1 /*
2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/cardTable.hpp"
30 #include "gc/shared/cardTableBarrierSet.hpp"
31 #include "gc/shared/collectedHeap.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "runtime/deoptimization.hpp"
47 #include "runtime/sharedRuntime.hpp"
48 #if INCLUDE_G1GC
49 #include "gc/g1/g1CardTable.hpp"
50 #include "gc/g1/g1ThreadLocalData.hpp"
51 #include "gc/g1/heapRegion.hpp"
52 #endif // INCLUDE_ALL_GCS
53
54 //----------------------------GraphKit-----------------------------------------
55 // Main utility constructor.
56 GraphKit::GraphKit(JVMState* jvms)
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn())
60 {
61 _exceptions = jvms->map()->next_exception();
62 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
63 set_jvms(jvms);
64 }
65
66 // Private constructor for parser.
67 GraphKit::GraphKit()
68 : Phase(Phase::Parser),
69 _env(C->env()),
70 _gvn(*C->initial_gvn())
71 {
72 _exceptions = NULL;
73 set_map(NULL);
74 debug_only(_sp = -99);
75 debug_only(set_bci(-99));
76 }
77
78
79
80 //---------------------------clean_stack---------------------------------------
81 // Clear away rubbish from the stack area of the JVM state.
82 // This destroys any arguments that may be waiting on the stack.
83 void GraphKit::clean_stack(int from_sp) {
84 SafePointNode* map = this->map();
85 JVMState* jvms = this->jvms();
86 int stk_size = jvms->stk_size();
87 int stkoff = jvms->stkoff();
88 Node* top = this->top();
89 for (int i = from_sp; i < stk_size; i++) {
90 if (map->in(stkoff + i) != top) {
91 map->set_req(stkoff + i, top);
92 }
93 }
94 }
95
96
97 //--------------------------------sync_jvms-----------------------------------
98 // Make sure our current jvms agrees with our parse state.
99 JVMState* GraphKit::sync_jvms() const {
100 JVMState* jvms = this->jvms();
101 jvms->set_bci(bci()); // Record the new bci in the JVMState
102 jvms->set_sp(sp()); // Record the new sp in the JVMState
103 assert(jvms_in_sync(), "jvms is now in sync");
104 return jvms;
105 }
106
107 //--------------------------------sync_jvms_for_reexecute---------------------
108 // Make sure our current jvms agrees with our parse state. This version
109 // uses the reexecute_sp for reexecuting bytecodes.
110 JVMState* GraphKit::sync_jvms_for_reexecute() {
111 JVMState* jvms = this->jvms();
112 jvms->set_bci(bci()); // Record the new bci in the JVMState
113 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
114 return jvms;
115 }
116
117 #ifdef ASSERT
118 bool GraphKit::jvms_in_sync() const {
119 Parse* parse = is_Parse();
120 if (parse == NULL) {
121 if (bci() != jvms()->bci()) return false;
122 if (sp() != (int)jvms()->sp()) return false;
123 return true;
124 }
125 if (jvms()->method() != parse->method()) return false;
126 if (jvms()->bci() != parse->bci()) return false;
127 int jvms_sp = jvms()->sp();
128 if (jvms_sp != parse->sp()) return false;
129 int jvms_depth = jvms()->depth();
130 if (jvms_depth != parse->depth()) return false;
131 return true;
132 }
133
134 // Local helper checks for special internal merge points
135 // used to accumulate and merge exception states.
136 // They are marked by the region's in(0) edge being the map itself.
137 // Such merge points must never "escape" into the parser at large,
138 // until they have been handed to gvn.transform.
139 static bool is_hidden_merge(Node* reg) {
140 if (reg == NULL) return false;
141 if (reg->is_Phi()) {
142 reg = reg->in(0);
143 if (reg == NULL) return false;
144 }
145 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
146 }
147
148 void GraphKit::verify_map() const {
149 if (map() == NULL) return; // null map is OK
150 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
151 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
152 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
153 }
154
155 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
156 assert(ex_map->next_exception() == NULL, "not already part of a chain");
157 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
158 }
159 #endif
160
161 //---------------------------stop_and_kill_map---------------------------------
162 // Set _map to NULL, signalling a stop to further bytecode execution.
163 // First smash the current map's control to a constant, to mark it dead.
164 void GraphKit::stop_and_kill_map() {
165 SafePointNode* dead_map = stop();
166 if (dead_map != NULL) {
167 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
168 assert(dead_map->is_killed(), "must be so marked");
169 }
170 }
171
172
173 //--------------------------------stopped--------------------------------------
174 // Tell if _map is NULL, or control is top.
175 bool GraphKit::stopped() {
176 if (map() == NULL) return true;
177 else if (control() == top()) return true;
178 else return false;
179 }
180
181
182 //-----------------------------has_ex_handler----------------------------------
183 // Tell if this method or any caller method has exception handlers.
184 bool GraphKit::has_ex_handler() {
185 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
186 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
187 return true;
188 }
189 }
190 return false;
191 }
192
193 //------------------------------save_ex_oop------------------------------------
194 // Save an exception without blowing stack contents or other JVM state.
195 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
196 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
197 ex_map->add_req(ex_oop);
198 debug_only(verify_exception_state(ex_map));
199 }
200
201 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
202 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
203 Node* ex_oop = ex_map->in(ex_map->req()-1);
204 if (clear_it) ex_map->del_req(ex_map->req()-1);
205 return ex_oop;
206 }
207
208 //-----------------------------saved_ex_oop------------------------------------
209 // Recover a saved exception from its map.
210 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
211 return common_saved_ex_oop(ex_map, false);
212 }
213
214 //--------------------------clear_saved_ex_oop---------------------------------
215 // Erase a previously saved exception from its map.
216 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
217 return common_saved_ex_oop(ex_map, true);
218 }
219
220 #ifdef ASSERT
221 //---------------------------has_saved_ex_oop----------------------------------
222 // Erase a previously saved exception from its map.
223 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
224 return ex_map->req() == ex_map->jvms()->endoff()+1;
225 }
226 #endif
227
228 //-------------------------make_exception_state--------------------------------
229 // Turn the current JVM state into an exception state, appending the ex_oop.
230 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
231 sync_jvms();
232 SafePointNode* ex_map = stop(); // do not manipulate this map any more
233 set_saved_ex_oop(ex_map, ex_oop);
234 return ex_map;
235 }
236
237
238 //--------------------------add_exception_state--------------------------------
239 // Add an exception to my list of exceptions.
240 void GraphKit::add_exception_state(SafePointNode* ex_map) {
241 if (ex_map == NULL || ex_map->control() == top()) {
242 return;
243 }
244 #ifdef ASSERT
245 verify_exception_state(ex_map);
246 if (has_exceptions()) {
247 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
248 }
249 #endif
250
251 // If there is already an exception of exactly this type, merge with it.
252 // In particular, null-checks and other low-level exceptions common up here.
253 Node* ex_oop = saved_ex_oop(ex_map);
254 const Type* ex_type = _gvn.type(ex_oop);
255 if (ex_oop == top()) {
256 // No action needed.
257 return;
258 }
259 assert(ex_type->isa_instptr(), "exception must be an instance");
260 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
261 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
262 // We check sp also because call bytecodes can generate exceptions
263 // both before and after arguments are popped!
264 if (ex_type2 == ex_type
265 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
266 combine_exception_states(ex_map, e2);
267 return;
268 }
269 }
270
271 // No pre-existing exception of the same type. Chain it on the list.
272 push_exception_state(ex_map);
273 }
274
275 //-----------------------add_exception_states_from-----------------------------
276 void GraphKit::add_exception_states_from(JVMState* jvms) {
277 SafePointNode* ex_map = jvms->map()->next_exception();
278 if (ex_map != NULL) {
279 jvms->map()->set_next_exception(NULL);
280 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
281 next_map = ex_map->next_exception();
282 ex_map->set_next_exception(NULL);
283 add_exception_state(ex_map);
284 }
285 }
286 }
287
288 //-----------------------transfer_exceptions_into_jvms-------------------------
289 JVMState* GraphKit::transfer_exceptions_into_jvms() {
290 if (map() == NULL) {
291 // We need a JVMS to carry the exceptions, but the map has gone away.
292 // Create a scratch JVMS, cloned from any of the exception states...
293 if (has_exceptions()) {
294 _map = _exceptions;
295 _map = clone_map();
296 _map->set_next_exception(NULL);
297 clear_saved_ex_oop(_map);
298 debug_only(verify_map());
299 } else {
300 // ...or created from scratch
301 JVMState* jvms = new (C) JVMState(_method, NULL);
302 jvms->set_bci(_bci);
303 jvms->set_sp(_sp);
304 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
305 set_jvms(jvms);
306 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
307 set_all_memory(top());
308 while (map()->req() < jvms->endoff()) map()->add_req(top());
309 }
310 // (This is a kludge, in case you didn't notice.)
311 set_control(top());
312 }
313 JVMState* jvms = sync_jvms();
314 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
315 jvms->map()->set_next_exception(_exceptions);
316 _exceptions = NULL; // done with this set of exceptions
317 return jvms;
318 }
319
320 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
321 assert(is_hidden_merge(dstphi), "must be a special merge node");
322 assert(is_hidden_merge(srcphi), "must be a special merge node");
323 uint limit = srcphi->req();
324 for (uint i = PhiNode::Input; i < limit; i++) {
325 dstphi->add_req(srcphi->in(i));
326 }
327 }
328 static inline void add_one_req(Node* dstphi, Node* src) {
329 assert(is_hidden_merge(dstphi), "must be a special merge node");
330 assert(!is_hidden_merge(src), "must not be a special merge node");
331 dstphi->add_req(src);
332 }
333
334 //-----------------------combine_exception_states------------------------------
335 // This helper function combines exception states by building phis on a
336 // specially marked state-merging region. These regions and phis are
337 // untransformed, and can build up gradually. The region is marked by
338 // having a control input of its exception map, rather than NULL. Such
339 // regions do not appear except in this function, and in use_exception_state.
340 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
341 if (failing()) return; // dying anyway...
342 JVMState* ex_jvms = ex_map->_jvms;
343 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
344 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
345 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
346 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
347 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
348 assert(ex_map->req() == phi_map->req(), "matching maps");
349 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
350 Node* hidden_merge_mark = root();
351 Node* region = phi_map->control();
352 MergeMemNode* phi_mem = phi_map->merged_memory();
353 MergeMemNode* ex_mem = ex_map->merged_memory();
354 if (region->in(0) != hidden_merge_mark) {
355 // The control input is not (yet) a specially-marked region in phi_map.
356 // Make it so, and build some phis.
357 region = new RegionNode(2);
358 _gvn.set_type(region, Type::CONTROL);
359 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
360 region->init_req(1, phi_map->control());
361 phi_map->set_control(region);
362 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
363 record_for_igvn(io_phi);
364 _gvn.set_type(io_phi, Type::ABIO);
365 phi_map->set_i_o(io_phi);
366 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
367 Node* m = mms.memory();
368 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
369 record_for_igvn(m_phi);
370 _gvn.set_type(m_phi, Type::MEMORY);
371 mms.set_memory(m_phi);
372 }
373 }
374
375 // Either or both of phi_map and ex_map might already be converted into phis.
376 Node* ex_control = ex_map->control();
377 // if there is special marking on ex_map also, we add multiple edges from src
378 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
379 // how wide was the destination phi_map, originally?
380 uint orig_width = region->req();
381
382 if (add_multiple) {
383 add_n_reqs(region, ex_control);
384 add_n_reqs(phi_map->i_o(), ex_map->i_o());
385 } else {
386 // ex_map has no merges, so we just add single edges everywhere
387 add_one_req(region, ex_control);
388 add_one_req(phi_map->i_o(), ex_map->i_o());
389 }
390 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
391 if (mms.is_empty()) {
392 // get a copy of the base memory, and patch some inputs into it
393 const TypePtr* adr_type = mms.adr_type(C);
394 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
395 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
396 mms.set_memory(phi);
397 // Prepare to append interesting stuff onto the newly sliced phi:
398 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
399 }
400 // Append stuff from ex_map:
401 if (add_multiple) {
402 add_n_reqs(mms.memory(), mms.memory2());
403 } else {
404 add_one_req(mms.memory(), mms.memory2());
405 }
406 }
407 uint limit = ex_map->req();
408 for (uint i = TypeFunc::Parms; i < limit; i++) {
409 // Skip everything in the JVMS after tos. (The ex_oop follows.)
410 if (i == tos) i = ex_jvms->monoff();
411 Node* src = ex_map->in(i);
412 Node* dst = phi_map->in(i);
413 if (src != dst) {
414 PhiNode* phi;
415 if (dst->in(0) != region) {
416 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
417 record_for_igvn(phi);
418 _gvn.set_type(phi, phi->type());
419 phi_map->set_req(i, dst);
420 // Prepare to append interesting stuff onto the new phi:
421 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
422 } else {
423 assert(dst->is_Phi(), "nobody else uses a hidden region");
424 phi = dst->as_Phi();
425 }
426 if (add_multiple && src->in(0) == ex_control) {
427 // Both are phis.
428 add_n_reqs(dst, src);
429 } else {
430 while (dst->req() < region->req()) add_one_req(dst, src);
431 }
432 const Type* srctype = _gvn.type(src);
433 if (phi->type() != srctype) {
434 const Type* dsttype = phi->type()->meet_speculative(srctype);
435 if (phi->type() != dsttype) {
436 phi->set_type(dsttype);
437 _gvn.set_type(phi, dsttype);
438 }
439 }
440 }
441 }
442 phi_map->merge_replaced_nodes_with(ex_map);
443 }
444
445 //--------------------------use_exception_state--------------------------------
446 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
447 if (failing()) { stop(); return top(); }
448 Node* region = phi_map->control();
449 Node* hidden_merge_mark = root();
450 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
451 Node* ex_oop = clear_saved_ex_oop(phi_map);
452 if (region->in(0) == hidden_merge_mark) {
453 // Special marking for internal ex-states. Process the phis now.
454 region->set_req(0, region); // now it's an ordinary region
455 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
456 // Note: Setting the jvms also sets the bci and sp.
457 set_control(_gvn.transform(region));
458 uint tos = jvms()->stkoff() + sp();
459 for (uint i = 1; i < tos; i++) {
460 Node* x = phi_map->in(i);
461 if (x->in(0) == region) {
462 assert(x->is_Phi(), "expected a special phi");
463 phi_map->set_req(i, _gvn.transform(x));
464 }
465 }
466 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
467 Node* x = mms.memory();
468 if (x->in(0) == region) {
469 assert(x->is_Phi(), "nobody else uses a hidden region");
470 mms.set_memory(_gvn.transform(x));
471 }
472 }
473 if (ex_oop->in(0) == region) {
474 assert(ex_oop->is_Phi(), "expected a special phi");
475 ex_oop = _gvn.transform(ex_oop);
476 }
477 } else {
478 set_jvms(phi_map->jvms());
479 }
480
481 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
482 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
483 return ex_oop;
484 }
485
486 //---------------------------------java_bc-------------------------------------
487 Bytecodes::Code GraphKit::java_bc() const {
488 ciMethod* method = this->method();
489 int bci = this->bci();
490 if (method != NULL && bci != InvocationEntryBci)
491 return method->java_code_at_bci(bci);
492 else
493 return Bytecodes::_illegal;
494 }
495
496 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
497 bool must_throw) {
498 // if the exception capability is set, then we will generate code
499 // to check the JavaThread.should_post_on_exceptions flag to see
500 // if we actually need to report exception events (for this
501 // thread). If we don't need to report exception events, we will
502 // take the normal fast path provided by add_exception_events. If
503 // exception event reporting is enabled for this thread, we will
504 // take the uncommon_trap in the BuildCutout below.
505
506 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
507 Node* jthread = _gvn.transform(new ThreadLocalNode());
508 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
509 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
510
511 // Test the should_post_on_exceptions_flag vs. 0
512 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
513 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
514
515 // Branch to slow_path if should_post_on_exceptions_flag was true
516 { BuildCutout unless(this, tst, PROB_MAX);
517 // Do not try anything fancy if we're notifying the VM on every throw.
518 // Cf. case Bytecodes::_athrow in parse2.cpp.
519 uncommon_trap(reason, Deoptimization::Action_none,
520 (ciKlass*)NULL, (char*)NULL, must_throw);
521 }
522
523 }
524
525 //------------------------------builtin_throw----------------------------------
526 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
527 bool must_throw = true;
528
529 if (env()->jvmti_can_post_on_exceptions()) {
530 // check if we must post exception events, take uncommon trap if so
531 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
532 // here if should_post_on_exceptions is false
533 // continue on with the normal codegen
534 }
535
536 // If this particular condition has not yet happened at this
537 // bytecode, then use the uncommon trap mechanism, and allow for
538 // a future recompilation if several traps occur here.
539 // If the throw is hot, try to use a more complicated inline mechanism
540 // which keeps execution inside the compiled code.
541 bool treat_throw_as_hot = false;
542 ciMethodData* md = method()->method_data();
543
544 if (ProfileTraps) {
545 if (too_many_traps(reason)) {
546 treat_throw_as_hot = true;
547 }
548 // (If there is no MDO at all, assume it is early in
549 // execution, and that any deopts are part of the
550 // startup transient, and don't need to be remembered.)
551
552 // Also, if there is a local exception handler, treat all throws
553 // as hot if there has been at least one in this method.
554 if (C->trap_count(reason) != 0
555 && method()->method_data()->trap_count(reason) != 0
556 && has_ex_handler()) {
557 treat_throw_as_hot = true;
558 }
559 }
560
561 // If this throw happens frequently, an uncommon trap might cause
562 // a performance pothole. If there is a local exception handler,
563 // and if this particular bytecode appears to be deoptimizing often,
564 // let us handle the throw inline, with a preconstructed instance.
565 // Note: If the deopt count has blown up, the uncommon trap
566 // runtime is going to flush this nmethod, not matter what.
567 if (treat_throw_as_hot
568 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
569 // If the throw is local, we use a pre-existing instance and
570 // punt on the backtrace. This would lead to a missing backtrace
571 // (a repeat of 4292742) if the backtrace object is ever asked
572 // for its backtrace.
573 // Fixing this remaining case of 4292742 requires some flavor of
574 // escape analysis. Leave that for the future.
