3499 unlock = _gvn.transform(unlock)->as_Unlock();
3500
3501 Node* mem = reset_memory();
3502
3503 // unlock has no side-effects, sets few values
3504 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3505
3506 // Kill monitor from debug info
3507 map()->pop_monitor( );
3508 }
3509
3510 //-------------------------------get_layout_helper-----------------------------
3511 // If the given klass is a constant or known to be an array,
3512 // fetch the constant layout helper value into constant_value
3513 // and return (Node*)NULL. Otherwise, load the non-constant
3514 // layout helper value, and return the node which represents it.
3515 // This two-faced routine is useful because allocation sites
3516 // almost always feature constant types.
3517 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3518 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3519 if (!StressReflectiveCode && inst_klass != NULL) {
3520 ciKlass* klass = inst_klass->klass();
3521 assert(klass != NULL, "klass should not be NULL");
3522 bool xklass = inst_klass->klass_is_exact();
3523 bool can_be_value_array = false;
3524 if (klass->is_array_klass() && EnableValhalla && ValueArrayFlatten) {
3525 ciKlass* elem = klass->as_array_klass()->element_klass();
3526 can_be_value_array = elem != NULL && (elem->is_java_lang_Object() || elem->is_interface());
3527 }
3528 if (xklass || (klass->is_array_klass() && !can_be_value_array)) {
3529 jint lhelper = klass->layout_helper();
3530 if (lhelper != Klass::_lh_neutral_value) {
3531 constant_value = lhelper;
3532 return (Node*) NULL;
3533 }
3534 }
3535 }
3536 constant_value = Klass::_lh_neutral_value; // put in a known value
3537 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3538 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3539 }
3540
3541 // We just put in an allocate/initialize with a big raw-memory effect.
3542 // Hook selected additional alias categories on the initialization.
3543 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3544 MergeMemNode* init_in_merge,
3545 Node* init_out_raw) {
3546 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3547 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3548
3549 Node* prevmem = kit.memory(alias_idx);
3550 init_in_merge->set_memory_at(alias_idx, prevmem);
3551 kit.set_memory(init_out_raw, alias_idx);
3552 }
3553
3554 //---------------------------set_output_for_allocation-------------------------
3555 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3556 const TypeOopPtr* oop_type,
3557 bool deoptimize_on_exception) {
3558 int rawidx = Compile::AliasIdxRaw;
3559 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3560 add_safepoint_edges(alloc);
3561 Node* allocx = _gvn.transform(alloc);
3562 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3563 // create memory projection for i_o
3564 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3565 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3566
3567 // create a memory projection as for the normal control path
3568 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3569 set_memory(malloc, rawidx);
3570
3571 // a normal slow-call doesn't change i_o, but an allocation does
3572 // we create a separate i_o projection for the normal control path
3573 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3574 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3575
3576 // put in an initialization barrier
3577 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3578 rawoop)->as_Initialize();
3579 assert(alloc->initialization() == init, "2-way macro link must work");
3580 assert(init ->allocation() == alloc, "2-way macro link must work");
3581 {
3582 // Extract memory strands which may participate in the new object's
3583 // initialization, and source them from the new InitializeNode.
3584 // This will allow us to observe initializations when they occur,
3585 // and link them properly (as a group) to the InitializeNode.
3586 assert(init->in(InitializeNode::Memory) == malloc, "");
3587 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3588 init->set_req(InitializeNode::Memory, minit_in);
3589 record_for_igvn(minit_in); // fold it up later, if possible
3590 Node* minit_out = memory(rawidx);
3591 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3592 if (oop_type->isa_aryptr()) {
3593 const TypeAryPtr* arytype = oop_type->is_aryptr();
3594 if (arytype->klass()->is_value_array_klass()) {
3595 ciValueArrayKlass* vak = arytype->klass()->as_value_array_klass();
3596 ciValueKlass* vk = vak->element_klass()->as_value_klass();
3597 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
3598 ciField* field = vk->nonstatic_field_at(i);
3599 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3600 continue; // do not bother to track really large numbers of fields
3601 int off_in_vt = field->offset() - vk->first_field_offset();
3602 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
3603 int fieldidx = C->get_alias_index(adr_type);
3604 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3605 }
3606 } else {
3607 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3608 int elemidx = C->get_alias_index(telemref);
3609 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3610 }
3611 } else if (oop_type->isa_instptr()) {
3612 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3613 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3614 ciField* field = ik->nonstatic_field_at(i);
3615 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3616 continue; // do not bother to track really large numbers of fields
3617 // Find (or create) the alias category for this field:
3618 int fieldidx = C->alias_type(field)->index();
3619 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3620 }
3621 }
3622 }
3623
3624 // Cast raw oop to the real thing...
