1 /*
2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/g1/g1BarrierSet.hpp"
29 #include "gc/g1/g1CardTable.hpp"
30 #include "gc/g1/heapRegion.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/cardTable.hpp"
33 #include "gc/shared/cardTableModRefBS.hpp"
34 #include "gc/shared/collectedHeap.hpp"
35 #include "memory/resourceArea.hpp"
36 #include "opto/addnode.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/graphKit.hpp"
40 #include "opto/idealKit.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/machnode.hpp"
44 #include "opto/opaquenode.hpp"
45 #include "opto/parse.hpp"
46 #include "opto/rootnode.hpp"
47 #include "opto/runtime.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/sharedRuntime.hpp"
50
51 //----------------------------GraphKit-----------------------------------------
52 // Main utility constructor.
53 GraphKit::GraphKit(JVMState* jvms)
54 : Phase(Phase::Parser),
55 _env(C->env()),
56 _gvn(*C->initial_gvn())
57 {
58 _exceptions = jvms->map()->next_exception();
59 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
60 set_jvms(jvms);
61 }
62
63 // Private constructor for parser.
64 GraphKit::GraphKit()
65 : Phase(Phase::Parser),
66 _env(C->env()),
67 _gvn(*C->initial_gvn())
68 {
69 _exceptions = NULL;
70 set_map(NULL);
71 debug_only(_sp = -99);
72 debug_only(set_bci(-99));
73 }
74
75
76
77 //---------------------------clean_stack---------------------------------------
78 // Clear away rubbish from the stack area of the JVM state.
79 // This destroys any arguments that may be waiting on the stack.
80 void GraphKit::clean_stack(int from_sp) {
81 SafePointNode* map = this->map();
82 JVMState* jvms = this->jvms();
83 int stk_size = jvms->stk_size();
84 int stkoff = jvms->stkoff();
85 Node* top = this->top();
86 for (int i = from_sp; i < stk_size; i++) {
87 if (map->in(stkoff + i) != top) {
88 map->set_req(stkoff + i, top);
89 }
90 }
91 }
92
93
94 //--------------------------------sync_jvms-----------------------------------
95 // Make sure our current jvms agrees with our parse state.
96 JVMState* GraphKit::sync_jvms() const {
97 JVMState* jvms = this->jvms();
98 jvms->set_bci(bci()); // Record the new bci in the JVMState
99 jvms->set_sp(sp()); // Record the new sp in the JVMState
100 assert(jvms_in_sync(), "jvms is now in sync");
101 return jvms;
102 }
103
104 //--------------------------------sync_jvms_for_reexecute---------------------
105 // Make sure our current jvms agrees with our parse state. This version
106 // uses the reexecute_sp for reexecuting bytecodes.
107 JVMState* GraphKit::sync_jvms_for_reexecute() {
108 JVMState* jvms = this->jvms();
109 jvms->set_bci(bci()); // Record the new bci in the JVMState
110 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
111 return jvms;
112 }
113
114 #ifdef ASSERT
115 bool GraphKit::jvms_in_sync() const {
116 Parse* parse = is_Parse();
117 if (parse == NULL) {
118 if (bci() != jvms()->bci()) return false;
119 if (sp() != (int)jvms()->sp()) return false;
120 return true;
121 }
122 if (jvms()->method() != parse->method()) return false;
123 if (jvms()->bci() != parse->bci()) return false;
124 int jvms_sp = jvms()->sp();
125 if (jvms_sp != parse->sp()) return false;
126 int jvms_depth = jvms()->depth();
127 if (jvms_depth != parse->depth()) return false;
128 return true;
129 }
130
131 // Local helper checks for special internal merge points
132 // used to accumulate and merge exception states.
133 // They are marked by the region's in(0) edge being the map itself.
134 // Such merge points must never "escape" into the parser at large,
135 // until they have been handed to gvn.transform.
136 static bool is_hidden_merge(Node* reg) {
137 if (reg == NULL) return false;
138 if (reg->is_Phi()) {
139 reg = reg->in(0);
140 if (reg == NULL) return false;
141 }
142 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
143 }
144
145 void GraphKit::verify_map() const {
146 if (map() == NULL) return; // null map is OK
147 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
148 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
149 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
150 }
151
152 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
153 assert(ex_map->next_exception() == NULL, "not already part of a chain");
154 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
155 }
156 #endif
157
158 //---------------------------stop_and_kill_map---------------------------------
159 // Set _map to NULL, signalling a stop to further bytecode execution.
160 // First smash the current map's control to a constant, to mark it dead.
161 void GraphKit::stop_and_kill_map() {
162 SafePointNode* dead_map = stop();
163 if (dead_map != NULL) {
164 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
165 assert(dead_map->is_killed(), "must be so marked");
166 }
167 }
168
169
170 //--------------------------------stopped--------------------------------------
171 // Tell if _map is NULL, or control is top.
172 bool GraphKit::stopped() {
173 if (map() == NULL) return true;
174 else if (control() == top()) return true;
175 else return false;
176 }
177
178
179 //-----------------------------has_ex_handler----------------------------------
180 // Tell if this method or any caller method has exception handlers.
181 bool GraphKit::has_ex_handler() {
182 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
183 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
184 return true;
185 }
186 }
187 return false;
188 }
189
190 //------------------------------save_ex_oop------------------------------------
191 // Save an exception without blowing stack contents or other JVM state.
192 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
193 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
194 ex_map->add_req(ex_oop);
195 debug_only(verify_exception_state(ex_map));
196 }
197
198 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
199 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
200 Node* ex_oop = ex_map->in(ex_map->req()-1);
201 if (clear_it) ex_map->del_req(ex_map->req()-1);
202 return ex_oop;
203 }
204
205 //-----------------------------saved_ex_oop------------------------------------
206 // Recover a saved exception from its map.
207 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
208 return common_saved_ex_oop(ex_map, false);
209 }
210
211 //--------------------------clear_saved_ex_oop---------------------------------
212 // Erase a previously saved exception from its map.
213 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
214 return common_saved_ex_oop(ex_map, true);
215 }
216
217 #ifdef ASSERT
218 //---------------------------has_saved_ex_oop----------------------------------
219 // Erase a previously saved exception from its map.
220 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
221 return ex_map->req() == ex_map->jvms()->endoff()+1;
222 }
223 #endif
224
225 //-------------------------make_exception_state--------------------------------
226 // Turn the current JVM state into an exception state, appending the ex_oop.
227 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
228 sync_jvms();
229 SafePointNode* ex_map = stop(); // do not manipulate this map any more
230 set_saved_ex_oop(ex_map, ex_oop);
231 return ex_map;
232 }
233
234
235 //--------------------------add_exception_state--------------------------------
236 // Add an exception to my list of exceptions.
237 void GraphKit::add_exception_state(SafePointNode* ex_map) {
238 if (ex_map == NULL || ex_map->control() == top()) {
239 return;
240 }
241 #ifdef ASSERT
242 verify_exception_state(ex_map);
243 if (has_exceptions()) {
244 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
245 }
246 #endif
247
248 // If there is already an exception of exactly this type, merge with it.
249 // In particular, null-checks and other low-level exceptions common up here.
250 Node* ex_oop = saved_ex_oop(ex_map);
251 const Type* ex_type = _gvn.type(ex_oop);
252 if (ex_oop == top()) {
253 // No action needed.
254 return;
255 }
256 assert(ex_type->isa_instptr(), "exception must be an instance");
257 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
258 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
259 // We check sp also because call bytecodes can generate exceptions
260 // both before and after arguments are popped!
261 if (ex_type2 == ex_type
262 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
263 combine_exception_states(ex_map, e2);
264 return;
265 }
266 }
267
268 // No pre-existing exception of the same type. Chain it on the list.
269 push_exception_state(ex_map);
270 }
271
272 //-----------------------add_exception_states_from-----------------------------
273 void GraphKit::add_exception_states_from(JVMState* jvms) {
274 SafePointNode* ex_map = jvms->map()->next_exception();
275 if (ex_map != NULL) {
276 jvms->map()->set_next_exception(NULL);
277 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
278 next_map = ex_map->next_exception();
279 ex_map->set_next_exception(NULL);
280 add_exception_state(ex_map);
281 }
282 }
283 }
284
285 //-----------------------transfer_exceptions_into_jvms-------------------------
286 JVMState* GraphKit::transfer_exceptions_into_jvms() {
287 if (map() == NULL) {
288 // We need a JVMS to carry the exceptions, but the map has gone away.
289 // Create a scratch JVMS, cloned from any of the exception states...
290 if (has_exceptions()) {
291 _map = _exceptions;
292 _map = clone_map();
293 _map->set_next_exception(NULL);
294 clear_saved_ex_oop(_map);
295 debug_only(verify_map());
296 } else {
297 // ...or created from scratch
298 JVMState* jvms = new (C) JVMState(_method, NULL);
299 jvms->set_bci(_bci);
300 jvms->set_sp(_sp);
301 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
302 set_jvms(jvms);
303 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
304 set_all_memory(top());
305 while (map()->req() < jvms->endoff()) map()->add_req(top());
306 }
307 // (This is a kludge, in case you didn't notice.)
308 set_control(top());
309 }
310 JVMState* jvms = sync_jvms();
311 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
312 jvms->map()->set_next_exception(_exceptions);
313 _exceptions = NULL; // done with this set of exceptions
314 return jvms;
315 }
316
317 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
318 assert(is_hidden_merge(dstphi), "must be a special merge node");
319 assert(is_hidden_merge(srcphi), "must be a special merge node");
320 uint limit = srcphi->req();
321 for (uint i = PhiNode::Input; i < limit; i++) {
322 dstphi->add_req(srcphi->in(i));
323 }
324 }
325 static inline void add_one_req(Node* dstphi, Node* src) {
326 assert(is_hidden_merge(dstphi), "must be a special merge node");
327 assert(!is_hidden_merge(src), "must not be a special merge node");
328 dstphi->add_req(src);
329 }
330
331 //-----------------------combine_exception_states------------------------------
332 // This helper function combines exception states by building phis on a
333 // specially marked state-merging region. These regions and phis are
334 // untransformed, and can build up gradually. The region is marked by
335 // having a control input of its exception map, rather than NULL. Such
336 // regions do not appear except in this function, and in use_exception_state.
337 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
338 if (failing()) return; // dying anyway...
339 JVMState* ex_jvms = ex_map->_jvms;
340 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
341 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
342 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
343 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
344 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
345 assert(ex_map->req() == phi_map->req(), "matching maps");
346 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
347 Node* hidden_merge_mark = root();
348 Node* region = phi_map->control();
349 MergeMemNode* phi_mem = phi_map->merged_memory();
350 MergeMemNode* ex_mem = ex_map->merged_memory();
351 if (region->in(0) != hidden_merge_mark) {
352 // The control input is not (yet) a specially-marked region in phi_map.
353 // Make it so, and build some phis.
354 region = new RegionNode(2);
355 _gvn.set_type(region, Type::CONTROL);
356 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
357 region->init_req(1, phi_map->control());
358 phi_map->set_control(region);
359 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
360 record_for_igvn(io_phi);
361 _gvn.set_type(io_phi, Type::ABIO);
362 phi_map->set_i_o(io_phi);
363 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
364 Node* m = mms.memory();
365 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
366 record_for_igvn(m_phi);
367 _gvn.set_type(m_phi, Type::MEMORY);
368 mms.set_memory(m_phi);
369 }
370 }
371
372 // Either or both of phi_map and ex_map might already be converted into phis.
373 Node* ex_control = ex_map->control();
374 // if there is special marking on ex_map also, we add multiple edges from src
375 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
376 // how wide was the destination phi_map, originally?
377 uint orig_width = region->req();
378
379 if (add_multiple) {
380 add_n_reqs(region, ex_control);
381 add_n_reqs(phi_map->i_o(), ex_map->i_o());
382 } else {
383 // ex_map has no merges, so we just add single edges everywhere
384 add_one_req(region, ex_control);
385 add_one_req(phi_map->i_o(), ex_map->i_o());
386 }
387 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
388 if (mms.is_empty()) {
389 // get a copy of the base memory, and patch some inputs into it
390 const TypePtr* adr_type = mms.adr_type(C);
391 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
392 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
393 mms.set_memory(phi);
394 // Prepare to append interesting stuff onto the newly sliced phi:
395 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
396 }
397 // Append stuff from ex_map:
398 if (add_multiple) {
399 add_n_reqs(mms.memory(), mms.memory2());
400 } else {
401 add_one_req(mms.memory(), mms.memory2());
402 }
403 }
404 uint limit = ex_map->req();
405 for (uint i = TypeFunc::Parms; i < limit; i++) {
406 // Skip everything in the JVMS after tos. (The ex_oop follows.)
407 if (i == tos) i = ex_jvms->monoff();
408 Node* src = ex_map->in(i);
409 Node* dst = phi_map->in(i);
410 if (src != dst) {
411 PhiNode* phi;
412 if (dst->in(0) != region) {
413 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
414 record_for_igvn(phi);
415 _gvn.set_type(phi, phi->type());
416 phi_map->set_req(i, dst);
417 // Prepare to append interesting stuff onto the new phi:
418 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
419 } else {
420 assert(dst->is_Phi(), "nobody else uses a hidden region");
421 phi = dst->as_Phi();
422 }
423 if (add_multiple && src->in(0) == ex_control) {
424 // Both are phis.
425 add_n_reqs(dst, src);
426 } else {
427 while (dst->req() < region->req()) add_one_req(dst, src);
428 }
429 const Type* srctype = _gvn.type(src);
430 if (phi->type() != srctype) {
431 const Type* dsttype = phi->type()->meet_speculative(srctype);
432 if (phi->type() != dsttype) {
433 phi->set_type(dsttype);
434 _gvn.set_type(phi, dsttype);
435 }
436 }
437 }
438 }
439 phi_map->merge_replaced_nodes_with(ex_map);
440 }
441
442 //--------------------------use_exception_state--------------------------------
443 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
444 if (failing()) { stop(); return top(); }
445 Node* region = phi_map->control();
446 Node* hidden_merge_mark = root();
447 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
448 Node* ex_oop = clear_saved_ex_oop(phi_map);
449 if (region->in(0) == hidden_merge_mark) {
450 // Special marking for internal ex-states. Process the phis now.
451 region->set_req(0, region); // now it's an ordinary region
452 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
453 // Note: Setting the jvms also sets the bci and sp.
454 set_control(_gvn.transform(region));
455 uint tos = jvms()->stkoff() + sp();
456 for (uint i = 1; i < tos; i++) {
457 Node* x = phi_map->in(i);
458 if (x->in(0) == region) {
459 assert(x->is_Phi(), "expected a special phi");
460 phi_map->set_req(i, _gvn.transform(x));
461 }
462 }
463 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
464 Node* x = mms.memory();
465 if (x->in(0) == region) {
466 assert(x->is_Phi(), "nobody else uses a hidden region");
467 mms.set_memory(_gvn.transform(x));
468 }
469 }
470 if (ex_oop->in(0) == region) {
471 assert(ex_oop->is_Phi(), "expected a special phi");
472 ex_oop = _gvn.transform(ex_oop);
473 }
474 } else {
475 set_jvms(phi_map->jvms());
476 }
477
478 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
479 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
480 return ex_oop;
481 }
482
483 //---------------------------------java_bc-------------------------------------
484 Bytecodes::Code GraphKit::java_bc() const {
485 ciMethod* method = this->method();
486 int bci = this->bci();
487 if (method != NULL && bci != InvocationEntryBci)
488 return method->java_code_at_bci(bci);
489 else
490 return Bytecodes::_illegal;
491 }
492
493 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
494 bool must_throw) {
495 // if the exception capability is set, then we will generate code
496 // to check the JavaThread.should_post_on_exceptions flag to see
497 // if we actually need to report exception events (for this
498 // thread). If we don't need to report exception events, we will
499 // take the normal fast path provided by add_exception_events. If
500 // exception event reporting is enabled for this thread, we will
501 // take the uncommon_trap in the BuildCutout below.
502
503 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
504 Node* jthread = _gvn.transform(new ThreadLocalNode());
505 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
506 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
507
508 // Test the should_post_on_exceptions_flag vs. 0
509 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
510 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
511
512 // Branch to slow_path if should_post_on_exceptions_flag was true
513 { BuildCutout unless(this, tst, PROB_MAX);
514 // Do not try anything fancy if we're notifying the VM on every throw.
515 // Cf. case Bytecodes::_athrow in parse2.cpp.
516 uncommon_trap(reason, Deoptimization::Action_none,
517 (ciKlass*)NULL, (char*)NULL, must_throw);
518 }
519
520 }
521
522 //------------------------------builtin_throw----------------------------------
523 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
524 bool must_throw = true;
525
526 if (env()->jvmti_can_post_on_exceptions()) {
527 // check if we must post exception events, take uncommon trap if so
528 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
529 // here if should_post_on_exceptions is false
530 // continue on with the normal codegen
531 }
532
533 // If this particular condition has not yet happened at this
534 // bytecode, then use the uncommon trap mechanism, and allow for
535 // a future recompilation if several traps occur here.
536 // If the throw is hot, try to use a more complicated inline mechanism
537 // which keeps execution inside the compiled code.
538 bool treat_throw_as_hot = false;
539 ciMethodData* md = method()->method_data();
540
541 if (ProfileTraps) {
542 if (too_many_traps(reason)) {
543 treat_throw_as_hot = true;
544 }
545 // (If there is no MDO at all, assume it is early in
546 // execution, and that any deopts are part of the
547 // startup transient, and don't need to be remembered.)
548
549 // Also, if there is a local exception handler, treat all throws
550 // as hot if there has been at least one in this method.
551 if (C->trap_count(reason) != 0
552 && method()->method_data()->trap_count(reason) != 0
553 && has_ex_handler()) {
554 treat_throw_as_hot = true;
555 }
556 }
557
558 // If this throw happens frequently, an uncommon trap might cause
559 // a performance pothole. If there is a local exception handler,
560 // and if this particular bytecode appears to be deoptimizing often,
561 // let us handle the throw inline, with a preconstructed instance.
562 // Note: If the deopt count has blown up, the uncommon trap
563 // runtime is going to flush this nmethod, not matter what.
564 if (treat_throw_as_hot
565 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
566 // If the throw is local, we use a pre-existing instance and
567 // punt on the backtrace. This would lead to a missing backtrace
568 // (a repeat of 4292742) if the backtrace object is ever asked
569 // for its backtrace.