575 ciInstance* ex_obj = NULL;
576 switch (reason) {
577 case Deoptimization::Reason_null_check:
578 ex_obj = env()->NullPointerException_instance();
579 break;
580 case Deoptimization::Reason_div0_check:
581 ex_obj = env()->ArithmeticException_instance();
582 break;
583 case Deoptimization::Reason_range_check:
584 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
585 break;
586 case Deoptimization::Reason_class_check:
587 if (java_bc() == Bytecodes::_aastore) {
588 ex_obj = env()->ArrayStoreException_instance();
589 } else {
590 ex_obj = env()->ClassCastException_instance();
591 }
592 break;
593 default:
594 break;
595 }
596 if (failing()) { stop(); return; } // exception allocation might fail
597 if (ex_obj != NULL) {
598 // Cheat with a preallocated exception object.
599 if (C->log() != NULL)
600 C->log()->elem("hot_throw preallocated='1' reason='%s'",
601 Deoptimization::trap_reason_name(reason));
602 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
603 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
604
605 // Clear the detail message of the preallocated exception object.
606 // Weblogic sometimes mutates the detail message of exceptions
607 // using reflection.
608 int offset = java_lang_Throwable::get_detailMessage_offset();
609 const TypePtr* adr_typ = ex_con->add_offset(offset);
610
611 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
612 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
613 // Conservatively release stores of object references.
614 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
615
616 add_exception_state(make_exception_state(ex_node));
617 return;
618 }
619 }
620
621 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
622 // It won't be much cheaper than bailing to the interp., since we'll
623 // have to pass up all the debug-info, and the runtime will have to
624 // create the stack trace.
625
626 // Usual case: Bail to interpreter.
627 // Reserve the right to recompile if we haven't seen anything yet.
628
629 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
630 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
631 if (treat_throw_as_hot
632 && (method()->method_data()->trap_recompiled_at(bci(), m)
633 || C->too_many_traps(reason))) {
634 // We cannot afford to take more traps here. Suffer in the interpreter.
635 if (C->log() != NULL)
636 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
637 Deoptimization::trap_reason_name(reason),
638 C->trap_count(reason));
639 action = Deoptimization::Action_none;
640 }
641
642 // "must_throw" prunes the JVM state to include only the stack, if there
643 // are no local exception handlers. This should cut down on register
644 // allocation time and code size, by drastically reducing the number
645 // of in-edges on the call to the uncommon trap.
646
647 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
648 }
649
650
651 //----------------------------PreserveJVMState---------------------------------
652 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
653 debug_only(kit->verify_map());
654 _kit = kit;
655 _map = kit->map(); // preserve the map
656 _sp = kit->sp();
657 kit->set_map(clone_map ? kit->clone_map() : NULL);
658 #ifdef ASSERT
659 _bci = kit->bci();
660 Parse* parser = kit->is_Parse();
661 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
662 _block = block;
663 #endif
664 }
665 PreserveJVMState::~PreserveJVMState() {
666 GraphKit* kit = _kit;
667 #ifdef ASSERT
668 assert(kit->bci() == _bci, "bci must not shift");
669 Parse* parser = kit->is_Parse();
670 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
671 assert(block == _block, "block must not shift");
672 #endif
673 kit->set_map(_map);
674 kit->set_sp(_sp);
675 }
676
677
678 //-----------------------------BuildCutout-------------------------------------
679 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
680 : PreserveJVMState(kit)
681 {
682 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
683 SafePointNode* outer_map = _map; // preserved map is caller's
684 SafePointNode* inner_map = kit->map();
685 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
686 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
687 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
688 }
689 BuildCutout::~BuildCutout() {
690 GraphKit* kit = _kit;
691 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
692 }
693
694 //---------------------------PreserveReexecuteState----------------------------
695 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
696 assert(!kit->stopped(), "must call stopped() before");
697 _kit = kit;
698 _sp = kit->sp();
699 _reexecute = kit->jvms()->_reexecute;
700 }
701 PreserveReexecuteState::~PreserveReexecuteState() {
702 if (_kit->stopped()) return;
703 _kit->jvms()->_reexecute = _reexecute;
704 _kit->set_sp(_sp);
705 }
706
707 //------------------------------clone_map--------------------------------------
708 // Implementation of PreserveJVMState
709 //
710 // Only clone_map(...) here. If this function is only used in the
711 // PreserveJVMState class we may want to get rid of this extra
712 // function eventually and do it all there.
713
714 SafePointNode* GraphKit::clone_map() {
715 if (map() == NULL) return NULL;
716
717 // Clone the memory edge first
718 Node* mem = MergeMemNode::make(map()->memory());
719 gvn().set_type_bottom(mem);
720
721 SafePointNode *clonemap = (SafePointNode*)map()->clone();
722 JVMState* jvms = this->jvms();
723 JVMState* clonejvms = jvms->clone_shallow(C);
724 clonemap->set_memory(mem);
725 clonemap->set_jvms(clonejvms);
726 clonejvms->set_map(clonemap);
727 record_for_igvn(clonemap);
728 gvn().set_type_bottom(clonemap);
729 return clonemap;
730 }
731
732
733 //-----------------------------set_map_clone-----------------------------------
734 void GraphKit::set_map_clone(SafePointNode* m) {
735 _map = m;
736 _map = clone_map();
737 _map->set_next_exception(NULL);
738 debug_only(verify_map());
739 }
740
741
742 //----------------------------kill_dead_locals---------------------------------
743 // Detect any locals which are known to be dead, and force them to top.
744 void GraphKit::kill_dead_locals() {
745 // Consult the liveness information for the locals. If any
746 // of them are unused, then they can be replaced by top(). This
747 // should help register allocation time and cut down on the size
748 // of the deoptimization information.
749
750 // This call is made from many of the bytecode handling
751 // subroutines called from the Big Switch in do_one_bytecode.
752 // Every bytecode which might include a slow path is responsible
753 // for killing its dead locals. The more consistent we
754 // are about killing deads, the fewer useless phis will be
755 // constructed for them at various merge points.
756
757 // bci can be -1 (InvocationEntryBci). We return the entry
758 // liveness for the method.
759
760 if (method() == NULL || method()->code_size() == 0) {
761 // We are building a graph for a call to a native method.
762 // All locals are live.
763 return;
764 }
765
766 ResourceMark rm;
767
768 // Consult the liveness information for the locals. If any
769 // of them are unused, then they can be replaced by top(). This
770 // should help register allocation time and cut down on the size
771 // of the deoptimization information.
772 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
773
774 int len = (int)live_locals.size();
775 assert(len <= jvms()->loc_size(), "too many live locals");
776 for (int local = 0; local < len; local++) {
777 if (!live_locals.at(local)) {
778 set_local(local, top());
779 }
780 }
781 }
782
783 #ifdef ASSERT
784 //-------------------------dead_locals_are_killed------------------------------
785 // Return true if all dead locals are set to top in the map.
786 // Used to assert "clean" debug info at various points.
787 bool GraphKit::dead_locals_are_killed() {
788 if (method() == NULL || method()->code_size() == 0) {
789 // No locals need to be dead, so all is as it should be.
790 return true;
791 }
792
793 // Make sure somebody called kill_dead_locals upstream.
794 ResourceMark rm;
795 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
796 if (jvms->loc_size() == 0) continue; // no locals to consult
797 SafePointNode* map = jvms->map();
798 ciMethod* method = jvms->method();
799 int bci = jvms->bci();
800 if (jvms == this->jvms()) {
801 bci = this->bci(); // it might not yet be synched
802 }
803 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
804 int len = (int)live_locals.size();
805 if (!live_locals.is_valid() || len == 0)
806 // This method is trivial, or is poisoned by a breakpoint.
807 return true;
808 assert(len == jvms->loc_size(), "live map consistent with locals map");
809 for (int local = 0; local < len; local++) {
810 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
811 if (PrintMiscellaneous && (Verbose || WizardMode)) {
812 tty->print_cr("Zombie local %d: ", local);
813 jvms->dump();
814 }
815 return false;
816 }
817 }
818 }
819 return true;
820 }
821
822 #endif //ASSERT
823
824 // Helper function for enforcing certain bytecodes to reexecute if
825 // deoptimization happens
826 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
827 ciMethod* cur_method = jvms->method();
828 int cur_bci = jvms->bci();
829 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
830 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
831 return Interpreter::bytecode_should_reexecute(code) ||
832 (is_anewarray && code == Bytecodes::_multianewarray);
833 // Reexecute _multianewarray bytecode which was replaced with
834 // sequence of [a]newarray. See Parse::do_multianewarray().
835 //
836 // Note: interpreter should not have it set since this optimization
837 // is limited by dimensions and guarded by flag so in some cases
838 // multianewarray() runtime calls will be generated and
839 // the bytecode should not be reexecutes (stack will not be reset).
840 } else
841 return false;
842 }
843
844 // Helper function for adding JVMState and debug information to node
845 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
846 // Add the safepoint edges to the call (or other safepoint).
847
848 // Make sure dead locals are set to top. This
849 // should help register allocation time and cut down on the size
850 // of the deoptimization information.
851 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
852
853 // Walk the inline list to fill in the correct set of JVMState's
854 // Also fill in the associated edges for each JVMState.
855
856 // If the bytecode needs to be reexecuted we need to put
857 // the arguments back on the stack.
858 const bool should_reexecute = jvms()->should_reexecute();
859 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
860
861 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
862 // undefined if the bci is different. This is normal for Parse but it
863 // should not happen for LibraryCallKit because only one bci is processed.
864 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
865 "in LibraryCallKit the reexecute bit should not change");
866
867 // If we are guaranteed to throw, we can prune everything but the
868 // input to the current bytecode.
869 bool can_prune_locals = false;
870 uint stack_slots_not_pruned = 0;
871 int inputs = 0, depth = 0;
872 if (must_throw) {
873 assert(method() == youngest_jvms->method(), "sanity");
874 if (compute_stack_effects(inputs, depth)) {
875 can_prune_locals = true;
876 stack_slots_not_pruned = inputs;
877 }
878 }
879
880 if (env()->should_retain_local_variables()) {
881 // At any safepoint, this method can get breakpointed, which would
882 // then require an immediate deoptimization.
883 can_prune_locals = false; // do not prune locals
884 stack_slots_not_pruned = 0;
885 }
886
887 // do not scribble on the input jvms
888 JVMState* out_jvms = youngest_jvms->clone_deep(C);
889 call->set_jvms(out_jvms); // Start jvms list for call node
890
891 // For a known set of bytecodes, the interpreter should reexecute them if
892 // deoptimization happens. We set the reexecute state for them here
893 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
894 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
895 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
896 }
897
898 // Presize the call:
899 DEBUG_ONLY(uint non_debug_edges = call->req());
900 call->add_req_batch(top(), youngest_jvms->debug_depth());
901 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
902
903 // Set up edges so that the call looks like this:
904 // Call [state:] ctl io mem fptr retadr
905 // [parms:] parm0 ... parmN
906 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
907 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
908 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
909 // Note that caller debug info precedes callee debug info.
910
911 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
912 uint debug_ptr = call->req();
913
914 // Loop over the map input edges associated with jvms, add them
915 // to the call node, & reset all offsets to match call node array.
916 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
917 uint debug_end = debug_ptr;
918 uint debug_start = debug_ptr - in_jvms->debug_size();
919 debug_ptr = debug_start; // back up the ptr
920
921 uint p = debug_start; // walks forward in [debug_start, debug_end)
922 uint j, k, l;
923 SafePointNode* in_map = in_jvms->map();
924 out_jvms->set_map(call);
925
926 if (can_prune_locals) {
927 assert(in_jvms->method() == out_jvms->method(), "sanity");
928 // If the current throw can reach an exception handler in this JVMS,
929 // then we must keep everything live that can reach that handler.
930 // As a quick and dirty approximation, we look for any handlers at all.
931 if (in_jvms->method()->has_exception_handlers()) {
932 can_prune_locals = false;
933 }
934 }
935
936 // Add the Locals
937 k = in_jvms->locoff();
938 l = in_jvms->loc_size();
939 out_jvms->set_locoff(p);
940 if (!can_prune_locals) {
941 for (j = 0; j < l; j++)
942 call->set_req(p++, in_map->in(k+j));
943 } else {
944 p += l; // already set to top above by add_req_batch
945 }
946
947 // Add the Expression Stack
948 k = in_jvms->stkoff();
949 l = in_jvms->sp();
950 out_jvms->set_stkoff(p);
951 if (!can_prune_locals) {
952 for (j = 0; j < l; j++)
953 call->set_req(p++, in_map->in(k+j));
954 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
955 // Divide stack into {S0,...,S1}, where S0 is set to top.
956 uint s1 = stack_slots_not_pruned;
957 stack_slots_not_pruned = 0; // for next iteration
958 if (s1 > l) s1 = l;
959 uint s0 = l - s1;
960 p += s0; // skip the tops preinstalled by add_req_batch
961 for (j = s0; j < l; j++)
962 call->set_req(p++, in_map->in(k+j));
963 } else {
964 p += l; // already set to top above by add_req_batch
965 }
966
967 // Add the Monitors
968 k = in_jvms->monoff();
969 l = in_jvms->mon_size();
970 out_jvms->set_monoff(p);
971 for (j = 0; j < l; j++)
972 call->set_req(p++, in_map->in(k+j));
973
974 // Copy any scalar object fields.
975 k = in_jvms->scloff();
976 l = in_jvms->scl_size();
977 out_jvms->set_scloff(p);
978 for (j = 0; j < l; j++)
979 call->set_req(p++, in_map->in(k+j));
980
981 // Finish the new jvms.
982 out_jvms->set_endoff(p);
983
984 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
985 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
986 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
987 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
988 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
989 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
990
991 // Update the two tail pointers in parallel.
992 out_jvms = out_jvms->caller();
993 in_jvms = in_jvms->caller();
994 }
995
996 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
997
998 // Test the correctness of JVMState::debug_xxx accessors:
999 assert(call->jvms()->debug_start() == non_debug_edges, "");
1000 assert(call->jvms()->debug_end() == call->req(), "");
1001 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1002 }
1003
1004 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1005 Bytecodes::Code code = java_bc();
1006 if (code == Bytecodes::_wide) {
1007 code = method()->java_code_at_bci(bci() + 1);
1008 }
1009
1010 BasicType rtype = T_ILLEGAL;
1011 int rsize = 0;
1012
1013 if (code != Bytecodes::_illegal) {
1014 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1015 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1016 if (rtype < T_CONFLICT)
1017 rsize = type2size[rtype];
1018 }
1019
1020 switch (code) {
1021 case Bytecodes::_illegal:
1022 return false;
1023
1024 case Bytecodes::_ldc:
1025 case Bytecodes::_ldc_w:
1026 case Bytecodes::_ldc2_w:
1027 inputs = 0;
1028 break;
1029
1030 case Bytecodes::_dup: inputs = 1; break;
1031 case Bytecodes::_dup_x1: inputs = 2; break;
1032 case Bytecodes::_dup_x2: inputs = 3; break;
1033 case Bytecodes::_dup2: inputs = 2; break;
1034 case Bytecodes::_dup2_x1: inputs = 3; break;
1035 case Bytecodes::_dup2_x2: inputs = 4; break;
1036 case Bytecodes::_swap: inputs = 2; break;
1037 case Bytecodes::_arraylength: inputs = 1; break;
1038
1039 case Bytecodes::_getstatic:
1040 case Bytecodes::_putstatic:
1041 case Bytecodes::_getfield:
1042 case Bytecodes::_putfield:
1043 {
1044 bool ignored_will_link;
1045 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1046 int size = field->type()->size();
1047 bool is_get = (depth >= 0), is_static = (depth & 1);
1048 inputs = (is_static ? 0 : 1);
1049 if (is_get) {
1050 depth = size - inputs;
1051 } else {
1052 inputs += size; // putxxx pops the value from the stack
1053 depth = - inputs;
1054 }
1055 }
1056 break;
1057
1058 case Bytecodes::_invokevirtual:
1059 case Bytecodes::_invokespecial:
1060 case Bytecodes::_invokestatic:
1061 case Bytecodes::_invokedynamic:
1062 case Bytecodes::_invokeinterface:
1063 {
1064 bool ignored_will_link;
1065 ciSignature* declared_signature = NULL;
1066 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1067 assert(declared_signature != NULL, "cannot be null");
1068 inputs = declared_signature->arg_size_for_bc(code);
1069 int size = declared_signature->return_type()->size();
1070 depth = size - inputs;
1071 }
1072 break;
1073
1074 case Bytecodes::_multianewarray:
1075 {
1076 ciBytecodeStream iter(method());
1077 iter.reset_to_bci(bci());
1078 iter.next();
1079 inputs = iter.get_dimensions();
1080 assert(rsize == 1, "");
1081 depth = rsize - inputs;
1082 }
1083 break;
1084
1085 case Bytecodes::_ireturn:
1086 case Bytecodes::_lreturn:
1087 case Bytecodes::_freturn:
1088 case Bytecodes::_dreturn:
1089 case Bytecodes::_areturn:
1090 assert(rsize == -depth, "");
1091 inputs = rsize;
1092 break;
1093
1094 case Bytecodes::_jsr:
1095 case Bytecodes::_jsr_w:
1096 inputs = 0;
1097 depth = 1; // S.B. depth=1, not zero
1098 break;
1099
1100 default:
1101 // bytecode produces a typed result
1102 inputs = rsize - depth;
1103 assert(inputs >= 0, "");
1104 break;
1105 }
1106
1107 #ifdef ASSERT
1108 // spot check
1109 int outputs = depth + inputs;
1110 assert(outputs >= 0, "sanity");
1111 switch (code) {
1112 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1113 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1114 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1115 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1116 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1117 default: break;
1118 }
1119 #endif //ASSERT
1120
1121 return true;
1122 }
1123
1124
1125
1126 //------------------------------basic_plus_adr---------------------------------
1127 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1128 // short-circuit a common case
1129 if (offset == intcon(0)) return ptr;
1130 return _gvn.transform( new AddPNode(base, ptr, offset) );
1131 }
1132
1133 Node* GraphKit::ConvI2L(Node* offset) {
1134 // short-circuit a common case
1135 jint offset_con = find_int_con(offset, Type::OffsetBot);
1136 if (offset_con != Type::OffsetBot) {
1137 return longcon((jlong) offset_con);
1138 }
1139 return _gvn.transform( new ConvI2LNode(offset));
1140 }
1141
1142 Node* GraphKit::ConvI2UL(Node* offset) {
1143 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1144 if (offset_con != (juint) Type::OffsetBot) {
1145 return longcon((julong) offset_con);
1146 }
1147 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1148 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1149 return _gvn.transform( new AndLNode(conv, mask) );
1150 }
1151
1152 Node* GraphKit::ConvL2I(Node* offset) {
1153 // short-circuit a common case
1154 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1155 if (offset_con != (jlong)Type::OffsetBot) {
1156 return intcon((int) offset_con);
1157 }
1158 return _gvn.transform( new ConvL2INode(offset));
1159 }
1160
1161 //-------------------------load_object_klass-----------------------------------
1162 Node* GraphKit::load_object_klass(Node* obj) {
1163 // Special-case a fresh allocation to avoid building nodes:
1164 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1165 if (akls != NULL) return akls;
1166 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1167 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1168 }
1169
1170 //-------------------------load_array_length-----------------------------------
1171 Node* GraphKit::load_array_length(Node* array) {
1172 // Special-case a fresh allocation to avoid building nodes:
1173 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1174 Node *alen;
1175 if (alloc == NULL) {
1176 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1177 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1178 } else {
1179 alen = alloc->Ideal_length();
1180 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1181 if (ccast != alen) {
1182 alen = _gvn.transform(ccast);
1183 }
1184 }
1185 return alen;
1186 }
1187
1188 //------------------------------do_null_check----------------------------------
1189 // Helper function to do a NULL pointer check. Returned value is
1190 // the incoming address with NULL casted away. You are allowed to use the
1191 // not-null value only if you are control dependent on the test.