3625 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3626 javaoop = _gvn.transform(javaoop);
3627 C->set_recent_alloc(control(), javaoop);
3628 assert(just_allocated_object(control()) == javaoop, "just allocated");
3629
3630 #ifdef ASSERT
3631 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3632 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3633 "Ideal_allocation works");
3634 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3635 "Ideal_allocation works");
3636 if (alloc->is_AllocateArray()) {
3637 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3638 "Ideal_allocation works");
3639 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3640 "Ideal_allocation works");
3641 } else {
3642 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3643 }
3644 }
3645 #endif //ASSERT
3646
3647 return javaoop;
3648 }
3649
3650 //---------------------------new_instance--------------------------------------
3651 // This routine takes a klass_node which may be constant (for a static type)
3652 // or may be non-constant (for reflective code). It will work equally well
3653 // for either, and the graph will fold nicely if the optimizer later reduces
3654 // the type to a constant.
3655 // The optional arguments are for specialized use by intrinsics:
3656 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3657 // - If 'return_size_val', report the the total object size to the caller.
3658 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3659 Node* GraphKit::new_instance(Node* klass_node,
3660 Node* extra_slow_test,
3661 Node* *return_size_val,
3662 bool deoptimize_on_exception,
3663 ValueTypeBaseNode* value_node) {
3664 // Compute size in doublewords
3665 // The size is always an integral number of doublewords, represented
3666 // as a positive bytewise size stored in the klass's layout_helper.
3667 // The layout_helper also encodes (in a low bit) the need for a slow path.
3668 jint layout_con = Klass::_lh_neutral_value;
3669 Node* layout_val = get_layout_helper(klass_node, layout_con);
3670 bool layout_is_con = (layout_val == NULL);
3671
3672 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3673 // Generate the initial go-slow test. It's either ALWAYS (return a
3674 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3675 // case) a computed value derived from the layout_helper.
3676 Node* initial_slow_test = NULL;
3677 if (layout_is_con) {
3678 assert(!StressReflectiveCode, "stress mode does not use these paths");
3679 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3680 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3681 } else { // reflective case
3682 // This reflective path is used by Unsafe.allocateInstance.
3683 // (It may be stress-tested by specifying StressReflectiveCode.)
3684 // Basically, we want to get into the VM is there's an illegal argument.
3685 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3686 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3687 if (extra_slow_test != intcon(0)) {
3688 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3689 }
3690 // (Macro-expander will further convert this to a Bool, if necessary.)
3691 }
3692
3693 // Find the size in bytes. This is easy; it's the layout_helper.
3694 // The size value must be valid even if the slow path is taken.
3695 Node* size = NULL;
3696 if (layout_is_con) {
3697 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3698 } else { // reflective case
3699 // This reflective path is used by clone and Unsafe.allocateInstance.
3700 size = ConvI2X(layout_val);
3701
3702 // Clear the low bits to extract layout_helper_size_in_bytes:
3703 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3704 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3705 size = _gvn.transform( new AndXNode(size, mask) );
3706 }
3707 if (return_size_val != NULL) {
3708 (*return_size_val) = size;
3709 }
3710
3711 // This is a precise notnull oop of the klass.
3712 // (Actually, it need not be precise if this is a reflective allocation.)
3713 // It's what we cast the result to.
3714 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3715 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3716 const TypeOopPtr* oop_type = tklass->as_instance_type();
3717
3718 // Now generate allocation code
3719
3720 // The entire memory state is needed for slow path of the allocation
3721 // since GC and deoptimization can happen.
