1 /*
2 * Copyright (c) 2001, 2015, 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 "compiler/compileLog.hpp"
27 #include "ci/ciValueKlass.hpp"
28 #include "gc/g1/g1SATBCardTableModRefBS.hpp"
29 #include "gc/g1/heapRegion.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/cardTableModRefBS.hpp"
32 #include "gc/shared/collectedHeap.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/idealKit.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/opaquenode.hpp"
42 #include "opto/parse.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/valuetypenode.hpp"
46 #include "runtime/deoptimization.hpp"
47 #include "runtime/sharedRuntime.hpp"
48
49 //----------------------------GraphKit-----------------------------------------
50 // Main utility constructor.
51 GraphKit::GraphKit(JVMState* jvms)
52 : Phase(Phase::Parser),
53 _env(C->env()),
54 _gvn(*C->initial_gvn())
55 {
56 _exceptions = jvms->map()->next_exception();
57 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
58 set_jvms(jvms);
59 }
60
61 // Private constructor for parser.
62 GraphKit::GraphKit()
63 : Phase(Phase::Parser),
64 _env(C->env()),
65 _gvn(*C->initial_gvn())
66 {
67 _exceptions = NULL;
68 set_map(NULL);
69 debug_only(_sp = -99);
70 debug_only(set_bci(-99));
71 }
72
73
74
75 //---------------------------clean_stack---------------------------------------
76 // Clear away rubbish from the stack area of the JVM state.
77 // This destroys any arguments that may be waiting on the stack.
78 void GraphKit::clean_stack(int from_sp) {
79 SafePointNode* map = this->map();
80 JVMState* jvms = this->jvms();
81 int stk_size = jvms->stk_size();
82 int stkoff = jvms->stkoff();
83 Node* top = this->top();
84 for (int i = from_sp; i < stk_size; i++) {
85 if (map->in(stkoff + i) != top) {
86 map->set_req(stkoff + i, top);
87 }
88 }
89 }
90
91
92 //--------------------------------sync_jvms-----------------------------------
93 // Make sure our current jvms agrees with our parse state.
94 JVMState* GraphKit::sync_jvms() const {
95 JVMState* jvms = this->jvms();
96 jvms->set_bci(bci()); // Record the new bci in the JVMState
97 jvms->set_sp(sp()); // Record the new sp in the JVMState
98 assert(jvms_in_sync(), "jvms is now in sync");
99 return jvms;
100 }
101
102 //--------------------------------sync_jvms_for_reexecute---------------------
103 // Make sure our current jvms agrees with our parse state. This version
104 // uses the reexecute_sp for reexecuting bytecodes.
105 JVMState* GraphKit::sync_jvms_for_reexecute() {
106 JVMState* jvms = this->jvms();
107 jvms->set_bci(bci()); // Record the new bci in the JVMState
108 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
109 return jvms;
110 }
111
112 #ifdef ASSERT
113 bool GraphKit::jvms_in_sync() const {
114 Parse* parse = is_Parse();
115 if (parse == NULL) {
116 if (bci() != jvms()->bci()) return false;
117 if (sp() != (int)jvms()->sp()) return false;
118 return true;
119 }
120 if (jvms()->method() != parse->method()) return false;
121 if (jvms()->bci() != parse->bci()) return false;
122 int jvms_sp = jvms()->sp();
123 if (jvms_sp != parse->sp()) return false;
124 int jvms_depth = jvms()->depth();
125 if (jvms_depth != parse->depth()) return false;
126 return true;
127 }
128
129 // Local helper checks for special internal merge points
130 // used to accumulate and merge exception states.
131 // They are marked by the region's in(0) edge being the map itself.
132 // Such merge points must never "escape" into the parser at large,
133 // until they have been handed to gvn.transform.
134 static bool is_hidden_merge(Node* reg) {
135 if (reg == NULL) return false;
136 if (reg->is_Phi()) {
137 reg = reg->in(0);
138 if (reg == NULL) return false;
139 }
140 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
141 }
142
143 void GraphKit::verify_map() const {
144 if (map() == NULL) return; // null map is OK
145 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
146 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
147 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
148 }
149
150 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
151 assert(ex_map->next_exception() == NULL, "not already part of a chain");
152 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
153 }
154 #endif
155
156 //---------------------------stop_and_kill_map---------------------------------
157 // Set _map to NULL, signalling a stop to further bytecode execution.
158 // First smash the current map's control to a constant, to mark it dead.
159 void GraphKit::stop_and_kill_map() {
160 SafePointNode* dead_map = stop();
161 if (dead_map != NULL) {
162 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
163 assert(dead_map->is_killed(), "must be so marked");
164 }
165 }
166
167
168 //--------------------------------stopped--------------------------------------
169 // Tell if _map is NULL, or control is top.
170 bool GraphKit::stopped() {
171 if (map() == NULL) return true;
172 else if (control() == top()) return true;
173 else return false;
174 }
175
176
177 //-----------------------------has_ex_handler----------------------------------
178 // Tell if this method or any caller method has exception handlers.
179 bool GraphKit::has_ex_handler() {
180 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
181 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
182 return true;
183 }
184 }
185 return false;
186 }
187
188 //------------------------------save_ex_oop------------------------------------
189 // Save an exception without blowing stack contents or other JVM state.
190 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
191 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
192 ex_map->add_req(ex_oop);
193 debug_only(verify_exception_state(ex_map));
194 }
195
196 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
197 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
198 Node* ex_oop = ex_map->in(ex_map->req()-1);
199 if (clear_it) ex_map->del_req(ex_map->req()-1);
200 return ex_oop;
201 }
202
203 //-----------------------------saved_ex_oop------------------------------------
204 // Recover a saved exception from its map.
205 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
206 return common_saved_ex_oop(ex_map, false);
207 }
208
209 //--------------------------clear_saved_ex_oop---------------------------------
210 // Erase a previously saved exception from its map.
211 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
212 return common_saved_ex_oop(ex_map, true);
213 }
214
215 #ifdef ASSERT
216 //---------------------------has_saved_ex_oop----------------------------------
217 // Erase a previously saved exception from its map.
218 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
219 return ex_map->req() == ex_map->jvms()->endoff()+1;
220 }
221 #endif
222
223 //-------------------------make_exception_state--------------------------------
224 // Turn the current JVM state into an exception state, appending the ex_oop.
225 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
226 sync_jvms();
227 SafePointNode* ex_map = stop(); // do not manipulate this map any more
228 set_saved_ex_oop(ex_map, ex_oop);
229 return ex_map;
230 }
231
232
233 //--------------------------add_exception_state--------------------------------
234 // Add an exception to my list of exceptions.
235 void GraphKit::add_exception_state(SafePointNode* ex_map) {
236 if (ex_map == NULL || ex_map->control() == top()) {
237 return;
238 }
239 #ifdef ASSERT
240 verify_exception_state(ex_map);
241 if (has_exceptions()) {
242 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
243 }
244 #endif
245
246 // If there is already an exception of exactly this type, merge with it.
247 // In particular, null-checks and other low-level exceptions common up here.
248 Node* ex_oop = saved_ex_oop(ex_map);
249 const Type* ex_type = _gvn.type(ex_oop);
250 if (ex_oop == top()) {
251 // No action needed.
252 return;
253 }
254 assert(ex_type->isa_instptr(), "exception must be an instance");
255 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
256 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
257 // We check sp also because call bytecodes can generate exceptions
258 // both before and after arguments are popped!
259 if (ex_type2 == ex_type
260 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
261 combine_exception_states(ex_map, e2);
262 return;
263 }
264 }
265
266 // No pre-existing exception of the same type. Chain it on the list.
267 push_exception_state(ex_map);
268 }
269
270 //-----------------------add_exception_states_from-----------------------------
271 void GraphKit::add_exception_states_from(JVMState* jvms) {
272 SafePointNode* ex_map = jvms->map()->next_exception();
273 if (ex_map != NULL) {
274 jvms->map()->set_next_exception(NULL);
275 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
276 next_map = ex_map->next_exception();
277 ex_map->set_next_exception(NULL);
278 add_exception_state(ex_map);
279 }
280 }
281 }
282
283 //-----------------------transfer_exceptions_into_jvms-------------------------
284 JVMState* GraphKit::transfer_exceptions_into_jvms() {
285 if (map() == NULL) {
286 // We need a JVMS to carry the exceptions, but the map has gone away.
287 // Create a scratch JVMS, cloned from any of the exception states...
288 if (has_exceptions()) {
289 _map = _exceptions;
290 _map = clone_map();
291 _map->set_next_exception(NULL);
292 clear_saved_ex_oop(_map);
293 debug_only(verify_map());
294 } else {
295 // ...or created from scratch
296 JVMState* jvms = new (C) JVMState(_method, NULL);
297 jvms->set_bci(_bci);
298 jvms->set_sp(_sp);
299 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
300 set_jvms(jvms);
301 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
302 set_all_memory(top());
303 while (map()->req() < jvms->endoff()) map()->add_req(top());
304 }
305 // (This is a kludge, in case you didn't notice.)
306 set_control(top());
307 }
308 JVMState* jvms = sync_jvms();
309 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
310 jvms->map()->set_next_exception(_exceptions);
311 _exceptions = NULL; // done with this set of exceptions
312 return jvms;
313 }
314
315 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
316 assert(is_hidden_merge(dstphi), "must be a special merge node");
317 assert(is_hidden_merge(srcphi), "must be a special merge node");
318 uint limit = srcphi->req();
319 for (uint i = PhiNode::Input; i < limit; i++) {
320 dstphi->add_req(srcphi->in(i));
321 }
322 }
323 static inline void add_one_req(Node* dstphi, Node* src) {
324 assert(is_hidden_merge(dstphi), "must be a special merge node");
325 assert(!is_hidden_merge(src), "must not be a special merge node");
326 dstphi->add_req(src);
327 }
328
329 //-----------------------combine_exception_states------------------------------
330 // This helper function combines exception states by building phis on a
331 // specially marked state-merging region. These regions and phis are
332 // untransformed, and can build up gradually. The region is marked by
333 // having a control input of its exception map, rather than NULL. Such
334 // regions do not appear except in this function, and in use_exception_state.
335 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
336 if (failing()) return; // dying anyway...
337 JVMState* ex_jvms = ex_map->_jvms;
338 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
339 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
340 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
341 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
342 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
343 assert(ex_map->req() == phi_map->req(), "matching maps");
344 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
345 Node* hidden_merge_mark = root();
346 Node* region = phi_map->control();
347 MergeMemNode* phi_mem = phi_map->merged_memory();
348 MergeMemNode* ex_mem = ex_map->merged_memory();
349 if (region->in(0) != hidden_merge_mark) {
350 // The control input is not (yet) a specially-marked region in phi_map.
351 // Make it so, and build some phis.
352 region = new RegionNode(2);
353 _gvn.set_type(region, Type::CONTROL);
354 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
355 region->init_req(1, phi_map->control());
356 phi_map->set_control(region);
357 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
358 record_for_igvn(io_phi);
359 _gvn.set_type(io_phi, Type::ABIO);
360 phi_map->set_i_o(io_phi);
361 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
362 Node* m = mms.memory();
363 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
364 record_for_igvn(m_phi);
365 _gvn.set_type(m_phi, Type::MEMORY);
366 mms.set_memory(m_phi);
367 }
368 }
369
370 // Either or both of phi_map and ex_map might already be converted into phis.
371 Node* ex_control = ex_map->control();
372 // if there is special marking on ex_map also, we add multiple edges from src
373 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
374 // how wide was the destination phi_map, originally?
375 uint orig_width = region->req();
376
377 if (add_multiple) {
378 add_n_reqs(region, ex_control);
379 add_n_reqs(phi_map->i_o(), ex_map->i_o());
380 } else {
381 // ex_map has no merges, so we just add single edges everywhere
382 add_one_req(region, ex_control);
383 add_one_req(phi_map->i_o(), ex_map->i_o());
384 }
385 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
386 if (mms.is_empty()) {
387 // get a copy of the base memory, and patch some inputs into it
388 const TypePtr* adr_type = mms.adr_type(C);
389 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
390 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
391 mms.set_memory(phi);
392 // Prepare to append interesting stuff onto the newly sliced phi:
393 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
394 }
395 // Append stuff from ex_map:
396 if (add_multiple) {
397 add_n_reqs(mms.memory(), mms.memory2());
398 } else {
399 add_one_req(mms.memory(), mms.memory2());
400 }
401 }
402 uint limit = ex_map->req();
403 for (uint i = TypeFunc::Parms; i < limit; i++) {
404 // Skip everything in the JVMS after tos. (The ex_oop follows.)
405 if (i == tos) i = ex_jvms->monoff();
406 Node* src = ex_map->in(i);
407 Node* dst = phi_map->in(i);
408 if (src != dst) {
409 PhiNode* phi;
410 if (dst->in(0) != region) {
411 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
412 record_for_igvn(phi);
413 _gvn.set_type(phi, phi->type());
414 phi_map->set_req(i, dst);
415 // Prepare to append interesting stuff onto the new phi:
416 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
417 } else {
418 assert(dst->is_Phi(), "nobody else uses a hidden region");
419 phi = dst->as_Phi();
420 }
421 if (add_multiple && src->in(0) == ex_control) {
422 // Both are phis.
423 add_n_reqs(dst, src);
424 } else {
425 while (dst->req() < region->req()) add_one_req(dst, src);
426 }
427 const Type* srctype = _gvn.type(src);
428 if (phi->type() != srctype) {
429 const Type* dsttype = phi->type()->meet_speculative(srctype);
430 if (phi->type() != dsttype) {
431 phi->set_type(dsttype);
432 _gvn.set_type(phi, dsttype);
433 }
434 }
435 }
436 }
437 phi_map->merge_replaced_nodes_with(ex_map);
438 }
439
440 //--------------------------use_exception_state--------------------------------
441 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
442 if (failing()) { stop(); return top(); }
443 Node* region = phi_map->control();
444 Node* hidden_merge_mark = root();
445 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
446 Node* ex_oop = clear_saved_ex_oop(phi_map);
447 if (region->in(0) == hidden_merge_mark) {
448 // Special marking for internal ex-states. Process the phis now.
449 region->set_req(0, region); // now it's an ordinary region
450 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
451 // Note: Setting the jvms also sets the bci and sp.
452 set_control(_gvn.transform(region));
453 uint tos = jvms()->stkoff() + sp();
454 for (uint i = 1; i < tos; i++) {
455 Node* x = phi_map->in(i);
456 if (x->in(0) == region) {
457 assert(x->is_Phi(), "expected a special phi");
458 phi_map->set_req(i, _gvn.transform(x));
459 }
460 }
461 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
462 Node* x = mms.memory();
463 if (x->in(0) == region) {
464 assert(x->is_Phi(), "nobody else uses a hidden region");
465 mms.set_memory(_gvn.transform(x));
466 }
467 }
468 if (ex_oop->in(0) == region) {
469 assert(ex_oop->is_Phi(), "expected a special phi");
470 ex_oop = _gvn.transform(ex_oop);
471 }
472 } else {
473 set_jvms(phi_map->jvms());
474 }
475
476 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
477 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
478 return ex_oop;
479 }
480
481 //---------------------------------java_bc-------------------------------------
482 Bytecodes::Code GraphKit::java_bc() const {
483 ciMethod* method = this->method();
484 int bci = this->bci();
485 if (method != NULL && bci != InvocationEntryBci)
486 return method->java_code_at_bci(bci);
487 else
488 return Bytecodes::_illegal;
489 }
490
491 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
492 bool must_throw) {
493 // if the exception capability is set, then we will generate code
494 // to check the JavaThread.should_post_on_exceptions flag to see
495 // if we actually need to report exception events (for this
496 // thread). If we don't need to report exception events, we will
497 // take the normal fast path provided by add_exception_events. If
498 // exception event reporting is enabled for this thread, we will
499 // take the uncommon_trap in the BuildCutout below.
500
501 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
502 Node* jthread = _gvn.transform(new ThreadLocalNode());
503 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
504 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
505
506 // Test the should_post_on_exceptions_flag vs. 0
507 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
508 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
509
510 // Branch to slow_path if should_post_on_exceptions_flag was true
511 { BuildCutout unless(this, tst, PROB_MAX);
512 // Do not try anything fancy if we're notifying the VM on every throw.
513 // Cf. case Bytecodes::_athrow in parse2.cpp.
514 uncommon_trap(reason, Deoptimization::Action_none,
515 (ciKlass*)NULL, (char*)NULL, must_throw);
516 }
517
518 }
519
520 //------------------------------builtin_throw----------------------------------
521 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
522 bool must_throw = true;
523
524 if (env()->jvmti_can_post_on_exceptions()) {
525 // check if we must post exception events, take uncommon trap if so
526 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
527 // here if should_post_on_exceptions is false
528 // continue on with the normal codegen
529 }
530
531 // If this particular condition has not yet happened at this
532 // bytecode, then use the uncommon trap mechanism, and allow for
533 // a future recompilation if several traps occur here.
534 // If the throw is hot, try to use a more complicated inline mechanism
535 // which keeps execution inside the compiled code.
536 bool treat_throw_as_hot = false;
537 ciMethodData* md = method()->method_data();
538
539 if (ProfileTraps) {
540 if (too_many_traps(reason)) {
541 treat_throw_as_hot = true;
542 }
543 // (If there is no MDO at all, assume it is early in
544 // execution, and that any deopts are part of the
545 // startup transient, and don't need to be remembered.)
546
547 // Also, if there is a local exception handler, treat all throws
548 // as hot if there has been at least one in this method.
549 if (C->trap_count(reason) != 0
550 && method()->method_data()->trap_count(reason) != 0
551 && has_ex_handler()) {
552 treat_throw_as_hot = true;
553 }
554
555 // TODO: Check if and how profiling can be used for vbox/vunbox
556 if (java_bc() == Bytecodes::_vbox || java_bc() == Bytecodes::_vunbox) {
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 }
594 if (failing()) { stop(); return; } // exception allocation might fail
595 if (ex_obj != NULL) {
596 // Cheat with a preallocated exception object.
597 if (C->log() != NULL)
598 C->log()->elem("hot_throw preallocated='1' reason='%s'",
599 Deoptimization::trap_reason_name(reason));
600 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
601 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
602
603 // Clear the detail message of the preallocated exception object.
604 // Weblogic sometimes mutates the detail message of exceptions
605 // using reflection.
606 int offset = java_lang_Throwable::get_detailMessage_offset();
607 const TypePtr* adr_typ = ex_con->add_offset(offset);
608
609 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
610 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
611 // Conservatively release stores of object references.
612 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
613
614 add_exception_state(make_exception_state(ex_node));
615 return;
616 }
617 }
618
619 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
620 // It won't be much cheaper than bailing to the interp., since we'll
621 // have to pass up all the debug-info, and the runtime will have to
622 // create the stack trace.
623
624 // Usual case: Bail to interpreter.
625 // Reserve the right to recompile if we haven't seen anything yet.