1192 #ifndef PRODUCT
1193 extern int explicit_null_checks_inserted,
1194 explicit_null_checks_elided;
1195 #endif
1196 Node* GraphKit::null_check_common(Node* value, BasicType type,
1197 // optional arguments for variations:
1198 bool assert_null,
1199 Node* *null_control,
1200 bool speculative) {
1201 assert(!assert_null || null_control == NULL, "not both at once");
1202 if (stopped()) return top();
1203 NOT_PRODUCT(explicit_null_checks_inserted++);
1204
1205 // Construct NULL check
1206 Node *chk = NULL;
1207 switch(type) {
1208 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1209 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1210 case T_ARRAY : // fall through
1211 type = T_OBJECT; // simplify further tests
1212 case T_OBJECT : {
1213 const Type *t = _gvn.type( value );
1214
1215 const TypeOopPtr* tp = t->isa_oopptr();
1216 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1217 // Only for do_null_check, not any of its siblings:
1218 && !assert_null && null_control == NULL) {
1219 // Usually, any field access or invocation on an unloaded oop type
1220 // will simply fail to link, since the statically linked class is
1221 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1222 // the static class is loaded but the sharper oop type is not.
1223 // Rather than checking for this obscure case in lots of places,
1224 // we simply observe that a null check on an unloaded class
1225 // will always be followed by a nonsense operation, so we
1226 // can just issue the uncommon trap here.
1227 // Our access to the unloaded class will only be correct
1228 // after it has been loaded and initialized, which requires
1229 // a trip through the interpreter.
1230 #ifndef PRODUCT
1231 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1232 #endif
1233 uncommon_trap(Deoptimization::Reason_unloaded,
1234 Deoptimization::Action_reinterpret,
1235 tp->klass(), "!loaded");
1236 return top();
1237 }
1238
1239 if (assert_null) {
1240 // See if the type is contained in NULL_PTR.
1241 // If so, then the value is already null.
1242 if (t->higher_equal(TypePtr::NULL_PTR)) {
1243 NOT_PRODUCT(explicit_null_checks_elided++);
1244 return value; // Elided null assert quickly!
1245 }
1246 } else {
1247 // See if mixing in the NULL pointer changes type.
1248 // If so, then the NULL pointer was not allowed in the original
1249 // type. In other words, "value" was not-null.
1250 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1251 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1252 NOT_PRODUCT(explicit_null_checks_elided++);
1253 return value; // Elided null check quickly!
1254 }
1255 }
1256 chk = new CmpPNode( value, null() );
1257 break;
1258 }
1259
1260 default:
1261 fatal("unexpected type: %s", type2name(type));
1262 }
1263 assert(chk != NULL, "sanity check");
1264 chk = _gvn.transform(chk);
1265
1266 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1267 BoolNode *btst = new BoolNode( chk, btest);
1268 Node *tst = _gvn.transform( btst );
1269
1270 //-----------
1271 // if peephole optimizations occurred, a prior test existed.
1272 // If a prior test existed, maybe it dominates as we can avoid this test.
1273 if (tst != btst && type == T_OBJECT) {
1274 // At this point we want to scan up the CFG to see if we can
1275 // find an identical test (and so avoid this test altogether).
1276 Node *cfg = control();
1277 int depth = 0;
1278 while( depth < 16 ) { // Limit search depth for speed
1279 if( cfg->Opcode() == Op_IfTrue &&
1280 cfg->in(0)->in(1) == tst ) {
1281 // Found prior test. Use "cast_not_null" to construct an identical
1282 // CastPP (and hence hash to) as already exists for the prior test.
1283 // Return that casted value.
1284 if (assert_null) {
1285 replace_in_map(value, null());
1286 return null(); // do not issue the redundant test
1287 }
1288 Node *oldcontrol = control();
1289 set_control(cfg);
1290 Node *res = cast_not_null(value);
1291 set_control(oldcontrol);
1292 NOT_PRODUCT(explicit_null_checks_elided++);
1293 return res;
1294 }
1295 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1296 if (cfg == NULL) break; // Quit at region nodes
1297 depth++;
1298 }
1299 }
1300
1301 //-----------
1302 // Branch to failure if null
1303 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1304 Deoptimization::DeoptReason reason;
1305 if (assert_null) {
1306 reason = Deoptimization::reason_null_assert(speculative);
1307 } else if (type == T_OBJECT) {
1308 reason = Deoptimization::reason_null_check(speculative);
1309 } else {
1310 reason = Deoptimization::Reason_div0_check;
1311 }
1312 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1313 // ciMethodData::has_trap_at will return a conservative -1 if any
1314 // must-be-null assertion has failed. This could cause performance
1315 // problems for a method after its first do_null_assert failure.
1316 // Consider using 'Reason_class_check' instead?
1317
1318 // To cause an implicit null check, we set the not-null probability
1319 // to the maximum (PROB_MAX). For an explicit check the probability
1320 // is set to a smaller value.
1321 if (null_control != NULL || too_many_traps(reason)) {
1322 // probability is less likely
1323 ok_prob = PROB_LIKELY_MAG(3);
1324 } else if (!assert_null &&
1325 (ImplicitNullCheckThreshold > 0) &&
1326 method() != NULL &&
1327 (method()->method_data()->trap_count(reason)
1328 >= (uint)ImplicitNullCheckThreshold)) {
1329 ok_prob = PROB_LIKELY_MAG(3);
1330 }
1331
1332 if (null_control != NULL) {
1333 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1334 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1335 set_control( _gvn.transform( new IfTrueNode(iff)));
1336 #ifndef PRODUCT
1337 if (null_true == top()) {
1338 explicit_null_checks_elided++;
1339 }
1340 #endif
1341 (*null_control) = null_true;
1342 } else {
1343 BuildCutout unless(this, tst, ok_prob);
1344 // Check for optimizer eliding test at parse time
1345 if (stopped()) {
1346 // Failure not possible; do not bother making uncommon trap.
1347 NOT_PRODUCT(explicit_null_checks_elided++);
1348 } else if (assert_null) {
1349 uncommon_trap(reason,
1350 Deoptimization::Action_make_not_entrant,
1351 NULL, "assert_null");
1352 } else {
1353 replace_in_map(value, zerocon(type));
1354 builtin_throw(reason);
1355 }
1356 }
1357
1358 // Must throw exception, fall-thru not possible?
1359 if (stopped()) {
1360 return top(); // No result
1361 }
1362
1363 if (assert_null) {
1364 // Cast obj to null on this path.
1365 replace_in_map(value, zerocon(type));
1366 return zerocon(type);
1367 }
1368
1369 // Cast obj to not-null on this path, if there is no null_control.
1370 // (If there is a null_control, a non-null value may come back to haunt us.)
1371 if (type == T_OBJECT) {
1372 Node* cast = cast_not_null(value, false);
1373 if (null_control == NULL || (*null_control) == top())
1374 replace_in_map(value, cast);
1375 value = cast;
1376 }
1377
1378 return value;
1379 }
1380
1381
1382 //------------------------------cast_not_null----------------------------------
1383 // Cast obj to not-null on this path
1384 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1385 const Type *t = _gvn.type(obj);
1386 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1387 // Object is already not-null?
1388 if( t == t_not_null ) return obj;
1389
1390 Node *cast = new CastPPNode(obj,t_not_null);
1391 cast->init_req(0, control());
1392 cast = _gvn.transform( cast );
1393
1394 // Scan for instances of 'obj' in the current JVM mapping.
1395 // These instances are known to be not-null after the test.
1396 if (do_replace_in_map)
1397 replace_in_map(obj, cast);
1398
1399 return cast; // Return casted value
1400 }
1401
1402 // Sometimes in intrinsics, we implicitly know an object is not null
1403 // (there's no actual null check) so we can cast it to not null. In
1404 // the course of optimizations, the input to the cast can become null.
1405 // In that case that data path will die and we need the control path
1406 // to become dead as well to keep the graph consistent. So we have to
1407 // add a check for null for which one branch can't be taken. It uses
1408 // an Opaque4 node that will cause the check to be removed after loop
1409 // opts so the test goes away and the compiled code doesn't execute a
1410 // useless check.
1411 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1412 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1413 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1414 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1415 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1416 _gvn.set_type(iff, iff->Value(&_gvn));
1417 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1418 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1419 Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1420 C->root()->add_req(halt);
1421 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1422 set_control(if_t);
1423 return cast_not_null(value, do_replace_in_map);
1424 }
1425
1426
1427 //--------------------------replace_in_map-------------------------------------
1428 void GraphKit::replace_in_map(Node* old, Node* neww) {
1429 if (old == neww) {
1430 return;
1431 }
1432
1433 map()->replace_edge(old, neww);
1434
1435 // Note: This operation potentially replaces any edge
1436 // on the map. This includes locals, stack, and monitors
1437 // of the current (innermost) JVM state.
1438
1439 // don't let inconsistent types from profiling escape this
1440 // method
1441
1442 const Type* told = _gvn.type(old);
1443 const Type* tnew = _gvn.type(neww);
1444
1445 if (!tnew->higher_equal(told)) {
1446 return;
1447 }
1448
1449 map()->record_replaced_node(old, neww);
1450 }
1451
1452
1453 //=============================================================================
1454 //--------------------------------memory---------------------------------------
1455 Node* GraphKit::memory(uint alias_idx) {
1456 MergeMemNode* mem = merged_memory();
1457 Node* p = mem->memory_at(alias_idx);
1458 _gvn.set_type(p, Type::MEMORY); // must be mapped
1459 return p;
1460 }
1461
1462 //-----------------------------reset_memory------------------------------------
1463 Node* GraphKit::reset_memory() {
1464 Node* mem = map()->memory();
1465 // do not use this node for any more parsing!
1466 debug_only( map()->set_memory((Node*)NULL) );
1467 return _gvn.transform( mem );
1468 }
1469
1470 //------------------------------set_all_memory---------------------------------
1471 void GraphKit::set_all_memory(Node* newmem) {
1472 Node* mergemem = MergeMemNode::make(newmem);
1473 gvn().set_type_bottom(mergemem);
1474 map()->set_memory(mergemem);
1475 }
1476
1477 //------------------------------set_all_memory_call----------------------------
1478 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1479 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1480 set_all_memory(newmem);
1481 }
1482
1483 //=============================================================================
1484 //
1485 // parser factory methods for MemNodes
1486 //
1487 // These are layered on top of the factory methods in LoadNode and StoreNode,
1488 // and integrate with the parser's memory state and _gvn engine.
1489 //
1490
1491 // factory methods in "int adr_idx"
1492 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1493 int adr_idx,
1494 MemNode::MemOrd mo,
1495 LoadNode::ControlDependency control_dependency,
1496 bool require_atomic_access,
1497 bool unaligned,
1498 bool mismatched) {
1499 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1500 const TypePtr* adr_type = NULL; // debug-mode-only argument
1501 debug_only(adr_type = C->get_adr_type(adr_idx));
1502 Node* mem = memory(adr_idx);
1503 Node* ld;
1504 if (require_atomic_access && bt == T_LONG) {
1505 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1506 } else if (require_atomic_access && bt == T_DOUBLE) {
1507 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1508 } else {
1509 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1510 }
1511 ld = _gvn.transform(ld);
1512 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1513 // Improve graph before escape analysis and boxing elimination.
1514 record_for_igvn(ld);
1515 }
1516 return ld;
1517 }
1518
1519 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1520 int adr_idx,
1521 MemNode::MemOrd mo,
1522 bool require_atomic_access,
1523 bool unaligned,
1524 bool mismatched) {
1525 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1526 const TypePtr* adr_type = NULL;
1527 debug_only(adr_type = C->get_adr_type(adr_idx));
1528 Node *mem = memory(adr_idx);
1529 Node* st;
1530 if (require_atomic_access && bt == T_LONG) {
1531 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1532 } else if (require_atomic_access && bt == T_DOUBLE) {
1533 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1534 } else {
1535 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1536 }
1537 if (unaligned) {
1538 st->as_Store()->set_unaligned_access();
1539 }
1540 if (mismatched) {
1541 st->as_Store()->set_mismatched_access();
1542 }
1543 st = _gvn.transform(st);
1544 set_memory(st, adr_idx);
1545 // Back-to-back stores can only remove intermediate store with DU info
1546 // so push on worklist for optimizer.
1547 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1548 record_for_igvn(st);
1549
1550 return st;
1551 }
1552
1553
1554 void GraphKit::pre_barrier(bool do_load,
1555 Node* ctl,
1556 Node* obj,
1557 Node* adr,
1558 uint adr_idx,
1559 Node* val,
1560 const TypeOopPtr* val_type,
1561 Node* pre_val,
1562 BasicType bt) {
1563
1564 BarrierSet* bs = BarrierSet::barrier_set();
1565 set_control(ctl);
1566 switch (bs->kind()) {
1567
1568 #if INCLUDE_G1GC
1569 case BarrierSet::G1BarrierSet:
1570 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1571 break;
1572 #endif
1573
1574 #if INCLUDE_EPSILONGC
1575 case BarrierSet::EpsilonBarrierSet:
1576 break;
1577 #endif
1578
1579 case BarrierSet::CardTableBarrierSet:
1580 break;
1581
1582 default :
1583 ShouldNotReachHere();
1584
1585 }
1586 }
1587
1588 bool GraphKit::can_move_pre_barrier() const {
1589 BarrierSet* bs = BarrierSet::barrier_set();
1590 switch (bs->kind()) {
1591
1592 #if INCLUDE_G1GC
1593 case BarrierSet::G1BarrierSet:
1594 return true; // Can move it if no safepoint
1595 #endif
1596
1597 #if INCLUDE_EPSILONGC
1598 case BarrierSet::EpsilonBarrierSet:
1599 return true; // There is no pre-barrier
1600 #endif
1601
1602 case BarrierSet::CardTableBarrierSet:
1603 return true; // There is no pre-barrier
1604
1605 default :
1606 ShouldNotReachHere();
1607 }
1608 return false;
1609 }
1610
1611 void GraphKit::post_barrier(Node* ctl,
1612 Node* store,
1613 Node* obj,
1614 Node* adr,
1615 uint adr_idx,
1616 Node* val,
1617 BasicType bt,
1618 bool use_precise) {
1619 BarrierSet* bs = BarrierSet::barrier_set();
1620 set_control(ctl);
1621 switch (bs->kind()) {
1622 #if INCLUDE_G1GC
1623 case BarrierSet::G1BarrierSet:
1624 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1625 break;
1626 #endif
1627
1628 #if INCLUDE_EPSILONGC
1629 case BarrierSet::EpsilonBarrierSet:
1630 break;
1631 #endif
1632
1633 case BarrierSet::CardTableBarrierSet:
1634 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1635 break;
1636
1637 default :
1638 ShouldNotReachHere();
1639
1640 }
1641 }
1642
1643 Node* GraphKit::store_oop(Node* ctl,
1644 Node* obj,
1645 Node* adr,
1646 const TypePtr* adr_type,
1647 Node* val,
1648 const TypeOopPtr* val_type,
1649 BasicType bt,
1650 bool use_precise,
1651 MemNode::MemOrd mo,
1652 bool mismatched) {
1653 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1654 // could be delayed during Parse (for example, in adjust_map_after_if()).
1655 // Execute transformation here to avoid barrier generation in such case.
1656 if (_gvn.type(val) == TypePtr::NULL_PTR)
1657 val = _gvn.makecon(TypePtr::NULL_PTR);
1658
1659 set_control(ctl);
1660 if (stopped()) return top(); // Dead path ?
1661
1662 assert(bt == T_OBJECT, "sanity");
1663 assert(val != NULL, "not dead path");
1664 uint adr_idx = C->get_alias_index(adr_type);
1665 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1666
1667 pre_barrier(true /* do_load */,
1668 control(), obj, adr, adr_idx, val, val_type,
1669 NULL /* pre_val */,
1670 bt);
1671
1672 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1673 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1674 return store;
1675 }
1676
1677 // Could be an array or object we don't know at compile time (unsafe ref.)
1678 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1679 Node* obj, // containing obj
1680 Node* adr, // actual adress to store val at
1681 const TypePtr* adr_type,
1682 Node* val,
1683 BasicType bt,
1684 MemNode::MemOrd mo,
1685 bool mismatched) {
1686 Compile::AliasType* at = C->alias_type(adr_type);
1687 const TypeOopPtr* val_type = NULL;
1688 if (adr_type->isa_instptr()) {
1689 if (at->field() != NULL) {
1690 // known field. This code is a copy of the do_put_xxx logic.
1691 ciField* field = at->field();
1692 if (!field->type()->is_loaded()) {
1693 val_type = TypeInstPtr::BOTTOM;
1694 } else {
1695 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1696 }
1697 }
1698 } else if (adr_type->isa_aryptr()) {
1699 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1700 }
1701 if (val_type == NULL) {
1702 val_type = TypeInstPtr::BOTTOM;
1703 }
1704 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1705 }
1706
1707
1708 //-------------------------array_element_address-------------------------
1709 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1710 const TypeInt* sizetype, Node* ctrl) {
1711 uint shift = exact_log2(type2aelembytes(elembt));
1712 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1713
1714 // short-circuit a common case (saves lots of confusing waste motion)
1715 jint idx_con = find_int_con(idx, -1);
1716 if (idx_con >= 0) {
1717 intptr_t offset = header + ((intptr_t)idx_con << shift);
1718 return basic_plus_adr(ary, offset);
1719 }
1720
1721 // must be correct type for alignment purposes
1722 Node* base = basic_plus_adr(ary, header);
1723 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1724 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1725 return basic_plus_adr(ary, base, scale);
1726 }
1727
1728 //-------------------------load_array_element-------------------------
1729 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1730 const Type* elemtype = arytype->elem();
1731 BasicType elembt = elemtype->array_element_basic_type();
1732 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1733 if (elembt == T_NARROWOOP) {
1734 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1735 }
1736 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1737 return ld;
1738 }
1739
1740 //-------------------------set_arguments_for_java_call-------------------------
1741 // Arguments (pre-popped from the stack) are taken from the JVMS.
