1 /*
2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "ci/ciValueKlass.hpp"
28 #include "gc/g1/g1SATBCardTableModRefBS.hpp"
29 #include "gc/g1/heapRegion.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/cardTableModRefBS.hpp"
32 #include "gc/shared/collectedHeap.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/idealKit.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/opaquenode.hpp"
42 #include "opto/parse.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/valuetypenode.hpp"
46 #include "runtime/deoptimization.hpp"
47 #include "runtime/sharedRuntime.hpp"
48
49 //----------------------------GraphKit-----------------------------------------
50 // Main utility constructor.
51 GraphKit::GraphKit(JVMState* jvms)
52 : Phase(Phase::Parser),
53 _env(C->env()),
54 _gvn(*C->initial_gvn())
55 {
56 _exceptions = jvms->map()->next_exception();
57 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
58 set_jvms(jvms);
59 }
60
61 // Private constructor for parser.
62 GraphKit::GraphKit()
63 : Phase(Phase::Parser),
64 _env(C->env()),
65 _gvn(*C->initial_gvn())
66 {
67 _exceptions = NULL;
68 set_map(NULL);
69 debug_only(_sp = -99);
70 debug_only(set_bci(-99));
71 }
72
73
74
75 //---------------------------clean_stack---------------------------------------
76 // Clear away rubbish from the stack area of the JVM state.
77 // This destroys any arguments that may be waiting on the stack.
78 void GraphKit::clean_stack(int from_sp) {
79 SafePointNode* map = this->map();
80 JVMState* jvms = this->jvms();
81 int stk_size = jvms->stk_size();
82 int stkoff = jvms->stkoff();
83 Node* top = this->top();
84 for (int i = from_sp; i < stk_size; i++) {
85 if (map->in(stkoff + i) != top) {
86 map->set_req(stkoff + i, top);
87 }
88 }
89 }
90
91
92 //--------------------------------sync_jvms-----------------------------------
93 // Make sure our current jvms agrees with our parse state.
94 JVMState* GraphKit::sync_jvms() const {
95 JVMState* jvms = this->jvms();
96 jvms->set_bci(bci()); // Record the new bci in the JVMState
97 jvms->set_sp(sp()); // Record the new sp in the JVMState
98 assert(jvms_in_sync(), "jvms is now in sync");
99 return jvms;
100 }
101
102 //--------------------------------sync_jvms_for_reexecute---------------------
103 // Make sure our current jvms agrees with our parse state. This version
104 // uses the reexecute_sp for reexecuting bytecodes.
105 JVMState* GraphKit::sync_jvms_for_reexecute() {
106 JVMState* jvms = this->jvms();
107 jvms->set_bci(bci()); // Record the new bci in the JVMState
108 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
109 return jvms;
110 }
111
112 #ifdef ASSERT
113 bool GraphKit::jvms_in_sync() const {
114 Parse* parse = is_Parse();
115 if (parse == NULL) {
116 if (bci() != jvms()->bci()) return false;
117 if (sp() != (int)jvms()->sp()) return false;
118 return true;
119 }
120 if (jvms()->method() != parse->method()) return false;
121 if (jvms()->bci() != parse->bci()) return false;
122 int jvms_sp = jvms()->sp();
123 if (jvms_sp != parse->sp()) return false;
124 int jvms_depth = jvms()->depth();
125 if (jvms_depth != parse->depth()) return false;
126 return true;
127 }
128
129 // Local helper checks for special internal merge points
130 // used to accumulate and merge exception states.
131 // They are marked by the region's in(0) edge being the map itself.
132 // Such merge points must never "escape" into the parser at large,
133 // until they have been handed to gvn.transform.
134 static bool is_hidden_merge(Node* reg) {
135 if (reg == NULL) return false;
136 if (reg->is_Phi()) {
137 reg = reg->in(0);
138 if (reg == NULL) return false;
139 }
140 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
141 }
142
143 void GraphKit::verify_map() const {
144 if (map() == NULL) return; // null map is OK
145 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
146 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
147 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
148 }
149
150 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
151 assert(ex_map->next_exception() == NULL, "not already part of a chain");
152 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
153 }
154 #endif
155
156 //---------------------------stop_and_kill_map---------------------------------
157 // Set _map to NULL, signalling a stop to further bytecode execution.
158 // First smash the current map's control to a constant, to mark it dead.
159 void GraphKit::stop_and_kill_map() {
160 SafePointNode* dead_map = stop();
161 if (dead_map != NULL) {
162 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
163 assert(dead_map->is_killed(), "must be so marked");
164 }
165 }
166
167
168 //--------------------------------stopped--------------------------------------
169 // Tell if _map is NULL, or control is top.
170 bool GraphKit::stopped() {
171 if (map() == NULL) return true;
172 else if (control() == top()) return true;
173 else return false;
174 }
175
176
177 //-----------------------------has_ex_handler----------------------------------
178 // Tell if this method or any caller method has exception handlers.
179 bool GraphKit::has_ex_handler() {
180 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
181 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
182 return true;
183 }
184 }
185 return false;
186 }
187
188 //------------------------------save_ex_oop------------------------------------
189 // Save an exception without blowing stack contents or other JVM state.
190 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
191 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
192 ex_map->add_req(ex_oop);
193 debug_only(verify_exception_state(ex_map));
194 }
195
196 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
197 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
198 Node* ex_oop = ex_map->in(ex_map->req()-1);
199 if (clear_it) ex_map->del_req(ex_map->req()-1);
200 return ex_oop;
201 }
202
203 //-----------------------------saved_ex_oop------------------------------------
204 // Recover a saved exception from its map.
205 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
206 return common_saved_ex_oop(ex_map, false);
207 }
208
209 //--------------------------clear_saved_ex_oop---------------------------------
210 // Erase a previously saved exception from its map.
211 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
212 return common_saved_ex_oop(ex_map, true);
213 }
214
215 #ifdef ASSERT
216 //---------------------------has_saved_ex_oop----------------------------------
217 // Erase a previously saved exception from its map.
218 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
219 return ex_map->req() == ex_map->jvms()->endoff()+1;
220 }
221 #endif
222
223 //-------------------------make_exception_state--------------------------------
224 // Turn the current JVM state into an exception state, appending the ex_oop.
225 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
226 sync_jvms();
227 SafePointNode* ex_map = stop(); // do not manipulate this map any more
228 set_saved_ex_oop(ex_map, ex_oop);
229 return ex_map;
230 }
231
232
233 //--------------------------add_exception_state--------------------------------
234 // Add an exception to my list of exceptions.
235 void GraphKit::add_exception_state(SafePointNode* ex_map) {
236 if (ex_map == NULL || ex_map->control() == top()) {
237 return;
238 }
239 #ifdef ASSERT
240 verify_exception_state(ex_map);
241 if (has_exceptions()) {
242 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
243 }
244 #endif
245
246 // If there is already an exception of exactly this type, merge with it.
247 // In particular, null-checks and other low-level exceptions common up here.
248 Node* ex_oop = saved_ex_oop(ex_map);
249 const Type* ex_type = _gvn.type(ex_oop);
250 if (ex_oop == top()) {
251 // No action needed.
252 return;
253 }
254 assert(ex_type->isa_instptr(), "exception must be an instance");
255 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
256 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
257 // We check sp also because call bytecodes can generate exceptions
258 // both before and after arguments are popped!
259 if (ex_type2 == ex_type
260 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
261 combine_exception_states(ex_map, e2);
262 return;
263 }
264 }
265
266 // No pre-existing exception of the same type. Chain it on the list.
267 push_exception_state(ex_map);
268 }
269
270 //-----------------------add_exception_states_from-----------------------------
271 void GraphKit::add_exception_states_from(JVMState* jvms) {
272 SafePointNode* ex_map = jvms->map()->next_exception();
273 if (ex_map != NULL) {
274 jvms->map()->set_next_exception(NULL);
275 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
276 next_map = ex_map->next_exception();
277 ex_map->set_next_exception(NULL);
278 add_exception_state(ex_map);
279 }
280 }
281 }
282
283 //-----------------------transfer_exceptions_into_jvms-------------------------
284 JVMState* GraphKit::transfer_exceptions_into_jvms() {
285 if (map() == NULL) {
286 // We need a JVMS to carry the exceptions, but the map has gone away.
287 // Create a scratch JVMS, cloned from any of the exception states...
288 if (has_exceptions()) {
289 _map = _exceptions;
290 _map = clone_map();
291 _map->set_next_exception(NULL);
292 clear_saved_ex_oop(_map);
293 debug_only(verify_map());
294 } else {
295 // ...or created from scratch
296 JVMState* jvms = new (C) JVMState(_method, NULL);
297 jvms->set_bci(_bci);
298 jvms->set_sp(_sp);
299 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
300 set_jvms(jvms);
301 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
302 set_all_memory(top());
303 while (map()->req() < jvms->endoff()) map()->add_req(top());
304 }
305 // (This is a kludge, in case you didn't notice.)
306 set_control(top());
307 }
308 JVMState* jvms = sync_jvms();
309 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
310 jvms->map()->set_next_exception(_exceptions);
311 _exceptions = NULL; // done with this set of exceptions
312 return jvms;
313 }
314
315 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
316 assert(is_hidden_merge(dstphi), "must be a special merge node");
317 assert(is_hidden_merge(srcphi), "must be a special merge node");
318 uint limit = srcphi->req();
319 for (uint i = PhiNode::Input; i < limit; i++) {
320 dstphi->add_req(srcphi->in(i));
321 }
322 }
323 static inline void add_one_req(Node* dstphi, Node* src) {
324 assert(is_hidden_merge(dstphi), "must be a special merge node");
325 assert(!is_hidden_merge(src), "must not be a special merge node");
326 dstphi->add_req(src);
327 }
328
329 //-----------------------combine_exception_states------------------------------
330 // This helper function combines exception states by building phis on a
331 // specially marked state-merging region. These regions and phis are
332 // untransformed, and can build up gradually. The region is marked by
333 // having a control input of its exception map, rather than NULL. Such
334 // regions do not appear except in this function, and in use_exception_state.
335 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
336 if (failing()) return; // dying anyway...
337 JVMState* ex_jvms = ex_map->_jvms;
338 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
339 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
340 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
341 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
342 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
343 assert(ex_map->req() == phi_map->req(), "matching maps");
344 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
345 Node* hidden_merge_mark = root();
346 Node* region = phi_map->control();
347 MergeMemNode* phi_mem = phi_map->merged_memory();
348 MergeMemNode* ex_mem = ex_map->merged_memory();
349 if (region->in(0) != hidden_merge_mark) {
350 // The control input is not (yet) a specially-marked region in phi_map.
351 // Make it so, and build some phis.
352 region = new RegionNode(2);
353 _gvn.set_type(region, Type::CONTROL);
354 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
355 region->init_req(1, phi_map->control());
356 phi_map->set_control(region);
357 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
358 record_for_igvn(io_phi);
359 _gvn.set_type(io_phi, Type::ABIO);
360 phi_map->set_i_o(io_phi);
361 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
362 Node* m = mms.memory();
363 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
364 record_for_igvn(m_phi);
365 _gvn.set_type(m_phi, Type::MEMORY);
366 mms.set_memory(m_phi);
367 }
368 }
369
370 // Either or both of phi_map and ex_map might already be converted into phis.
371 Node* ex_control = ex_map->control();
372 // if there is special marking on ex_map also, we add multiple edges from src
373 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
374 // how wide was the destination phi_map, originally?
375 uint orig_width = region->req();
376
377 if (add_multiple) {
378 add_n_reqs(region, ex_control);
379 add_n_reqs(phi_map->i_o(), ex_map->i_o());
380 } else {
381 // ex_map has no merges, so we just add single edges everywhere
382 add_one_req(region, ex_control);
383 add_one_req(phi_map->i_o(), ex_map->i_o());
384 }
385 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
386 if (mms.is_empty()) {
387 // get a copy of the base memory, and patch some inputs into it
388 const TypePtr* adr_type = mms.adr_type(C);
389 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
390 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
391 mms.set_memory(phi);
392 // Prepare to append interesting stuff onto the newly sliced phi:
393 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
394 }
395 // Append stuff from ex_map:
396 if (add_multiple) {
397 add_n_reqs(mms.memory(), mms.memory2());
398 } else {
399 add_one_req(mms.memory(), mms.memory2());
400 }
401 }
402 uint limit = ex_map->req();
403 for (uint i = TypeFunc::Parms; i < limit; i++) {
404 // Skip everything in the JVMS after tos. (The ex_oop follows.)
405 if (i == tos) i = ex_jvms->monoff();
406 Node* src = ex_map->in(i);
407 Node* dst = phi_map->in(i);
408 if (src != dst) {
409 PhiNode* phi;
410 if (dst->in(0) != region) {
411 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
412 record_for_igvn(phi);
413 _gvn.set_type(phi, phi->type());
414 phi_map->set_req(i, dst);
415 // Prepare to append interesting stuff onto the new phi:
416 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
417 } else {
418 assert(dst->is_Phi(), "nobody else uses a hidden region");
419 phi = dst->as_Phi();
420 }
421 if (add_multiple && src->in(0) == ex_control) {
422 // Both are phis.
