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/cardTable.hpp"
31 #include "gc/shared/cardTableBarrierSet.hpp"
32 #include "gc/shared/collectedHeap.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/idealKit.hpp"
40 #include "opto/intrinsicnode.hpp"
41 #include "opto/locknode.hpp"
42 #include "opto/machnode.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/parse.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/valuetypenode.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/sharedRuntime.hpp"
50 #if INCLUDE_G1GC
51 #include "gc/g1/g1CardTable.hpp"
52 #include "gc/g1/g1ThreadLocalData.hpp"
53 #include "gc/g1/heapRegion.hpp"
54 #endif // INCLUDE_ALL_GCS
55
56 //----------------------------GraphKit-----------------------------------------
57 // Main utility constructor.
58 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
59 : Phase(Phase::Parser),
60 _env(C->env()),
61 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn())
62 {
63 _exceptions = jvms->map()->next_exception();
64 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
65 set_jvms(jvms);
66 #ifdef ASSERT
67 if (_gvn.is_IterGVN() != NULL) {
68 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
69 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
70 _worklist_size = _gvn.C->for_igvn()->size();
71 }
72 #endif
73 }
74
75 // Private constructor for parser.
76 GraphKit::GraphKit()
77 : Phase(Phase::Parser),
78 _env(C->env()),
79 _gvn(*C->initial_gvn())
80 {
81 _exceptions = NULL;
82 set_map(NULL);
83 debug_only(_sp = -99);
84 debug_only(set_bci(-99));
85 }
86
87
88
89 //---------------------------clean_stack---------------------------------------
90 // Clear away rubbish from the stack area of the JVM state.
91 // This destroys any arguments that may be waiting on the stack.
92 void GraphKit::clean_stack(int from_sp) {
93 SafePointNode* map = this->map();
94 JVMState* jvms = this->jvms();
95 int stk_size = jvms->stk_size();
96 int stkoff = jvms->stkoff();
97 Node* top = this->top();
98 for (int i = from_sp; i < stk_size; i++) {
99 if (map->in(stkoff + i) != top) {
100 map->set_req(stkoff + i, top);
101 }
102 }
103 }
104
105
106 //--------------------------------sync_jvms-----------------------------------
107 // Make sure our current jvms agrees with our parse state.
108 JVMState* GraphKit::sync_jvms() const {
109 JVMState* jvms = this->jvms();
110 jvms->set_bci(bci()); // Record the new bci in the JVMState
111 jvms->set_sp(sp()); // Record the new sp in the JVMState
112 assert(jvms_in_sync(), "jvms is now in sync");
113 return jvms;
114 }
115
116 //--------------------------------sync_jvms_for_reexecute---------------------
117 // Make sure our current jvms agrees with our parse state. This version
118 // uses the reexecute_sp for reexecuting bytecodes.
119 JVMState* GraphKit::sync_jvms_for_reexecute() {
120 JVMState* jvms = this->jvms();
121 jvms->set_bci(bci()); // Record the new bci in the JVMState
122 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
123 return jvms;
124 }
125
126 #ifdef ASSERT
127 bool GraphKit::jvms_in_sync() const {
128 Parse* parse = is_Parse();
129 if (parse == NULL) {
130 if (bci() != jvms()->bci()) return false;
131 if (sp() != (int)jvms()->sp()) return false;
132 return true;
133 }
134 if (jvms()->method() != parse->method()) return false;
135 if (jvms()->bci() != parse->bci()) return false;
136 int jvms_sp = jvms()->sp();
137 if (jvms_sp != parse->sp()) return false;
138 int jvms_depth = jvms()->depth();
139 if (jvms_depth != parse->depth()) return false;
140 return true;
141 }
142
143 // Local helper checks for special internal merge points
144 // used to accumulate and merge exception states.
145 // They are marked by the region's in(0) edge being the map itself.
146 // Such merge points must never "escape" into the parser at large,
147 // until they have been handed to gvn.transform.
148 static bool is_hidden_merge(Node* reg) {
149 if (reg == NULL) return false;
150 if (reg->is_Phi()) {
151 reg = reg->in(0);
152 if (reg == NULL) return false;
153 }
154 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
155 }
156
157 void GraphKit::verify_map() const {
158 if (map() == NULL) return; // null map is OK
159 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
160 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
161 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
162 }
163
164 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
165 assert(ex_map->next_exception() == NULL, "not already part of a chain");
166 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
167 }
168 #endif
169
170 //---------------------------stop_and_kill_map---------------------------------
171 // Set _map to NULL, signalling a stop to further bytecode execution.
172 // First smash the current map's control to a constant, to mark it dead.
173 void GraphKit::stop_and_kill_map() {
174 SafePointNode* dead_map = stop();
175 if (dead_map != NULL) {
176 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
177 assert(dead_map->is_killed(), "must be so marked");
178 }
179 }
180
181
182 //--------------------------------stopped--------------------------------------
183 // Tell if _map is NULL, or control is top.
184 bool GraphKit::stopped() {
185 if (map() == NULL) return true;
186 else if (control() == top()) return true;
187 else return false;
188 }
189
190
191 //-----------------------------has_ex_handler----------------------------------
192 // Tell if this method or any caller method has exception handlers.
193 bool GraphKit::has_ex_handler() {
194 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
195 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
196 return true;
197 }
198 }
199 return false;
200 }
201
202 //------------------------------save_ex_oop------------------------------------
203 // Save an exception without blowing stack contents or other JVM state.
204 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
205 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
206 ex_map->add_req(ex_oop);
207 debug_only(verify_exception_state(ex_map));
208 }
209
210 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
211 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
212 Node* ex_oop = ex_map->in(ex_map->req()-1);
213 if (clear_it) ex_map->del_req(ex_map->req()-1);
214 return ex_oop;
215 }
216
217 //-----------------------------saved_ex_oop------------------------------------
218 // Recover a saved exception from its map.
219 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
220 return common_saved_ex_oop(ex_map, false);
221 }
222
223 //--------------------------clear_saved_ex_oop---------------------------------
224 // Erase a previously saved exception from its map.
225 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
226 return common_saved_ex_oop(ex_map, true);
227 }
228
229 #ifdef ASSERT
230 //---------------------------has_saved_ex_oop----------------------------------
231 // Erase a previously saved exception from its map.
232 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
233 return ex_map->req() == ex_map->jvms()->endoff()+1;
234 }
235 #endif
236
237 //-------------------------make_exception_state--------------------------------
238 // Turn the current JVM state into an exception state, appending the ex_oop.
239 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
240 sync_jvms();
241 SafePointNode* ex_map = stop(); // do not manipulate this map any more
242 set_saved_ex_oop(ex_map, ex_oop);
243 return ex_map;
244 }
245
246
247 //--------------------------add_exception_state--------------------------------
248 // Add an exception to my list of exceptions.
249 void GraphKit::add_exception_state(SafePointNode* ex_map) {
250 if (ex_map == NULL || ex_map->control() == top()) {
251 return;
252 }
253 #ifdef ASSERT
254 verify_exception_state(ex_map);
255 if (has_exceptions()) {
256 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
257 }
258 #endif
259
260 // If there is already an exception of exactly this type, merge with it.
261 // In particular, null-checks and other low-level exceptions common up here.
262 Node* ex_oop = saved_ex_oop(ex_map);
263 const Type* ex_type = _gvn.type(ex_oop);
264 if (ex_oop == top()) {
265 // No action needed.
266 return;
267 }
268 assert(ex_type->isa_instptr(), "exception must be an instance");
269 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
270 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
271 // We check sp also because call bytecodes can generate exceptions
272 // both before and after arguments are popped!
273 if (ex_type2 == ex_type
274 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
275 combine_exception_states(ex_map, e2);
276 return;
277 }
278 }
279
280 // No pre-existing exception of the same type. Chain it on the list.
281 push_exception_state(ex_map);
282 }
283
284 //-----------------------add_exception_states_from-----------------------------
285 void GraphKit::add_exception_states_from(JVMState* jvms) {
286 SafePointNode* ex_map = jvms->map()->next_exception();
287 if (ex_map != NULL) {
288 jvms->map()->set_next_exception(NULL);
289 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
290 next_map = ex_map->next_exception();
291 ex_map->set_next_exception(NULL);
292 add_exception_state(ex_map);
293 }
294 }
295 }
296
297 //-----------------------transfer_exceptions_into_jvms-------------------------
298 JVMState* GraphKit::transfer_exceptions_into_jvms() {
299 if (map() == NULL) {
300 // We need a JVMS to carry the exceptions, but the map has gone away.
301 // Create a scratch JVMS, cloned from any of the exception states...
302 if (has_exceptions()) {
303 _map = _exceptions;
304 _map = clone_map();
305 _map->set_next_exception(NULL);
306 clear_saved_ex_oop(_map);
307 debug_only(verify_map());
308 } else {
309 // ...or created from scratch
310 JVMState* jvms = new (C) JVMState(_method, NULL);
311 jvms->set_bci(_bci);
312 jvms->set_sp(_sp);
313 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
314 set_jvms(jvms);
315 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
316 set_all_memory(top());
317 while (map()->req() < jvms->endoff()) map()->add_req(top());
318 }
319 // (This is a kludge, in case you didn't notice.)
320 set_control(top());
321 }
322 JVMState* jvms = sync_jvms();
323 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
324 jvms->map()->set_next_exception(_exceptions);
325 _exceptions = NULL; // done with this set of exceptions
326 return jvms;
327 }
328
329 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
330 assert(is_hidden_merge(dstphi), "must be a special merge node");
331 assert(is_hidden_merge(srcphi), "must be a special merge node");
332 uint limit = srcphi->req();
333 for (uint i = PhiNode::Input; i < limit; i++) {
334 dstphi->add_req(srcphi->in(i));
335 }
336 }
337 static inline void add_one_req(Node* dstphi, Node* src) {
338 assert(is_hidden_merge(dstphi), "must be a special merge node");
339 assert(!is_hidden_merge(src), "must not be a special merge node");
340 dstphi->add_req(src);
341 }
342
343 //-----------------------combine_exception_states------------------------------
344 // This helper function combines exception states by building phis on a
345 // specially marked state-merging region. These regions and phis are
346 // untransformed, and can build up gradually. The region is marked by
347 // having a control input of its exception map, rather than NULL. Such
348 // regions do not appear except in this function, and in use_exception_state.
349 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
350 if (failing()) return; // dying anyway...
351 JVMState* ex_jvms = ex_map->_jvms;
352 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
353 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
354 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
355 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
356 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
357 assert(ex_map->req() == phi_map->req(), "matching maps");
358 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
359 Node* hidden_merge_mark = root();
360 Node* region = phi_map->control();
361 MergeMemNode* phi_mem = phi_map->merged_memory();
362 MergeMemNode* ex_mem = ex_map->merged_memory();
363 if (region->in(0) != hidden_merge_mark) {
364 // The control input is not (yet) a specially-marked region in phi_map.
365 // Make it so, and build some phis.
366 region = new RegionNode(2);
367 _gvn.set_type(region, Type::CONTROL);
368 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
369 region->init_req(1, phi_map->control());
370 phi_map->set_control(region);
371 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
372 record_for_igvn(io_phi);
373 _gvn.set_type(io_phi, Type::ABIO);
374 phi_map->set_i_o(io_phi);
375 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
376 Node* m = mms.memory();
377 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
378 record_for_igvn(m_phi);
379 _gvn.set_type(m_phi, Type::MEMORY);
380 mms.set_memory(m_phi);
381 }
382 }
383
384 // Either or both of phi_map and ex_map might already be converted into phis.
385 Node* ex_control = ex_map->control();
386 // if there is special marking on ex_map also, we add multiple edges from src
387 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
388 // how wide was the destination phi_map, originally?
389 uint orig_width = region->req();
390
391 if (add_multiple) {
392 add_n_reqs(region, ex_control);
393 add_n_reqs(phi_map->i_o(), ex_map->i_o());
394 } else {
395 // ex_map has no merges, so we just add single edges everywhere
396 add_one_req(region, ex_control);
397 add_one_req(phi_map->i_o(), ex_map->i_o());
398 }
399 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
400 if (mms.is_empty()) {
401 // get a copy of the base memory, and patch some inputs into it
402 const TypePtr* adr_type = mms.adr_type(C);
403 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
404 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
405 mms.set_memory(phi);
406 // Prepare to append interesting stuff onto the newly sliced phi:
407 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
408 }
409 // Append stuff from ex_map:
410 if (add_multiple) {
411 add_n_reqs(mms.memory(), mms.memory2());
412 } else {
413 add_one_req(mms.memory(), mms.memory2());
414 }
415 }
416 uint limit = ex_map->req();
417 for (uint i = TypeFunc::Parms; i < limit; i++) {
418 // Skip everything in the JVMS after tos. (The ex_oop follows.)
419 if (i == tos) i = ex_jvms->monoff();
420 Node* src = ex_map->in(i);
421 Node* dst = phi_map->in(i);
422 if (src != dst) {
423 PhiNode* phi;
424 if (dst->in(0) != region) {
425 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
426 record_for_igvn(phi);
427 _gvn.set_type(phi, phi->type());
428 phi_map->set_req(i, dst);
429 // Prepare to append interesting stuff onto the new phi:
430 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
431 } else {
432 assert(dst->is_Phi(), "nobody else uses a hidden region");
433 phi = dst->as_Phi();
434 }
435 if (add_multiple && src->in(0) == ex_control) {
436 // Both are phis.
437 add_n_reqs(dst, src);
438 } else {
439 while (dst->req() < region->req()) add_one_req(dst, src);
440 }
441 const Type* srctype = _gvn.type(src);
442 if (phi->type() != srctype) {
443 const Type* dsttype = phi->type()->meet_speculative(srctype);
444 if (phi->type() != dsttype) {
445 phi->set_type(dsttype);
446 _gvn.set_type(phi, dsttype);
447 }
448 }
449 }
450 }
451 phi_map->merge_replaced_nodes_with(ex_map);
452 }
453
454 //--------------------------use_exception_state--------------------------------
455 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
456 if (failing()) { stop(); return top(); }
457 Node* region = phi_map->control();
458 Node* hidden_merge_mark = root();
459 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
460 Node* ex_oop = clear_saved_ex_oop(phi_map);
461 if (region->in(0) == hidden_merge_mark) {
462 // Special marking for internal ex-states. Process the phis now.
463 region->set_req(0, region); // now it's an ordinary region
464 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
465 // Note: Setting the jvms also sets the bci and sp.
466 set_control(_gvn.transform(region));
467 uint tos = jvms()->stkoff() + sp();
468 for (uint i = 1; i < tos; i++) {
469 Node* x = phi_map->in(i);
470 if (x->in(0) == region) {
471 assert(x->is_Phi(), "expected a special phi");
472 phi_map->set_req(i, _gvn.transform(x));
473 }
474 }
475 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
476 Node* x = mms.memory();
477 if (x->in(0) == region) {
478 assert(x->is_Phi(), "nobody else uses a hidden region");
479 mms.set_memory(_gvn.transform(x));
480 }
481 }
482 if (ex_oop->in(0) == region) {
483 assert(ex_oop->is_Phi(), "expected a special phi");
484 ex_oop = _gvn.transform(ex_oop);
485 }
486 } else {
487 set_jvms(phi_map->jvms());
488 }
489
490 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
491 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
492 return ex_oop;
493 }
494
495 //---------------------------------java_bc-------------------------------------
496 Bytecodes::Code GraphKit::java_bc() const {
497 ciMethod* method = this->method();
498 int bci = this->bci();
499 if (method != NULL && bci != InvocationEntryBci)
500 return method->java_code_at_bci(bci);
501 else
502 return Bytecodes::_illegal;
503 }
504
505 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
506 bool must_throw) {
507 // if the exception capability is set, then we will generate code
508 // to check the JavaThread.should_post_on_exceptions flag to see
509 // if we actually need to report exception events (for this
510 // thread). If we don't need to report exception events, we will
511 // take the normal fast path provided by add_exception_events. If
512 // exception event reporting is enabled for this thread, we will
513 // take the uncommon_trap in the BuildCutout below.
514
515 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
516 Node* jthread = _gvn.transform(new ThreadLocalNode());
517 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
518 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
519
520 // Test the should_post_on_exceptions_flag vs. 0
521 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
522 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
523
524 // Branch to slow_path if should_post_on_exceptions_flag was true
525 { BuildCutout unless(this, tst, PROB_MAX);
526 // Do not try anything fancy if we're notifying the VM on every throw.
527 // Cf. case Bytecodes::_athrow in parse2.cpp.
528 uncommon_trap(reason, Deoptimization::Action_none,
529 (ciKlass*)NULL, (char*)NULL, must_throw);
530 }
531
532 }
533
534 //------------------------------builtin_throw----------------------------------
535 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
536 bool must_throw = true;
537
538 if (env()->jvmti_can_post_on_exceptions()) {
539 // check if we must post exception events, take uncommon trap if so
540 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
541 // here if should_post_on_exceptions is false
542 // continue on with the normal codegen
543 }
544
545 // If this particular condition has not yet happened at this
546 // bytecode, then use the uncommon trap mechanism, and allow for
547 // a future recompilation if several traps occur here.
548 // If the throw is hot, try to use a more complicated inline mechanism
549 // which keeps execution inside the compiled code.
550 bool treat_throw_as_hot = false;
551 ciMethodData* md = method()->method_data();
552
553 if (ProfileTraps) {
554 if (too_many_traps(reason)) {
555 treat_throw_as_hot = true;
556 }
557 // (If there is no MDO at all, assume it is early in
558 // execution, and that any deopts are part of the
559 // startup transient, and don't need to be remembered.)
560
561 // Also, if there is a local exception handler, treat all throws
562 // as hot if there has been at least one in this method.
563 if (C->trap_count(reason) != 0
564 && method()->method_data()->trap_count(reason) != 0
565 && has_ex_handler()) {
566 treat_throw_as_hot = true;
567 }
568 }
569
570 // If this throw happens frequently, an uncommon trap might cause
571 // a performance pothole. If there is a local exception handler,
572 // and if this particular bytecode appears to be deoptimizing often,
573 // let us handle the throw inline, with a preconstructed instance.
574 // Note: If the deopt count has blown up, the uncommon trap
575 // runtime is going to flush this nmethod, not matter what.
576 if (treat_throw_as_hot
577 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
578 // If the throw is local, we use a pre-existing instance and
579 // punt on the backtrace. This would lead to a missing backtrace
580 // (a repeat of 4292742) if the backtrace object is ever asked
581 // for its backtrace.
582 // Fixing this remaining case of 4292742 requires some flavor of
583 // escape analysis. Leave that for the future.
584 ciInstance* ex_obj = NULL;
585 switch (reason) {
586 case Deoptimization::Reason_null_check:
587 ex_obj = env()->NullPointerException_instance();
588 break;
589 case Deoptimization::Reason_div0_check:
590 ex_obj = env()->ArithmeticException_instance();
591 break;
592 case Deoptimization::Reason_range_check:
593 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
594 break;
595 case Deoptimization::Reason_class_check:
596 if (java_bc() == Bytecodes::_aastore) {
597 ex_obj = env()->ArrayStoreException_instance();
598 } else {
599 ex_obj = env()->ClassCastException_instance();
600 }
601 break;
602 default:
603 break;
604 }
605 if (failing()) { stop(); return; } // exception allocation might fail
606 if (ex_obj != NULL) {
607 // Cheat with a preallocated exception object.
608 if (C->log() != NULL)
609 C->log()->elem("hot_throw preallocated='1' reason='%s'",
610 Deoptimization::trap_reason_name(reason));
611 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
612 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
613
614 // Clear the detail message of the preallocated exception object.
615 // Weblogic sometimes mutates the detail message of exceptions
616 // using reflection.
617 int offset = java_lang_Throwable::get_detailMessage_offset();
618 const TypePtr* adr_typ = ex_con->add_offset(offset);
619
620 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
621 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
622 // Conservatively release stores of object references.
623 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
624
625 add_exception_state(make_exception_state(ex_node));
626 return;
627 }
628 }
629
630 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
631 // It won't be much cheaper than bailing to the interp., since we'll
632 // have to pass up all the debug-info, and the runtime will have to
633 // create the stack trace.
634
635 // Usual case: Bail to interpreter.
636 // Reserve the right to recompile if we haven't seen anything yet.