1742 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1743 // Add the call arguments:
1744 uint nargs = call->method()->arg_size();
1745 for (uint i = 0; i < nargs; i++) {
1746 Node* arg = argument(i);
1747 call->init_req(i + TypeFunc::Parms, arg);
1748 }
1749 }
1750
1751 //---------------------------set_edges_for_java_call---------------------------
1752 // Connect a newly created call into the current JVMS.
1753 // A return value node (if any) is returned from set_edges_for_java_call.
1754 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1755
1756 // Add the predefined inputs:
1757 call->init_req( TypeFunc::Control, control() );
1758 call->init_req( TypeFunc::I_O , i_o() );
1759 call->init_req( TypeFunc::Memory , reset_memory() );
1760 call->init_req( TypeFunc::FramePtr, frameptr() );
1761 call->init_req( TypeFunc::ReturnAdr, top() );
1762
1763 add_safepoint_edges(call, must_throw);
1764
1765 Node* xcall = _gvn.transform(call);
1766
1767 if (xcall == top()) {
1768 set_control(top());
1769 return;
1770 }
1771 assert(xcall == call, "call identity is stable");
1772
1773 // Re-use the current map to produce the result.
1774
1775 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1776 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1777 set_all_memory_call(xcall, separate_io_proj);
1778
1779 //return xcall; // no need, caller already has it
1780 }
1781
1782 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1783 if (stopped()) return top(); // maybe the call folded up?
1784
1785 // Capture the return value, if any.
1786 Node* ret;
1787 if (call->method() == NULL ||
1788 call->method()->return_type()->basic_type() == T_VOID)
1789 ret = top();
1790 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1791
1792 // Note: Since any out-of-line call can produce an exception,
1793 // we always insert an I_O projection from the call into the result.
1794
1795 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1796
1797 if (separate_io_proj) {
1798 // The caller requested separate projections be used by the fall
1799 // through and exceptional paths, so replace the projections for
1800 // the fall through path.
1801 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1802 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1803 }
1804 return ret;
1805 }
1806
1807 //--------------------set_predefined_input_for_runtime_call--------------------
1808 // Reading and setting the memory state is way conservative here.
1809 // The real problem is that I am not doing real Type analysis on memory,
1810 // so I cannot distinguish card mark stores from other stores. Across a GC
1811 // point the Store Barrier and the card mark memory has to agree. I cannot
1812 // have a card mark store and its barrier split across the GC point from
1813 // either above or below. Here I get that to happen by reading ALL of memory.
1814 // A better answer would be to separate out card marks from other memory.
1815 // For now, return the input memory state, so that it can be reused
1816 // after the call, if this call has restricted memory effects.
1817 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1818 // Set fixed predefined input arguments
1819 Node* memory = reset_memory();
1820 call->init_req( TypeFunc::Control, control() );
1821 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1822 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1823 call->init_req( TypeFunc::FramePtr, frameptr() );
1824 call->init_req( TypeFunc::ReturnAdr, top() );
1825 return memory;
1826 }
1827
1828 //-------------------set_predefined_output_for_runtime_call--------------------
1829 // Set control and memory (not i_o) from the call.
1830 // If keep_mem is not NULL, use it for the output state,
1831 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1832 // If hook_mem is NULL, this call produces no memory effects at all.
1833 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1834 // then only that memory slice is taken from the call.
1835 // In the last case, we must put an appropriate memory barrier before
1836 // the call, so as to create the correct anti-dependencies on loads
1837 // preceding the call.
1838 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1839 Node* keep_mem,
1840 const TypePtr* hook_mem) {
1841 // no i/o
1842 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1843 if (keep_mem) {
1844 // First clone the existing memory state
1845 set_all_memory(keep_mem);
1846 if (hook_mem != NULL) {
1847 // Make memory for the call
1848 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1849 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1850 // We also use hook_mem to extract specific effects from arraycopy stubs.
1851 set_memory(mem, hook_mem);
1852 }
1853 // ...else the call has NO memory effects.
1854
1855 // Make sure the call advertises its memory effects precisely.
1856 // This lets us build accurate anti-dependences in gcm.cpp.
1857 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1858 "call node must be constructed correctly");
1859 } else {
1860 assert(hook_mem == NULL, "");
1861 // This is not a "slow path" call; all memory comes from the call.
1862 set_all_memory_call(call);
1863 }
1864 }
1865
1866
1867 // Replace the call with the current state of the kit.
1868 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1869 JVMState* ejvms = NULL;
1870 if (has_exceptions()) {
1871 ejvms = transfer_exceptions_into_jvms();
1872 }
1873
1874 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1875 ReplacedNodes replaced_nodes_exception;
1876 Node* ex_ctl = top();
1877
1878 SafePointNode* final_state = stop();
1879
1880 // Find all the needed outputs of this call
1881 CallProjections callprojs;
1882 call->extract_projections(&callprojs, true);
1883
1884 Node* init_mem = call->in(TypeFunc::Memory);
1885 Node* final_mem = final_state->in(TypeFunc::Memory);
1886 Node* final_ctl = final_state->in(TypeFunc::Control);
1887 Node* final_io = final_state->in(TypeFunc::I_O);
1888
1889 // Replace all the old call edges with the edges from the inlining result
1890 if (callprojs.fallthrough_catchproj != NULL) {
1891 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1892 }
1893 if (callprojs.fallthrough_memproj != NULL) {
1894 if (final_mem->is_MergeMem()) {
1895 // Parser's exits MergeMem was not transformed but may be optimized
1896 final_mem = _gvn.transform(final_mem);
1897 }
1898 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1899 }
1900 if (callprojs.fallthrough_ioproj != NULL) {
1901 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1902 }
1903
1904 // Replace the result with the new result if it exists and is used
1905 if (callprojs.resproj != NULL && result != NULL) {
1906 C->gvn_replace_by(callprojs.resproj, result);
1907 }
1908
1909 if (ejvms == NULL) {
1910 // No exception edges to simply kill off those paths
1911 if (callprojs.catchall_catchproj != NULL) {
1912 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1913 }
1914 if (callprojs.catchall_memproj != NULL) {
1915 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1916 }
1917 if (callprojs.catchall_ioproj != NULL) {
1918 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1919 }
1920 // Replace the old exception object with top
1921 if (callprojs.exobj != NULL) {
1922 C->gvn_replace_by(callprojs.exobj, C->top());
1923 }
1924 } else {
1925 GraphKit ekit(ejvms);
1926
1927 // Load my combined exception state into the kit, with all phis transformed:
1928 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1929 replaced_nodes_exception = ex_map->replaced_nodes();
1930
1931 Node* ex_oop = ekit.use_exception_state(ex_map);
1932
1933 if (callprojs.catchall_catchproj != NULL) {
1934 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1935 ex_ctl = ekit.control();
1936 }
1937 if (callprojs.catchall_memproj != NULL) {
1938 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1939 }
1940 if (callprojs.catchall_ioproj != NULL) {
1941 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1942 }
1943
1944 // Replace the old exception object with the newly created one
1945 if (callprojs.exobj != NULL) {
1946 C->gvn_replace_by(callprojs.exobj, ex_oop);
1947 }
1948 }
1949
1950 // Disconnect the call from the graph
1951 call->disconnect_inputs(NULL, C);
1952 C->gvn_replace_by(call, C->top());
1953
1954 // Clean up any MergeMems that feed other MergeMems since the
1955 // optimizer doesn't like that.
1956 if (final_mem->is_MergeMem()) {
1957 Node_List wl;
1958 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1959 Node* m = i.get();
1960 if (m->is_MergeMem() && !wl.contains(m)) {
1961 wl.push(m);
1962 }
1963 }
1964 while (wl.size() > 0) {
1965 _gvn.transform(wl.pop());
1966 }
1967 }
1968
1969 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1970 replaced_nodes.apply(C, final_ctl);
1971 }
1972 if (!ex_ctl->is_top() && do_replaced_nodes) {
1973 replaced_nodes_exception.apply(C, ex_ctl);
1974 }
1975 }
1976
1977
1978 //------------------------------increment_counter------------------------------
1979 // for statistics: increment a VM counter by 1
1980
1981 void GraphKit::increment_counter(address counter_addr) {
1982 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1983 increment_counter(adr1);
1984 }
1985
1986 void GraphKit::increment_counter(Node* counter_addr) {
1987 int adr_type = Compile::AliasIdxRaw;
1988 Node* ctrl = control();
1989 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1990 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1991 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1992 }
1993
1994
1995 //------------------------------uncommon_trap----------------------------------
1996 // Bail out to the interpreter in mid-method. Implemented by calling the
1997 // uncommon_trap blob. This helper function inserts a runtime call with the
1998 // right debug info.
1999 void GraphKit::uncommon_trap(int trap_request,
2000 ciKlass* klass, const char* comment,
2001 bool must_throw,
2002 bool keep_exact_action) {
2003 if (failing()) stop();
2004 if (stopped()) return; // trap reachable?
2005
2006 // Note: If ProfileTraps is true, and if a deopt. actually
2007 // occurs here, the runtime will make sure an MDO exists. There is
2008 // no need to call method()->ensure_method_data() at this point.
2009
2010 // Set the stack pointer to the right value for reexecution:
2011 set_sp(reexecute_sp());
2012
2013 #ifdef ASSERT
2014 if (!must_throw) {
2015 // Make sure the stack has at least enough depth to execute
2016 // the current bytecode.
2017 int inputs, ignored_depth;
2018 if (compute_stack_effects(inputs, ignored_depth)) {
2019 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2020 Bytecodes::name(java_bc()), sp(), inputs);
2021 }
2022 }
2023 #endif
2024
2025 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2026 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2027
2028 switch (action) {
2029 case Deoptimization::Action_maybe_recompile:
2030 case Deoptimization::Action_reinterpret:
2031 // Temporary fix for 6529811 to allow virtual calls to be sure they
2032 // get the chance to go from mono->bi->mega
2033 if (!keep_exact_action &&
2034 Deoptimization::trap_request_index(trap_request) < 0 &&
2035 too_many_recompiles(reason)) {
2036 // This BCI is causing too many recompilations.
2037 if (C->log() != NULL) {
2038 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2039 Deoptimization::trap_reason_name(reason),
2040 Deoptimization::trap_action_name(action));
2041 }
2042 action = Deoptimization::Action_none;
2043 trap_request = Deoptimization::make_trap_request(reason, action);
2044 } else {
2045 C->set_trap_can_recompile(true);
2046 }
2047 break;
2048 case Deoptimization::Action_make_not_entrant:
2049 C->set_trap_can_recompile(true);
2050 break;
2051 case Deoptimization::Action_none:
2052 case Deoptimization::Action_make_not_compilable:
2053 break;
2054 default:
2055 #ifdef ASSERT
2056 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2057 #endif
2058 break;
2059 }
2060
2061 if (TraceOptoParse) {
2062 char buf[100];
2063 tty->print_cr("Uncommon trap %s at bci:%d",
2064 Deoptimization::format_trap_request(buf, sizeof(buf),
2065 trap_request), bci());
2066 }
2067
2068 CompileLog* log = C->log();
2069 if (log != NULL) {
2070 int kid = (klass == NULL)? -1: log->identify(klass);
2071 log->begin_elem("uncommon_trap bci='%d'", bci());
2072 char buf[100];
2073 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2074 trap_request));
2075 if (kid >= 0) log->print(" klass='%d'", kid);
2076 if (comment != NULL) log->print(" comment='%s'", comment);
2077 log->end_elem();
2078 }
2079
2080 // Make sure any guarding test views this path as very unlikely
2081 Node *i0 = control()->in(0);
2082 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2083 IfNode *iff = i0->as_If();
2084 float f = iff->_prob; // Get prob
2085 if (control()->Opcode() == Op_IfTrue) {
2086 if (f > PROB_UNLIKELY_MAG(4))
2087 iff->_prob = PROB_MIN;
2088 } else {
2089 if (f < PROB_LIKELY_MAG(4))
2090 iff->_prob = PROB_MAX;
2091 }
2092 }
2093
2094 // Clear out dead values from the debug info.
2095 kill_dead_locals();
2096
2097 // Now insert the uncommon trap subroutine call
2098 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2099 const TypePtr* no_memory_effects = NULL;
2100 // Pass the index of the class to be loaded
2101 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2102 (must_throw ? RC_MUST_THROW : 0),
2103 OptoRuntime::uncommon_trap_Type(),
2104 call_addr, "uncommon_trap", no_memory_effects,
2105 intcon(trap_request));
2106 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2107 "must extract request correctly from the graph");
2108 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2109
2110 call->set_req(TypeFunc::ReturnAdr, returnadr());
2111 // The debug info is the only real input to this call.
2112
2113 // Halt-and-catch fire here. The above call should never return!
2114 HaltNode* halt = new HaltNode(control(), frameptr());
2115 _gvn.set_type_bottom(halt);
2116 root()->add_req(halt);
2117
2118 stop_and_kill_map();
2119 }
2120
2121
2122 //--------------------------just_allocated_object------------------------------
2123 // Report the object that was just allocated.
2124 // It must be the case that there are no intervening safepoints.
2125 // We use this to determine if an object is so "fresh" that
2126 // it does not require card marks.
2127 Node* GraphKit::just_allocated_object(Node* current_control) {
2128 if (C->recent_alloc_ctl() == current_control)
2129 return C->recent_alloc_obj();
2130 return NULL;
2131 }
2132
2133
2134 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2135 // (Note: TypeFunc::make has a cache that makes this fast.)
2136 const TypeFunc* tf = TypeFunc::make(dest_method);
2137 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2138 for (int j = 0; j < nargs; j++) {
2139 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2140 if( targ->basic_type() == T_DOUBLE ) {
2141 // If any parameters are doubles, they must be rounded before
2142 // the call, dstore_rounding does gvn.transform
2143 Node *arg = argument(j);
2144 arg = dstore_rounding(arg);
2145 set_argument(j, arg);
2146 }
2147 }
2148 }
2149
2150 /**
2151 * Record profiling data exact_kls for Node n with the type system so
2152 * that it can propagate it (speculation)
2153 *
2154 * @param n node that the type applies to
2155 * @param exact_kls type from profiling
2156 * @param maybe_null did profiling see null?
2157 *
2158 * @return node with improved type
2159 */
2160 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2161 const Type* current_type = _gvn.type(n);
2162 assert(UseTypeSpeculation, "type speculation must be on");
2163
2164 const TypePtr* speculative = current_type->speculative();
2165
2166 // Should the klass from the profile be recorded in the speculative type?
2167 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2168 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2169 const TypeOopPtr* xtype = tklass->as_instance_type();
2170 assert(xtype->klass_is_exact(), "Should be exact");
2171 // Any reason to believe n is not null (from this profiling or a previous one)?
2172 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2173 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2174 // record the new speculative type's depth
2175 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2176 speculative = speculative->with_inline_depth(jvms()->depth());
2177 } else if (current_type->would_improve_ptr(ptr_kind)) {
2178 // Profiling report that null was never seen so we can change the
2179 // speculative type to non null ptr.
2180 if (ptr_kind == ProfileAlwaysNull) {
2181 speculative = TypePtr::NULL_PTR;
2182 } else {
2183 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2184 const TypePtr* ptr = TypePtr::NOTNULL;
2185 if (speculative != NULL) {
2186 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2187 } else {
2188 speculative = ptr;
2189 }
2190 }
2191 }
2192
2193 if (speculative != current_type->speculative()) {
2194 // Build a type with a speculative type (what we think we know
2195 // about the type but will need a guard when we use it)
2196 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2197 // We're changing the type, we need a new CheckCast node to carry
2198 // the new type. The new type depends on the control: what
2199 // profiling tells us is only valid from here as far as we can
2200 // tell.
2201 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2202 cast = _gvn.transform(cast);
2203 replace_in_map(n, cast);
2204 n = cast;
2205 }
2206
2207 return n;
2208 }
2209
2210 /**
2211 * Record profiling data from receiver profiling at an invoke with the
2212 * type system so that it can propagate it (speculation)
2213 *
2214 * @param n receiver node
2215 *
2216 * @return node with improved type
2217 */
2218 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2219 if (!UseTypeSpeculation) {
2220 return n;
2221 }
2222 ciKlass* exact_kls = profile_has_unique_klass();
2223 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2224 if ((java_bc() == Bytecodes::_checkcast ||
2225 java_bc() == Bytecodes::_instanceof ||
2226 java_bc() == Bytecodes::_aastore) &&
2227 method()->method_data()->is_mature()) {
2228 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2229 if (data != NULL) {
2230 if (!data->as_BitData()->null_seen()) {
2231 ptr_kind = ProfileNeverNull;
2232 } else {
2233 assert(data->is_ReceiverTypeData(), "bad profile data type");
2234 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2235 uint i = 0;
2236 for (; i < call->row_limit(); i++) {
2237 ciKlass* receiver = call->receiver(i);
2238 if (receiver != NULL) {
2239 break;
2240 }
2241 }
2242 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2243 }
2244 }
2245 }
2246 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2247 }
2248
2249 /**
2250 * Record profiling data from argument profiling at an invoke with the
2251 * type system so that it can propagate it (speculation)
2252 *
2253 * @param dest_method target method for the call
2254 * @param bc what invoke bytecode is this?