3722 Node *mem = reset_memory();
3723 set_all_memory(mem); // Create new memory state
3724
3725 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3726 control(), mem, i_o(),
3727 size, klass_node,
3728 initial_slow_test, value_node);
3729
3730 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3731 }
3732
3733 //-------------------------------new_array-------------------------------------
3734 // helper for newarray and anewarray
3735 // The 'length' parameter is (obviously) the length of the array.
3736 // See comments on new_instance for the meaning of the other arguments.
3737 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3738 Node* length, // number of array elements
3739 int nargs, // number of arguments to push back for uncommon trap
3740 Node* *return_size_val,
3741 bool deoptimize_on_exception) {
3742 jint layout_con = Klass::_lh_neutral_value;
3743 Node* layout_val = get_layout_helper(klass_node, layout_con);
3744 bool layout_is_con = (layout_val == NULL);
3745
3746 if (!layout_is_con && !StressReflectiveCode &&
3747 !too_many_traps(Deoptimization::Reason_class_check)) {
3748 // This is a reflective array creation site.
3749 // Optimistically assume that it is a subtype of Object[],
3750 // so that we can fold up all the address arithmetic.
3751 layout_con = Klass::array_layout_helper(T_OBJECT);
3752 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3753 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3754 { BuildCutout unless(this, bol_lh, PROB_MAX);
3755 inc_sp(nargs);
3756 uncommon_trap(Deoptimization::Reason_class_check,
3757 Deoptimization::Action_maybe_recompile);
3758 }
3759 layout_val = NULL;
3760 layout_is_con = true;
3761 }
3762
3763 // Generate the initial go-slow test. Make sure we do not overflow
3764 // if length is huge (near 2Gig) or negative! We do not need
3765 // exact double-words here, just a close approximation of needed
3766 // double-words. We can't add any offset or rounding bits, lest we
3767 // take a size -1 of bytes and make it positive. Use an unsigned
3768 // compare, so negative sizes look hugely positive.
3769 int fast_size_limit = FastAllocateSizeLimit;
3770 if (layout_is_con) {
3771 assert(!StressReflectiveCode, "stress mode does not use these paths");
3772 // Increase the size limit if we have exact knowledge of array type.
3773 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3774 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
3775 }
3776
3777 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3778 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3779
3780 // --- Size Computation ---
3781 // array_size = round_to_heap(array_header + (length << elem_shift));
3782 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3783 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3784 // The rounding mask is strength-reduced, if possible.
3785 int round_mask = MinObjAlignmentInBytes - 1;
3786 Node* header_size = NULL;
3787 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3788 // (T_BYTE has the weakest alignment and size restrictions...)
3789 if (layout_is_con) {
3790 int hsize = Klass::layout_helper_header_size(layout_con);
3791 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3792 BasicType etype = Klass::layout_helper_element_type(layout_con);
3793 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con);
3794 if ((round_mask & ~right_n_bits(eshift)) == 0)
3795 round_mask = 0; // strength-reduce it if it goes away completely
3796 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3797 assert(header_size_min <= hsize, "generic minimum is smallest");
3798 header_size_min = hsize;
3799 header_size = intcon(hsize + round_mask);
3800 } else {
3801 Node* hss = intcon(Klass::_lh_header_size_shift);
3802 Node* hsm = intcon(Klass::_lh_header_size_mask);
3803 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3804 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3805 Node* mask = intcon(round_mask);
3806 header_size = _gvn.transform( new AddINode(hsize, mask) );
3807 }
3808
3809 Node* elem_shift = NULL;
3810 if (layout_is_con) {
3811 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3812 if (eshift != 0)
3813 elem_shift = intcon(eshift);
3814 } else {
3815 // There is no need to mask or shift this value.
3816 // The semantics of LShiftINode include an implicit mask to 0x1F.
3817 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3818 elem_shift = layout_val;
3819 }
3820
3821 // Transition to native address size for all offset calculations:
3822 Node* lengthx = ConvI2X(length);
3823 Node* headerx = ConvI2X(header_size);
3824 #ifdef _LP64
3825 { const TypeInt* tilen = _gvn.find_int_type(length);
3826 if (tilen != NULL && tilen->_lo < 0) {
3827 // Add a manual constraint to a positive range. Cf. array_element_address.