626
627 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
628 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
629 if (treat_throw_as_hot
630 && (method()->method_data()->trap_recompiled_at(bci(), m)
631 || C->too_many_traps(reason))) {
632 // We cannot afford to take more traps here. Suffer in the interpreter.
633 if (C->log() != NULL)
634 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
635 Deoptimization::trap_reason_name(reason),
636 C->trap_count(reason));
637 action = Deoptimization::Action_none;
638 }
639
640 // "must_throw" prunes the JVM state to include only the stack, if there
641 // are no local exception handlers. This should cut down on register
642 // allocation time and code size, by drastically reducing the number
643 // of in-edges on the call to the uncommon trap.
644
645 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
646 }
647
648
649 //----------------------------PreserveJVMState---------------------------------
650 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
651 debug_only(kit->verify_map());
652 _kit = kit;
653 _map = kit->map(); // preserve the map
654 _sp = kit->sp();
655 kit->set_map(clone_map ? kit->clone_map() : NULL);
656 #ifdef ASSERT
657 _bci = kit->bci();
658 Parse* parser = kit->is_Parse();
659 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
660 _block = block;
661 #endif
662 }
663 PreserveJVMState::~PreserveJVMState() {
664 GraphKit* kit = _kit;
665 #ifdef ASSERT
666 assert(kit->bci() == _bci, "bci must not shift");
667 Parse* parser = kit->is_Parse();
668 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
669 assert(block == _block, "block must not shift");
670 #endif
671 kit->set_map(_map);
672 kit->set_sp(_sp);
673 }
674
675
676 //-----------------------------BuildCutout-------------------------------------
677 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
678 : PreserveJVMState(kit)
679 {
680 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
681 SafePointNode* outer_map = _map; // preserved map is caller's
682 SafePointNode* inner_map = kit->map();
683 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
684 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
685 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
686 }
687 BuildCutout::~BuildCutout() {
688 GraphKit* kit = _kit;
689 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
690 }
691
692 //---------------------------PreserveReexecuteState----------------------------
693 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
694 assert(!kit->stopped(), "must call stopped() before");
695 _kit = kit;
696 _sp = kit->sp();
697 _reexecute = kit->jvms()->_reexecute;
698 }
699 PreserveReexecuteState::~PreserveReexecuteState() {
700 if (_kit->stopped()) return;
701 _kit->jvms()->_reexecute = _reexecute;
702 _kit->set_sp(_sp);
703 }
704
705 //------------------------------clone_map--------------------------------------
706 // Implementation of PreserveJVMState
707 //
708 // Only clone_map(...) here. If this function is only used in the
709 // PreserveJVMState class we may want to get rid of this extra
710 // function eventually and do it all there.
711
712 SafePointNode* GraphKit::clone_map() {
713 if (map() == NULL) return NULL;
714
715 // Clone the memory edge first
716 Node* mem = MergeMemNode::make(map()->memory());
717 gvn().set_type_bottom(mem);
718
719 SafePointNode *clonemap = (SafePointNode*)map()->clone();
720 JVMState* jvms = this->jvms();
721 JVMState* clonejvms = jvms->clone_shallow(C);
722 clonemap->set_memory(mem);
723 clonemap->set_jvms(clonejvms);
724 clonejvms->set_map(clonemap);
725 record_for_igvn(clonemap);
726 gvn().set_type_bottom(clonemap);
727 return clonemap;
728 }
729
730
731 //-----------------------------set_map_clone-----------------------------------
732 void GraphKit::set_map_clone(SafePointNode* m) {
733 _map = m;
734 _map = clone_map();
735 _map->set_next_exception(NULL);
736 debug_only(verify_map());
737 }
738
739
740 //----------------------------kill_dead_locals---------------------------------
741 // Detect any locals which are known to be dead, and force them to top.
742 void GraphKit::kill_dead_locals() {
743 // Consult the liveness information for the locals. If any
744 // of them are unused, then they can be replaced by top(). This
745 // should help register allocation time and cut down on the size
746 // of the deoptimization information.
747
748 // This call is made from many of the bytecode handling
749 // subroutines called from the Big Switch in do_one_bytecode.
750 // Every bytecode which might include a slow path is responsible
751 // for killing its dead locals. The more consistent we
752 // are about killing deads, the fewer useless phis will be
753 // constructed for them at various merge points.
754
755 // bci can be -1 (InvocationEntryBci). We return the entry
756 // liveness for the method.
757
758 if (method() == NULL || method()->code_size() == 0) {
759 // We are building a graph for a call to a native method.
760 // All locals are live.
761 return;
762 }
763
764 ResourceMark rm;
765
766 // Consult the liveness information for the locals. If any
767 // of them are unused, then they can be replaced by top(). This
768 // should help register allocation time and cut down on the size
769 // of the deoptimization information.
770 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
771
772 int len = (int)live_locals.size();
773 assert(len <= jvms()->loc_size(), "too many live locals");
774 for (int local = 0; local < len; local++) {
775 if (!live_locals.at(local)) {
776 set_local(local, top());
777 }
778 }
779 }
780
781 #ifdef ASSERT
782 //-------------------------dead_locals_are_killed------------------------------
783 // Return true if all dead locals are set to top in the map.
784 // Used to assert "clean" debug info at various points.
785 bool GraphKit::dead_locals_are_killed() {
786 if (method() == NULL || method()->code_size() == 0) {
787 // No locals need to be dead, so all is as it should be.
788 return true;
789 }
790
791 // Make sure somebody called kill_dead_locals upstream.
792 ResourceMark rm;
793 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
794 if (jvms->loc_size() == 0) continue; // no locals to consult
795 SafePointNode* map = jvms->map();
796 ciMethod* method = jvms->method();
797 int bci = jvms->bci();
798 if (jvms == this->jvms()) {
799 bci = this->bci(); // it might not yet be synched
800 }
801 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
802 int len = (int)live_locals.size();
803 if (!live_locals.is_valid() || len == 0)
804 // This method is trivial, or is poisoned by a breakpoint.
805 return true;
806 assert(len == jvms->loc_size(), "live map consistent with locals map");
807 for (int local = 0; local < len; local++) {
808 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
809 if (PrintMiscellaneous && (Verbose || WizardMode)) {
810 tty->print_cr("Zombie local %d: ", local);
811 jvms->dump();
812 }
813 return false;
814 }
815 }
816 }
817 return true;
818 }
819
820 #endif //ASSERT
821
822 // Helper function for enforcing certain bytecodes to reexecute if
823 // deoptimization happens
824 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
825 ciMethod* cur_method = jvms->method();
826 int cur_bci = jvms->bci();
827 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
828 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
829 return Interpreter::bytecode_should_reexecute(code) ||
830 is_anewarray && code == Bytecodes::_multianewarray;
831 // Reexecute _multianewarray bytecode which was replaced with
832 // sequence of [a]newarray. See Parse::do_multianewarray().
833 //
834 // Note: interpreter should not have it set since this optimization
835 // is limited by dimensions and guarded by flag so in some cases
836 // multianewarray() runtime calls will be generated and
837 // the bytecode should not be reexecutes (stack will not be reset).
838 } else
839 return false;
840 }
841
842 // Helper function for adding JVMState and debug information to node
843 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
844 // Add the safepoint edges to the call (or other safepoint).
845
846 // Make sure dead locals are set to top. This
847 // should help register allocation time and cut down on the size
848 // of the deoptimization information.
849 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
850
851 // Walk the inline list to fill in the correct set of JVMState's
852 // Also fill in the associated edges for each JVMState.
853
854 // If the bytecode needs to be reexecuted we need to put
855 // the arguments back on the stack.
856 const bool should_reexecute = jvms()->should_reexecute();
857 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
858
859 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
860 // undefined if the bci is different. This is normal for Parse but it
861 // should not happen for LibraryCallKit because only one bci is processed.
862 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
863 "in LibraryCallKit the reexecute bit should not change");
864
865 // If we are guaranteed to throw, we can prune everything but the
866 // input to the current bytecode.
867 bool can_prune_locals = false;
868 uint stack_slots_not_pruned = 0;
869 int inputs = 0, depth = 0;
870 if (must_throw) {
871 assert(method() == youngest_jvms->method(), "sanity");
872 if (compute_stack_effects(inputs, depth)) {
873 can_prune_locals = true;
874 stack_slots_not_pruned = inputs;
875 }
876 }
877
878 if (env()->should_retain_local_variables()) {
879 // At any safepoint, this method can get breakpointed, which would
880 // then require an immediate deoptimization.
881 can_prune_locals = false; // do not prune locals
882 stack_slots_not_pruned = 0;
883 }
884
885 // do not scribble on the input jvms
886 JVMState* out_jvms = youngest_jvms->clone_deep(C);
887 call->set_jvms(out_jvms); // Start jvms list for call node
888
889 // For a known set of bytecodes, the interpreter should reexecute them if
890 // deoptimization happens. We set the reexecute state for them here
891 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
892 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
893 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
894 }
895
896 // Presize the call:
897 DEBUG_ONLY(uint non_debug_edges = call->req());
898 call->add_req_batch(top(), youngest_jvms->debug_depth());
899 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
900
901 // Set up edges so that the call looks like this:
902 // Call [state:] ctl io mem fptr retadr
903 // [parms:] parm0 ... parmN
904 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
905 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
906 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
907 // Note that caller debug info precedes callee debug info.
908
909 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
910 uint debug_ptr = call->req();
911
912 // Loop over the map input edges associated with jvms, add them
913 // to the call node, & reset all offsets to match call node array.
914 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
915 uint debug_end = debug_ptr;
916 uint debug_start = debug_ptr - in_jvms->debug_size();
917 debug_ptr = debug_start; // back up the ptr
918
919 uint p = debug_start; // walks forward in [debug_start, debug_end)
920 uint j, k, l;
921 SafePointNode* in_map = in_jvms->map();
922 out_jvms->set_map(call);
923
924 if (can_prune_locals) {
925 assert(in_jvms->method() == out_jvms->method(), "sanity");
926 // If the current throw can reach an exception handler in this JVMS,
927 // then we must keep everything live that can reach that handler.
928 // As a quick and dirty approximation, we look for any handlers at all.
929 if (in_jvms->method()->has_exception_handlers()) {
930 can_prune_locals = false;
931 }
932 }
933
934 // Add the Locals
935 k = in_jvms->locoff();
936 l = in_jvms->loc_size();
937 out_jvms->set_locoff(p);
938 if (!can_prune_locals) {
939 for (j = 0; j < l; j++)
940 call->set_req(p++, in_map->in(k+j));
941 } else {
942 p += l; // already set to top above by add_req_batch
943 }
944
945 // Add the Expression Stack
946 k = in_jvms->stkoff();
947 l = in_jvms->sp();
948 out_jvms->set_stkoff(p);
949 if (!can_prune_locals) {
950 for (j = 0; j < l; j++)
951 call->set_req(p++, in_map->in(k+j));
952 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
953 // Divide stack into {S0,...,S1}, where S0 is set to top.
954 uint s1 = stack_slots_not_pruned;
955 stack_slots_not_pruned = 0; // for next iteration
956 if (s1 > l) s1 = l;
957 uint s0 = l - s1;
958 p += s0; // skip the tops preinstalled by add_req_batch
959 for (j = s0; j < l; j++)
960 call->set_req(p++, in_map->in(k+j));
961 } else {
962 p += l; // already set to top above by add_req_batch
963 }
964
965 // Add the Monitors
966 k = in_jvms->monoff();
967 l = in_jvms->mon_size();
968 out_jvms->set_monoff(p);
969 for (j = 0; j < l; j++)
970 call->set_req(p++, in_map->in(k+j));
971
972 // Copy any scalar object fields.
973 k = in_jvms->scloff();
974 l = in_jvms->scl_size();
975 out_jvms->set_scloff(p);
976 for (j = 0; j < l; j++)
977 call->set_req(p++, in_map->in(k+j));
978
979 // Finish the new jvms.
980 out_jvms->set_endoff(p);
981
982 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
983 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
984 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
985 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
986 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
987 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
988
989 // Update the two tail pointers in parallel.
990 out_jvms = out_jvms->caller();
991 in_jvms = in_jvms->caller();
992 }
993
994 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
995
996 // Test the correctness of JVMState::debug_xxx accessors:
997 assert(call->jvms()->debug_start() == non_debug_edges, "");
998 assert(call->jvms()->debug_end() == call->req(), "");
999 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1000 }
1001
1002 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1003 Bytecodes::Code code = java_bc();
1004 if (code == Bytecodes::_wide) {
1005 code = method()->java_code_at_bci(bci() + 1);
1006 }
1007
1008 BasicType rtype = T_ILLEGAL;
1009 int rsize = 0;
1010
1011 if (code != Bytecodes::_illegal) {
1012 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1013 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1014 if (rtype < T_CONFLICT)
1015 rsize = type2size[rtype];
1016 }
1017
1018 switch (code) {
1019 case Bytecodes::_illegal:
1020 return false;
1021
1022 case Bytecodes::_ldc:
1023 case Bytecodes::_ldc_w:
1024 case Bytecodes::_ldc2_w:
1025 inputs = 0;
1026 break;
1027
1028 case Bytecodes::_dup: inputs = 1; break;
1029 case Bytecodes::_dup_x1: inputs = 2; break;
1030 case Bytecodes::_dup_x2: inputs = 3; break;
1031 case Bytecodes::_dup2: inputs = 2; break;
1032 case Bytecodes::_dup2_x1: inputs = 3; break;
1033 case Bytecodes::_dup2_x2: inputs = 4; break;
1034 case Bytecodes::_swap: inputs = 2; break;
1035 case Bytecodes::_arraylength: inputs = 1; break;
1036
1037 case Bytecodes::_getstatic:
1038 case Bytecodes::_putstatic:
1039 case Bytecodes::_getfield:
1040 case Bytecodes::_vgetfield:
1041 case Bytecodes::_putfield:
1042 {
1043 bool ignored_will_link;
1044 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1045 int size = field->type()->size();
1046 bool is_get = (depth >= 0), is_static = (depth & 1);
1047 inputs = (is_static ? 0 : 1);
1048 if (is_get) {
1049 depth = size - inputs;
1050 } else {
1051 inputs += size; // putxxx pops the value from the stack
1052 depth = - inputs;
1053 }
1054 }
1055 break;
1056
1057 case Bytecodes::_invokevirtual:
1058 case Bytecodes::_invokedirect:
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::_vnew: {
1086 // vnew pops the values from the stack
1087 ciValueKlass* vk = method()->holder()->as_value_klass();
1088 inputs = vk->field_size();
1089 depth = rsize - inputs;
1090 break;
1091 }
1092 case Bytecodes::_vwithfield: {
1093 bool ignored_will_link;
1094 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1095 int size = field->type()->size();
1096 inputs = size+1;
1097 depth = rsize - inputs;
1098 break;
1099 }
1100
1101 case Bytecodes::_ireturn:
1102 case Bytecodes::_lreturn:
1103 case Bytecodes::_freturn:
1104 case Bytecodes::_dreturn:
1105 case Bytecodes::_areturn:
1106 case Bytecodes::_vreturn:
1107 assert(rsize = -depth, "");
1108 inputs = rsize;
1109 break;
1110
1111 case Bytecodes::_jsr:
1112 case Bytecodes::_jsr_w:
1113 inputs = 0;
1114 depth = 1; // S.B. depth=1, not zero
1115 break;
1116
1117 default:
1118 // bytecode produces a typed result
1119 inputs = rsize - depth;
1120 assert(inputs >= 0, "");
1121 break;
1122 }
1123
1124 #ifdef ASSERT
1125 // spot check
1126 int outputs = depth + inputs;
1127 assert(outputs >= 0, "sanity");
1128 switch (code) {
1129 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1130 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1131 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1132 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1133 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1134 }
1135 #endif //ASSERT
1136
1137 return true;
1138 }
1139
1140
1141
1142 //------------------------------basic_plus_adr---------------------------------
1143 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1144 // short-circuit a common case
1145 if (offset == intcon(0)) return ptr;
1146 return _gvn.transform( new AddPNode(base, ptr, offset) );
1147 }
1148
1149 Node* GraphKit::ConvI2L(Node* offset) {
1150 // short-circuit a common case
1151 jint offset_con = find_int_con(offset, Type::OffsetBot);
1152 if (offset_con != Type::OffsetBot) {
1153 return longcon((jlong) offset_con);
1154 }
1155 return _gvn.transform( new ConvI2LNode(offset));
1156 }
1157
1158 Node* GraphKit::ConvI2UL(Node* offset) {
1159 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1160 if (offset_con != (juint) Type::OffsetBot) {
1161 return longcon((julong) offset_con);
1162 }
1163 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1164 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1165 return _gvn.transform( new AndLNode(conv, mask) );
1166 }
1167
1168 Node* GraphKit::ConvL2I(Node* offset) {
1169 // short-circuit a common case
1170 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1171 if (offset_con != (jlong)Type::OffsetBot) {
1172 return intcon((int) offset_con);
1173 }
1174 return _gvn.transform( new ConvL2INode(offset));
1175 }
1176
1177 //-------------------------load_object_klass-----------------------------------
1178 Node* GraphKit::load_object_klass(Node* obj) {
1179 // Special-case a fresh allocation to avoid building nodes:
1180 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1181 if (akls != NULL) return akls;
1182 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1183 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1184 }
1185
1186 //-------------------------load_array_length-----------------------------------
1187 Node* GraphKit::load_array_length(Node* array) {
1188 // Special-case a fresh allocation to avoid building nodes:
1189 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1190 Node *alen;
1191 if (alloc == NULL) {
1192 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1193 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1194 } else {
1195 alen = alloc->Ideal_length();
1196 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1197 if (ccast != alen) {
1198 alen = _gvn.transform(ccast);
1199 }
1200 }
1201 return alen;
1202 }
1203
1204 //------------------------------do_null_check----------------------------------
1205 // Helper function to do a NULL pointer check. Returned value is
1206 // the incoming address with NULL casted away. You are allowed to use the
1207 // not-null value only if you are control dependent on the test.
1208 #ifndef PRODUCT
1209 extern int explicit_null_checks_inserted,
1210 explicit_null_checks_elided;
1211 #endif
1212 Node* GraphKit::null_check_common(Node* value, BasicType type,
1213 // optional arguments for variations:
1214 bool assert_null,
1215 Node* *null_control,
1216 bool speculative) {
1217 assert(!assert_null || null_control == NULL, "not both at once");
1218 if (stopped()) return top();
1219 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1220 // For some performance testing, we may wish to suppress null checking.
1221 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1222 return value;
1223 }
1224 NOT_PRODUCT(explicit_null_checks_inserted++);
1225
1226 // Construct NULL check
1227 Node *chk = NULL;
1228 switch(type) {
1229 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1230 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1231 case T_ARRAY : // fall through
1232 type = T_OBJECT; // simplify further tests
1233 case T_OBJECT : {
1234 const Type *t = _gvn.type( value );
1235
1236 const TypeOopPtr* tp = t->isa_oopptr();
1237 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1238 // Only for do_null_check, not any of its siblings:
1239 && !assert_null && null_control == NULL) {
1240 // Usually, any field access or invocation on an unloaded oop type
1241 // will simply fail to link, since the statically linked class is
1242 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1243 // the static class is loaded but the sharper oop type is not.