570 // Fixing this remaining case of 4292742 requires some flavor of
571 // escape analysis. Leave that for the future.
572 ciInstance* ex_obj = NULL;
573 switch (reason) {
574 case Deoptimization::Reason_null_check:
575 ex_obj = env()->NullPointerException_instance();
576 break;
577 case Deoptimization::Reason_div0_check:
578 ex_obj = env()->ArithmeticException_instance();
579 break;
580 case Deoptimization::Reason_range_check:
581 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
582 break;
583 case Deoptimization::Reason_class_check:
584 if (java_bc() == Bytecodes::_aastore) {
585 ex_obj = env()->ArrayStoreException_instance();
586 } else {
587 ex_obj = env()->ClassCastException_instance();
588 }
589 break;
590 default:
591 break;
592 }
593 if (failing()) { stop(); return; } // exception allocation might fail
594 if (ex_obj != NULL) {
595 // Cheat with a preallocated exception object.
596 if (C->log() != NULL)
597 C->log()->elem("hot_throw preallocated='1' reason='%s'",
598 Deoptimization::trap_reason_name(reason));
599 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
600 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
601
602 // Clear the detail message of the preallocated exception object.
603 // Weblogic sometimes mutates the detail message of exceptions
604 // using reflection.
605 int offset = java_lang_Throwable::get_detailMessage_offset();
606 const TypePtr* adr_typ = ex_con->add_offset(offset);
607
608 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
609 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
610 // Conservatively release stores of object references.
611 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
612
613 add_exception_state(make_exception_state(ex_node));
614 return;
615 }
616 }
617
618 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
619 // It won't be much cheaper than bailing to the interp., since we'll
620 // have to pass up all the debug-info, and the runtime will have to
621 // create the stack trace.
622
623 // Usual case: Bail to interpreter.
624 // Reserve the right to recompile if we haven't seen anything yet.
625
626 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
627 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
628 if (treat_throw_as_hot
629 && (method()->method_data()->trap_recompiled_at(bci(), m)
630 || C->too_many_traps(reason))) {
631 // We cannot afford to take more traps here. Suffer in the interpreter.
632 if (C->log() != NULL)
633 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
634 Deoptimization::trap_reason_name(reason),
635 C->trap_count(reason));
636 action = Deoptimization::Action_none;
637 }
638
639 // "must_throw" prunes the JVM state to include only the stack, if there
640 // are no local exception handlers. This should cut down on register
641 // allocation time and code size, by drastically reducing the number
642 // of in-edges on the call to the uncommon trap.
643
644 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
645 }
646
647
648 //----------------------------PreserveJVMState---------------------------------
649 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
650 debug_only(kit->verify_map());
651 _kit = kit;
652 _map = kit->map(); // preserve the map
653 _sp = kit->sp();
654 kit->set_map(clone_map ? kit->clone_map() : NULL);
655 #ifdef ASSERT
656 _bci = kit->bci();
657 Parse* parser = kit->is_Parse();
658 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
659 _block = block;
660 #endif
661 }
662 PreserveJVMState::~PreserveJVMState() {
663 GraphKit* kit = _kit;
664 #ifdef ASSERT
665 assert(kit->bci() == _bci, "bci must not shift");
666 Parse* parser = kit->is_Parse();
667 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
668 assert(block == _block, "block must not shift");
669 #endif
670 kit->set_map(_map);
671 kit->set_sp(_sp);
672 }
673
674
675 //-----------------------------BuildCutout-------------------------------------
676 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
677 : PreserveJVMState(kit)
678 {
679 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
680 SafePointNode* outer_map = _map; // preserved map is caller's
681 SafePointNode* inner_map = kit->map();
682 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
683 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
684 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
685 }
686 BuildCutout::~BuildCutout() {
687 GraphKit* kit = _kit;
688 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
689 }
690
691 //---------------------------PreserveReexecuteState----------------------------
692 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
693 assert(!kit->stopped(), "must call stopped() before");
694 _kit = kit;
695 _sp = kit->sp();
696 _reexecute = kit->jvms()->_reexecute;
697 }
698 PreserveReexecuteState::~PreserveReexecuteState() {
699 if (_kit->stopped()) return;
700 _kit->jvms()->_reexecute = _reexecute;
701 _kit->set_sp(_sp);
702 }
703
704 //------------------------------clone_map--------------------------------------
705 // Implementation of PreserveJVMState
706 //
707 // Only clone_map(...) here. If this function is only used in the
708 // PreserveJVMState class we may want to get rid of this extra
709 // function eventually and do it all there.
710
711 SafePointNode* GraphKit::clone_map() {
712 if (map() == NULL) return NULL;
713
714 // Clone the memory edge first
715 Node* mem = MergeMemNode::make(map()->memory());
716 gvn().set_type_bottom(mem);
717
718 SafePointNode *clonemap = (SafePointNode*)map()->clone();
719 JVMState* jvms = this->jvms();
720 JVMState* clonejvms = jvms->clone_shallow(C);
721 clonemap->set_memory(mem);
722 clonemap->set_jvms(clonejvms);
723 clonejvms->set_map(clonemap);
724 record_for_igvn(clonemap);
725 gvn().set_type_bottom(clonemap);
726 return clonemap;
727 }
728
729
730 //-----------------------------set_map_clone-----------------------------------
731 void GraphKit::set_map_clone(SafePointNode* m) {
732 _map = m;
733 _map = clone_map();
734 _map->set_next_exception(NULL);
735 debug_only(verify_map());
736 }
737
738
739 //----------------------------kill_dead_locals---------------------------------
740 // Detect any locals which are known to be dead, and force them to top.
741 void GraphKit::kill_dead_locals() {
742 // Consult the liveness information for the locals. If any
743 // of them are unused, then they can be replaced by top(). This
744 // should help register allocation time and cut down on the size
745 // of the deoptimization information.
746
747 // This call is made from many of the bytecode handling
748 // subroutines called from the Big Switch in do_one_bytecode.
749 // Every bytecode which might include a slow path is responsible
750 // for killing its dead locals. The more consistent we
751 // are about killing deads, the fewer useless phis will be
752 // constructed for them at various merge points.
753
754 // bci can be -1 (InvocationEntryBci). We return the entry
755 // liveness for the method.
756
757 if (method() == NULL || method()->code_size() == 0) {
758 // We are building a graph for a call to a native method.
759 // All locals are live.
760 return;
761 }
762
763 ResourceMark rm;
764
765 // Consult the liveness information for the locals. If any
766 // of them are unused, then they can be replaced by top(). This
767 // should help register allocation time and cut down on the size
768 // of the deoptimization information.
769 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
770
771 int len = (int)live_locals.size();
772 assert(len <= jvms()->loc_size(), "too many live locals");
773 for (int local = 0; local < len; local++) {
774 if (!live_locals.at(local)) {
775 set_local(local, top());
776 }
777 }
778 }
779
780 #ifdef ASSERT
781 //-------------------------dead_locals_are_killed------------------------------
782 // Return true if all dead locals are set to top in the map.
783 // Used to assert "clean" debug info at various points.
784 bool GraphKit::dead_locals_are_killed() {
785 if (method() == NULL || method()->code_size() == 0) {
786 // No locals need to be dead, so all is as it should be.
787 return true;
788 }
789
790 // Make sure somebody called kill_dead_locals upstream.
791 ResourceMark rm;
792 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
793 if (jvms->loc_size() == 0) continue; // no locals to consult
794 SafePointNode* map = jvms->map();
795 ciMethod* method = jvms->method();
796 int bci = jvms->bci();
797 if (jvms == this->jvms()) {
798 bci = this->bci(); // it might not yet be synched
799 }
800 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
801 int len = (int)live_locals.size();
802 if (!live_locals.is_valid() || len == 0)
803 // This method is trivial, or is poisoned by a breakpoint.
804 return true;
805 assert(len == jvms->loc_size(), "live map consistent with locals map");
806 for (int local = 0; local < len; local++) {
807 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
808 if (PrintMiscellaneous && (Verbose || WizardMode)) {
809 tty->print_cr("Zombie local %d: ", local);
810 jvms->dump();
811 }
812 return false;
813 }
814 }
815 }
816 return true;
817 }
818
819 #endif //ASSERT
820
821 // Helper function for enforcing certain bytecodes to reexecute if
822 // deoptimization happens
823 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
824 ciMethod* cur_method = jvms->method();
825 int cur_bci = jvms->bci();
826 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
827 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
828 return Interpreter::bytecode_should_reexecute(code) ||
829 (is_anewarray && code == Bytecodes::_multianewarray);
830 // Reexecute _multianewarray bytecode which was replaced with
831 // sequence of [a]newarray. See Parse::do_multianewarray().
832 //
833 // Note: interpreter should not have it set since this optimization
834 // is limited by dimensions and guarded by flag so in some cases
835 // multianewarray() runtime calls will be generated and
836 // the bytecode should not be reexecutes (stack will not be reset).
837 } else
838 return false;
839 }
840
841 // Helper function for adding JVMState and debug information to node
842 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
843 // Add the safepoint edges to the call (or other safepoint).
844
845 // Make sure dead locals are set to top. This
846 // should help register allocation time and cut down on the size
847 // of the deoptimization information.
848 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
849
850 // Walk the inline list to fill in the correct set of JVMState's
851 // Also fill in the associated edges for each JVMState.
852
853 // If the bytecode needs to be reexecuted we need to put
854 // the arguments back on the stack.
855 const bool should_reexecute = jvms()->should_reexecute();
856 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
857
858 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
859 // undefined if the bci is different. This is normal for Parse but it
860 // should not happen for LibraryCallKit because only one bci is processed.
861 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
862 "in LibraryCallKit the reexecute bit should not change");
863
864 // If we are guaranteed to throw, we can prune everything but the
865 // input to the current bytecode.
866 bool can_prune_locals = false;
867 uint stack_slots_not_pruned = 0;
868 int inputs = 0, depth = 0;
869 if (must_throw) {
870 assert(method() == youngest_jvms->method(), "sanity");
871 if (compute_stack_effects(inputs, depth)) {
872 can_prune_locals = true;
873 stack_slots_not_pruned = inputs;
874 }
875 }
876
877 if (env()->should_retain_local_variables()) {
878 // At any safepoint, this method can get breakpointed, which would
879 // then require an immediate deoptimization.
880 can_prune_locals = false; // do not prune locals
881 stack_slots_not_pruned = 0;
882 }
883
884 // do not scribble on the input jvms
885 JVMState* out_jvms = youngest_jvms->clone_deep(C);
886 call->set_jvms(out_jvms); // Start jvms list for call node
887
888 // For a known set of bytecodes, the interpreter should reexecute them if
889 // deoptimization happens. We set the reexecute state for them here
890 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
891 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
892 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
893 }
894
895 // Presize the call:
896 DEBUG_ONLY(uint non_debug_edges = call->req());
897 call->add_req_batch(top(), youngest_jvms->debug_depth());
898 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
899
900 // Set up edges so that the call looks like this:
901 // Call [state:] ctl io mem fptr retadr
902 // [parms:] parm0 ... parmN
903 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
904 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
905 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
906 // Note that caller debug info precedes callee debug info.
907
908 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
909 uint debug_ptr = call->req();
910
911 // Loop over the map input edges associated with jvms, add them
912 // to the call node, & reset all offsets to match call node array.
913 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
914 uint debug_end = debug_ptr;
915 uint debug_start = debug_ptr - in_jvms->debug_size();
916 debug_ptr = debug_start; // back up the ptr
917
918 uint p = debug_start; // walks forward in [debug_start, debug_end)
919 uint j, k, l;
920 SafePointNode* in_map = in_jvms->map();
921 out_jvms->set_map(call);
922
923 if (can_prune_locals) {
924 assert(in_jvms->method() == out_jvms->method(), "sanity");
925 // If the current throw can reach an exception handler in this JVMS,
926 // then we must keep everything live that can reach that handler.
927 // As a quick and dirty approximation, we look for any handlers at all.
928 if (in_jvms->method()->has_exception_handlers()) {
929 can_prune_locals = false;
930 }
931 }
932
933 // Add the Locals
934 k = in_jvms->locoff();
935 l = in_jvms->loc_size();
936 out_jvms->set_locoff(p);
937 if (!can_prune_locals) {
938 for (j = 0; j < l; j++)
939 call->set_req(p++, in_map->in(k+j));
940 } else {
941 p += l; // already set to top above by add_req_batch
942 }
943
944 // Add the Expression Stack
945 k = in_jvms->stkoff();
946 l = in_jvms->sp();
947 out_jvms->set_stkoff(p);
948 if (!can_prune_locals) {
949 for (j = 0; j < l; j++)
950 call->set_req(p++, in_map->in(k+j));
951 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
952 // Divide stack into {S0,...,S1}, where S0 is set to top.
953 uint s1 = stack_slots_not_pruned;
954 stack_slots_not_pruned = 0; // for next iteration
955 if (s1 > l) s1 = l;
956 uint s0 = l - s1;
957 p += s0; // skip the tops preinstalled by add_req_batch
958 for (j = s0; j < l; j++)
959 call->set_req(p++, in_map->in(k+j));
960 } else {
961 p += l; // already set to top above by add_req_batch
962 }
963
964 // Add the Monitors
965 k = in_jvms->monoff();
966 l = in_jvms->mon_size();
967 out_jvms->set_monoff(p);
968 for (j = 0; j < l; j++)
969 call->set_req(p++, in_map->in(k+j));
970
971 // Copy any scalar object fields.
972 k = in_jvms->scloff();
973 l = in_jvms->scl_size();
974 out_jvms->set_scloff(p);
975 for (j = 0; j < l; j++)
976 call->set_req(p++, in_map->in(k+j));
977
978 // Finish the new jvms.
979 out_jvms->set_endoff(p);
980
981 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
982 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
983 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
984 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
985 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
986 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
987
988 // Update the two tail pointers in parallel.
989 out_jvms = out_jvms->caller();
990 in_jvms = in_jvms->caller();
991 }
992
993 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
994
995 // Test the correctness of JVMState::debug_xxx accessors:
996 assert(call->jvms()->debug_start() == non_debug_edges, "");
997 assert(call->jvms()->debug_end() == call->req(), "");
998 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
999 }
1000
1001 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1002 Bytecodes::Code code = java_bc();
1003 if (code == Bytecodes::_wide) {
1004 code = method()->java_code_at_bci(bci() + 1);
1005 }
1006
1007 BasicType rtype = T_ILLEGAL;
1008 int rsize = 0;
1009
1010 if (code != Bytecodes::_illegal) {
1011 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1012 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1013 if (rtype < T_CONFLICT)
1014 rsize = type2size[rtype];
1015 }
1016
1017 switch (code) {
1018 case Bytecodes::_illegal:
1019 return false;
1020
1021 case Bytecodes::_ldc:
1022 case Bytecodes::_ldc_w:
1023 case Bytecodes::_ldc2_w:
1024 inputs = 0;
1025 break;
1026
1027 case Bytecodes::_dup: inputs = 1; break;
1028 case Bytecodes::_dup_x1: inputs = 2; break;
1029 case Bytecodes::_dup_x2: inputs = 3; break;
1030 case Bytecodes::_dup2: inputs = 2; break;
1031 case Bytecodes::_dup2_x1: inputs = 3; break;
1032 case Bytecodes::_dup2_x2: inputs = 4; break;
1033 case Bytecodes::_swap: inputs = 2; break;
1034 case Bytecodes::_arraylength: inputs = 1; break;
1035
1036 case Bytecodes::_getstatic:
1037 case Bytecodes::_putstatic:
1038 case Bytecodes::_getfield:
1039 case Bytecodes::_putfield:
1040 {
1041 bool ignored_will_link;
1042 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1043 int size = field->type()->size();
1044 bool is_get = (depth >= 0), is_static = (depth & 1);
1045 inputs = (is_static ? 0 : 1);
1046 if (is_get) {
1047 depth = size - inputs;
1048 } else {
1049 inputs += size; // putxxx pops the value from the stack
1050 depth = - inputs;
1051 }
1052 }
1053 break;
1054
1055 case Bytecodes::_invokevirtual:
1056 case Bytecodes::_invokespecial:
1057 case Bytecodes::_invokestatic:
1058 case Bytecodes::_invokedynamic:
1059 case Bytecodes::_invokeinterface:
1060 {
1061 bool ignored_will_link;
1062 ciSignature* declared_signature = NULL;
1063 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1064 assert(declared_signature != NULL, "cannot be null");
1065 inputs = declared_signature->arg_size_for_bc(code);
1066 int size = declared_signature->return_type()->size();
1067 depth = size - inputs;
1068 }
1069 break;
1070
1071 case Bytecodes::_multianewarray:
1072 {
1073 ciBytecodeStream iter(method());
1074 iter.reset_to_bci(bci());
1075 iter.next();
1076 inputs = iter.get_dimensions();
1077 assert(rsize == 1, "");
1078 depth = rsize - inputs;
1079 }
1080 break;
1081
1082 case Bytecodes::_ireturn:
1083 case Bytecodes::_lreturn:
1084 case Bytecodes::_freturn:
1085 case Bytecodes::_dreturn:
1086 case Bytecodes::_areturn:
1087 assert(rsize == -depth, "");
1088 inputs = rsize;
1089 break;
1090
1091 case Bytecodes::_jsr:
1092 case Bytecodes::_jsr_w:
1093 inputs = 0;
1094 depth = 1; // S.B. depth=1, not zero
1095 break;
1096
1097 default:
1098 // bytecode produces a typed result
1099 inputs = rsize - depth;
1100 assert(inputs >= 0, "");
1101 break;
1102 }
1103
1104 #ifdef ASSERT
1105 // spot check
1106 int outputs = depth + inputs;
1107 assert(outputs >= 0, "sanity");
1108 switch (code) {
1109 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1110 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1111 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1112 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1113 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1114 default: break;
1115 }
1116 #endif //ASSERT
1117
1118 return true;
1119 }
1120
1121
1122
1123 //------------------------------basic_plus_adr---------------------------------
1124 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1125 // short-circuit a common case
1126 if (offset == intcon(0)) return ptr;
1127 return _gvn.transform( new AddPNode(base, ptr, offset) );
1128 }
1129
1130 Node* GraphKit::ConvI2L(Node* offset) {
1131 // short-circuit a common case
1132 jint offset_con = find_int_con(offset, Type::OffsetBot);
1133 if (offset_con != Type::OffsetBot) {
1134 return longcon((jlong) offset_con);
1135 }
1136 return _gvn.transform( new ConvI2LNode(offset));
1137 }
1138
1139 Node* GraphKit::ConvI2UL(Node* offset) {
1140 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1141 if (offset_con != (juint) Type::OffsetBot) {
1142 return longcon((julong) offset_con);
1143 }
1144 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1145 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1146 return _gvn.transform( new AndLNode(conv, mask) );
1147 }
1148
1149 Node* GraphKit::ConvL2I(Node* offset) {
1150 // short-circuit a common case
1151 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1152 if (offset_con != (jlong)Type::OffsetBot) {
1153 return intcon((int) offset_con);
1154 }
1155 return _gvn.transform( new ConvL2INode(offset));
1156 }
1157
1158 //-------------------------load_object_klass-----------------------------------
1159 Node* GraphKit::load_object_klass(Node* obj) {
1160 // Special-case a fresh allocation to avoid building nodes:
1161 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1162 if (akls != NULL) return akls;
1163 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1164 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1165 }
1166
1167 //-------------------------load_array_length-----------------------------------
1168 Node* GraphKit::load_array_length(Node* array) {
1169 // Special-case a fresh allocation to avoid building nodes:
1170 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1171 Node *alen;
1172 if (alloc == NULL) {
1173 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1174 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1175 } else {
1176 alen = alloc->Ideal_length();
1177 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1178 if (ccast != alen) {
1179 alen = _gvn.transform(ccast);
1180 }
1181 }
1182 return alen;
1183 }
1184
1185 //------------------------------do_null_check----------------------------------
1186 // Helper function to do a NULL pointer check. Returned value is
1187 // the incoming address with NULL casted away. You are allowed to use the
1188 // not-null value only if you are control dependent on the test.