2255 */
2256 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2257 if (!UseTypeSpeculation) {
2258 return;
2259 }
2260 const TypeFunc* tf = TypeFunc::make(dest_method);
2261 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2262 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2263 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2264 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2265 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2266 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2267 ciKlass* better_type = NULL;
2268 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2269 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2270 }
2271 i++;
2272 }
2273 }
2274 }
2275
2276 /**
2277 * Record profiling data from parameter profiling at an invoke with
2278 * the type system so that it can propagate it (speculation)
2279 */
2280 void GraphKit::record_profiled_parameters_for_speculation() {
2281 if (!UseTypeSpeculation) {
2282 return;
2283 }
2284 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2285 if (_gvn.type(local(i))->isa_oopptr()) {
2286 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2287 ciKlass* better_type = NULL;
2288 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2289 record_profile_for_speculation(local(i), better_type, ptr_kind);
2290 }
2291 j++;
2292 }
2293 }
2294 }
2295
2296 /**
2297 * Record profiling data from return value profiling at an invoke with
2298 * the type system so that it can propagate it (speculation)
2299 */
2300 void GraphKit::record_profiled_return_for_speculation() {
2301 if (!UseTypeSpeculation) {
2302 return;
2303 }
2304 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2305 ciKlass* better_type = NULL;
2306 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2307 // If profiling reports a single type for the return value,
2308 // feed it to the type system so it can propagate it as a
2309 // speculative type
2310 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2311 }
2312 }
2313
2314 void GraphKit::round_double_result(ciMethod* dest_method) {
2315 // A non-strict method may return a double value which has an extended
2316 // exponent, but this must not be visible in a caller which is 'strict'
2317 // If a strict caller invokes a non-strict callee, round a double result
2318
2319 BasicType result_type = dest_method->return_type()->basic_type();
2320 assert( method() != NULL, "must have caller context");
2321 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2322 // Destination method's return value is on top of stack
2323 // dstore_rounding() does gvn.transform
2324 Node *result = pop_pair();
2325 result = dstore_rounding(result);
2326 push_pair(result);
2327 }
2328 }
2329
2330 // rounding for strict float precision conformance
2331 Node* GraphKit::precision_rounding(Node* n) {
2332 return UseStrictFP && _method->flags().is_strict()
2333 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2334 ? _gvn.transform( new RoundFloatNode(0, n) )
2335 : n;
2336 }
2337
2338 // rounding for strict double precision conformance
2339 Node* GraphKit::dprecision_rounding(Node *n) {
2340 return UseStrictFP && _method->flags().is_strict()
2341 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2342 ? _gvn.transform( new RoundDoubleNode(0, n) )
2343 : n;
2344 }
2345
2346 // rounding for non-strict double stores
2347 Node* GraphKit::dstore_rounding(Node* n) {
2348 return Matcher::strict_fp_requires_explicit_rounding
2349 && UseSSE <= 1
2350 ? _gvn.transform( new RoundDoubleNode(0, n) )
2351 : n;
2352 }
2353
2354 //=============================================================================
2355 // Generate a fast path/slow path idiom. Graph looks like:
2356 // [foo] indicates that 'foo' is a parameter
2357 //
2358 // [in] NULL
2359 // \ /
2360 // CmpP
2361 // Bool ne
2362 // If
2363 // / \
2364 // True False-<2>
2365 // / |
2366 // / cast_not_null
2367 // Load | | ^
2368 // [fast_test] | |
2369 // gvn to opt_test | |
2370 // / \ | <1>
2371 // True False |
2372 // | \\ |
2373 // [slow_call] \[fast_result]
2374 // Ctl Val \ \
2375 // | \ \
2376 // Catch <1> \ \
2377 // / \ ^ \ \
2378 // Ex No_Ex | \ \
2379 // | \ \ | \ <2> \
2380 // ... \ [slow_res] | | \ [null_result]
2381 // \ \--+--+--- | |
2382 // \ | / \ | /
2383 // --------Region Phi
2384 //
2385 //=============================================================================
2386 // Code is structured as a series of driver functions all called 'do_XXX' that
2387 // call a set of helper functions. Helper functions first, then drivers.
2388
2389 //------------------------------null_check_oop---------------------------------
2390 // Null check oop. Set null-path control into Region in slot 3.
2391 // Make a cast-not-nullness use the other not-null control. Return cast.
2392 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2393 bool never_see_null,
2394 bool safe_for_replace,
2395 bool speculative) {
2396 // Initial NULL check taken path
2397 (*null_control) = top();
2398 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2399
2400 // Generate uncommon_trap:
2401 if (never_see_null && (*null_control) != top()) {
2402 // If we see an unexpected null at a check-cast we record it and force a
2403 // recompile; the offending check-cast will be compiled to handle NULLs.
2404 // If we see more than one offending BCI, then all checkcasts in the
2405 // method will be compiled to handle NULLs.
2406 PreserveJVMState pjvms(this);
2407 set_control(*null_control);
2408 replace_in_map(value, null());
2409 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2410 uncommon_trap(reason,
2411 Deoptimization::Action_make_not_entrant);
2412 (*null_control) = top(); // NULL path is dead
2413 }
2414 if ((*null_control) == top() && safe_for_replace) {
2415 replace_in_map(value, cast);
2416 }
2417
2418 // Cast away null-ness on the result
2419 return cast;
2420 }
2421
2422 //------------------------------opt_iff----------------------------------------
2423 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2424 // Return slow-path control.
2425 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2426 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2427
2428 // Fast path taken; set region slot 2
2429 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2430 region->init_req(2,fast_taken); // Capture fast-control
2431
2432 // Fast path not-taken, i.e. slow path
2433 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2434 return slow_taken;
2435 }
2436
2437 //-----------------------------make_runtime_call-------------------------------
2438 Node* GraphKit::make_runtime_call(int flags,
2439 const TypeFunc* call_type, address call_addr,
2440 const char* call_name,
2441 const TypePtr* adr_type,
2442 // The following parms are all optional.
2443 // The first NULL ends the list.
2444 Node* parm0, Node* parm1,
2445 Node* parm2, Node* parm3,
2446 Node* parm4, Node* parm5,
2447 Node* parm6, Node* parm7) {
2448 // Slow-path call
2449 bool is_leaf = !(flags & RC_NO_LEAF);
2450 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2451 if (call_name == NULL) {
2452 assert(!is_leaf, "must supply name for leaf");
2453 call_name = OptoRuntime::stub_name(call_addr);
2454 }
2455 CallNode* call;
2456 if (!is_leaf) {
2457 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2458 bci(), adr_type);
2459 } else if (flags & RC_NO_FP) {
2460 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2461 } else {
2462 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2463 }
2464
2465 // The following is similar to set_edges_for_java_call,
2466 // except that the memory effects of the call are restricted to AliasIdxRaw.
2467
2468 // Slow path call has no side-effects, uses few values
2469 bool wide_in = !(flags & RC_NARROW_MEM);
2470 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2471
2472 Node* prev_mem = NULL;
2473 if (wide_in) {
2474 prev_mem = set_predefined_input_for_runtime_call(call);
2475 } else {
2476 assert(!wide_out, "narrow in => narrow out");
2477 Node* narrow_mem = memory(adr_type);
2478 prev_mem = reset_memory();
2479 map()->set_memory(narrow_mem);
2480 set_predefined_input_for_runtime_call(call);
2481 }
2482
2483 // Hook each parm in order. Stop looking at the first NULL.
2484 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2485 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2486 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2487 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2488 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2489 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2490 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2491 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2492 /* close each nested if ===> */ } } } } } } } }
2493 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2494
2495 if (!is_leaf) {
2496 // Non-leaves can block and take safepoints:
2497 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2498 }
2499 // Non-leaves can throw exceptions:
2500 if (has_io) {
2501 call->set_req(TypeFunc::I_O, i_o());
2502 }
2503
2504 if (flags & RC_UNCOMMON) {
2505 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2506 // (An "if" probability corresponds roughly to an unconditional count.
2507 // Sort of.)
2508 call->set_cnt(PROB_UNLIKELY_MAG(4));
2509 }
2510
2511 Node* c = _gvn.transform(call);
2512 assert(c == call, "cannot disappear");
2513
2514 if (wide_out) {
2515 // Slow path call has full side-effects.
2516 set_predefined_output_for_runtime_call(call);
2517 } else {
2518 // Slow path call has few side-effects, and/or sets few values.
2519 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2520 }
2521
2522 if (has_io) {
2523 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2524 }
2525 return call;
2526
2527 }
2528
2529 //------------------------------merge_memory-----------------------------------
2530 // Merge memory from one path into the current memory state.
2531 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2532 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2533 Node* old_slice = mms.force_memory();
2534 Node* new_slice = mms.memory2();
2535 if (old_slice != new_slice) {
2536 PhiNode* phi;
2537 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2538 if (mms.is_empty()) {
2539 // clone base memory Phi's inputs for this memory slice
2540 assert(old_slice == mms.base_memory(), "sanity");
2541 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2542 _gvn.set_type(phi, Type::MEMORY);
2543 for (uint i = 1; i < phi->req(); i++) {
2544 phi->init_req(i, old_slice->in(i));
2545 }
2546 } else {
2547 phi = old_slice->as_Phi(); // Phi was generated already
2548 }
2549 } else {
2550 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2551 _gvn.set_type(phi, Type::MEMORY);
2552 }
2553 phi->set_req(new_path, new_slice);
2554 mms.set_memory(phi);
2555 }
2556 }
2557 }
2558
2559 //------------------------------make_slow_call_ex------------------------------
2560 // Make the exception handler hookups for the slow call
2561 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2562 if (stopped()) return;
2563
2564 // Make a catch node with just two handlers: fall-through and catch-all
2565 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2566 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2567 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2568 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2569
2570 { PreserveJVMState pjvms(this);
2571 set_control(excp);
2572 set_i_o(i_o);
2573
2574 if (excp != top()) {
2575 if (deoptimize) {
2576 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2577 uncommon_trap(Deoptimization::Reason_unhandled,
2578 Deoptimization::Action_none);
2579 } else {
2580 // Create an exception state also.
2581 // Use an exact type if the caller has specified a specific exception.
2582 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2583 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2584 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2585 }
2586 }
2587 }
2588
2589 // Get the no-exception control from the CatchNode.
2590 set_control(norm);
2591 }
2592
2593 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2594 Node* cmp = NULL;
2595 switch(bt) {
2596 case T_INT: cmp = new CmpINode(in1, in2); break;
2597 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2598 default: fatal("unexpected comparison type %s", type2name(bt));
2599 }
2600 gvn->transform(cmp);
2601 Node* bol = gvn->transform(new BoolNode(cmp, test));
2602 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2603 gvn->transform(iff);
2604 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2605 return iff;
2606 }
2607
2608
2609 //-------------------------------gen_subtype_check-----------------------------
2610 // Generate a subtyping check. Takes as input the subtype and supertype.
2611 // Returns 2 values: sets the default control() to the true path and returns
2612 // the false path. Only reads invariant memory; sets no (visible) memory.
2613 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2614 // but that's not exposed to the optimizer. This call also doesn't take in an
2615 // Object; if you wish to check an Object you need to load the Object's class
2616 // prior to coming here.
2617 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2618 Compile* C = gvn->C;
2619
2620 if ((*ctrl)->is_top()) {
2621 return C->top();
2622 }
2623
2624 // Fast check for identical types, perhaps identical constants.
2625 // The types can even be identical non-constants, in cases
2626 // involving Array.newInstance, Object.clone, etc.
2627 if (subklass == superklass)
2628 return C->top(); // false path is dead; no test needed.
2629
2630 if (gvn->type(superklass)->singleton()) {
2631 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2632 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2633
2634 // In the common case of an exact superklass, try to fold up the
2635 // test before generating code. You may ask, why not just generate
2636 // the code and then let it fold up? The answer is that the generated
2637 // code will necessarily include null checks, which do not always
2638 // completely fold away. If they are also needless, then they turn
2639 // into a performance loss. Example:
2640 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2641 // Here, the type of 'fa' is often exact, so the store check
2642 // of fa[1]=x will fold up, without testing the nullness of x.
2643 switch (C->static_subtype_check(superk, subk)) {
2644 case Compile::SSC_always_false:
2645 {
2646 Node* always_fail = *ctrl;
2647 *ctrl = gvn->C->top();
2648 return always_fail;
2649 }
2650 case Compile::SSC_always_true:
2651 return C->top();
2652 case Compile::SSC_easy_test:
2653 {
2654 // Just do a direct pointer compare and be done.
2655 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2656 *ctrl = gvn->transform(new IfTrueNode(iff));
2657 return gvn->transform(new IfFalseNode(iff));
2658 }
2659 case Compile::SSC_full_test:
2660 break;
2661 default:
2662 ShouldNotReachHere();
2663 }
2664 }
2665
2666 // %%% Possible further optimization: Even if the superklass is not exact,
2667 // if the subklass is the unique subtype of the superklass, the check
2668 // will always succeed. We could leave a dependency behind to ensure this.
2669
2670 // First load the super-klass's check-offset
2671 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2672 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2673 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2674 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2675 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2676
2677 // Load from the sub-klass's super-class display list, or a 1-word cache of
2678 // the secondary superclass list, or a failing value with a sentinel offset
2679 // if the super-klass is an interface or exceptionally deep in the Java
2680 // hierarchy and we have to scan the secondary superclass list the hard way.
2681 // Worst-case type is a little odd: NULL is allowed as a result (usually
2682 // klass loads can never produce a NULL).
2683 Node *chk_off_X = chk_off;
2684 #ifdef _LP64
2685 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2686 #endif
2687 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2688 // For some types like interfaces the following loadKlass is from a 1-word
2689 // cache which is mutable so can't use immutable memory. Other
2690 // types load from the super-class display table which is immutable.
2691 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2692 Node *kmem = might_be_cache ? m : C->immutable_memory();
2693 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2694
2695 // Compile speed common case: ARE a subtype and we canNOT fail
2696 if( superklass == nkls )
2697 return C->top(); // false path is dead; no test needed.
2698
2699 // See if we get an immediate positive hit. Happens roughly 83% of the
2700 // time. Test to see if the value loaded just previously from the subklass
2701 // is exactly the superklass.
2702 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2703 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2704 *ctrl = gvn->transform(new IfFalseNode(iff1));
2705
2706 // Compile speed common case: Check for being deterministic right now. If
2707 // chk_off is a constant and not equal to cacheoff then we are NOT a
2708 // subklass. In this case we need exactly the 1 test above and we can
2709 // return those results immediately.
2710 if (!might_be_cache) {
2711 Node* not_subtype_ctrl = *ctrl;
2712 *ctrl = iftrue1; // We need exactly the 1 test above
2713 return not_subtype_ctrl;
2714 }
2715
2716 // Gather the various success & failures here
2717 RegionNode *r_ok_subtype = new RegionNode(4);
2718 gvn->record_for_igvn(r_ok_subtype);
2719 RegionNode *r_not_subtype = new RegionNode(3);
2720 gvn->record_for_igvn(r_not_subtype);
2721
2722 r_ok_subtype->init_req(1, iftrue1);
2723
2724 // Check for immediate negative hit. Happens roughly 11% of the time (which
2725 // is roughly 63% of the remaining cases). Test to see if the loaded
2726 // check-offset points into the subklass display list or the 1-element
2727 // cache. If it points to the display (and NOT the cache) and the display
2728 // missed then it's not a subtype.
2729 Node *cacheoff = gvn->intcon(cacheoff_con);
2730 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2731 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2732 *ctrl = gvn->transform(new IfFalseNode(iff2));
2733
2734 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2735 // No performance impact (too rare) but allows sharing of secondary arrays
2736 // which has some footprint reduction.
2737 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2738 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2739 *ctrl = gvn->transform(new IfFalseNode(iff3));
2740
2741 // -- Roads not taken here: --
2742 // We could also have chosen to perform the self-check at the beginning
2743 // of this code sequence, as the assembler does. This would not pay off
2744 // the same way, since the optimizer, unlike the assembler, can perform
2745 // static type analysis to fold away many successful self-checks.
2746 // Non-foldable self checks work better here in second position, because
2747 // the initial primary superclass check subsumes a self-check for most
2748 // types. An exception would be a secondary type like array-of-interface,
2749 // which does not appear in its own primary supertype display.
2750 // Finally, we could have chosen to move the self-check into the
2751 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2752 // dependent manner. But it is worthwhile to have the check here,
2753 // where it can be perhaps be optimized. The cost in code space is
2754 // small (register compare, branch).
2755
2756 // Now do a linear scan of the secondary super-klass array. Again, no real
2757 // performance impact (too rare) but it's gotta be done.
2758 // Since the code is rarely used, there is no penalty for moving it
2759 // out of line, and it can only improve I-cache density.
2760 // The decision to inline or out-of-line this final check is platform
2761 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2762 Node* psc = gvn->transform(
2763 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2764
2765 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2766 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2767 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2768
2769 // Return false path; set default control to true path.
2770 *ctrl = gvn->transform(r_ok_subtype);
2771 return gvn->transform(r_not_subtype);
2772 }
2773
2774 // Profile-driven exact type check:
2775 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2776 float prob,
2777 Node* *casted_receiver) {
2778 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2779 Node* recv_klass = load_object_klass(receiver);
2780 Node* want_klass = makecon(tklass);
2781 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2782 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2783 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2784 set_control( _gvn.transform( new IfTrueNode (iff) ));
2785 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2786
2787 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2788 assert(recv_xtype->klass_is_exact(), "");
2789
2790 // Subsume downstream occurrences of receiver with a cast to
2791 // recv_xtype, since now we know what the type will be.
2792 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2793 (*casted_receiver) = _gvn.transform(cast);
2794 // (User must make the replace_in_map call.)
2795
2796 return fail;
2797 }
2798
2799
2800 //------------------------------seems_never_null-------------------------------
2801 // Use null_seen information if it is available from the profile.
2802 // If we see an unexpected null at a type check we record it and force a
2803 // recompile; the offending check will be recompiled to handle NULLs.
2804 // If we see several offending BCIs, then all checks in the
2805 // method will be recompiled.
2806 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2807 speculating = !_gvn.type(obj)->speculative_maybe_null();
2808 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2809 if (UncommonNullCast // Cutout for this technique
2810 && obj != null() // And not the -Xcomp stupid case?
2811 && !too_many_traps(reason)
2812 ) {
2813 if (speculating) {
2814 return true;
2815 }
2816 if (data == NULL)
2817 // Edge case: no mature data. Be optimistic here.
2818 return true;
2819 // If the profile has not seen a null, assume it won't happen.
2820 assert(java_bc() == Bytecodes::_checkcast ||
2821 java_bc() == Bytecodes::_instanceof ||
2822 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2823 return !data->as_BitData()->null_seen();
2824 }
2825 speculating = false;
2826 return false;
2827 }
2828
2829 //------------------------maybe_cast_profiled_receiver-------------------------
2830 // If the profile has seen exactly one type, narrow to exactly that type.
2831 // Subsequent type checks will always fold up.
2832 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2833 ciKlass* require_klass,
2834 ciKlass* spec_klass,
2835 bool safe_for_replace) {
2836 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2837
2838 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2839
2840 // Make sure we haven't already deoptimized from this tactic.
2841 if (too_many_traps(reason) || too_many_recompiles(reason))
2842 return NULL;
2843
2844 // (No, this isn't a call, but it's enough like a virtual call
2845 // to use the same ciMethod accessor to get the profile info...)
2846 // If we have a speculative type use it instead of profiling (which
2847 // may not help us)
2848 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2849 if (exact_kls != NULL) {// no cast failures here
2850 if (require_klass == NULL ||
2851 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2852 // If we narrow the type to match what the type profile sees or
2853 // the speculative type, we can then remove the rest of the
2854 // cast.