1244 // Rather than checking for this obscure case in lots of places,
1245 // we simply observe that a null check on an unloaded class
1246 // will always be followed by a nonsense operation, so we
1247 // can just issue the uncommon trap here.
1248 // Our access to the unloaded class will only be correct
1249 // after it has been loaded and initialized, which requires
1250 // a trip through the interpreter.
1251 #ifndef PRODUCT
1252 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1253 #endif
1254 uncommon_trap(Deoptimization::Reason_unloaded,
1255 Deoptimization::Action_reinterpret,
1256 tp->klass(), "!loaded");
1257 return top();
1258 }
1259
1260 if (assert_null) {
1261 // See if the type is contained in NULL_PTR.
1262 // If so, then the value is already null.
1263 if (t->higher_equal(TypePtr::NULL_PTR)) {
1264 NOT_PRODUCT(explicit_null_checks_elided++);
1265 return value; // Elided null assert quickly!
1266 }
1267 } else {
1268 // See if mixing in the NULL pointer changes type.
1269 // If so, then the NULL pointer was not allowed in the original
1270 // type. In other words, "value" was not-null.
1271 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1272 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1273 NOT_PRODUCT(explicit_null_checks_elided++);
1274 return value; // Elided null check quickly!
1275 }
1276 }
1277 chk = new CmpPNode( value, null() );
1278 break;
1279 }
1280
1281 default:
1282 fatal("unexpected type: %s", type2name(type));
1283 }
1284 assert(chk != NULL, "sanity check");
1285 chk = _gvn.transform(chk);
1286
1287 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1288 BoolNode *btst = new BoolNode( chk, btest);
1289 Node *tst = _gvn.transform( btst );
1290
1291 //-----------
1292 // if peephole optimizations occurred, a prior test existed.
1293 // If a prior test existed, maybe it dominates as we can avoid this test.
1294 if (tst != btst && type == T_OBJECT) {
1295 // At this point we want to scan up the CFG to see if we can
1296 // find an identical test (and so avoid this test altogether).
1297 Node *cfg = control();
1298 int depth = 0;
1299 while( depth < 16 ) { // Limit search depth for speed
1300 if( cfg->Opcode() == Op_IfTrue &&
1301 cfg->in(0)->in(1) == tst ) {
1302 // Found prior test. Use "cast_not_null" to construct an identical
1303 // CastPP (and hence hash to) as already exists for the prior test.
1304 // Return that casted value.
1305 if (assert_null) {
1306 replace_in_map(value, null());
1307 return null(); // do not issue the redundant test
1308 }
1309 Node *oldcontrol = control();
1310 set_control(cfg);
1311 Node *res = cast_not_null(value);
1312 set_control(oldcontrol);
1313 NOT_PRODUCT(explicit_null_checks_elided++);
1314 return res;
1315 }
1316 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1317 if (cfg == NULL) break; // Quit at region nodes
1318 depth++;
1319 }
1320 }
1321
1322 //-----------
1323 // Branch to failure if null
1324 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1325 Deoptimization::DeoptReason reason;
1326 if (assert_null) {
1327 reason = Deoptimization::Reason_null_assert;
1328 } else if (type == T_OBJECT) {
1329 reason = Deoptimization::reason_null_check(speculative);
1330 } else {
1331 reason = Deoptimization::Reason_div0_check;
1332 }
1333 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1334 // ciMethodData::has_trap_at will return a conservative -1 if any
1335 // must-be-null assertion has failed. This could cause performance
1336 // problems for a method after its first do_null_assert failure.
1337 // Consider using 'Reason_class_check' instead?
1338
1339 // To cause an implicit null check, we set the not-null probability
1340 // to the maximum (PROB_MAX). For an explicit check the probability
1341 // is set to a smaller value.
1342 if (null_control != NULL || too_many_traps(reason)) {
1343 // probability is less likely
1344 ok_prob = PROB_LIKELY_MAG(3);
1345 } else if (!assert_null &&
1346 (ImplicitNullCheckThreshold > 0) &&
1347 method() != NULL &&
1348 (method()->method_data()->trap_count(reason)
1349 >= (uint)ImplicitNullCheckThreshold)) {
1350 ok_prob = PROB_LIKELY_MAG(3);
1351 }
1352
1353 if (null_control != NULL) {
1354 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1355 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1356 set_control( _gvn.transform( new IfTrueNode(iff)));
1357 #ifndef PRODUCT
1358 if (null_true == top()) {
1359 explicit_null_checks_elided++;
1360 }
1361 #endif
1362 (*null_control) = null_true;
1363 } else {
1364 BuildCutout unless(this, tst, ok_prob);
1365 // Check for optimizer eliding test at parse time
1366 if (stopped()) {
1367 // Failure not possible; do not bother making uncommon trap.
1368 NOT_PRODUCT(explicit_null_checks_elided++);
1369 } else if (assert_null) {
1370 uncommon_trap(reason,
1371 Deoptimization::Action_make_not_entrant,
1372 NULL, "assert_null");
1373 } else {
1374 replace_in_map(value, zerocon(type));
1375 builtin_throw(reason);
1376 }
1377 }
1378
1379 // Must throw exception, fall-thru not possible?
1380 if (stopped()) {
1381 return top(); // No result
1382 }
1383
1384 if (assert_null) {
1385 // Cast obj to null on this path.
1386 replace_in_map(value, zerocon(type));
1387 return zerocon(type);
1388 }
1389
1390 // Cast obj to not-null on this path, if there is no null_control.
1391 // (If there is a null_control, a non-null value may come back to haunt us.)
1392 if (type == T_OBJECT) {
1393 Node* cast = cast_not_null(value, false);
1394 if (null_control == NULL || (*null_control) == top())
1395 replace_in_map(value, cast);
1396 value = cast;
1397 }
1398
1399 return value;
1400 }
1401
1402
1403 //------------------------------cast_not_null----------------------------------
1404 // Cast obj to not-null on this path
1405 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1406 const Type *t = _gvn.type(obj);
1407 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1408 // Object is already not-null?
1409 if( t == t_not_null ) return obj;
1410
1411 Node *cast = new CastPPNode(obj,t_not_null);
1412 cast->init_req(0, control());
1413 cast = _gvn.transform( cast );
1414
1415 // Scan for instances of 'obj' in the current JVM mapping.
1416 // These instances are known to be not-null after the test.
1417 if (do_replace_in_map)
1418 replace_in_map(obj, cast);
1419
1420 return cast; // Return casted value
1421 }
1422
1423
1424 //--------------------------replace_in_map-------------------------------------
1425 void GraphKit::replace_in_map(Node* old, Node* neww) {
1426 if (old == neww) {
1427 return;
1428 }
1429
1430 map()->replace_edge(old, neww);
1431
1432 // Note: This operation potentially replaces any edge
1433 // on the map. This includes locals, stack, and monitors
1434 // of the current (innermost) JVM state.
1435
1436 // don't let inconsistent types from profiling escape this
1437 // method
1438
1439 const Type* told = _gvn.type(old);
1440 const Type* tnew = _gvn.type(neww);
1441
1442 if (!tnew->higher_equal(told)) {
1443 return;
1444 }
1445
1446 map()->record_replaced_node(old, neww);
1447 }
1448
1449
1450 //=============================================================================
1451 //--------------------------------memory---------------------------------------
1452 Node* GraphKit::memory(uint alias_idx) {
1453 MergeMemNode* mem = merged_memory();
1454 Node* p = mem->memory_at(alias_idx);
1455 _gvn.set_type(p, Type::MEMORY); // must be mapped
1456 return p;
1457 }
1458
1459 //-----------------------------reset_memory------------------------------------
1460 Node* GraphKit::reset_memory() {
1461 Node* mem = map()->memory();
1462 // do not use this node for any more parsing!
1463 debug_only( map()->set_memory((Node*)NULL) );
1464 return _gvn.transform( mem );
1465 }
1466
1467 //------------------------------set_all_memory---------------------------------
1468 void GraphKit::set_all_memory(Node* newmem) {
1469 Node* mergemem = MergeMemNode::make(newmem);
1470 gvn().set_type_bottom(mergemem);
1471 map()->set_memory(mergemem);
1472 }
1473
1474 //------------------------------set_all_memory_call----------------------------
1475 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1476 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1477 set_all_memory(newmem);
1478 }
1479
1480 //=============================================================================
1481 //
1482 // parser factory methods for MemNodes
1483 //
1484 // These are layered on top of the factory methods in LoadNode and StoreNode,
1485 // and integrate with the parser's memory state and _gvn engine.
1486 //
1487
1488 // factory methods in "int adr_idx"
1489 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1490 int adr_idx,
1491 MemNode::MemOrd mo,
1492 LoadNode::ControlDependency control_dependency,
1493 bool require_atomic_access,
1494 bool unaligned,
1495 bool mismatched) {
1496 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1497 const TypePtr* adr_type = NULL; // debug-mode-only argument
1498 debug_only(adr_type = C->get_adr_type(adr_idx));
1499 Node* mem = memory(adr_idx);
1500 Node* ld;
1501 if (require_atomic_access && bt == T_LONG) {
1502 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1503 } else if (require_atomic_access && bt == T_DOUBLE) {
1504 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1505 } else {
1506 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1507 }
1508 ld = _gvn.transform(ld);
1509
1510 if (bt == T_VALUETYPE) {
1511 // Load value type from oop
1512 ld = ValueTypeNode::make(gvn(), map()->memory(), ld);
1513 } else if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1514 // Improve graph before escape analysis and boxing elimination.
1515 record_for_igvn(ld);
1516 }
1517 return ld;
1518 }
1519
1520 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1521 int adr_idx,
1522 MemNode::MemOrd mo,
1523 bool require_atomic_access,
1524 bool unaligned,
1525 bool mismatched) {
1526 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1527 const TypePtr* adr_type = NULL;
1528 debug_only(adr_type = C->get_adr_type(adr_idx));
1529 Node *mem = memory(adr_idx);
1530 Node* st;
1531 if (require_atomic_access && bt == T_LONG) {
1532 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1533 } else if (require_atomic_access && bt == T_DOUBLE) {
1534 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1535 } else {
1536 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1537 }
1538 if (unaligned) {
1539 st->as_Store()->set_unaligned_access();
1540 }
1541 if (mismatched) {
1542 st->as_Store()->set_mismatched_access();
1543 }
1544 st = _gvn.transform(st);
1545 set_memory(st, adr_idx);
1546 // Back-to-back stores can only remove intermediate store with DU info
1547 // so push on worklist for optimizer.
1548 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1549 record_for_igvn(st);
1550
1551 return st;
1552 }
1553
1554
1555 void GraphKit::pre_barrier(bool do_load,
1556 Node* ctl,
1557 Node* obj,
1558 Node* adr,
1559 uint adr_idx,
1560 Node* val,
1561 const TypeOopPtr* val_type,
1562 Node* pre_val,
1563 BasicType bt) {
1564
1565 BarrierSet* bs = Universe::heap()->barrier_set();
1566 set_control(ctl);
1567 switch (bs->kind()) {
1568 case BarrierSet::G1SATBCTLogging:
1569 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1570 break;
1571
1572 case BarrierSet::CardTableForRS:
1573 case BarrierSet::CardTableExtension:
1574 case BarrierSet::ModRef:
1575 break;
1576
1577 default :
1578 ShouldNotReachHere();
1579
1580 }
1581 }
1582
1583 bool GraphKit::can_move_pre_barrier() const {
1584 BarrierSet* bs = Universe::heap()->barrier_set();
1585 switch (bs->kind()) {
1586 case BarrierSet::G1SATBCTLogging:
1587 return true; // Can move it if no safepoint
1588
1589 case BarrierSet::CardTableForRS:
1590 case BarrierSet::CardTableExtension:
1591 case BarrierSet::ModRef:
1592 return true; // There is no pre-barrier
1593
1594 default :
1595 ShouldNotReachHere();
1596 }
1597 return false;
1598 }
1599
1600 void GraphKit::post_barrier(Node* ctl,
1601 Node* store,
1602 Node* obj,
1603 Node* adr,
1604 uint adr_idx,
1605 Node* val,
1606 BasicType bt,
1607 bool use_precise) {
1608 BarrierSet* bs = Universe::heap()->barrier_set();
1609 set_control(ctl);
1610 switch (bs->kind()) {
1611 case BarrierSet::G1SATBCTLogging:
1612 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1613 break;
1614
1615 case BarrierSet::CardTableForRS:
1616 case BarrierSet::CardTableExtension:
1617 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1618 break;
1619
1620 case BarrierSet::ModRef:
1621 break;
1622
1623 default :
1624 ShouldNotReachHere();
1625
1626 }
1627 }
1628
1629 Node* GraphKit::store_oop(Node* ctl,
1630 Node* obj,
1631 Node* adr,
1632 const TypePtr* adr_type,
1633 Node* val,
1634 const TypeOopPtr* val_type,
1635 BasicType bt,
1636 bool use_precise,
1637 MemNode::MemOrd mo,
1638 bool mismatched) {
1639 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1640 // could be delayed during Parse (for example, in adjust_map_after_if()).
1641 // Execute transformation here to avoid barrier generation in such case.
1642 if (_gvn.type(val) == TypePtr::NULL_PTR)
1643 val = _gvn.makecon(TypePtr::NULL_PTR);
1644
1645 set_control(ctl);
1646 if (stopped()) return top(); // Dead path ?
1647
1648 assert(bt == T_OBJECT || bt == T_VALUETYPE, "sanity");
1649 assert(val != NULL, "not dead path");
1650 uint adr_idx = C->get_alias_index(adr_type);
1651 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1652
1653 if (bt == T_VALUETYPE) {
1654 // Allocate value type and store oop
1655 val = val->as_ValueType()->store_to_memory(this);
1656 }
1657
1658 pre_barrier(true /* do_load */,
1659 control(), obj, adr, adr_idx, val, val_type,
1660 NULL /* pre_val */,
1661 bt);
1662
1663 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1664 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1665 return store;
1666 }
1667
1668 // Could be an array or object we don't know at compile time (unsafe ref.)
1669 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1670 Node* obj, // containing obj
1671 Node* adr, // actual adress to store val at
1672 const TypePtr* adr_type,
1673 Node* val,
1674 BasicType bt,
1675 MemNode::MemOrd mo,
1676 bool mismatched) {
1677 Compile::AliasType* at = C->alias_type(adr_type);
1678 const TypeOopPtr* val_type = NULL;
1679 if (adr_type->isa_instptr()) {
1680 if (at->field() != NULL) {
1681 // known field. This code is a copy of the do_put_xxx logic.
1682 ciField* field = at->field();
1683 if (!field->type()->is_loaded()) {
1684 val_type = TypeInstPtr::BOTTOM;
1685 } else {
1686 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1687 }
1688 }
1689 } else if (adr_type->isa_aryptr()) {
1690 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1691 }
1692 if (val_type == NULL) {
1693 val_type = TypeInstPtr::BOTTOM;
1694 }
1695 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1696 }
1697
1698
1699 //-------------------------array_element_address-------------------------
1700 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1701 const TypeInt* sizetype, Node* ctrl) {
1702 uint shift = exact_log2(type2aelembytes(elembt));
1703 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1704 if (arytype_klass->is_value_array_klass()) {
1705 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass();
1706 shift = vak->log2_element_size();
1707 }
1708 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1709
1710 // short-circuit a common case (saves lots of confusing waste motion)
1711 jint idx_con = find_int_con(idx, -1);
1712 if (idx_con >= 0) {
1713 intptr_t offset = header + ((intptr_t)idx_con << shift);
1714 return basic_plus_adr(ary, offset);
1715 }
1716
1717 // must be correct type for alignment purposes
1718 Node* base = basic_plus_adr(ary, header);
1719 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1720 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1721 return basic_plus_adr(ary, base, scale);
1722 }
1723
1724 //-------------------------load_array_element-------------------------
1725 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1726 const Type* elemtype = arytype->elem();
1727 BasicType elembt = elemtype->array_element_basic_type();
1728 assert(elembt != T_VALUETYPE, "value types are not supported by this method");
1729 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1730 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1731 return ld;
1732 }
1733
1734 //-------------------------set_arguments_for_java_call-------------------------
1735 // Arguments (pre-popped from the stack) are taken from the JVMS.
1736 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1737 // Add the call arguments:
1738 uint nargs = call->method()->arg_size();
1739 for (uint i = 0, idx = 0; i < nargs; i++) {
1740 Node* arg = argument(i);
1741 if (ValueTypePassFieldsAsArgs) {
1742 if (arg->is_ValueType()) {
1743 ValueTypeNode* vt = arg->as_ValueType();
1744 // We don't pass value type arguments by reference but instead
1745 // pass each field of the value type
1746 idx += vt->set_arguments_for_java_call(call, idx + TypeFunc::Parms, *this);
1747 } else {
1748 call->init_req(idx + TypeFunc::Parms, arg);
1749 idx++;
1750 }
1751 } else {
1752 if (arg->is_ValueType()) {
1753 // Pass value type argument via oop to callee
1754 arg = arg->as_ValueType()->store_to_memory(this);
1755 }
1756 call->init_req(i + TypeFunc::Parms, arg);
1757 }
1758 }
1759 }
1760
1761 //---------------------------set_edges_for_java_call---------------------------
1762 // Connect a newly created call into the current JVMS.
1763 // A return value node (if any) is returned from set_edges_for_java_call.
1764 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1765
1766 // Add the predefined inputs:
1767 call->init_req( TypeFunc::Control, control() );
1768 call->init_req( TypeFunc::I_O , i_o() );
1769 call->init_req( TypeFunc::Memory , reset_memory() );
1770 call->init_req( TypeFunc::FramePtr, frameptr() );
1771 call->init_req( TypeFunc::ReturnAdr, top() );
1772
1773 add_safepoint_edges(call, must_throw);
1774
1775 Node* xcall = _gvn.transform(call);
1776
1777 if (xcall == top()) {
1778 set_control(top());
1779 return;
1780 }
1781 assert(xcall == call, "call identity is stable");
1782
1783 // Re-use the current map to produce the result.
1784
1785 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1786 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1787 set_all_memory_call(xcall, separate_io_proj);
1788
1789 //return xcall; // no need, caller already has it
1790 }
1791
1792 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1793 if (stopped()) return top(); // maybe the call folded up?
1794
1795 // Capture the return value, if any.
1796 Node* ret;
1797 if (call->method() == NULL ||
1798 call->method()->return_type()->basic_type() == T_VOID)
1799 ret = top();
1800 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1801
1802 // Note: Since any out-of-line call can produce an exception,
1803 // we always insert an I_O projection from the call into the result.
1804
1805 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1806
1807 if (separate_io_proj) {
1808 // The caller requested separate projections be used by the fall
1809 // through and exceptional paths, so replace the projections for
1810 // the fall through path.