1189 #ifndef PRODUCT
1190 extern int explicit_null_checks_inserted,
1191 explicit_null_checks_elided;
1192 #endif
1193 Node* GraphKit::null_check_common(Node* value, BasicType type,
1194 // optional arguments for variations:
1195 bool assert_null,
1196 Node* *null_control,
1197 bool speculative) {
1198 assert(!assert_null || null_control == NULL, "not both at once");
1199 if (stopped()) return top();
1200 NOT_PRODUCT(explicit_null_checks_inserted++);
1201
1202 // Construct NULL check
1203 Node *chk = NULL;
1204 switch(type) {
1205 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1206 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1207 case T_ARRAY : // fall through
1208 type = T_OBJECT; // simplify further tests
1209 case T_OBJECT : {
1210 const Type *t = _gvn.type( value );
1211
1212 const TypeOopPtr* tp = t->isa_oopptr();
1213 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1214 // Only for do_null_check, not any of its siblings:
1215 && !assert_null && null_control == NULL) {
1216 // Usually, any field access or invocation on an unloaded oop type
1217 // will simply fail to link, since the statically linked class is
1218 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1219 // the static class is loaded but the sharper oop type is not.
1220 // Rather than checking for this obscure case in lots of places,
1221 // we simply observe that a null check on an unloaded class
1222 // will always be followed by a nonsense operation, so we
1223 // can just issue the uncommon trap here.
1224 // Our access to the unloaded class will only be correct
1225 // after it has been loaded and initialized, which requires
1226 // a trip through the interpreter.
1227 #ifndef PRODUCT
1228 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1229 #endif
1230 uncommon_trap(Deoptimization::Reason_unloaded,
1231 Deoptimization::Action_reinterpret,
1232 tp->klass(), "!loaded");
1233 return top();
1234 }
1235
1236 if (assert_null) {
1237 // See if the type is contained in NULL_PTR.
1238 // If so, then the value is already null.
1239 if (t->higher_equal(TypePtr::NULL_PTR)) {
1240 NOT_PRODUCT(explicit_null_checks_elided++);
1241 return value; // Elided null assert quickly!
1242 }
1243 } else {
1244 // See if mixing in the NULL pointer changes type.
1245 // If so, then the NULL pointer was not allowed in the original
1246 // type. In other words, "value" was not-null.
1247 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1248 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1249 NOT_PRODUCT(explicit_null_checks_elided++);
1250 return value; // Elided null check quickly!
1251 }
1252 }
1253 chk = new CmpPNode( value, null() );
1254 break;
1255 }
1256
1257 default:
1258 fatal("unexpected type: %s", type2name(type));
1259 }
1260 assert(chk != NULL, "sanity check");
1261 chk = _gvn.transform(chk);
1262
1263 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1264 BoolNode *btst = new BoolNode( chk, btest);
1265 Node *tst = _gvn.transform( btst );
1266
1267 //-----------
1268 // if peephole optimizations occurred, a prior test existed.
1269 // If a prior test existed, maybe it dominates as we can avoid this test.
1270 if (tst != btst && type == T_OBJECT) {
1271 // At this point we want to scan up the CFG to see if we can
1272 // find an identical test (and so avoid this test altogether).
1273 Node *cfg = control();
1274 int depth = 0;
1275 while( depth < 16 ) { // Limit search depth for speed
1276 if( cfg->Opcode() == Op_IfTrue &&
1277 cfg->in(0)->in(1) == tst ) {
1278 // Found prior test. Use "cast_not_null" to construct an identical
1279 // CastPP (and hence hash to) as already exists for the prior test.
1280 // Return that casted value.
1281 if (assert_null) {
1282 replace_in_map(value, null());
1283 return null(); // do not issue the redundant test
1284 }
1285 Node *oldcontrol = control();
1286 set_control(cfg);
1287 Node *res = cast_not_null(value);
1288 set_control(oldcontrol);
1289 NOT_PRODUCT(explicit_null_checks_elided++);
1290 return res;
1291 }
1292 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1293 if (cfg == NULL) break; // Quit at region nodes
1294 depth++;
1295 }
1296 }
1297
1298 //-----------
1299 // Branch to failure if null
1300 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1301 Deoptimization::DeoptReason reason;
1302 if (assert_null) {
1303 reason = Deoptimization::reason_null_assert(speculative);
1304 } else if (type == T_OBJECT) {
1305 reason = Deoptimization::reason_null_check(speculative);
1306 } else {
1307 reason = Deoptimization::Reason_div0_check;
1308 }
1309 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1310 // ciMethodData::has_trap_at will return a conservative -1 if any
1311 // must-be-null assertion has failed. This could cause performance
1312 // problems for a method after its first do_null_assert failure.
1313 // Consider using 'Reason_class_check' instead?
1314
1315 // To cause an implicit null check, we set the not-null probability
1316 // to the maximum (PROB_MAX). For an explicit check the probability
1317 // is set to a smaller value.
1318 if (null_control != NULL || too_many_traps(reason)) {
1319 // probability is less likely
1320 ok_prob = PROB_LIKELY_MAG(3);
1321 } else if (!assert_null &&
1322 (ImplicitNullCheckThreshold > 0) &&
1323 method() != NULL &&
1324 (method()->method_data()->trap_count(reason)
1325 >= (uint)ImplicitNullCheckThreshold)) {
1326 ok_prob = PROB_LIKELY_MAG(3);
1327 }
1328
1329 if (null_control != NULL) {
1330 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1331 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1332 set_control( _gvn.transform( new IfTrueNode(iff)));
1333 #ifndef PRODUCT
1334 if (null_true == top()) {
1335 explicit_null_checks_elided++;
1336 }
1337 #endif
1338 (*null_control) = null_true;
1339 } else {
1340 BuildCutout unless(this, tst, ok_prob);
1341 // Check for optimizer eliding test at parse time
1342 if (stopped()) {
1343 // Failure not possible; do not bother making uncommon trap.
1344 NOT_PRODUCT(explicit_null_checks_elided++);
1345 } else if (assert_null) {
1346 uncommon_trap(reason,
1347 Deoptimization::Action_make_not_entrant,
1348 NULL, "assert_null");
1349 } else {
1350 replace_in_map(value, zerocon(type));
1351 builtin_throw(reason);
1352 }
1353 }
1354
1355 // Must throw exception, fall-thru not possible?
1356 if (stopped()) {
1357 return top(); // No result
1358 }
1359
1360 if (assert_null) {
1361 // Cast obj to null on this path.
1362 replace_in_map(value, zerocon(type));
1363 return zerocon(type);
1364 }
1365
1366 // Cast obj to not-null on this path, if there is no null_control.
1367 // (If there is a null_control, a non-null value may come back to haunt us.)
1368 if (type == T_OBJECT) {
1369 Node* cast = cast_not_null(value, false);
1370 if (null_control == NULL || (*null_control) == top())
1371 replace_in_map(value, cast);
1372 value = cast;
1373 }
1374
1375 return value;
1376 }
1377
1378
1379 //------------------------------cast_not_null----------------------------------
1380 // Cast obj to not-null on this path
1381 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1382 const Type *t = _gvn.type(obj);
1383 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1384 // Object is already not-null?
1385 if( t == t_not_null ) return obj;
1386
1387 Node *cast = new CastPPNode(obj,t_not_null);
1388 cast->init_req(0, control());
1389 cast = _gvn.transform( cast );
1390
1391 // Scan for instances of 'obj' in the current JVM mapping.
1392 // These instances are known to be not-null after the test.
1393 if (do_replace_in_map)
1394 replace_in_map(obj, cast);
1395
1396 return cast; // Return casted value
1397 }
1398
1399 // Sometimes in intrinsics, we implicitly know an object is not null
1400 // (there's no actual null check) so we can cast it to not null. In
1401 // the course of optimizations, the input to the cast can become null.
1402 // In that case that data path will die and we need the control path
1403 // to become dead as well to keep the graph consistent. So we have to
1404 // add a check for null for which one branch can't be taken. It uses
1405 // an Opaque4 node that will cause the check to be removed after loop
1406 // opts so the test goes away and the compiled code doesn't execute a
1407 // useless check.
1408 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1409 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1410 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1411 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1412 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1413 _gvn.set_type(iff, iff->Value(&_gvn));
1414 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1415 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1416 Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1417 C->root()->add_req(halt);
1418 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1419 set_control(if_t);
1420 return cast_not_null(value, do_replace_in_map);
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 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1510 // Improve graph before escape analysis and boxing elimination.
1511 record_for_igvn(ld);
1512 }
1513 return ld;
1514 }
1515
1516 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1517 int adr_idx,
1518 MemNode::MemOrd mo,
1519 bool require_atomic_access,
1520 bool unaligned,
1521 bool mismatched) {
1522 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1523 const TypePtr* adr_type = NULL;
1524 debug_only(adr_type = C->get_adr_type(adr_idx));
1525 Node *mem = memory(adr_idx);
1526 Node* st;
1527 if (require_atomic_access && bt == T_LONG) {
1528 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1529 } else if (require_atomic_access && bt == T_DOUBLE) {
1530 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1531 } else {
1532 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1533 }
1534 if (unaligned) {
1535 st->as_Store()->set_unaligned_access();
1536 }
1537 if (mismatched) {
1538 st->as_Store()->set_mismatched_access();
1539 }
1540 st = _gvn.transform(st);
1541 set_memory(st, adr_idx);
1542 // Back-to-back stores can only remove intermediate store with DU info
1543 // so push on worklist for optimizer.
1544 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1545 record_for_igvn(st);
1546
1547 return st;
1548 }
1549
1550
1551 void GraphKit::pre_barrier(bool do_load,
1552 Node* ctl,
1553 Node* obj,
1554 Node* adr,
1555 uint adr_idx,
1556 Node* val,
1557 const TypeOopPtr* val_type,
1558 Node* pre_val,
1559 BasicType bt) {
1560
1561 BarrierSet* bs = Universe::heap()->barrier_set();
1562 set_control(ctl);
1563 switch (bs->kind()) {
1564 case BarrierSet::G1BarrierSet:
1565 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1566 break;
1567
1568 case BarrierSet::CardTableModRef:
1569 case BarrierSet::Z:
1570 break;
1571
1572 default :
1573 ShouldNotReachHere();
1574
1575 }
1576 }
1577
1578 bool GraphKit::can_move_pre_barrier() const {
1579 BarrierSet* bs = Universe::heap()->barrier_set();
1580 switch (bs->kind()) {
1581 case BarrierSet::G1BarrierSet:
1582 return true; // Can move it if no safepoint
1583
1584 case BarrierSet::CardTableModRef:
1585 case BarrierSet::Z:
1586 return true; // There is no pre-barrier
1587
1588 default :
1589 ShouldNotReachHere();
1590 }
1591 return false;
1592 }
1593
1594 bool GraphKit::has_post_barrier() {
1595 BarrierSet* bs = Universe::heap()->barrier_set();
1596 switch (bs->kind()) {
1597 case BarrierSet::G1BarrierSet:
1598 case BarrierSet::CardTableModRef:
1599 return true;
1600 case BarrierSet::Z:
1601 return false;
1602 default:
1603 ShouldNotReachHere();
1604 return false;
1605 }
1606 }
1607
1608 void GraphKit::post_barrier(Node* ctl,
1609 Node* store,
1610 Node* obj,
1611 Node* adr,
1612 uint adr_idx,
1613 Node* val,
1614 BasicType bt,
1615 bool use_precise) {
1616 BarrierSet* bs = Universe::heap()->barrier_set();
1617 set_control(ctl);
1618 switch (bs->kind()) {
1619 case BarrierSet::G1BarrierSet:
1620 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1621 break;
1622
1623 case BarrierSet::CardTableModRef:
1624 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1625 break;
1626
1627 case BarrierSet::Z:
1628 break;
1629
1630 default :
1631 ShouldNotReachHere();
1632
1633 }
1634 }
1635
1636 Node* GraphKit::store_oop(Node* ctl,
1637 Node* obj,
1638 Node* adr,
1639 const TypePtr* adr_type,
1640 Node* val,
1641 const TypeOopPtr* val_type,
1642 BasicType bt,
1643 bool use_precise,
1644 MemNode::MemOrd mo,
1645 bool mismatched) {
1646 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1647 // could be delayed during Parse (for example, in adjust_map_after_if()).
1648 // Execute transformation here to avoid barrier generation in such case.
1649 if (_gvn.type(val) == TypePtr::NULL_PTR)
1650 val = _gvn.makecon(TypePtr::NULL_PTR);
1651
1652 set_control(ctl);
1653 if (stopped()) return top(); // Dead path ?
1654
1655 assert(bt == T_OBJECT, "sanity");
1656 assert(val != NULL, "not dead path");
1657 uint adr_idx = C->get_alias_index(adr_type);
1658 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1659
1660 pre_barrier(true /* do_load */,
1661 control(), obj, adr, adr_idx, val, val_type,
1662 NULL /* pre_val */,
1663 bt);
1664
1665 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1666 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1667 return store;
1668 }
1669
1670 // Could be an array or object we don't know at compile time (unsafe ref.)
1671 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1672 Node* obj, // containing obj
1673 Node* adr, // actual adress to store val at
1674 const TypePtr* adr_type,
1675 Node* val,
1676 BasicType bt,
1677 MemNode::MemOrd mo,
1678 bool mismatched) {
1679 Compile::AliasType* at = C->alias_type(adr_type);
1680 const TypeOopPtr* val_type = NULL;
1681 if (adr_type->isa_instptr()) {
1682 if (at->field() != NULL) {
1683 // known field. This code is a copy of the do_put_xxx logic.
1684 ciField* field = at->field();
1685 if (!field->type()->is_loaded()) {
1686 val_type = TypeInstPtr::BOTTOM;
1687 } else {
1688 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1689 }
1690 }
1691 } else if (adr_type->isa_aryptr()) {
1692 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1693 }
1694 if (val_type == NULL) {
1695 val_type = TypeInstPtr::BOTTOM;
1696 }
1697 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1698 }
1699
1700 Node* GraphKit::load_barrier(Node* val, Node* adr, bool weak, bool writeback, bool oop_reload_allowed) {
1701 assert(UseZGC, "invariant");
1702 Node* barrier = new LoadBarrierNode(C, control(), memory(TypeRawPtr::BOTTOM), val, adr, weak, writeback, oop_reload_allowed);
1703 Node* transformed_barrier = _gvn.transform(barrier);
1704
1705 if (transformed_barrier->is_LoadBarrier()) {
1706 if (barrier == transformed_barrier) {
1707 set_control(_gvn.transform(new ProjNode(barrier, LoadBarrierNode::Control)));
1708 }
1709 return _gvn.transform(new ProjNode(transformed_barrier, LoadBarrierNode::Oop));
1710 } else {
1711 return val;
1712 }
1713 }
1714
1715 //-------------------------array_element_address-------------------------
1716 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1717 const TypeInt* sizetype, Node* ctrl) {
1718 uint shift = exact_log2(type2aelembytes(elembt));
1719 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1720
1721 // short-circuit a common case (saves lots of confusing waste motion)
1722 jint idx_con = find_int_con(idx, -1);
1723 if (idx_con >= 0) {
1724 intptr_t offset = header + ((intptr_t)idx_con << shift);
1725 return basic_plus_adr(ary, offset);
1726 }
1727
1728 // must be correct type for alignment purposes
1729 Node* base = basic_plus_adr(ary, header);
1730 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1731 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1732 return basic_plus_adr(ary, base, scale);
1733 }
1734
1735 //-------------------------load_array_element-------------------------
1736 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1737 const Type* elemtype = arytype->elem();
1738 BasicType elembt = elemtype->array_element_basic_type();
1739 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1740 if (elembt == T_NARROWOOP) {
1741 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1742 }
1743 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1744 return ld;
1745 }
1746
1747 //-------------------------set_arguments_for_java_call-------------------------
1748 // Arguments (pre-popped from the stack) are taken from the JVMS.
1749 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1750 // Add the call arguments:
1751 uint nargs = call->method()->arg_size();
1752 for (uint i = 0; i < nargs; i++) {
1753 Node* arg = argument(i);
1754 call->init_req(i + TypeFunc::Parms, arg);
1755 }
1756 }
1757
1758 //---------------------------set_edges_for_java_call---------------------------
1759 // Connect a newly created call into the current JVMS.
1760 // A return value node (if any) is returned from set_edges_for_java_call.
1761 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1762
1763 // Add the predefined inputs:
1764 call->init_req( TypeFunc::Control, control() );
1765 call->init_req( TypeFunc::I_O , i_o() );
1766 call->init_req( TypeFunc::Memory , reset_memory() );
1767 call->init_req( TypeFunc::FramePtr, frameptr() );
1768 call->init_req( TypeFunc::ReturnAdr, top() );
1769
1770 add_safepoint_edges(call, must_throw);
1771
1772 Node* xcall = _gvn.transform(call);
1773
1774 if (xcall == top()) {
1775 set_control(top());
1776 return;
1777 }
1778 assert(xcall == call, "call identity is stable");
1779
1780 // Re-use the current map to produce the result.