423 add_n_reqs(dst, src);
424 } else {
425 while (dst->req() < region->req()) add_one_req(dst, src);
426 }
427 const Type* srctype = _gvn.type(src);
428 if (phi->type() != srctype) {
429 const Type* dsttype = phi->type()->meet_speculative(srctype);
430 if (phi->type() != dsttype) {
431 phi->set_type(dsttype);
432 _gvn.set_type(phi, dsttype);
433 }
434 }
435 }
436 }
437 phi_map->merge_replaced_nodes_with(ex_map);
438 }
439
440 //--------------------------use_exception_state--------------------------------
441 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
442 if (failing()) { stop(); return top(); }
443 Node* region = phi_map->control();
444 Node* hidden_merge_mark = root();
445 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
446 Node* ex_oop = clear_saved_ex_oop(phi_map);
447 if (region->in(0) == hidden_merge_mark) {
448 // Special marking for internal ex-states. Process the phis now.
449 region->set_req(0, region); // now it's an ordinary region
450 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
451 // Note: Setting the jvms also sets the bci and sp.
452 set_control(_gvn.transform(region));
453 uint tos = jvms()->stkoff() + sp();
454 for (uint i = 1; i < tos; i++) {
455 Node* x = phi_map->in(i);
456 if (x->in(0) == region) {
457 assert(x->is_Phi(), "expected a special phi");
458 phi_map->set_req(i, _gvn.transform(x));
459 }
460 }
461 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
462 Node* x = mms.memory();
463 if (x->in(0) == region) {
464 assert(x->is_Phi(), "nobody else uses a hidden region");
465 mms.set_memory(_gvn.transform(x));
466 }
467 }
468 if (ex_oop->in(0) == region) {
469 assert(ex_oop->is_Phi(), "expected a special phi");
470 ex_oop = _gvn.transform(ex_oop);
471 }
472 } else {
473 set_jvms(phi_map->jvms());
474 }
475
476 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
477 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
478 return ex_oop;
479 }
480
481 //---------------------------------java_bc-------------------------------------
482 Bytecodes::Code GraphKit::java_bc() const {
483 ciMethod* method = this->method();
484 int bci = this->bci();
485 if (method != NULL && bci != InvocationEntryBci)
486 return method->java_code_at_bci(bci);
487 else
488 return Bytecodes::_illegal;
489 }
490
491 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
492 bool must_throw) {
493 // if the exception capability is set, then we will generate code
494 // to check the JavaThread.should_post_on_exceptions flag to see
495 // if we actually need to report exception events (for this
496 // thread). If we don't need to report exception events, we will
497 // take the normal fast path provided by add_exception_events. If
498 // exception event reporting is enabled for this thread, we will
499 // take the uncommon_trap in the BuildCutout below.
500
501 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
502 Node* jthread = _gvn.transform(new ThreadLocalNode());
503 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
504 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
505
506 // Test the should_post_on_exceptions_flag vs. 0
507 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
508 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
509
510 // Branch to slow_path if should_post_on_exceptions_flag was true
511 { BuildCutout unless(this, tst, PROB_MAX);
512 // Do not try anything fancy if we're notifying the VM on every throw.
513 // Cf. case Bytecodes::_athrow in parse2.cpp.
514 uncommon_trap(reason, Deoptimization::Action_none,
515 (ciKlass*)NULL, (char*)NULL, must_throw);
516 }
517
518 }
519
520 //------------------------------builtin_throw----------------------------------
521 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
522 bool must_throw = true;
523
524 if (env()->jvmti_can_post_on_exceptions()) {
525 // check if we must post exception events, take uncommon trap if so
526 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
527 // here if should_post_on_exceptions is false
528 // continue on with the normal codegen
529 }
530
531 // If this particular condition has not yet happened at this
532 // bytecode, then use the uncommon trap mechanism, and allow for
533 // a future recompilation if several traps occur here.
534 // If the throw is hot, try to use a more complicated inline mechanism
535 // which keeps execution inside the compiled code.
536 bool treat_throw_as_hot = false;
537 ciMethodData* md = method()->method_data();
538
539 if (ProfileTraps) {
540 if (too_many_traps(reason)) {
541 treat_throw_as_hot = true;
542 }
543 // (If there is no MDO at all, assume it is early in
544 // execution, and that any deopts are part of the
545 // startup transient, and don't need to be remembered.)
546
547 // Also, if there is a local exception handler, treat all throws
548 // as hot if there has been at least one in this method.
549 if (C->trap_count(reason) != 0
550 && method()->method_data()->trap_count(reason) != 0
551 && has_ex_handler()) {
552 treat_throw_as_hot = true;
553 }
554 }
555
556 // If this throw happens frequently, an uncommon trap might cause
557 // a performance pothole. If there is a local exception handler,
558 // and if this particular bytecode appears to be deoptimizing often,
559 // let us handle the throw inline, with a preconstructed instance.
560 // Note: If the deopt count has blown up, the uncommon trap
561 // runtime is going to flush this nmethod, not matter what.
562 if (treat_throw_as_hot
563 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
564 // If the throw is local, we use a pre-existing instance and
565 // punt on the backtrace. This would lead to a missing backtrace
566 // (a repeat of 4292742) if the backtrace object is ever asked
567 // for its backtrace.
568 // Fixing this remaining case of 4292742 requires some flavor of
569 // escape analysis. Leave that for the future.
570 ciInstance* ex_obj = NULL;
571 switch (reason) {
572 case Deoptimization::Reason_null_check:
573 ex_obj = env()->NullPointerException_instance();
574 break;
575 case Deoptimization::Reason_div0_check:
576 ex_obj = env()->ArithmeticException_instance();
577 break;
578 case Deoptimization::Reason_range_check:
579 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
580 break;
581 case Deoptimization::Reason_class_check:
582 if (java_bc() == Bytecodes::_aastore) {
583 ex_obj = env()->ArrayStoreException_instance();
584 } else {
585 ex_obj = env()->ClassCastException_instance();
586 }
587 break;
588 }
589 if (failing()) { stop(); return; } // exception allocation might fail
590 if (ex_obj != NULL) {
591 // Cheat with a preallocated exception object.
592 if (C->log() != NULL)
593 C->log()->elem("hot_throw preallocated='1' reason='%s'",
594 Deoptimization::trap_reason_name(reason));
595 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
596 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
597
598 // Clear the detail message of the preallocated exception object.
599 // Weblogic sometimes mutates the detail message of exceptions
600 // using reflection.
601 int offset = java_lang_Throwable::get_detailMessage_offset();
602 const TypePtr* adr_typ = ex_con->add_offset(offset);
603
604 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
605 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
606 // Conservatively release stores of object references.
607 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
608
609 add_exception_state(make_exception_state(ex_node));
610 return;
611 }
612 }
613
614 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
615 // It won't be much cheaper than bailing to the interp., since we'll
616 // have to pass up all the debug-info, and the runtime will have to
617 // create the stack trace.
618
619 // Usual case: Bail to interpreter.
620 // Reserve the right to recompile if we haven't seen anything yet.
621
622 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
623 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
624 if (treat_throw_as_hot
625 && (method()->method_data()->trap_recompiled_at(bci(), m)
626 || C->too_many_traps(reason))) {
627 // We cannot afford to take more traps here. Suffer in the interpreter.
628 if (C->log() != NULL)
629 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
630 Deoptimization::trap_reason_name(reason),
631 C->trap_count(reason));
632 action = Deoptimization::Action_none;
633 }
634
635 // "must_throw" prunes the JVM state to include only the stack, if there
636 // are no local exception handlers. This should cut down on register
637 // allocation time and code size, by drastically reducing the number
638 // of in-edges on the call to the uncommon trap.
639
640 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
641 }
642
643
644 //----------------------------PreserveJVMState---------------------------------
645 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
646 debug_only(kit->verify_map());
647 _kit = kit;
648 _map = kit->map(); // preserve the map
649 _sp = kit->sp();
650 kit->set_map(clone_map ? kit->clone_map() : NULL);
651 #ifdef ASSERT
652 _bci = kit->bci();
653 Parse* parser = kit->is_Parse();
654 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
655 _block = block;
656 #endif
657 }
658 PreserveJVMState::~PreserveJVMState() {
659 GraphKit* kit = _kit;
660 #ifdef ASSERT
661 assert(kit->bci() == _bci, "bci must not shift");
662 Parse* parser = kit->is_Parse();
663 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
664 assert(block == _block, "block must not shift");
665 #endif
666 kit->set_map(_map);
667 kit->set_sp(_sp);
668 }
669
670
671 //-----------------------------BuildCutout-------------------------------------
672 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
673 : PreserveJVMState(kit)
674 {
675 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
676 SafePointNode* outer_map = _map; // preserved map is caller's
677 SafePointNode* inner_map = kit->map();
678 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
679 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
680 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
681 }
682 BuildCutout::~BuildCutout() {
683 GraphKit* kit = _kit;
684 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
685 }
686
687 //---------------------------PreserveReexecuteState----------------------------
688 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
689 assert(!kit->stopped(), "must call stopped() before");
690 _kit = kit;
691 _sp = kit->sp();
692 _reexecute = kit->jvms()->_reexecute;
693 }
694 PreserveReexecuteState::~PreserveReexecuteState() {
695 if (_kit->stopped()) return;
696 _kit->jvms()->_reexecute = _reexecute;
697 _kit->set_sp(_sp);
698 }
699
700 //------------------------------clone_map--------------------------------------
701 // Implementation of PreserveJVMState
702 //
703 // Only clone_map(...) here. If this function is only used in the
704 // PreserveJVMState class we may want to get rid of this extra
705 // function eventually and do it all there.
706
707 SafePointNode* GraphKit::clone_map() {
708 if (map() == NULL) return NULL;
709
710 // Clone the memory edge first
711 Node* mem = MergeMemNode::make(map()->memory());
712 gvn().set_type_bottom(mem);
713
714 SafePointNode *clonemap = (SafePointNode*)map()->clone();
715 JVMState* jvms = this->jvms();
716 JVMState* clonejvms = jvms->clone_shallow(C);
717 clonemap->set_memory(mem);
718 clonemap->set_jvms(clonejvms);
719 clonejvms->set_map(clonemap);
720 record_for_igvn(clonemap);
721 gvn().set_type_bottom(clonemap);
722 return clonemap;
723 }
724
725
726 //-----------------------------set_map_clone-----------------------------------
727 void GraphKit::set_map_clone(SafePointNode* m) {
728 _map = m;
729 _map = clone_map();
730 _map->set_next_exception(NULL);
731 debug_only(verify_map());
732 }
733
734
735 //----------------------------kill_dead_locals---------------------------------
736 // Detect any locals which are known to be dead, and force them to top.
737 void GraphKit::kill_dead_locals() {
738 // Consult the liveness information for the locals. If any
739 // of them are unused, then they can be replaced by top(). This
740 // should help register allocation time and cut down on the size
741 // of the deoptimization information.
742
743 // This call is made from many of the bytecode handling
744 // subroutines called from the Big Switch in do_one_bytecode.
745 // Every bytecode which might include a slow path is responsible
746 // for killing its dead locals. The more consistent we
747 // are about killing deads, the fewer useless phis will be
748 // constructed for them at various merge points.
749
750 // bci can be -1 (InvocationEntryBci). We return the entry
751 // liveness for the method.
752
753 if (method() == NULL || method()->code_size() == 0) {
754 // We are building a graph for a call to a native method.
755 // All locals are live.
756 return;
757 }
758
759 ResourceMark rm;
760
761 // Consult the liveness information for the locals. If any
762 // of them are unused, then they can be replaced by top(). This
763 // should help register allocation time and cut down on the size
764 // of the deoptimization information.
765 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
766
767 int len = (int)live_locals.size();
768 assert(len <= jvms()->loc_size(), "too many live locals");
769 for (int local = 0; local < len; local++) {
770 if (!live_locals.at(local)) {
771 set_local(local, top());
772 }
773 }
774 }
775
776 #ifdef ASSERT
777 //-------------------------dead_locals_are_killed------------------------------
778 // Return true if all dead locals are set to top in the map.
779 // Used to assert "clean" debug info at various points.
780 bool GraphKit::dead_locals_are_killed() {
781 if (method() == NULL || method()->code_size() == 0) {
782 // No locals need to be dead, so all is as it should be.
783 return true;
784 }
785
786 // Make sure somebody called kill_dead_locals upstream.