1811 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1812 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1813 }
1814 return ret;
1815 }
1816
1817 //--------------------set_predefined_input_for_runtime_call--------------------
1818 // Reading and setting the memory state is way conservative here.
1819 // The real problem is that I am not doing real Type analysis on memory,
1820 // so I cannot distinguish card mark stores from other stores. Across a GC
1821 // point the Store Barrier and the card mark memory has to agree. I cannot
1822 // have a card mark store and its barrier split across the GC point from
1823 // either above or below. Here I get that to happen by reading ALL of memory.
1824 // A better answer would be to separate out card marks from other memory.
1825 // For now, return the input memory state, so that it can be reused
1826 // after the call, if this call has restricted memory effects.
1827 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1828 // Set fixed predefined input arguments
1829 Node* memory = reset_memory();
1830 call->init_req( TypeFunc::Control, control() );
1831 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1832 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1833 call->init_req( TypeFunc::FramePtr, frameptr() );
1834 call->init_req( TypeFunc::ReturnAdr, top() );
1835 return memory;
1836 }
1837
1838 //-------------------set_predefined_output_for_runtime_call--------------------
1839 // Set control and memory (not i_o) from the call.
1840 // If keep_mem is not NULL, use it for the output state,
1841 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1842 // If hook_mem is NULL, this call produces no memory effects at all.
1843 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1844 // then only that memory slice is taken from the call.
1845 // In the last case, we must put an appropriate memory barrier before
1846 // the call, so as to create the correct anti-dependencies on loads
1847 // preceding the call.
1848 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1849 Node* keep_mem,
1850 const TypePtr* hook_mem) {
1851 // no i/o
1852 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1853 if (keep_mem) {
1854 // First clone the existing memory state
1855 set_all_memory(keep_mem);
1856 if (hook_mem != NULL) {
1857 // Make memory for the call
1858 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1859 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1860 // We also use hook_mem to extract specific effects from arraycopy stubs.
1861 set_memory(mem, hook_mem);
1862 }
1863 // ...else the call has NO memory effects.
1864
1865 // Make sure the call advertises its memory effects precisely.
1866 // This lets us build accurate anti-dependences in gcm.cpp.
1867 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1868 "call node must be constructed correctly");
1869 } else {
1870 assert(hook_mem == NULL, "");
1871 // This is not a "slow path" call; all memory comes from the call.
1872 set_all_memory_call(call);
1873 }
1874 }
1875
1876
1877 // Replace the call with the current state of the kit.
1878 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1879 JVMState* ejvms = NULL;
1880 if (has_exceptions()) {
1881 ejvms = transfer_exceptions_into_jvms();
1882 }
1883
1884 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1885 ReplacedNodes replaced_nodes_exception;
1886 Node* ex_ctl = top();
1887
1888 SafePointNode* final_state = stop();
1889
1890 // Find all the needed outputs of this call
1891 CallProjections callprojs;
1892 call->extract_projections(&callprojs, true);
1893
1894 Node* init_mem = call->in(TypeFunc::Memory);
1895 Node* final_mem = final_state->in(TypeFunc::Memory);
1896 Node* final_ctl = final_state->in(TypeFunc::Control);
1897 Node* final_io = final_state->in(TypeFunc::I_O);
1898
1899 // Replace all the old call edges with the edges from the inlining result
1900 if (callprojs.fallthrough_catchproj != NULL) {
1901 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1902 }
1903 if (callprojs.fallthrough_memproj != NULL) {
1904 if (final_mem->is_MergeMem()) {
1905 // Parser's exits MergeMem was not transformed but may be optimized
1906 final_mem = _gvn.transform(final_mem);
1907 }
1908 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1909 }
1910 if (callprojs.fallthrough_ioproj != NULL) {
1911 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1912 }
1913
1914 // Replace the result with the new result if it exists and is used
1915 if (callprojs.resproj != NULL && result != NULL) {
1916 C->gvn_replace_by(callprojs.resproj, result);
1917 }
1918
1919 if (ejvms == NULL) {
1920 // No exception edges to simply kill off those paths
1921 if (callprojs.catchall_catchproj != NULL) {
1922 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1923 }
1924 if (callprojs.catchall_memproj != NULL) {
1925 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1926 }
1927 if (callprojs.catchall_ioproj != NULL) {
1928 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1929 }
1930 // Replace the old exception object with top
1931 if (callprojs.exobj != NULL) {
1932 C->gvn_replace_by(callprojs.exobj, C->top());
1933 }
1934 } else {
1935 GraphKit ekit(ejvms);
1936
1937 // Load my combined exception state into the kit, with all phis transformed:
1938 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1939 replaced_nodes_exception = ex_map->replaced_nodes();
1940
1941 Node* ex_oop = ekit.use_exception_state(ex_map);
1942
1943 if (callprojs.catchall_catchproj != NULL) {
1944 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1945 ex_ctl = ekit.control();
1946 }
1947 if (callprojs.catchall_memproj != NULL) {
1948 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1949 }
1950 if (callprojs.catchall_ioproj != NULL) {
1951 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1952 }
1953
1954 // Replace the old exception object with the newly created one
1955 if (callprojs.exobj != NULL) {
1956 C->gvn_replace_by(callprojs.exobj, ex_oop);
1957 }
1958 }
1959
1960 // Disconnect the call from the graph
1961 call->disconnect_inputs(NULL, C);
1962 C->gvn_replace_by(call, C->top());
1963
1964 // Clean up any MergeMems that feed other MergeMems since the
1965 // optimizer doesn't like that.
1966 if (final_mem->is_MergeMem()) {
1967 Node_List wl;
1968 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1969 Node* m = i.get();
1970 if (m->is_MergeMem() && !wl.contains(m)) {
1971 wl.push(m);
1972 }
1973 }
1974 while (wl.size() > 0) {
1975 _gvn.transform(wl.pop());
1976 }
1977 }
1978
1979 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1980 replaced_nodes.apply(C, final_ctl);
1981 }
1982 if (!ex_ctl->is_top() && do_replaced_nodes) {
1983 replaced_nodes_exception.apply(C, ex_ctl);
1984 }
1985 }
1986
1987
1988 //------------------------------increment_counter------------------------------
1989 // for statistics: increment a VM counter by 1
1990
1991 void GraphKit::increment_counter(address counter_addr) {
1992 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1993 increment_counter(adr1);
1994 }
1995
1996 void GraphKit::increment_counter(Node* counter_addr) {
1997 int adr_type = Compile::AliasIdxRaw;
1998 Node* ctrl = control();
1999 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2000 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
2001 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
2002 }
2003
2004
2005 //------------------------------uncommon_trap----------------------------------
2006 // Bail out to the interpreter in mid-method. Implemented by calling the
2007 // uncommon_trap blob. This helper function inserts a runtime call with the
2008 // right debug info.
2009 void GraphKit::uncommon_trap(int trap_request,
2010 ciKlass* klass, const char* comment,
2011 bool must_throw,
2012 bool keep_exact_action) {
2013 if (failing()) stop();
2014 if (stopped()) return; // trap reachable?
2015
2016 // Note: If ProfileTraps is true, and if a deopt. actually
2017 // occurs here, the runtime will make sure an MDO exists. There is
2018 // no need to call method()->ensure_method_data() at this point.
2019
2020 // Set the stack pointer to the right value for reexecution:
2021 set_sp(reexecute_sp());
2022
2023 #ifdef ASSERT
2024 if (!must_throw) {
2025 // Make sure the stack has at least enough depth to execute
2026 // the current bytecode.
2027 int inputs, ignored_depth;
2028 if (compute_stack_effects(inputs, ignored_depth)) {
2029 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2030 Bytecodes::name(java_bc()), sp(), inputs);
2031 }
2032 }
2033 #endif
2034
2035 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2036 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2037
2038 switch (action) {
2039 case Deoptimization::Action_maybe_recompile:
2040 case Deoptimization::Action_reinterpret:
2041 // Temporary fix for 6529811 to allow virtual calls to be sure they
2042 // get the chance to go from mono->bi->mega
2043 if (!keep_exact_action &&
2044 Deoptimization::trap_request_index(trap_request) < 0 &&
2045 too_many_recompiles(reason)) {
2046 // This BCI is causing too many recompilations.
2047 if (C->log() != NULL) {
2048 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2049 Deoptimization::trap_reason_name(reason),
2050 Deoptimization::trap_action_name(action));
2051 }
2052 action = Deoptimization::Action_none;
2053 trap_request = Deoptimization::make_trap_request(reason, action);
2054 } else {
2055 C->set_trap_can_recompile(true);
2056 }
2057 break;
2058 case Deoptimization::Action_make_not_entrant:
2059 C->set_trap_can_recompile(true);
2060 break;
2061 #ifdef ASSERT
2062 case Deoptimization::Action_none:
2063 case Deoptimization::Action_make_not_compilable:
2064 break;
2065 default:
2066 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2067 break;
2068 #endif
2069 }
2070
2071 if (TraceOptoParse) {
2072 char buf[100];
2073 tty->print_cr("Uncommon trap %s at bci:%d",
2074 Deoptimization::format_trap_request(buf, sizeof(buf),
2075 trap_request), bci());
2076 }
2077
2078 CompileLog* log = C->log();
2079 if (log != NULL) {
2080 int kid = (klass == NULL)? -1: log->identify(klass);
2081 log->begin_elem("uncommon_trap bci='%d'", bci());
2082 char buf[100];
2083 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2084 trap_request));
2085 if (kid >= 0) log->print(" klass='%d'", kid);
2086 if (comment != NULL) log->print(" comment='%s'", comment);
2087 log->end_elem();
2088 }
2089
2090 // Make sure any guarding test views this path as very unlikely
2091 Node *i0 = control()->in(0);
2092 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2093 IfNode *iff = i0->as_If();
2094 float f = iff->_prob; // Get prob
2095 if (control()->Opcode() == Op_IfTrue) {
2096 if (f > PROB_UNLIKELY_MAG(4))
2097 iff->_prob = PROB_MIN;
2098 } else {
2099 if (f < PROB_LIKELY_MAG(4))
2100 iff->_prob = PROB_MAX;
2101 }
2102 }
2103
2104 // Clear out dead values from the debug info.
2105 kill_dead_locals();
2106
2107 // Now insert the uncommon trap subroutine call
2108 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2109 const TypePtr* no_memory_effects = NULL;
2110 // Pass the index of the class to be loaded
2111 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2112 (must_throw ? RC_MUST_THROW : 0),
2113 OptoRuntime::uncommon_trap_Type(),
2114 call_addr, "uncommon_trap", no_memory_effects,
2115 intcon(trap_request));
2116 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2117 "must extract request correctly from the graph");
2118 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2119
2120 call->set_req(TypeFunc::ReturnAdr, returnadr());
2121 // The debug info is the only real input to this call.
2122
2123 // Halt-and-catch fire here. The above call should never return!
2124 HaltNode* halt = new HaltNode(control(), frameptr());
2125 _gvn.set_type_bottom(halt);
2126 root()->add_req(halt);
2127
2128 stop_and_kill_map();
2129 }
2130
2131
2132 //--------------------------just_allocated_object------------------------------
2133 // Report the object that was just allocated.
2134 // It must be the case that there are no intervening safepoints.
2135 // We use this to determine if an object is so "fresh" that
2136 // it does not require card marks.
2137 Node* GraphKit::just_allocated_object(Node* current_control) {
2138 if (C->recent_alloc_ctl() == current_control)
2139 return C->recent_alloc_obj();
2140 return NULL;
2141 }
2142
2143
2144 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2145 // (Note: TypeFunc::make has a cache that makes this fast.)
2146 const TypeFunc* tf = TypeFunc::make(dest_method);
2147 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2148 for (int j = 0; j < nargs; j++) {
2149 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2150 if( targ->basic_type() == T_DOUBLE ) {
2151 // If any parameters are doubles, they must be rounded before
2152 // the call, dstore_rounding does gvn.transform
2153 Node *arg = argument(j);
2154 arg = dstore_rounding(arg);
2155 set_argument(j, arg);
2156 }
2157 }
2158 }
2159
2160 /**
2161 * Record profiling data exact_kls for Node n with the type system so
2162 * that it can propagate it (speculation)
2163 *
2164 * @param n node that the type applies to
2165 * @param exact_kls type from profiling
2166 * @param maybe_null did profiling see null?
2167 *
2168 * @return node with improved type
2169 */
2170 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) {
2171 const Type* current_type = _gvn.type(n);
2172 assert(UseTypeSpeculation, "type speculation must be on");
2173
2174 const TypePtr* speculative = current_type->speculative();
2175
2176 // Should the klass from the profile be recorded in the speculative type?
2177 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2178 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2179 const TypeOopPtr* xtype = tklass->as_instance_type();
2180 assert(xtype->klass_is_exact(), "Should be exact");
2181 // Any reason to believe n is not null (from this profiling or a previous one)?
2182 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2183 // record the new speculative type's depth
2184 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2185 speculative = speculative->with_inline_depth(jvms()->depth());
2186 } else if (current_type->would_improve_ptr(maybe_null)) {
2187 // Profiling report that null was never seen so we can change the
2188 // speculative type to non null ptr.
2189 assert(!maybe_null, "nothing to improve");
2190 if (speculative == NULL) {
2191 speculative = TypePtr::NOTNULL;
2192 } else {
2193 const TypePtr* ptr = TypePtr::NOTNULL;
2194 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2195 }
2196 }
2197
2198 if (speculative != current_type->speculative()) {
2199 // Build a type with a speculative type (what we think we know
2200 // about the type but will need a guard when we use it)
2201 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2202 // We're changing the type, we need a new CheckCast node to carry
2203 // the new type. The new type depends on the control: what
2204 // profiling tells us is only valid from here as far as we can
2205 // tell.
2206 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2207 cast = _gvn.transform(cast);
2208 replace_in_map(n, cast);
2209 n = cast;
2210 }
2211
2212 return n;
2213 }
2214
2215 /**
2216 * Record profiling data from receiver profiling at an invoke with the
2217 * type system so that it can propagate it (speculation)
2218 *
2219 * @param n receiver node
2220 *
2221 * @return node with improved type
2222 */
2223 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2224 if (!UseTypeSpeculation) {
2225 return n;
2226 }
2227 ciKlass* exact_kls = profile_has_unique_klass();
2228 bool maybe_null = true;
2229 if (java_bc() == Bytecodes::_checkcast ||
2230 java_bc() == Bytecodes::_instanceof ||
2231 java_bc() == Bytecodes::_aastore) {
2232 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2233 bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen();
2234 }
2235 return record_profile_for_speculation(n, exact_kls, maybe_null);
2236 return n;
2237 }
2238
2239 /**
2240 * Record profiling data from argument profiling at an invoke with the
2241 * type system so that it can propagate it (speculation)
2242 *
2243 * @param dest_method target method for the call
2244 * @param bc what invoke bytecode is this?
2245 */
2246 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2247 if (!UseTypeSpeculation) {
2248 return;
2249 }
2250 const TypeFunc* tf = TypeFunc::make(dest_method);
2251 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2252 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2253 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2254 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2255 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2256 bool maybe_null = true;
2257 ciKlass* better_type = NULL;
2258 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) {
2259 record_profile_for_speculation(argument(j), better_type, maybe_null);
2260 }
2261 i++;
2262 }
2263 }
2264 }
2265
2266 /**
2267 * Record profiling data from parameter profiling at an invoke with
2268 * the type system so that it can propagate it (speculation)
2269 */
2270 void GraphKit::record_profiled_parameters_for_speculation() {
2271 if (!UseTypeSpeculation) {
2272 return;
2273 }
2274 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2275 if (_gvn.type(local(i))->isa_oopptr()) {
2276 bool maybe_null = true;
2277 ciKlass* better_type = NULL;
2278 if (method()->parameter_profiled_type(j, better_type, maybe_null)) {
2279 record_profile_for_speculation(local(i), better_type, maybe_null);
2280 }
2281 j++;
2282 }
2283 }
2284 }
2285
2286 /**
2287 * Record profiling data from return value profiling at an invoke with
2288 * the type system so that it can propagate it (speculation)
2289 */
2290 void GraphKit::record_profiled_return_for_speculation() {
2291 if (!UseTypeSpeculation) {
2292 return;
2293 }
2294 bool maybe_null = true;
2295 ciKlass* better_type = NULL;
2296 if (method()->return_profiled_type(bci(), better_type, maybe_null)) {
2297 // If profiling reports a single type for the return value,
2298 // feed it to the type system so it can propagate it as a
2299 // speculative type
2300 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null);
2301 }
2302 }
2303
2304 void GraphKit::round_double_result(ciMethod* dest_method) {
2305 // A non-strict method may return a double value which has an extended
2306 // exponent, but this must not be visible in a caller which is 'strict'
2307 // If a strict caller invokes a non-strict callee, round a double result
2308
2309 BasicType result_type = dest_method->return_type()->basic_type();
2310 assert( method() != NULL, "must have caller context");
2311 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2312 // Destination method's return value is on top of stack
2313 // dstore_rounding() does gvn.transform
2314 Node *result = pop_pair();
2315 result = dstore_rounding(result);
2316 push_pair(result);
2317 }
2318 }
2319
2320 // rounding for strict float precision conformance
2321 Node* GraphKit::precision_rounding(Node* n) {
2322 return UseStrictFP && _method->flags().is_strict()
2323 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2324 ? _gvn.transform( new RoundFloatNode(0, n) )
2325 : n;
2326 }
2327
2328 // rounding for strict double precision conformance
2329 Node* GraphKit::dprecision_rounding(Node *n) {
2330 return UseStrictFP && _method->flags().is_strict()
2331 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2332 ? _gvn.transform( new RoundDoubleNode(0, n) )
2333 : n;
2334 }
2335
2336 // rounding for non-strict double stores
2337 Node* GraphKit::dstore_rounding(Node* n) {
2338 return Matcher::strict_fp_requires_explicit_rounding
2339 && UseSSE <= 1
2340 ? _gvn.transform( new RoundDoubleNode(0, n) )
2341 : n;
2342 }
2343
2344 //=============================================================================
2345 // Generate a fast path/slow path idiom. Graph looks like:
2346 // [foo] indicates that 'foo' is a parameter
2347 //
2348 // [in] NULL
2349 // \ /
2350 // CmpP
2351 // Bool ne
2352 // If
2353 // / \
2354 // True False-<2>
2355 // / |
2356 // / cast_not_null
2357 // Load | | ^
2358 // [fast_test] | |
2359 // gvn to opt_test | |
2360 // / \ | <1>
2361 // True False |
2362 // | \\ |
2363 // [slow_call] \[fast_result]
2364 // Ctl Val \ \
2365 // | \ \
2366 // Catch <1> \ \
2367 // / \ ^ \ \
2368 // Ex No_Ex | \ \
2369 // | \ \ | \ <2> \
2370 // ... \ [slow_res] | | \ [null_result]
2371 // \ \--+--+--- | |
2372 // \ | / \ | /
2373 // --------Region Phi
2374 //
2375 //=============================================================================
2376 // Code is structured as a series of driver functions all called 'do_XXX' that
2377 // call a set of helper functions. Helper functions first, then drivers.