1781
1782 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1783 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1784 set_all_memory_call(xcall, separate_io_proj);
1785
1786 //return xcall; // no need, caller already has it
1787 }
1788
1789 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1790 if (stopped()) return top(); // maybe the call folded up?
1791
1792 // Capture the return value, if any.
1793 Node* ret;
1794 if (call->method() == NULL ||
1795 call->method()->return_type()->basic_type() == T_VOID)
1796 ret = top();
1797 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1798
1799 // Note: Since any out-of-line call can produce an exception,
1800 // we always insert an I_O projection from the call into the result.
1801
1802 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1803
1804 if (separate_io_proj) {
1805 // The caller requested separate projections be used by the fall
1806 // through and exceptional paths, so replace the projections for
1807 // the fall through path.
1808 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1809 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1810 }
1811 return ret;
1812 }
1813
1814 //--------------------set_predefined_input_for_runtime_call--------------------
1815 // Reading and setting the memory state is way conservative here.
1816 // The real problem is that I am not doing real Type analysis on memory,
1817 // so I cannot distinguish card mark stores from other stores. Across a GC
1818 // point the Store Barrier and the card mark memory has to agree. I cannot
1819 // have a card mark store and its barrier split across the GC point from
1820 // either above or below. Here I get that to happen by reading ALL of memory.
1821 // A better answer would be to separate out card marks from other memory.
1822 // For now, return the input memory state, so that it can be reused
1823 // after the call, if this call has restricted memory effects.
1824 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1825 // Set fixed predefined input arguments
1826 Node* memory = reset_memory();
1827 call->init_req( TypeFunc::Control, control() );
1828 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1829 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1830 call->init_req( TypeFunc::FramePtr, frameptr() );
1831 call->init_req( TypeFunc::ReturnAdr, top() );
1832 return memory;
1833 }
1834
1835 //-------------------set_predefined_output_for_runtime_call--------------------
1836 // Set control and memory (not i_o) from the call.
1837 // If keep_mem is not NULL, use it for the output state,
1838 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1839 // If hook_mem is NULL, this call produces no memory effects at all.
1840 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1841 // then only that memory slice is taken from the call.
1842 // In the last case, we must put an appropriate memory barrier before
1843 // the call, so as to create the correct anti-dependencies on loads
1844 // preceding the call.
1845 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1846 Node* keep_mem,
1847 const TypePtr* hook_mem) {
1848 // no i/o
1849 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1850 if (keep_mem) {
1851 // First clone the existing memory state
1852 set_all_memory(keep_mem);
1853 if (hook_mem != NULL) {
1854 // Make memory for the call
1855 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1856 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1857 // We also use hook_mem to extract specific effects from arraycopy stubs.
1858 set_memory(mem, hook_mem);
1859 }
1860 // ...else the call has NO memory effects.
1861
1862 // Make sure the call advertises its memory effects precisely.
1863 // This lets us build accurate anti-dependences in gcm.cpp.
1864 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1865 "call node must be constructed correctly");
1866 } else {
1867 assert(hook_mem == NULL, "");
1868 // This is not a "slow path" call; all memory comes from the call.
1869 set_all_memory_call(call);
1870 }
1871 }
1872
1873
1874 // Replace the call with the current state of the kit.
1875 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1876 JVMState* ejvms = NULL;
1877 if (has_exceptions()) {
1878 ejvms = transfer_exceptions_into_jvms();
1879 }
1880
1881 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1882 ReplacedNodes replaced_nodes_exception;
1883 Node* ex_ctl = top();
1884
1885 SafePointNode* final_state = stop();
1886
1887 // Find all the needed outputs of this call
1888 CallProjections callprojs;
1889 call->extract_projections(&callprojs, true);
1890
1891 Node* init_mem = call->in(TypeFunc::Memory);
1892 Node* final_mem = final_state->in(TypeFunc::Memory);
1893 Node* final_ctl = final_state->in(TypeFunc::Control);
1894 Node* final_io = final_state->in(TypeFunc::I_O);
1895
1896 // Replace all the old call edges with the edges from the inlining result
1897 if (callprojs.fallthrough_catchproj != NULL) {
1898 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1899 }
1900 if (callprojs.fallthrough_memproj != NULL) {
1901 if (final_mem->is_MergeMem()) {
1902 // Parser's exits MergeMem was not transformed but may be optimized
1903 final_mem = _gvn.transform(final_mem);
1904 }
1905 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1906 }
1907 if (callprojs.fallthrough_ioproj != NULL) {
1908 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1909 }
1910
1911 // Replace the result with the new result if it exists and is used
1912 if (callprojs.resproj != NULL && result != NULL) {
1913 C->gvn_replace_by(callprojs.resproj, result);
1914 }
1915
1916 if (ejvms == NULL) {
1917 // No exception edges to simply kill off those paths
1918 if (callprojs.catchall_catchproj != NULL) {
1919 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1920 }
1921 if (callprojs.catchall_memproj != NULL) {
1922 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1923 }
1924 if (callprojs.catchall_ioproj != NULL) {
1925 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1926 }
1927 // Replace the old exception object with top
1928 if (callprojs.exobj != NULL) {
1929 C->gvn_replace_by(callprojs.exobj, C->top());
1930 }
1931 } else {
1932 GraphKit ekit(ejvms);
1933
1934 // Load my combined exception state into the kit, with all phis transformed:
1935 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1936 replaced_nodes_exception = ex_map->replaced_nodes();
1937
1938 Node* ex_oop = ekit.use_exception_state(ex_map);
1939
1940 if (callprojs.catchall_catchproj != NULL) {
1941 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1942 ex_ctl = ekit.control();
1943 }
1944 if (callprojs.catchall_memproj != NULL) {
1945 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1946 }
1947 if (callprojs.catchall_ioproj != NULL) {
1948 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1949 }
1950
1951 // Replace the old exception object with the newly created one
1952 if (callprojs.exobj != NULL) {
1953 C->gvn_replace_by(callprojs.exobj, ex_oop);
1954 }
1955 }
1956
1957 // Disconnect the call from the graph
1958 call->disconnect_inputs(NULL, C);
1959 C->gvn_replace_by(call, C->top());
1960
1961 // Clean up any MergeMems that feed other MergeMems since the
1962 // optimizer doesn't like that.
1963 if (final_mem->is_MergeMem()) {
1964 Node_List wl;
1965 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1966 Node* m = i.get();
1967 if (m->is_MergeMem() && !wl.contains(m)) {
1968 wl.push(m);
1969 }
1970 }
1971 while (wl.size() > 0) {
1972 _gvn.transform(wl.pop());
1973 }
1974 }
1975
1976 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1977 replaced_nodes.apply(C, final_ctl);
1978 }
1979 if (!ex_ctl->is_top() && do_replaced_nodes) {
1980 replaced_nodes_exception.apply(C, ex_ctl);
1981 }
1982 }
1983
1984
1985 //------------------------------increment_counter------------------------------
1986 // for statistics: increment a VM counter by 1
1987
1988 void GraphKit::increment_counter(address counter_addr) {
1989 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1990 increment_counter(adr1);
1991 }
1992
1993 void GraphKit::increment_counter(Node* counter_addr) {
1994 int adr_type = Compile::AliasIdxRaw;
1995 Node* ctrl = control();
1996 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1997 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1998 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1999 }
2000
2001
2002 //------------------------------uncommon_trap----------------------------------
2003 // Bail out to the interpreter in mid-method. Implemented by calling the
2004 // uncommon_trap blob. This helper function inserts a runtime call with the
2005 // right debug info.
2006 void GraphKit::uncommon_trap(int trap_request,
2007 ciKlass* klass, const char* comment,
2008 bool must_throw,
2009 bool keep_exact_action) {
2010 if (failing()) stop();
2011 if (stopped()) return; // trap reachable?
2012
2013 // Note: If ProfileTraps is true, and if a deopt. actually
2014 // occurs here, the runtime will make sure an MDO exists. There is
2015 // no need to call method()->ensure_method_data() at this point.
2016
2017 // Set the stack pointer to the right value for reexecution:
2018 set_sp(reexecute_sp());
2019
2020 #ifdef ASSERT
2021 if (!must_throw) {
2022 // Make sure the stack has at least enough depth to execute
2023 // the current bytecode.
2024 int inputs, ignored_depth;
2025 if (compute_stack_effects(inputs, ignored_depth)) {
2026 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2027 Bytecodes::name(java_bc()), sp(), inputs);
2028 }
2029 }
2030 #endif
2031
2032 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2033 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2034
2035 switch (action) {
2036 case Deoptimization::Action_maybe_recompile:
2037 case Deoptimization::Action_reinterpret:
2038 // Temporary fix for 6529811 to allow virtual calls to be sure they
2039 // get the chance to go from mono->bi->mega
2040 if (!keep_exact_action &&
2041 Deoptimization::trap_request_index(trap_request) < 0 &&
2042 too_many_recompiles(reason)) {
2043 // This BCI is causing too many recompilations.
2044 if (C->log() != NULL) {
2045 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2046 Deoptimization::trap_reason_name(reason),
2047 Deoptimization::trap_action_name(action));
2048 }
2049 action = Deoptimization::Action_none;
2050 trap_request = Deoptimization::make_trap_request(reason, action);
2051 } else {
2052 C->set_trap_can_recompile(true);
2053 }
2054 break;
2055 case Deoptimization::Action_make_not_entrant:
2056 C->set_trap_can_recompile(true);
2057 break;
2058 case Deoptimization::Action_none:
2059 case Deoptimization::Action_make_not_compilable:
2060 break;
2061 default:
2062 #ifdef ASSERT
2063 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2064 #endif
2065 break;
2066 }
2067
2068 if (TraceOptoParse) {
2069 char buf[100];
2070 tty->print_cr("Uncommon trap %s at bci:%d",
2071 Deoptimization::format_trap_request(buf, sizeof(buf),
2072 trap_request), bci());
2073 }
2074
2075 CompileLog* log = C->log();
2076 if (log != NULL) {
2077 int kid = (klass == NULL)? -1: log->identify(klass);
2078 log->begin_elem("uncommon_trap bci='%d'", bci());
2079 char buf[100];
2080 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2081 trap_request));
2082 if (kid >= 0) log->print(" klass='%d'", kid);
2083 if (comment != NULL) log->print(" comment='%s'", comment);
2084 log->end_elem();
2085 }
2086
2087 // Make sure any guarding test views this path as very unlikely
2088 Node *i0 = control()->in(0);
2089 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2090 IfNode *iff = i0->as_If();
2091 float f = iff->_prob; // Get prob
2092 if (control()->Opcode() == Op_IfTrue) {
2093 if (f > PROB_UNLIKELY_MAG(4))
2094 iff->_prob = PROB_MIN;
2095 } else {
2096 if (f < PROB_LIKELY_MAG(4))
2097 iff->_prob = PROB_MAX;
2098 }
2099 }
2100
2101 // Clear out dead values from the debug info.
2102 kill_dead_locals();
2103
2104 // Now insert the uncommon trap subroutine call
2105 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2106 const TypePtr* no_memory_effects = NULL;
2107 // Pass the index of the class to be loaded
2108 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2109 (must_throw ? RC_MUST_THROW : 0),
2110 OptoRuntime::uncommon_trap_Type(),
2111 call_addr, "uncommon_trap", no_memory_effects,
2112 intcon(trap_request));
2113 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2114 "must extract request correctly from the graph");
2115 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2116
2117 call->set_req(TypeFunc::ReturnAdr, returnadr());
2118 // The debug info is the only real input to this call.
2119
2120 // Halt-and-catch fire here. The above call should never return!
2121 HaltNode* halt = new HaltNode(control(), frameptr());
2122 _gvn.set_type_bottom(halt);
2123 root()->add_req(halt);
2124
2125 stop_and_kill_map();
2126 }
2127
2128
2129 //--------------------------just_allocated_object------------------------------
2130 // Report the object that was just allocated.
2131 // It must be the case that there are no intervening safepoints.
2132 // We use this to determine if an object is so "fresh" that
2133 // it does not require card marks.
2134 Node* GraphKit::just_allocated_object(Node* current_control) {
2135 if (C->recent_alloc_ctl() == current_control)
2136 return C->recent_alloc_obj();
2137 return NULL;
2138 }
2139
2140
2141 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2142 // (Note: TypeFunc::make has a cache that makes this fast.)
2143 const TypeFunc* tf = TypeFunc::make(dest_method);
2144 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2145 for (int j = 0; j < nargs; j++) {
2146 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2147 if( targ->basic_type() == T_DOUBLE ) {
2148 // If any parameters are doubles, they must be rounded before
2149 // the call, dstore_rounding does gvn.transform
2150 Node *arg = argument(j);
2151 arg = dstore_rounding(arg);
2152 set_argument(j, arg);
2153 }
2154 }
2155 }
2156
2157 /**
2158 * Record profiling data exact_kls for Node n with the type system so
2159 * that it can propagate it (speculation)
2160 *
2161 * @param n node that the type applies to
2162 * @param exact_kls type from profiling
2163 * @param maybe_null did profiling see null?
2164 *
2165 * @return node with improved type
2166 */
2167 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2168 const Type* current_type = _gvn.type(n);
2169 assert(UseTypeSpeculation, "type speculation must be on");
2170
2171 const TypePtr* speculative = current_type->speculative();
2172
2173 // Should the klass from the profile be recorded in the speculative type?
2174 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2175 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2176 const TypeOopPtr* xtype = tklass->as_instance_type();
2177 assert(xtype->klass_is_exact(), "Should be exact");
2178 // Any reason to believe n is not null (from this profiling or a previous one)?
2179 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2180 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2181 // record the new speculative type's depth
2182 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2183 speculative = speculative->with_inline_depth(jvms()->depth());
2184 } else if (current_type->would_improve_ptr(ptr_kind)) {
2185 // Profiling report that null was never seen so we can change the
2186 // speculative type to non null ptr.
2187 if (ptr_kind == ProfileAlwaysNull) {
2188 speculative = TypePtr::NULL_PTR;
2189 } else {
2190 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2191 const TypePtr* ptr = TypePtr::NOTNULL;
2192 if (speculative != NULL) {
2193 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2194 } else {
2195 speculative = ptr;
2196 }
2197 }
2198 }
2199
2200 if (speculative != current_type->speculative()) {
2201 // Build a type with a speculative type (what we think we know
2202 // about the type but will need a guard when we use it)
2203 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2204 // We're changing the type, we need a new CheckCast node to carry
2205 // the new type. The new type depends on the control: what
2206 // profiling tells us is only valid from here as far as we can
2207 // tell.
2208 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2209 cast = _gvn.transform(cast);
2210 replace_in_map(n, cast);
2211 n = cast;
2212 }
2213
2214 return n;
2215 }
2216
2217 /**
2218 * Record profiling data from receiver profiling at an invoke with the
2219 * type system so that it can propagate it (speculation)
2220 *
2221 * @param n receiver node
2222 *
2223 * @return node with improved type
2224 */
2225 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2226 if (!UseTypeSpeculation) {
2227 return n;
2228 }
2229 ciKlass* exact_kls = profile_has_unique_klass();
2230 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2231 if ((java_bc() == Bytecodes::_checkcast ||
2232 java_bc() == Bytecodes::_instanceof ||
2233 java_bc() == Bytecodes::_aastore) &&
2234 method()->method_data()->is_mature()) {
2235 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2236 if (data != NULL) {
2237 if (!data->as_BitData()->null_seen()) {
2238 ptr_kind = ProfileNeverNull;
2239 } else {
2240 assert(data->is_ReceiverTypeData(), "bad profile data type");
2241 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2242 uint i = 0;
2243 for (; i < call->row_limit(); i++) {
2244 ciKlass* receiver = call->receiver(i);
2245 if (receiver != NULL) {
2246 break;
2247 }
2248 }
2249 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2250 }
2251 }
2252 }
2253 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2254 }
2255
2256 /**
2257 * Record profiling data from argument profiling at an invoke with the
2258 * type system so that it can propagate it (speculation)
2259 *
2260 * @param dest_method target method for the call
2261 * @param bc what invoke bytecode is this?
2262 */
2263 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2264 if (!UseTypeSpeculation) {
2265 return;
2266 }
2267 const TypeFunc* tf = TypeFunc::make(dest_method);
2268 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2269 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2270 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2271 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2272 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2273 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2274 ciKlass* better_type = NULL;
2275 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2276 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2277 }
2278 i++;
2279 }
2280 }
2281 }
2282
2283 /**
2284 * Record profiling data from parameter profiling at an invoke with
2285 * the type system so that it can propagate it (speculation)
2286 */
2287 void GraphKit::record_profiled_parameters_for_speculation() {
2288 if (!UseTypeSpeculation) {
2289 return;
2290 }
2291 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2292 if (_gvn.type(local(i))->isa_oopptr()) {
2293 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2294 ciKlass* better_type = NULL;
2295 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2296 record_profile_for_speculation(local(i), better_type, ptr_kind);
2297 }
2298 j++;
2299 }
2300 }
2301 }
2302
2303 /**
2304 * Record profiling data from return value profiling at an invoke with
2305 * the type system so that it can propagate it (speculation)
2306 */
2307 void GraphKit::record_profiled_return_for_speculation() {
2308 if (!UseTypeSpeculation) {
2309 return;
2310 }
2311 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2312 ciKlass* better_type = NULL;
2313 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2314 // If profiling reports a single type for the return value,
2315 // feed it to the type system so it can propagate it as a
2316 // speculative type
2317 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2318 }
2319 }
2320
2321 void GraphKit::round_double_result(ciMethod* dest_method) {
2322 // A non-strict method may return a double value which has an extended
2323 // exponent, but this must not be visible in a caller which is 'strict'
2324 // If a strict caller invokes a non-strict callee, round a double result
2325
2326 BasicType result_type = dest_method->return_type()->basic_type();
2327 assert( method() != NULL, "must have caller context");
2328 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2329 // Destination method's return value is on top of stack
2330 // dstore_rounding() does gvn.transform
2331 Node *result = pop_pair();
2332 result = dstore_rounding(result);
2333 push_pair(result);
2334 }
2335 }
2336
2337 // rounding for strict float precision conformance
2338 Node* GraphKit::precision_rounding(Node* n) {
2339 return UseStrictFP && _method->flags().is_strict()
2340 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2341 ? _gvn.transform( new RoundFloatNode(0, n) )
2342 : n;
2343 }
2344
2345 // rounding for strict double precision conformance
2346 Node* GraphKit::dprecision_rounding(Node *n) {
2347 return UseStrictFP && _method->flags().is_strict()
2348 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2349 ? _gvn.transform( new RoundDoubleNode(0, n) )
2350 : n;
2351 }
2352
2353 // rounding for non-strict double stores
2354 Node* GraphKit::dstore_rounding(Node* n) {
2355 return Matcher::strict_fp_requires_explicit_rounding
2356 && UseSSE <= 1
2357 ? _gvn.transform( new RoundDoubleNode(0, n) )
2358 : n;
2359 }
2360
2361 //=============================================================================
2362 // Generate a fast path/slow path idiom. Graph looks like:
2363 // [foo] indicates that 'foo' is a parameter
2364 //
2365 // [in] NULL
2366 // \ /
2367 // CmpP
2368 // Bool ne
2369 // If
2370 // / \
2371 // True False-<2>
2372 // / |
2373 // / cast_not_null
2374 // Load | | ^
2375 // [fast_test] | |
2376 // gvn to opt_test | |
2377 // / \ | <1>
2378 // True False |
2379 // | \\ |
2380 // [slow_call] \[fast_result]
2381 // Ctl Val \ \
2382 // | \ \
2383 // Catch <1> \ \
2384 // / \ ^ \ \
2385 // Ex No_Ex | \ \
2386 // | \ \ | \ <2> \
2387 // ... \ [slow_res] | | \ [null_result]
2388 // \ \--+--+--- | |
2389 // \ | / \ | /
2390 // --------Region Phi
2391 //
2392 //=============================================================================
2393 // Code is structured as a series of driver functions all called 'do_XXX' that
2394 // call a set of helper functions. Helper functions first, then drivers.