787 ResourceMark rm;
788 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
789 if (jvms->loc_size() == 0) continue; // no locals to consult
790 SafePointNode* map = jvms->map();
791 ciMethod* method = jvms->method();
792 int bci = jvms->bci();
793 if (jvms == this->jvms()) {
794 bci = this->bci(); // it might not yet be synched
795 }
796 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
797 int len = (int)live_locals.size();
798 if (!live_locals.is_valid() || len == 0)
799 // This method is trivial, or is poisoned by a breakpoint.
800 return true;
801 assert(len == jvms->loc_size(), "live map consistent with locals map");
802 for (int local = 0; local < len; local++) {
803 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
804 if (PrintMiscellaneous && (Verbose || WizardMode)) {
805 tty->print_cr("Zombie local %d: ", local);
806 jvms->dump();
807 }
808 return false;
809 }
810 }
811 }
812 return true;
813 }
814
815 #endif //ASSERT
816
817 // Helper function for enforcing certain bytecodes to reexecute if
818 // deoptimization happens
819 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
820 ciMethod* cur_method = jvms->method();
821 int cur_bci = jvms->bci();
822 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
823 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
824 return Interpreter::bytecode_should_reexecute(code) ||
825 is_anewarray && code == Bytecodes::_multianewarray;
826 // Reexecute _multianewarray bytecode which was replaced with
827 // sequence of [a]newarray. See Parse::do_multianewarray().
828 //
829 // Note: interpreter should not have it set since this optimization
830 // is limited by dimensions and guarded by flag so in some cases
831 // multianewarray() runtime calls will be generated and
832 // the bytecode should not be reexecutes (stack will not be reset).
833 } else
834 return false;
835 }
836
837 // Helper function for adding JVMState and debug information to node
838 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
839 // Add the safepoint edges to the call (or other safepoint).
840
841 // Make sure dead locals are set to top. This
842 // should help register allocation time and cut down on the size
843 // of the deoptimization information.
844 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
845
846 // Walk the inline list to fill in the correct set of JVMState's
847 // Also fill in the associated edges for each JVMState.
848
849 // If the bytecode needs to be reexecuted we need to put
850 // the arguments back on the stack.
851 const bool should_reexecute = jvms()->should_reexecute();
852 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
853
854 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
855 // undefined if the bci is different. This is normal for Parse but it
856 // should not happen for LibraryCallKit because only one bci is processed.
857 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
858 "in LibraryCallKit the reexecute bit should not change");
859
860 // If we are guaranteed to throw, we can prune everything but the
861 // input to the current bytecode.
862 bool can_prune_locals = false;
863 uint stack_slots_not_pruned = 0;
864 int inputs = 0, depth = 0;
865 if (must_throw) {
866 assert(method() == youngest_jvms->method(), "sanity");
867 if (compute_stack_effects(inputs, depth)) {
868 can_prune_locals = true;
869 stack_slots_not_pruned = inputs;
870 }
871 }
872
873 if (env()->should_retain_local_variables()) {
874 // At any safepoint, this method can get breakpointed, which would
875 // then require an immediate deoptimization.
876 can_prune_locals = false; // do not prune locals
877 stack_slots_not_pruned = 0;
878 }
879
880 // do not scribble on the input jvms
881 JVMState* out_jvms = youngest_jvms->clone_deep(C);
882 call->set_jvms(out_jvms); // Start jvms list for call node
883
884 // For a known set of bytecodes, the interpreter should reexecute them if
885 // deoptimization happens. We set the reexecute state for them here
886 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
887 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
888 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
889 }
890
891 // Presize the call:
892 DEBUG_ONLY(uint non_debug_edges = call->req());
893 call->add_req_batch(top(), youngest_jvms->debug_depth());
894 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
895
896 // Set up edges so that the call looks like this:
897 // Call [state:] ctl io mem fptr retadr
898 // [parms:] parm0 ... parmN
899 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
900 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
901 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
902 // Note that caller debug info precedes callee debug info.
903
904 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
905 uint debug_ptr = call->req();
906
907 // Loop over the map input edges associated with jvms, add them
908 // to the call node, & reset all offsets to match call node array.
909 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
910 uint debug_end = debug_ptr;
911 uint debug_start = debug_ptr - in_jvms->debug_size();
912 debug_ptr = debug_start; // back up the ptr
913
914 uint p = debug_start; // walks forward in [debug_start, debug_end)
915 uint j, k, l;
916 SafePointNode* in_map = in_jvms->map();
917 out_jvms->set_map(call);
918
919 if (can_prune_locals) {
920 assert(in_jvms->method() == out_jvms->method(), "sanity");
921 // If the current throw can reach an exception handler in this JVMS,
922 // then we must keep everything live that can reach that handler.
923 // As a quick and dirty approximation, we look for any handlers at all.
924 if (in_jvms->method()->has_exception_handlers()) {
925 can_prune_locals = false;
926 }
927 }
928
929 // Add the Locals
930 k = in_jvms->locoff();
931 l = in_jvms->loc_size();
932 out_jvms->set_locoff(p);
933 if (!can_prune_locals) {
934 for (j = 0; j < l; j++)
935 call->set_req(p++, in_map->in(k+j));
936 } else {
937 p += l; // already set to top above by add_req_batch
938 }
939
940 // Add the Expression Stack
941 k = in_jvms->stkoff();
942 l = in_jvms->sp();
943 out_jvms->set_stkoff(p);
944 if (!can_prune_locals) {
945 for (j = 0; j < l; j++)
946 call->set_req(p++, in_map->in(k+j));
947 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
948 // Divide stack into {S0,...,S1}, where S0 is set to top.
949 uint s1 = stack_slots_not_pruned;
950 stack_slots_not_pruned = 0; // for next iteration
951 if (s1 > l) s1 = l;
952 uint s0 = l - s1;
953 p += s0; // skip the tops preinstalled by add_req_batch
954 for (j = s0; j < l; j++)
955 call->set_req(p++, in_map->in(k+j));
956 } else {
957 p += l; // already set to top above by add_req_batch
958 }
959
960 // Add the Monitors
961 k = in_jvms->monoff();
962 l = in_jvms->mon_size();
963 out_jvms->set_monoff(p);
964 for (j = 0; j < l; j++)
965 call->set_req(p++, in_map->in(k+j));
966
967 // Copy any scalar object fields.
968 k = in_jvms->scloff();
969 l = in_jvms->scl_size();
970 out_jvms->set_scloff(p);
971 for (j = 0; j < l; j++)
972 call->set_req(p++, in_map->in(k+j));
973
974 // Finish the new jvms.
975 out_jvms->set_endoff(p);
976
977 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
978 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
979 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
980 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
981 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
982 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
983
984 // Update the two tail pointers in parallel.
985 out_jvms = out_jvms->caller();
986 in_jvms = in_jvms->caller();
987 }
988
989 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
990
991 // Test the correctness of JVMState::debug_xxx accessors:
992 assert(call->jvms()->debug_start() == non_debug_edges, "");
993 assert(call->jvms()->debug_end() == call->req(), "");
994 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
995 }
996
997 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
998 Bytecodes::Code code = java_bc();
999 if (code == Bytecodes::_wide) {
1000 code = method()->java_code_at_bci(bci() + 1);
1001 }
1002
1003 BasicType rtype = T_ILLEGAL;
1004 int rsize = 0;
1005
1006 if (code != Bytecodes::_illegal) {
1007 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1008 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1009 if (rtype < T_CONFLICT)
1010 rsize = type2size[rtype];
1011 }
1012
1013 switch (code) {
1014 case Bytecodes::_illegal:
1015 return false;
1016
1017 case Bytecodes::_ldc:
1018 case Bytecodes::_ldc_w:
1019 case Bytecodes::_ldc2_w:
1020 inputs = 0;
1021 break;
1022
1023 case Bytecodes::_dup: inputs = 1; break;
1024 case Bytecodes::_dup_x1: inputs = 2; break;
1025 case Bytecodes::_dup_x2: inputs = 3; break;
1026 case Bytecodes::_dup2: inputs = 2; break;
1027 case Bytecodes::_dup2_x1: inputs = 3; break;
1028 case Bytecodes::_dup2_x2: inputs = 4; break;
1029 case Bytecodes::_swap: inputs = 2; break;
1030 case Bytecodes::_arraylength: inputs = 1; break;
1031
1032 case Bytecodes::_getstatic:
1033 case Bytecodes::_putstatic:
1034 case Bytecodes::_getfield:
1035 case Bytecodes::_vgetfield:
1036 case Bytecodes::_putfield:
1037 {
1038 bool ignored_will_link;
1039 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1040 int size = field->type()->size();
1041 bool is_get = (depth >= 0), is_static = (depth & 1);
1042 inputs = (is_static ? 0 : 1);
1043 if (is_get) {
1044 depth = size - inputs;
1045 } else {
1046 inputs += size; // putxxx pops the value from the stack
1047 depth = - inputs;
1048 }
1049 }
1050 break;
1051
1052 case Bytecodes::_invokevirtual:
1053 case Bytecodes::_invokedirect:
1054 case Bytecodes::_invokespecial:
1055 case Bytecodes::_invokestatic:
1056 case Bytecodes::_invokedynamic:
1057 case Bytecodes::_invokeinterface:
1058 {
1059 bool ignored_will_link;
1060 ciSignature* declared_signature = NULL;
1061 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1062 assert(declared_signature != NULL, "cannot be null");
1063 inputs = declared_signature->arg_size_for_bc(code);
1064 int size = declared_signature->return_type()->size();
1065 depth = size - inputs;
1066 }
1067 break;
1068
1069 case Bytecodes::_multianewarray:
1070 {
1071 ciBytecodeStream iter(method());
1072 iter.reset_to_bci(bci());
1073 iter.next();
1074 inputs = iter.get_dimensions();
1075 assert(rsize == 1, "");
1076 depth = rsize - inputs;
1077 }
1078 break;
1079
1080 case Bytecodes::_vnew: {
1081 // vnew pops the values from the stack
1082 ciValueKlass* vk = method()->holder()->as_value_klass();
1083 inputs = vk->field_size();
1084 depth = rsize - inputs;
1085 break;
1086 }
1087 case Bytecodes::_vwithfield: {
1088 bool ignored_will_link;
1089 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1090 int size = field->type()->size();
1091 inputs = size+1;
1092 depth = rsize - inputs;
1093 break;
1094 }
1095
1096 case Bytecodes::_ireturn:
1097 case Bytecodes::_lreturn:
1098 case Bytecodes::_freturn:
1099 case Bytecodes::_dreturn:
1100 case Bytecodes::_areturn:
1101 case Bytecodes::_vreturn:
1102 assert(rsize = -depth, "");
1103 inputs = rsize;
1104 break;
1105
1106 case Bytecodes::_jsr:
1107 case Bytecodes::_jsr_w:
1108 inputs = 0;
1109 depth = 1; // S.B. depth=1, not zero
1110 break;
1111
1112 default:
1113 // bytecode produces a typed result
1114 inputs = rsize - depth;
1115 assert(inputs >= 0, "");
1116 break;
1117 }
1118
1119 #ifdef ASSERT
1120 // spot check
1121 int outputs = depth + inputs;
1122 assert(outputs >= 0, "sanity");
1123 switch (code) {
1124 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1125 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1126 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1127 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1128 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1129 }
1130 #endif //ASSERT
1131
1132 return true;
1133 }
1134
1135
1136
1137 //------------------------------basic_plus_adr---------------------------------
1138 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1139 // short-circuit a common case
1140 if (offset == intcon(0)) return ptr;
1141 return _gvn.transform( new AddPNode(base, ptr, offset) );
1142 }
1143
1144 Node* GraphKit::ConvI2L(Node* offset) {
1145 // short-circuit a common case
1146 jint offset_con = find_int_con(offset, Type::OffsetBot);
1147 if (offset_con != Type::OffsetBot) {
1148 return longcon((jlong) offset_con);
1149 }
1150 return _gvn.transform( new ConvI2LNode(offset));
1151 }
1152
1153 Node* GraphKit::ConvI2UL(Node* offset) {
1154 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1155 if (offset_con != (juint) Type::OffsetBot) {
1156 return longcon((julong) offset_con);
1157 }
1158 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1159 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1160 return _gvn.transform( new AndLNode(conv, mask) );
1161 }
1162
1163 Node* GraphKit::ConvL2I(Node* offset) {
1164 // short-circuit a common case
1165 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1166 if (offset_con != (jlong)Type::OffsetBot) {
1167 return intcon((int) offset_con);
1168 }
1169 return _gvn.transform( new ConvL2INode(offset));
1170 }
1171
1172 //-------------------------load_object_klass-----------------------------------
1173 Node* GraphKit::load_object_klass(Node* obj) {
1174 // Special-case a fresh allocation to avoid building nodes:
1175 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1176 if (akls != NULL) return akls;
1177 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1178 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1179 }
1180
1181 //-------------------------load_array_length-----------------------------------
1182 Node* GraphKit::load_array_length(Node* array) {
1183 // Special-case a fresh allocation to avoid building nodes:
1184 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1185 Node *alen;
1186 if (alloc == NULL) {
1187 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1188 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1189 } else {
1190 alen = alloc->Ideal_length();
1191 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1192 if (ccast != alen) {
1193 alen = _gvn.transform(ccast);
1194 }
1195 }
1196 return alen;
1197 }
1198
1199 //------------------------------do_null_check----------------------------------
1200 // Helper function to do a NULL pointer check. Returned value is
1201 // the incoming address with NULL casted away. You are allowed to use the
1202 // not-null value only if you are control dependent on the test.