2855 // This is a win, even if the exact_kls is very specific,
2856 // because downstream operations, such as method calls,
2857 // will often benefit from the sharper type.
2858 Node* exact_obj = not_null_obj; // will get updated in place...
2859 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2860 &exact_obj);
2861 { PreserveJVMState pjvms(this);
2862 set_control(slow_ctl);
2863 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2864 }
2865 if (safe_for_replace) {
2866 replace_in_map(not_null_obj, exact_obj);
2867 }
2868 return exact_obj;
2869 }
2870 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2871 }
2872
2873 return NULL;
2874 }
2875
2876 /**
2877 * Cast obj to type and emit guard unless we had too many traps here
2878 * already
2879 *
2880 * @param obj node being casted
2881 * @param type type to cast the node to
2882 * @param not_null true if we know node cannot be null
2883 */
2884 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2885 ciKlass* type,
2886 bool not_null) {
2887 if (stopped()) {
2888 return obj;
2889 }
2890
2891 // type == NULL if profiling tells us this object is always null
2892 if (type != NULL) {
2893 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2894 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2895
2896 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2897 !too_many_traps(class_reason) &&
2898 !too_many_recompiles(class_reason)) {
2899 Node* not_null_obj = NULL;
2900 // not_null is true if we know the object is not null and
2901 // there's no need for a null check
2902 if (!not_null) {
2903 Node* null_ctl = top();
2904 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2905 assert(null_ctl->is_top(), "no null control here");
2906 } else {
2907 not_null_obj = obj;
2908 }
2909
2910 Node* exact_obj = not_null_obj;
2911 ciKlass* exact_kls = type;
2912 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2913 &exact_obj);
2914 {
2915 PreserveJVMState pjvms(this);
2916 set_control(slow_ctl);
2917 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2918 }
2919 replace_in_map(not_null_obj, exact_obj);
2920 obj = exact_obj;
2921 }
2922 } else {
2923 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2924 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2925 Node* exact_obj = null_assert(obj);
2926 replace_in_map(obj, exact_obj);
2927 obj = exact_obj;
2928 }
2929 }
2930 return obj;
2931 }
2932
2933 //-------------------------------gen_instanceof--------------------------------
2934 // Generate an instance-of idiom. Used by both the instance-of bytecode
2935 // and the reflective instance-of call.
2936 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2937 kill_dead_locals(); // Benefit all the uncommon traps
2938 assert( !stopped(), "dead parse path should be checked in callers" );
2939 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2940 "must check for not-null not-dead klass in callers");
2941
2942 // Make the merge point
2943 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2944 RegionNode* region = new RegionNode(PATH_LIMIT);
2945 Node* phi = new PhiNode(region, TypeInt::BOOL);
2946 C->set_has_split_ifs(true); // Has chance for split-if optimization
2947
2948 ciProfileData* data = NULL;
2949 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2950 data = method()->method_data()->bci_to_data(bci());
2951 }
2952 bool speculative_not_null = false;
2953 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2954 && seems_never_null(obj, data, speculative_not_null));
2955
2956 // Null check; get casted pointer; set region slot 3
2957 Node* null_ctl = top();
2958 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2959
2960 // If not_null_obj is dead, only null-path is taken
2961 if (stopped()) { // Doing instance-of on a NULL?
2962 set_control(null_ctl);
2963 return intcon(0);
2964 }
2965 region->init_req(_null_path, null_ctl);
2966 phi ->init_req(_null_path, intcon(0)); // Set null path value
2967 if (null_ctl == top()) {
2968 // Do this eagerly, so that pattern matches like is_diamond_phi
2969 // will work even during parsing.
2970 assert(_null_path == PATH_LIMIT-1, "delete last");
2971 region->del_req(_null_path);
2972 phi ->del_req(_null_path);
2973 }
2974
2975 // Do we know the type check always succeed?
2976 bool known_statically = false;
2977 if (_gvn.type(superklass)->singleton()) {
2978 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2979 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2980 if (subk != NULL && subk->is_loaded()) {
2981 int static_res = C->static_subtype_check(superk, subk);
2982 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2983 }
2984 }
2985
2986 if (!known_statically) {
2987 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2988 // We may not have profiling here or it may not help us. If we
2989 // have a speculative type use it to perform an exact cast.
2990 ciKlass* spec_obj_type = obj_type->speculative_type();
2991 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2992 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2993 if (stopped()) { // Profile disagrees with this path.
2994 set_control(null_ctl); // Null is the only remaining possibility.
2995 return intcon(0);
2996 }
2997 if (cast_obj != NULL) {
2998 not_null_obj = cast_obj;
2999 }
3000 }
3001 }
3002
3003 // Load the object's klass
3004 Node* obj_klass = load_object_klass(not_null_obj);
3005
3006 // Generate the subtype check
3007 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3008
3009 // Plug in the success path to the general merge in slot 1.
3010 region->init_req(_obj_path, control());
3011 phi ->init_req(_obj_path, intcon(1));
3012
3013 // Plug in the failing path to the general merge in slot 2.
3014 region->init_req(_fail_path, not_subtype_ctrl);
3015 phi ->init_req(_fail_path, intcon(0));
3016
3017 // Return final merged results
3018 set_control( _gvn.transform(region) );
3019 record_for_igvn(region);
3020
3021 // If we know the type check always succeeds then we don't use the
3022 // profiling data at this bytecode. Don't lose it, feed it to the
3023 // type system as a speculative type.
3024 if (safe_for_replace) {
3025 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3026 replace_in_map(obj, casted_obj);
3027 }
3028
3029 return _gvn.transform(phi);
3030 }
3031
3032 //-------------------------------gen_checkcast---------------------------------
3033 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3034 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3035 // uncommon-trap paths work. Adjust stack after this call.
3036 // If failure_control is supplied and not null, it is filled in with
3037 // the control edge for the cast failure. Otherwise, an appropriate
3038 // uncommon trap or exception is thrown.
3039 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3040 Node* *failure_control) {
3041 kill_dead_locals(); // Benefit all the uncommon traps
3042 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3043 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3044
3045 // Fast cutout: Check the case that the cast is vacuously true.
3046 // This detects the common cases where the test will short-circuit
3047 // away completely. We do this before we perform the null check,
3048 // because if the test is going to turn into zero code, we don't
3049 // want a residual null check left around. (Causes a slowdown,
3050 // for example, in some objArray manipulations, such as a[i]=a[j].)
3051 if (tk->singleton()) {
3052 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3053 if (objtp != NULL && objtp->klass() != NULL) {
3054 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3055 case Compile::SSC_always_true:
3056 // If we know the type check always succeed then we don't use
3057 // the profiling data at this bytecode. Don't lose it, feed it
3058 // to the type system as a speculative type.
3059 return record_profiled_receiver_for_speculation(obj);
3060 case Compile::SSC_always_false:
3061 // It needs a null check because a null will *pass* the cast check.
3062 // A non-null value will always produce an exception.
3063 return null_assert(obj);
3064 }
3065 }
3066 }
3067
3068 ciProfileData* data = NULL;
3069 bool safe_for_replace = false;
3070 if (failure_control == NULL) { // use MDO in regular case only
3071 assert(java_bc() == Bytecodes::_aastore ||
3072 java_bc() == Bytecodes::_checkcast,
3073 "interpreter profiles type checks only for these BCs");
3074 data = method()->method_data()->bci_to_data(bci());
3075 safe_for_replace = true;
3076 }
3077
3078 // Make the merge point
3079 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3080 RegionNode* region = new RegionNode(PATH_LIMIT);
3081 Node* phi = new PhiNode(region, toop);
3082 C->set_has_split_ifs(true); // Has chance for split-if optimization
3083
3084 // Use null-cast information if it is available
3085 bool speculative_not_null = false;
3086 bool never_see_null = ((failure_control == NULL) // regular case only
3087 && seems_never_null(obj, data, speculative_not_null));
3088
3089 // Null check; get casted pointer; set region slot 3
3090 Node* null_ctl = top();
3091 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3092
3093 // If not_null_obj is dead, only null-path is taken
3094 if (stopped()) { // Doing instance-of on a NULL?
3095 set_control(null_ctl);
3096 return null();
3097 }
3098 region->init_req(_null_path, null_ctl);
3099 phi ->init_req(_null_path, null()); // Set null path value
3100 if (null_ctl == top()) {
3101 // Do this eagerly, so that pattern matches like is_diamond_phi
3102 // will work even during parsing.
3103 assert(_null_path == PATH_LIMIT-1, "delete last");
3104 region->del_req(_null_path);
3105 phi ->del_req(_null_path);
3106 }
3107
3108 Node* cast_obj = NULL;
3109 if (tk->klass_is_exact()) {
3110 // The following optimization tries to statically cast the speculative type of the object
3111 // (for example obtained during profiling) to the type of the superklass and then do a
3112 // dynamic check that the type of the object is what we expect. To work correctly
3113 // for checkcast and aastore the type of superklass should be exact.
3114 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3115 // We may not have profiling here or it may not help us. If we have
3116 // a speculative type use it to perform an exact cast.
3117 ciKlass* spec_obj_type = obj_type->speculative_type();
3118 if (spec_obj_type != NULL || data != NULL) {
3119 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3120 if (cast_obj != NULL) {
3121 if (failure_control != NULL) // failure is now impossible
3122 (*failure_control) = top();
3123 // adjust the type of the phi to the exact klass:
3124 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3125 }
3126 }
3127 }
3128
3129 if (cast_obj == NULL) {
3130 // Load the object's klass
3131 Node* obj_klass = load_object_klass(not_null_obj);
3132
3133 // Generate the subtype check
3134 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3135
3136 // Plug in success path into the merge
3137 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3138 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3139 if (failure_control == NULL) {
3140 if (not_subtype_ctrl != top()) { // If failure is possible
3141 PreserveJVMState pjvms(this);
3142 set_control(not_subtype_ctrl);
3143 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3144 }
3145 } else {
3146 (*failure_control) = not_subtype_ctrl;
3147 }
3148 }
3149
3150 region->init_req(_obj_path, control());
3151 phi ->init_req(_obj_path, cast_obj);
3152
3153 // A merge of NULL or Casted-NotNull obj
3154 Node* res = _gvn.transform(phi);
3155
3156 // Note I do NOT always 'replace_in_map(obj,result)' here.
3157 // if( tk->klass()->can_be_primary_super() )
3158 // This means that if I successfully store an Object into an array-of-String
3159 // I 'forget' that the Object is really now known to be a String. I have to
3160 // do this because we don't have true union types for interfaces - if I store
3161 // a Baz into an array-of-Interface and then tell the optimizer it's an
3162 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3163 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3164 // replace_in_map( obj, res );
3165
3166 // Return final merged results
3167 set_control( _gvn.transform(region) );
3168 record_for_igvn(region);
3169
3170 return record_profiled_receiver_for_speculation(res);
3171 }
3172
3173 //------------------------------next_monitor-----------------------------------
3174 // What number should be given to the next monitor?
3175 int GraphKit::next_monitor() {
3176 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3177 int next = current + C->sync_stack_slots();
3178 // Keep the toplevel high water mark current:
3179 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3180 return current;
3181 }
3182
3183 //------------------------------insert_mem_bar---------------------------------
3184 // Memory barrier to avoid floating things around
3185 // The membar serves as a pinch point between both control and all memory slices.
3186 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3187 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3188 mb->init_req(TypeFunc::Control, control());
3189 mb->init_req(TypeFunc::Memory, reset_memory());
3190 Node* membar = _gvn.transform(mb);
3191 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3192 set_all_memory_call(membar);
3193 return membar;
3194 }
3195
3196 //-------------------------insert_mem_bar_volatile----------------------------
3197 // Memory barrier to avoid floating things around
3198 // The membar serves as a pinch point between both control and memory(alias_idx).
3199 // If you want to make a pinch point on all memory slices, do not use this
3200 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3201 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3202 // When Parse::do_put_xxx updates a volatile field, it appends a series
3203 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3204 // The first membar is on the same memory slice as the field store opcode.
3205 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3206 // All the other membars (for other volatile slices, including AliasIdxBot,
3207 // which stands for all unknown volatile slices) are control-dependent
3208 // on the first membar. This prevents later volatile loads or stores
3209 // from sliding up past the just-emitted store.
3210
3211 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3212 mb->set_req(TypeFunc::Control,control());
3213 if (alias_idx == Compile::AliasIdxBot) {
3214 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3215 } else {
3216 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3217 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3218 }
3219 Node* membar = _gvn.transform(mb);
3220 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3221 if (alias_idx == Compile::AliasIdxBot) {
3222 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3223 } else {
3224 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3225 }
3226 return membar;
3227 }
3228
3229 //------------------------------shared_lock------------------------------------
3230 // Emit locking code.
3231 FastLockNode* GraphKit::shared_lock(Node* obj) {
3232 // bci is either a monitorenter bc or InvocationEntryBci
3233 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3234 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3235
3236 if( !GenerateSynchronizationCode )
3237 return NULL; // Not locking things?
3238 if (stopped()) // Dead monitor?
3239 return NULL;
3240
3241 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3242
3243 // Box the stack location
3244 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3245 Node* mem = reset_memory();
3246
3247 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3248 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3249 // Create the counters for this fast lock.
3250 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3251 }
3252
3253 // Create the rtm counters for this fast lock if needed.
3254 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3255
3256 // Add monitor to debug info for the slow path. If we block inside the
3257 // slow path and de-opt, we need the monitor hanging around
3258 map()->push_monitor( flock );
3259
3260 const TypeFunc *tf = LockNode::lock_type();
3261 LockNode *lock = new LockNode(C, tf);
3262
3263 lock->init_req( TypeFunc::Control, control() );
3264 lock->init_req( TypeFunc::Memory , mem );
3265 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3266 lock->init_req( TypeFunc::FramePtr, frameptr() );
3267 lock->init_req( TypeFunc::ReturnAdr, top() );
3268
3269 lock->init_req(TypeFunc::Parms + 0, obj);
3270 lock->init_req(TypeFunc::Parms + 1, box);
3271 lock->init_req(TypeFunc::Parms + 2, flock);
3272 add_safepoint_edges(lock);
3273
3274 lock = _gvn.transform( lock )->as_Lock();
3275
3276 // lock has no side-effects, sets few values
3277 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3278
3279 insert_mem_bar(Op_MemBarAcquireLock);
3280
3281 // Add this to the worklist so that the lock can be eliminated
3282 record_for_igvn(lock);
3283
3284 #ifndef PRODUCT
3285 if (PrintLockStatistics) {
3286 // Update the counter for this lock. Don't bother using an atomic
3287 // operation since we don't require absolute accuracy.
3288 lock->create_lock_counter(map()->jvms());
3289 increment_counter(lock->counter()->addr());
3290 }
3291 #endif
3292
3293 return flock;
3294 }
3295
3296
3297 //------------------------------shared_unlock----------------------------------
3298 // Emit unlocking code.
3299 void GraphKit::shared_unlock(Node* box, Node* obj) {
3300 // bci is either a monitorenter bc or InvocationEntryBci
3301 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3302 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3303
3304 if( !GenerateSynchronizationCode )
3305 return;
3306 if (stopped()) { // Dead monitor?
3307 map()->pop_monitor(); // Kill monitor from debug info
3308 return;
3309 }
3310
3311 // Memory barrier to avoid floating things down past the locked region
3312 insert_mem_bar(Op_MemBarReleaseLock);
3313
3314 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3315 UnlockNode *unlock = new UnlockNode(C, tf);
3316 #ifdef ASSERT
3317 unlock->set_dbg_jvms(sync_jvms());
3318 #endif
3319 uint raw_idx = Compile::AliasIdxRaw;
3320 unlock->init_req( TypeFunc::Control, control() );
3321 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3322 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3323 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3324 unlock->init_req( TypeFunc::ReturnAdr, top() );
3325
3326 unlock->init_req(TypeFunc::Parms + 0, obj);
3327 unlock->init_req(TypeFunc::Parms + 1, box);
3328 unlock = _gvn.transform(unlock)->as_Unlock();
3329
3330 Node* mem = reset_memory();
3331
3332 // unlock has no side-effects, sets few values
3333 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3334
3335 // Kill monitor from debug info
3336 map()->pop_monitor( );
3337 }
3338
3339 //-------------------------------get_layout_helper-----------------------------
3340 // If the given klass is a constant or known to be an array,
3341 // fetch the constant layout helper value into constant_value
3342 // and return (Node*)NULL. Otherwise, load the non-constant
3343 // layout helper value, and return the node which represents it.
3344 // This two-faced routine is useful because allocation sites
3345 // almost always feature constant types.
3346 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3347 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3348 if (!StressReflectiveCode && inst_klass != NULL) {
3349 ciKlass* klass = inst_klass->klass();
3350 bool xklass = inst_klass->klass_is_exact();
3351 if (xklass || klass->is_array_klass()) {
3352 jint lhelper = klass->layout_helper();
3353 if (lhelper != Klass::_lh_neutral_value) {
3354 constant_value = lhelper;
3355 return (Node*) NULL;
3356 }
3357 }
3358 }
3359 constant_value = Klass::_lh_neutral_value; // put in a known value
3360 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3361 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3362 }
3363
3364 // We just put in an allocate/initialize with a big raw-memory effect.
3365 // Hook selected additional alias categories on the initialization.
3366 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3367 MergeMemNode* init_in_merge,
3368 Node* init_out_raw) {
3369 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3370 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3371
3372 Node* prevmem = kit.memory(alias_idx);
3373 init_in_merge->set_memory_at(alias_idx, prevmem);
3374 kit.set_memory(init_out_raw, alias_idx);
3375 }
3376
3377 //---------------------------set_output_for_allocation-------------------------
3378 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3379 const TypeOopPtr* oop_type,
3380 bool deoptimize_on_exception) {
3381 int rawidx = Compile::AliasIdxRaw;
3382 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3383 add_safepoint_edges(alloc);
3384 Node* allocx = _gvn.transform(alloc);
3385 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3386 // create memory projection for i_o
3387 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3388 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3389
3390 // create a memory projection as for the normal control path
3391 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3392 set_memory(malloc, rawidx);
3393
3394 // a normal slow-call doesn't change i_o, but an allocation does
3395 // we create a separate i_o projection for the normal control path
3396 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3397 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3398
3399 // put in an initialization barrier
3400 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3401 rawoop)->as_Initialize();
3402 assert(alloc->initialization() == init, "2-way macro link must work");
3403 assert(init ->allocation() == alloc, "2-way macro link must work");
3404 {
3405 // Extract memory strands which may participate in the new object's
3406 // initialization, and source them from the new InitializeNode.