2378
2379 //------------------------------null_check_oop---------------------------------
2380 // Null check oop. Set null-path control into Region in slot 3.
2381 // Make a cast-not-nullness use the other not-null control. Return cast.
2382 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2383 bool never_see_null,
2384 bool safe_for_replace,
2385 bool speculative) {
2386 // Initial NULL check taken path
2387 (*null_control) = top();
2388 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2389
2390 // Generate uncommon_trap:
2391 if (never_see_null && (*null_control) != top()) {
2392 // If we see an unexpected null at a check-cast we record it and force a
2393 // recompile; the offending check-cast will be compiled to handle NULLs.
2394 // If we see more than one offending BCI, then all checkcasts in the
2395 // method will be compiled to handle NULLs.
2396 PreserveJVMState pjvms(this);
2397 set_control(*null_control);
2398 replace_in_map(value, null());
2399 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2400 uncommon_trap(reason,
2401 Deoptimization::Action_make_not_entrant);
2402 (*null_control) = top(); // NULL path is dead
2403 }
2404 if ((*null_control) == top() && safe_for_replace) {
2405 replace_in_map(value, cast);
2406 }
2407
2408 // Cast away null-ness on the result
2409 return cast;
2410 }
2411
2412 //------------------------------opt_iff----------------------------------------
2413 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2414 // Return slow-path control.
2415 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2416 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2417
2418 // Fast path taken; set region slot 2
2419 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2420 region->init_req(2,fast_taken); // Capture fast-control
2421
2422 // Fast path not-taken, i.e. slow path
2423 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2424 return slow_taken;
2425 }
2426
2427 //-----------------------------make_runtime_call-------------------------------
2428 Node* GraphKit::make_runtime_call(int flags,
2429 const TypeFunc* call_type, address call_addr,
2430 const char* call_name,
2431 const TypePtr* adr_type,
2432 // The following parms are all optional.
2433 // The first NULL ends the list.
2434 Node* parm0, Node* parm1,
2435 Node* parm2, Node* parm3,
2436 Node* parm4, Node* parm5,
2437 Node* parm6, Node* parm7) {
2438 // Slow-path call
2439 bool is_leaf = !(flags & RC_NO_LEAF);
2440 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2441 if (call_name == NULL) {
2442 assert(!is_leaf, "must supply name for leaf");
2443 call_name = OptoRuntime::stub_name(call_addr);
2444 }
2445 CallNode* call;
2446 if (!is_leaf) {
2447 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2448 bci(), adr_type);
2449 } else if (flags & RC_NO_FP) {
2450 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2451 } else {
2452 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2453 }
2454
2455 // The following is similar to set_edges_for_java_call,
2456 // except that the memory effects of the call are restricted to AliasIdxRaw.
2457
2458 // Slow path call has no side-effects, uses few values
2459 bool wide_in = !(flags & RC_NARROW_MEM);
2460 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2461
2462 Node* prev_mem = NULL;
2463 if (wide_in) {
2464 prev_mem = set_predefined_input_for_runtime_call(call);
2465 } else {
2466 assert(!wide_out, "narrow in => narrow out");
2467 Node* narrow_mem = memory(adr_type);
2468 prev_mem = reset_memory();
2469 map()->set_memory(narrow_mem);
2470 set_predefined_input_for_runtime_call(call);
2471 }
2472
2473 // Hook each parm in order. Stop looking at the first NULL.
2474 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2475 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2476 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2477 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2478 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2479 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2480 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2481 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2482 /* close each nested if ===> */ } } } } } } } }
2483 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2484
2485 if (!is_leaf) {
2486 // Non-leaves can block and take safepoints:
2487 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2488 }
2489 // Non-leaves can throw exceptions:
2490 if (has_io) {
2491 call->set_req(TypeFunc::I_O, i_o());
2492 }
2493
2494 if (flags & RC_UNCOMMON) {
2495 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2496 // (An "if" probability corresponds roughly to an unconditional count.
2497 // Sort of.)
2498 call->set_cnt(PROB_UNLIKELY_MAG(4));
2499 }
2500
2501 Node* c = _gvn.transform(call);
2502 assert(c == call, "cannot disappear");
2503
2504 if (wide_out) {
2505 // Slow path call has full side-effects.
2506 set_predefined_output_for_runtime_call(call);
2507 } else {
2508 // Slow path call has few side-effects, and/or sets few values.
2509 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2510 }
2511
2512 if (has_io) {
2513 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2514 }
2515 return call;
2516
2517 }
2518
2519 //------------------------------merge_memory-----------------------------------
2520 // Merge memory from one path into the current memory state.
2521 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2522 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2523 Node* old_slice = mms.force_memory();
2524 Node* new_slice = mms.memory2();
2525 if (old_slice != new_slice) {
2526 PhiNode* phi;
2527 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2528 if (mms.is_empty()) {
2529 // clone base memory Phi's inputs for this memory slice
2530 assert(old_slice == mms.base_memory(), "sanity");
2531 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2532 _gvn.set_type(phi, Type::MEMORY);
2533 for (uint i = 1; i < phi->req(); i++) {
2534 phi->init_req(i, old_slice->in(i));
2535 }
2536 } else {
2537 phi = old_slice->as_Phi(); // Phi was generated already
2538 }
2539 } else {
2540 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2541 _gvn.set_type(phi, Type::MEMORY);
2542 }
2543 phi->set_req(new_path, new_slice);
2544 mms.set_memory(phi);
2545 }
2546 }
2547 }
2548
2549 //------------------------------make_slow_call_ex------------------------------
2550 // Make the exception handler hookups for the slow call
2551 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2552 if (stopped()) return;
2553
2554 // Make a catch node with just two handlers: fall-through and catch-all
2555 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2556 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2557 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2558 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2559
2560 { PreserveJVMState pjvms(this);
2561 set_control(excp);
2562 set_i_o(i_o);
2563
2564 if (excp != top()) {
2565 if (deoptimize) {
2566 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2567 uncommon_trap(Deoptimization::Reason_unhandled,
2568 Deoptimization::Action_none);
2569 } else {
2570 // Create an exception state also.
2571 // Use an exact type if the caller has specified a specific exception.
2572 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2573 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2574 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2575 }
2576 }
2577 }
2578
2579 // Get the no-exception control from the CatchNode.
2580 set_control(norm);
2581 }
2582
2583 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2584 Node* cmp = NULL;
2585 switch(bt) {
2586 case T_INT: cmp = new CmpINode(in1, in2); break;
2587 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2588 default: fatal("unexpected comparison type %s", type2name(bt));
2589 }
2590 gvn->transform(cmp);
2591 Node* bol = gvn->transform(new BoolNode(cmp, test));
2592 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2593 gvn->transform(iff);
2594 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2595 return iff;
2596 }
2597
2598
2599 //-------------------------------gen_subtype_check-----------------------------
2600 // Generate a subtyping check. Takes as input the subtype and supertype.
2601 // Returns 2 values: sets the default control() to the true path and returns
2602 // the false path. Only reads invariant memory; sets no (visible) memory.
2603 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2604 // but that's not exposed to the optimizer. This call also doesn't take in an
2605 // Object; if you wish to check an Object you need to load the Object's class
2606 // prior to coming here.
2607 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2608 Compile* C = gvn->C;
2609
2610 if ((*ctrl)->is_top()) {
2611 return C->top();
2612 }
2613
2614 // Fast check for identical types, perhaps identical constants.
2615 // The types can even be identical non-constants, in cases
2616 // involving Array.newInstance, Object.clone, etc.
2617 if (subklass == superklass)
2618 return C->top(); // false path is dead; no test needed.
2619
2620 if (gvn->type(superklass)->singleton()) {
2621 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2622 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2623
2624 // In the common case of an exact superklass, try to fold up the
2625 // test before generating code. You may ask, why not just generate
2626 // the code and then let it fold up? The answer is that the generated
2627 // code will necessarily include null checks, which do not always
2628 // completely fold away. If they are also needless, then they turn
2629 // into a performance loss. Example:
2630 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2631 // Here, the type of 'fa' is often exact, so the store check
2632 // of fa[1]=x will fold up, without testing the nullness of x.
2633 switch (C->static_subtype_check(superk, subk)) {
2634 case Compile::SSC_always_false:
2635 {
2636 Node* always_fail = *ctrl;
2637 *ctrl = gvn->C->top();
2638 return always_fail;
2639 }
2640 case Compile::SSC_always_true:
2641 return C->top();
2642 case Compile::SSC_easy_test:
2643 {
2644 // Just do a direct pointer compare and be done.
2645 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2646 *ctrl = gvn->transform(new IfTrueNode(iff));
2647 return gvn->transform(new IfFalseNode(iff));
2648 }
2649 case Compile::SSC_full_test:
2650 break;
2651 default:
2652 ShouldNotReachHere();
2653 }
2654 }
2655
2656 // %%% Possible further optimization: Even if the superklass is not exact,
2657 // if the subklass is the unique subtype of the superklass, the check
2658 // will always succeed. We could leave a dependency behind to ensure this.
2659
2660 // First load the super-klass's check-offset
2661 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2662 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2663 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2664 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2665 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2666
2667 // Load from the sub-klass's super-class display list, or a 1-word cache of
2668 // the secondary superclass list, or a failing value with a sentinel offset
2669 // if the super-klass is an interface or exceptionally deep in the Java
2670 // hierarchy and we have to scan the secondary superclass list the hard way.
2671 // Worst-case type is a little odd: NULL is allowed as a result (usually
2672 // klass loads can never produce a NULL).
2673 Node *chk_off_X = chk_off;
2674 #ifdef _LP64
2675 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2676 #endif
2677 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2678 // For some types like interfaces the following loadKlass is from a 1-word
2679 // cache which is mutable so can't use immutable memory. Other
2680 // types load from the super-class display table which is immutable.
2681 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2682 Node *kmem = might_be_cache ? m : C->immutable_memory();
2683 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2684
2685 // Compile speed common case: ARE a subtype and we canNOT fail
2686 if( superklass == nkls )
2687 return C->top(); // false path is dead; no test needed.
2688
2689 // See if we get an immediate positive hit. Happens roughly 83% of the
2690 // time. Test to see if the value loaded just previously from the subklass
2691 // is exactly the superklass.
2692 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2693 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2694 *ctrl = gvn->transform(new IfFalseNode(iff1));
2695
2696 // Compile speed common case: Check for being deterministic right now. If
2697 // chk_off is a constant and not equal to cacheoff then we are NOT a
2698 // subklass. In this case we need exactly the 1 test above and we can
2699 // return those results immediately.
2700 if (!might_be_cache) {
2701 Node* not_subtype_ctrl = *ctrl;
2702 *ctrl = iftrue1; // We need exactly the 1 test above
2703 return not_subtype_ctrl;
2704 }
2705
2706 // Gather the various success & failures here
2707 RegionNode *r_ok_subtype = new RegionNode(4);
2708 gvn->record_for_igvn(r_ok_subtype);
2709 RegionNode *r_not_subtype = new RegionNode(3);
2710 gvn->record_for_igvn(r_not_subtype);
2711
2712 r_ok_subtype->init_req(1, iftrue1);
2713
2714 // Check for immediate negative hit. Happens roughly 11% of the time (which
2715 // is roughly 63% of the remaining cases). Test to see if the loaded
2716 // check-offset points into the subklass display list or the 1-element
2717 // cache. If it points to the display (and NOT the cache) and the display
2718 // missed then it's not a subtype.
2719 Node *cacheoff = gvn->intcon(cacheoff_con);
2720 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2721 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2722 *ctrl = gvn->transform(new IfFalseNode(iff2));
2723
2724 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2725 // No performance impact (too rare) but allows sharing of secondary arrays
2726 // which has some footprint reduction.
2727 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2728 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2729 *ctrl = gvn->transform(new IfFalseNode(iff3));
2730
2731 // -- Roads not taken here: --
2732 // We could also have chosen to perform the self-check at the beginning
2733 // of this code sequence, as the assembler does. This would not pay off
2734 // the same way, since the optimizer, unlike the assembler, can perform
2735 // static type analysis to fold away many successful self-checks.
2736 // Non-foldable self checks work better here in second position, because
2737 // the initial primary superclass check subsumes a self-check for most
2738 // types. An exception would be a secondary type like array-of-interface,
2739 // which does not appear in its own primary supertype display.
2740 // Finally, we could have chosen to move the self-check into the
2741 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2742 // dependent manner. But it is worthwhile to have the check here,
2743 // where it can be perhaps be optimized. The cost in code space is
2744 // small (register compare, branch).
2745
2746 // Now do a linear scan of the secondary super-klass array. Again, no real
2747 // performance impact (too rare) but it's gotta be done.
2748 // Since the code is rarely used, there is no penalty for moving it
2749 // out of line, and it can only improve I-cache density.
2750 // The decision to inline or out-of-line this final check is platform
2751 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2752 Node* psc = gvn->transform(
2753 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2754
2755 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2756 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2757 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2758
2759 // Return false path; set default control to true path.
2760 *ctrl = gvn->transform(r_ok_subtype);
2761 return gvn->transform(r_not_subtype);
2762 }
2763
2764 // Profile-driven exact type check:
2765 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2766 float prob,
2767 Node* *casted_receiver) {
2768 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2769 Node* recv_klass = load_object_klass(receiver);
2770 Node* want_klass = makecon(tklass);
2771 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2772 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2773 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2774 set_control( _gvn.transform( new IfTrueNode (iff) ));
2775 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2776
2777 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2778 assert(recv_xtype->klass_is_exact(), "");
2779
2780 // Subsume downstream occurrences of receiver with a cast to
2781 // recv_xtype, since now we know what the type will be.
2782 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2783 (*casted_receiver) = _gvn.transform(cast);
2784 // (User must make the replace_in_map call.)
2785
2786 return fail;
2787 }
2788
2789
2790 //------------------------------seems_never_null-------------------------------
2791 // Use null_seen information if it is available from the profile.
2792 // If we see an unexpected null at a type check we record it and force a
2793 // recompile; the offending check will be recompiled to handle NULLs.
2794 // If we see several offending BCIs, then all checks in the
2795 // method will be recompiled.
2796 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2797 speculating = !_gvn.type(obj)->speculative_maybe_null();
2798 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2799 if (UncommonNullCast // Cutout for this technique
2800 && obj != null() // And not the -Xcomp stupid case?
2801 && !too_many_traps(reason)
2802 ) {
2803 if (speculating) {
2804 return true;
2805 }
2806 if (data == NULL)
2807 // Edge case: no mature data. Be optimistic here.
2808 return true;
2809 // If the profile has not seen a null, assume it won't happen.
2810 assert(java_bc() == Bytecodes::_checkcast ||
2811 java_bc() == Bytecodes::_instanceof ||
2812 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2813 return !data->as_BitData()->null_seen();
2814 }
2815 speculating = false;
2816 return false;
2817 }
2818
2819 //------------------------maybe_cast_profiled_receiver-------------------------
2820 // If the profile has seen exactly one type, narrow to exactly that type.
2821 // Subsequent type checks will always fold up.
2822 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2823 ciKlass* require_klass,
2824 ciKlass* spec_klass,
2825 bool safe_for_replace) {
2826 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2827
2828 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2829
2830 // Make sure we haven't already deoptimized from this tactic.
2831 if (too_many_traps(reason) || too_many_recompiles(reason))
2832 return NULL;
2833
2834 // (No, this isn't a call, but it's enough like a virtual call
2835 // to use the same ciMethod accessor to get the profile info...)
2836 // If we have a speculative type use it instead of profiling (which
2837 // may not help us)
2838 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2839 if (exact_kls != NULL) {// no cast failures here
2840 if (require_klass == NULL ||
2841 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2842 // If we narrow the type to match what the type profile sees or
2843 // the speculative type, we can then remove the rest of the
2844 // cast.
2845 // This is a win, even if the exact_kls is very specific,
2846 // because downstream operations, such as method calls,
2847 // will often benefit from the sharper type.
2848 Node* exact_obj = not_null_obj; // will get updated in place...
2849 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2850 &exact_obj);
2851 { PreserveJVMState pjvms(this);
2852 set_control(slow_ctl);
2853 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2854 }
2855 if (safe_for_replace) {
2856 replace_in_map(not_null_obj, exact_obj);
2857 }
2858 return exact_obj;
2859 }
2860 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2861 }
2862
2863 return NULL;
2864 }
2865
2866 /**
2867 * Cast obj to type and emit guard unless we had too many traps here
2868 * already
2869 *
2870 * @param obj node being casted
2871 * @param type type to cast the node to
2872 * @param not_null true if we know node cannot be null
2873 */
2874 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2875 ciKlass* type,
2876 bool not_null) {
2877 if (stopped()) {
2878 return obj;
2879 }
2880
2881 // type == NULL if profiling tells us this object is always null
2882 if (type != NULL) {
2883 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2884 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2885
2886 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2887 !too_many_traps(class_reason) &&
2888 !too_many_recompiles(class_reason)) {
2889 Node* not_null_obj = NULL;
2890 // not_null is true if we know the object is not null and
2891 // there's no need for a null check
2892 if (!not_null) {
2893 Node* null_ctl = top();
2894 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2895 assert(null_ctl->is_top(), "no null control here");
2896 } else {
2897 not_null_obj = obj;
2898 }
2899
2900 Node* exact_obj = not_null_obj;
2901 ciKlass* exact_kls = type;
2902 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2903 &exact_obj);
2904 {
2905 PreserveJVMState pjvms(this);
2906 set_control(slow_ctl);
2907 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2908 }
2909 replace_in_map(not_null_obj, exact_obj);
2910 obj = exact_obj;
2911 }
2912 } else {
2913 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2914 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2915 Node* exact_obj = null_assert(obj);
2916 replace_in_map(obj, exact_obj);
2917 obj = exact_obj;
2918 }
2919 }
2920 return obj;
2921 }
2922
2923 //-------------------------------gen_instanceof--------------------------------
2924 // Generate an instance-of idiom. Used by both the instance-of bytecode
2925 // and the reflective instance-of call.
2926 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2927 kill_dead_locals(); // Benefit all the uncommon traps
2928 assert( !stopped(), "dead parse path should be checked in callers" );
2929 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2930 "must check for not-null not-dead klass in callers");
2931
2932 // Make the merge point
2933 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2934 RegionNode* region = new RegionNode(PATH_LIMIT);
2935 Node* phi = new PhiNode(region, TypeInt::BOOL);
2936 C->set_has_split_ifs(true); // Has chance for split-if optimization
2937
2938 ciProfileData* data = NULL;
2939 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2940 data = method()->method_data()->bci_to_data(bci());
2941 }
2942 bool speculative_not_null = false;
2943 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2944 && seems_never_null(obj, data, speculative_not_null));
2945
2946 // Null check; get casted pointer; set region slot 3
2947 Node* null_ctl = top();
2948 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2949
2950 // If not_null_obj is dead, only null-path is taken
2951 if (stopped()) { // Doing instance-of on a NULL?