1203 #ifndef PRODUCT
1204 extern int explicit_null_checks_inserted,
1205 explicit_null_checks_elided;
1206 #endif
1207 Node* GraphKit::null_check_common(Node* value, BasicType type,
1208 // optional arguments for variations:
1209 bool assert_null,
1210 Node* *null_control,
1211 bool speculative) {
1212 assert(!assert_null || null_control == NULL, "not both at once");
1213 if (stopped()) return top();
1214 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1215 // For some performance testing, we may wish to suppress null checking.
1216 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1217 return value;
1218 }
1219 NOT_PRODUCT(explicit_null_checks_inserted++);
1220
1221 // Construct NULL check
1222 Node *chk = NULL;
1223 switch(type) {
1224 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1225 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1226 case T_ARRAY : // fall through
1227 type = T_OBJECT; // simplify further tests
1228 case T_OBJECT : {
1229 const Type *t = _gvn.type( value );
1230
1231 const TypeOopPtr* tp = t->isa_oopptr();
1232 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1233 // Only for do_null_check, not any of its siblings:
1234 && !assert_null && null_control == NULL) {
1235 // Usually, any field access or invocation on an unloaded oop type
1236 // will simply fail to link, since the statically linked class is
1237 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1238 // the static class is loaded but the sharper oop type is not.
1239 // Rather than checking for this obscure case in lots of places,
1240 // we simply observe that a null check on an unloaded class
1241 // will always be followed by a nonsense operation, so we
1242 // can just issue the uncommon trap here.
1243 // Our access to the unloaded class will only be correct
1244 // after it has been loaded and initialized, which requires
1245 // a trip through the interpreter.
1246 #ifndef PRODUCT
1247 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1248 #endif
1249 uncommon_trap(Deoptimization::Reason_unloaded,
1250 Deoptimization::Action_reinterpret,
1251 tp->klass(), "!loaded");
1252 return top();
1253 }
1254
1255 if (assert_null) {
1256 // See if the type is contained in NULL_PTR.
1257 // If so, then the value is already null.
1258 if (t->higher_equal(TypePtr::NULL_PTR)) {
1259 NOT_PRODUCT(explicit_null_checks_elided++);
1260 return value; // Elided null assert quickly!
1261 }
1262 } else {
1263 // See if mixing in the NULL pointer changes type.
1264 // If so, then the NULL pointer was not allowed in the original
1265 // type. In other words, "value" was not-null.
1266 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1267 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1268 NOT_PRODUCT(explicit_null_checks_elided++);
1269 return value; // Elided null check quickly!
1270 }
1271 }
1272 chk = new CmpPNode( value, null() );
1273 break;
1274 }
1275
1276 default:
1277 fatal("unexpected type: %s", type2name(type));
1278 }
1279 assert(chk != NULL, "sanity check");
1280 chk = _gvn.transform(chk);
1281
1282 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1283 BoolNode *btst = new BoolNode( chk, btest);
1284 Node *tst = _gvn.transform( btst );
1285
1286 //-----------
1287 // if peephole optimizations occurred, a prior test existed.
1288 // If a prior test existed, maybe it dominates as we can avoid this test.
1289 if (tst != btst && type == T_OBJECT) {
1290 // At this point we want to scan up the CFG to see if we can
1291 // find an identical test (and so avoid this test altogether).
1292 Node *cfg = control();
1293 int depth = 0;
1294 while( depth < 16 ) { // Limit search depth for speed
1295 if( cfg->Opcode() == Op_IfTrue &&
1296 cfg->in(0)->in(1) == tst ) {
1297 // Found prior test. Use "cast_not_null" to construct an identical
1298 // CastPP (and hence hash to) as already exists for the prior test.
1299 // Return that casted value.
1300 if (assert_null) {
1301 replace_in_map(value, null());
1302 return null(); // do not issue the redundant test
1303 }
1304 Node *oldcontrol = control();
1305 set_control(cfg);
1306 Node *res = cast_not_null(value);
1307 set_control(oldcontrol);
1308 NOT_PRODUCT(explicit_null_checks_elided++);
1309 return res;
1310 }
1311 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1312 if (cfg == NULL) break; // Quit at region nodes
1313 depth++;
1314 }
1315 }
1316
1317 //-----------
1318 // Branch to failure if null
1319 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1320 Deoptimization::DeoptReason reason;
1321 if (assert_null) {
1322 reason = Deoptimization::Reason_null_assert;
1323 } else if (type == T_OBJECT) {
1324 reason = Deoptimization::reason_null_check(speculative);
1325 } else {
1326 reason = Deoptimization::Reason_div0_check;
1327 }
1328 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1329 // ciMethodData::has_trap_at will return a conservative -1 if any
1330 // must-be-null assertion has failed. This could cause performance
1331 // problems for a method after its first do_null_assert failure.
1332 // Consider using 'Reason_class_check' instead?
1333
1334 // To cause an implicit null check, we set the not-null probability
1335 // to the maximum (PROB_MAX). For an explicit check the probability
1336 // is set to a smaller value.
1337 if (null_control != NULL || too_many_traps(reason)) {
1338 // probability is less likely
1339 ok_prob = PROB_LIKELY_MAG(3);
1340 } else if (!assert_null &&
1341 (ImplicitNullCheckThreshold > 0) &&
1342 method() != NULL &&
1343 (method()->method_data()->trap_count(reason)
1344 >= (uint)ImplicitNullCheckThreshold)) {
1345 ok_prob = PROB_LIKELY_MAG(3);
1346 }
1347
1348 if (null_control != NULL) {
1349 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1350 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1351 set_control( _gvn.transform( new IfTrueNode(iff)));
1352 #ifndef PRODUCT
1353 if (null_true == top()) {
1354 explicit_null_checks_elided++;
1355 }
1356 #endif
1357 (*null_control) = null_true;
1358 } else {
1359 BuildCutout unless(this, tst, ok_prob);
1360 // Check for optimizer eliding test at parse time
1361 if (stopped()) {
1362 // Failure not possible; do not bother making uncommon trap.
1363 NOT_PRODUCT(explicit_null_checks_elided++);
1364 } else if (assert_null) {
1365 uncommon_trap(reason,
1366 Deoptimization::Action_make_not_entrant,
1367 NULL, "assert_null");
1368 } else {
1369 replace_in_map(value, zerocon(type));
1370 builtin_throw(reason);
1371 }
1372 }
1373
1374 // Must throw exception, fall-thru not possible?
1375 if (stopped()) {
1376 return top(); // No result
1377 }
1378
1379 if (assert_null) {
1380 // Cast obj to null on this path.
1381 replace_in_map(value, zerocon(type));
1382 return zerocon(type);
1383 }
1384
1385 // Cast obj to not-null on this path, if there is no null_control.
1386 // (If there is a null_control, a non-null value may come back to haunt us.)
1387 if (type == T_OBJECT) {
1388 Node* cast = cast_not_null(value, false);
1389 if (null_control == NULL || (*null_control) == top())
1390 replace_in_map(value, cast);
1391 value = cast;
1392 }
1393
1394 return value;
1395 }
1396
1397
1398 //------------------------------cast_not_null----------------------------------
1399 // Cast obj to not-null on this path
1400 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1401 const Type *t = _gvn.type(obj);
1402 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1403 // Object is already not-null?
1404 if( t == t_not_null ) return obj;
1405
1406 Node *cast = new CastPPNode(obj,t_not_null);
1407 cast->init_req(0, control());
1408 cast = _gvn.transform( cast );
1409
1410 // Scan for instances of 'obj' in the current JVM mapping.
1411 // These instances are known to be not-null after the test.
1412 if (do_replace_in_map)
1413 replace_in_map(obj, cast);
1414
1415 return cast; // Return casted value
1416 }
1417
1418
1419 //--------------------------replace_in_map-------------------------------------
1420 void GraphKit::replace_in_map(Node* old, Node* neww) {
1421 if (old == neww) {
1422 return;
1423 }
1424
1425 map()->replace_edge(old, neww);
1426
1427 // Note: This operation potentially replaces any edge
1428 // on the map. This includes locals, stack, and monitors
1429 // of the current (innermost) JVM state.
1430
1431 // don't let inconsistent types from profiling escape this
1432 // method
1433
1434 const Type* told = _gvn.type(old);
1435 const Type* tnew = _gvn.type(neww);
1436
1437 if (!tnew->higher_equal(told)) {
1438 return;
1439 }
1440
1441 map()->record_replaced_node(old, neww);
1442 }
1443
1444
1445 //=============================================================================
1446 //--------------------------------memory---------------------------------------
1447 Node* GraphKit::memory(uint alias_idx) {
1448 MergeMemNode* mem = merged_memory();
1449 Node* p = mem->memory_at(alias_idx);
1450 _gvn.set_type(p, Type::MEMORY); // must be mapped
1451 return p;
1452 }
1453
1454 //-----------------------------reset_memory------------------------------------
1455 Node* GraphKit::reset_memory() {
1456 Node* mem = map()->memory();
1457 // do not use this node for any more parsing!
1458 debug_only( map()->set_memory((Node*)NULL) );
1459 return _gvn.transform( mem );
1460 }
1461
1462 //------------------------------set_all_memory---------------------------------
1463 void GraphKit::set_all_memory(Node* newmem) {
1464 Node* mergemem = MergeMemNode::make(newmem);
1465 gvn().set_type_bottom(mergemem);
1466 map()->set_memory(mergemem);
1467 }
1468
1469 //------------------------------set_all_memory_call----------------------------
1470 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1471 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1472 set_all_memory(newmem);
1473 }
1474
1475 //=============================================================================
1476 //
1477 // parser factory methods for MemNodes
1478 //
1479 // These are layered on top of the factory methods in LoadNode and StoreNode,
1480 // and integrate with the parser's memory state and _gvn engine.
1481 //
1482
1483 // factory methods in "int adr_idx"
1484 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1485 int adr_idx,
1486 MemNode::MemOrd mo,
1487 LoadNode::ControlDependency control_dependency,
1488 bool require_atomic_access,
1489 bool unaligned,
1490 bool mismatched) {
1491 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1492 const TypePtr* adr_type = NULL; // debug-mode-only argument
1493 debug_only(adr_type = C->get_adr_type(adr_idx));
1494 Node* mem = memory(adr_idx);
1495 Node* ld;
1496 if (require_atomic_access && bt == T_LONG) {
1497 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1498 } else if (require_atomic_access && bt == T_DOUBLE) {
1499 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1500 } else {
1501 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1502 }
1503 ld = _gvn.transform(ld);
1504
1505 if (bt == T_VALUETYPE) {
1506 // Load value type from oop
1507 ld = ValueTypeNode::make(gvn(), map()->memory(), ld);
1508 } else if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1509 // Improve graph before escape analysis and boxing elimination.
1510 record_for_igvn(ld);
1511 }
1512 return ld;
1513 }
1514
1515 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1516 int adr_idx,
1517 MemNode::MemOrd mo,
1518 bool require_atomic_access,
1519 bool unaligned,
1520 bool mismatched) {
1521 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1522 const TypePtr* adr_type = NULL;
1523 debug_only(adr_type = C->get_adr_type(adr_idx));
1524 Node *mem = memory(adr_idx);
1525 Node* st;
1526 if (require_atomic_access && bt == T_LONG) {
1527 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1528 } else if (require_atomic_access && bt == T_DOUBLE) {
1529 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1530 } else {
1531 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1532 }
1533 if (unaligned) {
1534 st->as_Store()->set_unaligned_access();
1535 }
1536 if (mismatched) {
1537 st->as_Store()->set_mismatched_access();
1538 }
1539 st = _gvn.transform(st);
1540 set_memory(st, adr_idx);
1541 // Back-to-back stores can only remove intermediate store with DU info
1542 // so push on worklist for optimizer.