2395
2396 //------------------------------null_check_oop---------------------------------
2397 // Null check oop. Set null-path control into Region in slot 3.
2398 // Make a cast-not-nullness use the other not-null control. Return cast.
2399 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2400 bool never_see_null,
2401 bool safe_for_replace,
2402 bool speculative) {
2403 // Initial NULL check taken path
2404 (*null_control) = top();
2405 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2406
2407 // Generate uncommon_trap:
2408 if (never_see_null && (*null_control) != top()) {
2409 // If we see an unexpected null at a check-cast we record it and force a
2410 // recompile; the offending check-cast will be compiled to handle NULLs.
2411 // If we see more than one offending BCI, then all checkcasts in the
2412 // method will be compiled to handle NULLs.
2413 PreserveJVMState pjvms(this);
2414 set_control(*null_control);
2415 replace_in_map(value, null());
2416 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2417 uncommon_trap(reason,
2418 Deoptimization::Action_make_not_entrant);
2419 (*null_control) = top(); // NULL path is dead
2420 }
2421 if ((*null_control) == top() && safe_for_replace) {
2422 replace_in_map(value, cast);
2423 }
2424
2425 // Cast away null-ness on the result
2426 return cast;
2427 }
2428
2429 //------------------------------opt_iff----------------------------------------
2430 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2431 // Return slow-path control.
2432 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2433 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2434
2435 // Fast path taken; set region slot 2
2436 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2437 region->init_req(2,fast_taken); // Capture fast-control
2438
2439 // Fast path not-taken, i.e. slow path
2440 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2441 return slow_taken;
2442 }
2443
2444 //-----------------------------make_runtime_call-------------------------------
2445 Node* GraphKit::make_runtime_call(int flags,
2446 const TypeFunc* call_type, address call_addr,
2447 const char* call_name,
2448 const TypePtr* adr_type,
2449 // The following parms are all optional.
2450 // The first NULL ends the list.
2451 Node* parm0, Node* parm1,
2452 Node* parm2, Node* parm3,
2453 Node* parm4, Node* parm5,
2454 Node* parm6, Node* parm7) {
2455 // Slow-path call
2456 bool is_leaf = !(flags & RC_NO_LEAF);
2457 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2458 if (call_name == NULL) {
2459 assert(!is_leaf, "must supply name for leaf");
2460 call_name = OptoRuntime::stub_name(call_addr);
2461 }
2462 CallNode* call;
2463 if (!is_leaf) {
2464 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2465 bci(), adr_type);
2466 } else if (flags & RC_NO_FP) {
2467 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2468 } else {
2469 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2470 }
2471
2472 // The following is similar to set_edges_for_java_call,
2473 // except that the memory effects of the call are restricted to AliasIdxRaw.
2474
2475 // Slow path call has no side-effects, uses few values
2476 bool wide_in = !(flags & RC_NARROW_MEM);
2477 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2478
2479 Node* prev_mem = NULL;
2480 if (wide_in) {
2481 prev_mem = set_predefined_input_for_runtime_call(call);
2482 } else {
2483 assert(!wide_out, "narrow in => narrow out");
2484 Node* narrow_mem = memory(adr_type);
2485 prev_mem = reset_memory();
2486 map()->set_memory(narrow_mem);
2487 set_predefined_input_for_runtime_call(call);
2488 }
2489
2490 // Hook each parm in order. Stop looking at the first NULL.
2491 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2492 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2493 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2494 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2495 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2496 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2497 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2498 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2499 /* close each nested if ===> */ } } } } } } } }
2500 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2501
2502 if (!is_leaf) {
2503 // Non-leaves can block and take safepoints:
2504 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2505 }
2506 // Non-leaves can throw exceptions:
2507 if (has_io) {
2508 call->set_req(TypeFunc::I_O, i_o());
2509 }
2510
2511 if (flags & RC_UNCOMMON) {
2512 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2513 // (An "if" probability corresponds roughly to an unconditional count.
2514 // Sort of.)
2515 call->set_cnt(PROB_UNLIKELY_MAG(4));
2516 }
2517
2518 Node* c = _gvn.transform(call);
2519 assert(c == call, "cannot disappear");
2520
2521 if (wide_out) {
2522 // Slow path call has full side-effects.
2523 set_predefined_output_for_runtime_call(call);
2524 } else {
2525 // Slow path call has few side-effects, and/or sets few values.
2526 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2527 }
2528
2529 if (has_io) {
2530 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2531 }
2532 return call;
2533
2534 }
2535
2536 //------------------------------merge_memory-----------------------------------
2537 // Merge memory from one path into the current memory state.
2538 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2539 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2540 Node* old_slice = mms.force_memory();
2541 Node* new_slice = mms.memory2();
2542 if (old_slice != new_slice) {
2543 PhiNode* phi;
2544 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2545 if (mms.is_empty()) {
2546 // clone base memory Phi's inputs for this memory slice
2547 assert(old_slice == mms.base_memory(), "sanity");
2548 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2549 _gvn.set_type(phi, Type::MEMORY);
2550 for (uint i = 1; i < phi->req(); i++) {
2551 phi->init_req(i, old_slice->in(i));
2552 }
2553 } else {
2554 phi = old_slice->as_Phi(); // Phi was generated already
2555 }
2556 } else {
2557 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2558 _gvn.set_type(phi, Type::MEMORY);
2559 }
2560 phi->set_req(new_path, new_slice);
2561 mms.set_memory(phi);
2562 }
2563 }
2564 }
2565
2566 //------------------------------make_slow_call_ex------------------------------
2567 // Make the exception handler hookups for the slow call
2568 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2569 if (stopped()) return;
2570
2571 // Make a catch node with just two handlers: fall-through and catch-all
2572 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2573 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2574 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2575 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2576
2577 { PreserveJVMState pjvms(this);
2578 set_control(excp);
2579 set_i_o(i_o);
2580
2581 if (excp != top()) {
2582 if (deoptimize) {
2583 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2584 uncommon_trap(Deoptimization::Reason_unhandled,
2585 Deoptimization::Action_none);
2586 } else {
2587 // Create an exception state also.
2588 // Use an exact type if the caller has specified a specific exception.
2589 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2590 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2591 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2592 }
2593 }
2594 }
2595
2596 // Get the no-exception control from the CatchNode.
2597 set_control(norm);
2598 }
2599
2600 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2601 Node* cmp = NULL;
2602 switch(bt) {
2603 case T_INT: cmp = new CmpINode(in1, in2); break;
2604 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2605 default: fatal("unexpected comparison type %s", type2name(bt));
2606 }
2607 gvn->transform(cmp);
2608 Node* bol = gvn->transform(new BoolNode(cmp, test));
2609 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2610 gvn->transform(iff);
2611 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2612 return iff;
2613 }
2614
2615
2616 //-------------------------------gen_subtype_check-----------------------------
2617 // Generate a subtyping check. Takes as input the subtype and supertype.
2618 // Returns 2 values: sets the default control() to the true path and returns
2619 // the false path. Only reads invariant memory; sets no (visible) memory.
2620 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2621 // but that's not exposed to the optimizer. This call also doesn't take in an
2622 // Object; if you wish to check an Object you need to load the Object's class
2623 // prior to coming here.
2624 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2625 Compile* C = gvn->C;
2626
2627 if ((*ctrl)->is_top()) {
2628 return C->top();
2629 }
2630
2631 // Fast check for identical types, perhaps identical constants.
2632 // The types can even be identical non-constants, in cases
2633 // involving Array.newInstance, Object.clone, etc.
2634 if (subklass == superklass)
2635 return C->top(); // false path is dead; no test needed.
2636
2637 if (gvn->type(superklass)->singleton()) {
2638 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2639 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2640
2641 // In the common case of an exact superklass, try to fold up the
2642 // test before generating code. You may ask, why not just generate
2643 // the code and then let it fold up? The answer is that the generated
2644 // code will necessarily include null checks, which do not always
2645 // completely fold away. If they are also needless, then they turn
2646 // into a performance loss. Example:
2647 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2648 // Here, the type of 'fa' is often exact, so the store check
2649 // of fa[1]=x will fold up, without testing the nullness of x.
2650 switch (C->static_subtype_check(superk, subk)) {
2651 case Compile::SSC_always_false:
2652 {
2653 Node* always_fail = *ctrl;
2654 *ctrl = gvn->C->top();
2655 return always_fail;
2656 }
2657 case Compile::SSC_always_true:
2658 return C->top();
2659 case Compile::SSC_easy_test:
2660 {
2661 // Just do a direct pointer compare and be done.
2662 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2663 *ctrl = gvn->transform(new IfTrueNode(iff));
2664 return gvn->transform(new IfFalseNode(iff));
2665 }
2666 case Compile::SSC_full_test:
2667 break;
2668 default:
2669 ShouldNotReachHere();
2670 }
2671 }
2672
2673 // %%% Possible further optimization: Even if the superklass is not exact,
2674 // if the subklass is the unique subtype of the superklass, the check
2675 // will always succeed. We could leave a dependency behind to ensure this.
2676
2677 // First load the super-klass's check-offset
2678 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2679 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2680 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2681 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2682 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2683
2684 // Load from the sub-klass's super-class display list, or a 1-word cache of
2685 // the secondary superclass list, or a failing value with a sentinel offset
2686 // if the super-klass is an interface or exceptionally deep in the Java
2687 // hierarchy and we have to scan the secondary superclass list the hard way.
2688 // Worst-case type is a little odd: NULL is allowed as a result (usually
2689 // klass loads can never produce a NULL).
2690 Node *chk_off_X = chk_off;
2691 #ifdef _LP64
2692 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2693 #endif
2694 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2695 // For some types like interfaces the following loadKlass is from a 1-word
2696 // cache which is mutable so can't use immutable memory. Other
2697 // types load from the super-class display table which is immutable.
2698 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2699 Node *kmem = might_be_cache ? m : C->immutable_memory();
2700 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2701
2702 // Compile speed common case: ARE a subtype and we canNOT fail
2703 if( superklass == nkls )
2704 return C->top(); // false path is dead; no test needed.
2705
2706 // See if we get an immediate positive hit. Happens roughly 83% of the
2707 // time. Test to see if the value loaded just previously from the subklass
2708 // is exactly the superklass.
2709 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2710 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2711 *ctrl = gvn->transform(new IfFalseNode(iff1));
2712
2713 // Compile speed common case: Check for being deterministic right now. If
2714 // chk_off is a constant and not equal to cacheoff then we are NOT a
2715 // subklass. In this case we need exactly the 1 test above and we can
2716 // return those results immediately.
2717 if (!might_be_cache) {
2718 Node* not_subtype_ctrl = *ctrl;
2719 *ctrl = iftrue1; // We need exactly the 1 test above
2720 return not_subtype_ctrl;
2721 }
2722
2723 // Gather the various success & failures here
2724 RegionNode *r_ok_subtype = new RegionNode(4);
2725 gvn->record_for_igvn(r_ok_subtype);
2726 RegionNode *r_not_subtype = new RegionNode(3);
2727 gvn->record_for_igvn(r_not_subtype);
2728
2729 r_ok_subtype->init_req(1, iftrue1);
2730
2731 // Check for immediate negative hit. Happens roughly 11% of the time (which
2732 // is roughly 63% of the remaining cases). Test to see if the loaded
2733 // check-offset points into the subklass display list or the 1-element
2734 // cache. If it points to the display (and NOT the cache) and the display
2735 // missed then it's not a subtype.
2736 Node *cacheoff = gvn->intcon(cacheoff_con);
2737 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2738 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2739 *ctrl = gvn->transform(new IfFalseNode(iff2));
2740
2741 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2742 // No performance impact (too rare) but allows sharing of secondary arrays
2743 // which has some footprint reduction.
2744 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2745 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2746 *ctrl = gvn->transform(new IfFalseNode(iff3));
2747
2748 // -- Roads not taken here: --
2749 // We could also have chosen to perform the self-check at the beginning
2750 // of this code sequence, as the assembler does. This would not pay off
2751 // the same way, since the optimizer, unlike the assembler, can perform
2752 // static type analysis to fold away many successful self-checks.
2753 // Non-foldable self checks work better here in second position, because
2754 // the initial primary superclass check subsumes a self-check for most
2755 // types. An exception would be a secondary type like array-of-interface,
2756 // which does not appear in its own primary supertype display.
2757 // Finally, we could have chosen to move the self-check into the
2758 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2759 // dependent manner. But it is worthwhile to have the check here,
2760 // where it can be perhaps be optimized. The cost in code space is
2761 // small (register compare, branch).
2762
2763 // Now do a linear scan of the secondary super-klass array. Again, no real
2764 // performance impact (too rare) but it's gotta be done.
2765 // Since the code is rarely used, there is no penalty for moving it
2766 // out of line, and it can only improve I-cache density.
2767 // The decision to inline or out-of-line this final check is platform
2768 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2769 Node* psc = gvn->transform(
2770 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2771
2772 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2773 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2774 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2775
2776 // Return false path; set default control to true path.
2777 *ctrl = gvn->transform(r_ok_subtype);
2778 return gvn->transform(r_not_subtype);
2779 }
2780
2781 // Profile-driven exact type check:
2782 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2783 float prob,
2784 Node* *casted_receiver) {
2785 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2786 Node* recv_klass = load_object_klass(receiver);
2787 Node* want_klass = makecon(tklass);
2788 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2789 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2790 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2791 set_control( _gvn.transform( new IfTrueNode (iff) ));
2792 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2793
2794 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2795 assert(recv_xtype->klass_is_exact(), "");
2796
2797 // Subsume downstream occurrences of receiver with a cast to
2798 // recv_xtype, since now we know what the type will be.
2799 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2800 (*casted_receiver) = _gvn.transform(cast);
2801 // (User must make the replace_in_map call.)
2802
2803 return fail;
2804 }
2805
2806
2807 //------------------------------seems_never_null-------------------------------
2808 // Use null_seen information if it is available from the profile.
2809 // If we see an unexpected null at a type check we record it and force a
2810 // recompile; the offending check will be recompiled to handle NULLs.
2811 // If we see several offending BCIs, then all checks in the
2812 // method will be recompiled.
2813 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2814 speculating = !_gvn.type(obj)->speculative_maybe_null();
2815 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2816 if (UncommonNullCast // Cutout for this technique
2817 && obj != null() // And not the -Xcomp stupid case?
2818 && !too_many_traps(reason)
2819 ) {
2820 if (speculating) {
2821 return true;
2822 }
2823 if (data == NULL)
2824 // Edge case: no mature data. Be optimistic here.
2825 return true;
2826 // If the profile has not seen a null, assume it won't happen.
2827 assert(java_bc() == Bytecodes::_checkcast ||
2828 java_bc() == Bytecodes::_instanceof ||
2829 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2830 return !data->as_BitData()->null_seen();
2831 }
2832 speculating = false;
2833 return false;
2834 }
2835
2836 //------------------------maybe_cast_profiled_receiver-------------------------
2837 // If the profile has seen exactly one type, narrow to exactly that type.
2838 // Subsequent type checks will always fold up.
2839 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2840 ciKlass* require_klass,
2841 ciKlass* spec_klass,
2842 bool safe_for_replace) {
2843 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2844
2845 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2846
2847 // Make sure we haven't already deoptimized from this tactic.
2848 if (too_many_traps(reason) || too_many_recompiles(reason))
2849 return NULL;
2850
2851 // (No, this isn't a call, but it's enough like a virtual call
2852 // to use the same ciMethod accessor to get the profile info...)
2853 // If we have a speculative type use it instead of profiling (which
2854 // may not help us)
2855 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2856 if (exact_kls != NULL) {// no cast failures here
2857 if (require_klass == NULL ||
2858 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2859 // If we narrow the type to match what the type profile sees or
2860 // the speculative type, we can then remove the rest of the
2861 // cast.
2862 // This is a win, even if the exact_kls is very specific,
2863 // because downstream operations, such as method calls,
2864 // will often benefit from the sharper type.