2952 set_control(null_ctl);
2953 return intcon(0);
2954 }
2955 region->init_req(_null_path, null_ctl);
2956 phi ->init_req(_null_path, intcon(0)); // Set null path value
2957 if (null_ctl == top()) {
2958 // Do this eagerly, so that pattern matches like is_diamond_phi
2959 // will work even during parsing.
2960 assert(_null_path == PATH_LIMIT-1, "delete last");
2961 region->del_req(_null_path);
2962 phi ->del_req(_null_path);
2963 }
2964
2965 // Do we know the type check always succeed?
2966 bool known_statically = false;
2967 if (_gvn.type(superklass)->singleton()) {
2968 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2969 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2970 if (subk != NULL && subk->is_loaded()) {
2971 int static_res = C->static_subtype_check(superk, subk);
2972 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2973 }
2974 }
2975
2976 if (known_statically && UseTypeSpeculation) {
2977 // If we know the type check always succeeds then we don't use the
2978 // profiling data at this bytecode. Don't lose it, feed it to the
2979 // type system as a speculative type.
2980 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2981 } else {
2982 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2983 // We may not have profiling here or it may not help us. If we
2984 // have a speculative type use it to perform an exact cast.
2985 ciKlass* spec_obj_type = obj_type->speculative_type();
2986 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2987 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2988 if (stopped()) { // Profile disagrees with this path.
2989 set_control(null_ctl); // Null is the only remaining possibility.
2990 return intcon(0);
2991 }
2992 if (cast_obj != NULL) {
2993 not_null_obj = cast_obj;
2994 }
2995 }
2996 }
2997
2998 // Load the object's klass
2999 Node* obj_klass = load_object_klass(not_null_obj);
3000
3001 // Generate the subtype check
3002 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3003
3004 // Plug in the success path to the general merge in slot 1.
3005 region->init_req(_obj_path, control());
3006 phi ->init_req(_obj_path, intcon(1));
3007
3008 // Plug in the failing path to the general merge in slot 2.
3009 region->init_req(_fail_path, not_subtype_ctrl);
3010 phi ->init_req(_fail_path, intcon(0));
3011
3012 // Return final merged results
3013 set_control( _gvn.transform(region) );
3014 record_for_igvn(region);
3015 return _gvn.transform(phi);
3016 }
3017
3018 //-------------------------------gen_checkcast---------------------------------
3019 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3020 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3021 // uncommon-trap paths work. Adjust stack after this call.
3022 // If failure_control is supplied and not null, it is filled in with
3023 // the control edge for the cast failure. Otherwise, an appropriate
3024 // uncommon trap or exception is thrown.
3025 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3026 Node* *failure_control) {
3027 kill_dead_locals(); // Benefit all the uncommon traps
3028 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3029 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3030
3031 // Fast cutout: Check the case that the cast is vacuously true.
3032 // This detects the common cases where the test will short-circuit
3033 // away completely. We do this before we perform the null check,
3034 // because if the test is going to turn into zero code, we don't
3035 // want a residual null check left around. (Causes a slowdown,
3036 // for example, in some objArray manipulations, such as a[i]=a[j].)
3037 if (tk->singleton()) {
3038 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3039 if (objtp != NULL && objtp->klass() != NULL) {
3040 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3041 case Compile::SSC_always_true:
3042 // If we know the type check always succeed then we don't use
3043 // the profiling data at this bytecode. Don't lose it, feed it
3044 // to the type system as a speculative type.
3045 return record_profiled_receiver_for_speculation(obj);
3046 case Compile::SSC_always_false:
3047 // It needs a null check because a null will *pass* the cast check.
3048 // A non-null value will always produce an exception.
3049 return null_assert(obj);
3050 }
3051 }
3052 }
3053
3054 ciProfileData* data = NULL;
3055 bool safe_for_replace = false;
3056 if (failure_control == NULL &&
3057 java_bc() != Bytecodes::_vbox &&
3058 java_bc() != Bytecodes::_vunbox) { // use MDO in regular case only
3059 // Don't use MDO for the vunbox and vbox as no profile
3060 // information is recorded for that bytecode.
3061 // TOOD: Implement profiling for vunbox.
3062 assert(java_bc() == Bytecodes::_aastore ||
3063 java_bc() == Bytecodes::_checkcast,
3064 "interpreter profiles type checks only for these BCs");
3065 data = method()->method_data()->bci_to_data(bci());
3066 safe_for_replace = true;
3067 }
3068
3069 // Make the merge point
3070 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3071 RegionNode* region = new RegionNode(PATH_LIMIT);
3072 Node* phi = new PhiNode(region, toop);
3073 C->set_has_split_ifs(true); // Has chance for split-if optimization
3074
3075 // Use null-cast information if it is available
3076 bool speculative_not_null = false;
3077 bool never_see_null = ((failure_control == NULL) // regular case only
3078 && java_bc() != Bytecodes::_vunbox // TODO: Implement profiling for vunbox and vbox.
3079 && java_bc() != Bytecodes::_vbox
3080 && seems_never_null(obj, data, speculative_not_null));
3081
3082 // Null check; get casted pointer; set region slot 3
3083 Node* null_ctl = top();
3084 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3085
3086 // If not_null_obj is dead, only null-path is taken
3087 if (stopped()) { // Doing instance-of on a NULL?
3088 set_control(null_ctl);
3089 return null();
3090 }
3091 region->init_req(_null_path, null_ctl);
3092 phi ->init_req(_null_path, null()); // Set null path value
3093 if (null_ctl == top()) {
3094 // Do this eagerly, so that pattern matches like is_diamond_phi
3095 // will work even during parsing.
3096 assert(_null_path == PATH_LIMIT-1, "delete last");
3097 region->del_req(_null_path);
3098 phi ->del_req(_null_path);
3099 }
3100
3101 Node* cast_obj = NULL;
3102 if (tk->klass_is_exact()) {
3103 // The following optimization tries to statically cast the speculative type of the object
3104 // (for example obtained during profiling) to the type of the superklass and then do a
3105 // dynamic check that the type of the object is what we expect. To work correctly
3106 // for checkcast and aastore the type of superklass should be exact.
3107 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3108 // We may not have profiling here or it may not help us. If we have
3109 // a speculative type use it to perform an exact cast.
3110 ciKlass* spec_obj_type = obj_type->speculative_type();
3111 if (spec_obj_type != NULL || data != NULL) {
3112 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3113 if (cast_obj != NULL) {
3114 if (failure_control != NULL) // failure is now impossible
3115 (*failure_control) = top();
3116 // adjust the type of the phi to the exact klass:
3117 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3118 }
3119 }
3120 }
3121
3122 if (cast_obj == NULL) {
3123 // Load the object's klass
3124 Node* obj_klass = load_object_klass(not_null_obj);
3125
3126 // Generate the subtype check
3127 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3128
3129 // Plug in success path into the merge
3130 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3131 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3132 if (failure_control == NULL) {
3133 if (not_subtype_ctrl != top()) { // If failure is possible
3134 PreserveJVMState pjvms(this);
3135 set_control(not_subtype_ctrl);
3136 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3137 }
3138 } else {
3139 (*failure_control) = not_subtype_ctrl;
3140 }
3141 }
3142
3143 region->init_req(_obj_path, control());
3144 phi ->init_req(_obj_path, cast_obj);
3145
3146 // A merge of NULL or Casted-NotNull obj
3147 Node* res = _gvn.transform(phi);
3148
3149 // Note I do NOT always 'replace_in_map(obj,result)' here.
3150 // if( tk->klass()->can_be_primary_super() )
3151 // This means that if I successfully store an Object into an array-of-String
3152 // I 'forget' that the Object is really now known to be a String. I have to
3153 // do this because we don't have true union types for interfaces - if I store
3154 // a Baz into an array-of-Interface and then tell the optimizer it's an
3155 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3156 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3157 // replace_in_map( obj, res );
3158
3159 // Return final merged results
3160 set_control( _gvn.transform(region) );
3161 record_for_igvn(region);
3162 return res;
3163 }
3164
3165 //------------------------------next_monitor-----------------------------------
3166 // What number should be given to the next monitor?
3167 int GraphKit::next_monitor() {
3168 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3169 int next = current + C->sync_stack_slots();
3170 // Keep the toplevel high water mark current:
3171 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3172 return current;
3173 }
3174
3175 //------------------------------insert_mem_bar---------------------------------
3176 // Memory barrier to avoid floating things around
3177 // The membar serves as a pinch point between both control and all memory slices.
3178 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3179 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3180 mb->init_req(TypeFunc::Control, control());
3181 mb->init_req(TypeFunc::Memory, reset_memory());
3182 Node* membar = _gvn.transform(mb);
3183 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3184 set_all_memory_call(membar);
3185 return membar;
3186 }
3187
3188 //-------------------------insert_mem_bar_volatile----------------------------
3189 // Memory barrier to avoid floating things around
3190 // The membar serves as a pinch point between both control and memory(alias_idx).
3191 // If you want to make a pinch point on all memory slices, do not use this
3192 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3193 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3194 // When Parse::do_put_xxx updates a volatile field, it appends a series
3195 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3196 // The first membar is on the same memory slice as the field store opcode.
3197 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3198 // All the other membars (for other volatile slices, including AliasIdxBot,
3199 // which stands for all unknown volatile slices) are control-dependent
3200 // on the first membar. This prevents later volatile loads or stores
3201 // from sliding up past the just-emitted store.
3202
3203 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3204 mb->set_req(TypeFunc::Control,control());
3205 if (alias_idx == Compile::AliasIdxBot) {
3206 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3207 } else {
3208 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3209 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3210 }
3211 Node* membar = _gvn.transform(mb);
3212 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3213 if (alias_idx == Compile::AliasIdxBot) {
3214 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3215 } else {
3216 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3217 }
3218 return membar;
3219 }
3220
3221 void GraphKit::insert_store_load_for_barrier() {
3222 Node* mem = reset_memory();
3223 MemBarNode* mb = MemBarNode::make(C, Op_MemBarVolatile, Compile::AliasIdxBot);
3224 mb->init_req(TypeFunc::Control, control());
3225 mb->init_req(TypeFunc::Memory, mem);
3226 Node* membar = _gvn.transform(mb);
3227 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3228 Node* newmem = _gvn.transform(new ProjNode(membar, TypeFunc::Memory));
3229 set_all_memory(mem);
3230 set_memory(newmem, Compile::AliasIdxRaw);
3231 }
3232
3233
3234 //------------------------------shared_lock------------------------------------
3235 // Emit locking code.
3236 FastLockNode* GraphKit::shared_lock(Node* obj) {
3237 // bci is either a monitorenter bc or InvocationEntryBci
3238 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3239 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3240
3241 if( !GenerateSynchronizationCode )
3242 return NULL; // Not locking things?
3243 if (stopped()) // Dead monitor?
3244 return NULL;
3245
3246 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3247
3248 // Box the stack location
3249 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3250 Node* mem = reset_memory();
3251
3252 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3253 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3254 // Create the counters for this fast lock.
3255 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3256 }
3257
3258 // Create the rtm counters for this fast lock if needed.
3259 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3260
3261 // Add monitor to debug info for the slow path. If we block inside the
3262 // slow path and de-opt, we need the monitor hanging around
3263 map()->push_monitor( flock );
3264
3265 const TypeFunc *tf = LockNode::lock_type();
3266 LockNode *lock = new LockNode(C, tf);
3267
3268 lock->init_req( TypeFunc::Control, control() );
3269 lock->init_req( TypeFunc::Memory , mem );
3270 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3271 lock->init_req( TypeFunc::FramePtr, frameptr() );
3272 lock->init_req( TypeFunc::ReturnAdr, top() );
3273
3274 lock->init_req(TypeFunc::Parms + 0, obj);
3275 lock->init_req(TypeFunc::Parms + 1, box);
3276 lock->init_req(TypeFunc::Parms + 2, flock);
3277 add_safepoint_edges(lock);
3278
3279 lock = _gvn.transform( lock )->as_Lock();
3280
3281 // lock has no side-effects, sets few values
3282 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3283
3284 insert_mem_bar(Op_MemBarAcquireLock);
3285
3286 // Add this to the worklist so that the lock can be eliminated
3287 record_for_igvn(lock);
3288
3289 #ifndef PRODUCT
3290 if (PrintLockStatistics) {
3291 // Update the counter for this lock. Don't bother using an atomic
3292 // operation since we don't require absolute accuracy.
3293 lock->create_lock_counter(map()->jvms());
3294 increment_counter(lock->counter()->addr());
3295 }
3296 #endif
3297
3298 return flock;
3299 }
3300
3301
3302 //------------------------------shared_unlock----------------------------------
3303 // Emit unlocking code.
3304 void GraphKit::shared_unlock(Node* box, Node* obj) {
3305 // bci is either a monitorenter bc or InvocationEntryBci
3306 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3307 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3308
3309 if( !GenerateSynchronizationCode )
3310 return;
3311 if (stopped()) { // Dead monitor?
3312 map()->pop_monitor(); // Kill monitor from debug info
3313 return;
3314 }
3315
3316 // Memory barrier to avoid floating things down past the locked region
3317 insert_mem_bar(Op_MemBarReleaseLock);
3318
3319 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3320 UnlockNode *unlock = new UnlockNode(C, tf);
3321 #ifdef ASSERT
3322 unlock->set_dbg_jvms(sync_jvms());
3323 #endif
3324 uint raw_idx = Compile::AliasIdxRaw;
3325 unlock->init_req( TypeFunc::Control, control() );
3326 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3327 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3328 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3329 unlock->init_req( TypeFunc::ReturnAdr, top() );
3330
3331 unlock->init_req(TypeFunc::Parms + 0, obj);
3332 unlock->init_req(TypeFunc::Parms + 1, box);
3333 unlock = _gvn.transform(unlock)->as_Unlock();
3334
3335 Node* mem = reset_memory();
3336
3337 // unlock has no side-effects, sets few values
3338 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3339
3340 // Kill monitor from debug info
3341 map()->pop_monitor( );
3342 }
3343
3344 //-------------------------------get_layout_helper-----------------------------
3345 // If the given klass is a constant or known to be an array,
3346 // fetch the constant layout helper value into constant_value
3347 // and return (Node*)NULL. Otherwise, load the non-constant
3348 // layout helper value, and return the node which represents it.
3349 // This two-faced routine is useful because allocation sites
3350 // almost always feature constant types.
3351 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3352 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3353 if (!StressReflectiveCode && inst_klass != NULL) {
3354 ciKlass* klass = inst_klass->klass();
3355 bool xklass = inst_klass->klass_is_exact();
3356 if (xklass || klass->is_array_klass()) {
3357 jint lhelper = klass->layout_helper();
3358 if (lhelper != Klass::_lh_neutral_value) {
3359 constant_value = lhelper;
3360 return (Node*) NULL;
3361 }
3362 }
3363 }
3364 constant_value = Klass::_lh_neutral_value; // put in a known value
3365 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3366 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3367 }
3368
3369 // We just put in an allocate/initialize with a big raw-memory effect.
3370 // Hook selected additional alias categories on the initialization.
3371 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3372 MergeMemNode* init_in_merge,
3373 Node* init_out_raw) {
3374 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3375 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3376
3377 Node* prevmem = kit.memory(alias_idx);
3378 init_in_merge->set_memory_at(alias_idx, prevmem);
3379 kit.set_memory(init_out_raw, alias_idx);
3380 }
3381
3382 //---------------------------set_output_for_allocation-------------------------
3383 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3384 const TypeOopPtr* oop_type,
3385 bool deoptimize_on_exception) {
3386 int rawidx = Compile::AliasIdxRaw;
3387 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3388 add_safepoint_edges(alloc);
3389 Node* allocx = _gvn.transform(alloc);
3390 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3391 // create memory projection for i_o
3392 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3393 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3394
3395 // create a memory projection as for the normal control path
3396 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3397 set_memory(malloc, rawidx);
3398
3399 // a normal slow-call doesn't change i_o, but an allocation does
3400 // we create a separate i_o projection for the normal control path
3401 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3402 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3403
3404 // put in an initialization barrier
3405 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3406 rawoop)->as_Initialize();
3407 assert(alloc->initialization() == init, "2-way macro link must work");
3408 assert(init ->allocation() == alloc, "2-way macro link must work");
3409 {
3410 // Extract memory strands which may participate in the new object's
3411 // initialization, and source them from the new InitializeNode.
3412 // This will allow us to observe initializations when they occur,
3413 // and link them properly (as a group) to the InitializeNode.
3414 assert(init->in(InitializeNode::Memory) == malloc, "");
3415 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3416 init->set_req(InitializeNode::Memory, minit_in);
3417 record_for_igvn(minit_in); // fold it up later, if possible
3418 Node* minit_out = memory(rawidx);
3419 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3420 if (oop_type->isa_aryptr()) {
3421 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3422 int elemidx = C->get_alias_index(telemref);
3423 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3424 } else if (oop_type->isa_instptr() || oop_type->isa_valuetypeptr()) {
3425 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3426 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3427 ciField* field = ik->nonstatic_field_at(i);
3428 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3429 continue; // do not bother to track really large numbers of fields
3430 // Find (or create) the alias category for this field:
3431 int fieldidx = C->alias_type(field)->index();
3432 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3433 }
3434 }
3435 }
3436
3437 // Cast raw oop to the real thing...
3438 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3439 javaoop = _gvn.transform(javaoop);
3440 C->set_recent_alloc(control(), javaoop);
3441 assert(just_allocated_object(control()) == javaoop, "just allocated");
3442
3443 #ifdef ASSERT
3444 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3445 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3446 "Ideal_allocation works");
3447 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3448 "Ideal_allocation works");
3449 if (alloc->is_AllocateArray()) {
3450 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3451 "Ideal_allocation works");
3452 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3453 "Ideal_allocation works");
3454 } else {
3455 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3456 }
3457 }
3458 #endif //ASSERT
3459
3460 return javaoop;
3461 }
3462
3463 //---------------------------new_instance--------------------------------------
3464 // This routine takes a klass_node which may be constant (for a static type)
3465 // or may be non-constant (for reflective code). It will work equally well
3466 // for either, and the graph will fold nicely if the optimizer later reduces
3467 // the type to a constant.
3468 // The optional arguments are for specialized use by intrinsics:
3469 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3470 // - If 'return_size_val', report the the total object size to the caller.
3471 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3472 Node* GraphKit::new_instance(Node* klass_node,
3473 Node* extra_slow_test,
3474 Node* *return_size_val,
3475 bool deoptimize_on_exception) {
3476 // Compute size in doublewords
3477 // The size is always an integral number of doublewords, represented
3478 // as a positive bytewise size stored in the klass's layout_helper.
3479 // The layout_helper also encodes (in a low bit) the need for a slow path.