637
638 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
639 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
640 if (treat_throw_as_hot
641 && (method()->method_data()->trap_recompiled_at(bci(), m)
642 || C->too_many_traps(reason))) {
643 // We cannot afford to take more traps here. Suffer in the interpreter.
644 if (C->log() != NULL)
645 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
646 Deoptimization::trap_reason_name(reason),
647 C->trap_count(reason));
648 action = Deoptimization::Action_none;
649 }
650
651 // "must_throw" prunes the JVM state to include only the stack, if there
652 // are no local exception handlers. This should cut down on register
653 // allocation time and code size, by drastically reducing the number
654 // of in-edges on the call to the uncommon trap.
655
656 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
657 }
658
659
660 //----------------------------PreserveJVMState---------------------------------
661 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
662 debug_only(kit->verify_map());
663 _kit = kit;
664 _map = kit->map(); // preserve the map
665 _sp = kit->sp();
666 kit->set_map(clone_map ? kit->clone_map() : NULL);
667 #ifdef ASSERT
668 _bci = kit->bci();
669 Parse* parser = kit->is_Parse();
670 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
671 _block = block;
672 #endif
673 }
674 PreserveJVMState::~PreserveJVMState() {
675 GraphKit* kit = _kit;
676 #ifdef ASSERT
677 assert(kit->bci() == _bci, "bci must not shift");
678 Parse* parser = kit->is_Parse();
679 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
680 assert(block == _block, "block must not shift");
681 #endif
682 kit->set_map(_map);
683 kit->set_sp(_sp);
684 }
685
686
687 //-----------------------------BuildCutout-------------------------------------
688 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
689 : PreserveJVMState(kit)
690 {
691 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
692 SafePointNode* outer_map = _map; // preserved map is caller's
693 SafePointNode* inner_map = kit->map();
694 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
695 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
696 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
697 }
698 BuildCutout::~BuildCutout() {
699 GraphKit* kit = _kit;
700 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
701 }
702
703 //---------------------------PreserveReexecuteState----------------------------
704 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
705 assert(!kit->stopped(), "must call stopped() before");
706 _kit = kit;
707 _sp = kit->sp();
708 _reexecute = kit->jvms()->_reexecute;
709 }
710 PreserveReexecuteState::~PreserveReexecuteState() {
711 if (_kit->stopped()) return;
712 _kit->jvms()->_reexecute = _reexecute;
713 _kit->set_sp(_sp);
714 }
715
716 //------------------------------clone_map--------------------------------------
717 // Implementation of PreserveJVMState
718 //
719 // Only clone_map(...) here. If this function is only used in the
720 // PreserveJVMState class we may want to get rid of this extra
721 // function eventually and do it all there.
722
723 SafePointNode* GraphKit::clone_map() {
724 if (map() == NULL) return NULL;
725
726 // Clone the memory edge first
727 Node* mem = MergeMemNode::make(map()->memory());
728 gvn().set_type_bottom(mem);
729
730 SafePointNode *clonemap = (SafePointNode*)map()->clone();
731 JVMState* jvms = this->jvms();
732 JVMState* clonejvms = jvms->clone_shallow(C);
733 clonemap->set_memory(mem);
734 clonemap->set_jvms(clonejvms);
735 clonejvms->set_map(clonemap);
736 record_for_igvn(clonemap);
737 gvn().set_type_bottom(clonemap);
738 return clonemap;
739 }
740
741
742 //-----------------------------set_map_clone-----------------------------------
743 void GraphKit::set_map_clone(SafePointNode* m) {
744 _map = m;
745 _map = clone_map();
746 _map->set_next_exception(NULL);
747 debug_only(verify_map());
748 }
749
750
751 //----------------------------kill_dead_locals---------------------------------
752 // Detect any locals which are known to be dead, and force them to top.
753 void GraphKit::kill_dead_locals() {
754 // Consult the liveness information for the locals. If any
755 // of them are unused, then they can be replaced by top(). This
756 // should help register allocation time and cut down on the size
757 // of the deoptimization information.
758
759 // This call is made from many of the bytecode handling
760 // subroutines called from the Big Switch in do_one_bytecode.
761 // Every bytecode which might include a slow path is responsible
762 // for killing its dead locals. The more consistent we
763 // are about killing deads, the fewer useless phis will be
764 // constructed for them at various merge points.
765
766 // bci can be -1 (InvocationEntryBci). We return the entry
767 // liveness for the method.
768
769 if (method() == NULL || method()->code_size() == 0) {
770 // We are building a graph for a call to a native method.
771 // All locals are live.
772 return;
773 }
774
775 ResourceMark rm;
776
777 // Consult the liveness information for the locals. If any
778 // of them are unused, then they can be replaced by top(). This
779 // should help register allocation time and cut down on the size
780 // of the deoptimization information.
781 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
782
783 int len = (int)live_locals.size();
784 assert(len <= jvms()->loc_size(), "too many live locals");
785 for (int local = 0; local < len; local++) {
786 if (!live_locals.at(local)) {
787 set_local(local, top());
788 }
789 }
790 }
791
792 #ifdef ASSERT
793 //-------------------------dead_locals_are_killed------------------------------
794 // Return true if all dead locals are set to top in the map.
795 // Used to assert "clean" debug info at various points.
796 bool GraphKit::dead_locals_are_killed() {
797 if (method() == NULL || method()->code_size() == 0) {
798 // No locals need to be dead, so all is as it should be.
799 return true;
800 }
801
802 // Make sure somebody called kill_dead_locals upstream.
803 ResourceMark rm;
804 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
805 if (jvms->loc_size() == 0) continue; // no locals to consult
806 SafePointNode* map = jvms->map();
807 ciMethod* method = jvms->method();
808 int bci = jvms->bci();
809 if (jvms == this->jvms()) {
810 bci = this->bci(); // it might not yet be synched
811 }
812 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
813 int len = (int)live_locals.size();
814 if (!live_locals.is_valid() || len == 0)
815 // This method is trivial, or is poisoned by a breakpoint.
816 return true;
817 assert(len == jvms->loc_size(), "live map consistent with locals map");
818 for (int local = 0; local < len; local++) {
819 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
820 if (PrintMiscellaneous && (Verbose || WizardMode)) {
821 tty->print_cr("Zombie local %d: ", local);
822 jvms->dump();
823 }
824 return false;
825 }
826 }
827 }
828 return true;
829 }
830
831 #endif //ASSERT
832
833 // Helper function for enforcing certain bytecodes to reexecute if
834 // deoptimization happens
835 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
836 ciMethod* cur_method = jvms->method();
837 int cur_bci = jvms->bci();
838 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
839 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
840 return Interpreter::bytecode_should_reexecute(code) ||
841 (is_anewarray && (code == Bytecodes::_multianewarray));
842 // Reexecute _multianewarray bytecode which was replaced with
843 // sequence of [a]newarray. See Parse::do_multianewarray().
844 //
845 // Note: interpreter should not have it set since this optimization
846 // is limited by dimensions and guarded by flag so in some cases
847 // multianewarray() runtime calls will be generated and
848 // the bytecode should not be reexecutes (stack will not be reset).
849 } else {
850 return false;
851 }
852 }
853
854 // Helper function for adding JVMState and debug information to node
855 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
856 // Add the safepoint edges to the call (or other safepoint).
857
858 // Make sure dead locals are set to top. This
859 // should help register allocation time and cut down on the size
860 // of the deoptimization information.
861 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
862
863 // Walk the inline list to fill in the correct set of JVMState's
864 // Also fill in the associated edges for each JVMState.
865
866 // If the bytecode needs to be reexecuted we need to put
867 // the arguments back on the stack.
868 const bool should_reexecute = jvms()->should_reexecute();
869 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
870
871 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
872 // undefined if the bci is different. This is normal for Parse but it
873 // should not happen for LibraryCallKit because only one bci is processed.
874 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
875 "in LibraryCallKit the reexecute bit should not change");
876
877 // If we are guaranteed to throw, we can prune everything but the
878 // input to the current bytecode.
879 bool can_prune_locals = false;
880 uint stack_slots_not_pruned = 0;
881 int inputs = 0, depth = 0;
882 if (must_throw) {
883 assert(method() == youngest_jvms->method(), "sanity");
884 if (compute_stack_effects(inputs, depth)) {
885 can_prune_locals = true;
886 stack_slots_not_pruned = inputs;
887 }
888 }
889
890 if (env()->should_retain_local_variables()) {
891 // At any safepoint, this method can get breakpointed, which would
892 // then require an immediate deoptimization.
893 can_prune_locals = false; // do not prune locals
894 stack_slots_not_pruned = 0;
895 }
896
897 // do not scribble on the input jvms
898 JVMState* out_jvms = youngest_jvms->clone_deep(C);
899 call->set_jvms(out_jvms); // Start jvms list for call node
900
901 // For a known set of bytecodes, the interpreter should reexecute them if
902 // deoptimization happens. We set the reexecute state for them here
903 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
904 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
905 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
906 }
907
908 // Presize the call:
909 DEBUG_ONLY(uint non_debug_edges = call->req());
910 call->add_req_batch(top(), youngest_jvms->debug_depth());
911 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
912
913 // Set up edges so that the call looks like this:
914 // Call [state:] ctl io mem fptr retadr
915 // [parms:] parm0 ... parmN
916 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
917 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
918 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
919 // Note that caller debug info precedes callee debug info.
920
921 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
922 uint debug_ptr = call->req();
923
924 // Loop over the map input edges associated with jvms, add them
925 // to the call node, & reset all offsets to match call node array.
926 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
927 uint debug_end = debug_ptr;
928 uint debug_start = debug_ptr - in_jvms->debug_size();
929 debug_ptr = debug_start; // back up the ptr
930
931 uint p = debug_start; // walks forward in [debug_start, debug_end)
932 uint j, k, l;
933 SafePointNode* in_map = in_jvms->map();
934 out_jvms->set_map(call);
935
936 if (can_prune_locals) {
937 assert(in_jvms->method() == out_jvms->method(), "sanity");
938 // If the current throw can reach an exception handler in this JVMS,
939 // then we must keep everything live that can reach that handler.
940 // As a quick and dirty approximation, we look for any handlers at all.
941 if (in_jvms->method()->has_exception_handlers()) {
942 can_prune_locals = false;
943 }
944 }
945
946 // Add the Locals
947 k = in_jvms->locoff();
948 l = in_jvms->loc_size();
949 out_jvms->set_locoff(p);
950 if (!can_prune_locals) {
951 for (j = 0; j < l; j++)
952 call->set_req(p++, in_map->in(k+j));
953 } else {
954 p += l; // already set to top above by add_req_batch
955 }
956
957 // Add the Expression Stack
958 k = in_jvms->stkoff();
959 l = in_jvms->sp();
960 out_jvms->set_stkoff(p);
961 if (!can_prune_locals) {
962 for (j = 0; j < l; j++)
963 call->set_req(p++, in_map->in(k+j));
964 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
965 // Divide stack into {S0,...,S1}, where S0 is set to top.
966 uint s1 = stack_slots_not_pruned;
967 stack_slots_not_pruned = 0; // for next iteration
968 if (s1 > l) s1 = l;
969 uint s0 = l - s1;
970 p += s0; // skip the tops preinstalled by add_req_batch
971 for (j = s0; j < l; j++)
972 call->set_req(p++, in_map->in(k+j));
973 } else {
974 p += l; // already set to top above by add_req_batch
975 }
976
977 // Add the Monitors
978 k = in_jvms->monoff();
979 l = in_jvms->mon_size();
980 out_jvms->set_monoff(p);
981 for (j = 0; j < l; j++)
982 call->set_req(p++, in_map->in(k+j));
983
984 // Copy any scalar object fields.
985 k = in_jvms->scloff();
986 l = in_jvms->scl_size();
987 out_jvms->set_scloff(p);
988 for (j = 0; j < l; j++)
989 call->set_req(p++, in_map->in(k+j));
990
991 // Finish the new jvms.
992 out_jvms->set_endoff(p);
993
994 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
995 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
996 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
997 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
998 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
999 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1000
1001 // Update the two tail pointers in parallel.
1002 out_jvms = out_jvms->caller();
1003 in_jvms = in_jvms->caller();
1004 }
1005
1006 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1007
1008 // Test the correctness of JVMState::debug_xxx accessors:
1009 assert(call->jvms()->debug_start() == non_debug_edges, "");
1010 assert(call->jvms()->debug_end() == call->req(), "");
1011 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1012 }
1013
1014 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1015 Bytecodes::Code code = java_bc();
1016 if (code == Bytecodes::_wide) {
1017 code = method()->java_code_at_bci(bci() + 1);
1018 }
1019
1020 BasicType rtype = T_ILLEGAL;
1021 int rsize = 0;
1022
1023 if (code != Bytecodes::_illegal) {
1024 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1025 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1026 if (rtype < T_CONFLICT)
1027 rsize = type2size[rtype];
1028 }
1029
1030 switch (code) {
1031 case Bytecodes::_illegal:
1032 return false;
1033
1034 case Bytecodes::_ldc:
1035 case Bytecodes::_ldc_w:
1036 case Bytecodes::_ldc2_w:
1037 inputs = 0;
1038 break;
1039
1040 case Bytecodes::_dup: inputs = 1; break;
1041 case Bytecodes::_dup_x1: inputs = 2; break;
1042 case Bytecodes::_dup_x2: inputs = 3; break;
1043 case Bytecodes::_dup2: inputs = 2; break;
1044 case Bytecodes::_dup2_x1: inputs = 3; break;
1045 case Bytecodes::_dup2_x2: inputs = 4; break;
1046 case Bytecodes::_swap: inputs = 2; break;
1047 case Bytecodes::_arraylength: inputs = 1; break;
1048
1049 case Bytecodes::_getstatic:
1050 case Bytecodes::_putstatic:
1051 case Bytecodes::_getfield:
1052 case Bytecodes::_putfield:
1053 {
1054 bool ignored_will_link;
1055 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1056 int size = field->type()->size();
1057 bool is_get = (depth >= 0), is_static = (depth & 1);
1058 inputs = (is_static ? 0 : 1);
1059 if (is_get) {
1060 depth = size - inputs;
1061 } else {
1062 inputs += size; // putxxx pops the value from the stack
1063 depth = - inputs;
1064 }
1065 }
1066 break;
1067
1068 case Bytecodes::_invokevirtual:
1069 case Bytecodes::_invokespecial:
1070 case Bytecodes::_invokestatic:
1071 case Bytecodes::_invokedynamic:
1072 case Bytecodes::_invokeinterface:
1073 {
1074 bool ignored_will_link;
1075 ciSignature* declared_signature = NULL;
1076 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1077 assert(declared_signature != NULL, "cannot be null");
1078 inputs = declared_signature->arg_size_for_bc(code);
1079 int size = declared_signature->return_type()->size();
1080 depth = size - inputs;
1081 }
1082 break;
1083
1084 case Bytecodes::_multianewarray:
1085 {
1086 ciBytecodeStream iter(method());
1087 iter.reset_to_bci(bci());
1088 iter.next();
1089 inputs = iter.get_dimensions();
1090 assert(rsize == 1, "");
1091 depth = rsize - inputs;
1092 }
1093 break;
1094
1095 case Bytecodes::_withfield: {
1096 bool ignored_will_link;
1097 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1098 int size = field->type()->size();
1099 inputs = size+1;
1100 depth = rsize - inputs;
1101 break;
1102 }
1103
1104 case Bytecodes::_ireturn:
1105 case Bytecodes::_lreturn:
1106 case Bytecodes::_freturn:
1107 case Bytecodes::_dreturn:
1108 case Bytecodes::_areturn:
1109 assert(rsize == -depth, "");
1110 inputs = rsize;
1111 break;
1112
1113 case Bytecodes::_jsr:
1114 case Bytecodes::_jsr_w:
1115 inputs = 0;
1116 depth = 1; // S.B. depth=1, not zero
1117 break;
1118
1119 default:
1120 // bytecode produces a typed result
1121 inputs = rsize - depth;
1122 assert(inputs >= 0, "");
1123 break;
1124 }
1125
1126 #ifdef ASSERT
1127 // spot check
1128 int outputs = depth + inputs;
1129 assert(outputs >= 0, "sanity");
1130 switch (code) {
1131 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1132 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1133 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1134 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1135 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1136 default: break;
1137 }
1138 #endif //ASSERT
1139
1140 return true;
1141 }
1142
1143
1144
1145 //------------------------------basic_plus_adr---------------------------------
1146 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1147 // short-circuit a common case
1148 if (offset == intcon(0)) return ptr;
1149 return _gvn.transform( new AddPNode(base, ptr, offset) );
1150 }
1151
1152 Node* GraphKit::ConvI2L(Node* offset) {
1153 // short-circuit a common case
1154 jint offset_con = find_int_con(offset, Type::OffsetBot);
1155 if (offset_con != Type::OffsetBot) {
1156 return longcon((jlong) offset_con);
1157 }
1158 return _gvn.transform( new ConvI2LNode(offset));
1159 }
1160
1161 Node* GraphKit::ConvI2UL(Node* offset) {
1162 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1163 if (offset_con != (juint) Type::OffsetBot) {
1164 return longcon((julong) offset_con);
1165 }
1166 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1167 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1168 return _gvn.transform( new AndLNode(conv, mask) );
1169 }
1170
1171 Node* GraphKit::ConvL2I(Node* offset) {
1172 // short-circuit a common case
1173 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1174 if (offset_con != (jlong)Type::OffsetBot) {
1175 return intcon((int) offset_con);
1176 }
1177 return _gvn.transform( new ConvL2INode(offset));
1178 }
1179
1180 //-------------------------load_object_klass-----------------------------------
1181 Node* GraphKit::load_object_klass(Node* obj) {
1182 // Special-case a fresh allocation to avoid building nodes:
1183 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1184 if (akls != NULL) return akls;
1185 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1186 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1187 }
1188
1189 //-------------------------load_array_length-----------------------------------
1190 Node* GraphKit::load_array_length(Node* array) {
1191 // Special-case a fresh allocation to avoid building nodes:
1192 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1193 Node *alen;
1194 if (alloc == NULL) {
1195 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1196 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1197 } else {
1198 alen = alloc->Ideal_length();
1199 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1200 if (ccast != alen) {
1201 alen = _gvn.transform(ccast);
1202 }
1203 }
1204 return alen;
1205 }
1206
1207 //------------------------------do_null_check----------------------------------
1208 // Helper function to do a NULL pointer check. Returned value is
1209 // the incoming address with NULL casted away. You are allowed to use the
1210 // not-null value only if you are control dependent on the test.
1211 #ifndef PRODUCT
1212 extern int explicit_null_checks_inserted,
1213 explicit_null_checks_elided;
1214 #endif
1215 Node* GraphKit::null_check_common(Node* value, BasicType type,
1216 // optional arguments for variations:
1217 bool assert_null,
1218 Node* *null_control,
1219 bool speculative) {
1220 assert(!assert_null || null_control == NULL, "not both at once");
1221 if (stopped()) return top();
1222 NOT_PRODUCT(explicit_null_checks_inserted++);
1223
1224 // Construct NULL check
1225 Node *chk = NULL;
1226 switch(type) {
1227 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1228 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1229 case T_VALUETYPE : // fall through
1230 case T_ARRAY : // fall through
1231 type = T_OBJECT; // simplify further tests
1232 case T_OBJECT : {
1233 const Type *t = _gvn.type( value );
1234
1235 const TypeOopPtr* tp = t->isa_oopptr();
1236 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1237 // Only for do_null_check, not any of its siblings:
1238 && !assert_null && null_control == NULL) {
1239 // Usually, any field access or invocation on an unloaded oop type
1240 // will simply fail to link, since the statically linked class is
1241 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1242 // the static class is loaded but the sharper oop type is not.
1243 // Rather than checking for this obscure case in lots of places,
1244 // we simply observe that a null check on an unloaded class
1245 // will always be followed by a nonsense operation, so we
1246 // can just issue the uncommon trap here.
1247 // Our access to the unloaded class will only be correct
1248 // after it has been loaded and initialized, which requires
1249 // a trip through the interpreter.
1250 #ifndef PRODUCT
1251 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1252 #endif
1253 uncommon_trap(Deoptimization::Reason_unloaded,
1254 Deoptimization::Action_reinterpret,
1255 tp->klass(), "!loaded");
1256 return top();
1257 }
1258
1259 if (assert_null) {
1260 // See if the type is contained in NULL_PTR.
1261 // If so, then the value is already null.