2865 Node* exact_obj = not_null_obj; // will get updated in place...
2866 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2867 &exact_obj);
2868 { PreserveJVMState pjvms(this);
2869 set_control(slow_ctl);
2870 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2871 }
2872 if (safe_for_replace) {
2873 replace_in_map(not_null_obj, exact_obj);
2874 }
2875 return exact_obj;
2876 }
2877 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2878 }
2879
2880 return NULL;
2881 }
2882
2883 /**
2884 * Cast obj to type and emit guard unless we had too many traps here
2885 * already
2886 *
2887 * @param obj node being casted
2888 * @param type type to cast the node to
2889 * @param not_null true if we know node cannot be null
2890 */
2891 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2892 ciKlass* type,
2893 bool not_null) {
2894 if (stopped()) {
2895 return obj;
2896 }
2897
2898 // type == NULL if profiling tells us this object is always null
2899 if (type != NULL) {
2900 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2901 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2902
2903 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2904 !too_many_traps(class_reason) &&
2905 !too_many_recompiles(class_reason)) {
2906 Node* not_null_obj = NULL;
2907 // not_null is true if we know the object is not null and
2908 // there's no need for a null check
2909 if (!not_null) {
2910 Node* null_ctl = top();
2911 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2912 assert(null_ctl->is_top(), "no null control here");
2913 } else {
2914 not_null_obj = obj;
2915 }
2916
2917 Node* exact_obj = not_null_obj;
2918 ciKlass* exact_kls = type;
2919 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2920 &exact_obj);
2921 {
2922 PreserveJVMState pjvms(this);
2923 set_control(slow_ctl);
2924 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2925 }
2926 replace_in_map(not_null_obj, exact_obj);
2927 obj = exact_obj;
2928 }
2929 } else {
2930 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2931 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2932 Node* exact_obj = null_assert(obj);
2933 replace_in_map(obj, exact_obj);
2934 obj = exact_obj;
2935 }
2936 }
2937 return obj;
2938 }
2939
2940 //-------------------------------gen_instanceof--------------------------------
2941 // Generate an instance-of idiom. Used by both the instance-of bytecode
2942 // and the reflective instance-of call.
2943 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2944 kill_dead_locals(); // Benefit all the uncommon traps
2945 assert( !stopped(), "dead parse path should be checked in callers" );
2946 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2947 "must check for not-null not-dead klass in callers");
2948
2949 // Make the merge point
2950 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2951 RegionNode* region = new RegionNode(PATH_LIMIT);
2952 Node* phi = new PhiNode(region, TypeInt::BOOL);
2953 C->set_has_split_ifs(true); // Has chance for split-if optimization
2954
2955 ciProfileData* data = NULL;
2956 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2957 data = method()->method_data()->bci_to_data(bci());
2958 }
2959 bool speculative_not_null = false;
2960 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2961 && seems_never_null(obj, data, speculative_not_null));
2962
2963 // Null check; get casted pointer; set region slot 3
2964 Node* null_ctl = top();
2965 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2966
2967 // If not_null_obj is dead, only null-path is taken
2968 if (stopped()) { // Doing instance-of on a NULL?
2969 set_control(null_ctl);
2970 return intcon(0);
2971 }
2972 region->init_req(_null_path, null_ctl);
2973 phi ->init_req(_null_path, intcon(0)); // Set null path value
2974 if (null_ctl == top()) {
2975 // Do this eagerly, so that pattern matches like is_diamond_phi
2976 // will work even during parsing.
2977 assert(_null_path == PATH_LIMIT-1, "delete last");
2978 region->del_req(_null_path);
2979 phi ->del_req(_null_path);
2980 }
2981
2982 // Do we know the type check always succeed?
2983 bool known_statically = false;
2984 if (_gvn.type(superklass)->singleton()) {
2985 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2986 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2987 if (subk != NULL && subk->is_loaded()) {
2988 int static_res = C->static_subtype_check(superk, subk);
2989 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2990 }
2991 }
2992
2993 if (!known_statically) {
2994 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2995 // We may not have profiling here or it may not help us. If we
2996 // have a speculative type use it to perform an exact cast.
2997 ciKlass* spec_obj_type = obj_type->speculative_type();
2998 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2999 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3000 if (stopped()) { // Profile disagrees with this path.
3001 set_control(null_ctl); // Null is the only remaining possibility.
3002 return intcon(0);
3003 }
3004 if (cast_obj != NULL) {
3005 not_null_obj = cast_obj;
3006 }
3007 }
3008 }
3009
3010 // Load the object's klass
3011 Node* obj_klass = load_object_klass(not_null_obj);
3012
3013 // Generate the subtype check
3014 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3015
3016 // Plug in the success path to the general merge in slot 1.
3017 region->init_req(_obj_path, control());
3018 phi ->init_req(_obj_path, intcon(1));
3019
3020 // Plug in the failing path to the general merge in slot 2.
3021 region->init_req(_fail_path, not_subtype_ctrl);
3022 phi ->init_req(_fail_path, intcon(0));
3023
3024 // Return final merged results
3025 set_control( _gvn.transform(region) );
3026 record_for_igvn(region);
3027
3028 // If we know the type check always succeeds then we don't use the
3029 // profiling data at this bytecode. Don't lose it, feed it to the
3030 // type system as a speculative type.
3031 if (safe_for_replace) {
3032 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3033 replace_in_map(obj, casted_obj);
3034 }
3035
3036 return _gvn.transform(phi);
3037 }
3038
3039 //-------------------------------gen_checkcast---------------------------------
3040 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3041 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3042 // uncommon-trap paths work. Adjust stack after this call.
3043 // If failure_control is supplied and not null, it is filled in with
3044 // the control edge for the cast failure. Otherwise, an appropriate
3045 // uncommon trap or exception is thrown.
3046 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3047 Node* *failure_control) {
3048 kill_dead_locals(); // Benefit all the uncommon traps
3049 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3050 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3051
3052 // Fast cutout: Check the case that the cast is vacuously true.
3053 // This detects the common cases where the test will short-circuit
3054 // away completely. We do this before we perform the null check,
3055 // because if the test is going to turn into zero code, we don't
3056 // want a residual null check left around. (Causes a slowdown,
3057 // for example, in some objArray manipulations, such as a[i]=a[j].)
3058 if (tk->singleton()) {
3059 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3060 if (objtp != NULL && objtp->klass() != NULL) {
3061 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3062 case Compile::SSC_always_true:
3063 // If we know the type check always succeed then we don't use
3064 // the profiling data at this bytecode. Don't lose it, feed it
3065 // to the type system as a speculative type.
3066 return record_profiled_receiver_for_speculation(obj);
3067 case Compile::SSC_always_false:
3068 // It needs a null check because a null will *pass* the cast check.
3069 // A non-null value will always produce an exception.
3070 return null_assert(obj);
3071 }
3072 }
3073 }
3074
3075 ciProfileData* data = NULL;
3076 bool safe_for_replace = false;
3077 if (failure_control == NULL) { // use MDO in regular case only
3078 assert(java_bc() == Bytecodes::_aastore ||
3079 java_bc() == Bytecodes::_checkcast,
3080 "interpreter profiles type checks only for these BCs");
3081 data = method()->method_data()->bci_to_data(bci());
3082 safe_for_replace = true;
3083 }
3084
3085 // Make the merge point
3086 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3087 RegionNode* region = new RegionNode(PATH_LIMIT);
3088 Node* phi = new PhiNode(region, toop);
3089 C->set_has_split_ifs(true); // Has chance for split-if optimization
3090
3091 // Use null-cast information if it is available
3092 bool speculative_not_null = false;
3093 bool never_see_null = ((failure_control == NULL) // regular case only
3094 && seems_never_null(obj, data, speculative_not_null));
3095
3096 // Null check; get casted pointer; set region slot 3
3097 Node* null_ctl = top();
3098 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3099
3100 // If not_null_obj is dead, only null-path is taken
3101 if (stopped()) { // Doing instance-of on a NULL?
3102 set_control(null_ctl);
3103 return null();
3104 }
3105 region->init_req(_null_path, null_ctl);
3106 phi ->init_req(_null_path, null()); // Set null path value
3107 if (null_ctl == top()) {
3108 // Do this eagerly, so that pattern matches like is_diamond_phi
3109 // will work even during parsing.
3110 assert(_null_path == PATH_LIMIT-1, "delete last");
3111 region->del_req(_null_path);
3112 phi ->del_req(_null_path);
3113 }
3114
3115 Node* cast_obj = NULL;
3116 if (tk->klass_is_exact()) {
3117 // The following optimization tries to statically cast the speculative type of the object
3118 // (for example obtained during profiling) to the type of the superklass and then do a
3119 // dynamic check that the type of the object is what we expect. To work correctly
3120 // for checkcast and aastore the type of superklass should be exact.
3121 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3122 // We may not have profiling here or it may not help us. If we have
3123 // a speculative type use it to perform an exact cast.
3124 ciKlass* spec_obj_type = obj_type->speculative_type();
3125 if (spec_obj_type != NULL || data != NULL) {
3126 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3127 if (cast_obj != NULL) {
3128 if (failure_control != NULL) // failure is now impossible
3129 (*failure_control) = top();
3130 // adjust the type of the phi to the exact klass:
3131 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3132 }
3133 }
3134 }
3135
3136 if (cast_obj == NULL) {
3137 // Load the object's klass
3138 Node* obj_klass = load_object_klass(not_null_obj);
3139
3140 // Generate the subtype check
3141 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3142
3143 // Plug in success path into the merge
3144 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3145 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3146 if (failure_control == NULL) {
3147 if (not_subtype_ctrl != top()) { // If failure is possible
3148 PreserveJVMState pjvms(this);
3149 set_control(not_subtype_ctrl);
3150 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3151 }
3152 } else {
3153 (*failure_control) = not_subtype_ctrl;
3154 }
3155 }
3156
3157 region->init_req(_obj_path, control());
3158 phi ->init_req(_obj_path, cast_obj);
3159
3160 // A merge of NULL or Casted-NotNull obj
3161 Node* res = _gvn.transform(phi);
3162
3163 // Note I do NOT always 'replace_in_map(obj,result)' here.
3164 // if( tk->klass()->can_be_primary_super() )
3165 // This means that if I successfully store an Object into an array-of-String
3166 // I 'forget' that the Object is really now known to be a String. I have to
3167 // do this because we don't have true union types for interfaces - if I store
3168 // a Baz into an array-of-Interface and then tell the optimizer it's an
3169 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3170 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3171 // replace_in_map( obj, res );
3172
3173 // Return final merged results
3174 set_control( _gvn.transform(region) );
3175 record_for_igvn(region);
3176
3177 return record_profiled_receiver_for_speculation(res);
3178 }
3179
3180 //------------------------------next_monitor-----------------------------------
3181 // What number should be given to the next monitor?
3182 int GraphKit::next_monitor() {
3183 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3184 int next = current + C->sync_stack_slots();
3185 // Keep the toplevel high water mark current:
3186 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3187 return current;
3188 }
3189
3190 //------------------------------insert_mem_bar---------------------------------
3191 // Memory barrier to avoid floating things around
3192 // The membar serves as a pinch point between both control and all memory slices.
3193 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3194 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3195 mb->init_req(TypeFunc::Control, control());
3196 mb->init_req(TypeFunc::Memory, reset_memory());
3197 Node* membar = _gvn.transform(mb);
3198 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3199 set_all_memory_call(membar);
3200 return membar;
3201 }
3202
3203 //-------------------------insert_mem_bar_volatile----------------------------
3204 // Memory barrier to avoid floating things around
3205 // The membar serves as a pinch point between both control and memory(alias_idx).
3206 // If you want to make a pinch point on all memory slices, do not use this
3207 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3208 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3209 // When Parse::do_put_xxx updates a volatile field, it appends a series
3210 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3211 // The first membar is on the same memory slice as the field store opcode.
3212 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3213 // All the other membars (for other volatile slices, including AliasIdxBot,
3214 // which stands for all unknown volatile slices) are control-dependent
3215 // on the first membar. This prevents later volatile loads or stores
3216 // from sliding up past the just-emitted store.
3217
3218 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3219 mb->set_req(TypeFunc::Control,control());
3220 if (alias_idx == Compile::AliasIdxBot) {
3221 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3222 } else {
3223 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3224 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3225 }
3226 Node* membar = _gvn.transform(mb);
3227 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3228 if (alias_idx == Compile::AliasIdxBot) {
3229 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3230 } else {
3231 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3232 }
3233 return membar;
3234 }
3235
3236 //------------------------------shared_lock------------------------------------
3237 // Emit locking code.
3238 FastLockNode* GraphKit::shared_lock(Node* obj) {
3239 // bci is either a monitorenter bc or InvocationEntryBci
3240 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3241 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3242
3243 if( !GenerateSynchronizationCode )
3244 return NULL; // Not locking things?
3245 if (stopped()) // Dead monitor?
3246 return NULL;
3247
3248 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3249
3250 // Box the stack location
3251 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3252 Node* mem = reset_memory();
3253
3254 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3255 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3256 // Create the counters for this fast lock.
3257 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3258 }
3259
3260 // Create the rtm counters for this fast lock if needed.
3261 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3262
3263 // Add monitor to debug info for the slow path. If we block inside the
3264 // slow path and de-opt, we need the monitor hanging around
3265 map()->push_monitor( flock );
3266
3267 const TypeFunc *tf = LockNode::lock_type();
3268 LockNode *lock = new LockNode(C, tf);
3269
3270 lock->init_req( TypeFunc::Control, control() );
3271 lock->init_req( TypeFunc::Memory , mem );
3272 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3273 lock->init_req( TypeFunc::FramePtr, frameptr() );
3274 lock->init_req( TypeFunc::ReturnAdr, top() );
3275
3276 lock->init_req(TypeFunc::Parms + 0, obj);
3277 lock->init_req(TypeFunc::Parms + 1, box);
3278 lock->init_req(TypeFunc::Parms + 2, flock);
3279 add_safepoint_edges(lock);
3280
3281 lock = _gvn.transform( lock )->as_Lock();
3282
3283 // lock has no side-effects, sets few values
3284 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3285
3286 insert_mem_bar(Op_MemBarAcquireLock);
3287
3288 // Add this to the worklist so that the lock can be eliminated
3289 record_for_igvn(lock);
3290
3291 #ifndef PRODUCT
3292 if (PrintLockStatistics) {
3293 // Update the counter for this lock. Don't bother using an atomic
3294 // operation since we don't require absolute accuracy.
3295 lock->create_lock_counter(map()->jvms());
3296 increment_counter(lock->counter()->addr());
3297 }
3298 #endif
3299
3300 return flock;
3301 }
3302
3303
3304 //------------------------------shared_unlock----------------------------------
3305 // Emit unlocking code.
3306 void GraphKit::shared_unlock(Node* box, Node* obj) {
3307 // bci is either a monitorenter bc or InvocationEntryBci
3308 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3309 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3310
3311 if( !GenerateSynchronizationCode )
3312 return;
3313 if (stopped()) { // Dead monitor?
3314 map()->pop_monitor(); // Kill monitor from debug info
3315 return;
3316 }
3317
3318 // Memory barrier to avoid floating things down past the locked region
3319 insert_mem_bar(Op_MemBarReleaseLock);
3320
3321 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3322 UnlockNode *unlock = new UnlockNode(C, tf);
3323 #ifdef ASSERT
3324 unlock->set_dbg_jvms(sync_jvms());
3325 #endif
3326 uint raw_idx = Compile::AliasIdxRaw;
3327 unlock->init_req( TypeFunc::Control, control() );
3328 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3329 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3330 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3331 unlock->init_req( TypeFunc::ReturnAdr, top() );
3332
3333 unlock->init_req(TypeFunc::Parms + 0, obj);
3334 unlock->init_req(TypeFunc::Parms + 1, box);
3335 unlock = _gvn.transform(unlock)->as_Unlock();
3336
3337 Node* mem = reset_memory();
3338
3339 // unlock has no side-effects, sets few values
3340 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3341
3342 // Kill monitor from debug info
3343 map()->pop_monitor( );
3344 }
3345
3346 //-------------------------------get_layout_helper-----------------------------
3347 // If the given klass is a constant or known to be an array,
3348 // fetch the constant layout helper value into constant_value
3349 // and return (Node*)NULL. Otherwise, load the non-constant
3350 // layout helper value, and return the node which represents it.
3351 // This two-faced routine is useful because allocation sites
3352 // almost always feature constant types.
3353 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3354 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3355 if (!StressReflectiveCode && inst_klass != NULL) {
3356 ciKlass* klass = inst_klass->klass();
3357 bool xklass = inst_klass->klass_is_exact();
3358 if (xklass || klass->is_array_klass()) {
3359 jint lhelper = klass->layout_helper();
3360 if (lhelper != Klass::_lh_neutral_value) {
3361 constant_value = lhelper;
3362 return (Node*) NULL;
3363 }
3364 }
3365 }
3366 constant_value = Klass::_lh_neutral_value; // put in a known value
3367 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3368 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3369 }
3370
3371 // We just put in an allocate/initialize with a big raw-memory effect.
3372 // Hook selected additional alias categories on the initialization.
3373 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3374 MergeMemNode* init_in_merge,
3375 Node* init_out_raw) {
3376 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3377 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3378
3379 Node* prevmem = kit.memory(alias_idx);
3380 init_in_merge->set_memory_at(alias_idx, prevmem);
3381 kit.set_memory(init_out_raw, alias_idx);
3382 }
3383
3384 //---------------------------set_output_for_allocation-------------------------
3385 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3386 const TypeOopPtr* oop_type,
3387 bool deoptimize_on_exception) {
3388 int rawidx = Compile::AliasIdxRaw;
3389 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3390 add_safepoint_edges(alloc);
3391 Node* allocx = _gvn.transform(alloc);
3392 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3393 // create memory projection for i_o
3394 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3395 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3396
3397 // create a memory projection as for the normal control path
3398 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3399 set_memory(malloc, rawidx);
3400
3401 // a normal slow-call doesn't change i_o, but an allocation does
3402 // we create a separate i_o projection for the normal control path
3403 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3404 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3405
3406 // put in an initialization barrier
3407 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3408 rawoop)->as_Initialize();
3409 assert(alloc->initialization() == init, "2-way macro link must work");
3410 assert(init ->allocation() == alloc, "2-way macro link must work");
3411 {
3412 // Extract memory strands which may participate in the new object's
3413 // initialization, and source them from the new InitializeNode.
3414 // This will allow us to observe initializations when they occur,
3415 // and link them properly (as a group) to the InitializeNode.
3416 assert(init->in(InitializeNode::Memory) == malloc, "");
3417 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3418 init->set_req(InitializeNode::Memory, minit_in);
3419 record_for_igvn(minit_in); // fold it up later, if possible
3420 Node* minit_out = memory(rawidx);
3421 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3422 if (oop_type->isa_aryptr()) {
3423 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3424 int elemidx = C->get_alias_index(telemref);
3425 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3426 } else if (oop_type->isa_instptr()) {
3427 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3428 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3429 ciField* field = ik->nonstatic_field_at(i);
3430 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3431 continue; // do not bother to track really large numbers of fields
3432 // Find (or create) the alias category for this field:
3433 int fieldidx = C->alias_type(field)->index();
3434 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3435 }
3436 }
3437 }
3438
3439 // Cast raw oop to the real thing...