1543 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1544 record_for_igvn(st);
1545
1546 return st;
1547 }
1548
1549
1550 void GraphKit::pre_barrier(bool do_load,
1551 Node* ctl,
1552 Node* obj,
1553 Node* adr,
1554 uint adr_idx,
1555 Node* val,
1556 const TypeOopPtr* val_type,
1557 Node* pre_val,
1558 BasicType bt) {
1559
1560 BarrierSet* bs = Universe::heap()->barrier_set();
1561 set_control(ctl);
1562 switch (bs->kind()) {
1563 case BarrierSet::G1SATBCTLogging:
1564 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1565 break;
1566
1567 case BarrierSet::CardTableForRS:
1568 case BarrierSet::CardTableExtension:
1569 case BarrierSet::ModRef:
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::G1SATBCTLogging:
1582 return true; // Can move it if no safepoint
1583
1584 case BarrierSet::CardTableForRS:
1585 case BarrierSet::CardTableExtension:
1586 case BarrierSet::ModRef:
1587 return true; // There is no pre-barrier
1588
1589 default :
1590 ShouldNotReachHere();
1591 }
1592 return false;
1593 }
1594
1595 void GraphKit::post_barrier(Node* ctl,
1596 Node* store,
1597 Node* obj,
1598 Node* adr,
1599 uint adr_idx,
1600 Node* val,
1601 BasicType bt,
1602 bool use_precise) {
1603 BarrierSet* bs = Universe::heap()->barrier_set();
1604 set_control(ctl);
1605 switch (bs->kind()) {
1606 case BarrierSet::G1SATBCTLogging:
1607 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1608 break;
1609
1610 case BarrierSet::CardTableForRS:
1611 case BarrierSet::CardTableExtension:
1612 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1613 break;
1614
1615 case BarrierSet::ModRef:
1616 break;
1617
1618 default :
1619 ShouldNotReachHere();
1620
1621 }
1622 }
1623
1624 Node* GraphKit::store_oop(Node* ctl,
1625 Node* obj,
1626 Node* adr,
1627 const TypePtr* adr_type,
1628 Node* val,
1629 const TypeOopPtr* val_type,
1630 BasicType bt,
1631 bool use_precise,
1632 MemNode::MemOrd mo,
1633 bool mismatched) {
1634 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1635 // could be delayed during Parse (for example, in adjust_map_after_if()).
1636 // Execute transformation here to avoid barrier generation in such case.
1637 if (_gvn.type(val) == TypePtr::NULL_PTR)
1638 val = _gvn.makecon(TypePtr::NULL_PTR);
1639
1640 set_control(ctl);
1641 if (stopped()) return top(); // Dead path ?
1642
1643 assert(bt == T_OBJECT || bt == T_VALUETYPE, "sanity");
1644 assert(val != NULL, "not dead path");
1645 uint adr_idx = C->get_alias_index(adr_type);
1646 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1647
1648 if (bt == T_VALUETYPE) {
1649 // Allocate value type and store oop
1650 val = val->as_ValueType()->store_to_memory(this);
1651 }
1652
1653 pre_barrier(true /* do_load */,
1654 control(), obj, adr, adr_idx, val, val_type,
1655 NULL /* pre_val */,
1656 bt);
1657
1658 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1659 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1660 return store;
1661 }
1662
1663 // Could be an array or object we don't know at compile time (unsafe ref.)
1664 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1665 Node* obj, // containing obj
1666 Node* adr, // actual adress to store val at
1667 const TypePtr* adr_type,
1668 Node* val,
1669 BasicType bt,
1670 MemNode::MemOrd mo,
1671 bool mismatched) {
1672 Compile::AliasType* at = C->alias_type(adr_type);
1673 const TypeOopPtr* val_type = NULL;
1674 if (adr_type->isa_instptr()) {
1675 if (at->field() != NULL) {
1676 // known field. This code is a copy of the do_put_xxx logic.
1677 ciField* field = at->field();
1678 if (!field->type()->is_loaded()) {
1679 val_type = TypeInstPtr::BOTTOM;
1680 } else {
1681 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1682 }
1683 }
1684 } else if (adr_type->isa_aryptr()) {
1685 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1686 }
1687 if (val_type == NULL) {
1688 val_type = TypeInstPtr::BOTTOM;
1689 }
1690 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1691 }
1692
1693
1694 //-------------------------array_element_address-------------------------
1695 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1696 const TypeInt* sizetype, Node* ctrl) {
1697 uint shift = exact_log2(type2aelembytes(elembt));
1698 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1699 if (arytype_klass->is_value_array_klass()) {
1700 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass();
1701 shift = vak->log2_element_size();
1702 }
1703 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1704
1705 // short-circuit a common case (saves lots of confusing waste motion)
1706 jint idx_con = find_int_con(idx, -1);
1707 if (idx_con >= 0) {
1708 intptr_t offset = header + ((intptr_t)idx_con << shift);
1709 return basic_plus_adr(ary, offset);
1710 }
1711
1712 // must be correct type for alignment purposes
1713 Node* base = basic_plus_adr(ary, header);
1714 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1715 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1716 return basic_plus_adr(ary, base, scale);
1717 }
1718
1719 //-------------------------load_array_element-------------------------
1720 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1721 const Type* elemtype = arytype->elem();
1722 BasicType elembt = elemtype->array_element_basic_type();
1723 assert(elembt != T_VALUETYPE, "value types are not supported by this method");
1724 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1725 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1726 return ld;
1727 }
1728
1729 //-------------------------set_arguments_for_java_call-------------------------
1730 // Arguments (pre-popped from the stack) are taken from the JVMS.
1731 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1732 // Add the call arguments:
1733 uint nargs = call->method()->arg_size();
1734 for (uint i = 0, idx = 0; i < nargs; i++) {
1735 Node* arg = argument(i);
1736 if (ValueTypePassFieldsAsArgs) {
1737 if (arg->is_ValueType()) {
1738 ValueTypeNode* vt = arg->as_ValueType();
1739 // We don't pass value type arguments by reference but instead
1740 // pass each field of the value type
1741 idx += vt->set_arguments_for_java_call(call, idx + TypeFunc::Parms, *this);
1742 } else {
1743 call->init_req(idx + TypeFunc::Parms, arg);
1744 idx++;
1745 }
1746 } else {
1747 if (arg->is_ValueType()) {
1748 // Pass value type argument via oop to callee
1749 arg = arg->as_ValueType()->store_to_memory(this);
1750 }
1751 call->init_req(i + TypeFunc::Parms, arg);
1752 }
1753 }
1754 }
1755
1756 //---------------------------set_edges_for_java_call---------------------------
1757 // Connect a newly created call into the current JVMS.
1758 // A return value node (if any) is returned from set_edges_for_java_call.
1759 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1760
1761 // Add the predefined inputs:
1762 call->init_req( TypeFunc::Control, control() );
1763 call->init_req( TypeFunc::I_O , i_o() );
1764 call->init_req( TypeFunc::Memory , reset_memory() );
1765 call->init_req( TypeFunc::FramePtr, frameptr() );
1766 call->init_req( TypeFunc::ReturnAdr, top() );
1767
1768 add_safepoint_edges(call, must_throw);
1769
1770 Node* xcall = _gvn.transform(call);
1771
1772 if (xcall == top()) {
1773 set_control(top());
1774 return;
1775 }
1776 assert(xcall == call, "call identity is stable");
1777
1778 // Re-use the current map to produce the result.
1779
1780 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1781 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1782 set_all_memory_call(xcall, separate_io_proj);
1783
1784 //return xcall; // no need, caller already has it
1785 }
1786
1787 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1788 if (stopped()) return top(); // maybe the call folded up?
1789
1790 // Capture the return value, if any.
1791 Node* ret;
1792 if (call->method() == NULL ||
1793 call->method()->return_type()->basic_type() == T_VOID)
1794 ret = top();
1795 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1796
1797 // Note: Since any out-of-line call can produce an exception,
1798 // we always insert an I_O projection from the call into the result.
1799
1800 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1801
1802 if (separate_io_proj) {
1803 // The caller requested separate projections be used by the fall
1804 // through and exceptional paths, so replace the projections for
1805 // the fall through path.
1806 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1807 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1808 }
1809 return ret;
1810 }
1811
1812 //--------------------set_predefined_input_for_runtime_call--------------------
1813 // Reading and setting the memory state is way conservative here.
1814 // The real problem is that I am not doing real Type analysis on memory,
1815 // so I cannot distinguish card mark stores from other stores. Across a GC
1816 // point the Store Barrier and the card mark memory has to agree. I cannot
1817 // have a card mark store and its barrier split across the GC point from
1818 // either above or below. Here I get that to happen by reading ALL of memory.
1819 // A better answer would be to separate out card marks from other memory.
1820 // For now, return the input memory state, so that it can be reused
1821 // after the call, if this call has restricted memory effects.
1822 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1823 // Set fixed predefined input arguments
1824 Node* memory = reset_memory();
1825 call->init_req( TypeFunc::Control, control() );
1826 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1827 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1828 call->init_req( TypeFunc::FramePtr, frameptr() );
1829 call->init_req( TypeFunc::ReturnAdr, top() );
1830 return memory;
1831 }
1832
1833 //-------------------set_predefined_output_for_runtime_call--------------------
1834 // Set control and memory (not i_o) from the call.
1835 // If keep_mem is not NULL, use it for the output state,
1836 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1837 // If hook_mem is NULL, this call produces no memory effects at all.
1838 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1839 // then only that memory slice is taken from the call.
1840 // In the last case, we must put an appropriate memory barrier before
1841 // the call, so as to create the correct anti-dependencies on loads
1842 // preceding the call.
1843 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1844 Node* keep_mem,
1845 const TypePtr* hook_mem) {
1846 // no i/o
1847 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1848 if (keep_mem) {
1849 // First clone the existing memory state
1850 set_all_memory(keep_mem);
1851 if (hook_mem != NULL) {
1852 // Make memory for the call
1853 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1854 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1855 // We also use hook_mem to extract specific effects from arraycopy stubs.
1856 set_memory(mem, hook_mem);
1857 }
1858 // ...else the call has NO memory effects.
1859
1860 // Make sure the call advertises its memory effects precisely.
1861 // This lets us build accurate anti-dependences in gcm.cpp.
1862 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1863 "call node must be constructed correctly");
1864 } else {
1865 assert(hook_mem == NULL, "");
1866 // This is not a "slow path" call; all memory comes from the call.
1867 set_all_memory_call(call);
1868 }
1869 }
1870
1871
1872 // Replace the call with the current state of the kit.
1873 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1874 JVMState* ejvms = NULL;
1875 if (has_exceptions()) {
1876 ejvms = transfer_exceptions_into_jvms();
1877 }
1878
1879 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1880 ReplacedNodes replaced_nodes_exception;
1881 Node* ex_ctl = top();
1882
1883 SafePointNode* final_state = stop();
1884
1885 // Find all the needed outputs of this call
1886 CallProjections callprojs;
1887 call->extract_projections(&callprojs, true);
1888
1889 Node* init_mem = call->in(TypeFunc::Memory);
1890 Node* final_mem = final_state->in(TypeFunc::Memory);
1891 Node* final_ctl = final_state->in(TypeFunc::Control);
1892 Node* final_io = final_state->in(TypeFunc::I_O);
1893
1894 // Replace all the old call edges with the edges from the inlining result
1895 if (callprojs.fallthrough_catchproj != NULL) {
1896 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1897 }
1898 if (callprojs.fallthrough_memproj != NULL) {
1899 if (final_mem->is_MergeMem()) {
1900 // Parser's exits MergeMem was not transformed but may be optimized
1901 final_mem = _gvn.transform(final_mem);
1902 }
1903 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1904 }
1905 if (callprojs.fallthrough_ioproj != NULL) {
1906 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1907 }
1908
1909 // Replace the result with the new result if it exists and is used
1910 if (callprojs.resproj != NULL && result != NULL) {
1911 C->gvn_replace_by(callprojs.resproj, result);
1912 }
1913
1914 if (ejvms == NULL) {
1915 // No exception edges to simply kill off those paths
1916 if (callprojs.catchall_catchproj != NULL) {
1917 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1918 }
1919 if (callprojs.catchall_memproj != NULL) {
1920 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1921 }
1922 if (callprojs.catchall_ioproj != NULL) {
1923 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1924 }
1925 // Replace the old exception object with top
1926 if (callprojs.exobj != NULL) {
1927 C->gvn_replace_by(callprojs.exobj, C->top());
1928 }
1929 } else {
1930 GraphKit ekit(ejvms);
1931
1932 // Load my combined exception state into the kit, with all phis transformed:
1933 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1934 replaced_nodes_exception = ex_map->replaced_nodes();
1935
1936 Node* ex_oop = ekit.use_exception_state(ex_map);
1937
1938 if (callprojs.catchall_catchproj != NULL) {
1939 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1940 ex_ctl = ekit.control();
1941 }
1942 if (callprojs.catchall_memproj != NULL) {
1943 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1944 }
1945 if (callprojs.catchall_ioproj != NULL) {
1946 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1947 }
1948
1949 // Replace the old exception object with the newly created one
1950 if (callprojs.exobj != NULL) {
1951 C->gvn_replace_by(callprojs.exobj, ex_oop);
1952 }
1953 }
1954
1955 // Disconnect the call from the graph
1956 call->disconnect_inputs(NULL, C);
1957 C->gvn_replace_by(call, C->top());
1958
1959 // Clean up any MergeMems that feed other MergeMems since the
1960 // optimizer doesn't like that.