1262 if (t->higher_equal(TypePtr::NULL_PTR)) {
1263 NOT_PRODUCT(explicit_null_checks_elided++);
1264 return value; // Elided null assert quickly!
1265 }
1266 } else {
1267 // See if mixing in the NULL pointer changes type.
1268 // If so, then the NULL pointer was not allowed in the original
1269 // type. In other words, "value" was not-null.
1270 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1271 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1272 NOT_PRODUCT(explicit_null_checks_elided++);
1273 return value; // Elided null check quickly!
1274 }
1275 }
1276 chk = new CmpPNode( value, null() );
1277 break;
1278 }
1279
1280 default:
1281 fatal("unexpected type: %s", type2name(type));
1282 }
1283 assert(chk != NULL, "sanity check");
1284 chk = _gvn.transform(chk);
1285
1286 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1287 BoolNode *btst = new BoolNode( chk, btest);
1288 Node *tst = _gvn.transform( btst );
1289
1290 //-----------
1291 // if peephole optimizations occurred, a prior test existed.
1292 // If a prior test existed, maybe it dominates as we can avoid this test.
1293 if (tst != btst && type == T_OBJECT) {
1294 // At this point we want to scan up the CFG to see if we can
1295 // find an identical test (and so avoid this test altogether).
1296 Node *cfg = control();
1297 int depth = 0;
1298 while( depth < 16 ) { // Limit search depth for speed
1299 if( cfg->Opcode() == Op_IfTrue &&
1300 cfg->in(0)->in(1) == tst ) {
1301 // Found prior test. Use "cast_not_null" to construct an identical
1302 // CastPP (and hence hash to) as already exists for the prior test.
1303 // Return that casted value.
1304 if (assert_null) {
1305 replace_in_map(value, null());
1306 return null(); // do not issue the redundant test
1307 }
1308 Node *oldcontrol = control();
1309 set_control(cfg);
1310 Node *res = cast_not_null(value);
1311 set_control(oldcontrol);
1312 NOT_PRODUCT(explicit_null_checks_elided++);
1313 return res;
1314 }
1315 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1316 if (cfg == NULL) break; // Quit at region nodes
1317 depth++;
1318 }
1319 }
1320
1321 //-----------
1322 // Branch to failure if null
1323 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1324 Deoptimization::DeoptReason reason;
1325 if (assert_null) {
1326 reason = Deoptimization::reason_null_assert(speculative);
1327 } else if (type == T_OBJECT) {
1328 reason = Deoptimization::reason_null_check(speculative);
1329 } else {
1330 reason = Deoptimization::Reason_div0_check;
1331 }
1332 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1333 // ciMethodData::has_trap_at will return a conservative -1 if any
1334 // must-be-null assertion has failed. This could cause performance
1335 // problems for a method after its first do_null_assert failure.
1336 // Consider using 'Reason_class_check' instead?
1337
1338 // To cause an implicit null check, we set the not-null probability
1339 // to the maximum (PROB_MAX). For an explicit check the probability
1340 // is set to a smaller value.
1341 if (null_control != NULL || too_many_traps(reason)) {
1342 // probability is less likely
1343 ok_prob = PROB_LIKELY_MAG(3);
1344 } else if (!assert_null &&
1345 (ImplicitNullCheckThreshold > 0) &&
1346 method() != NULL &&
1347 (method()->method_data()->trap_count(reason)
1348 >= (uint)ImplicitNullCheckThreshold)) {
1349 ok_prob = PROB_LIKELY_MAG(3);
1350 }
1351
1352 if (null_control != NULL) {
1353 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1354 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1355 set_control( _gvn.transform( new IfTrueNode(iff)));
1356 #ifndef PRODUCT
1357 if (null_true == top()) {
1358 explicit_null_checks_elided++;
1359 }
1360 #endif
1361 (*null_control) = null_true;
1362 } else {
1363 BuildCutout unless(this, tst, ok_prob);
1364 // Check for optimizer eliding test at parse time
1365 if (stopped()) {
1366 // Failure not possible; do not bother making uncommon trap.
1367 NOT_PRODUCT(explicit_null_checks_elided++);
1368 } else if (assert_null) {
1369 uncommon_trap(reason,
1370 Deoptimization::Action_make_not_entrant,
1371 NULL, "assert_null");
1372 } else {
1373 replace_in_map(value, zerocon(type));
1374 builtin_throw(reason);
1375 }
1376 }
1377
1378 // Must throw exception, fall-thru not possible?
1379 if (stopped()) {
1380 return top(); // No result
1381 }
1382
1383 if (assert_null) {
1384 // Cast obj to null on this path.
1385 replace_in_map(value, zerocon(type));
1386 return zerocon(type);
1387 }
1388
1389 // Cast obj to not-null on this path, if there is no null_control.
1390 // (If there is a null_control, a non-null value may come back to haunt us.)
1391 if (type == T_OBJECT) {
1392 Node* cast = cast_not_null(value, false);
1393 if (null_control == NULL || (*null_control) == top())
1394 replace_in_map(value, cast);
1395 value = cast;
1396 }
1397
1398 return value;
1399 }
1400
1401
1402 //------------------------------cast_not_null----------------------------------
1403 // Cast obj to not-null on this path
1404 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1405 const Type *t = _gvn.type(obj);
1406 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1407 // Object is already not-null?
1408 if( t == t_not_null ) return obj;
1409
1410 Node *cast = new CastPPNode(obj,t_not_null);
1411 cast->init_req(0, control());
1412 cast = _gvn.transform( cast );
1413
1414 // Scan for instances of 'obj' in the current JVM mapping.
1415 // These instances are known to be not-null after the test.
1416 if (do_replace_in_map)
1417 replace_in_map(obj, cast);
1418
1419 return cast; // Return casted value
1420 }
1421
1422 // Sometimes in intrinsics, we implicitly know an object is not null
1423 // (there's no actual null check) so we can cast it to not null. In
1424 // the course of optimizations, the input to the cast can become null.
1425 // In that case that data path will die and we need the control path
1426 // to become dead as well to keep the graph consistent. So we have to
1427 // add a check for null for which one branch can't be taken. It uses
1428 // an Opaque4 node that will cause the check to be removed after loop
1429 // opts so the test goes away and the compiled code doesn't execute a
1430 // useless check.
1431 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1432 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1433 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1434 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1435 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1436 _gvn.set_type(iff, iff->Value(&_gvn));
1437 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1438 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1439 Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1440 C->root()->add_req(halt);
1441 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1442 set_control(if_t);
1443 return cast_not_null(value, do_replace_in_map);
1444 }
1445
1446
1447 //--------------------------replace_in_map-------------------------------------
1448 void GraphKit::replace_in_map(Node* old, Node* neww) {
1449 if (old == neww) {
1450 return;
1451 }
1452
1453 map()->replace_edge(old, neww);
1454
1455 // Note: This operation potentially replaces any edge
1456 // on the map. This includes locals, stack, and monitors
1457 // of the current (innermost) JVM state.
1458
1459 // don't let inconsistent types from profiling escape this
1460 // method
1461
1462 const Type* told = _gvn.type(old);
1463 const Type* tnew = _gvn.type(neww);
1464
1465 if (!tnew->higher_equal(told)) {
1466 return;
1467 }
1468
1469 map()->record_replaced_node(old, neww);
1470 }
1471
1472
1473 //=============================================================================
1474 //--------------------------------memory---------------------------------------
1475 Node* GraphKit::memory(uint alias_idx) {
1476 MergeMemNode* mem = merged_memory();
1477 Node* p = mem->memory_at(alias_idx);
1478 _gvn.set_type(p, Type::MEMORY); // must be mapped
1479 return p;
1480 }
1481
1482 //-----------------------------reset_memory------------------------------------
1483 Node* GraphKit::reset_memory() {
1484 Node* mem = map()->memory();
1485 // do not use this node for any more parsing!
1486 debug_only( map()->set_memory((Node*)NULL) );
1487 return _gvn.transform( mem );
1488 }
1489
1490 //------------------------------set_all_memory---------------------------------
1491 void GraphKit::set_all_memory(Node* newmem) {
1492 Node* mergemem = MergeMemNode::make(newmem);
1493 gvn().set_type_bottom(mergemem);
1494 map()->set_memory(mergemem);
1495 }
1496
1497 //------------------------------set_all_memory_call----------------------------
1498 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1499 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1500 set_all_memory(newmem);
1501 }
1502
1503 //=============================================================================
1504 //
1505 // parser factory methods for MemNodes
1506 //
1507 // These are layered on top of the factory methods in LoadNode and StoreNode,
1508 // and integrate with the parser's memory state and _gvn engine.
1509 //
1510
1511 // factory methods in "int adr_idx"
1512 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1513 int adr_idx,
1514 MemNode::MemOrd mo,
1515 LoadNode::ControlDependency control_dependency,
1516 bool require_atomic_access,
1517 bool unaligned,
1518 bool mismatched) {
1519 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1520 const TypePtr* adr_type = NULL; // debug-mode-only argument
1521 debug_only(adr_type = C->get_adr_type(adr_idx));
1522 Node* mem = memory(adr_idx);
1523 Node* ld;
1524 if (require_atomic_access && bt == T_LONG) {
1525 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1526 } else if (require_atomic_access && bt == T_DOUBLE) {
1527 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1528 } else {
1529 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1530 }
1531 ld = _gvn.transform(ld);
1532 if (bt == T_VALUETYPE) {
1533 // Loading a non-flattened (but flattenable) value type from memory
1534 // We need to return the default value type if the field is null
1535 ld = ValueTypeNode::make_from_oop(this, ld, t->make_oopptr()->value_klass(), true /* null check */);
1536 } else if (bt == T_VALUETYPEPTR) {
1537 // Loading non-flattenable value type from memory
1538 inc_sp(1);
1539 Node* null_ctl = top();
1540 Node* not_null_obj = null_check_common(ld, T_VALUETYPE, false, &null_ctl, false);
1541 if (null_ctl != top()) {
1542 // TODO For now, we just deoptimize if value type is NULL
1543 PreserveJVMState pjvms(this);
1544 set_control(null_ctl);
1545 replace_in_map(ld, null());
1546 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none);
1547 }
1548 dec_sp(1);
1549 ld = ValueTypeNode::make_from_oop(this, not_null_obj, t->make_oopptr()->value_klass());
1550 } else if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1551 // Improve graph before escape analysis and boxing elimination.
1552 record_for_igvn(ld);
1553 }
1554 return ld;
1555 }
1556
1557 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1558 int adr_idx,
1559 MemNode::MemOrd mo,
1560 bool require_atomic_access,
1561 bool unaligned,
1562 bool mismatched) {
1563 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1564 const TypePtr* adr_type = NULL;
1565 debug_only(adr_type = C->get_adr_type(adr_idx));
1566 Node *mem = memory(adr_idx);
1567 Node* st;
1568 if (require_atomic_access && bt == T_LONG) {
1569 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1570 } else if (require_atomic_access && bt == T_DOUBLE) {
1571 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1572 } else {
1573 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1574 }
1575 if (unaligned) {
1576 st->as_Store()->set_unaligned_access();
1577 }
1578 if (mismatched) {
1579 st->as_Store()->set_mismatched_access();
1580 }
1581 st = _gvn.transform(st);
1582 set_memory(st, adr_idx);
1583 // Back-to-back stores can only remove intermediate store with DU info
1584 // so push on worklist for optimizer.
1585 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1586 record_for_igvn(st);
1587
1588 return st;
1589 }
1590
1591
1592 void GraphKit::pre_barrier(bool do_load,
1593 Node* ctl,
1594 Node* obj,
1595 Node* adr,
1596 uint adr_idx,
1597 Node* val,
1598 const TypeOopPtr* val_type,
1599 Node* pre_val,
1600 BasicType bt) {
1601
1602 BarrierSet* bs = BarrierSet::barrier_set();
1603 set_control(ctl);
1604 switch (bs->kind()) {
1605
1606 #if INCLUDE_G1GC
1607 case BarrierSet::G1BarrierSet:
1608 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1609 break;
1610 #endif
1611
1612 case BarrierSet::CardTableBarrierSet:
1613 break;
1614
1615 default :
1616 ShouldNotReachHere();
1617
1618 }
1619 }
1620
1621 bool GraphKit::can_move_pre_barrier() const {
1622 BarrierSet* bs = BarrierSet::barrier_set();
1623 switch (bs->kind()) {
1624
1625 #if INCLUDE_G1GC
1626 case BarrierSet::G1BarrierSet:
1627 return true; // Can move it if no safepoint
1628 #endif
1629
1630 case BarrierSet::CardTableBarrierSet:
1631 return true; // There is no pre-barrier
1632
1633 default :
1634 ShouldNotReachHere();
1635 }
1636 return false;
1637 }
1638
1639 void GraphKit::post_barrier(Node* ctl,
1640 Node* store,
1641 Node* obj,
1642 Node* adr,
1643 uint adr_idx,
1644 Node* val,
1645 BasicType bt,
1646 bool use_precise) {
1647 BarrierSet* bs = BarrierSet::barrier_set();
1648 set_control(ctl);
1649 switch (bs->kind()) {
1650 #if INCLUDE_G1GC
1651 case BarrierSet::G1BarrierSet:
1652 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1653 break;
1654 #endif
1655
1656 case BarrierSet::CardTableBarrierSet:
1657 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1658 break;
1659
1660 default :
1661 ShouldNotReachHere();
1662
1663 }
1664 }
1665
1666 Node* GraphKit::store_oop(Node* ctl,
1667 Node* obj,
1668 Node* adr,
1669 const TypePtr* adr_type,
1670 Node* val,
1671 const TypeOopPtr* val_type,
1672 BasicType bt,
1673 bool use_precise,
1674 MemNode::MemOrd mo,
1675 bool mismatched,
1676 bool deoptimize_on_exception) {
1677 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1678 // could be delayed during Parse (for example, in adjust_map_after_if()).
1679 // Execute transformation here to avoid barrier generation in such case.
1680 if (_gvn.type(val) == TypePtr::NULL_PTR)
1681 val = _gvn.makecon(TypePtr::NULL_PTR);
1682
1683 set_control(ctl);
1684 if (stopped()) return top(); // Dead path ?
1685
1686 assert(bt == T_OBJECT || bt == T_VALUETYPE, "sanity");
1687 assert(val != NULL, "not dead path");
1688 uint adr_idx = C->get_alias_index(adr_type);
1689 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1690
1691 if (val->is_ValueType()) {
1692 // Allocate value type and get oop
1693 val = val->as_ValueType()->allocate(this, deoptimize_on_exception)->get_oop();
1694 }
1695
1696 pre_barrier(true /* do_load */,
1697 control(), obj, adr, adr_idx, val, val_type,
1698 NULL /* pre_val */,
1699 bt);
1700
1701 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1702 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1703 return store;
1704 }
1705
1706 // Could be an array or object we don't know at compile time (unsafe ref.)
1707 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1708 Node* obj, // containing obj
1709 Node* adr, // actual adress to store val at
1710 const TypePtr* adr_type,
1711 Node* val,
1712 BasicType bt,
1713 MemNode::MemOrd mo,
1714 bool mismatched) {
1715 Compile::AliasType* at = C->alias_type(adr_type);
1716 const TypeOopPtr* val_type = NULL;
1717 if (adr_type->isa_instptr()) {
1718 if (at->field() != NULL) {
1719 // known field. This code is a copy of the do_put_xxx logic.
1720 ciField* field = at->field();
1721 if (!field->type()->is_loaded()) {
1722 val_type = TypeInstPtr::BOTTOM;
1723 } else {
1724 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1725 }
1726 }
1727 } else if (adr_type->isa_aryptr()) {
1728 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1729 }
1730 if (val_type == NULL) {
1731 val_type = TypeInstPtr::BOTTOM;
1732 }
1733 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1734 }
1735
1736
1737 //-------------------------array_element_address-------------------------
1738 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1739 const TypeInt* sizetype, Node* ctrl) {
1740 uint shift = exact_log2(type2aelembytes(elembt));
1741 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1742 if (arytype_klass->is_value_array_klass()) {
1743 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass();
1744 shift = vak->log2_element_size();
1745 }
1746 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1747
1748 // short-circuit a common case (saves lots of confusing waste motion)
1749 jint idx_con = find_int_con(idx, -1);
1750 if (idx_con >= 0) {
1751 intptr_t offset = header + ((intptr_t)idx_con << shift);
1752 return basic_plus_adr(ary, offset);
1753 }
1754
1755 // must be correct type for alignment purposes
1756 Node* base = basic_plus_adr(ary, header);
1757 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1758 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1759 return basic_plus_adr(ary, base, scale);
1760 }
1761
1762 //-------------------------load_array_element-------------------------
1763 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1764 const Type* elemtype = arytype->elem();
1765 BasicType elembt = elemtype->array_element_basic_type();
1766 assert(elembt != T_VALUETYPE, "value types are not supported by this method");
1767 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1768 if (elembt == T_NARROWOOP) {
1769 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1770 }
1771 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1772 return ld;
1773 }
1774
1775 //-------------------------set_arguments_for_java_call-------------------------
1776 // Arguments (pre-popped from the stack) are taken from the JVMS.
1777 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1778 // Add the call arguments:
1779 const TypeTuple* domain = call->tf()->domain_sig();
1780 uint nargs = domain->cnt();
1781 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1782 Node* arg = argument(i-TypeFunc::Parms);
1783 if (ValueTypePassFieldsAsArgs) {
1784 if (arg->is_ValueType()) {
1785 ValueTypeNode* vt = arg->as_ValueType();
1786 // TODO fix this with the calling convention changes
1787 if (true /*!domain->field_at(i)->is_valuetypeptr()->is__Value()*/) {
1788 // We don't pass value type arguments by reference but instead
1789 // pass each field of the value type
1790 idx += vt->pass_fields(call, idx, *this);
1791 // If a value type argument is passed as fields, attach the Method* to the call site
1792 // to be able to access the extended signature later via attached_method_before_pc().
1793 // For example, see CompiledMethod::preserve_callee_argument_oops().
1794 call->set_override_symbolic_info(true);
1795 } else {
1796 arg = arg->as_ValueType()->allocate(this)->get_oop();
1797 call->init_req(idx, arg);
1798 idx++;
1799 }
1800 } else {
1801 call->init_req(idx, arg);
1802 idx++;
1803 }
1804 } else {
1805 if (arg->is_ValueType()) {
1806 // Pass value type argument via oop to callee
1807 arg = arg->as_ValueType()->allocate(this)->get_oop();
1808 }
1809 call->init_req(i, arg);
1810 }
1811 }
1812 }
1813
1814 //---------------------------set_edges_for_java_call---------------------------
1815 // Connect a newly created call into the current JVMS.
1816 // A return value node (if any) is returned from set_edges_for_java_call.
1817 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1818
1819 // Add the predefined inputs:
1820 call->init_req( TypeFunc::Control, control() );
1821 call->init_req( TypeFunc::I_O , i_o() );
1822 call->init_req( TypeFunc::Memory , reset_memory() );
1823 call->init_req( TypeFunc::FramePtr, frameptr() );
1824 call->init_req( TypeFunc::ReturnAdr, top() );
1825
1826 add_safepoint_edges(call, must_throw);
1827
1828 Node* xcall = _gvn.transform(call);
1829
1830 if (xcall == top()) {
1831 set_control(top());
1832 return;
1833 }
1834 assert(xcall == call, "call identity is stable");
1835
1836 // Re-use the current map to produce the result.
1837
1838 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1839 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1840 set_all_memory_call(xcall, separate_io_proj);
1841
1842 //return xcall; // no need, caller already has it
1843 }
1844
1845 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1846 if (stopped()) return top(); // maybe the call folded up?
1847
1848 // Note: Since any out-of-line call can produce an exception,
1849 // we always insert an I_O projection from the call into the result.
1850
1851 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1852
1853 if (separate_io_proj) {
1854 // The caller requested separate projections be used by the fall
1855 // through and exceptional paths, so replace the projections for
1856 // the fall through path.
1857 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1858 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1859 }
1860
1861 // Capture the return value, if any.
1862 Node* ret;
1863 if (call->method() == NULL ||
1864 call->method()->return_type()->basic_type() == T_VOID) {
1865 ret = top();
1866 } else {
1867 if (!call->tf()->returns_value_type_as_fields()) {
1868 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1869 } else {
1870 // Return of multiple values (value type fields): we create a
1871 // ValueType node, each field is a projection from the call.