3480 jint layout_con = Klass::_lh_neutral_value;
3481 Node* layout_val = get_layout_helper(klass_node, layout_con);
3482 int layout_is_con = (layout_val == NULL);
3483
3484 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3485 // Generate the initial go-slow test. It's either ALWAYS (return a
3486 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3487 // case) a computed value derived from the layout_helper.
3488 Node* initial_slow_test = NULL;
3489 if (layout_is_con) {
3490 assert(!StressReflectiveCode, "stress mode does not use these paths");
3491 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3492 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3493 } else { // reflective case
3494 // This reflective path is used by Unsafe.allocateInstance.
3495 // (It may be stress-tested by specifying StressReflectiveCode.)
3496 // Basically, we want to get into the VM is there's an illegal argument.
3497 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3498 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3499 if (extra_slow_test != intcon(0)) {
3500 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3501 }
3502 // (Macro-expander will further convert this to a Bool, if necessary.)
3503 }
3504
3505 // Find the size in bytes. This is easy; it's the layout_helper.
3506 // The size value must be valid even if the slow path is taken.
3507 Node* size = NULL;
3508 if (layout_is_con) {
3509 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3510 } else { // reflective case
3511 // This reflective path is used by clone and Unsafe.allocateInstance.
3512 size = ConvI2X(layout_val);
3513
3514 // Clear the low bits to extract layout_helper_size_in_bytes:
3515 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3516 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3517 size = _gvn.transform( new AndXNode(size, mask) );
3518 }
3519 if (return_size_val != NULL) {
3520 (*return_size_val) = size;
3521 }
3522
3523 // This is a precise notnull oop of the klass.
3524 // (Actually, it need not be precise if this is a reflective allocation.)
3525 // It's what we cast the result to.
3526 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3527 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3528 const TypeOopPtr* oop_type = tklass->as_instance_type();
3529
3530 // Now generate allocation code
3531
3532 // The entire memory state is needed for slow path of the allocation
3533 // since GC and deoptimization can happen.
3534 Node *mem = reset_memory();
3535 set_all_memory(mem); // Create new memory state
3536
3537 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3538 control(), mem, i_o(),
3539 size, klass_node,
3540 initial_slow_test);
3541
3542 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3543 }
3544
3545 //-------------------------------new_array-------------------------------------
3546 // helper for newarray, anewarray and vnewarray
3547 // The 'length' parameter is (obviously) the length of the array.
3548 // See comments on new_instance for the meaning of the other arguments.
3549 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3550 Node* length, // number of array elements
3551 int nargs, // number of arguments to push back for uncommon trap
3552 Node* *return_size_val,
3553 bool deoptimize_on_exception) {
3554 jint layout_con = Klass::_lh_neutral_value;
3555 Node* layout_val = get_layout_helper(klass_node, layout_con);
3556 int layout_is_con = (layout_val == NULL);
3557
3558 if (!layout_is_con && !StressReflectiveCode &&
3559 !too_many_traps(Deoptimization::Reason_class_check)) {
3560 // This is a reflective array creation site.
3561 // Optimistically assume that it is a subtype of Object[],
3562 // so that we can fold up all the address arithmetic.
3563 layout_con = Klass::array_layout_helper(T_OBJECT);
3564 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3565 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3566 { BuildCutout unless(this, bol_lh, PROB_MAX);
3567 inc_sp(nargs);
3568 uncommon_trap(Deoptimization::Reason_class_check,
3569 Deoptimization::Action_maybe_recompile);
3570 }
3571 layout_val = NULL;
3572 layout_is_con = true;
3573 }
3574
3575 // Generate the initial go-slow test. Make sure we do not overflow
3576 // if length is huge (near 2Gig) or negative! We do not need
3577 // exact double-words here, just a close approximation of needed
3578 // double-words. We can't add any offset or rounding bits, lest we
3579 // take a size -1 of bytes and make it positive. Use an unsigned
3580 // compare, so negative sizes look hugely positive.
3581 int fast_size_limit = FastAllocateSizeLimit;
3582 if (layout_is_con) {
3583 assert(!StressReflectiveCode, "stress mode does not use these paths");
3584 // Increase the size limit if we have exact knowledge of array type.
3585 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3586 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3587 }
3588
3589 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3590 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3591
3592 // --- Size Computation ---
3593 // array_size = round_to_heap(array_header + (length << elem_shift));
3594 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3595 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3596 // The rounding mask is strength-reduced, if possible.
3597 int round_mask = MinObjAlignmentInBytes - 1;
3598 Node* header_size = NULL;
3599 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3600 // (T_BYTE has the weakest alignment and size restrictions...)
3601 if (layout_is_con) {
3602 int hsize = Klass::layout_helper_header_size(layout_con);
3603 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3604 BasicType etype = Klass::layout_helper_element_type(layout_con);
3605 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con);
3606 if ((round_mask & ~right_n_bits(eshift)) == 0)
3607 round_mask = 0; // strength-reduce it if it goes away completely
3608 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3609 assert(header_size_min <= hsize, "generic minimum is smallest");
3610 header_size_min = hsize;
3611 header_size = intcon(hsize + round_mask);
3612 } else {
3613 Node* hss = intcon(Klass::_lh_header_size_shift);
3614 Node* hsm = intcon(Klass::_lh_header_size_mask);
3615 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3616 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3617 Node* mask = intcon(round_mask);
3618 header_size = _gvn.transform( new AddINode(hsize, mask) );
3619 }
3620
3621 Node* elem_shift = NULL;
3622 if (layout_is_con) {
3623 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3624 if (eshift != 0)
3625 elem_shift = intcon(eshift);
3626 } else {
3627 // There is no need to mask or shift this value.
3628 // The semantics of LShiftINode include an implicit mask to 0x1F.
3629 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3630 elem_shift = layout_val;
3631 }
3632
3633 // Transition to native address size for all offset calculations:
3634 Node* lengthx = ConvI2X(length);
3635 Node* headerx = ConvI2X(header_size);
3636 #ifdef _LP64
3637 { const TypeInt* tilen = _gvn.find_int_type(length);
3638 if (tilen != NULL && tilen->_lo < 0) {
3639 // Add a manual constraint to a positive range. Cf. array_element_address.
3640 jint size_max = fast_size_limit;
3641 if (size_max > tilen->_hi) size_max = tilen->_hi;
3642 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3643
3644 // Only do a narrow I2L conversion if the range check passed.
3645 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3646 _gvn.transform(iff);
3647 RegionNode* region = new RegionNode(3);
3648 _gvn.set_type(region, Type::CONTROL);
3649 lengthx = new PhiNode(region, TypeLong::LONG);
3650 _gvn.set_type(lengthx, TypeLong::LONG);
3651
3652 // Range check passed. Use ConvI2L node with narrow type.
3653 Node* passed = IfFalse(iff);
3654 region->init_req(1, passed);
3655 // Make I2L conversion control dependent to prevent it from
3656 // floating above the range check during loop optimizations.
3657 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3658
3659 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3660 region->init_req(2, IfTrue(iff));
3661 lengthx->init_req(2, ConvI2X(length));
3662
3663 set_control(region);
3664 record_for_igvn(region);
3665 record_for_igvn(lengthx);
3666 }
3667 }
3668 #endif
3669
3670 // Combine header size (plus rounding) and body size. Then round down.
3671 // This computation cannot overflow, because it is used only in two
3672 // places, one where the length is sharply limited, and the other
3673 // after a successful allocation.
3674 Node* abody = lengthx;
3675 if (elem_shift != NULL)
3676 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3677 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3678 if (round_mask != 0) {
3679 Node* mask = MakeConX(~round_mask);
3680 size = _gvn.transform( new AndXNode(size, mask) );
3681 }
3682 // else if round_mask == 0, the size computation is self-rounding
3683
3684 if (return_size_val != NULL) {
3685 // This is the size
3686 (*return_size_val) = size;
3687 }
3688
3689 // Now generate allocation code
3690
3691 // The entire memory state is needed for slow path of the allocation
3692 // since GC and deoptimization can happen.
3693 Node *mem = reset_memory();
3694 set_all_memory(mem); // Create new memory state
3695
3696 if (initial_slow_test->is_Bool()) {
3697 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3698 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3699 }
3700
3701 // Create the AllocateArrayNode and its result projections
3702 AllocateArrayNode* alloc
3703 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3704 control(), mem, i_o(),
3705 size, klass_node,
3706 initial_slow_test,
3707 length);
3708
3709 // Cast to correct type. Note that the klass_node may be constant or not,
3710 // and in the latter case the actual array type will be inexact also.
3711 // (This happens via a non-constant argument to inline_native_newArray.)
3712 // In any case, the value of klass_node provides the desired array type.
3713 const TypeInt* length_type = _gvn.find_int_type(length);
3714 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3715 if (ary_type->isa_aryptr() && length_type != NULL) {
3716 // Try to get a better type than POS for the size
3717 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3718 }
3719
3720 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3721
3722 // Cast length on remaining path to be as narrow as possible
3723 if (map()->find_edge(length) >= 0) {
3724 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3725 if (ccast != length) {
3726 _gvn.set_type_bottom(ccast);
3727 record_for_igvn(ccast);
3728 replace_in_map(length, ccast);
3729 }
3730 }
3731
3732 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
3733 ciKlass* elem_klass = ary_ptr != NULL ? ary_ptr->klass()->as_array_klass()->element_klass() : NULL;
3734 if (elem_klass != NULL && elem_klass->is_valuetype()) {
3735 ciValueKlass* vk = elem_klass->as_value_klass();
3736 if (!vk->flatten_array()) {
3737 // Non-flattened value type arrays need to be initialized with default value type oops
3738 initialize_value_type_array(javaoop, length, elem_klass->as_value_klass(), nargs);
3739 InitializeNode* init = alloc->initialization();
3740 init->set_complete_with_arraycopy();
3741 }
3742 }
3743
3744 return javaoop;
3745 }
3746
3747 void GraphKit::initialize_value_type_array(Node* array, Node* length, ciValueKlass* vk, int nargs) {
3748 // Check for zero length
3749 Node* null_ctl = top();
3750 null_check_common(length, T_INT, false, &null_ctl, false);
3751 if (stopped()) {
3752 set_control(null_ctl); // Always zero
3753 return;
3754 }
3755
3756 // Prepare for merging control and IO
3757 RegionNode* res_ctl = new RegionNode(3);
3758 res_ctl->init_req(1, null_ctl);
3759 gvn().set_type(res_ctl, Type::CONTROL);
3760 record_for_igvn(res_ctl);
3761 Node* res_io = PhiNode::make(res_ctl, i_o(), Type::ABIO);
3762 gvn().set_type(res_io, Type::ABIO);
3763 record_for_igvn(res_io);
3764
3765 // TODO comment
3766 SafePointNode* loop_map = NULL;
3767 {
3768 PreserveJVMState pjvms(this);
3769 // Create default value type and store it to memory
3770 Node* oop = ValueTypeNode::make_default(gvn(), vk);
3771 oop = oop->as_ValueType()->store_to_memory(this);
3772
3773 length = SubI(length, intcon(1));
3774 add_predicate(nargs);
3775 RegionNode* loop = new RegionNode(3);
3776 loop->init_req(1, control());
3777 gvn().set_type(loop, Type::CONTROL);
3778 record_for_igvn(loop);
3779
3780 Node* index = new PhiNode(loop, TypeInt::INT);
3781 index->init_req(1, intcon(0));
3782 gvn().set_type(index, TypeInt::INT);
3783 record_for_igvn(index);
3784
3785 // TODO explain why we need to capture all memory
3786 PhiNode* mem = new PhiNode(loop, Type::MEMORY, TypePtr::BOTTOM);
3787 mem->init_req(1, reset_memory());
3788 gvn().set_type(mem, Type::MEMORY);
3789 record_for_igvn(mem);
3790 set_control(loop);
3791 set_all_memory(mem);
3792 // Initialize array element
3793 Node* adr = array_element_address(array, index, T_OBJECT);
3794 const TypeOopPtr* elemtype = TypeValueTypePtr::make(TypePtr::NotNull, vk);
3795 Node* store = store_oop_to_array(control(), array, adr, TypeAryPtr::OOPS, oop, elemtype, T_OBJECT, MemNode::release);
3796
3797 IfNode* iff = create_and_map_if(control(), Bool(CmpI(index, length), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
3798 loop->init_req(2, IfTrue(iff));
3799 mem->init_req(2, merged_memory());
3800 index->init_req(2, AddI(index, intcon(1)));
3801
3802 res_ctl->init_req(2, IfFalse(iff));
3803 res_io->set_req(2, i_o());
3804 loop_map = stop();
3805 }
3806 // Set merged control, IO and memory
3807 set_control(res_ctl);
3808 set_i_o(res_io);
3809 merge_memory(loop_map->merged_memory(), res_ctl, 2);
3810
3811 // Transform new memory Phis.
3812 for (MergeMemStream mms(merged_memory()); mms.next_non_empty();) {
3813 Node* phi = mms.memory();
3814 if (phi->is_Phi() && phi->in(0) == res_ctl) {
3815 mms.set_memory(gvn().transform(phi));
3816 }
3817 }
3818 }
3819
3820 // The following "Ideal_foo" functions are placed here because they recognize
3821 // the graph shapes created by the functions immediately above.
3822
3823 //---------------------------Ideal_allocation----------------------------------
3824 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3825 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3826 if (ptr == NULL) { // reduce dumb test in callers
3827 return NULL;
3828 }
3829 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3830 ptr = ptr->in(1);
3831 if (ptr == NULL) return NULL;
3832 }
3833 // Return NULL for allocations with several casts:
3834 // j.l.reflect.Array.newInstance(jobject, jint)
3835 // Object.clone()
3836 // to keep more precise type from last cast.
3837 if (ptr->is_Proj()) {
3838 Node* allo = ptr->in(0);
3839 if (allo != NULL && allo->is_Allocate()) {
3840 return allo->as_Allocate();
3841 }
3842 }
3843 // Report failure to match.
3844 return NULL;
3845 }
3846
3847 // Fancy version which also strips off an offset (and reports it to caller).
3848 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3849 intptr_t& offset) {
3850 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3851 if (base == NULL) return NULL;
3852 return Ideal_allocation(base, phase);
3853 }
3854
3855 // Trace Initialize <- Proj[Parm] <- Allocate
3856 AllocateNode* InitializeNode::allocation() {
3857 Node* rawoop = in(InitializeNode::RawAddress);
3858 if (rawoop->is_Proj()) {
3859 Node* alloc = rawoop->in(0);
3860 if (alloc->is_Allocate()) {
3861 return alloc->as_Allocate();
3862 }
3863 }
3864 return NULL;
3865 }
3866
3867 // Trace Allocate -> Proj[Parm] -> Initialize
3868 InitializeNode* AllocateNode::initialization() {
3869 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3870 if (rawoop == NULL) return NULL;
3871 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3872 Node* init = rawoop->fast_out(i);
3873 if (init->is_Initialize()) {
3874 assert(init->as_Initialize()->allocation() == this, "2-way link");
3875 return init->as_Initialize();
3876 }
3877 }
3878 return NULL;
3879 }
3880
3881 //----------------------------- loop predicates ---------------------------
3882
3883 //------------------------------add_predicate_impl----------------------------
3884 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3885 // Too many traps seen?
3886 if (too_many_traps(reason)) {
3887 #ifdef ASSERT
3888 if (TraceLoopPredicate) {
3889 int tc = C->trap_count(reason);
3890 tty->print("too many traps=%s tcount=%d in ",
3891 Deoptimization::trap_reason_name(reason), tc);
3892 method()->print(); // which method has too many predicate traps
3893 tty->cr();
3894 }
3895 #endif
3896 // We cannot afford to take more traps here,
3897 // do not generate predicate.
3898 return;
3899 }
3900
3901 Node *cont = _gvn.intcon(1);
3902 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
3903 Node *bol = _gvn.transform(new Conv2BNode(opq));
3904 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3905 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3906 C->add_predicate_opaq(opq);
3907 {
3908 PreserveJVMState pjvms(this);
3909 set_control(iffalse);
3910 inc_sp(nargs);
3911 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3912 }
3913 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3914 set_control(iftrue);
3915 }
3916
3917 //------------------------------add_predicate---------------------------------
3918 void GraphKit::add_predicate(int nargs) {
3919 if (UseLoopPredicate) {
3920 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3921 }
3922 // loop's limit check predicate should be near the loop.
3923 if (LoopLimitCheck) {
3924 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3925 }
3926 }
3927
3928 //----------------------------- store barriers ----------------------------
3929 #define __ ideal.
3930
3931 void GraphKit::sync_kit(IdealKit& ideal) {
3932 set_all_memory(__ merged_memory());
3933 set_i_o(__ i_o());
3934 set_control(__ ctrl());
3935 }
3936
3937 void GraphKit::final_sync(IdealKit& ideal) {
3938 // Final sync IdealKit and graphKit.
3939 sync_kit(ideal);
3940 }
3941
3942 Node* GraphKit::byte_map_base_node() {
3943 // Get base of card map
3944 CardTableModRefBS* ct =
3945 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
3946 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code");
3947 if (ct->byte_map_base != NULL) {
3948 return makecon(TypeRawPtr::make((address)ct->byte_map_base));
3949 } else {
3950 return null();
3951 }
3952 }
3953
3954 // vanilla/CMS post barrier
3955 // Insert a write-barrier store. This is to let generational GC work; we have
3956 // to flag all oop-stores before the next GC point.
3957 void GraphKit::write_barrier_post(Node* oop_store,
3958 Node* obj,
3959 Node* adr,
3960 uint adr_idx,
3961 Node* val,
3962 bool use_precise) {
3963 // No store check needed if we're storing a NULL or an old object
3964 // (latter case is probably a string constant). The concurrent
3965 // mark sweep garbage collector, however, needs to have all nonNull
3966 // oop updates flagged via card-marks.
3967 if (val != NULL && val->is_Con()) {
3968 // must be either an oop or NULL
3969 const Type* t = val->bottom_type();
3970 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3971 // stores of null never (?) need barriers
3972 return;
3973 }
3974
3975 if (use_ReduceInitialCardMarks()
3976 && obj == just_allocated_object(control())) {
3977 // We can skip marks on a freshly-allocated object in Eden.
3978 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3979 // That routine informs GC to take appropriate compensating steps,
3980 // upon a slow-path allocation, so as to make this card-mark
3981 // elision safe.
3982 return;
3983 }
3984
3985 if (!use_precise) {
3986 // All card marks for a (non-array) instance are in one place:
3987 adr = obj;
3988 }
3989 // (Else it's an array (or unknown), and we want more precise card marks.)