3828 jint size_max = fast_size_limit;
3829 if (size_max > tilen->_hi) size_max = tilen->_hi;
3830 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3831
3832 // Only do a narrow I2L conversion if the range check passed.
3833 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3834 _gvn.transform(iff);
3835 RegionNode* region = new RegionNode(3);
3836 _gvn.set_type(region, Type::CONTROL);
3837 lengthx = new PhiNode(region, TypeLong::LONG);
3838 _gvn.set_type(lengthx, TypeLong::LONG);
3839
3840 // Range check passed. Use ConvI2L node with narrow type.
3841 Node* passed = IfFalse(iff);
3842 region->init_req(1, passed);
3843 // Make I2L conversion control dependent to prevent it from
3844 // floating above the range check during loop optimizations.
3845 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3846
3847 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3848 region->init_req(2, IfTrue(iff));
3849 lengthx->init_req(2, ConvI2X(length));
3850
3851 set_control(region);
3852 record_for_igvn(region);
3853 record_for_igvn(lengthx);
3854 }
3855 }
3856 #endif
3857
3858 // Combine header size (plus rounding) and body size. Then round down.
3859 // This computation cannot overflow, because it is used only in two
3860 // places, one where the length is sharply limited, and the other
3861 // after a successful allocation.
3862 Node* abody = lengthx;
3863 if (elem_shift != NULL)
3864 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3865 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3866 if (round_mask != 0) {
3867 Node* mask = MakeConX(~round_mask);
3868 size = _gvn.transform( new AndXNode(size, mask) );
3869 }
3870 // else if round_mask == 0, the size computation is self-rounding
3871
3872 if (return_size_val != NULL) {
3873 // This is the size
3874 (*return_size_val) = size;
3875 }
3876
3877 // Now generate allocation code
3878
3879 // The entire memory state is needed for slow path of the allocation
3880 // since GC and deoptimization can happen.
3881 Node *mem = reset_memory();
3882 set_all_memory(mem); // Create new memory state
3883
3884 if (initial_slow_test->is_Bool()) {
3885 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3886 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3887 }
3888
3889 // Create the AllocateArrayNode and its result projections
3890 AllocateArrayNode* alloc
3891 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3892 control(), mem, i_o(),
3893 size, klass_node,
3894 initial_slow_test,
3895 length);
3896
3897 // Cast to correct type. Note that the klass_node may be constant or not,
3898 // and in the latter case the actual array type will be inexact also.
3899 // (This happens via a non-constant argument to inline_native_newArray.)
3900 // In any case, the value of klass_node provides the desired array type.
3901 const TypeInt* length_type = _gvn.find_int_type(length);
3902 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3903 if (ary_type->isa_aryptr() && length_type != NULL) {
3904 // Try to get a better type than POS for the size
3905 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3906 }
3907
3908 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3909
3910 // Cast length on remaining path to be as narrow as possible
3911 if (map()->find_edge(length) >= 0) {
3912 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3913 if (ccast != length) {
3914 _gvn.set_type_bottom(ccast);
3915 record_for_igvn(ccast);
3916 replace_in_map(length, ccast);
3917 }
3918 }
3919
3920 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
3921 ciKlass* elem_klass = ary_ptr != NULL ? ary_ptr->klass()->as_array_klass()->element_klass() : NULL;
3922 if (elem_klass != NULL && elem_klass->is_valuetype()) {
3923 ciValueKlass* vk = elem_klass->as_value_klass();
3924 if (!vk->flatten_array()) {
3925 // Non-flattened value type arrays need to be initialized with default value type oops
3926 initialize_value_type_array(javaoop, length, elem_klass->as_value_klass(), nargs);