1872 const TypeTuple* range_sig = call->tf()->range_sig();
1873 const Type* t = range_sig->field_at(TypeFunc::Parms);
1874 assert(t->is_valuetypeptr(), "only value types for multiple return values");
1875 ciValueKlass* vk = t->value_klass();
1876 Node* ctl = control();
1877 ret = ValueTypeNode::make_from_multi(_gvn, ctl, merged_memory(), call, vk, TypeFunc::Parms+1, false);
1878 set_control(ctl);
1879 }
1880 }
1881
1882 return ret;
1883 }
1884
1885 //--------------------set_predefined_input_for_runtime_call--------------------
1886 // Reading and setting the memory state is way conservative here.
1887 // The real problem is that I am not doing real Type analysis on memory,
1888 // so I cannot distinguish card mark stores from other stores. Across a GC
1889 // point the Store Barrier and the card mark memory has to agree. I cannot
1890 // have a card mark store and its barrier split across the GC point from
1891 // either above or below. Here I get that to happen by reading ALL of memory.
1892 // A better answer would be to separate out card marks from other memory.
1893 // For now, return the input memory state, so that it can be reused
1894 // after the call, if this call has restricted memory effects.
1895 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1896 // Set fixed predefined input arguments
1897 Node* memory = reset_memory();
1898 call->init_req( TypeFunc::Control, control() );
1899 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1900 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1901 call->init_req( TypeFunc::FramePtr, frameptr() );
1902 call->init_req( TypeFunc::ReturnAdr, top() );
1903 return memory;
1904 }
1905
1906 //-------------------set_predefined_output_for_runtime_call--------------------
1907 // Set control and memory (not i_o) from the call.
1908 // If keep_mem is not NULL, use it for the output state,
1909 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1910 // If hook_mem is NULL, this call produces no memory effects at all.
1911 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1912 // then only that memory slice is taken from the call.
1913 // In the last case, we must put an appropriate memory barrier before
1914 // the call, so as to create the correct anti-dependencies on loads
1915 // preceding the call.
1916 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1917 Node* keep_mem,
1918 const TypePtr* hook_mem) {
1919 // no i/o
1920 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1921 if (keep_mem) {
1922 // First clone the existing memory state
1923 set_all_memory(keep_mem);
1924 if (hook_mem != NULL) {
1925 // Make memory for the call
1926 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1927 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1928 // We also use hook_mem to extract specific effects from arraycopy stubs.
1929 set_memory(mem, hook_mem);
1930 }
1931 // ...else the call has NO memory effects.
1932
1933 // Make sure the call advertises its memory effects precisely.
1934 // This lets us build accurate anti-dependences in gcm.cpp.
1935 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1936 "call node must be constructed correctly");
1937 } else {
1938 assert(hook_mem == NULL, "");
1939 // This is not a "slow path" call; all memory comes from the call.
1940 set_all_memory_call(call);
1941 }
1942 }
1943
1944
1945 // Replace the call with the current state of the kit.
1946 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1947 JVMState* ejvms = NULL;
1948 if (has_exceptions()) {
1949 ejvms = transfer_exceptions_into_jvms();
1950 }
1951
1952 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1953 ReplacedNodes replaced_nodes_exception;
1954 Node* ex_ctl = top();
1955
1956 SafePointNode* final_state = stop();
1957
1958 // Find all the needed outputs of this call
1959 CallProjections* callprojs = call->extract_projections(true);
1960
1961 Node* init_mem = call->in(TypeFunc::Memory);
1962 Node* final_mem = final_state->in(TypeFunc::Memory);
1963 Node* final_ctl = final_state->in(TypeFunc::Control);
1964 Node* final_io = final_state->in(TypeFunc::I_O);
1965
1966 // Replace all the old call edges with the edges from the inlining result
1967 if (callprojs->fallthrough_catchproj != NULL) {
1968 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
1969 }
1970 if (callprojs->fallthrough_memproj != NULL) {
1971 if (final_mem->is_MergeMem()) {
1972 // Parser's exits MergeMem was not transformed but may be optimized
1973 final_mem = _gvn.transform(final_mem);
1974 }
1975 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
1976 }
1977 if (callprojs->fallthrough_ioproj != NULL) {
1978 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
1979 }
1980
1981 // Replace the result with the new result if it exists and is used
1982 if (callprojs->resproj[0] != NULL && result != NULL) {
1983 assert(callprojs->nb_resproj == 1, "unexpected number of results");
1984 C->gvn_replace_by(callprojs->resproj[0], result);
1985 }
1986
1987 if (ejvms == NULL) {
1988 // No exception edges to simply kill off those paths
1989 if (callprojs->catchall_catchproj != NULL) {
1990 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
1991 }
1992 if (callprojs->catchall_memproj != NULL) {
1993 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
1994 }
1995 if (callprojs->catchall_ioproj != NULL) {
1996 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
1997 }
1998 // Replace the old exception object with top
1999 if (callprojs->exobj != NULL) {
2000 C->gvn_replace_by(callprojs->exobj, C->top());
2001 }
2002 } else {
2003 GraphKit ekit(ejvms);
2004
2005 // Load my combined exception state into the kit, with all phis transformed:
2006 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2007 replaced_nodes_exception = ex_map->replaced_nodes();
2008
2009 Node* ex_oop = ekit.use_exception_state(ex_map);
2010
2011 if (callprojs->catchall_catchproj != NULL) {
2012 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2013 ex_ctl = ekit.control();
2014 }
2015 if (callprojs->catchall_memproj != NULL) {
2016 C->gvn_replace_by(callprojs->catchall_memproj, ekit.reset_memory());
2017 }
2018 if (callprojs->catchall_ioproj != NULL) {
2019 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2020 }
2021
2022 // Replace the old exception object with the newly created one
2023 if (callprojs->exobj != NULL) {
2024 C->gvn_replace_by(callprojs->exobj, ex_oop);
2025 }
2026 }
2027
2028 // Disconnect the call from the graph
2029 call->disconnect_inputs(NULL, C);
2030 C->gvn_replace_by(call, C->top());
2031
2032 // Clean up any MergeMems that feed other MergeMems since the
2033 // optimizer doesn't like that.
2034 if (final_mem->is_MergeMem()) {
2035 Node_List wl;
2036 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
2037 Node* m = i.get();
2038 if (m->is_MergeMem() && !wl.contains(m)) {
2039 wl.push(m);
2040 }
2041 }
2042 while (wl.size() > 0) {
2043 _gvn.transform(wl.pop());
2044 }
2045 }
2046
2047 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2048 replaced_nodes.apply(C, final_ctl);
2049 }
2050 if (!ex_ctl->is_top() && do_replaced_nodes) {
2051 replaced_nodes_exception.apply(C, ex_ctl);
2052 }
2053 }
2054
2055
2056 //------------------------------increment_counter------------------------------
2057 // for statistics: increment a VM counter by 1
2058
2059 void GraphKit::increment_counter(address counter_addr) {
2060 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2061 increment_counter(adr1);
2062 }
2063
2064 void GraphKit::increment_counter(Node* counter_addr) {
2065 int adr_type = Compile::AliasIdxRaw;
2066 Node* ctrl = control();
2067 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2068 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
2069 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
2070 }
2071
2072
2073 //------------------------------uncommon_trap----------------------------------
2074 // Bail out to the interpreter in mid-method. Implemented by calling the
2075 // uncommon_trap blob. This helper function inserts a runtime call with the
2076 // right debug info.
2077 void GraphKit::uncommon_trap(int trap_request,
2078 ciKlass* klass, const char* comment,
2079 bool must_throw,
2080 bool keep_exact_action) {
2081 if (failing()) stop();
2082 if (stopped()) return; // trap reachable?
2083
2084 // Note: If ProfileTraps is true, and if a deopt. actually
2085 // occurs here, the runtime will make sure an MDO exists. There is
2086 // no need to call method()->ensure_method_data() at this point.
2087
2088 // Set the stack pointer to the right value for reexecution:
2089 set_sp(reexecute_sp());
2090
2091 #ifdef ASSERT
2092 if (!must_throw) {
2093 // Make sure the stack has at least enough depth to execute
2094 // the current bytecode.
2095 int inputs, ignored_depth;
2096 if (compute_stack_effects(inputs, ignored_depth)) {
2097 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2098 Bytecodes::name(java_bc()), sp(), inputs);
2099 }
2100 }
2101 #endif
2102
2103 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2104 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2105
2106 switch (action) {
2107 case Deoptimization::Action_maybe_recompile:
2108 case Deoptimization::Action_reinterpret:
2109 // Temporary fix for 6529811 to allow virtual calls to be sure they
2110 // get the chance to go from mono->bi->mega
2111 if (!keep_exact_action &&
2112 Deoptimization::trap_request_index(trap_request) < 0 &&
2113 too_many_recompiles(reason)) {
2114 // This BCI is causing too many recompilations.
2115 if (C->log() != NULL) {
2116 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2117 Deoptimization::trap_reason_name(reason),
2118 Deoptimization::trap_action_name(action));
2119 }
2120 action = Deoptimization::Action_none;
2121 trap_request = Deoptimization::make_trap_request(reason, action);
2122 } else {
2123 C->set_trap_can_recompile(true);
2124 }
2125 break;
2126 case Deoptimization::Action_make_not_entrant:
2127 C->set_trap_can_recompile(true);
2128 break;
2129 case Deoptimization::Action_none:
2130 case Deoptimization::Action_make_not_compilable:
2131 break;
2132 default:
2133 #ifdef ASSERT
2134 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2135 #endif
2136 break;
2137 }
2138
2139 if (TraceOptoParse) {
2140 char buf[100];
2141 tty->print_cr("Uncommon trap %s at bci:%d",
2142 Deoptimization::format_trap_request(buf, sizeof(buf),
2143 trap_request), bci());
2144 }
2145
2146 CompileLog* log = C->log();
2147 if (log != NULL) {
2148 int kid = (klass == NULL)? -1: log->identify(klass);
2149 log->begin_elem("uncommon_trap bci='%d'", bci());
2150 char buf[100];
2151 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2152 trap_request));
2153 if (kid >= 0) log->print(" klass='%d'", kid);
2154 if (comment != NULL) log->print(" comment='%s'", comment);
2155 log->end_elem();
2156 }
2157
2158 // Make sure any guarding test views this path as very unlikely
2159 Node *i0 = control()->in(0);
2160 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2161 IfNode *iff = i0->as_If();
2162 float f = iff->_prob; // Get prob
2163 if (control()->Opcode() == Op_IfTrue) {
2164 if (f > PROB_UNLIKELY_MAG(4))
2165 iff->_prob = PROB_MIN;
2166 } else {
2167 if (f < PROB_LIKELY_MAG(4))
2168 iff->_prob = PROB_MAX;
2169 }
2170 }
2171
2172 // Clear out dead values from the debug info.
2173 kill_dead_locals();
2174
2175 // Now insert the uncommon trap subroutine call
2176 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2177 const TypePtr* no_memory_effects = NULL;
2178 // Pass the index of the class to be loaded
2179 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2180 (must_throw ? RC_MUST_THROW : 0),
2181 OptoRuntime::uncommon_trap_Type(),
2182 call_addr, "uncommon_trap", no_memory_effects,
2183 intcon(trap_request));
2184 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2185 "must extract request correctly from the graph");
2186 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2187
2188 call->set_req(TypeFunc::ReturnAdr, returnadr());
2189 // The debug info is the only real input to this call.
2190
2191 // Halt-and-catch fire here. The above call should never return!
2192 HaltNode* halt = new HaltNode(control(), frameptr());
2193 _gvn.set_type_bottom(halt);
2194 root()->add_req(halt);
2195
2196 stop_and_kill_map();
2197 }
2198
2199
2200 //--------------------------just_allocated_object------------------------------
2201 // Report the object that was just allocated.
2202 // It must be the case that there are no intervening safepoints.
2203 // We use this to determine if an object is so "fresh" that
2204 // it does not require card marks.
2205 Node* GraphKit::just_allocated_object(Node* current_control) {
2206 if (C->recent_alloc_ctl() == current_control)
2207 return C->recent_alloc_obj();
2208 return NULL;
2209 }
2210
2211
2212 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2213 // (Note: TypeFunc::make has a cache that makes this fast.)
2214 const TypeFunc* tf = TypeFunc::make(dest_method);
2215 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2216 for (int j = 0; j < nargs; j++) {
2217 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2218 if( targ->basic_type() == T_DOUBLE ) {
2219 // If any parameters are doubles, they must be rounded before
2220 // the call, dstore_rounding does gvn.transform
2221 Node *arg = argument(j);
2222 arg = dstore_rounding(arg);
2223 set_argument(j, arg);
2224 }
2225 }
2226 }
2227
2228 /**
2229 * Record profiling data exact_kls for Node n with the type system so
2230 * that it can propagate it (speculation)
2231 *
2232 * @param n node that the type applies to
2233 * @param exact_kls type from profiling
2234 * @param maybe_null did profiling see null?
2235 *
2236 * @return node with improved type
2237 */
2238 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2239 const Type* current_type = _gvn.type(n);
2240 assert(UseTypeSpeculation, "type speculation must be on");
2241
2242 const TypePtr* speculative = current_type->speculative();
2243
2244 // Should the klass from the profile be recorded in the speculative type?
2245 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2246 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2247 const TypeOopPtr* xtype = tklass->as_instance_type();
2248 assert(xtype->klass_is_exact(), "Should be exact");
2249 // Any reason to believe n is not null (from this profiling or a previous one)?
2250 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2251 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2252 // record the new speculative type's depth
2253 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2254 speculative = speculative->with_inline_depth(jvms()->depth());
2255 } else if (current_type->would_improve_ptr(ptr_kind)) {
2256 // Profiling report that null was never seen so we can change the
2257 // speculative type to non null ptr.
2258 if (ptr_kind == ProfileAlwaysNull) {
2259 speculative = TypePtr::NULL_PTR;
2260 } else {
2261 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2262 const TypePtr* ptr = TypePtr::NOTNULL;
2263 if (speculative != NULL) {
2264 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2265 } else {
2266 speculative = ptr;
2267 }
2268 }
2269 }
2270
2271 if (speculative != current_type->speculative()) {
2272 // Build a type with a speculative type (what we think we know
2273 // about the type but will need a guard when we use it)
2274 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2275 // We're changing the type, we need a new CheckCast node to carry
2276 // the new type. The new type depends on the control: what
2277 // profiling tells us is only valid from here as far as we can
2278 // tell.
2279 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2280 cast = _gvn.transform(cast);
2281 replace_in_map(n, cast);
2282 n = cast;
2283 }
2284
2285 return n;
2286 }
2287
2288 /**
2289 * Record profiling data from receiver profiling at an invoke with the
2290 * type system so that it can propagate it (speculation)
2291 *
2292 * @param n receiver node
2293 *
2294 * @return node with improved type
2295 */
2296 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2297 if (!UseTypeSpeculation) {
2298 return n;
2299 }
2300 ciKlass* exact_kls = profile_has_unique_klass();
2301 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2302 if ((java_bc() == Bytecodes::_checkcast ||
2303 java_bc() == Bytecodes::_instanceof ||
2304 java_bc() == Bytecodes::_aastore) &&
2305 method()->method_data()->is_mature()) {
2306 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2307 if (data != NULL) {
2308 if (!data->as_BitData()->null_seen()) {
2309 ptr_kind = ProfileNeverNull;
2310 } else {
2311 assert(data->is_ReceiverTypeData(), "bad profile data type");
2312 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2313 uint i = 0;
2314 for (; i < call->row_limit(); i++) {
2315 ciKlass* receiver = call->receiver(i);
2316 if (receiver != NULL) {
2317 break;
2318 }
2319 }
2320 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2321 }
2322 }
2323 }
2324 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2325 }
2326
2327 /**
2328 * Record profiling data from argument profiling at an invoke with the
2329 * type system so that it can propagate it (speculation)
2330 *
2331 * @param dest_method target method for the call
2332 * @param bc what invoke bytecode is this?
2333 */
2334 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2335 if (!UseTypeSpeculation) {
2336 return;
2337 }
2338 const TypeFunc* tf = TypeFunc::make(dest_method);
2339 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2340 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2341 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2342 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2343 if (targ->isa_oopptr()) {
2344 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2345 ciKlass* better_type = NULL;
2346 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2347 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2348 }
2349 i++;
2350 }
2351 }
2352 }
2353
2354 /**
2355 * Record profiling data from parameter profiling at an invoke with
2356 * the type system so that it can propagate it (speculation)
2357 */
2358 void GraphKit::record_profiled_parameters_for_speculation() {
2359 if (!UseTypeSpeculation) {
2360 return;
2361 }
2362 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2363 if (_gvn.type(local(i))->isa_oopptr()) {
2364 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2365 ciKlass* better_type = NULL;
2366 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2367 record_profile_for_speculation(local(i), better_type, ptr_kind);
2368 }
2369 j++;
2370 }
2371 }
2372 }
2373
2374 /**
2375 * Record profiling data from return value profiling at an invoke with
2376 * the type system so that it can propagate it (speculation)
2377 */
2378 void GraphKit::record_profiled_return_for_speculation() {
2379 if (!UseTypeSpeculation) {
2380 return;
2381 }
2382 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2383 ciKlass* better_type = NULL;
2384 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2385 // If profiling reports a single type for the return value,
2386 // feed it to the type system so it can propagate it as a
2387 // speculative type
2388 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2389 }
2390 }
2391
2392 void GraphKit::round_double_result(ciMethod* dest_method) {
2393 // A non-strict method may return a double value which has an extended
2394 // exponent, but this must not be visible in a caller which is 'strict'
2395 // If a strict caller invokes a non-strict callee, round a double result
2396
2397 BasicType result_type = dest_method->return_type()->basic_type();
2398 assert( method() != NULL, "must have caller context");
2399 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2400 // Destination method's return value is on top of stack
2401 // dstore_rounding() does gvn.transform
2402 Node *result = pop_pair();
2403 result = dstore_rounding(result);
2404 push_pair(result);
2405 }
2406 }
2407
2408 // rounding for strict float precision conformance
2409 Node* GraphKit::precision_rounding(Node* n) {
2410 return UseStrictFP && _method->flags().is_strict()
2411 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2412 ? _gvn.transform( new RoundFloatNode(0, n) )
2413 : n;
2414 }
2415
2416 // rounding for strict double precision conformance
2417 Node* GraphKit::dprecision_rounding(Node *n) {
2418 return UseStrictFP && _method->flags().is_strict()
2419 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2420 ? _gvn.transform( new RoundDoubleNode(0, n) )
2421 : n;
2422 }
2423
2424 // rounding for non-strict double stores
2425 Node* GraphKit::dstore_rounding(Node* n) {
2426 return Matcher::strict_fp_requires_explicit_rounding
2427 && UseSSE <= 1
2428 ? _gvn.transform( new RoundDoubleNode(0, n) )
2429 : n;
2430 }
2431
2432 //=============================================================================
2433 // Generate a fast path/slow path idiom. Graph looks like:
2434 // [foo] indicates that 'foo' is a parameter
2435 //
2436 // [in] NULL
2437 // \ /
2438 // CmpP
2439 // Bool ne
2440 // If
2441 // / \
2442 // True False-<2>
2443 // / |
2444 // / cast_not_null
2445 // Load | | ^
2446 // [fast_test] | |
2447 // gvn to opt_test | |
2448 // / \ | <1>
2449 // True False |
2450 // | \\ |
2451 // [slow_call] \[fast_result]
2452 // Ctl Val \ \
2453 // | \ \
2454 // Catch <1> \ \
2455 // / \ ^ \ \
2456 // Ex No_Ex | \ \
2457 // | \ \ | \ <2> \
2458 // ... \ [slow_res] | | \ [null_result]
2459 // \ \--+--+--- | |
2460 // \ | / \ | /
2461 // --------Region Phi
2462 //
2463 //=============================================================================
2464 // Code is structured as a series of driver functions all called 'do_XXX' that
2465 // call a set of helper functions. Helper functions first, then drivers.
2466
2467 //------------------------------null_check_oop---------------------------------
2468 // Null check oop. Set null-path control into Region in slot 3.
2469 // Make a cast-not-nullness use the other not-null control. Return cast.