3990 assert(adr != NULL, "");
3991
3992 IdealKit ideal(this, true);
3993
3994 // Convert the pointer to an int prior to doing math on it
3995 Node* cast = __ CastPX(__ ctrl(), adr);
3996
3997 // Divide by card size
3998 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef),
3999 "Only one we handle so far.");
4000 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4001
4002 // Combine card table base and card offset
4003 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
4004
4005 // Get the alias_index for raw card-mark memory
4006 int adr_type = Compile::AliasIdxRaw;
4007 Node* zero = __ ConI(0); // Dirty card value
4008 BasicType bt = T_BYTE;
4009
4010 if (UseConcMarkSweepGC && UseCondCardMark) {
4011 insert_store_load_for_barrier();
4012 __ sync_kit(this);
4013 }
4014
4015 if (UseCondCardMark) {
4016 // The classic GC reference write barrier is typically implemented
4017 // as a store into the global card mark table. Unfortunately
4018 // unconditional stores can result in false sharing and excessive
4019 // coherence traffic as well as false transactional aborts.
4020 // UseCondCardMark enables MP "polite" conditional card mark
4021 // stores. In theory we could relax the load from ctrl() to
4022 // no_ctrl, but that doesn't buy much latitude.
4023 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
4024 __ if_then(card_val, BoolTest::ne, zero);
4025 }
4026
4027 // Smash zero into card
4028 if( !UseConcMarkSweepGC ) {
4029 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
4030 } else {
4031 // Specialized path for CM store barrier
4032 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
4033 }
4034
4035 if (UseCondCardMark) {
4036 __ end_if();
4037 }
4038
4039 // Final sync IdealKit and GraphKit.
4040 final_sync(ideal);
4041 }
4042 /*
4043 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
4044 * required by SATB to make sure all objects live at the start of the
4045 * marking are kept alive, all reference updates need to any previous
4046 * reference stored before writing.
4047 *
4048 * If the previous value is NULL there is no need to save the old value.
4049 * References that are NULL are filtered during runtime by the barrier
4050 * code to avoid unnecessary queuing.
4051 *
4052 * However in the case of newly allocated objects it might be possible to
4053 * prove that the reference about to be overwritten is NULL during compile
4054 * time and avoid adding the barrier code completely.
4055 *
4056 * The compiler needs to determine that the object in which a field is about
4057 * to be written is newly allocated, and that no prior store to the same field
4058 * has happened since the allocation.
4059 *
4060 * Returns true if the pre-barrier can be removed
4061 */
4062 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
4063 BasicType bt, uint adr_idx) {
4064 intptr_t offset = 0;
4065 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4066 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4067
4068 if (offset == Type::OffsetBot) {
4069 return false; // cannot unalias unless there are precise offsets
4070 }
4071
4072 if (alloc == NULL) {
4073 return false; // No allocation found
4074 }
4075
4076 intptr_t size_in_bytes = type2aelembytes(bt);
4077
4078 Node* mem = memory(adr_idx); // start searching here...
4079
4080 for (int cnt = 0; cnt < 50; cnt++) {
4081
4082 if (mem->is_Store()) {
4083
4084 Node* st_adr = mem->in(MemNode::Address);
4085 intptr_t st_offset = 0;
4086 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
4087
4088 if (st_base == NULL) {
4089 break; // inscrutable pointer
4090 }
4091
4092 // Break we have found a store with same base and offset as ours so break
4093 if (st_base == base && st_offset == offset) {
4094 break;
4095 }
4096
4097 if (st_offset != offset && st_offset != Type::OffsetBot) {
4098 const int MAX_STORE = BytesPerLong;
4099 if (st_offset >= offset + size_in_bytes ||
4100 st_offset <= offset - MAX_STORE ||
4101 st_offset <= offset - mem->as_Store()->memory_size()) {
4102 // Success: The offsets are provably independent.
4103 // (You may ask, why not just test st_offset != offset and be done?
4104 // The answer is that stores of different sizes can co-exist
4105 // in the same sequence of RawMem effects. We sometimes initialize
4106 // a whole 'tile' of array elements with a single jint or jlong.)
4107 mem = mem->in(MemNode::Memory);
4108 continue; // advance through independent store memory
4109 }
4110 }
4111
4112 if (st_base != base
4113 && MemNode::detect_ptr_independence(base, alloc, st_base,
4114 AllocateNode::Ideal_allocation(st_base, phase),
4115 phase)) {
4116 // Success: The bases are provably independent.
4117 mem = mem->in(MemNode::Memory);
4118 continue; // advance through independent store memory
4119 }
4120 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4121
4122 InitializeNode* st_init = mem->in(0)->as_Initialize();
4123 AllocateNode* st_alloc = st_init->allocation();
4124
4125 // Make sure that we are looking at the same allocation site.
4126 // The alloc variable is guaranteed to not be null here from earlier check.
4127 if (alloc == st_alloc) {
4128 // Check that the initialization is storing NULL so that no previous store
4129 // has been moved up and directly write a reference
4130 Node* captured_store = st_init->find_captured_store(offset,
4131 type2aelembytes(T_OBJECT),
4132 phase);
4133 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4134 return true;
4135 }
4136 }
4137 }
4138
4139 // Unless there is an explicit 'continue', we must bail out here,
4140 // because 'mem' is an inscrutable memory state (e.g., a call).
4141 break;
4142 }
4143
4144 return false;
4145 }
4146
4147 // G1 pre/post barriers
4148 void GraphKit::g1_write_barrier_pre(bool do_load,
4149 Node* obj,
4150 Node* adr,
4151 uint alias_idx,
4152 Node* val,
4153 const TypeOopPtr* val_type,
4154 Node* pre_val,
4155 BasicType bt) {
4156
4157 // Some sanity checks
4158 // Note: val is unused in this routine.
4159
4160 if (do_load) {
4161 // We need to generate the load of the previous value
4162 assert(obj != NULL, "must have a base");
4163 assert(adr != NULL, "where are loading from?");
4164 assert(pre_val == NULL, "loaded already?");
4165 assert(val_type != NULL, "need a type");
4166
4167 if (use_ReduceInitialCardMarks()
4168 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4169 return;
4170 }
4171
4172 } else {
4173 // In this case both val_type and alias_idx are unused.
4174 assert(pre_val != NULL, "must be loaded already");
4175 // Nothing to be done if pre_val is null.
4176 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4177 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4178 }
4179 assert(bt == T_OBJECT || bt == T_VALUETYPE, "or we shouldn't be here");
4180
4181 IdealKit ideal(this, true);
4182
4183 Node* tls = __ thread(); // ThreadLocalStorage
4184
4185 Node* no_ctrl = NULL;
4186 Node* no_base = __ top();
4187 Node* zero = __ ConI(0);
4188 Node* zeroX = __ ConX(0);
4189
4190 float likely = PROB_LIKELY(0.999);
4191 float unlikely = PROB_UNLIKELY(0.999);
4192
4193 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4194 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4195
4196 // Offsets into the thread
4197 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
4198 SATBMarkQueue::byte_offset_of_active());
4199 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
4200 SATBMarkQueue::byte_offset_of_index());
4201 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
4202 SATBMarkQueue::byte_offset_of_buf());
4203
4204 // Now the actual pointers into the thread
4205 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4206 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4207 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4208
4209 // Now some of the values
4210 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4211
4212 // if (!marking)
4213 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4214 BasicType index_bt = TypeX_X->basic_type();
4215 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4216 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4217
4218 if (do_load) {
4219 // load original value
4220 // alias_idx correct??
4221 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4222 }
4223
4224 // if (pre_val != NULL)
4225 __ if_then(pre_val, BoolTest::ne, null()); {
4226 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4227
4228 // is the queue for this thread full?
4229 __ if_then(index, BoolTest::ne, zeroX, likely); {
4230
4231 // decrement the index
4232 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4233
4234 // Now get the buffer location we will log the previous value into and store it
4235 Node *log_addr = __ AddP(no_base, buffer, next_index);
4236 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4237 // update the index
4238 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4239
4240 } __ else_(); {
4241
4242 // logging buffer is full, call the runtime
4243 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4244 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4245 } __ end_if(); // (!index)
4246 } __ end_if(); // (pre_val != NULL)
4247 } __ end_if(); // (!marking)
4248
4249 // Final sync IdealKit and GraphKit.
4250 final_sync(ideal);
4251 }
4252
4253 /*
4254 * G1 similar to any GC with a Young Generation requires a way to keep track of
4255 * references from Old Generation to Young Generation to make sure all live
4256 * objects are found. G1 also requires to keep track of object references
4257 * between different regions to enable evacuation of old regions, which is done
4258 * as part of mixed collections. References are tracked in remembered sets and
4259 * is continuously updated as reference are written to with the help of the
4260 * post-barrier.
4261 *
4262 * To reduce the number of updates to the remembered set the post-barrier
4263 * filters updates to fields in objects located in the Young Generation,
4264 * the same region as the reference, when the NULL is being written or
4265 * if the card is already marked as dirty by an earlier write.
4266 *
4267 * Under certain circumstances it is possible to avoid generating the
4268 * post-barrier completely if it is possible during compile time to prove
4269 * the object is newly allocated and that no safepoint exists between the
4270 * allocation and the store.
4271 *
4272 * In the case of slow allocation the allocation code must handle the barrier
4273 * as part of the allocation in the case the allocated object is not located
4274 * in the nursery, this would happen for humongous objects. This is similar to
4275 * how CMS is required to handle this case, see the comments for the method
4276 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier.
4277 * A deferred card mark is required for these objects and handled in the above
4278 * mentioned methods.
4279 *
4280 * Returns true if the post barrier can be removed
4281 */
4282 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4283 Node* adr) {
4284 intptr_t offset = 0;
4285 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4286 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4287
4288 if (offset == Type::OffsetBot) {
4289 return false; // cannot unalias unless there are precise offsets
4290 }
4291
4292 if (alloc == NULL) {
4293 return false; // No allocation found
4294 }
4295
4296 // Start search from Store node
4297 Node* mem = store->in(MemNode::Control);
4298 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4299
4300 InitializeNode* st_init = mem->in(0)->as_Initialize();
4301 AllocateNode* st_alloc = st_init->allocation();
4302
4303 // Make sure we are looking at the same allocation
4304 if (alloc == st_alloc) {
4305 return true;
4306 }
4307 }
4308
4309 return false;
4310 }
4311
4312 //
4313 // Update the card table and add card address to the queue
4314 //
4315 void GraphKit::g1_mark_card(IdealKit& ideal,
4316 Node* card_adr,
4317 Node* oop_store,
4318 uint oop_alias_idx,
4319 Node* index,
4320 Node* index_adr,
4321 Node* buffer,
4322 const TypeFunc* tf) {
4323
4324 Node* zero = __ ConI(0);
4325 Node* zeroX = __ ConX(0);
4326 Node* no_base = __ top();
4327 BasicType card_bt = T_BYTE;
4328 // Smash zero into card. MUST BE ORDERED WRT TO STORE
4329 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4330
4331 // Now do the queue work
4332 __ if_then(index, BoolTest::ne, zeroX); {
4333
4334 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4335 Node* log_addr = __ AddP(no_base, buffer, next_index);
4336
4337 // Order, see storeCM.
4338 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4339 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4340
4341 } __ else_(); {
4342 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4343 } __ end_if();
4344
4345 }
4346
4347 void GraphKit::g1_write_barrier_post(Node* oop_store,
4348 Node* obj,
4349 Node* adr,
4350 uint alias_idx,
4351 Node* val,
4352 BasicType bt,
4353 bool use_precise) {
4354 // If we are writing a NULL then we need no post barrier
4355
4356 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4357 // Must be NULL
4358 const Type* t = val->bottom_type();
4359 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4360 // No post barrier if writing NULLx
4361 return;
4362 }
4363
4364 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4365 // We can skip marks on a freshly-allocated object in Eden.
4366 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
4367 // That routine informs GC to take appropriate compensating steps,
4368 // upon a slow-path allocation, so as to make this card-mark
4369 // elision safe.
4370 return;
4371 }
4372
4373 if (use_ReduceInitialCardMarks()
4374 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4375 return;
4376 }
4377
4378 if (!use_precise) {
4379 // All card marks for a (non-array) instance are in one place:
4380 adr = obj;
4381 }
4382 // (Else it's an array (or unknown), and we want more precise card marks.)
4383 assert(adr != NULL, "");
4384
4385 IdealKit ideal(this, true);
4386
4387 Node* tls = __ thread(); // ThreadLocalStorage
4388
4389 Node* no_base = __ top();
4390 float likely = PROB_LIKELY(0.999);
4391 float unlikely = PROB_UNLIKELY(0.999);
4392 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
4393 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
4394 Node* zeroX = __ ConX(0);
4395
4396 // Get the alias_index for raw card-mark memory
4397 const TypePtr* card_type = TypeRawPtr::BOTTOM;
4398
4399 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4400
4401 // Offsets into the thread
4402 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4403 DirtyCardQueue::byte_offset_of_index());
4404 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4405 DirtyCardQueue::byte_offset_of_buf());
4406
4407 // Pointers into the thread
4408
4409 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4410 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4411
4412 // Now some values
4413 // Use ctrl to avoid hoisting these values past a safepoint, which could
4414 // potentially reset these fields in the JavaThread.
4415 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4416 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4417
4418 // Convert the store obj pointer to an int prior to doing math on it
4419 // Must use ctrl to prevent "integerized oop" existing across safepoint
4420 Node* cast = __ CastPX(__ ctrl(), adr);
4421
4422 // Divide pointer by card size
4423 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4424
4425 // Combine card table base and card offset
4426 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4427
4428 // If we know the value being stored does it cross regions?
4429
4430 if (val != NULL) {
4431 // Does the store cause us to cross regions?
4432
4433 // Should be able to do an unsigned compare of region_size instead of
4434 // and extra shift. Do we have an unsigned compare??
4435 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4436 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4437
4438 // if (xor_res == 0) same region so skip
4439 __ if_then(xor_res, BoolTest::ne, zeroX); {
4440
4441 // No barrier if we are storing a NULL
4442 __ if_then(val, BoolTest::ne, null(), unlikely); {
4443
4444 // Ok must mark the card if not already dirty
4445
4446 // load the original value of the card
4447 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4448
4449 __ if_then(card_val, BoolTest::ne, young_card); {
4450 sync_kit(ideal);
4451 insert_store_load_for_barrier();
4452 __ sync_kit(this);
4453
4454 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4455 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4456 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4457 } __ end_if();
4458 } __ end_if();
4459 } __ end_if();
4460 } __ end_if();
4461 } else {
4462 // Object.clone() instrinsic uses this path.
4463 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4464 }
4465
4466 // Final sync IdealKit and GraphKit.
4467 final_sync(ideal);
4468 }
4469 #undef __
4470
4471
4472 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4473 Node* len = load_array_length(load_String_value(ctrl, str));
4474 Node* coder = load_String_coder(ctrl, str);
4475 // Divide length by 2 if coder is UTF16
4476 return _gvn.transform(new RShiftINode(len, coder));
4477 }
4478
4479 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4480 int value_offset = java_lang_String::value_offset_in_bytes();
4481 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4482 false, NULL, Type::Offset(0));
4483 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4484 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4485 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4486 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4487 int value_field_idx = C->get_alias_index(value_field_type);
4488 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4489 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4490 // String.value field is known to be @Stable.
4491 if (UseImplicitStableValues) {
4492 load = cast_array_to_stable(load, value_type);
4493 }
4494 return load;
4495 }
4496
4497 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4498 if (java_lang_String::has_coder_field()) {
4499 if (!CompactStrings) {
4500 return intcon(java_lang_String::CODER_UTF16);
4501 }
4502 int coder_offset = java_lang_String::coder_offset_in_bytes();
4503 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4504 false, NULL, Type::Offset(0));
4505 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4506 int coder_field_idx = C->get_alias_index(coder_field_type);
4507 return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4508 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4509 } else {
4510 return intcon(0); // false
4511 }
4512 }
4513
4514 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4515 int value_offset = java_lang_String::value_offset_in_bytes();
4516 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4517 false, NULL, Type::Offset(0));
4518 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4519 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4520 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4521 }
4522
4523 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4524 int coder_offset = java_lang_String::coder_offset_in_bytes();
4525 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4526 false, NULL, Type::Offset(0));
4527 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4528 int coder_field_idx = C->get_alias_index(coder_field_type);
4529 store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4530 value, T_BYTE, coder_field_idx, MemNode::unordered);
4531 }
4532
4533 // Capture src and dst memory state with a MergeMemNode
4534 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4535 if (src_type == dst_type) {
4536 // Types are equal, we don't need a MergeMemNode
4537 return memory(src_type);
4538 }
4539 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4540 record_for_igvn(merge); // fold it up later, if possible
4541 int src_idx = C->get_alias_index(src_type);
4542 int dst_idx = C->get_alias_index(dst_type);
4543 merge->set_memory_at(src_idx, memory(src_idx));
4544 merge->set_memory_at(dst_idx, memory(dst_idx));
4545 return merge;
4546 }
4547
4548 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4549 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4550 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4551 // If input and output memory types differ, capture both states to preserve
4552 // the dependency between preceding and subsequent loads/stores.
4553 // For example, the following program:
4554 // StoreB
4555 // compress_string
4556 // LoadB
4557 // has this memory graph (use->def):
4558 // LoadB -> compress_string -> CharMem
4559 // ... -> StoreB -> ByteMem
4560 // The intrinsic hides the dependency between LoadB and StoreB, causing
4561 // the load to read from memory not containing the result of the StoreB.
4562 // The correct memory graph should look like this:
4563 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4564 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4565 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4566 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4567 set_memory(res_mem, TypeAryPtr::BYTES);
4568 return str;
4569 }
4570
4571 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4572 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4573 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4574 // Capture src and dst memory (see comment in 'compress_string').
4575 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4576 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4577 set_memory(_gvn.transform(str), dst_type);
4578 }
4579
4580 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4581 /**
4582 * int i_char = start;
4583 * for (int i_byte = 0; i_byte < count; i_byte++) {
4584 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4585 * }
4586 */
4587 add_predicate();
4588 RegionNode* head = new RegionNode(3);
4589 head->init_req(1, control());
4590 gvn().set_type(head, Type::CONTROL);
4591 record_for_igvn(head);
4592
4593 Node* i_byte = new PhiNode(head, TypeInt::INT);
4594 i_byte->init_req(1, intcon(0));
4595 gvn().set_type(i_byte, TypeInt::INT);
4596 record_for_igvn(i_byte);
4597
4598 Node* i_char = new PhiNode(head, TypeInt::INT);
4599 i_char->init_req(1, start);
4600 gvn().set_type(i_char, TypeInt::INT);
4601 record_for_igvn(i_char);
4602
4603 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4604 gvn().set_type(mem, Type::MEMORY);
4605 record_for_igvn(mem);
4606 set_control(head);
4607 set_memory(mem, TypeAryPtr::BYTES);
4608 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4609 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4610 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4611 false, false, true /* mismatched */);
4612
4613 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4614 head->init_req(2, IfTrue(iff));
4615 mem->init_req(2, st);
4616 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4617 i_char->init_req(2, AddI(i_char, intcon(2)));
4618
4619 set_control(IfFalse(iff));
4620 set_memory(st, TypeAryPtr::BYTES);
4621 }
4622
4623 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4624 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4625 // assumption of CCP analysis.
4626 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4627 }