3927 // TODO re-enable once JDK-8189802 is fixed
3928 // InitializeNode* init = alloc->initialization();
3929 //init->set_complete_with_arraycopy();
3930 }
3931 } else if (EnableValhalla && (!layout_con || elem_klass == NULL || (elem_klass->is_java_lang_Object() && !ary_type->klass_is_exact()))) {
3932 InitializeNode* init = alloc->initialization();
3933 init->set_unknown_value();
3934 }
3935
3936 return javaoop;
3937 }
3938
3939 void GraphKit::initialize_value_type_array(Node* array, Node* length, ciValueKlass* vk, int nargs) {
3940 // Check for zero length
3941 Node* null_ctl = top();
3942 null_check_common(length, T_INT, false, &null_ctl, false);
3943 if (stopped()) {
3944 set_control(null_ctl); // Always zero
3945 return;
3946 }
3947
3948 RegionNode* res_ctl = new RegionNode(3);
3949 gvn().set_type(res_ctl, Type::CONTROL);
3950 record_for_igvn(res_ctl);
3951
3952 // Length is zero: don't execute initialization loop
3953 res_ctl->init_req(1, null_ctl);
3954 PhiNode* res_io = PhiNode::make(res_ctl, i_o(), Type::ABIO);
3955 PhiNode* res_mem = PhiNode::make(res_ctl, merged_memory(), Type::MEMORY, TypePtr::BOTTOM);
3956 gvn().set_type(res_io, Type::ABIO);
3957 gvn().set_type(res_mem, Type::MEMORY);
3958 record_for_igvn(res_io);
3959 record_for_igvn(res_mem);
3960
3961 // Length is non-zero: execute a loop that initializes the array with the default value type
3962 Node* oop = ValueTypeNode::load_default_oop(gvn(), vk);
3963
3964 add_predicate(nargs);
3965 RegionNode* loop = new RegionNode(3);
3966 loop->init_req(1, control());
3967 PhiNode* index = PhiNode::make(loop, intcon(0), TypeInt::INT);
3968 PhiNode* mem = PhiNode::make(loop, reset_memory(), Type::MEMORY, TypePtr::BOTTOM);
3969
3970 gvn().set_type(loop, Type::CONTROL);
3971 gvn().set_type(index, TypeInt::INT);
3972 gvn().set_type(mem, Type::MEMORY);
3973 record_for_igvn(loop);
3974 record_for_igvn(index);
3975 record_for_igvn(mem);
3976
3977 // Loop body: initialize array element at 'index'
3978 set_control(loop);
3979 set_all_memory(mem);
3980 Node* adr = array_element_address(array, index, T_OBJECT);
3981 const TypeOopPtr* elemtype = TypeInstPtr::make(TypePtr::NotNull, vk);
3982 access_store_at(control(), array, adr, TypeAryPtr::OOPS, oop, elemtype, T_VALUETYPE, MemNode::release);
3983
3984 // Check if we need to execute another loop iteration
3985 length = SubI(length, intcon(1));
3986 IfNode* iff = create_and_map_if(control(), Bool(CmpI(index, length), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
3987
3988 // Continue with next iteration
3989 loop->init_req(2, IfTrue(iff));
3990 index->init_req(2, AddI(index, intcon(1)));
3991 mem->init_req(2, merged_memory());
3992
3993 // Exit loop
3994 res_ctl->init_req(2, IfFalse(iff));
3995 res_io->set_req(2, i_o());
3996 res_mem->set_req(2, reset_memory());
3997
3998 // Set merged control, IO and memory
3999 set_control(res_ctl);
4000 set_i_o(res_io);
4001 set_all_memory(res_mem);
4002 }
4003
4004 // The following "Ideal_foo" functions are placed here because they recognize
4005 // the graph shapes created by the functions immediately above.
4006
4007 //---------------------------Ideal_allocation----------------------------------
4008 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4009 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
4010 if (ptr == NULL) { // reduce dumb test in callers
4011 return NULL;
4012 }
4013 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4014 ptr = ptr->in(1);
4015 if (ptr == NULL) return NULL;
4016 }
4017 // Return NULL for allocations with several casts:
4018 // j.l.reflect.Array.newInstance(jobject, jint)
4019 // Object.clone()
4020 // to keep more precise type from last cast.
4021 if (ptr->is_Proj()) {
4022 Node* allo = ptr->in(0);
4023 if (allo != NULL && allo->is_Allocate()) {
4024 return allo->as_Allocate();
4025 }
4026 }
4027 // Report failure to match.
4028 return NULL;
4029 }
4030
4031 // Fancy version which also strips off an offset (and reports it to caller).