2470 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2471 bool never_see_null,
2472 bool safe_for_replace,
2473 bool speculative) {
2474 // Initial NULL check taken path
2475 (*null_control) = top();
2476 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2477
2478 // Generate uncommon_trap:
2479 if (never_see_null && (*null_control) != top()) {
2480 // If we see an unexpected null at a check-cast we record it and force a
2481 // recompile; the offending check-cast will be compiled to handle NULLs.
2482 // If we see more than one offending BCI, then all checkcasts in the
2483 // method will be compiled to handle NULLs.
2484 PreserveJVMState pjvms(this);
2485 set_control(*null_control);
2486 replace_in_map(value, null());
2487 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2488 uncommon_trap(reason,
2489 Deoptimization::Action_make_not_entrant);
2490 (*null_control) = top(); // NULL path is dead
2491 }
2492 if ((*null_control) == top() && safe_for_replace) {
2493 replace_in_map(value, cast);
2494 }
2495
2496 // Cast away null-ness on the result
2497 return cast;
2498 }
2499
2500 //------------------------------opt_iff----------------------------------------
2501 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2502 // Return slow-path control.
2503 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2504 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2505
2506 // Fast path taken; set region slot 2
2507 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2508 region->init_req(2,fast_taken); // Capture fast-control
2509
2510 // Fast path not-taken, i.e. slow path
2511 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2512 return slow_taken;
2513 }
2514
2515 //-----------------------------make_runtime_call-------------------------------
2516 Node* GraphKit::make_runtime_call(int flags,
2517 const TypeFunc* call_type, address call_addr,
2518 const char* call_name,
2519 const TypePtr* adr_type,
2520 // The following parms are all optional.
2521 // The first NULL ends the list.
2522 Node* parm0, Node* parm1,
2523 Node* parm2, Node* parm3,
2524 Node* parm4, Node* parm5,
2525 Node* parm6, Node* parm7) {
2526 // Slow-path call
2527 bool is_leaf = !(flags & RC_NO_LEAF);
2528 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2529 if (call_name == NULL) {
2530 assert(!is_leaf, "must supply name for leaf");
2531 call_name = OptoRuntime::stub_name(call_addr);
2532 }
2533 CallNode* call;
2534 if (!is_leaf) {
2535 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2536 bci(), adr_type);
2537 } else if (flags & RC_NO_FP) {
2538 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2539 } else {
2540 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2541 }
2542
2543 // The following is similar to set_edges_for_java_call,
2544 // except that the memory effects of the call are restricted to AliasIdxRaw.
2545
2546 // Slow path call has no side-effects, uses few values
2547 bool wide_in = !(flags & RC_NARROW_MEM);
2548 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2549
2550 Node* prev_mem = NULL;
2551 if (wide_in) {
2552 prev_mem = set_predefined_input_for_runtime_call(call);
2553 } else {
2554 assert(!wide_out, "narrow in => narrow out");
2555 Node* narrow_mem = memory(adr_type);
2556 prev_mem = reset_memory();
2557 map()->set_memory(narrow_mem);
2558 set_predefined_input_for_runtime_call(call);
2559 }
2560
2561 // Hook each parm in order. Stop looking at the first NULL.
2562 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2563 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2564 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2565 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2566 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2567 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2568 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2569 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2570 /* close each nested if ===> */ } } } } } } } }
2571 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2572
2573 if (!is_leaf) {
2574 // Non-leaves can block and take safepoints:
2575 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2576 }
2577 // Non-leaves can throw exceptions:
2578 if (has_io) {
2579 call->set_req(TypeFunc::I_O, i_o());
2580 }
2581
2582 if (flags & RC_UNCOMMON) {
2583 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2584 // (An "if" probability corresponds roughly to an unconditional count.
2585 // Sort of.)
2586 call->set_cnt(PROB_UNLIKELY_MAG(4));
2587 }
2588
2589 Node* c = _gvn.transform(call);
2590 assert(c == call, "cannot disappear");
2591
2592 if (wide_out) {
2593 // Slow path call has full side-effects.
2594 set_predefined_output_for_runtime_call(call);
2595 } else {
2596 // Slow path call has few side-effects, and/or sets few values.
2597 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2598 }
2599
2600 if (has_io) {
2601 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2602 }
2603 return call;
2604
2605 }
2606
2607 //------------------------------merge_memory-----------------------------------
2608 // Merge memory from one path into the current memory state.
2609 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2610 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2611 Node* old_slice = mms.force_memory();
2612 Node* new_slice = mms.memory2();
2613 if (old_slice != new_slice) {
2614 PhiNode* phi;
2615 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2616 if (mms.is_empty()) {
2617 // clone base memory Phi's inputs for this memory slice
2618 assert(old_slice == mms.base_memory(), "sanity");
2619 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2620 _gvn.set_type(phi, Type::MEMORY);
2621 for (uint i = 1; i < phi->req(); i++) {
2622 phi->init_req(i, old_slice->in(i));
2623 }
2624 } else {
2625 phi = old_slice->as_Phi(); // Phi was generated already
2626 }
2627 } else {
2628 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2629 _gvn.set_type(phi, Type::MEMORY);
2630 }
2631 phi->set_req(new_path, new_slice);
2632 mms.set_memory(phi);
2633 }
2634 }
2635 }
2636
2637 //------------------------------make_slow_call_ex------------------------------
2638 // Make the exception handler hookups for the slow call
2639 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2640 if (stopped()) return;
2641
2642 // Make a catch node with just two handlers: fall-through and catch-all
2643 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2644 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2645 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2646 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2647
2648 { PreserveJVMState pjvms(this);
2649 set_control(excp);
2650 set_i_o(i_o);
2651
2652 if (excp != top()) {
2653 if (deoptimize) {
2654 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2655 uncommon_trap(Deoptimization::Reason_unhandled,
2656 Deoptimization::Action_none);
2657 } else {
2658 // Create an exception state also.
2659 // Use an exact type if the caller has specified a specific exception.
2660 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2661 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2662 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2663 }
2664 }
2665 }
2666
2667 // Get the no-exception control from the CatchNode.
2668 set_control(norm);
2669 }
2670
2671 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2672 Node* cmp = NULL;
2673 switch(bt) {
2674 case T_INT: cmp = new CmpINode(in1, in2); break;
2675 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2676 default: fatal("unexpected comparison type %s", type2name(bt));
2677 }
2678 gvn->transform(cmp);
2679 Node* bol = gvn->transform(new BoolNode(cmp, test));
2680 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2681 gvn->transform(iff);
2682 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2683 return iff;
2684 }
2685
2686
2687 //-------------------------------gen_subtype_check-----------------------------
2688 // Generate a subtyping check. Takes as input the subtype and supertype.
2689 // Returns 2 values: sets the default control() to the true path and returns
2690 // the false path. Only reads invariant memory; sets no (visible) memory.
2691 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2692 // but that's not exposed to the optimizer. This call also doesn't take in an
2693 // Object; if you wish to check an Object you need to load the Object's class
2694 // prior to coming here.
2695 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2696 Compile* C = gvn->C;
2697
2698 if ((*ctrl)->is_top()) {
2699 return C->top();
2700 }
2701
2702 // Fast check for identical types, perhaps identical constants.
2703 // The types can even be identical non-constants, in cases
2704 // involving Array.newInstance, Object.clone, etc.
2705 if (subklass == superklass)
2706 return C->top(); // false path is dead; no test needed.
2707
2708 if (gvn->type(superklass)->singleton()) {
2709 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2710 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2711
2712 // In the common case of an exact superklass, try to fold up the
2713 // test before generating code. You may ask, why not just generate
2714 // the code and then let it fold up? The answer is that the generated
2715 // code will necessarily include null checks, which do not always
2716 // completely fold away. If they are also needless, then they turn
2717 // into a performance loss. Example:
2718 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2719 // Here, the type of 'fa' is often exact, so the store check
2720 // of fa[1]=x will fold up, without testing the nullness of x.
2721 switch (C->static_subtype_check(superk, subk)) {
2722 case Compile::SSC_always_false:
2723 {
2724 Node* always_fail = *ctrl;
2725 *ctrl = gvn->C->top();
2726 return always_fail;
2727 }
2728 case Compile::SSC_always_true:
2729 return C->top();
2730 case Compile::SSC_easy_test:
2731 {
2732 // Just do a direct pointer compare and be done.
2733 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2734 *ctrl = gvn->transform(new IfTrueNode(iff));
2735 return gvn->transform(new IfFalseNode(iff));
2736 }
2737 case Compile::SSC_full_test:
2738 break;
2739 default:
2740 ShouldNotReachHere();
2741 }
2742 }
2743
2744 // %%% Possible further optimization: Even if the superklass is not exact,
2745 // if the subklass is the unique subtype of the superklass, the check
2746 // will always succeed. We could leave a dependency behind to ensure this.
2747
2748 // First load the super-klass's check-offset
2749 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2750 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2751 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2752 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2753 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2754
2755 // Load from the sub-klass's super-class display list, or a 1-word cache of
2756 // the secondary superclass list, or a failing value with a sentinel offset
2757 // if the super-klass is an interface or exceptionally deep in the Java
2758 // hierarchy and we have to scan the secondary superclass list the hard way.
2759 // Worst-case type is a little odd: NULL is allowed as a result (usually
2760 // klass loads can never produce a NULL).
2761 Node *chk_off_X = chk_off;
2762 #ifdef _LP64
2763 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2764 #endif
2765 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2766 // For some types like interfaces the following loadKlass is from a 1-word
2767 // cache which is mutable so can't use immutable memory. Other
2768 // types load from the super-class display table which is immutable.
2769 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2770 Node *kmem = might_be_cache ? m : C->immutable_memory();
2771 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::BOTTOM));
2772
2773 // Compile speed common case: ARE a subtype and we canNOT fail
2774 if( superklass == nkls )
2775 return C->top(); // false path is dead; no test needed.
2776
2777 // See if we get an immediate positive hit. Happens roughly 83% of the
2778 // time. Test to see if the value loaded just previously from the subklass
2779 // is exactly the superklass.
2780 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2781 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2782 *ctrl = gvn->transform(new IfFalseNode(iff1));
2783
2784 // Compile speed common case: Check for being deterministic right now. If
2785 // chk_off is a constant and not equal to cacheoff then we are NOT a
2786 // subklass. In this case we need exactly the 1 test above and we can
2787 // return those results immediately.
2788 if (!might_be_cache) {
2789 Node* not_subtype_ctrl = *ctrl;
2790 *ctrl = iftrue1; // We need exactly the 1 test above
2791 return not_subtype_ctrl;
2792 }
2793
2794 // Gather the various success & failures here
2795 RegionNode *r_ok_subtype = new RegionNode(4);
2796 gvn->record_for_igvn(r_ok_subtype);
2797 RegionNode *r_not_subtype = new RegionNode(3);
2798 gvn->record_for_igvn(r_not_subtype);
2799
2800 r_ok_subtype->init_req(1, iftrue1);
2801
2802 // Check for immediate negative hit. Happens roughly 11% of the time (which
2803 // is roughly 63% of the remaining cases). Test to see if the loaded
2804 // check-offset points into the subklass display list or the 1-element
2805 // cache. If it points to the display (and NOT the cache) and the display
2806 // missed then it's not a subtype.
2807 Node *cacheoff = gvn->intcon(cacheoff_con);
2808 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2809 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2810 *ctrl = gvn->transform(new IfFalseNode(iff2));
2811
2812 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2813 // No performance impact (too rare) but allows sharing of secondary arrays
2814 // which has some footprint reduction.
2815 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2816 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2817 *ctrl = gvn->transform(new IfFalseNode(iff3));
2818
2819 // -- Roads not taken here: --
2820 // We could also have chosen to perform the self-check at the beginning
2821 // of this code sequence, as the assembler does. This would not pay off
2822 // the same way, since the optimizer, unlike the assembler, can perform
2823 // static type analysis to fold away many successful self-checks.
2824 // Non-foldable self checks work better here in second position, because
2825 // the initial primary superclass check subsumes a self-check for most
2826 // types. An exception would be a secondary type like array-of-interface,
2827 // which does not appear in its own primary supertype display.
2828 // Finally, we could have chosen to move the self-check into the
2829 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2830 // dependent manner. But it is worthwhile to have the check here,
2831 // where it can be perhaps be optimized. The cost in code space is
2832 // small (register compare, branch).
2833
2834 // Now do a linear scan of the secondary super-klass array. Again, no real
2835 // performance impact (too rare) but it's gotta be done.
2836 // Since the code is rarely used, there is no penalty for moving it
2837 // out of line, and it can only improve I-cache density.
2838 // The decision to inline or out-of-line this final check is platform
2839 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2840 Node* psc = gvn->transform(
2841 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2842
2843 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2844 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2845 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2846
2847 // Return false path; set default control to true path.
2848 *ctrl = gvn->transform(r_ok_subtype);
2849 return gvn->transform(r_not_subtype);
2850 }
2851
2852 // Profile-driven exact type check:
2853 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2854 float prob,
2855 Node* *casted_receiver) {
2856 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2857 Node* recv_klass = load_object_klass(receiver);
2858 Node* fail = type_check(recv_klass, tklass, prob);
2859 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2860 assert(recv_xtype->klass_is_exact(), "");
2861
2862 // Subsume downstream occurrences of receiver with a cast to
2863 // recv_xtype, since now we know what the type will be.
2864 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2865 (*casted_receiver) = _gvn.transform(cast);
2866 // (User must make the replace_in_map call.)
2867
2868 return fail;
2869 }
2870
2871 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
2872 float prob) {
2873 //const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2874 Node* want_klass = makecon(tklass);
2875 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass));
2876 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2877 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2878 set_control( _gvn.transform( new IfTrueNode (iff)));
2879 Node* fail = _gvn.transform( new IfFalseNode(iff));
2880 return fail;
2881 }
2882
2883
2884 //------------------------------seems_never_null-------------------------------
2885 // Use null_seen information if it is available from the profile.
2886 // If we see an unexpected null at a type check we record it and force a
2887 // recompile; the offending check will be recompiled to handle NULLs.
2888 // If we see several offending BCIs, then all checks in the
2889 // method will be recompiled.
2890 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2891 speculating = !_gvn.type(obj)->speculative_maybe_null();
2892 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2893 if (UncommonNullCast // Cutout for this technique
2894 && obj != null() // And not the -Xcomp stupid case?
2895 && !too_many_traps(reason)
2896 ) {
2897 if (speculating) {
2898 return true;
2899 }
2900 if (data == NULL)
2901 // Edge case: no mature data. Be optimistic here.
2902 return true;
2903 // If the profile has not seen a null, assume it won't happen.
2904 assert(java_bc() == Bytecodes::_checkcast ||
2905 java_bc() == Bytecodes::_instanceof ||
2906 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2907 return !data->as_BitData()->null_seen();
2908 }
2909 speculating = false;
2910 return false;
2911 }
2912
2913 //------------------------maybe_cast_profiled_receiver-------------------------
2914 // If the profile has seen exactly one type, narrow to exactly that type.
2915 // Subsequent type checks will always fold up.
2916 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2917 ciKlass* require_klass,
2918 ciKlass* spec_klass,
2919 bool safe_for_replace) {
2920 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2921
2922 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2923
2924 // Make sure we haven't already deoptimized from this tactic.
2925 if (too_many_traps(reason) || too_many_recompiles(reason))
2926 return NULL;
2927
2928 // (No, this isn't a call, but it's enough like a virtual call
2929 // to use the same ciMethod accessor to get the profile info...)
2930 // If we have a speculative type use it instead of profiling (which
2931 // may not help us)
2932 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2933 if (exact_kls != NULL) {// no cast failures here
2934 if (require_klass == NULL ||
2935 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2936 // If we narrow the type to match what the type profile sees or
2937 // the speculative type, we can then remove the rest of the
2938 // cast.
2939 // This is a win, even if the exact_kls is very specific,
2940 // because downstream operations, such as method calls,
2941 // will often benefit from the sharper type.
2942 Node* exact_obj = not_null_obj; // will get updated in place...
2943 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2944 &exact_obj);
2945 { PreserveJVMState pjvms(this);
2946 set_control(slow_ctl);
2947 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2948 }
2949 if (safe_for_replace) {
2950 replace_in_map(not_null_obj, exact_obj);
2951 }
2952 return exact_obj;
2953 }
2954 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2955 }
2956
2957 return NULL;
2958 }
2959
2960 /**
2961 * Cast obj to type and emit guard unless we had too many traps here
2962 * already
2963 *
2964 * @param obj node being casted
2965 * @param type type to cast the node to
2966 * @param not_null true if we know node cannot be null
2967 */
2968 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2969 ciKlass* type,
2970 bool not_null) {
2971 if (stopped()) {
2972 return obj;
2973 }
2974
2975 // type == NULL if profiling tells us this object is always null
2976 if (type != NULL) {
2977 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2978 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2979
2980 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2981 !too_many_traps(class_reason) &&
2982 !too_many_recompiles(class_reason)) {
2983 Node* not_null_obj = NULL;
2984 // not_null is true if we know the object is not null and
2985 // there's no need for a null check
2986 if (!not_null) {
2987 Node* null_ctl = top();
2988 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2989 assert(null_ctl->is_top(), "no null control here");
2990 } else {
2991 not_null_obj = obj;
2992 }
2993
2994 Node* exact_obj = not_null_obj;
2995 ciKlass* exact_kls = type;
2996 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2997 &exact_obj);
2998 {
2999 PreserveJVMState pjvms(this);
3000 set_control(slow_ctl);
3001 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3002 }
3003 replace_in_map(not_null_obj, exact_obj);
3004 obj = exact_obj;
3005 }
3006 } else {
3007 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
3008 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
3009 Node* exact_obj = null_assert(obj);
3010 replace_in_map(obj, exact_obj);
3011 obj = exact_obj;
3012 }
3013 }
3014 return obj;
3015 }
3016
3017 //-------------------------------gen_instanceof--------------------------------
3018 // Generate an instance-of idiom. Used by both the instance-of bytecode
3019 // and the reflective instance-of call.
3020 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3021 kill_dead_locals(); // Benefit all the uncommon traps
3022 assert( !stopped(), "dead parse path should be checked in callers" );
3023 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3024 "must check for not-null not-dead klass in callers");
3025
3026 // Make the merge point
3027 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3028 RegionNode* region = new RegionNode(PATH_LIMIT);
3029 Node* phi = new PhiNode(region, TypeInt::BOOL);
3030 C->set_has_split_ifs(true); // Has chance for split-if optimization
3031
3032 ciProfileData* data = NULL;
3033 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3034 data = method()->method_data()->bci_to_data(bci());
3035 }
3036 bool speculative_not_null = false;
3037 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3038 && seems_never_null(obj, data, speculative_not_null));
3039
3040 bool is_value = obj->is_ValueType();
3041
3042 // Null check; get casted pointer; set region slot 3
3043 Node* null_ctl = top();
3044 Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3045
3046 // If not_null_obj is dead, only null-path is taken
3047 if (stopped()) { // Doing instance-of on a NULL?
3048 set_control(null_ctl);
3049 return intcon(0);
3050 }
3051 region->init_req(_null_path, null_ctl);
3052 phi ->init_req(_null_path, intcon(0)); // Set null path value
3053 if (null_ctl == top()) {
3054 // Do this eagerly, so that pattern matches like is_diamond_phi
3055 // will work even during parsing.
3056 assert(_null_path == PATH_LIMIT-1, "delete last");
3057 region->del_req(_null_path);
3058 phi ->del_req(_null_path);
3059 }
3060
3061 // Do we know the type check always succeed?
3062 if (!is_value) {
3063 bool known_statically = false;
3064 if (_gvn.type(superklass)->singleton()) {
3065 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3066 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3067 if (subk != NULL && subk->is_loaded()) {
3068 int static_res = C->static_subtype_check(superk, subk);
3069 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3070 }
3071 }
3072
3073 if (!known_statically) {
3074 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3075 // We may not have profiling here or it may not help us. If we
3076 // have a speculative type use it to perform an exact cast.
3077 ciKlass* spec_obj_type = obj_type->speculative_type();
3078 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3079 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3080 if (stopped()) { // Profile disagrees with this path.
3081 set_control(null_ctl); // Null is the only remaining possibility.
3082 return intcon(0);
3083 }
3084 if (cast_obj != NULL) {
3085 not_null_obj = cast_obj;
3086 }
3087 }
3088 }
3089 }
3090
3091 // Load the object's klass
3092 Node* obj_klass = NULL;
3093 if (is_value) {
3094 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(obj)->is_valuetype()->value_klass()));
3095 } else {
3096 obj_klass = load_object_klass(not_null_obj);
3097 }
3098
3099 // Generate the subtype check
3100 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3101
3102 // Plug in the success path to the general merge in slot 1.