3407 // This will allow us to observe initializations when they occur,
3408 // and link them properly (as a group) to the InitializeNode.
3409 assert(init->in(InitializeNode::Memory) == malloc, "");
3410 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3411 init->set_req(InitializeNode::Memory, minit_in);
3412 record_for_igvn(minit_in); // fold it up later, if possible
3413 Node* minit_out = memory(rawidx);
3414 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3415 if (oop_type->isa_aryptr()) {
3416 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3417 int elemidx = C->get_alias_index(telemref);
3418 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3419 } else if (oop_type->isa_instptr()) {
3420 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3421 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3422 ciField* field = ik->nonstatic_field_at(i);
3423 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3424 continue; // do not bother to track really large numbers of fields
3425 // Find (or create) the alias category for this field:
3426 int fieldidx = C->alias_type(field)->index();
3427 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3428 }
3429 }
3430 }
3431
3432 // Cast raw oop to the real thing...
3433 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3434 javaoop = _gvn.transform(javaoop);
3435 C->set_recent_alloc(control(), javaoop);
3436 assert(just_allocated_object(control()) == javaoop, "just allocated");
3437
3438 #ifdef ASSERT
3439 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3440 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3441 "Ideal_allocation works");
3442 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3443 "Ideal_allocation works");
3444 if (alloc->is_AllocateArray()) {
3445 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3446 "Ideal_allocation works");
3447 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3448 "Ideal_allocation works");
3449 } else {
3450 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3451 }
3452 }
3453 #endif //ASSERT
3454
3455 return javaoop;
3456 }
3457
3458 //---------------------------new_instance--------------------------------------
3459 // This routine takes a klass_node which may be constant (for a static type)
3460 // or may be non-constant (for reflective code). It will work equally well
3461 // for either, and the graph will fold nicely if the optimizer later reduces
3462 // the type to a constant.
3463 // The optional arguments are for specialized use by intrinsics:
3464 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3465 // - If 'return_size_val', report the the total object size to the caller.
3466 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3467 Node* GraphKit::new_instance(Node* klass_node,
3468 Node* extra_slow_test,
3469 Node* *return_size_val,
3470 bool deoptimize_on_exception) {
3471 // Compute size in doublewords
3472 // The size is always an integral number of doublewords, represented
3473 // as a positive bytewise size stored in the klass's layout_helper.
3474 // The layout_helper also encodes (in a low bit) the need for a slow path.
3475 jint layout_con = Klass::_lh_neutral_value;
3476 Node* layout_val = get_layout_helper(klass_node, layout_con);
3477 int layout_is_con = (layout_val == NULL);
3478
3479 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3480 // Generate the initial go-slow test. It's either ALWAYS (return a
3481 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3482 // case) a computed value derived from the layout_helper.
3483 Node* initial_slow_test = NULL;
3484 if (layout_is_con) {
3485 assert(!StressReflectiveCode, "stress mode does not use these paths");
3486 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3487 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3488 } else { // reflective case
3489 // This reflective path is used by Unsafe.allocateInstance.
3490 // (It may be stress-tested by specifying StressReflectiveCode.)
3491 // Basically, we want to get into the VM is there's an illegal argument.
3492 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3493 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3494 if (extra_slow_test != intcon(0)) {
3495 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3496 }
3497 // (Macro-expander will further convert this to a Bool, if necessary.)
3498 }
3499
3500 // Find the size in bytes. This is easy; it's the layout_helper.
3501 // The size value must be valid even if the slow path is taken.
3502 Node* size = NULL;
3503 if (layout_is_con) {
3504 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3505 } else { // reflective case
3506 // This reflective path is used by clone and Unsafe.allocateInstance.
3507 size = ConvI2X(layout_val);
3508
3509 // Clear the low bits to extract layout_helper_size_in_bytes:
3510 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3511 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3512 size = _gvn.transform( new AndXNode(size, mask) );
3513 }
3514 if (return_size_val != NULL) {
3515 (*return_size_val) = size;
3516 }
3517
3518 // This is a precise notnull oop of the klass.
3519 // (Actually, it need not be precise if this is a reflective allocation.)
3520 // It's what we cast the result to.
3521 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3522 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3523 const TypeOopPtr* oop_type = tklass->as_instance_type();
3524
3525 // Now generate allocation code
3526
3527 // The entire memory state is needed for slow path of the allocation
3528 // since GC and deoptimization can happened.
3529 Node *mem = reset_memory();
3530 set_all_memory(mem); // Create new memory state
3531
3532 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3533 control(), mem, i_o(),
3534 size, klass_node,
3535 initial_slow_test);
3536
3537 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3538 }
3539
3540 //-------------------------------new_array-------------------------------------
3541 // helper for both newarray and anewarray
3542 // The 'length' parameter is (obviously) the length of the array.
3543 // See comments on new_instance for the meaning of the other arguments.
3544 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3545 Node* length, // number of array elements
3546 int nargs, // number of arguments to push back for uncommon trap
3547 Node* *return_size_val,
3548 bool deoptimize_on_exception) {
3549 jint layout_con = Klass::_lh_neutral_value;
3550 Node* layout_val = get_layout_helper(klass_node, layout_con);
3551 int layout_is_con = (layout_val == NULL);
3552
3553 if (!layout_is_con && !StressReflectiveCode &&
3554 !too_many_traps(Deoptimization::Reason_class_check)) {
3555 // This is a reflective array creation site.
3556 // Optimistically assume that it is a subtype of Object[],
3557 // so that we can fold up all the address arithmetic.
3558 layout_con = Klass::array_layout_helper(T_OBJECT);
3559 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3560 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3561 { BuildCutout unless(this, bol_lh, PROB_MAX);
3562 inc_sp(nargs);
3563 uncommon_trap(Deoptimization::Reason_class_check,
3564 Deoptimization::Action_maybe_recompile);
3565 }
3566 layout_val = NULL;
3567 layout_is_con = true;
3568 }
3569
3570 // Generate the initial go-slow test. Make sure we do not overflow
3571 // if length is huge (near 2Gig) or negative! We do not need
3572 // exact double-words here, just a close approximation of needed
3573 // double-words. We can't add any offset or rounding bits, lest we
3574 // take a size -1 of bytes and make it positive. Use an unsigned
3575 // compare, so negative sizes look hugely positive.
3576 int fast_size_limit = FastAllocateSizeLimit;
3577 if (layout_is_con) {
3578 assert(!StressReflectiveCode, "stress mode does not use these paths");
3579 // Increase the size limit if we have exact knowledge of array type.
3580 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3581 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3582 }
3583
3584 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3585 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3586
3587 // --- Size Computation ---
3588 // array_size = round_to_heap(array_header + (length << elem_shift));
3589 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3590 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3591 // The rounding mask is strength-reduced, if possible.
3592 int round_mask = MinObjAlignmentInBytes - 1;
3593 Node* header_size = NULL;
3594 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3595 // (T_BYTE has the weakest alignment and size restrictions...)
3596 if (layout_is_con) {
3597 int hsize = Klass::layout_helper_header_size(layout_con);
3598 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3599 BasicType etype = Klass::layout_helper_element_type(layout_con);
3600 if ((round_mask & ~right_n_bits(eshift)) == 0)
3601 round_mask = 0; // strength-reduce it if it goes away completely
3602 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3603 assert(header_size_min <= hsize, "generic minimum is smallest");
3604 header_size_min = hsize;
3605 header_size = intcon(hsize + round_mask);
3606 } else {
3607 Node* hss = intcon(Klass::_lh_header_size_shift);
3608 Node* hsm = intcon(Klass::_lh_header_size_mask);
3609 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3610 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3611 Node* mask = intcon(round_mask);
3612 header_size = _gvn.transform( new AddINode(hsize, mask) );
3613 }
3614
3615 Node* elem_shift = NULL;
3616 if (layout_is_con) {
3617 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3618 if (eshift != 0)
3619 elem_shift = intcon(eshift);
3620 } else {
3621 // There is no need to mask or shift this value.
3622 // The semantics of LShiftINode include an implicit mask to 0x1F.
3623 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3624 elem_shift = layout_val;
3625 }
3626
3627 // Transition to native address size for all offset calculations:
3628 Node* lengthx = ConvI2X(length);
3629 Node* headerx = ConvI2X(header_size);
3630 #ifdef _LP64
3631 { const TypeInt* tilen = _gvn.find_int_type(length);
3632 if (tilen != NULL && tilen->_lo < 0) {
3633 // Add a manual constraint to a positive range. Cf. array_element_address.
3634 jint size_max = fast_size_limit;
3635 if (size_max > tilen->_hi) size_max = tilen->_hi;
3636 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3637
3638 // Only do a narrow I2L conversion if the range check passed.
3639 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3640 _gvn.transform(iff);
3641 RegionNode* region = new RegionNode(3);
3642 _gvn.set_type(region, Type::CONTROL);
3643 lengthx = new PhiNode(region, TypeLong::LONG);
3644 _gvn.set_type(lengthx, TypeLong::LONG);
3645
3646 // Range check passed. Use ConvI2L node with narrow type.
3647 Node* passed = IfFalse(iff);
3648 region->init_req(1, passed);
3649 // Make I2L conversion control dependent to prevent it from
3650 // floating above the range check during loop optimizations.
3651 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3652
3653 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3654 region->init_req(2, IfTrue(iff));
3655 lengthx->init_req(2, ConvI2X(length));
3656
3657 set_control(region);
3658 record_for_igvn(region);
3659 record_for_igvn(lengthx);
3660 }
3661 }
3662 #endif
3663
3664 // Combine header size (plus rounding) and body size. Then round down.
3665 // This computation cannot overflow, because it is used only in two
3666 // places, one where the length is sharply limited, and the other
3667 // after a successful allocation.
3668 Node* abody = lengthx;
3669 if (elem_shift != NULL)
3670 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3671 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3672 if (round_mask != 0) {
3673 Node* mask = MakeConX(~round_mask);
3674 size = _gvn.transform( new AndXNode(size, mask) );
3675 }
3676 // else if round_mask == 0, the size computation is self-rounding
3677
3678 if (return_size_val != NULL) {
3679 // This is the size
3680 (*return_size_val) = size;
3681 }
3682
3683 // Now generate allocation code
3684
3685 // The entire memory state is needed for slow path of the allocation
3686 // since GC and deoptimization can happened.
3687 Node *mem = reset_memory();
3688 set_all_memory(mem); // Create new memory state
3689
3690 if (initial_slow_test->is_Bool()) {
3691 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3692 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3693 }
3694
3695 // Create the AllocateArrayNode and its result projections
3696 AllocateArrayNode* alloc
3697 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3698 control(), mem, i_o(),
3699 size, klass_node,
3700 initial_slow_test,
3701 length);
3702
3703 // Cast to correct type. Note that the klass_node may be constant or not,
3704 // and in the latter case the actual array type will be inexact also.
3705 // (This happens via a non-constant argument to inline_native_newArray.)
3706 // In any case, the value of klass_node provides the desired array type.
3707 const TypeInt* length_type = _gvn.find_int_type(length);
3708 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3709 if (ary_type->isa_aryptr() && length_type != NULL) {
3710 // Try to get a better type than POS for the size
3711 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3712 }
3713
3714 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3715
3716 // Cast length on remaining path to be as narrow as possible
3717 if (map()->find_edge(length) >= 0) {
3718 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3719 if (ccast != length) {
3720 _gvn.set_type_bottom(ccast);
3721 record_for_igvn(ccast);
3722 replace_in_map(length, ccast);
3723 }
3724 }
3725
3726 return javaoop;
3727 }
3728
3729 // The following "Ideal_foo" functions are placed here because they recognize
3730 // the graph shapes created by the functions immediately above.
3731
3732 //---------------------------Ideal_allocation----------------------------------
3733 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3734 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3735 if (ptr == NULL) { // reduce dumb test in callers
3736 return NULL;
3737 }
3738 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3739 ptr = ptr->in(1);
3740 if (ptr == NULL) return NULL;
3741 }
3742 // Return NULL for allocations with several casts:
3743 // j.l.reflect.Array.newInstance(jobject, jint)
3744 // Object.clone()
3745 // to keep more precise type from last cast.
3746 if (ptr->is_Proj()) {
3747 Node* allo = ptr->in(0);
3748 if (allo != NULL && allo->is_Allocate()) {
3749 return allo->as_Allocate();
3750 }
3751 }
3752 // Report failure to match.
3753 return NULL;
3754 }
3755
3756 // Fancy version which also strips off an offset (and reports it to caller).
3757 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3758 intptr_t& offset) {
3759 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3760 if (base == NULL) return NULL;
3761 return Ideal_allocation(base, phase);
3762 }
3763
3764 // Trace Initialize <- Proj[Parm] <- Allocate
3765 AllocateNode* InitializeNode::allocation() {
3766 Node* rawoop = in(InitializeNode::RawAddress);
3767 if (rawoop->is_Proj()) {
3768 Node* alloc = rawoop->in(0);
3769 if (alloc->is_Allocate()) {
3770 return alloc->as_Allocate();
3771 }
3772 }
3773 return NULL;
3774 }
3775
3776 // Trace Allocate -> Proj[Parm] -> Initialize
3777 InitializeNode* AllocateNode::initialization() {
3778 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
3779 if (rawoop == NULL) return NULL;
3780 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3781 Node* init = rawoop->fast_out(i);
3782 if (init->is_Initialize()) {
3783 assert(init->as_Initialize()->allocation() == this, "2-way link");
3784 return init->as_Initialize();
3785 }
3786 }
3787 return NULL;
3788 }
3789
3790 //----------------------------- loop predicates ---------------------------
3791
3792 //------------------------------add_predicate_impl----------------------------
3793 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3794 // Too many traps seen?
3795 if (too_many_traps(reason)) {
3796 #ifdef ASSERT
3797 if (TraceLoopPredicate) {
3798 int tc = C->trap_count(reason);
3799 tty->print("too many traps=%s tcount=%d in ",
3800 Deoptimization::trap_reason_name(reason), tc);
3801 method()->print(); // which method has too many predicate traps
3802 tty->cr();
3803 }
3804 #endif
3805 // We cannot afford to take more traps here,
3806 // do not generate predicate.
3807 return;
3808 }
3809
3810 Node *cont = _gvn.intcon(1);
3811 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
3812 Node *bol = _gvn.transform(new Conv2BNode(opq));
3813 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3814 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3815 C->add_predicate_opaq(opq);
3816 {
3817 PreserveJVMState pjvms(this);
3818 set_control(iffalse);
3819 inc_sp(nargs);
3820 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3821 }
3822 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3823 set_control(iftrue);
3824 }
3825
3826 //------------------------------add_predicate---------------------------------
3827 void GraphKit::add_predicate(int nargs) {
3828 if (UseLoopPredicate) {
3829 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3830 }
3831 // loop's limit check predicate should be near the loop.
3832 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3833 }
3834
3835 //----------------------------- store barriers ----------------------------
3836 #define __ ideal.
3837
3838 bool GraphKit::use_ReduceInitialCardMarks() {
3839 BarrierSet *bs = BarrierSet::barrier_set();
3840 return bs->is_a(BarrierSet::CardTableBarrierSet)
3841 && barrier_set_cast<CardTableBarrierSet>(bs)->can_elide_tlab_store_barriers()
3842 && ReduceInitialCardMarks;
3843 }
3844
3845 void GraphKit::sync_kit(IdealKit& ideal) {
3846 set_all_memory(__ merged_memory());
3847 set_i_o(__ i_o());
3848 set_control(__ ctrl());
3849 }
3850
3851 void GraphKit::final_sync(IdealKit& ideal) {
3852 // Final sync IdealKit and graphKit.
3853 sync_kit(ideal);
3854 }
3855
3856 Node* GraphKit::byte_map_base_node() {
3857 // Get base of card map
3858 jbyte* card_table_base = ci_card_table_address();
3859 if (card_table_base != NULL) {
3860 return makecon(TypeRawPtr::make((address)card_table_base));
3861 } else {
3862 return null();
3863 }
3864 }
3865
3866 // vanilla/CMS post barrier
3867 // Insert a write-barrier store. This is to let generational GC work; we have
3868 // to flag all oop-stores before the next GC point.
3869 void GraphKit::write_barrier_post(Node* oop_store,
3870 Node* obj,
3871 Node* adr,
3872 uint adr_idx,
3873 Node* val,
3874 bool use_precise) {
3875 // No store check needed if we're storing a NULL or an old object
3876 // (latter case is probably a string constant). The concurrent
3877 // mark sweep garbage collector, however, needs to have all nonNull
3878 // oop updates flagged via card-marks.
3879 if (val != NULL && val->is_Con()) {
3880 // must be either an oop or NULL
3881 const Type* t = val->bottom_type();
3882 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3883 // stores of null never (?) need barriers
3884 return;
3885 }
3886
3887 if (use_ReduceInitialCardMarks()
3888 && obj == just_allocated_object(control())) {
3889 // We can skip marks on a freshly-allocated object in Eden.
3890 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
3891 // That routine informs GC to take appropriate compensating steps,
3892 // upon a slow-path allocation, so as to make this card-mark
3893 // elision safe.
3894 return;
3895 }
3896
3897 if (!use_precise) {
3898 // All card marks for a (non-array) instance are in one place:
3899 adr = obj;
3900 }
3901 // (Else it's an array (or unknown), and we want more precise card marks.)
3902 assert(adr != NULL, "");
3903
3904 IdealKit ideal(this, true);
3905
3906 // Convert the pointer to an int prior to doing math on it
3907 Node* cast = __ CastPX(__ ctrl(), adr);
3908
3909 // Divide by card size
3910 assert(BarrierSet::barrier_set()->is_a(BarrierSet::CardTableBarrierSet),
3911 "Only one we handle so far.");
3912 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
3913
3914 // Combine card table base and card offset
3915 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3916
3917 // Get the alias_index for raw card-mark memory
3918 int adr_type = Compile::AliasIdxRaw;
3919 Node* zero = __ ConI(0); // Dirty card value
3920 BasicType bt = T_BYTE;
3921
3922 if (UseConcMarkSweepGC && UseCondCardMark) {
3923 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
3924 __ sync_kit(this);
3925 }
3926
3927 if (UseCondCardMark) {
3928 // The classic GC reference write barrier is typically implemented
3929 // as a store into the global card mark table. Unfortunately
3930 // unconditional stores can result in false sharing and excessive
3931 // coherence traffic as well as false transactional aborts.
3932 // UseCondCardMark enables MP "polite" conditional card mark
3933 // stores. In theory we could relax the load from ctrl() to
3934 // no_ctrl, but that doesn't buy much latitude.