1961 if (final_mem->is_MergeMem()) {
1962 Node_List wl;
1963 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1964 Node* m = i.get();
1965 if (m->is_MergeMem() && !wl.contains(m)) {
1966 wl.push(m);
1967 }
1968 }
1969 while (wl.size() > 0) {
1970 _gvn.transform(wl.pop());
1971 }
1972 }
1973
1974 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1975 replaced_nodes.apply(C, final_ctl);
1976 }
1977 if (!ex_ctl->is_top() && do_replaced_nodes) {
1978 replaced_nodes_exception.apply(C, ex_ctl);
1979 }
1980 }
1981
1982
1983 //------------------------------increment_counter------------------------------
1984 // for statistics: increment a VM counter by 1
1985
1986 void GraphKit::increment_counter(address counter_addr) {
1987 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1988 increment_counter(adr1);
1989 }
1990
1991 void GraphKit::increment_counter(Node* counter_addr) {
1992 int adr_type = Compile::AliasIdxRaw;
1993 Node* ctrl = control();
1994 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1995 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1996 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1997 }
1998
1999
2000 //------------------------------uncommon_trap----------------------------------
2001 // Bail out to the interpreter in mid-method. Implemented by calling the
2002 // uncommon_trap blob. This helper function inserts a runtime call with the
2003 // right debug info.
2004 void GraphKit::uncommon_trap(int trap_request,
2005 ciKlass* klass, const char* comment,
2006 bool must_throw,
2007 bool keep_exact_action) {
2008 if (failing()) stop();
2009 if (stopped()) return; // trap reachable?
2010
2011 // Note: If ProfileTraps is true, and if a deopt. actually
2012 // occurs here, the runtime will make sure an MDO exists. There is
2013 // no need to call method()->ensure_method_data() at this point.
2014
2015 // Set the stack pointer to the right value for reexecution:
2016 set_sp(reexecute_sp());
2017
2018 #ifdef ASSERT
2019 if (!must_throw) {
2020 // Make sure the stack has at least enough depth to execute
2021 // the current bytecode.
2022 int inputs, ignored_depth;
2023 if (compute_stack_effects(inputs, ignored_depth)) {
2024 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2025 Bytecodes::name(java_bc()), sp(), inputs);
2026 }
2027 }
2028 #endif
2029
2030 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2031 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2032
2033 switch (action) {
2034 case Deoptimization::Action_maybe_recompile:
2035 case Deoptimization::Action_reinterpret:
2036 // Temporary fix for 6529811 to allow virtual calls to be sure they
2037 // get the chance to go from mono->bi->mega
2038 if (!keep_exact_action &&
2039 Deoptimization::trap_request_index(trap_request) < 0 &&
2040 too_many_recompiles(reason)) {
2041 // This BCI is causing too many recompilations.
2042 if (C->log() != NULL) {
2043 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2044 Deoptimization::trap_reason_name(reason),
2045 Deoptimization::trap_action_name(action));
2046 }
2047 action = Deoptimization::Action_none;
2048 trap_request = Deoptimization::make_trap_request(reason, action);
2049 } else {
2050 C->set_trap_can_recompile(true);
2051 }
2052 break;
2053 case Deoptimization::Action_make_not_entrant:
2054 C->set_trap_can_recompile(true);
2055 break;
2056 #ifdef ASSERT
2057 case Deoptimization::Action_none:
2058 case Deoptimization::Action_make_not_compilable:
2059 break;
2060 default:
2061 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2062 break;
2063 #endif
2064 }
2065
2066 if (TraceOptoParse) {
2067 char buf[100];
2068 tty->print_cr("Uncommon trap %s at bci:%d",
2069 Deoptimization::format_trap_request(buf, sizeof(buf),
2070 trap_request), bci());
2071 }
2072
2073 CompileLog* log = C->log();
2074 if (log != NULL) {
2075 int kid = (klass == NULL)? -1: log->identify(klass);
2076 log->begin_elem("uncommon_trap bci='%d'", bci());
2077 char buf[100];
2078 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2079 trap_request));
2080 if (kid >= 0) log->print(" klass='%d'", kid);
2081 if (comment != NULL) log->print(" comment='%s'", comment);
2082 log->end_elem();
2083 }
2084
2085 // Make sure any guarding test views this path as very unlikely
2086 Node *i0 = control()->in(0);
2087 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2088 IfNode *iff = i0->as_If();
2089 float f = iff->_prob; // Get prob
2090 if (control()->Opcode() == Op_IfTrue) {
2091 if (f > PROB_UNLIKELY_MAG(4))
2092 iff->_prob = PROB_MIN;
2093 } else {
2094 if (f < PROB_LIKELY_MAG(4))
2095 iff->_prob = PROB_MAX;
2096 }
2097 }
2098
2099 // Clear out dead values from the debug info.
2100 kill_dead_locals();
2101
2102 // Now insert the uncommon trap subroutine call
2103 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2104 const TypePtr* no_memory_effects = NULL;
2105 // Pass the index of the class to be loaded
2106 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2107 (must_throw ? RC_MUST_THROW : 0),
2108 OptoRuntime::uncommon_trap_Type(),
2109 call_addr, "uncommon_trap", no_memory_effects,
2110 intcon(trap_request));
2111 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2112 "must extract request correctly from the graph");
2113 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2114
2115 call->set_req(TypeFunc::ReturnAdr, returnadr());
2116 // The debug info is the only real input to this call.
2117
2118 // Halt-and-catch fire here. The above call should never return!
2119 HaltNode* halt = new HaltNode(control(), frameptr());
2120 _gvn.set_type_bottom(halt);
2121 root()->add_req(halt);
2122
2123 stop_and_kill_map();
2124 }
2125
2126
2127 //--------------------------just_allocated_object------------------------------
2128 // Report the object that was just allocated.
2129 // It must be the case that there are no intervening safepoints.
2130 // We use this to determine if an object is so "fresh" that
2131 // it does not require card marks.
2132 Node* GraphKit::just_allocated_object(Node* current_control) {
2133 if (C->recent_alloc_ctl() == current_control)
2134 return C->recent_alloc_obj();
2135 return NULL;
2136 }
2137
2138
2139 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2140 // (Note: TypeFunc::make has a cache that makes this fast.)
2141 const TypeFunc* tf = TypeFunc::make(dest_method);
2142 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2143 for (int j = 0; j < nargs; j++) {
2144 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2145 if( targ->basic_type() == T_DOUBLE ) {
2146 // If any parameters are doubles, they must be rounded before
2147 // the call, dstore_rounding does gvn.transform
2148 Node *arg = argument(j);
2149 arg = dstore_rounding(arg);
2150 set_argument(j, arg);
2151 }
2152 }
2153 }
2154
2155 /**
2156 * Record profiling data exact_kls for Node n with the type system so
2157 * that it can propagate it (speculation)
2158 *
2159 * @param n node that the type applies to
2160 * @param exact_kls type from profiling
2161 * @param maybe_null did profiling see null?
2162 *
2163 * @return node with improved type
2164 */
2165 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) {
2166 const Type* current_type = _gvn.type(n);
2167 assert(UseTypeSpeculation, "type speculation must be on");
2168
2169 const TypePtr* speculative = current_type->speculative();
2170
2171 // Should the klass from the profile be recorded in the speculative type?
2172 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2173 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2174 const TypeOopPtr* xtype = tklass->as_instance_type();
2175 assert(xtype->klass_is_exact(), "Should be exact");
2176 // Any reason to believe n is not null (from this profiling or a previous one)?
2177 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2178 // record the new speculative type's depth
2179 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2180 speculative = speculative->with_inline_depth(jvms()->depth());
2181 } else if (current_type->would_improve_ptr(maybe_null)) {
2182 // Profiling report that null was never seen so we can change the
2183 // speculative type to non null ptr.
2184 assert(!maybe_null, "nothing to improve");
2185 if (speculative == NULL) {
2186 speculative = TypePtr::NOTNULL;
2187 } else {
2188 const TypePtr* ptr = TypePtr::NOTNULL;
2189 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2190 }
2191 }
2192
2193 if (speculative != current_type->speculative()) {
2194 // Build a type with a speculative type (what we think we know
2195 // about the type but will need a guard when we use it)
2196 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2197 // We're changing the type, we need a new CheckCast node to carry
2198 // the new type. The new type depends on the control: what
2199 // profiling tells us is only valid from here as far as we can
2200 // tell.
2201 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2202 cast = _gvn.transform(cast);
2203 replace_in_map(n, cast);
2204 n = cast;
2205 }
2206
2207 return n;
2208 }
2209
2210 /**
2211 * Record profiling data from receiver profiling at an invoke with the
2212 * type system so that it can propagate it (speculation)
2213 *
2214 * @param n receiver node
2215 *
2216 * @return node with improved type
2217 */
2218 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2219 if (!UseTypeSpeculation) {
2220 return n;
2221 }
2222 ciKlass* exact_kls = profile_has_unique_klass();
2223 bool maybe_null = true;
2224 if (java_bc() == Bytecodes::_checkcast ||
2225 java_bc() == Bytecodes::_instanceof ||
2226 java_bc() == Bytecodes::_aastore) {
2227 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2228 bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen();
2229 }
2230 return record_profile_for_speculation(n, exact_kls, maybe_null);
2231 return n;
2232 }
2233
2234 /**
2235 * Record profiling data from argument profiling at an invoke with the
2236 * type system so that it can propagate it (speculation)
2237 *
2238 * @param dest_method target method for the call
2239 * @param bc what invoke bytecode is this?
2240 */
2241 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2242 if (!UseTypeSpeculation) {
2243 return;
2244 }
2245 const TypeFunc* tf = TypeFunc::make(dest_method);
2246 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2247 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2248 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2249 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2250 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2251 bool maybe_null = true;
2252 ciKlass* better_type = NULL;
2253 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) {
2254 record_profile_for_speculation(argument(j), better_type, maybe_null);
2255 }
2256 i++;
2257 }
2258 }
2259 }
2260
2261 /**
2262 * Record profiling data from parameter profiling at an invoke with
2263 * the type system so that it can propagate it (speculation)
2264 */
2265 void GraphKit::record_profiled_parameters_for_speculation() {
2266 if (!UseTypeSpeculation) {
2267 return;
2268 }
2269 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2270 if (_gvn.type(local(i))->isa_oopptr()) {
2271 bool maybe_null = true;
2272 ciKlass* better_type = NULL;
2273 if (method()->parameter_profiled_type(j, better_type, maybe_null)) {
2274 record_profile_for_speculation(local(i), better_type, maybe_null);
2275 }
2276 j++;
2277 }
2278 }
2279 }
2280
2281 /**
2282 * Record profiling data from return value profiling at an invoke with
2283 * the type system so that it can propagate it (speculation)
2284 */
2285 void GraphKit::record_profiled_return_for_speculation() {
2286 if (!UseTypeSpeculation) {
2287 return;
2288 }
2289 bool maybe_null = true;
2290 ciKlass* better_type = NULL;
2291 if (method()->return_profiled_type(bci(), better_type, maybe_null)) {
2292 // If profiling reports a single type for the return value,
2293 // feed it to the type system so it can propagate it as a
2294 // speculative type
2295 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null);
2296 }
2297 }
2298
2299 void GraphKit::round_double_result(ciMethod* dest_method) {
2300 // A non-strict method may return a double value which has an extended
2301 // exponent, but this must not be visible in a caller which is 'strict'
2302 // If a strict caller invokes a non-strict callee, round a double result
2303
2304 BasicType result_type = dest_method->return_type()->basic_type();
2305 assert( method() != NULL, "must have caller context");
2306 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2307 // Destination method's return value is on top of stack
2308 // dstore_rounding() does gvn.transform
2309 Node *result = pop_pair();
2310 result = dstore_rounding(result);
2311 push_pair(result);
2312 }
2313 }
2314
2315 // rounding for strict float precision conformance
2316 Node* GraphKit::precision_rounding(Node* n) {
2317 return UseStrictFP && _method->flags().is_strict()
2318 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2319 ? _gvn.transform( new RoundFloatNode(0, n) )
2320 : n;
2321 }
2322
2323 // rounding for strict double precision conformance
2324 Node* GraphKit::dprecision_rounding(Node *n) {
2325 return UseStrictFP && _method->flags().is_strict()
2326 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2327 ? _gvn.transform( new RoundDoubleNode(0, n) )
2328 : n;
2329 }
2330
2331 // rounding for non-strict double stores
2332 Node* GraphKit::dstore_rounding(Node* n) {
2333 return Matcher::strict_fp_requires_explicit_rounding
2334 && UseSSE <= 1
2335 ? _gvn.transform( new RoundDoubleNode(0, n) )
2336 : n;
2337 }
2338
2339 //=============================================================================
2340 // Generate a fast path/slow path idiom. Graph looks like:
2341 // [foo] indicates that 'foo' is a parameter
2342 //
2343 // [in] NULL
2344 // \ /
2345 // CmpP
2346 // Bool ne
2347 // If
2348 // / \
2349 // True False-<2>
2350 // / |
2351 // / cast_not_null
2352 // Load | | ^
2353 // [fast_test] | |
2354 // gvn to opt_test | |
2355 // / \ | <1>
2356 // True False |
2357 // | \\ |
2358 // [slow_call] \[fast_result]
2359 // Ctl Val \ \
2360 // | \ \
2361 // Catch <1> \ \
2362 // / \ ^ \ \
2363 // Ex No_Ex | \ \
2364 // | \ \ | \ <2> \
2365 // ... \ [slow_res] | | \ [null_result]
2366 // \ \--+--+--- | |
2367 // \ | / \ | /
2368 // --------Region Phi
2369 //
2370 //=============================================================================
2371 // Code is structured as a series of driver functions all called 'do_XXX' that
2372 // call a set of helper functions. Helper functions first, then drivers.