4032 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
4033 intptr_t& offset) {
4034 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4035 if (base == NULL) return NULL;
4036 return Ideal_allocation(base, phase);
4037 }
4038
4039 // Trace Initialize <- Proj[Parm] <- Allocate
4040 AllocateNode* InitializeNode::allocation() {
4041 Node* rawoop = in(InitializeNode::RawAddress);
4042 if (rawoop->is_Proj()) {
4043 Node* alloc = rawoop->in(0);
4044 if (alloc->is_Allocate()) {
4045 return alloc->as_Allocate();
4046 }
4047 }
4048 return NULL;
4049 }
4050
4051 // Trace Allocate -> Proj[Parm] -> Initialize
4052 InitializeNode* AllocateNode::initialization() {
4053 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4054 if (rawoop == NULL) return NULL;
4055 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4056 Node* init = rawoop->fast_out(i);
4057 if (init->is_Initialize()) {
4058 assert(init->as_Initialize()->allocation() == this, "2-way link");
4059 return init->as_Initialize();
4060 }
4061 }
4062 return NULL;
4063 }
4064
4065 //----------------------------- loop predicates ---------------------------
4066
4067 //------------------------------add_predicate_impl----------------------------
4068 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
4069 // Too many traps seen?
4070 if (too_many_traps(reason)) {
4071 #ifdef ASSERT
4072 if (TraceLoopPredicate) {
4073 int tc = C->trap_count(reason);
4074 tty->print("too many traps=%s tcount=%d in ",
4075 Deoptimization::trap_reason_name(reason), tc);
4076 method()->print(); // which method has too many predicate traps
4077 tty->cr();
4078 }
4079 #endif
4080 // We cannot afford to take more traps here,
4081 // do not generate predicate.
4082 return;
4083 }
4084
4085 Node *cont = _gvn.intcon(1);
4086 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
4087 Node *bol = _gvn.transform(new Conv2BNode(opq));
4088 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4089 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
4090 C->add_predicate_opaq(opq);
4091 {
4092 PreserveJVMState pjvms(this);
4093 set_control(iffalse);
4094 inc_sp(nargs);
4095 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4096 }
4097 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
4098 set_control(iftrue);
4099 }
4100
4101 //------------------------------add_predicate---------------------------------
4102 void GraphKit::add_predicate(int nargs) {
4103 if (UseLoopPredicate) {
4104 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
4105 }
4106 // loop's limit check predicate should be near the loop.
4107 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
4108 }
4109
4110 void GraphKit::sync_kit(IdealKit& ideal) {
4111 set_all_memory(ideal.merged_memory());
4112 set_i_o(ideal.i_o());
4113 set_control(ideal.ctrl());
4114 }
4115
4116 void GraphKit::final_sync(IdealKit& ideal) {
4117 // Final sync IdealKit and graphKit.
4118 sync_kit(ideal);
4119 }
4120
4121 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4122 Node* len = load_array_length(load_String_value(ctrl, str));
4123 Node* coder = load_String_coder(ctrl, str);
4124 // Divide length by 2 if coder is UTF16
4125 return _gvn.transform(new RShiftINode(len, coder));
4126 }
4127
4128 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4129 int value_offset = java_lang_String::value_offset_in_bytes();
4130 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4131 false, NULL, Type::Offset(0));
4132 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4133 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4134 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4135 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4136 Node* p = basic_plus_adr(str, str, value_offset);
4137 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4138 IN_HEAP | C2_CONTROL_DEPENDENT_LOAD);
4139 // String.value field is known to be @Stable.