3935 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3936 __ if_then(card_val, BoolTest::ne, zero);
3937 }
3938
3939 // Smash zero into card
3940 if( !UseConcMarkSweepGC ) {
3941 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3942 } else {
3943 // Specialized path for CM store barrier
3944 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3945 }
3946
3947 if (UseCondCardMark) {
3948 __ end_if();
3949 }
3950
3951 // Final sync IdealKit and GraphKit.
3952 final_sync(ideal);
3953 }
3954
3955 #if INCLUDE_G1GC
3956
3957 /*
3958 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
3959 * required by SATB to make sure all objects live at the start of the
3960 * marking are kept alive, all reference updates need to any previous
3961 * reference stored before writing.
3962 *
3963 * If the previous value is NULL there is no need to save the old value.
3964 * References that are NULL are filtered during runtime by the barrier
3965 * code to avoid unnecessary queuing.
3966 *
3967 * However in the case of newly allocated objects it might be possible to
3968 * prove that the reference about to be overwritten is NULL during compile
3969 * time and avoid adding the barrier code completely.
3970 *
3971 * The compiler needs to determine that the object in which a field is about
3972 * to be written is newly allocated, and that no prior store to the same field
3973 * has happened since the allocation.
3974 *
3975 * Returns true if the pre-barrier can be removed
3976 */
3977 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
3978 BasicType bt, uint adr_idx) {
3979 intptr_t offset = 0;
3980 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3981 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
3982
3983 if (offset == Type::OffsetBot) {
3984 return false; // cannot unalias unless there are precise offsets
3985 }
3986
3987 if (alloc == NULL) {
3988 return false; // No allocation found
3989 }
3990
3991 intptr_t size_in_bytes = type2aelembytes(bt);
3992
3993 Node* mem = memory(adr_idx); // start searching here...
3994
3995 for (int cnt = 0; cnt < 50; cnt++) {
3996
3997 if (mem->is_Store()) {
3998
3999 Node* st_adr = mem->in(MemNode::Address);
4000 intptr_t st_offset = 0;
4001 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
4002
4003 if (st_base == NULL) {
4004 break; // inscrutable pointer
4005 }
4006
4007 // Break we have found a store with same base and offset as ours so break
4008 if (st_base == base && st_offset == offset) {
4009 break;
4010 }
4011
4012 if (st_offset != offset && st_offset != Type::OffsetBot) {
4013 const int MAX_STORE = BytesPerLong;
4014 if (st_offset >= offset + size_in_bytes ||
4015 st_offset <= offset - MAX_STORE ||
4016 st_offset <= offset - mem->as_Store()->memory_size()) {
4017 // Success: The offsets are provably independent.
4018 // (You may ask, why not just test st_offset != offset and be done?
4019 // The answer is that stores of different sizes can co-exist
4020 // in the same sequence of RawMem effects. We sometimes initialize
4021 // a whole 'tile' of array elements with a single jint or jlong.)
4022 mem = mem->in(MemNode::Memory);
4023 continue; // advance through independent store memory
4024 }
4025 }
4026
4027 if (st_base != base
4028 && MemNode::detect_ptr_independence(base, alloc, st_base,
4029 AllocateNode::Ideal_allocation(st_base, phase),
4030 phase)) {
4031 // Success: The bases are provably independent.
4032 mem = mem->in(MemNode::Memory);
4033 continue; // advance through independent store memory
4034 }
4035 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4036
4037 InitializeNode* st_init = mem->in(0)->as_Initialize();
4038 AllocateNode* st_alloc = st_init->allocation();
4039
4040 // Make sure that we are looking at the same allocation site.
4041 // The alloc variable is guaranteed to not be null here from earlier check.
4042 if (alloc == st_alloc) {
4043 // Check that the initialization is storing NULL so that no previous store
4044 // has been moved up and directly write a reference
4045 Node* captured_store = st_init->find_captured_store(offset,
4046 type2aelembytes(T_OBJECT),
4047 phase);
4048 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4049 return true;
4050 }
4051 }
4052 }
4053
4054 // Unless there is an explicit 'continue', we must bail out here,
4055 // because 'mem' is an inscrutable memory state (e.g., a call).
4056 break;
4057 }
4058
4059 return false;
4060 }
4061
4062 // G1 pre/post barriers
4063 void GraphKit::g1_write_barrier_pre(bool do_load,
4064 Node* obj,
4065 Node* adr,
4066 uint alias_idx,
4067 Node* val,
4068 const TypeOopPtr* val_type,
4069 Node* pre_val,
4070 BasicType bt) {
4071
4072 // Some sanity checks
4073 // Note: val is unused in this routine.
4074
4075 if (do_load) {
4076 // We need to generate the load of the previous value
4077 assert(obj != NULL, "must have a base");
4078 assert(adr != NULL, "where are loading from?");
4079 assert(pre_val == NULL, "loaded already?");
4080 assert(val_type != NULL, "need a type");
4081
4082 if (use_ReduceInitialCardMarks()
4083 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4084 return;
4085 }
4086
4087 } else {
4088 // In this case both val_type and alias_idx are unused.
4089 assert(pre_val != NULL, "must be loaded already");
4090 // Nothing to be done if pre_val is null.
4091 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4092 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4093 }
4094 assert(bt == T_OBJECT, "or we shouldn't be here");
4095
4096 IdealKit ideal(this, true);
4097
4098 Node* tls = __ thread(); // ThreadLocalStorage
4099
4100 Node* no_ctrl = NULL;
4101 Node* no_base = __ top();
4102 Node* zero = __ ConI(0);
4103 Node* zeroX = __ ConX(0);
4104
4105 float likely = PROB_LIKELY(0.999);
4106 float unlikely = PROB_UNLIKELY(0.999);
4107
4108 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4109 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4110
4111 // Offsets into the thread
4112 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
4113 const int index_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset());
4114 const int buffer_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
4115
4116 // Now the actual pointers into the thread
4117 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4118 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4119 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4120
4121 // Now some of the values
4122 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4123
4124 // if (!marking)
4125 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4126 BasicType index_bt = TypeX_X->basic_type();
4127 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4128 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4129
4130 if (do_load) {
4131 // load original value
4132 // alias_idx correct??
4133 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4134 }
4135
4136 // if (pre_val != NULL)
4137 __ if_then(pre_val, BoolTest::ne, null()); {
4138 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4139
4140 // is the queue for this thread full?
4141 __ if_then(index, BoolTest::ne, zeroX, likely); {
4142
4143 // decrement the index
4144 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4145
4146 // Now get the buffer location we will log the previous value into and store it
4147 Node *log_addr = __ AddP(no_base, buffer, next_index);
4148 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4149 // update the index
4150 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4151
4152 } __ else_(); {
4153
4154 // logging buffer is full, call the runtime
4155 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4156 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4157 } __ end_if(); // (!index)
4158 } __ end_if(); // (pre_val != NULL)
4159 } __ end_if(); // (!marking)
4160
4161 // Final sync IdealKit and GraphKit.
4162 final_sync(ideal);
4163 }
4164
4165 /*
4166 * G1 similar to any GC with a Young Generation requires a way to keep track of
4167 * references from Old Generation to Young Generation to make sure all live
4168 * objects are found. G1 also requires to keep track of object references
4169 * between different regions to enable evacuation of old regions, which is done
4170 * as part of mixed collections. References are tracked in remembered sets and
4171 * is continuously updated as reference are written to with the help of the
4172 * post-barrier.
4173 *
4174 * To reduce the number of updates to the remembered set the post-barrier
4175 * filters updates to fields in objects located in the Young Generation,
4176 * the same region as the reference, when the NULL is being written or
4177 * if the card is already marked as dirty by an earlier write.
4178 *
4179 * Under certain circumstances it is possible to avoid generating the
4180 * post-barrier completely if it is possible during compile time to prove
4181 * the object is newly allocated and that no safepoint exists between the
4182 * allocation and the store.
4183 *
4184 * In the case of slow allocation the allocation code must handle the barrier
4185 * as part of the allocation in the case the allocated object is not located
4186 * in the nursery, this would happen for humongous objects. This is similar to
4187 * how CMS is required to handle this case, see the comments for the method
4188 * CardTableBarrierSet::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier.
4189 * A deferred card mark is required for these objects and handled in the above
4190 * mentioned methods.
4191 *
4192 * Returns true if the post barrier can be removed
4193 */
4194 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4195 Node* adr) {
4196 intptr_t offset = 0;
4197 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4198 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4199
4200 if (offset == Type::OffsetBot) {
4201 return false; // cannot unalias unless there are precise offsets
4202 }
4203
4204 if (alloc == NULL) {
4205 return false; // No allocation found
4206 }
4207
4208 // Start search from Store node
4209 Node* mem = store->in(MemNode::Control);
4210 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4211
4212 InitializeNode* st_init = mem->in(0)->as_Initialize();
4213 AllocateNode* st_alloc = st_init->allocation();
4214
4215 // Make sure we are looking at the same allocation
4216 if (alloc == st_alloc) {
4217 return true;
4218 }
4219 }
4220
4221 return false;
4222 }
4223
4224 //
4225 // Update the card table and add card address to the queue
4226 //
4227 void GraphKit::g1_mark_card(IdealKit& ideal,
4228 Node* card_adr,
4229 Node* oop_store,
4230 uint oop_alias_idx,
4231 Node* index,
4232 Node* index_adr,
4233 Node* buffer,
4234 const TypeFunc* tf) {
4235
4236 Node* zero = __ ConI(0);
4237 Node* zeroX = __ ConX(0);
4238 Node* no_base = __ top();
4239 BasicType card_bt = T_BYTE;
4240 // Smash zero into card. MUST BE ORDERED WRT TO STORE
4241 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4242
4243 // Now do the queue work
4244 __ if_then(index, BoolTest::ne, zeroX); {
4245
4246 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4247 Node* log_addr = __ AddP(no_base, buffer, next_index);
4248
4249 // Order, see storeCM.
4250 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4251 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4252
4253 } __ else_(); {
4254 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4255 } __ end_if();
4256
4257 }
4258
4259 void GraphKit::g1_write_barrier_post(Node* oop_store,
4260 Node* obj,
4261 Node* adr,
4262 uint alias_idx,
4263 Node* val,
4264 BasicType bt,
4265 bool use_precise) {
4266 // If we are writing a NULL then we need no post barrier
4267
4268 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4269 // Must be NULL
4270 const Type* t = val->bottom_type();
4271 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4272 // No post barrier if writing NULLx
4273 return;
4274 }
4275
4276 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4277 // We can skip marks on a freshly-allocated object in Eden.
4278 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
4279 // That routine informs GC to take appropriate compensating steps,
4280 // upon a slow-path allocation, so as to make this card-mark
4281 // elision safe.
4282 return;
4283 }
4284
4285 if (use_ReduceInitialCardMarks()
4286 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4287 return;
4288 }
4289
4290 if (!use_precise) {
4291 // All card marks for a (non-array) instance are in one place:
4292 adr = obj;
4293 }
4294 // (Else it's an array (or unknown), and we want more precise card marks.)
4295 assert(adr != NULL, "");
4296
4297 IdealKit ideal(this, true);
4298
4299 Node* tls = __ thread(); // ThreadLocalStorage
4300
4301 Node* no_base = __ top();
4302 float likely = PROB_LIKELY(0.999);
4303 float unlikely = PROB_UNLIKELY(0.999);
4304 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val());
4305 Node* dirty_card = __ ConI((jint)CardTable::dirty_card_val());
4306 Node* zeroX = __ ConX(0);
4307
4308 // Get the alias_index for raw card-mark memory
4309 const TypePtr* card_type = TypeRawPtr::BOTTOM;
4310
4311 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4312
4313 // Offsets into the thread
4314 const int index_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset());
4315 const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset());
4316
4317 // Pointers into the thread
4318
4319 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4320 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4321
4322 // Now some values
4323 // Use ctrl to avoid hoisting these values past a safepoint, which could
4324 // potentially reset these fields in the JavaThread.
4325 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4326 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4327
4328 // Convert the store obj pointer to an int prior to doing math on it
4329 // Must use ctrl to prevent "integerized oop" existing across safepoint
4330 Node* cast = __ CastPX(__ ctrl(), adr);
4331
4332 // Divide pointer by card size
4333 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
4334
4335 // Combine card table base and card offset
4336 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4337
4338 // If we know the value being stored does it cross regions?
4339
4340 if (val != NULL) {
4341 // Does the store cause us to cross regions?
4342
4343 // Should be able to do an unsigned compare of region_size instead of
4344 // and extra shift. Do we have an unsigned compare??
4345 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4346 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4347
4348 // if (xor_res == 0) same region so skip
4349 __ if_then(xor_res, BoolTest::ne, zeroX); {
4350
4351 // No barrier if we are storing a NULL
4352 __ if_then(val, BoolTest::ne, null(), unlikely); {
4353
4354 // Ok must mark the card if not already dirty
4355
4356 // load the original value of the card
4357 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4358
4359 __ if_then(card_val, BoolTest::ne, young_card); {
4360 sync_kit(ideal);
4361 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4362 insert_mem_bar(Op_MemBarVolatile, oop_store);
4363 __ sync_kit(this);
4364
4365 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4366 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4367 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4368 } __ end_if();
4369 } __ end_if();
4370 } __ end_if();
4371 } __ end_if();
4372 } else {
4373 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
4374 // We don't need a barrier here if the destination is a newly allocated object
4375 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
4376 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
4377 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
4378 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4379 __ if_then(card_val, BoolTest::ne, young_card); {
4380 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4381 } __ end_if();
4382 }
4383
4384 // Final sync IdealKit and GraphKit.
4385 final_sync(ideal);
4386 }
4387 #undef __
4388
4389 #endif // INCLUDE_G1GC
4390
4391 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4392 Node* len = load_array_length(load_String_value(ctrl, str));
4393 Node* coder = load_String_coder(ctrl, str);
4394 // Divide length by 2 if coder is UTF16
4395 return _gvn.transform(new RShiftINode(len, coder));
4396 }
4397
4398 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4399 int value_offset = java_lang_String::value_offset_in_bytes();
4400 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4401 false, NULL, 0);
4402 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4403 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4404 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4405 ciTypeArrayKlass::make(T_BYTE), true, 0);
4406 int value_field_idx = C->get_alias_index(value_field_type);
4407 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4408 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4409 // String.value field is known to be @Stable.
4410 if (UseImplicitStableValues) {
4411 load = cast_array_to_stable(load, value_type);
4412 }
4413 return load;
4414 }
4415
4416 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4417 if (!CompactStrings) {
4418 return intcon(java_lang_String::CODER_UTF16);
4419 }
4420 int coder_offset = java_lang_String::coder_offset_in_bytes();
4421 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4422 false, NULL, 0);
4423 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4424 int coder_field_idx = C->get_alias_index(coder_field_type);
4425 return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4426 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4427 }
4428
4429 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4430 int value_offset = java_lang_String::value_offset_in_bytes();
4431 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4432 false, NULL, 0);
4433 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4434 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4435 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4436 }
4437
4438 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4439 int coder_offset = java_lang_String::coder_offset_in_bytes();
4440 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4441 false, NULL, 0);
4442 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4443 int coder_field_idx = C->get_alias_index(coder_field_type);
4444 store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4445 value, T_BYTE, coder_field_idx, MemNode::unordered);
4446 }
4447
4448 // Capture src and dst memory state with a MergeMemNode
4449 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4450 if (src_type == dst_type) {
4451 // Types are equal, we don't need a MergeMemNode
4452 return memory(src_type);
4453 }
4454 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4455 record_for_igvn(merge); // fold it up later, if possible
4456 int src_idx = C->get_alias_index(src_type);
4457 int dst_idx = C->get_alias_index(dst_type);
4458 merge->set_memory_at(src_idx, memory(src_idx));
4459 merge->set_memory_at(dst_idx, memory(dst_idx));
4460 return merge;
4461 }
4462
4463 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4464 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4465 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4466 // If input and output memory types differ, capture both states to preserve
4467 // the dependency between preceding and subsequent loads/stores.
4468 // For example, the following program:
4469 // StoreB
4470 // compress_string
4471 // LoadB
4472 // has this memory graph (use->def):
4473 // LoadB -> compress_string -> CharMem
4474 // ... -> StoreB -> ByteMem
4475 // The intrinsic hides the dependency between LoadB and StoreB, causing
4476 // the load to read from memory not containing the result of the StoreB.
4477 // The correct memory graph should look like this:
4478 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4479 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4480 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4481 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4482 set_memory(res_mem, TypeAryPtr::BYTES);
4483 return str;
4484 }
4485
4486 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4487 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4488 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4489 // Capture src and dst memory (see comment in 'compress_string').
4490 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4491 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4492 set_memory(_gvn.transform(str), dst_type);
4493 }
4494
4495 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4496 /**
4497 * int i_char = start;
4498 * for (int i_byte = 0; i_byte < count; i_byte++) {
4499 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4500 * }
4501 */
4502 add_predicate();
4503 RegionNode* head = new RegionNode(3);
4504 head->init_req(1, control());
4505 gvn().set_type(head, Type::CONTROL);
4506 record_for_igvn(head);
4507
4508 Node* i_byte = new PhiNode(head, TypeInt::INT);
4509 i_byte->init_req(1, intcon(0));
4510 gvn().set_type(i_byte, TypeInt::INT);
4511 record_for_igvn(i_byte);
4512
4513 Node* i_char = new PhiNode(head, TypeInt::INT);
4514 i_char->init_req(1, start);
4515 gvn().set_type(i_char, TypeInt::INT);
4516 record_for_igvn(i_char);
4517
4518 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4519 gvn().set_type(mem, Type::MEMORY);
4520 record_for_igvn(mem);
4521 set_control(head);
4522 set_memory(mem, TypeAryPtr::BYTES);
4523 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4524 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4525 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4526 false, false, true /* mismatched */);
4527
4528 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4529 head->init_req(2, IfTrue(iff));
4530 mem->init_req(2, st);
4531 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4532 i_char->init_req(2, AddI(i_char, intcon(2)));
4533
4534 set_control(IfFalse(iff));
4535 set_memory(st, TypeAryPtr::BYTES);
4536 }
4537
4538 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4539 if (!field->is_constant()) {
4540 return NULL; // Field not marked as constant.
4541 }
4542 ciInstance* holder = NULL;
4543 if (!field->is_static()) {
4544 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4545 if (const_oop != NULL && const_oop->is_instance()) {
4546 holder = const_oop->as_instance();
4547 }
4548 }
4549 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4550 /*is_unsigned_load=*/false);
4551 if (con_type != NULL) {
4552 return makecon(con_type);
4553 }
4554 return NULL;
4555 }
4556
4557 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4558 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4559 // assumption of CCP analysis.
4560 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4561 }