3440 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3441 javaoop = _gvn.transform(javaoop);
3442 C->set_recent_alloc(control(), javaoop);
3443 assert(just_allocated_object(control()) == javaoop, "just allocated");
3444
3445 #ifdef ASSERT
3446 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3447 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3448 "Ideal_allocation works");
3449 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3450 "Ideal_allocation works");
3451 if (alloc->is_AllocateArray()) {
3452 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3453 "Ideal_allocation works");
3454 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3455 "Ideal_allocation works");
3456 } else {
3457 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3458 }
3459 }
3460 #endif //ASSERT
3461
3462 return javaoop;
3463 }
3464
3465 //---------------------------new_instance--------------------------------------
3466 // This routine takes a klass_node which may be constant (for a static type)
3467 // or may be non-constant (for reflective code). It will work equally well
3468 // for either, and the graph will fold nicely if the optimizer later reduces
3469 // the type to a constant.
3470 // The optional arguments are for specialized use by intrinsics:
3471 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3472 // - If 'return_size_val', report the the total object size to the caller.
3473 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3474 Node* GraphKit::new_instance(Node* klass_node,
3475 Node* extra_slow_test,
3476 Node* *return_size_val,
3477 bool deoptimize_on_exception) {
3478 // Compute size in doublewords
3479 // The size is always an integral number of doublewords, represented
3480 // as a positive bytewise size stored in the klass's layout_helper.
3481 // The layout_helper also encodes (in a low bit) the need for a slow path.
3482 jint layout_con = Klass::_lh_neutral_value;
3483 Node* layout_val = get_layout_helper(klass_node, layout_con);
3484 int layout_is_con = (layout_val == NULL);
3485
3486 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3487 // Generate the initial go-slow test. It's either ALWAYS (return a
3488 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3489 // case) a computed value derived from the layout_helper.
3490 Node* initial_slow_test = NULL;
3491 if (layout_is_con) {
3492 assert(!StressReflectiveCode, "stress mode does not use these paths");
3493 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3494 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3495 } else { // reflective case
3496 // This reflective path is used by Unsafe.allocateInstance.
3497 // (It may be stress-tested by specifying StressReflectiveCode.)
3498 // Basically, we want to get into the VM is there's an illegal argument.
3499 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3500 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3501 if (extra_slow_test != intcon(0)) {
3502 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3503 }
3504 // (Macro-expander will further convert this to a Bool, if necessary.)
3505 }
3506
3507 // Find the size in bytes. This is easy; it's the layout_helper.
3508 // The size value must be valid even if the slow path is taken.
3509 Node* size = NULL;
3510 if (layout_is_con) {
3511 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3512 } else { // reflective case
3513 // This reflective path is used by clone and Unsafe.allocateInstance.
3514 size = ConvI2X(layout_val);
3515
3516 // Clear the low bits to extract layout_helper_size_in_bytes:
3517 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3518 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3519 size = _gvn.transform( new AndXNode(size, mask) );
3520 }
3521 if (return_size_val != NULL) {
3522 (*return_size_val) = size;
3523 }
3524
3525 // This is a precise notnull oop of the klass.
3526 // (Actually, it need not be precise if this is a reflective allocation.)
3527 // It's what we cast the result to.
3528 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3529 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3530 const TypeOopPtr* oop_type = tklass->as_instance_type();
3531
3532 // Now generate allocation code
3533
3534 // The entire memory state is needed for slow path of the allocation
3535 // since GC and deoptimization can happened.
3536 Node *mem = reset_memory();
3537 set_all_memory(mem); // Create new memory state
3538
3539 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3540 control(), mem, i_o(),
3541 size, klass_node,
3542 initial_slow_test);
3543
3544 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3545 }
3546
3547 //-------------------------------new_array-------------------------------------
3548 // helper for both newarray and anewarray
3549 // The 'length' parameter is (obviously) the length of the array.
3550 // See comments on new_instance for the meaning of the other arguments.
3551 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3552 Node* length, // number of array elements
3553 int nargs, // number of arguments to push back for uncommon trap
3554 Node* *return_size_val,
3555 bool deoptimize_on_exception) {
3556 jint layout_con = Klass::_lh_neutral_value;
3557 Node* layout_val = get_layout_helper(klass_node, layout_con);
3558 int layout_is_con = (layout_val == NULL);
3559
3560 if (!layout_is_con && !StressReflectiveCode &&
3561 !too_many_traps(Deoptimization::Reason_class_check)) {
3562 // This is a reflective array creation site.
3563 // Optimistically assume that it is a subtype of Object[],
3564 // so that we can fold up all the address arithmetic.
3565 layout_con = Klass::array_layout_helper(T_OBJECT);
3566 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3567 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3568 { BuildCutout unless(this, bol_lh, PROB_MAX);
3569 inc_sp(nargs);
3570 uncommon_trap(Deoptimization::Reason_class_check,
3571 Deoptimization::Action_maybe_recompile);
3572 }
3573 layout_val = NULL;
3574 layout_is_con = true;
3575 }
3576
3577 // Generate the initial go-slow test. Make sure we do not overflow
3578 // if length is huge (near 2Gig) or negative! We do not need
3579 // exact double-words here, just a close approximation of needed
3580 // double-words. We can't add any offset or rounding bits, lest we
3581 // take a size -1 of bytes and make it positive. Use an unsigned
3582 // compare, so negative sizes look hugely positive.
3583 int fast_size_limit = FastAllocateSizeLimit;
3584 if (layout_is_con) {
3585 assert(!StressReflectiveCode, "stress mode does not use these paths");
3586 // Increase the size limit if we have exact knowledge of array type.
3587 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3588 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3589 }
3590
3591 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3592 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3593
3594 // --- Size Computation ---
3595 // array_size = round_to_heap(array_header + (length << elem_shift));
3596 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3597 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3598 // The rounding mask is strength-reduced, if possible.
3599 int round_mask = MinObjAlignmentInBytes - 1;
3600 Node* header_size = NULL;
3601 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3602 // (T_BYTE has the weakest alignment and size restrictions...)
3603 if (layout_is_con) {
3604 int hsize = Klass::layout_helper_header_size(layout_con);
3605 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3606 BasicType etype = Klass::layout_helper_element_type(layout_con);
3607 if ((round_mask & ~right_n_bits(eshift)) == 0)
3608 round_mask = 0; // strength-reduce it if it goes away completely
3609 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3610 assert(header_size_min <= hsize, "generic minimum is smallest");
3611 header_size_min = hsize;
3612 header_size = intcon(hsize + round_mask);
3613 } else {
3614 Node* hss = intcon(Klass::_lh_header_size_shift);
3615 Node* hsm = intcon(Klass::_lh_header_size_mask);
3616 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3617 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3618 Node* mask = intcon(round_mask);
3619 header_size = _gvn.transform( new AddINode(hsize, mask) );
3620 }
3621
3622 Node* elem_shift = NULL;
3623 if (layout_is_con) {
3624 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3625 if (eshift != 0)
3626 elem_shift = intcon(eshift);
3627 } else {
3628 // There is no need to mask or shift this value.
3629 // The semantics of LShiftINode include an implicit mask to 0x1F.
3630 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3631 elem_shift = layout_val;
3632 }
3633
3634 // Transition to native address size for all offset calculations:
3635 Node* lengthx = ConvI2X(length);
3636 Node* headerx = ConvI2X(header_size);
3637 #ifdef _LP64
3638 { const TypeInt* tilen = _gvn.find_int_type(length);
3639 if (tilen != NULL && tilen->_lo < 0) {
3640 // Add a manual constraint to a positive range. Cf. array_element_address.
3641 jint size_max = fast_size_limit;
3642 if (size_max > tilen->_hi) size_max = tilen->_hi;
3643 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3644
3645 // Only do a narrow I2L conversion if the range check passed.
3646 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3647 _gvn.transform(iff);
3648 RegionNode* region = new RegionNode(3);
3649 _gvn.set_type(region, Type::CONTROL);
3650 lengthx = new PhiNode(region, TypeLong::LONG);
3651 _gvn.set_type(lengthx, TypeLong::LONG);
3652
3653 // Range check passed. Use ConvI2L node with narrow type.
3654 Node* passed = IfFalse(iff);
3655 region->init_req(1, passed);
3656 // Make I2L conversion control dependent to prevent it from
3657 // floating above the range check during loop optimizations.
3658 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3659
3660 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3661 region->init_req(2, IfTrue(iff));
3662 lengthx->init_req(2, ConvI2X(length));
3663
3664 set_control(region);
3665 record_for_igvn(region);
3666 record_for_igvn(lengthx);
3667 }
3668 }
3669 #endif
3670
3671 // Combine header size (plus rounding) and body size. Then round down.
3672 // This computation cannot overflow, because it is used only in two
3673 // places, one where the length is sharply limited, and the other
3674 // after a successful allocation.
3675 Node* abody = lengthx;
3676 if (elem_shift != NULL)
3677 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3678 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3679 if (round_mask != 0) {
3680 Node* mask = MakeConX(~round_mask);
3681 size = _gvn.transform( new AndXNode(size, mask) );
3682 }
3683 // else if round_mask == 0, the size computation is self-rounding
3684
3685 if (return_size_val != NULL) {
3686 // This is the size
3687 (*return_size_val) = size;
3688 }
3689
3690 // Now generate allocation code
3691
3692 // The entire memory state is needed for slow path of the allocation
3693 // since GC and deoptimization can happened.
3694 Node *mem = reset_memory();
3695 set_all_memory(mem); // Create new memory state
3696
3697 if (initial_slow_test->is_Bool()) {
3698 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3699 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3700 }
3701
3702 // Create the AllocateArrayNode and its result projections
3703 AllocateArrayNode* alloc
3704 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3705 control(), mem, i_o(),
3706 size, klass_node,
3707 initial_slow_test,
3708 length);
3709
3710 // Cast to correct type. Note that the klass_node may be constant or not,
3711 // and in the latter case the actual array type will be inexact also.
3712 // (This happens via a non-constant argument to inline_native_newArray.)
3713 // In any case, the value of klass_node provides the desired array type.
3714 const TypeInt* length_type = _gvn.find_int_type(length);
3715 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3716 if (ary_type->isa_aryptr() && length_type != NULL) {
3717 // Try to get a better type than POS for the size
3718 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3719 }
3720
3721 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3722
3723 // Cast length on remaining path to be as narrow as possible
3724 if (map()->find_edge(length) >= 0) {
3725 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3726 if (ccast != length) {
3727 _gvn.set_type_bottom(ccast);
3728 record_for_igvn(ccast);
3729 replace_in_map(length, ccast);
3730 }
3731 }
3732
3733 return javaoop;
3734 }
3735
3736 // The following "Ideal_foo" functions are placed here because they recognize
3737 // the graph shapes created by the functions immediately above.
3738
3739 //---------------------------Ideal_allocation----------------------------------
3740 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3741 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3742 if (ptr == NULL) { // reduce dumb test in callers
3743 return NULL;
3744 }
3745 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3746 ptr = ptr->in(1);
3747 if (ptr == NULL) return NULL;
3748 }
3749 // Return NULL for allocations with several casts:
3750 // j.l.reflect.Array.newInstance(jobject, jint)
3751 // Object.clone()
3752 // to keep more precise type from last cast.
3753 if (ptr->is_Proj()) {
3754 Node* allo = ptr->in(0);
3755 if (allo != NULL && allo->is_Allocate()) {
3756 return allo->as_Allocate();
3757 }
3758 }
3759 // Report failure to match.
3760 return NULL;
3761 }
3762
3763 // Fancy version which also strips off an offset (and reports it to caller).
3764 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3765 intptr_t& offset) {
3766 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3767 if (base == NULL) return NULL;
3768 return Ideal_allocation(base, phase);
3769 }
3770
3771 // Trace Initialize <- Proj[Parm] <- Allocate
3772 AllocateNode* InitializeNode::allocation() {
3773 Node* rawoop = in(InitializeNode::RawAddress);
3774 if (rawoop->is_Proj()) {
3775 Node* alloc = rawoop->in(0);
3776 if (alloc->is_Allocate()) {
3777 return alloc->as_Allocate();
3778 }
3779 }
3780 return NULL;
3781 }
3782
3783 // Trace Allocate -> Proj[Parm] -> Initialize
3784 InitializeNode* AllocateNode::initialization() {
3785 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
3786 if (rawoop == NULL) return NULL;
3787 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3788 Node* init = rawoop->fast_out(i);
3789 if (init->is_Initialize()) {
3790 assert(init->as_Initialize()->allocation() == this, "2-way link");
3791 return init->as_Initialize();
3792 }
3793 }
3794 return NULL;
3795 }
3796
3797 //----------------------------- loop predicates ---------------------------
3798
3799 //------------------------------add_predicate_impl----------------------------
3800 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3801 // Too many traps seen?
3802 if (too_many_traps(reason)) {
3803 #ifdef ASSERT
3804 if (TraceLoopPredicate) {
3805 int tc = C->trap_count(reason);
3806 tty->print("too many traps=%s tcount=%d in ",
3807 Deoptimization::trap_reason_name(reason), tc);
3808 method()->print(); // which method has too many predicate traps
3809 tty->cr();
3810 }
3811 #endif
3812 // We cannot afford to take more traps here,
3813 // do not generate predicate.
3814 return;
3815 }
3816
3817 Node *cont = _gvn.intcon(1);
3818 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
3819 Node *bol = _gvn.transform(new Conv2BNode(opq));
3820 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3821 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3822 C->add_predicate_opaq(opq);
3823 {
3824 PreserveJVMState pjvms(this);
3825 set_control(iffalse);
3826 inc_sp(nargs);
3827 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3828 }
3829 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3830 set_control(iftrue);
3831 }
3832
3833 //------------------------------add_predicate---------------------------------
3834 void GraphKit::add_predicate(int nargs) {
3835 if (UseLoopPredicate) {
3836 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3837 }
3838 // loop's limit check predicate should be near the loop.
3839 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3840 }
3841
3842 //----------------------------- store barriers ----------------------------
3843 #define __ ideal.
3844
3845 bool GraphKit::use_ReduceInitialCardMarks() {
3846 BarrierSet *bs = Universe::heap()->barrier_set();
3847 return bs->is_a(BarrierSet::CardTableModRef)
3848 && barrier_set_cast<CardTableModRefBS>(bs)->can_elide_tlab_store_barriers()
3849 && ReduceInitialCardMarks;
3850 }
3851
3852 void GraphKit::sync_kit(IdealKit& ideal) {
3853 set_all_memory(__ merged_memory());
3854 set_i_o(__ i_o());
3855 set_control(__ ctrl());
3856 }
3857
3858 void GraphKit::final_sync(IdealKit& ideal) {
3859 // Final sync IdealKit and graphKit.
3860 sync_kit(ideal);
3861 }
3862
3863 Node* GraphKit::byte_map_base_node() {
3864 // Get base of card map
3865 jbyte* card_table_base = ci_card_table_address();
3866 if (card_table_base != NULL) {
3867 return makecon(TypeRawPtr::make((address)card_table_base));
3868 } else {
3869 return null();
3870 }
3871 }
3872
3873 // vanilla/CMS post barrier
3874 // Insert a write-barrier store. This is to let generational GC work; we have
3875 // to flag all oop-stores before the next GC point.
3876 void GraphKit::write_barrier_post(Node* oop_store,
3877 Node* obj,
3878 Node* adr,
3879 uint adr_idx,
3880 Node* val,
3881 bool use_precise) {
3882 // No store check needed if we're storing a NULL or an old object
3883 // (latter case is probably a string constant). The concurrent
3884 // mark sweep garbage collector, however, needs to have all nonNull
3885 // oop updates flagged via card-marks.
3886 if (val != NULL && val->is_Con()) {
3887 // must be either an oop or NULL
3888 const Type* t = val->bottom_type();
3889 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3890 // stores of null never (?) need barriers
3891 return;
3892 }
3893
3894 if (use_ReduceInitialCardMarks()
3895 && obj == just_allocated_object(control())) {
3896 // We can skip marks on a freshly-allocated object in Eden.
3897 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
3898 // That routine informs GC to take appropriate compensating steps,
3899 // upon a slow-path allocation, so as to make this card-mark
3900 // elision safe.
3901 return;
3902 }
3903
3904 if (!use_precise) {
3905 // All card marks for a (non-array) instance are in one place:
3906 adr = obj;
3907 }
3908 // (Else it's an array (or unknown), and we want more precise card marks.)
3909 assert(adr != NULL, "");
3910
3911 IdealKit ideal(this, true);
3912
3913 // Convert the pointer to an int prior to doing math on it
3914 Node* cast = __ CastPX(__ ctrl(), adr);
3915
3916 // Divide by card size
3917 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef),
3918 "Only one we handle so far.");
3919 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
3920
3921 // Combine card table base and card offset
3922 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3923
3924 // Get the alias_index for raw card-mark memory
3925 int adr_type = Compile::AliasIdxRaw;
3926 Node* zero = __ ConI(0); // Dirty card value
3927 BasicType bt = T_BYTE;
3928
3929 if (UseConcMarkSweepGC && UseCondCardMark) {
3930 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
3931 __ sync_kit(this);
3932 }
3933
3934 if (UseCondCardMark) {
3935 // The classic GC reference write barrier is typically implemented
3936 // as a store into the global card mark table. Unfortunately
3937 // unconditional stores can result in false sharing and excessive
3938 // coherence traffic as well as false transactional aborts.
3939 // UseCondCardMark enables MP "polite" conditional card mark
3940 // stores. In theory we could relax the load from ctrl() to
3941 // no_ctrl, but that doesn't buy much latitude.
3942 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3943 __ if_then(card_val, BoolTest::ne, zero);
3944 }
3945
3946 // Smash zero into card
3947 if( !UseConcMarkSweepGC ) {
3948 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3949 } else {
3950 // Specialized path for CM store barrier
3951 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3952 }
3953
3954 if (UseCondCardMark) {
3955 __ end_if();
3956 }
3957
3958 // Final sync IdealKit and GraphKit.
3959 final_sync(ideal);
3960 }
3961 /*
3962 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
3963 * required by SATB to make sure all objects live at the start of the
3964 * marking are kept alive, all reference updates need to any previous
3965 * reference stored before writing.