4140 if (UseImplicitStableValues) {
4141 load = cast_array_to_stable(load, value_type);
4142 }
4143 return load;
4144 }
4145
4146 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4147 if (!CompactStrings) {
4148 return intcon(java_lang_String::CODER_UTF16);
4149 }
4150 int coder_offset = java_lang_String::coder_offset_in_bytes();
4151 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4152 false, NULL, Type::Offset(0));
4153 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4154 int coder_field_idx = C->get_alias_index(coder_field_type);
4155 return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4156 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4157 }
4158
4159 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4160 int value_offset = java_lang_String::value_offset_in_bytes();
4161 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4162 false, NULL, Type::Offset(0));
4163 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4164 access_store_at(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4165 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP);
4166 }
4167
4168 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4169 int coder_offset = java_lang_String::coder_offset_in_bytes();
4170 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4171 false, NULL, Type::Offset(0));
4172 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4173 int coder_field_idx = C->get_alias_index(coder_field_type);
4174 store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4175 value, T_BYTE, coder_field_idx, MemNode::unordered);
4176 }
4177
4178 // Capture src and dst memory state with a MergeMemNode
4179 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4180 if (src_type == dst_type) {
4181 // Types are equal, we don't need a MergeMemNode
4182 return memory(src_type);
4183 }
4184 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4185 record_for_igvn(merge); // fold it up later, if possible
4186 int src_idx = C->get_alias_index(src_type);
4187 int dst_idx = C->get_alias_index(dst_type);
4188 merge->set_memory_at(src_idx, memory(src_idx));
4189 merge->set_memory_at(dst_idx, memory(dst_idx));
4190 return merge;
4191 }
4192
4193 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4194 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4195 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4196 // If input and output memory types differ, capture both states to preserve
4197 // the dependency between preceding and subsequent loads/stores.
4198 // For example, the following program:
4199 // StoreB
4200 // compress_string
4201 // LoadB
4202 // has this memory graph (use->def):
4203 // LoadB -> compress_string -> CharMem
4204 // ... -> StoreB -> ByteMem
4205 // The intrinsic hides the dependency between LoadB and StoreB, causing
4206 // the load to read from memory not containing the result of the StoreB.
4207 // The correct memory graph should look like this:
4208 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4209 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4210 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4211 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4212 set_memory(res_mem, TypeAryPtr::BYTES);
4213 return str;
4214 }
4215
4216 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4217 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4218 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4219 // Capture src and dst memory (see comment in 'compress_string').
4220 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4221 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4222 set_memory(_gvn.transform(str), dst_type);
4223 }
4224
4225 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4226 /**
4227 * int i_char = start;
4228 * for (int i_byte = 0; i_byte < count; i_byte++) {
4229 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4230 * }
4231 */
4232 add_predicate();
4233 RegionNode* head = new RegionNode(3);
4234 head->init_req(1, control());
4235 gvn().set_type(head, Type::CONTROL);
4236 record_for_igvn(head);
4237
4238 Node* i_byte = new PhiNode(head, TypeInt::INT);
4239 i_byte->init_req(1, intcon(0));
4240 gvn().set_type(i_byte, TypeInt::INT);
4241 record_for_igvn(i_byte);
4242
4243 Node* i_char = new PhiNode(head, TypeInt::INT);
4244 i_char->init_req(1, start);
4245 gvn().set_type(i_char, TypeInt::INT);
4246 record_for_igvn(i_char);
4247
4248 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4249 gvn().set_type(mem, Type::MEMORY);
4250 record_for_igvn(mem);
4251 set_control(head);
4252 set_memory(mem, TypeAryPtr::BYTES);
4253 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4254 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4255 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4256 false, false, true /* mismatched */);
4257
4258 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4259 head->init_req(2, IfTrue(iff));
4260 mem->init_req(2, st);
4261 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4262 i_char->init_req(2, AddI(i_char, intcon(2)));
4263
4264 set_control(IfFalse(iff));
4265 set_memory(st, TypeAryPtr::BYTES);
4266 }
4267
4268 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4269 if (!field->is_constant()) {
4270 return NULL; // Field not marked as constant.
4271 }
4272 ciInstance* holder = NULL;
4273 if (!field->is_static()) {
4274 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4275 if (const_oop != NULL && const_oop->is_instance()) {
4276 holder = const_oop->as_instance();
4277 }
4278 }
4279 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4280 /*is_unsigned_load=*/false);
4281 if (con_type != NULL) {
4282 Node* con = makecon(con_type);
4283 if (field->layout_type() == T_VALUETYPE) {
4284 // Load value type from constant oop
4285 con = ValueTypeNode::make_from_oop(this, con, field->type()->as_value_klass());
4286 }
4287 return con;
4288 }
4289 return NULL;
4290 }
4291
4292 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4293 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4294 // assumption of CCP analysis.
4295 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4296 }