3103 region->init_req(_obj_path, control());
3104 phi ->init_req(_obj_path, intcon(1));
3105
3106 // Plug in the failing path to the general merge in slot 2.
3107 region->init_req(_fail_path, not_subtype_ctrl);
3108 phi ->init_req(_fail_path, intcon(0));
3109
3110 // Return final merged results
3111 set_control( _gvn.transform(region) );
3112 record_for_igvn(region);
3113
3114 // If we know the type check always succeeds then we don't use the
3115 // profiling data at this bytecode. Don't lose it, feed it to the
3116 // type system as a speculative type.
3117 if (safe_for_replace && !is_value) {
3118 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3119 replace_in_map(obj, casted_obj);
3120 }
3121
3122 return _gvn.transform(phi);
3123 }
3124
3125 //-------------------------------gen_checkcast---------------------------------
3126 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3127 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3128 // uncommon-trap paths work. Adjust stack after this call.
3129 // If failure_control is supplied and not null, it is filled in with
3130 // the control edge for the cast failure. Otherwise, an appropriate
3131 // uncommon trap or exception is thrown.
3132 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3133 Node* *failure_control) {
3134 kill_dead_locals(); // Benefit all the uncommon traps
3135 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr();
3136 const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass());
3137 if (toop->is_valuetypeptr()) {
3138 if (obj->is_ValueType()) {
3139 const TypeValueType* tvt = _gvn.type(obj)->is_valuetype();
3140 if (toop->klass() == tvt->value_klass()) {
3141 return obj;
3142 } else {
3143 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(tvt->value_klass())));
3144 return top();
3145 }
3146 }
3147 Node* null_ctl = top();
3148 Node* not_null_obj = null_check_common(obj, T_VALUETYPE, false, &null_ctl, false);
3149 if (null_ctl != top()) {
3150 // TODO For now, we just deoptimize if value type is NULL
3151 PreserveJVMState pjvms(this);
3152 set_control(null_ctl);
3153 replace_in_map(obj, null());
3154 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none);
3155 null_ctl = top(); // NULL path is dead
3156 }
3157 replace_in_map(obj, not_null_obj);
3158 obj = not_null_obj;
3159 } else if (obj->is_ValueType()) {
3160 ValueTypeBaseNode* vt = obj->as_ValueType()->allocate(this, true);
3161 obj = ValueTypePtrNode::make_from_value_type(_gvn, vt->as_ValueType());
3162 }
3163
3164 // Fast cutout: Check the case that the cast is vacuously true.
3165 // This detects the common cases where the test will short-circuit
3166 // away completely. We do this before we perform the null check,
3167 // because if the test is going to turn into zero code, we don't
3168 // want a residual null check left around. (Causes a slowdown,
3169 // for example, in some objArray manipulations, such as a[i]=a[j].)
3170 if (tk->singleton()) {
3171 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3172 if (objtp != NULL && objtp->klass() != NULL) {
3173 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3174 case Compile::SSC_always_true:
3175 // If we know the type check always succeed then we don't use
3176 // the profiling data at this bytecode. Don't lose it, feed it
3177 // to the type system as a speculative type.
3178 obj = record_profiled_receiver_for_speculation(obj);
3179 if (toop->is_valuetypeptr()) {
3180 obj = ValueTypeNode::make_from_oop(this, obj, toop->value_klass());
3181 }
3182 return obj;
3183 case Compile::SSC_always_false:
3184 // It needs a null check because a null will *pass* the cast check.
3185 // A non-null value will always produce an exception.
3186 return null_assert(obj);
3187 }
3188 }
3189 }
3190
3191 ciProfileData* data = NULL;
3192 bool safe_for_replace = false;
3193 if (failure_control == NULL) { // use MDO in regular case only
3194 assert(java_bc() == Bytecodes::_aastore ||
3195 java_bc() == Bytecodes::_checkcast,
3196 "interpreter profiles type checks only for these BCs");
3197 data = method()->method_data()->bci_to_data(bci());
3198 safe_for_replace = true;
3199 }
3200
3201 // Make the merge point
3202 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3203 RegionNode* region = new RegionNode(PATH_LIMIT);
3204 Node* phi = new PhiNode(region, toop);
3205 C->set_has_split_ifs(true); // Has chance for split-if optimization
3206
3207 // Use null-cast information if it is available
3208 bool speculative_not_null = false;
3209 bool never_see_null = ((failure_control == NULL) // regular case only
3210 && seems_never_null(obj, data, speculative_not_null));
3211
3212 // Null check; get casted pointer; set region slot 3
3213 Node* null_ctl = top();
3214 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3215
3216 // If not_null_obj is dead, only null-path is taken
3217 if (stopped()) { // Doing instance-of on a NULL?
3218 set_control(null_ctl);
3219 return null();
3220 }
3221 region->init_req(_null_path, null_ctl);
3222 phi ->init_req(_null_path, null()); // Set null path value
3223 if (null_ctl == top()) {
3224 // Do this eagerly, so that pattern matches like is_diamond_phi
3225 // will work even during parsing.
3226 assert(_null_path == PATH_LIMIT-1, "delete last");
3227 region->del_req(_null_path);
3228 phi ->del_req(_null_path);
3229 }
3230
3231 Node* cast_obj = NULL;
3232 if (tk->klass_is_exact()) {
3233 // The following optimization tries to statically cast the speculative type of the object
3234 // (for example obtained during profiling) to the type of the superklass and then do a
3235 // dynamic check that the type of the object is what we expect. To work correctly
3236 // for checkcast and aastore the type of superklass should be exact.
3237 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3238 // We may not have profiling here or it may not help us. If we have
3239 // a speculative type use it to perform an exact cast.
3240 ciKlass* spec_obj_type = obj_type->speculative_type();
3241 if (spec_obj_type != NULL || data != NULL) {
3242 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3243 if (cast_obj != NULL) {
3244 if (failure_control != NULL) // failure is now impossible
3245 (*failure_control) = top();
3246 // adjust the type of the phi to the exact klass:
3247 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3248 }
3249 }
3250 }
3251
3252 if (cast_obj == NULL) {
3253 // Load the object's klass
3254 Node* obj_klass = load_object_klass(not_null_obj);
3255
3256 // Generate the subtype check
3257 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3258
3259 // Plug in success path into the merge
3260 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3261 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3262 if (failure_control == NULL) {
3263 if (not_subtype_ctrl != top()) { // If failure is possible
3264 PreserveJVMState pjvms(this);
3265 set_control(not_subtype_ctrl);
3266 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3267 }
3268 } else {
3269 (*failure_control) = not_subtype_ctrl;
3270 }
3271 }
3272
3273 region->init_req(_obj_path, control());
3274 phi ->init_req(_obj_path, cast_obj);
3275
3276 // A merge of NULL or Casted-NotNull obj
3277 Node* res = _gvn.transform(phi);
3278
3279 // Note I do NOT always 'replace_in_map(obj,result)' here.
3280 // if( tk->klass()->can_be_primary_super() )
3281 // This means that if I successfully store an Object into an array-of-String
3282 // I 'forget' that the Object is really now known to be a String. I have to
3283 // do this because we don't have true union types for interfaces - if I store
3284 // a Baz into an array-of-Interface and then tell the optimizer it's an
3285 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3286 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3287 // replace_in_map( obj, res );
3288
3289 // Return final merged results
3290 set_control( _gvn.transform(region) );
3291 record_for_igvn(region);
3292
3293 res = record_profiled_receiver_for_speculation(res);
3294 if (toop->is_valuetypeptr()) {
3295 res = ValueTypeNode::make_from_oop(this, res, toop->value_klass());
3296 }
3297 return res;
3298 }
3299
3300 // Deoptimize if 'ary' is flattened or if 'obj' is null and 'ary' is a value type array
3301 void GraphKit::gen_value_type_array_guard(Node* ary, Node* obj, Node* elem_klass) {
3302 assert(EnableValhalla, "should only be used if value types are enabled");
3303 if (elem_klass == NULL) {
3304 // Load array element klass
3305 Node* kls = load_object_klass(ary);
3306 Node* k_adr = basic_plus_adr(kls, kls, in_bytes(ArrayKlass::element_klass_offset()));
3307 elem_klass = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
3308 }
3309 // Check if element is a value type
3310 Node* flags_addr = basic_plus_adr(elem_klass, in_bytes(Klass::access_flags_offset()));
3311 Node* flags = make_load(NULL, flags_addr, TypeInt::INT, T_INT, MemNode::unordered);
3312 Node* is_value_elem = _gvn.transform(new AndINode(flags, intcon(JVM_ACC_VALUE)));
3313
3314 const Type* objtype = _gvn.type(obj);
3315 if (objtype == TypePtr::NULL_PTR) {
3316 // Object is always null, check if array is a value type array
3317 Node* cmp = _gvn.transform(new CmpINode(is_value_elem, intcon(0)));
3318 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3319 { BuildCutout unless(this, bol, PROB_MAX);
3320 // TODO just deoptimize for now if we store null to a value type array
3321 uncommon_trap(Deoptimization::Reason_array_check,
3322 Deoptimization::Action_none);
3323 }
3324 } else {
3325 // Check if array is flattened or if we are storing null to a value type array
3326 // TODO can we merge these checks?
3327 gen_flattened_array_guard(ary);
3328 if (objtype->meet(TypePtr::NULL_PTR) == objtype) {
3329 // Check if (is_value_elem && obj_is_null) <=> (!is_value_elem | !obj_is_null == 0)
3330 // TODO what if we later figure out that obj is never null?
3331 Node* not_value = _gvn.transform(new XorINode(is_value_elem, intcon(JVM_ACC_VALUE)));
3332 not_value = _gvn.transform(new ConvI2LNode(not_value));
3333 Node* not_null = _gvn.transform(new CastP2XNode(NULL, obj));
3334 Node* both = _gvn.transform(new OrLNode(not_null, not_value));
3335 Node* cmp = _gvn.transform(new CmpLNode(both, longcon(0)));
3336 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3337 { BuildCutout unless(this, bol, PROB_MAX);
3338 // TODO just deoptimize for now if we store null to a value type array
3339 uncommon_trap(Deoptimization::Reason_array_check,
3340 Deoptimization::Action_none);
3341 }
3342 }
3343 }
3344 }
3345
3346 // Deoptimize if 'ary' is a flattened value type array
3347 void GraphKit::gen_flattened_array_guard(Node* ary, int nargs) {
3348 assert(EnableValhalla, "should only be used if value types are enabled");
3349 if (ValueArrayFlatten) {
3350 // Cannot statically determine if array is flattened, emit runtime check
3351 Node* kls = load_object_klass(ary);
3352 Node* lhp = basic_plus_adr(kls, kls, in_bytes(Klass::layout_helper_offset()));
3353 Node* layout_val = make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3354 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3355 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(Klass::_lh_array_tag_vt_value)));
3356 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3357
3358 { BuildCutout unless(this, bol, PROB_MAX);
3359 // TODO just deoptimize for now if value type array is flattened
3360 inc_sp(nargs);
3361 uncommon_trap(Deoptimization::Reason_array_check,
3362 Deoptimization::Action_none);
3363 }
3364 }
3365 }
3366
3367 //------------------------------next_monitor-----------------------------------
3368 // What number should be given to the next monitor?
3369 int GraphKit::next_monitor() {
3370 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3371 int next = current + C->sync_stack_slots();
3372 // Keep the toplevel high water mark current:
3373 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3374 return current;
3375 }
3376
3377 //------------------------------insert_mem_bar---------------------------------
3378 // Memory barrier to avoid floating things around
3379 // The membar serves as a pinch point between both control and all memory slices.
3380 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3381 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3382 mb->init_req(TypeFunc::Control, control());
3383 mb->init_req(TypeFunc::Memory, reset_memory());
3384 Node* membar = _gvn.transform(mb);
3385 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3386 set_all_memory_call(membar);
3387 return membar;
3388 }
3389
3390 //-------------------------insert_mem_bar_volatile----------------------------
3391 // Memory barrier to avoid floating things around
3392 // The membar serves as a pinch point between both control and memory(alias_idx).
3393 // If you want to make a pinch point on all memory slices, do not use this
3394 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3395 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3396 // When Parse::do_put_xxx updates a volatile field, it appends a series
3397 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3398 // The first membar is on the same memory slice as the field store opcode.
3399 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3400 // All the other membars (for other volatile slices, including AliasIdxBot,
3401 // which stands for all unknown volatile slices) are control-dependent
3402 // on the first membar. This prevents later volatile loads or stores
3403 // from sliding up past the just-emitted store.
3404
3405 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3406 mb->set_req(TypeFunc::Control,control());
3407 if (alias_idx == Compile::AliasIdxBot) {
3408 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3409 } else {
3410 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3411 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3412 }
3413 Node* membar = _gvn.transform(mb);
3414 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3415 if (alias_idx == Compile::AliasIdxBot) {
3416 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3417 } else {
3418 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3419 }
3420 return membar;
3421 }
3422
3423 //------------------------------shared_lock------------------------------------
3424 // Emit locking code.
3425 FastLockNode* GraphKit::shared_lock(Node* obj) {
3426 // bci is either a monitorenter bc or InvocationEntryBci
3427 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3428 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3429
3430 if( !GenerateSynchronizationCode )
3431 return NULL; // Not locking things?
3432 if (stopped()) // Dead monitor?
3433 return NULL;
3434
3435 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3436
3437 // Box the stack location
3438 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3439 Node* mem = reset_memory();
3440
3441 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3442 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3443 // Create the counters for this fast lock.
3444 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3445 }
3446
3447 // Create the rtm counters for this fast lock if needed.
3448 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3449
3450 // Add monitor to debug info for the slow path. If we block inside the
3451 // slow path and de-opt, we need the monitor hanging around
3452 map()->push_monitor( flock );
3453
3454 const TypeFunc *tf = LockNode::lock_type();
3455 LockNode *lock = new LockNode(C, tf);
3456
3457 lock->init_req( TypeFunc::Control, control() );
3458 lock->init_req( TypeFunc::Memory , mem );
3459 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3460 lock->init_req( TypeFunc::FramePtr, frameptr() );
3461 lock->init_req( TypeFunc::ReturnAdr, top() );
3462
3463 lock->init_req(TypeFunc::Parms + 0, obj);
3464 lock->init_req(TypeFunc::Parms + 1, box);
3465 lock->init_req(TypeFunc::Parms + 2, flock);
3466 add_safepoint_edges(lock);
3467
3468 lock = _gvn.transform( lock )->as_Lock();
3469
3470 // lock has no side-effects, sets few values
3471 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3472
3473 insert_mem_bar(Op_MemBarAcquireLock);
3474
3475 // Add this to the worklist so that the lock can be eliminated
3476 record_for_igvn(lock);
3477
3478 #ifndef PRODUCT
3479 if (PrintLockStatistics) {
3480 // Update the counter for this lock. Don't bother using an atomic
3481 // operation since we don't require absolute accuracy.
3482 lock->create_lock_counter(map()->jvms());
3483 increment_counter(lock->counter()->addr());
3484 }
3485 #endif
3486
3487 return flock;
3488 }
3489
3490
3491 //------------------------------shared_unlock----------------------------------
3492 // Emit unlocking code.
3493 void GraphKit::shared_unlock(Node* box, Node* obj) {
3494 // bci is either a monitorenter bc or InvocationEntryBci
3495 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3496 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3497
3498 if( !GenerateSynchronizationCode )
3499 return;
3500 if (stopped()) { // Dead monitor?
3501 map()->pop_monitor(); // Kill monitor from debug info
3502 return;
3503 }
3504
3505 // Memory barrier to avoid floating things down past the locked region
3506 insert_mem_bar(Op_MemBarReleaseLock);
3507
3508 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3509 UnlockNode *unlock = new UnlockNode(C, tf);
3510 #ifdef ASSERT
3511 unlock->set_dbg_jvms(sync_jvms());
3512 #endif
3513 uint raw_idx = Compile::AliasIdxRaw;
3514 unlock->init_req( TypeFunc::Control, control() );
3515 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3516 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3517 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3518 unlock->init_req( TypeFunc::ReturnAdr, top() );
3519
3520 unlock->init_req(TypeFunc::Parms + 0, obj);
3521 unlock->init_req(TypeFunc::Parms + 1, box);
3522 unlock = _gvn.transform(unlock)->as_Unlock();
3523
3524 Node* mem = reset_memory();
3525
3526 // unlock has no side-effects, sets few values
3527 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3528
3529 // Kill monitor from debug info
3530 map()->pop_monitor( );
3531 }
3532
3533 //-------------------------------get_layout_helper-----------------------------
3534 // If the given klass is a constant or known to be an array,
3535 // fetch the constant layout helper value into constant_value
3536 // and return (Node*)NULL. Otherwise, load the non-constant
3537 // layout helper value, and return the node which represents it.
3538 // This two-faced routine is useful because allocation sites
3539 // almost always feature constant types.
3540 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3541 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3542 if (!StressReflectiveCode && inst_klass != NULL) {
3543 ciKlass* klass = inst_klass->klass();
3544 assert(klass != NULL, "klass should not be NULL");
3545 bool xklass = inst_klass->klass_is_exact();
3546 if (xklass || klass->is_array_klass()) {
3547 jint lhelper = klass->layout_helper();
3548 if (lhelper != Klass::_lh_neutral_value) {
3549 constant_value = lhelper;
3550 return (Node*) NULL;
3551 }
3552 }
3553 }
3554 constant_value = Klass::_lh_neutral_value; // put in a known value
3555 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3556 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3557 }
3558
3559 // We just put in an allocate/initialize with a big raw-memory effect.
3560 // Hook selected additional alias categories on the initialization.
3561 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3562 MergeMemNode* init_in_merge,
3563 Node* init_out_raw) {
3564 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3565 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3566
3567 Node* prevmem = kit.memory(alias_idx);
3568 init_in_merge->set_memory_at(alias_idx, prevmem);
3569 kit.set_memory(init_out_raw, alias_idx);
3570 }
3571
3572 //---------------------------set_output_for_allocation-------------------------
3573 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3574 const TypeOopPtr* oop_type,
3575 bool deoptimize_on_exception) {
3576 int rawidx = Compile::AliasIdxRaw;
3577 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3578 add_safepoint_edges(alloc);
3579 Node* allocx = _gvn.transform(alloc);
3580 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3581 // create memory projection for i_o
3582 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3583 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3584
3585 // create a memory projection as for the normal control path
3586 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3587 set_memory(malloc, rawidx);
3588
3589 // a normal slow-call doesn't change i_o, but an allocation does
3590 // we create a separate i_o projection for the normal control path
3591 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3592 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3593
3594 // put in an initialization barrier
3595 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3596 rawoop)->as_Initialize();
3597 assert(alloc->initialization() == init, "2-way macro link must work");
3598 assert(init ->allocation() == alloc, "2-way macro link must work");
3599 {
3600 // Extract memory strands which may participate in the new object's
3601 // initialization, and source them from the new InitializeNode.
3602 // This will allow us to observe initializations when they occur,
3603 // and link them properly (as a group) to the InitializeNode.
3604 assert(init->in(InitializeNode::Memory) == malloc, "");
3605 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3606 init->set_req(InitializeNode::Memory, minit_in);
3607 record_for_igvn(minit_in); // fold it up later, if possible
3608 Node* minit_out = memory(rawidx);
3609 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3610 if (oop_type->isa_aryptr()) {
3611 const TypeAryPtr* arytype = oop_type->is_aryptr();
3612 if (arytype->klass()->is_value_array_klass()) {
3613 ciValueArrayKlass* vak = arytype->klass()->as_value_array_klass();
3614 ciValueKlass* vk = vak->element_klass()->as_value_klass();
3615 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
3616 ciField* field = vk->nonstatic_field_at(i);
3617 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3618 continue; // do not bother to track really large numbers of fields
3619 int off_in_vt = field->offset() - vk->first_field_offset();
3620 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
3621 int fieldidx = C->get_alias_index(adr_type);
3622 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3623 }
3624 } else {
3625 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3626 int elemidx = C->get_alias_index(telemref);
3627 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3628 }
3629 } else if (oop_type->isa_instptr()) {
3630 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3631 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3632 ciField* field = ik->nonstatic_field_at(i);
3633 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3634 continue; // do not bother to track really large numbers of fields
3635 // Find (or create) the alias category for this field:
3636 int fieldidx = C->alias_type(field)->index();
3637 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3638 }
3639 }
3640 }
3641
3642 // Cast raw oop to the real thing...