2373
2374 //------------------------------null_check_oop---------------------------------
2375 // Null check oop. Set null-path control into Region in slot 3.
2376 // Make a cast-not-nullness use the other not-null control. Return cast.
2377 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2378 bool never_see_null,
2379 bool safe_for_replace,
2380 bool speculative) {
2381 // Initial NULL check taken path
2382 (*null_control) = top();
2383 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2384
2385 // Generate uncommon_trap:
2386 if (never_see_null && (*null_control) != top()) {
2387 // If we see an unexpected null at a check-cast we record it and force a
2388 // recompile; the offending check-cast will be compiled to handle NULLs.
2389 // If we see more than one offending BCI, then all checkcasts in the
2390 // method will be compiled to handle NULLs.
2391 PreserveJVMState pjvms(this);
2392 set_control(*null_control);
2393 replace_in_map(value, null());
2394 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2395 uncommon_trap(reason,
2396 Deoptimization::Action_make_not_entrant);
2397 (*null_control) = top(); // NULL path is dead
2398 }
2399 if ((*null_control) == top() && safe_for_replace) {
2400 replace_in_map(value, cast);
2401 }
2402
2403 // Cast away null-ness on the result
2404 return cast;
2405 }
2406
2407 //------------------------------opt_iff----------------------------------------
2408 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2409 // Return slow-path control.
2410 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2411 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2412
2413 // Fast path taken; set region slot 2
2414 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2415 region->init_req(2,fast_taken); // Capture fast-control
2416
2417 // Fast path not-taken, i.e. slow path
2418 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2419 return slow_taken;
2420 }
2421
2422 //-----------------------------make_runtime_call-------------------------------
2423 Node* GraphKit::make_runtime_call(int flags,
2424 const TypeFunc* call_type, address call_addr,
2425 const char* call_name,
2426 const TypePtr* adr_type,
2427 // The following parms are all optional.
2428 // The first NULL ends the list.
2429 Node* parm0, Node* parm1,
2430 Node* parm2, Node* parm3,
2431 Node* parm4, Node* parm5,
2432 Node* parm6, Node* parm7) {
2433 // Slow-path call
2434 bool is_leaf = !(flags & RC_NO_LEAF);
2435 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2436 if (call_name == NULL) {
2437 assert(!is_leaf, "must supply name for leaf");
2438 call_name = OptoRuntime::stub_name(call_addr);
2439 }
2440 CallNode* call;
2441 if (!is_leaf) {
2442 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2443 bci(), adr_type);
2444 } else if (flags & RC_NO_FP) {
2445 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2446 } else {
2447 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2448 }
2449
2450 // The following is similar to set_edges_for_java_call,
2451 // except that the memory effects of the call are restricted to AliasIdxRaw.
2452
2453 // Slow path call has no side-effects, uses few values
2454 bool wide_in = !(flags & RC_NARROW_MEM);
2455 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2456
2457 Node* prev_mem = NULL;
2458 if (wide_in) {
2459 prev_mem = set_predefined_input_for_runtime_call(call);
2460 } else {
2461 assert(!wide_out, "narrow in => narrow out");
2462 Node* narrow_mem = memory(adr_type);
2463 prev_mem = reset_memory();
2464 map()->set_memory(narrow_mem);
2465 set_predefined_input_for_runtime_call(call);
2466 }
2467
2468 // Hook each parm in order. Stop looking at the first NULL.
2469 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2470 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2471 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2472 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2473 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2474 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2475 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2476 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2477 /* close each nested if ===> */ } } } } } } } }
2478 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2479
2480 if (!is_leaf) {
2481 // Non-leaves can block and take safepoints:
2482 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2483 }
2484 // Non-leaves can throw exceptions:
2485 if (has_io) {
2486 call->set_req(TypeFunc::I_O, i_o());
2487 }
2488
2489 if (flags & RC_UNCOMMON) {
2490 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2491 // (An "if" probability corresponds roughly to an unconditional count.
2492 // Sort of.)
2493 call->set_cnt(PROB_UNLIKELY_MAG(4));
2494 }
2495
2496 Node* c = _gvn.transform(call);
2497 assert(c == call, "cannot disappear");
2498
2499 if (wide_out) {
2500 // Slow path call has full side-effects.
2501 set_predefined_output_for_runtime_call(call);
2502 } else {
2503 // Slow path call has few side-effects, and/or sets few values.
2504 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2505 }
2506
2507 if (has_io) {
2508 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2509 }
2510 return call;
2511
2512 }
2513
2514 //------------------------------merge_memory-----------------------------------
2515 // Merge memory from one path into the current memory state.
2516 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2517 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2518 Node* old_slice = mms.force_memory();
2519 Node* new_slice = mms.memory2();
2520 if (old_slice != new_slice) {
2521 PhiNode* phi;
2522 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2523 if (mms.is_empty()) {
2524 // clone base memory Phi's inputs for this memory slice
2525 assert(old_slice == mms.base_memory(), "sanity");
2526 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2527 _gvn.set_type(phi, Type::MEMORY);
2528 for (uint i = 1; i < phi->req(); i++) {
2529 phi->init_req(i, old_slice->in(i));
2530 }
2531 } else {
2532 phi = old_slice->as_Phi(); // Phi was generated already
2533 }
2534 } else {
2535 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2536 _gvn.set_type(phi, Type::MEMORY);
2537 }
2538 phi->set_req(new_path, new_slice);
2539 mms.set_memory(phi);
2540 }
2541 }
2542 }
2543
2544 //------------------------------make_slow_call_ex------------------------------
2545 // Make the exception handler hookups for the slow call
2546 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2547 if (stopped()) return;
2548
2549 // Make a catch node with just two handlers: fall-through and catch-all
2550 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2551 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2552 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2553 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2554
2555 { PreserveJVMState pjvms(this);
2556 set_control(excp);
2557 set_i_o(i_o);
2558
2559 if (excp != top()) {
2560 if (deoptimize) {
2561 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2562 uncommon_trap(Deoptimization::Reason_unhandled,
2563 Deoptimization::Action_none);
2564 } else {
2565 // Create an exception state also.
2566 // Use an exact type if the caller has specified a specific exception.
2567 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2568 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2569 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2570 }
2571 }
2572 }
2573
2574 // Get the no-exception control from the CatchNode.
2575 set_control(norm);
2576 }
2577
2578 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2579 Node* cmp = NULL;
2580 switch(bt) {
2581 case T_INT: cmp = new CmpINode(in1, in2); break;
2582 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2583 default: fatal("unexpected comparison type %s", type2name(bt));
2584 }
2585 gvn->transform(cmp);
2586 Node* bol = gvn->transform(new BoolNode(cmp, test));
2587 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2588 gvn->transform(iff);
2589 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2590 return iff;
2591 }
2592
2593
2594 //-------------------------------gen_subtype_check-----------------------------
2595 // Generate a subtyping check. Takes as input the subtype and supertype.
2596 // Returns 2 values: sets the default control() to the true path and returns
2597 // the false path. Only reads invariant memory; sets no (visible) memory.
2598 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2599 // but that's not exposed to the optimizer. This call also doesn't take in an
2600 // Object; if you wish to check an Object you need to load the Object's class
2601 // prior to coming here.
2602 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2603 Compile* C = gvn->C;
2604
2605 if ((*ctrl)->is_top()) {
2606 return C->top();
2607 }
2608
2609 // Fast check for identical types, perhaps identical constants.
2610 // The types can even be identical non-constants, in cases
2611 // involving Array.newInstance, Object.clone, etc.
2612 if (subklass == superklass)
2613 return C->top(); // false path is dead; no test needed.
2614
2615 if (gvn->type(superklass)->singleton()) {
2616 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2617 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2618
2619 // In the common case of an exact superklass, try to fold up the
2620 // test before generating code. You may ask, why not just generate
2621 // the code and then let it fold up? The answer is that the generated
2622 // code will necessarily include null checks, which do not always
2623 // completely fold away. If they are also needless, then they turn
2624 // into a performance loss. Example:
2625 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2626 // Here, the type of 'fa' is often exact, so the store check
2627 // of fa[1]=x will fold up, without testing the nullness of x.
2628 switch (C->static_subtype_check(superk, subk)) {
2629 case Compile::SSC_always_false:
2630 {
2631 Node* always_fail = *ctrl;
2632 *ctrl = gvn->C->top();
2633 return always_fail;
2634 }
2635 case Compile::SSC_always_true:
2636 return C->top();
2637 case Compile::SSC_easy_test:
2638 {
2639 // Just do a direct pointer compare and be done.
2640 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2641 *ctrl = gvn->transform(new IfTrueNode(iff));
2642 return gvn->transform(new IfFalseNode(iff));
2643 }
2644 case Compile::SSC_full_test:
2645 break;
2646 default:
2647 ShouldNotReachHere();
2648 }
2649 }
2650
2651 // %%% Possible further optimization: Even if the superklass is not exact,
2652 // if the subklass is the unique subtype of the superklass, the check
2653 // will always succeed. We could leave a dependency behind to ensure this.
2654
2655 // First load the super-klass's check-offset
2656 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2657 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2658 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2659 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2660 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2661
2662 // Load from the sub-klass's super-class display list, or a 1-word cache of
2663 // the secondary superclass list, or a failing value with a sentinel offset
2664 // if the super-klass is an interface or exceptionally deep in the Java
2665 // hierarchy and we have to scan the secondary superclass list the hard way.
2666 // Worst-case type is a little odd: NULL is allowed as a result (usually
2667 // klass loads can never produce a NULL).
2668 Node *chk_off_X = chk_off;
2669 #ifdef _LP64
2670 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2671 #endif
2672 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2673 // For some types like interfaces the following loadKlass is from a 1-word
2674 // cache which is mutable so can't use immutable memory. Other
2675 // types load from the super-class display table which is immutable.
2676 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2677 Node *kmem = might_be_cache ? m : C->immutable_memory();
2678 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2679
2680 // Compile speed common case: ARE a subtype and we canNOT fail
2681 if( superklass == nkls )
2682 return C->top(); // false path is dead; no test needed.
2683
2684 // See if we get an immediate positive hit. Happens roughly 83% of the
2685 // time. Test to see if the value loaded just previously from the subklass
2686 // is exactly the superklass.
2687 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2688 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2689 *ctrl = gvn->transform(new IfFalseNode(iff1));
2690
2691 // Compile speed common case: Check for being deterministic right now. If
2692 // chk_off is a constant and not equal to cacheoff then we are NOT a
2693 // subklass. In this case we need exactly the 1 test above and we can
2694 // return those results immediately.
2695 if (!might_be_cache) {
2696 Node* not_subtype_ctrl = *ctrl;
2697 *ctrl = iftrue1; // We need exactly the 1 test above
2698 return not_subtype_ctrl;
2699 }
2700
2701 // Gather the various success & failures here
2702 RegionNode *r_ok_subtype = new RegionNode(4);
2703 gvn->record_for_igvn(r_ok_subtype);
2704 RegionNode *r_not_subtype = new RegionNode(3);
2705 gvn->record_for_igvn(r_not_subtype);
2706
2707 r_ok_subtype->init_req(1, iftrue1);
2708
2709 // Check for immediate negative hit. Happens roughly 11% of the time (which
2710 // is roughly 63% of the remaining cases). Test to see if the loaded
2711 // check-offset points into the subklass display list or the 1-element
2712 // cache. If it points to the display (and NOT the cache) and the display
2713 // missed then it's not a subtype.
2714 Node *cacheoff = gvn->intcon(cacheoff_con);
2715 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2716 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2717 *ctrl = gvn->transform(new IfFalseNode(iff2));
2718
2719 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2720 // No performance impact (too rare) but allows sharing of secondary arrays
2721 // which has some footprint reduction.
2722 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2723 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2724 *ctrl = gvn->transform(new IfFalseNode(iff3));
2725
2726 // -- Roads not taken here: --
2727 // We could also have chosen to perform the self-check at the beginning
2728 // of this code sequence, as the assembler does. This would not pay off
2729 // the same way, since the optimizer, unlike the assembler, can perform
2730 // static type analysis to fold away many successful self-checks.