3966 *
3967 * If the previous value is NULL there is no need to save the old value.
3968 * References that are NULL are filtered during runtime by the barrier
3969 * code to avoid unnecessary queuing.
3970 *
3971 * However in the case of newly allocated objects it might be possible to
3972 * prove that the reference about to be overwritten is NULL during compile
3973 * time and avoid adding the barrier code completely.
3974 *
3975 * The compiler needs to determine that the object in which a field is about
3976 * to be written is newly allocated, and that no prior store to the same field
3977 * has happened since the allocation.
3978 *
3979 * Returns true if the pre-barrier can be removed
3980 */
3981 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
3982 BasicType bt, uint adr_idx) {
3983 intptr_t offset = 0;
3984 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3985 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
3986
3987 if (offset == Type::OffsetBot) {
3988 return false; // cannot unalias unless there are precise offsets
3989 }
3990
3991 if (alloc == NULL) {
3992 return false; // No allocation found
3993 }
3994
3995 intptr_t size_in_bytes = type2aelembytes(bt);
3996
3997 Node* mem = memory(adr_idx); // start searching here...
3998
3999 for (int cnt = 0; cnt < 50; cnt++) {
4000
4001 if (mem->is_Store()) {
4002
4003 Node* st_adr = mem->in(MemNode::Address);
4004 intptr_t st_offset = 0;
4005 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
4006
4007 if (st_base == NULL) {
4008 break; // inscrutable pointer
4009 }
4010
4011 // Break we have found a store with same base and offset as ours so break
4012 if (st_base == base && st_offset == offset) {
4013 break;
4014 }
4015
4016 if (st_offset != offset && st_offset != Type::OffsetBot) {
4017 const int MAX_STORE = BytesPerLong;
4018 if (st_offset >= offset + size_in_bytes ||
4019 st_offset <= offset - MAX_STORE ||
4020 st_offset <= offset - mem->as_Store()->memory_size()) {
4021 // Success: The offsets are provably independent.
4022 // (You may ask, why not just test st_offset != offset and be done?
4023 // The answer is that stores of different sizes can co-exist
4024 // in the same sequence of RawMem effects. We sometimes initialize
4025 // a whole 'tile' of array elements with a single jint or jlong.)
4026 mem = mem->in(MemNode::Memory);
4027 continue; // advance through independent store memory
4028 }
4029 }
4030
4031 if (st_base != base
4032 && MemNode::detect_ptr_independence(base, alloc, st_base,
4033 AllocateNode::Ideal_allocation(st_base, phase),
4034 phase)) {
4035 // Success: The bases are provably independent.
4036 mem = mem->in(MemNode::Memory);
4037 continue; // advance through independent store memory
4038 }
4039 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4040
4041 InitializeNode* st_init = mem->in(0)->as_Initialize();
4042 AllocateNode* st_alloc = st_init->allocation();
4043
4044 // Make sure that we are looking at the same allocation site.
4045 // The alloc variable is guaranteed to not be null here from earlier check.
4046 if (alloc == st_alloc) {
4047 // Check that the initialization is storing NULL so that no previous store
4048 // has been moved up and directly write a reference
4049 Node* captured_store = st_init->find_captured_store(offset,
4050 type2aelembytes(T_OBJECT),
4051 phase);
4052 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4053 return true;
4054 }
4055 }
4056 }
4057
4058 // Unless there is an explicit 'continue', we must bail out here,
4059 // because 'mem' is an inscrutable memory state (e.g., a call).
4060 break;
4061 }
4062
4063 return false;
4064 }
4065
4066 // G1 pre/post barriers
4067 void GraphKit::g1_write_barrier_pre(bool do_load,
4068 Node* obj,
4069 Node* adr,
4070 uint alias_idx,
4071 Node* val,
4072 const TypeOopPtr* val_type,
4073 Node* pre_val,
4074 BasicType bt) {
4075
4076 // Some sanity checks
4077 // Note: val is unused in this routine.
4078
4079 if (do_load) {
4080 // We need to generate the load of the previous value
4081 assert(obj != NULL, "must have a base");
4082 assert(adr != NULL, "where are loading from?");
4083 assert(pre_val == NULL, "loaded already?");
4084 assert(val_type != NULL, "need a type");
4085
4086 if (use_ReduceInitialCardMarks()
4087 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4088 return;
4089 }
4090
4091 } else {
4092 // In this case both val_type and alias_idx are unused.
4093 assert(pre_val != NULL, "must be loaded already");
4094 // Nothing to be done if pre_val is null.
4095 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4096 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4097 }
4098 assert(bt == T_OBJECT, "or we shouldn't be here");
4099
4100 IdealKit ideal(this, true);
4101
4102 Node* tls = __ thread(); // ThreadLocalStorage
4103
4104 Node* no_ctrl = NULL;
4105 Node* no_base = __ top();
4106 Node* zero = __ ConI(0);
4107 Node* zeroX = __ ConX(0);
4108
4109 float likely = PROB_LIKELY(0.999);
4110 float unlikely = PROB_UNLIKELY(0.999);
4111
4112 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4113 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4114
4115 // Offsets into the thread
4116 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
4117 SATBMarkQueue::byte_offset_of_active());
4118 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
4119 SATBMarkQueue::byte_offset_of_index());
4120 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
4121 SATBMarkQueue::byte_offset_of_buf());
4122
4123 // Now the actual pointers into the thread
4124 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4125 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4126 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4127
4128 // Now some of the values
4129 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4130
4131 // if (!marking)
4132 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4133 BasicType index_bt = TypeX_X->basic_type();
4134 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4135 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4136
4137 if (do_load) {
4138 // load original value
4139 // alias_idx correct??
4140 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4141 }
4142
4143 // if (pre_val != NULL)
4144 __ if_then(pre_val, BoolTest::ne, null()); {
4145 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4146
4147 // is the queue for this thread full?
4148 __ if_then(index, BoolTest::ne, zeroX, likely); {
4149
4150 // decrement the index
4151 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4152
4153 // Now get the buffer location we will log the previous value into and store it
4154 Node *log_addr = __ AddP(no_base, buffer, next_index);
4155 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4156 // update the index
4157 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4158
4159 } __ else_(); {
4160
4161 // logging buffer is full, call the runtime
4162 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4163 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4164 } __ end_if(); // (!index)
4165 } __ end_if(); // (pre_val != NULL)
4166 } __ end_if(); // (!marking)
4167
4168 // Final sync IdealKit and GraphKit.
4169 final_sync(ideal);
4170 }
4171
4172 /*
4173 * G1 similar to any GC with a Young Generation requires a way to keep track of
4174 * references from Old Generation to Young Generation to make sure all live
4175 * objects are found. G1 also requires to keep track of object references
4176 * between different regions to enable evacuation of old regions, which is done
4177 * as part of mixed collections. References are tracked in remembered sets and
4178 * is continuously updated as reference are written to with the help of the
4179 * post-barrier.
4180 *
4181 * To reduce the number of updates to the remembered set the post-barrier
4182 * filters updates to fields in objects located in the Young Generation,
4183 * the same region as the reference, when the NULL is being written or
4184 * if the card is already marked as dirty by an earlier write.
4185 *
4186 * Under certain circumstances it is possible to avoid generating the
4187 * post-barrier completely if it is possible during compile time to prove
4188 * the object is newly allocated and that no safepoint exists between the
4189 * allocation and the store.
4190 *
4191 * In the case of slow allocation the allocation code must handle the barrier
4192 * as part of the allocation in the case the allocated object is not located
4193 * in the nursery, this would happen for humongous objects. This is similar to
4194 * how CMS is required to handle this case, see the comments for the method
4195 * CardTableModRefBS::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier.
4196 * A deferred card mark is required for these objects and handled in the above
4197 * mentioned methods.
4198 *
4199 * Returns true if the post barrier can be removed
4200 */
4201 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4202 Node* adr) {
4203 intptr_t offset = 0;
4204 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4205 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4206
4207 if (offset == Type::OffsetBot) {
4208 return false; // cannot unalias unless there are precise offsets
4209 }
4210
4211 if (alloc == NULL) {
4212 return false; // No allocation found
4213 }
4214
4215 // Start search from Store node
4216 Node* mem = store->in(MemNode::Control);
4217 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4218
4219 InitializeNode* st_init = mem->in(0)->as_Initialize();
4220 AllocateNode* st_alloc = st_init->allocation();
4221
4222 // Make sure we are looking at the same allocation
4223 if (alloc == st_alloc) {
4224 return true;
4225 }
4226 }
4227
4228 return false;
4229 }
4230
4231 //
4232 // Update the card table and add card address to the queue
4233 //
4234 void GraphKit::g1_mark_card(IdealKit& ideal,
4235 Node* card_adr,
4236 Node* oop_store,
4237 uint oop_alias_idx,
4238 Node* index,
4239 Node* index_adr,
4240 Node* buffer,
4241 const TypeFunc* tf) {
4242
4243 Node* zero = __ ConI(0);
4244 Node* zeroX = __ ConX(0);
4245 Node* no_base = __ top();
4246 BasicType card_bt = T_BYTE;
4247 // Smash zero into card. MUST BE ORDERED WRT TO STORE
4248 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4249
4250 // Now do the queue work
4251 __ if_then(index, BoolTest::ne, zeroX); {
4252
4253 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4254 Node* log_addr = __ AddP(no_base, buffer, next_index);
4255
4256 // Order, see storeCM.
4257 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4258 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4259
4260 } __ else_(); {
4261 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4262 } __ end_if();
4263
4264 }
4265
4266 void GraphKit::g1_write_barrier_post(Node* oop_store,
4267 Node* obj,
4268 Node* adr,
4269 uint alias_idx,
4270 Node* val,
4271 BasicType bt,
4272 bool use_precise) {
4273 // If we are writing a NULL then we need no post barrier
4274
4275 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4276 // Must be NULL
4277 const Type* t = val->bottom_type();
4278 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4279 // No post barrier if writing NULLx
4280 return;
4281 }
4282
4283 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4284 // We can skip marks on a freshly-allocated object in Eden.
4285 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
4286 // That routine informs GC to take appropriate compensating steps,
4287 // upon a slow-path allocation, so as to make this card-mark
4288 // elision safe.
4289 return;
4290 }
4291
4292 if (use_ReduceInitialCardMarks()
4293 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4294 return;
4295 }
4296
4297 if (!use_precise) {
4298 // All card marks for a (non-array) instance are in one place:
4299 adr = obj;
4300 }
4301 // (Else it's an array (or unknown), and we want more precise card marks.)
4302 assert(adr != NULL, "");
4303
4304 IdealKit ideal(this, true);
4305
4306 Node* tls = __ thread(); // ThreadLocalStorage
4307
4308 Node* no_base = __ top();
4309 float likely = PROB_LIKELY(0.999);
4310 float unlikely = PROB_UNLIKELY(0.999);
4311 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val());
4312 Node* dirty_card = __ ConI((jint)CardTable::dirty_card_val());
4313 Node* zeroX = __ ConX(0);
4314
4315 // Get the alias_index for raw card-mark memory
4316 const TypePtr* card_type = TypeRawPtr::BOTTOM;
4317
4318 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4319
4320 // Offsets into the thread
4321 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4322 DirtyCardQueue::byte_offset_of_index());
4323 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4324 DirtyCardQueue::byte_offset_of_buf());
4325
4326 // Pointers into the thread
4327
4328 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4329 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4330
4331 // Now some values
4332 // Use ctrl to avoid hoisting these values past a safepoint, which could
4333 // potentially reset these fields in the JavaThread.
4334 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4335 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4336
4337 // Convert the store obj pointer to an int prior to doing math on it
4338 // Must use ctrl to prevent "integerized oop" existing across safepoint
4339 Node* cast = __ CastPX(__ ctrl(), adr);
4340
4341 // Divide pointer by card size
4342 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
4343
4344 // Combine card table base and card offset
4345 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4346
4347 // If we know the value being stored does it cross regions?
4348
4349 if (val != NULL) {
4350 // Does the store cause us to cross regions?
4351
4352 // Should be able to do an unsigned compare of region_size instead of
4353 // and extra shift. Do we have an unsigned compare??
4354 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4355 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4356
4357 // if (xor_res == 0) same region so skip
4358 __ if_then(xor_res, BoolTest::ne, zeroX); {
4359
4360 // No barrier if we are storing a NULL
4361 __ if_then(val, BoolTest::ne, null(), unlikely); {
4362
4363 // Ok must mark the card if not already dirty
4364
4365 // load the original value of the card
4366 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4367
4368 __ if_then(card_val, BoolTest::ne, young_card); {
4369 sync_kit(ideal);
4370 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4371 insert_mem_bar(Op_MemBarVolatile, oop_store);
4372 __ sync_kit(this);
4373
4374 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4375 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4376 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4377 } __ end_if();
4378 } __ end_if();
4379 } __ end_if();
4380 } __ end_if();
4381 } else {
4382 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
4383 // We don't need a barrier here if the destination is a newly allocated object
4384 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
4385 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
4386 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
4387 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4388 __ if_then(card_val, BoolTest::ne, young_card); {
4389 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4390 } __ end_if();
4391 }
4392
4393 // Final sync IdealKit and GraphKit.
4394 final_sync(ideal);
4395 }
4396 #undef __
4397
4398
4399 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4400 Node* len = load_array_length(load_String_value(ctrl, str));
4401 Node* coder = load_String_coder(ctrl, str);
4402 // Divide length by 2 if coder is UTF16
4403 return _gvn.transform(new RShiftINode(len, coder));
4404 }
4405
4406 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4407 int value_offset = java_lang_String::value_offset_in_bytes();
4408 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4409 false, NULL, 0);
4410 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4411 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4412 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4413 ciTypeArrayKlass::make(T_BYTE), true, 0);
4414 int value_field_idx = C->get_alias_index(value_field_type);
4415 Node* adr = basic_plus_adr(str, str, value_offset);
4416 Node* load = make_load(ctrl, adr,
4417 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4418 if (UseZGC) {
4419 load = load_barrier(load, adr);
4420 }
4421
4422 // String.value field is known to be @Stable.
4423 if (UseImplicitStableValues) {
4424 load = cast_array_to_stable(load, value_type);
4425 }
4426 return load;
4427 }
4428
4429 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4430 if (!CompactStrings) {
4431 return intcon(java_lang_String::CODER_UTF16);
4432 }
4433 int coder_offset = java_lang_String::coder_offset_in_bytes();
4434 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4435 false, NULL, 0);
4436 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4437 int coder_field_idx = C->get_alias_index(coder_field_type);
4438 return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4439 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4440 }
4441
4442 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4443 int value_offset = java_lang_String::value_offset_in_bytes();
4444 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4445 false, NULL, 0);
4446 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4447 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4448 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4449 }
4450
4451 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4452 int coder_offset = java_lang_String::coder_offset_in_bytes();
4453 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4454 false, NULL, 0);
4455 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4456 int coder_field_idx = C->get_alias_index(coder_field_type);
4457 store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4458 value, T_BYTE, coder_field_idx, MemNode::unordered);
4459 }
4460
4461 // Capture src and dst memory state with a MergeMemNode
4462 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4463 if (src_type == dst_type) {
4464 // Types are equal, we don't need a MergeMemNode
4465 return memory(src_type);
4466 }
4467 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4468 record_for_igvn(merge); // fold it up later, if possible
4469 int src_idx = C->get_alias_index(src_type);
4470 int dst_idx = C->get_alias_index(dst_type);
4471 merge->set_memory_at(src_idx, memory(src_idx));
4472 merge->set_memory_at(dst_idx, memory(dst_idx));
4473 return merge;
4474 }
4475
4476 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4477 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4478 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4479 // If input and output memory types differ, capture both states to preserve
4480 // the dependency between preceding and subsequent loads/stores.
4481 // For example, the following program:
4482 // StoreB
4483 // compress_string
4484 // LoadB
4485 // has this memory graph (use->def):
4486 // LoadB -> compress_string -> CharMem
4487 // ... -> StoreB -> ByteMem
4488 // The intrinsic hides the dependency between LoadB and StoreB, causing
4489 // the load to read from memory not containing the result of the StoreB.
4490 // The correct memory graph should look like this:
4491 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4492 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4493 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4494 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4495 set_memory(res_mem, TypeAryPtr::BYTES);
4496 return str;
4497 }
4498
4499 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4500 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4501 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4502 // Capture src and dst memory (see comment in 'compress_string').
4503 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4504 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4505 set_memory(_gvn.transform(str), dst_type);
4506 }
4507
4508 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4509 /**
4510 * int i_char = start;
4511 * for (int i_byte = 0; i_byte < count; i_byte++) {
4512 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4513 * }
4514 */
4515 add_predicate();
4516 RegionNode* head = new RegionNode(3);
4517 head->init_req(1, control());
4518 gvn().set_type(head, Type::CONTROL);
4519 record_for_igvn(head);
4520
4521 Node* i_byte = new PhiNode(head, TypeInt::INT);
4522 i_byte->init_req(1, intcon(0));
4523 gvn().set_type(i_byte, TypeInt::INT);
4524 record_for_igvn(i_byte);
4525
4526 Node* i_char = new PhiNode(head, TypeInt::INT);
4527 i_char->init_req(1, start);
4528 gvn().set_type(i_char, TypeInt::INT);
4529 record_for_igvn(i_char);
4530
4531 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4532 gvn().set_type(mem, Type::MEMORY);
4533 record_for_igvn(mem);
4534 set_control(head);
4535 set_memory(mem, TypeAryPtr::BYTES);
4536 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4537 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4538 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4539 false, false, true /* mismatched */);
4540
4541 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4542 head->init_req(2, IfTrue(iff));
4543 mem->init_req(2, st);
4544 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4545 i_char->init_req(2, AddI(i_char, intcon(2)));
4546
4547 set_control(IfFalse(iff));
4548 set_memory(st, TypeAryPtr::BYTES);
4549 }
4550
4551 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4552 if (!field->is_constant()) {
4553 return NULL; // Field not marked as constant.
4554 }
4555 ciInstance* holder = NULL;
4556 if (!field->is_static()) {
4557 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4558 if (const_oop != NULL && const_oop->is_instance()) {
4559 holder = const_oop->as_instance();
4560 }
4561 }
4562 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4563 /*is_unsigned_load=*/false);
4564 if (con_type != NULL) {
4565 return makecon(con_type);
4566 }
4567 return NULL;
4568 }
4569
4570 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4571 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4572 // assumption of CCP analysis.
4573 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4574 }