3643 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3644 javaoop = _gvn.transform(javaoop);
3645 C->set_recent_alloc(control(), javaoop);
3646 assert(just_allocated_object(control()) == javaoop, "just allocated");
3647
3648 #ifdef ASSERT
3649 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3650 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3651 "Ideal_allocation works");
3652 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3653 "Ideal_allocation works");
3654 if (alloc->is_AllocateArray()) {
3655 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3656 "Ideal_allocation works");
3657 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3658 "Ideal_allocation works");
3659 } else {
3660 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3661 }
3662 }
3663 #endif //ASSERT
3664
3665 return javaoop;
3666 }
3667
3668 //---------------------------new_instance--------------------------------------
3669 // This routine takes a klass_node which may be constant (for a static type)
3670 // or may be non-constant (for reflective code). It will work equally well
3671 // for either, and the graph will fold nicely if the optimizer later reduces
3672 // the type to a constant.
3673 // The optional arguments are for specialized use by intrinsics:
3674 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3675 // - If 'return_size_val', report the the total object size to the caller.
3676 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3677 Node* GraphKit::new_instance(Node* klass_node,
3678 Node* extra_slow_test,
3679 Node* *return_size_val,
3680 bool deoptimize_on_exception,
3681 ValueTypeBaseNode* value_node) {
3682 // Compute size in doublewords
3683 // The size is always an integral number of doublewords, represented
3684 // as a positive bytewise size stored in the klass's layout_helper.
3685 // The layout_helper also encodes (in a low bit) the need for a slow path.
3686 jint layout_con = Klass::_lh_neutral_value;
3687 Node* layout_val = get_layout_helper(klass_node, layout_con);
3688 bool layout_is_con = (layout_val == NULL);
3689
3690 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3691 // Generate the initial go-slow test. It's either ALWAYS (return a
3692 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3693 // case) a computed value derived from the layout_helper.
3694 Node* initial_slow_test = NULL;
3695 if (layout_is_con) {
3696 assert(!StressReflectiveCode, "stress mode does not use these paths");
3697 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3698 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3699 } else { // reflective case
3700 // This reflective path is used by Unsafe.allocateInstance.
3701 // (It may be stress-tested by specifying StressReflectiveCode.)
3702 // Basically, we want to get into the VM is there's an illegal argument.
3703 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3704 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3705 if (extra_slow_test != intcon(0)) {
3706 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3707 }
3708 // (Macro-expander will further convert this to a Bool, if necessary.)
3709 }
3710
3711 // Find the size in bytes. This is easy; it's the layout_helper.
3712 // The size value must be valid even if the slow path is taken.
3713 Node* size = NULL;
3714 if (layout_is_con) {
3715 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3716 } else { // reflective case
3717 // This reflective path is used by clone and Unsafe.allocateInstance.
3718 size = ConvI2X(layout_val);
3719
3720 // Clear the low bits to extract layout_helper_size_in_bytes:
3721 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3722 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3723 size = _gvn.transform( new AndXNode(size, mask) );
3724 }
3725 if (return_size_val != NULL) {
3726 (*return_size_val) = size;
3727 }
3728
3729 // This is a precise notnull oop of the klass.
3730 // (Actually, it need not be precise if this is a reflective allocation.)
3731 // It's what we cast the result to.
3732 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3733 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3734 const TypeOopPtr* oop_type = tklass->as_instance_type();
3735
3736 // Now generate allocation code
3737
3738 // The entire memory state is needed for slow path of the allocation
3739 // since GC and deoptimization can happen.
3740 Node *mem = reset_memory();
3741 set_all_memory(mem); // Create new memory state
3742
3743 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3744 control(), mem, i_o(),
3745 size, klass_node,
3746 initial_slow_test, value_node);
3747
3748 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3749 }
3750
3751 //-------------------------------new_array-------------------------------------
3752 // helper for newarray and anewarray
3753 // The 'length' parameter is (obviously) the length of the array.
3754 // See comments on new_instance for the meaning of the other arguments.
3755 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3756 Node* length, // number of array elements
3757 int nargs, // number of arguments to push back for uncommon trap
3758 Node* *return_size_val,
3759 bool deoptimize_on_exception) {
3760 jint layout_con = Klass::_lh_neutral_value;
3761 Node* layout_val = get_layout_helper(klass_node, layout_con);
3762 bool layout_is_con = (layout_val == NULL);
3763
3764 if (!layout_is_con && !StressReflectiveCode &&
3765 !too_many_traps(Deoptimization::Reason_class_check)) {
3766 // This is a reflective array creation site.
3767 // Optimistically assume that it is a subtype of Object[],
3768 // so that we can fold up all the address arithmetic.
3769 layout_con = Klass::array_layout_helper(T_OBJECT);
3770 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3771 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3772 { BuildCutout unless(this, bol_lh, PROB_MAX);
3773 inc_sp(nargs);
3774 uncommon_trap(Deoptimization::Reason_class_check,
3775 Deoptimization::Action_maybe_recompile);
3776 }
3777 layout_val = NULL;
3778 layout_is_con = true;
3779 }
3780
3781 // Generate the initial go-slow test. Make sure we do not overflow
3782 // if length is huge (near 2Gig) or negative! We do not need
3783 // exact double-words here, just a close approximation of needed
3784 // double-words. We can't add any offset or rounding bits, lest we
3785 // take a size -1 of bytes and make it positive. Use an unsigned
3786 // compare, so negative sizes look hugely positive.
3787 int fast_size_limit = FastAllocateSizeLimit;
3788 if (layout_is_con) {
3789 assert(!StressReflectiveCode, "stress mode does not use these paths");
3790 // Increase the size limit if we have exact knowledge of array type.
3791 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3792 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3793 }
3794
3795 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3796 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3797
3798 // --- Size Computation ---
3799 // array_size = round_to_heap(array_header + (length << elem_shift));
3800 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3801 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3802 // The rounding mask is strength-reduced, if possible.
3803 int round_mask = MinObjAlignmentInBytes - 1;
3804 Node* header_size = NULL;
3805 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3806 // (T_BYTE has the weakest alignment and size restrictions...)
3807 if (layout_is_con) {
3808 int hsize = Klass::layout_helper_header_size(layout_con);
3809 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3810 BasicType etype = Klass::layout_helper_element_type(layout_con);
3811 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con);
3812 if ((round_mask & ~right_n_bits(eshift)) == 0)
3813 round_mask = 0; // strength-reduce it if it goes away completely
3814 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3815 assert(header_size_min <= hsize, "generic minimum is smallest");
3816 header_size_min = hsize;
3817 header_size = intcon(hsize + round_mask);
3818 } else {
3819 Node* hss = intcon(Klass::_lh_header_size_shift);
3820 Node* hsm = intcon(Klass::_lh_header_size_mask);
3821 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3822 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3823 Node* mask = intcon(round_mask);
3824 header_size = _gvn.transform( new AddINode(hsize, mask) );
3825 }
3826
3827 Node* elem_shift = NULL;
3828 if (layout_is_con) {
3829 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3830 if (eshift != 0)
3831 elem_shift = intcon(eshift);
3832 } else {
3833 // There is no need to mask or shift this value.
3834 // The semantics of LShiftINode include an implicit mask to 0x1F.
3835 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3836 elem_shift = layout_val;
3837 }
3838
3839 // Transition to native address size for all offset calculations:
3840 Node* lengthx = ConvI2X(length);
3841 Node* headerx = ConvI2X(header_size);
3842 #ifdef _LP64
3843 { const TypeInt* tilen = _gvn.find_int_type(length);
3844 if (tilen != NULL && tilen->_lo < 0) {
3845 // Add a manual constraint to a positive range. Cf. array_element_address.
3846 jint size_max = fast_size_limit;
3847 if (size_max > tilen->_hi) size_max = tilen->_hi;
3848 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3849
3850 // Only do a narrow I2L conversion if the range check passed.
3851 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3852 _gvn.transform(iff);
3853 RegionNode* region = new RegionNode(3);
3854 _gvn.set_type(region, Type::CONTROL);
3855 lengthx = new PhiNode(region, TypeLong::LONG);
3856 _gvn.set_type(lengthx, TypeLong::LONG);
3857
3858 // Range check passed. Use ConvI2L node with narrow type.
3859 Node* passed = IfFalse(iff);
3860 region->init_req(1, passed);
3861 // Make I2L conversion control dependent to prevent it from
3862 // floating above the range check during loop optimizations.
3863 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3864
3865 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3866 region->init_req(2, IfTrue(iff));
3867 lengthx->init_req(2, ConvI2X(length));
3868
3869 set_control(region);
3870 record_for_igvn(region);
3871 record_for_igvn(lengthx);
3872 }
3873 }
3874 #endif
3875
3876 // Combine header size (plus rounding) and body size. Then round down.
3877 // This computation cannot overflow, because it is used only in two
3878 // places, one where the length is sharply limited, and the other
3879 // after a successful allocation.
3880 Node* abody = lengthx;
3881 if (elem_shift != NULL)
3882 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3883 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3884 if (round_mask != 0) {
3885 Node* mask = MakeConX(~round_mask);
3886 size = _gvn.transform( new AndXNode(size, mask) );
3887 }
3888 // else if round_mask == 0, the size computation is self-rounding
3889
3890 if (return_size_val != NULL) {
3891 // This is the size
3892 (*return_size_val) = size;
3893 }
3894
3895 // Now generate allocation code
3896
3897 // The entire memory state is needed for slow path of the allocation
3898 // since GC and deoptimization can happen.
3899 Node *mem = reset_memory();
3900 set_all_memory(mem); // Create new memory state
3901
3902 if (initial_slow_test->is_Bool()) {
3903 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3904 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3905 }
3906
3907 // Create the AllocateArrayNode and its result projections
3908 AllocateArrayNode* alloc
3909 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3910 control(), mem, i_o(),
3911 size, klass_node,
3912 initial_slow_test,
3913 length);
3914
3915 // Cast to correct type. Note that the klass_node may be constant or not,
3916 // and in the latter case the actual array type will be inexact also.
3917 // (This happens via a non-constant argument to inline_native_newArray.)
3918 // In any case, the value of klass_node provides the desired array type.
3919 const TypeInt* length_type = _gvn.find_int_type(length);
3920 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3921 if (ary_type->isa_aryptr() && length_type != NULL) {
3922 // Try to get a better type than POS for the size
3923 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3924 }
3925
3926 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3927
3928 // Cast length on remaining path to be as narrow as possible
3929 if (map()->find_edge(length) >= 0) {
3930 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3931 if (ccast != length) {
3932 _gvn.set_type_bottom(ccast);
3933 record_for_igvn(ccast);
3934 replace_in_map(length, ccast);
3935 }
3936 }
3937
3938 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
3939 ciKlass* elem_klass = ary_ptr != NULL ? ary_ptr->klass()->as_array_klass()->element_klass() : NULL;
3940 if (elem_klass != NULL && elem_klass->is_valuetype()) {
3941 ciValueKlass* vk = elem_klass->as_value_klass();
3942 if (!vk->flatten_array()) {
3943 // Non-flattened value type arrays need to be initialized with default value type oops
3944 initialize_value_type_array(javaoop, length, elem_klass->as_value_klass(), nargs);
3945 // TODO re-enable once JDK-8189802 is fixed
3946 // InitializeNode* init = alloc->initialization();
3947 //init->set_complete_with_arraycopy();
3948 }
3949 }
3950
3951 return javaoop;
3952 }
3953
3954 void GraphKit::initialize_value_type_array(Node* array, Node* length, ciValueKlass* vk, int nargs) {
3955 // Check for zero length
3956 Node* null_ctl = top();
3957 null_check_common(length, T_INT, false, &null_ctl, false);
3958 if (stopped()) {
3959 set_control(null_ctl); // Always zero
3960 return;
3961 }
3962
3963 RegionNode* res_ctl = new RegionNode(3);
3964 gvn().set_type(res_ctl, Type::CONTROL);
3965 record_for_igvn(res_ctl);
3966
3967 // Length is zero: don't execute initialization loop
3968 res_ctl->init_req(1, null_ctl);
3969 PhiNode* res_io = PhiNode::make(res_ctl, i_o(), Type::ABIO);
3970 PhiNode* res_mem = PhiNode::make(res_ctl, merged_memory(), Type::MEMORY, TypePtr::BOTTOM);
3971 gvn().set_type(res_io, Type::ABIO);
3972 gvn().set_type(res_mem, Type::MEMORY);
3973 record_for_igvn(res_io);
3974 record_for_igvn(res_mem);
3975
3976 // Length is non-zero: execute a loop that initializes the array with the default value type
3977 Node* oop = ValueTypeNode::load_default_oop(gvn(), vk);
3978
3979 add_predicate(nargs);
3980 RegionNode* loop = new RegionNode(3);
3981 loop->init_req(1, control());
3982 PhiNode* index = PhiNode::make(loop, intcon(0), TypeInt::INT);
3983 PhiNode* mem = PhiNode::make(loop, reset_memory(), Type::MEMORY, TypePtr::BOTTOM);
3984
3985 gvn().set_type(loop, Type::CONTROL);
3986 gvn().set_type(index, TypeInt::INT);
3987 gvn().set_type(mem, Type::MEMORY);
3988 record_for_igvn(loop);
3989 record_for_igvn(index);
3990 record_for_igvn(mem);
3991
3992 // Loop body: initialize array element at 'index'
3993 set_control(loop);
3994 set_all_memory(mem);
3995 Node* adr = array_element_address(array, index, T_OBJECT);
3996 const TypeOopPtr* elemtype = TypeInstPtr::make(TypePtr::NotNull, vk);
3997 store_oop_to_array(control(), array, adr, TypeAryPtr::OOPS, oop, elemtype, T_VALUETYPE, MemNode::release);
3998
3999 // Check if we need to execute another loop iteration
4000 length = SubI(length, intcon(1));
4001 IfNode* iff = create_and_map_if(control(), Bool(CmpI(index, length), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4002
4003 // Continue with next iteration
4004 loop->init_req(2, IfTrue(iff));
4005 index->init_req(2, AddI(index, intcon(1)));
4006 mem->init_req(2, merged_memory());
4007
4008 // Exit loop
4009 res_ctl->init_req(2, IfFalse(iff));
4010 res_io->set_req(2, i_o());
4011 res_mem->set_req(2, reset_memory());
4012
4013 // Set merged control, IO and memory
4014 set_control(res_ctl);
4015 set_i_o(res_io);
4016 set_all_memory(res_mem);
4017 }
4018
4019 // The following "Ideal_foo" functions are placed here because they recognize
4020 // the graph shapes created by the functions immediately above.
4021
4022 //---------------------------Ideal_allocation----------------------------------
4023 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4024 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
4025 if (ptr == NULL) { // reduce dumb test in callers
4026 return NULL;
4027 }
4028 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4029 ptr = ptr->in(1);
4030 if (ptr == NULL) return NULL;
4031 }
4032 // Return NULL for allocations with several casts:
4033 // j.l.reflect.Array.newInstance(jobject, jint)
4034 // Object.clone()
4035 // to keep more precise type from last cast.
4036 if (ptr->is_Proj()) {
4037 Node* allo = ptr->in(0);
4038 if (allo != NULL && allo->is_Allocate()) {
4039 return allo->as_Allocate();
4040 }
4041 }
4042 // Report failure to match.
4043 return NULL;
4044 }
4045
4046 // Fancy version which also strips off an offset (and reports it to caller).
4047 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
4048 intptr_t& offset) {
4049 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4050 if (base == NULL) return NULL;
4051 return Ideal_allocation(base, phase);
4052 }
4053
4054 // Trace Initialize <- Proj[Parm] <- Allocate
4055 AllocateNode* InitializeNode::allocation() {
4056 Node* rawoop = in(InitializeNode::RawAddress);
4057 if (rawoop->is_Proj()) {
4058 Node* alloc = rawoop->in(0);
4059 if (alloc->is_Allocate()) {
4060 return alloc->as_Allocate();
4061 }
4062 }
4063 return NULL;
4064 }
4065
4066 // Trace Allocate -> Proj[Parm] -> Initialize
4067 InitializeNode* AllocateNode::initialization() {
4068 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4069 if (rawoop == NULL) return NULL;
4070 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4071 Node* init = rawoop->fast_out(i);
4072 if (init->is_Initialize()) {
4073 assert(init->as_Initialize()->allocation() == this, "2-way link");
4074 return init->as_Initialize();
4075 }
4076 }
4077 return NULL;
4078 }
4079
4080 //----------------------------- loop predicates ---------------------------
4081
4082 //------------------------------add_predicate_impl----------------------------
4083 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
4084 // Too many traps seen?
4085 if (too_many_traps(reason)) {
4086 #ifdef ASSERT
4087 if (TraceLoopPredicate) {
4088 int tc = C->trap_count(reason);
4089 tty->print("too many traps=%s tcount=%d in ",
4090 Deoptimization::trap_reason_name(reason), tc);
4091 method()->print(); // which method has too many predicate traps
4092 tty->cr();
4093 }
4094 #endif
4095 // We cannot afford to take more traps here,
4096 // do not generate predicate.
4097 return;
4098 }
4099
4100 Node *cont = _gvn.intcon(1);
4101 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
4102 Node *bol = _gvn.transform(new Conv2BNode(opq));
4103 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4104 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
4105 C->add_predicate_opaq(opq);
4106 {
4107 PreserveJVMState pjvms(this);
4108 set_control(iffalse);
4109 inc_sp(nargs);
4110 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4111 }
4112 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
4113 set_control(iftrue);
4114 }
4115
4116 //------------------------------add_predicate---------------------------------
4117 void GraphKit::add_predicate(int nargs) {
4118 if (UseLoopPredicate) {
4119 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
4120 }
4121 // loop's limit check predicate should be near the loop.
4122 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
4123 }
4124
4125 //----------------------------- store barriers ----------------------------
4126 #define __ ideal.
4127
4128 bool GraphKit::use_ReduceInitialCardMarks() {
4129 BarrierSet *bs = BarrierSet::barrier_set();
4130 return bs->is_a(BarrierSet::CardTableBarrierSet)
4131 && barrier_set_cast<CardTableBarrierSet>(bs)->can_elide_tlab_store_barriers()
4132 && ReduceInitialCardMarks;
4133 }
4134
4135 void GraphKit::sync_kit(IdealKit& ideal) {
4136 set_all_memory(__ merged_memory());
4137 set_i_o(__ i_o());
4138 set_control(__ ctrl());
4139 }
4140
4141 void GraphKit::final_sync(IdealKit& ideal) {
4142 // Final sync IdealKit and graphKit.
4143 sync_kit(ideal);
4144 }
4145
4146 Node* GraphKit::byte_map_base_node() {
4147 // Get base of card map
4148 jbyte* card_table_base = ci_card_table_address();
4149 if (card_table_base != NULL) {
4150 return makecon(TypeRawPtr::make((address)card_table_base));
4151 } else {
4152 return null();
4153 }
4154 }
4155
4156 // vanilla/CMS post barrier
4157 // Insert a write-barrier store. This is to let generational GC work; we have
4158 // to flag all oop-stores before the next GC point.
4159 void GraphKit::write_barrier_post(Node* oop_store,
4160 Node* obj,
4161 Node* adr,
4162 uint adr_idx,
4163 Node* val,
4164 bool use_precise) {
4165 // No store check needed if we're storing a NULL or an old object
4166 // (latter case is probably a string constant). The concurrent
4167 // mark sweep garbage collector, however, needs to have all nonNull
4168 // oop updates flagged via card-marks.
4169 if (val != NULL && val->is_Con()) {
4170 // must be either an oop or NULL
4171 const Type* t = val->bottom_type();
4172 if (t == TypePtr::NULL_PTR || t == Type::TOP)
4173 // stores of null never (?) need barriers
4174 return;
4175 }
4176
4177 if (use_ReduceInitialCardMarks()
4178 && obj == just_allocated_object(control())) {
4179 // We can skip marks on a freshly-allocated object in Eden.
4180 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
4181 // That routine informs GC to take appropriate compensating steps,
4182 // upon a slow-path allocation, so as to make this card-mark
4183 // elision safe.
4184 return;
4185 }
4186
4187 if (!use_precise) {
4188 // All card marks for a (non-array) instance are in one place:
4189 adr = obj;
4190 }
4191 // (Else it's an array (or unknown), and we want more precise card marks.)
4192 assert(adr != NULL, "");
4193
4194 IdealKit ideal(this, true);
4195
4196 // Convert the pointer to an int prior to doing math on it
4197 Node* cast = __ CastPX(__ ctrl(), adr);
4198
4199 // Divide by card size
4200 assert(BarrierSet::barrier_set()->is_a(BarrierSet::CardTableBarrierSet),
4201 "Only one we handle so far.");
4202 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
4203
4204 // Combine card table base and card offset
4205 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
4206
4207 // Get the alias_index for raw card-mark memory
4208 int adr_type = Compile::AliasIdxRaw;
4209 Node* zero = __ ConI(0); // Dirty card value
4210 BasicType bt = T_BYTE;
4211
4212 if (UseConcMarkSweepGC && UseCondCardMark) {
4213 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
4214 __ sync_kit(this);
4215 }
4216
4217 if (UseCondCardMark) {
4218 // The classic GC reference write barrier is typically implemented
4219 // as a store into the global card mark table. Unfortunately
4220 // unconditional stores can result in false sharing and excessive
4221 // coherence traffic as well as false transactional aborts.
4222 // UseCondCardMark enables MP "polite" conditional card mark
4223 // stores. In theory we could relax the load from ctrl() to
4224 // no_ctrl, but that doesn't buy much latitude.
4225 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
4226 __ if_then(card_val, BoolTest::ne, zero);
4227 }
4228
4229 // Smash zero into card
4230 if( !UseConcMarkSweepGC ) {
4231 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
4232 } else {
4233 // Specialized path for CM store barrier
4234 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
4235 }
4236
4237 if (UseCondCardMark) {
4238 __ end_if();
4239 }
4240
4241 // Final sync IdealKit and GraphKit.
4242 final_sync(ideal);
4243 }
4244
4245 #if INCLUDE_G1GC
4246
4247 /*
4248 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
4249 * required by SATB to make sure all objects live at the start of the
4250 * marking are kept alive, all reference updates need to any previous
4251 * reference stored before writing.
4252 *
4253 * If the previous value is NULL there is no need to save the old value.
4254 * References that are NULL are filtered during runtime by the barrier
4255 * code to avoid unnecessary queuing.
4256 *
4257 * However in the case of newly allocated objects it might be possible to
4258 * prove that the reference about to be overwritten is NULL during compile
4259 * time and avoid adding the barrier code completely.
4260 *
4261 * The compiler needs to determine that the object in which a field is about
4262 * to be written is newly allocated, and that no prior store to the same field
4263 * has happened since the allocation.
4264 *
4265 * Returns true if the pre-barrier can be removed
4266 */
4267 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
4268 BasicType bt, uint adr_idx) {
4269 intptr_t offset = 0;
4270 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4271 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4272
4273 if (offset == Type::OffsetBot) {
4274 return false; // cannot unalias unless there are precise offsets
4275 }
4276
4277 if (alloc == NULL) {
4278 return false; // No allocation found
4279 }
4280
4281 intptr_t size_in_bytes = type2aelembytes(bt);
4282
4283 Node* mem = memory(adr_idx); // start searching here...
4284
4285 for (int cnt = 0; cnt < 50; cnt++) {
4286
4287 if (mem->is_Store()) {
4288
4289 Node* st_adr = mem->in(MemNode::Address);
4290 intptr_t st_offset = 0;
4291 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
4292
4293 if (st_base == NULL) {
4294 break; // inscrutable pointer
4295 }
4296
4297 // Break we have found a store with same base and offset as ours so break
4298 if (st_base == base && st_offset == offset) {
4299 break;
4300 }
4301
4302 if (st_offset != offset && st_offset != Type::OffsetBot) {
4303 const int MAX_STORE = BytesPerLong;
4304 if (st_offset >= offset + size_in_bytes ||
4305 st_offset <= offset - MAX_STORE ||
4306 st_offset <= offset - mem->as_Store()->memory_size()) {
4307 // Success: The offsets are provably independent.
4308 // (You may ask, why not just test st_offset != offset and be done?
4309 // The answer is that stores of different sizes can co-exist
4310 // in the same sequence of RawMem effects. We sometimes initialize
4311 // a whole 'tile' of array elements with a single jint or jlong.)
4312 mem = mem->in(MemNode::Memory);
4313 continue; // advance through independent store memory
4314 }
4315 }
4316
4317 if (st_base != base
4318 && MemNode::detect_ptr_independence(base, alloc, st_base,
4319 AllocateNode::Ideal_allocation(st_base, phase),
4320 phase)) {
4321 // Success: The bases are provably independent.
4322 mem = mem->in(MemNode::Memory);
4323 continue; // advance through independent store memory
4324 }
4325 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4326
4327 InitializeNode* st_init = mem->in(0)->as_Initialize();
4328 AllocateNode* st_alloc = st_init->allocation();
4329
4330 // Make sure that we are looking at the same allocation site.
4331 // The alloc variable is guaranteed to not be null here from earlier check.
4332 if (alloc == st_alloc) {
4333 // Check that the initialization is storing NULL so that no previous store
4334 // has been moved up and directly write a reference
4335 Node* captured_store = st_init->find_captured_store(offset,
4336 type2aelembytes(T_OBJECT),
4337 phase);
4338 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4339 return true;
4340 }
4341 }
4342 }
4343
4344 // Unless there is an explicit 'continue', we must bail out here,
4345 // because 'mem' is an inscrutable memory state (e.g., a call).
4346 break;
4347 }
4348
4349 return false;
4350 }
4351
4352 // G1 pre/post barriers
4353 void GraphKit::g1_write_barrier_pre(bool do_load,
4354 Node* obj,
4355 Node* adr,
4356 uint alias_idx,
4357 Node* val,
4358 const TypeOopPtr* val_type,
4359 Node* pre_val,
4360 BasicType bt) {
4361
4362 // Some sanity checks
4363 // Note: val is unused in this routine.
4364
4365 if (do_load) {
4366 // We need to generate the load of the previous value
4367 assert(obj != NULL, "must have a base");
4368 assert(adr != NULL, "where are loading from?");
4369 assert(pre_val == NULL, "loaded already?");
4370 assert(val_type != NULL, "need a type");
4371
4372 if (use_ReduceInitialCardMarks()
4373 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4374 return;
4375 }
4376
4377 } else {
4378 // In this case both val_type and alias_idx are unused.
4379 assert(pre_val != NULL, "must be loaded already");
4380 // Nothing to be done if pre_val is null.
4381 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4382 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4383 }
4384 assert(bt == T_OBJECT || bt == T_VALUETYPE, "or we shouldn't be here");
4385
4386 IdealKit ideal(this, true);
4387
4388 Node* tls = __ thread(); // ThreadLocalStorage
4389
4390 Node* no_ctrl = NULL;
4391 Node* no_base = __ top();
4392 Node* zero = __ ConI(0);
4393 Node* zeroX = __ ConX(0);
4394
4395 float likely = PROB_LIKELY(0.999);
4396 float unlikely = PROB_UNLIKELY(0.999);
4397
4398 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4399 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4400
4401 // Offsets into the thread
4402 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
4403 const int index_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset());
4404 const int buffer_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
4405
4406 // Now the actual pointers into the thread
4407 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4408 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4409 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4410
4411 // Now some of the values
4412 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4413
4414 // if (!marking)
4415 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4416 BasicType index_bt = TypeX_X->basic_type();
4417 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4418 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4419
4420 if (do_load) {
4421 // load original value
4422 // alias_idx correct??
4423 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4424 }
4425
4426 // if (pre_val != NULL)
4427 __ if_then(pre_val, BoolTest::ne, null()); {
4428 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4429
4430 // is the queue for this thread full?
4431 __ if_then(index, BoolTest::ne, zeroX, likely); {
4432
4433 // decrement the index
4434 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4435
4436 // Now get the buffer location we will log the previous value into and store it
4437 Node *log_addr = __ AddP(no_base, buffer, next_index);
4438 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4439 // update the index
4440 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4441
4442 } __ else_(); {
4443
4444 // logging buffer is full, call the runtime
4445 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4446 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4447 } __ end_if(); // (!index)
4448 } __ end_if(); // (pre_val != NULL)
4449 } __ end_if(); // (!marking)
4450
4451 // Final sync IdealKit and GraphKit.
4452 final_sync(ideal);
4453 }
4454
4455 /*
4456 * G1 similar to any GC with a Young Generation requires a way to keep track of
4457 * references from Old Generation to Young Generation to make sure all live
4458 * objects are found. G1 also requires to keep track of object references
4459 * between different regions to enable evacuation of old regions, which is done
4460 * as part of mixed collections. References are tracked in remembered sets and
4461 * is continuously updated as reference are written to with the help of the
4462 * post-barrier.
4463 *
4464 * To reduce the number of updates to the remembered set the post-barrier
4465 * filters updates to fields in objects located in the Young Generation,
4466 * the same region as the reference, when the NULL is being written or
4467 * if the card is already marked as dirty by an earlier write.
4468 *
4469 * Under certain circumstances it is possible to avoid generating the
4470 * post-barrier completely if it is possible during compile time to prove
4471 * the object is newly allocated and that no safepoint exists between the
4472 * allocation and the store.
4473 *
4474 * In the case of slow allocation the allocation code must handle the barrier
4475 * as part of the allocation in the case the allocated object is not located
4476 * in the nursery, this would happen for humongous objects. This is similar to
4477 * how CMS is required to handle this case, see the comments for the method
4478 * CardTableBarrierSet::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier.
4479 * A deferred card mark is required for these objects and handled in the above
4480 * mentioned methods.
4481 *
4482 * Returns true if the post barrier can be removed
4483 */
4484 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4485 Node* adr) {
4486 intptr_t offset = 0;
4487 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4488 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4489
4490 if (offset == Type::OffsetBot) {
4491 return false; // cannot unalias unless there are precise offsets
4492 }
4493
4494 if (alloc == NULL) {
4495 return false; // No allocation found
4496 }
4497
4498 // Start search from Store node
4499 Node* mem = store->in(MemNode::Control);
4500 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4501
4502 InitializeNode* st_init = mem->in(0)->as_Initialize();
4503 AllocateNode* st_alloc = st_init->allocation();
4504
4505 // Make sure we are looking at the same allocation
4506 if (alloc == st_alloc) {
4507 return true;
4508 }
4509 }
4510
4511 return false;
4512 }
4513
4514 //
4515 // Update the card table and add card address to the queue
4516 //
4517 void GraphKit::g1_mark_card(IdealKit& ideal,
4518 Node* card_adr,
4519 Node* oop_store,
4520 uint oop_alias_idx,
4521 Node* index,
4522 Node* index_adr,
4523 Node* buffer,
4524 const TypeFunc* tf) {
4525
4526 Node* zero = __ ConI(0);
4527 Node* zeroX = __ ConX(0);
4528 Node* no_base = __ top();
4529 BasicType card_bt = T_BYTE;
4530 // Smash zero into card. MUST BE ORDERED WRT TO STORE
4531 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4532
4533 // Now do the queue work
4534 __ if_then(index, BoolTest::ne, zeroX); {
4535
4536 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4537 Node* log_addr = __ AddP(no_base, buffer, next_index);
4538
4539 // Order, see storeCM.
4540 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4541 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4542
4543 } __ else_(); {
4544 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4545 } __ end_if();
4546
4547 }
4548
4549 void GraphKit::g1_write_barrier_post(Node* oop_store,
4550 Node* obj,
4551 Node* adr,
4552 uint alias_idx,
4553 Node* val,
4554 BasicType bt,
4555 bool use_precise) {
4556 // If we are writing a NULL then we need no post barrier
4557
4558 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4559 // Must be NULL
4560 const Type* t = val->bottom_type();
4561 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4562 // No post barrier if writing NULLx
4563 return;
4564 }
4565
4566 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4567 // We can skip marks on a freshly-allocated object in Eden.
4568 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
4569 // That routine informs GC to take appropriate compensating steps,
4570 // upon a slow-path allocation, so as to make this card-mark
4571 // elision safe.
4572 return;
4573 }
4574
4575 if (use_ReduceInitialCardMarks()
4576 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4577 return;
4578 }
4579
4580 if (!use_precise) {
4581 // All card marks for a (non-array) instance are in one place:
4582 adr = obj;
4583 }
4584 // (Else it's an array (or unknown), and we want more precise card marks.)
4585 assert(adr != NULL, "");
4586
4587 IdealKit ideal(this, true);
4588
4589 Node* tls = __ thread(); // ThreadLocalStorage
4590
4591 Node* no_base = __ top();
4592 float likely = PROB_LIKELY(0.999);
4593 float unlikely = PROB_UNLIKELY(0.999);
4594 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val());
4595 Node* dirty_card = __ ConI((jint)CardTable::dirty_card_val());
4596 Node* zeroX = __ ConX(0);
4597
4598 // Get the alias_index for raw card-mark memory
4599 const TypePtr* card_type = TypeRawPtr::BOTTOM;
4600
4601 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4602
4603 // Offsets into the thread
4604 const int index_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset());
4605 const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset());
4606
4607 // Pointers into the thread
4608
4609 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4610 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4611
4612 // Now some values
4613 // Use ctrl to avoid hoisting these values past a safepoint, which could
4614 // potentially reset these fields in the JavaThread.
4615 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4616 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4617
4618 // Convert the store obj pointer to an int prior to doing math on it
4619 // Must use ctrl to prevent "integerized oop" existing across safepoint
4620 Node* cast = __ CastPX(__ ctrl(), adr);
4621
4622 // Divide pointer by card size
4623 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
4624
4625 // Combine card table base and card offset
4626 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4627
4628 // If we know the value being stored does it cross regions?
4629
4630 if (val != NULL) {
4631 // Does the store cause us to cross regions?
4632
4633 // Should be able to do an unsigned compare of region_size instead of
4634 // and extra shift. Do we have an unsigned compare??
4635 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4636 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4637
4638 // if (xor_res == 0) same region so skip
4639 __ if_then(xor_res, BoolTest::ne, zeroX); {
4640
4641 // No barrier if we are storing a NULL
4642 __ if_then(val, BoolTest::ne, null(), unlikely); {
4643
4644 // Ok must mark the card if not already dirty
4645
4646 // load the original value of the card
4647 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4648
4649 __ if_then(card_val, BoolTest::ne, young_card); {
4650 sync_kit(ideal);
4651 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4652 insert_mem_bar(Op_MemBarVolatile, oop_store);
4653 __ sync_kit(this);
4654
4655 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4656 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4657 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4658 } __ end_if();
4659 } __ end_if();
4660 } __ end_if();
4661 } __ end_if();
4662 } else {
4663 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
4664 // We don't need a barrier here if the destination is a newly allocated object
4665 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
4666 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
4667 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
4668 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4669 __ if_then(card_val, BoolTest::ne, young_card); {
4670 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4671 } __ end_if();
4672 }
4673
4674 // Final sync IdealKit and GraphKit.
4675 final_sync(ideal);
4676 }
4677 #undef __
4678
4679 #endif // INCLUDE_G1GC
4680
4681 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4682 Node* len = load_array_length(load_String_value(ctrl, str));
4683 Node* coder = load_String_coder(ctrl, str);
4684 // Divide length by 2 if coder is UTF16
4685 return _gvn.transform(new RShiftINode(len, coder));
4686 }
4687
4688 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4689 int value_offset = java_lang_String::value_offset_in_bytes();
4690 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4691 false, NULL, Type::Offset(0));
4692 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4693 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4694 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4695 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4696 int value_field_idx = C->get_alias_index(value_field_type);
4697 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4698 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4699 // String.value field is known to be @Stable.
4700 if (UseImplicitStableValues) {
4701 load = cast_array_to_stable(load, value_type);
4702 }
4703 return load;
4704 }
4705
4706 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4707 if (!CompactStrings) {
4708 return intcon(java_lang_String::CODER_UTF16);
4709 }
4710 int coder_offset = java_lang_String::coder_offset_in_bytes();
4711 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4712 false, NULL, Type::Offset(0));
4713 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4714 int coder_field_idx = C->get_alias_index(coder_field_type);
4715 return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4716 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4717 }
4718
4719 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4720 int value_offset = java_lang_String::value_offset_in_bytes();
4721 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4722 false, NULL, Type::Offset(0));
4723 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4724 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4725 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4726 }
4727
4728 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4729 int coder_offset = java_lang_String::coder_offset_in_bytes();
4730 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4731 false, NULL, Type::Offset(0));
4732 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4733 int coder_field_idx = C->get_alias_index(coder_field_type);
4734 store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4735 value, T_BYTE, coder_field_idx, MemNode::unordered);
4736 }
4737
4738 // Capture src and dst memory state with a MergeMemNode
4739 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4740 if (src_type == dst_type) {
4741 // Types are equal, we don't need a MergeMemNode
4742 return memory(src_type);
4743 }
4744 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4745 record_for_igvn(merge); // fold it up later, if possible
4746 int src_idx = C->get_alias_index(src_type);
4747 int dst_idx = C->get_alias_index(dst_type);
4748 merge->set_memory_at(src_idx, memory(src_idx));
4749 merge->set_memory_at(dst_idx, memory(dst_idx));
4750 return merge;
4751 }
4752
4753 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4754 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4755 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4756 // If input and output memory types differ, capture both states to preserve
4757 // the dependency between preceding and subsequent loads/stores.
4758 // For example, the following program:
4759 // StoreB
4760 // compress_string
4761 // LoadB
4762 // has this memory graph (use->def):
4763 // LoadB -> compress_string -> CharMem
4764 // ... -> StoreB -> ByteMem
4765 // The intrinsic hides the dependency between LoadB and StoreB, causing
4766 // the load to read from memory not containing the result of the StoreB.
4767 // The correct memory graph should look like this:
4768 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4769 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4770 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4771 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4772 set_memory(res_mem, TypeAryPtr::BYTES);
4773 return str;
4774 }
4775
4776 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4777 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4778 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4779 // Capture src and dst memory (see comment in 'compress_string').
4780 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4781 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4782 set_memory(_gvn.transform(str), dst_type);
4783 }
4784
4785 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4786 /**
4787 * int i_char = start;
4788 * for (int i_byte = 0; i_byte < count; i_byte++) {
4789 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4790 * }
4791 */
4792 add_predicate();
4793 RegionNode* head = new RegionNode(3);
4794 head->init_req(1, control());
4795 gvn().set_type(head, Type::CONTROL);
4796 record_for_igvn(head);
4797
4798 Node* i_byte = new PhiNode(head, TypeInt::INT);
4799 i_byte->init_req(1, intcon(0));
4800 gvn().set_type(i_byte, TypeInt::INT);
4801 record_for_igvn(i_byte);
4802
4803 Node* i_char = new PhiNode(head, TypeInt::INT);
4804 i_char->init_req(1, start);
4805 gvn().set_type(i_char, TypeInt::INT);
4806 record_for_igvn(i_char);
4807
4808 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4809 gvn().set_type(mem, Type::MEMORY);
4810 record_for_igvn(mem);
4811 set_control(head);
4812 set_memory(mem, TypeAryPtr::BYTES);
4813 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4814 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4815 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4816 false, false, true /* mismatched */);
4817
4818 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4819 head->init_req(2, IfTrue(iff));
4820 mem->init_req(2, st);
4821 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4822 i_char->init_req(2, AddI(i_char, intcon(2)));
4823
4824 set_control(IfFalse(iff));
4825 set_memory(st, TypeAryPtr::BYTES);
4826 }
4827
4828 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4829 if (!field->is_constant()) {
4830 return NULL; // Field not marked as constant.
4831 }
4832 ciInstance* holder = NULL;
4833 if (!field->is_static()) {
4834 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4835 if (const_oop != NULL && const_oop->is_instance()) {
4836 holder = const_oop->as_instance();
4837 }
4838 }
4839 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4840 /*is_unsigned_load=*/false);
4841 if (con_type != NULL) {
4842 Node* con = makecon(con_type);
4843 if (field->layout_type() == T_VALUETYPE) {
4844 // Load value type from constant oop
4845 con = ValueTypeNode::make_from_oop(this, con, field->type()->as_value_klass());
4846 }
4847 return con;
4848 }
4849 return NULL;
4850 }
4851
4852 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4853 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4854 // assumption of CCP analysis.
4855 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4856 }