2731 // Non-foldable self checks work better here in second position, because
2732 // the initial primary superclass check subsumes a self-check for most
2733 // types. An exception would be a secondary type like array-of-interface,
2734 // which does not appear in its own primary supertype display.
2735 // Finally, we could have chosen to move the self-check into the
2736 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2737 // dependent manner. But it is worthwhile to have the check here,
2738 // where it can be perhaps be optimized. The cost in code space is
2739 // small (register compare, branch).
2740
2741 // Now do a linear scan of the secondary super-klass array. Again, no real
2742 // performance impact (too rare) but it's gotta be done.
2743 // Since the code is rarely used, there is no penalty for moving it
2744 // out of line, and it can only improve I-cache density.
2745 // The decision to inline or out-of-line this final check is platform
2746 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2747 Node* psc = gvn->transform(
2748 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2749
2750 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2751 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2752 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2753
2754 // Return false path; set default control to true path.
2755 *ctrl = gvn->transform(r_ok_subtype);
2756 return gvn->transform(r_not_subtype);
2757 }
2758
2759 // Profile-driven exact type check:
2760 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2761 float prob,
2762 Node* *casted_receiver) {
2763 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2764 Node* recv_klass = load_object_klass(receiver);
2765 Node* want_klass = makecon(tklass);
2766 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2767 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2768 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2769 set_control( _gvn.transform( new IfTrueNode (iff) ));
2770 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2771
2772 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2773 assert(recv_xtype->klass_is_exact(), "");
2774
2775 // Subsume downstream occurrences of receiver with a cast to
2776 // recv_xtype, since now we know what the type will be.
2777 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2778 (*casted_receiver) = _gvn.transform(cast);
2779 // (User must make the replace_in_map call.)
2780
2781 return fail;
2782 }
2783
2784
2785 //------------------------------seems_never_null-------------------------------
2786 // Use null_seen information if it is available from the profile.
2787 // If we see an unexpected null at a type check we record it and force a
2788 // recompile; the offending check will be recompiled to handle NULLs.
2789 // If we see several offending BCIs, then all checks in the
2790 // method will be recompiled.
2791 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2792 speculating = !_gvn.type(obj)->speculative_maybe_null();
2793 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2794 if (UncommonNullCast // Cutout for this technique
2795 && obj != null() // And not the -Xcomp stupid case?
2796 && !too_many_traps(reason)
2797 ) {
2798 if (speculating) {
2799 return true;
2800 }
2801 if (data == NULL)
2802 // Edge case: no mature data. Be optimistic here.
2803 return true;
2804 // If the profile has not seen a null, assume it won't happen.
2805 assert(java_bc() == Bytecodes::_checkcast ||
2806 java_bc() == Bytecodes::_instanceof ||
2807 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2808 return !data->as_BitData()->null_seen();
2809 }
2810 speculating = false;
2811 return false;
2812 }
2813
2814 //------------------------maybe_cast_profiled_receiver-------------------------
2815 // If the profile has seen exactly one type, narrow to exactly that type.
2816 // Subsequent type checks will always fold up.
2817 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2818 ciKlass* require_klass,
2819 ciKlass* spec_klass,
2820 bool safe_for_replace) {
2821 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2822
2823 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2824
2825 // Make sure we haven't already deoptimized from this tactic.
2826 if (too_many_traps(reason) || too_many_recompiles(reason))
2827 return NULL;
2828
2829 // (No, this isn't a call, but it's enough like a virtual call
2830 // to use the same ciMethod accessor to get the profile info...)
2831 // If we have a speculative type use it instead of profiling (which
2832 // may not help us)
2833 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2834 if (exact_kls != NULL) {// no cast failures here
2835 if (require_klass == NULL ||
2836 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2837 // If we narrow the type to match what the type profile sees or
2838 // the speculative type, we can then remove the rest of the
2839 // cast.
2840 // This is a win, even if the exact_kls is very specific,
2841 // because downstream operations, such as method calls,
2842 // will often benefit from the sharper type.
2843 Node* exact_obj = not_null_obj; // will get updated in place...
2844 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2845 &exact_obj);
2846 { PreserveJVMState pjvms(this);
2847 set_control(slow_ctl);
2848 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2849 }
2850 if (safe_for_replace) {
2851 replace_in_map(not_null_obj, exact_obj);
2852 }
2853 return exact_obj;
2854 }
2855 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2856 }
2857
2858 return NULL;
2859 }
2860
2861 /**
2862 * Cast obj to type and emit guard unless we had too many traps here
2863 * already
2864 *
2865 * @param obj node being casted
2866 * @param type type to cast the node to
2867 * @param not_null true if we know node cannot be null
2868 */
2869 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2870 ciKlass* type,
2871 bool not_null) {
2872 if (stopped()) {
2873 return obj;
2874 }
2875
2876 // type == NULL if profiling tells us this object is always null
2877 if (type != NULL) {
2878 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2879 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2880
2881 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2882 !too_many_traps(class_reason) &&
2883 !too_many_recompiles(class_reason)) {
2884 Node* not_null_obj = NULL;
2885 // not_null is true if we know the object is not null and
2886 // there's no need for a null check
2887 if (!not_null) {
2888 Node* null_ctl = top();
2889 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2890 assert(null_ctl->is_top(), "no null control here");
2891 } else {
2892 not_null_obj = obj;
2893 }
2894
2895 Node* exact_obj = not_null_obj;
2896 ciKlass* exact_kls = type;
2897 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2898 &exact_obj);
2899 {
2900 PreserveJVMState pjvms(this);
2901 set_control(slow_ctl);
2902 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2903 }
2904 replace_in_map(not_null_obj, exact_obj);
2905 obj = exact_obj;
2906 }
2907 } else {
2908 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2909 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2910 Node* exact_obj = null_assert(obj);
2911 replace_in_map(obj, exact_obj);
2912 obj = exact_obj;
2913 }
2914 }
2915 return obj;
2916 }
2917
2918 //-------------------------------gen_instanceof--------------------------------
2919 // Generate an instance-of idiom. Used by both the instance-of bytecode
2920 // and the reflective instance-of call.
2921 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2922 kill_dead_locals(); // Benefit all the uncommon traps
2923 assert( !stopped(), "dead parse path should be checked in callers" );
2924 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2925 "must check for not-null not-dead klass in callers");
2926
2927 // Make the merge point
2928 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2929 RegionNode* region = new RegionNode(PATH_LIMIT);
2930 Node* phi = new PhiNode(region, TypeInt::BOOL);
2931 C->set_has_split_ifs(true); // Has chance for split-if optimization
2932
2933 ciProfileData* data = NULL;
2934 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2935 data = method()->method_data()->bci_to_data(bci());
2936 }
2937 bool speculative_not_null = false;
2938 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2939 && seems_never_null(obj, data, speculative_not_null));
2940
2941 // Null check; get casted pointer; set region slot 3
2942 Node* null_ctl = top();
2943 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2944
2945 // If not_null_obj is dead, only null-path is taken
2946 if (stopped()) { // Doing instance-of on a NULL?
2947 set_control(null_ctl);
2948 return intcon(0);
2949 }
2950 region->init_req(_null_path, null_ctl);
2951 phi ->init_req(_null_path, intcon(0)); // Set null path value
2952 if (null_ctl == top()) {
2953 // Do this eagerly, so that pattern matches like is_diamond_phi
2954 // will work even during parsing.
2955 assert(_null_path == PATH_LIMIT-1, "delete last");
2956 region->del_req(_null_path);
2957 phi ->del_req(_null_path);
2958 }
2959
2960 // Do we know the type check always succeed?
2961 bool known_statically = false;
2962 if (_gvn.type(superklass)->singleton()) {
2963 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2964 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2965 if (subk != NULL && subk->is_loaded()) {
2966 int static_res = C->static_subtype_check(superk, subk);
2967 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2968 }
2969 }
2970
2971 if (known_statically && UseTypeSpeculation) {
2972 // If we know the type check always succeeds then we don't use the
2973 // profiling data at this bytecode. Don't lose it, feed it to the
2974 // type system as a speculative type.
2975 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2976 } else {
2977 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2978 // We may not have profiling here or it may not help us. If we
2979 // have a speculative type use it to perform an exact cast.
2980 ciKlass* spec_obj_type = obj_type->speculative_type();
2981 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2982 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2983 if (stopped()) { // Profile disagrees with this path.
2984 set_control(null_ctl); // Null is the only remaining possibility.
2985 return intcon(0);
2986 }
2987 if (cast_obj != NULL) {
2988 not_null_obj = cast_obj;
2989 }
2990 }
2991 }
2992
2993 // Load the object's klass
2994 Node* obj_klass = load_object_klass(not_null_obj);
2995
2996 // Generate the subtype check
2997 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2998
2999 // Plug in the success path to the general merge in slot 1.
3000 region->init_req(_obj_path, control());
3001 phi ->init_req(_obj_path, intcon(1));
3002
3003 // Plug in the failing path to the general merge in slot 2.
3004 region->init_req(_fail_path, not_subtype_ctrl);
3005 phi ->init_req(_fail_path, intcon(0));
3006
3007 // Return final merged results
3008 set_control( _gvn.transform(region) );
3009 record_for_igvn(region);
3010 return _gvn.transform(phi);
3011 }
3012
3013 //-------------------------------gen_checkcast---------------------------------
3014 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3015 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3016 // uncommon-trap paths work. Adjust stack after this call.
3017 // If failure_control is supplied and not null, it is filled in with
3018 // the control edge for the cast failure. Otherwise, an appropriate
3019 // uncommon trap or exception is thrown.
3020 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3021 Node* *failure_control) {
3022 kill_dead_locals(); // Benefit all the uncommon traps
3023 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3024 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3025
3026 // Fast cutout: Check the case that the cast is vacuously true.
3027 // This detects the common cases where the test will short-circuit
3028 // away completely. We do this before we perform the null check,
3029 // because if the test is going to turn into zero code, we don't
3030 // want a residual null check left around. (Causes a slowdown,
3031 // for example, in some objArray manipulations, such as a[i]=a[j].)
3032 if (tk->singleton()) {
3033 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3034 if (objtp != NULL && objtp->klass() != NULL) {
3035 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3036 case Compile::SSC_always_true:
3037 // If we know the type check always succeed then we don't use
3038 // the profiling data at this bytecode. Don't lose it, feed it
3039 // to the type system as a speculative type.
3040 return record_profiled_receiver_for_speculation(obj);
3041 case Compile::SSC_always_false:
3042 // It needs a null check because a null will *pass* the cast check.
3043 // A non-null value will always produce an exception.
3044 return null_assert(obj);
3045 }
3046 }
3047 }
3048
3049 ciProfileData* data = NULL;
3050 bool safe_for_replace = false;
3051 if (failure_control == NULL) { // use MDO in regular case only
3052 assert(java_bc() == Bytecodes::_aastore ||
3053 java_bc() == Bytecodes::_checkcast,
3054 "interpreter profiles type checks only for these BCs");
3055 data = method()->method_data()->bci_to_data(bci());
3056 safe_for_replace = true;
3057 }
3058
3059 // Make the merge point
3060 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3061 RegionNode* region = new RegionNode(PATH_LIMIT);
3062 Node* phi = new PhiNode(region, toop);
3063 C->set_has_split_ifs(true); // Has chance for split-if optimization
3064
3065 // Use null-cast information if it is available
3066 bool speculative_not_null = false;
3067 bool never_see_null = ((failure_control == NULL) // regular case only
3068 && seems_never_null(obj, data, speculative_not_null));
3069
3070 // Null check; get casted pointer; set region slot 3
3071 Node* null_ctl = top();
3072 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3073
3074 // If not_null_obj is dead, only null-path is taken
3075 if (stopped()) { // Doing instance-of on a NULL?
3076 set_control(null_ctl);
3077 return null();
3078 }
3079 region->init_req(_null_path, null_ctl);
3080 phi ->init_req(_null_path, null()); // Set null path value
3081 if (null_ctl == top()) {
3082 // Do this eagerly, so that pattern matches like is_diamond_phi
3083 // will work even during parsing.
3084 assert(_null_path == PATH_LIMIT-1, "delete last");
3085 region->del_req(_null_path);
3086 phi ->del_req(_null_path);
3087 }
3088
3089 Node* cast_obj = NULL;
3090 if (tk->klass_is_exact()) {
3091 // The following optimization tries to statically cast the speculative type of the object
3092 // (for example obtained during profiling) to the type of the superklass and then do a
3093 // dynamic check that the type of the object is what we expect. To work correctly
3094 // for checkcast and aastore the type of superklass should be exact.
3095 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3096 // We may not have profiling here or it may not help us. If we have
3097 // a speculative type use it to perform an exact cast.
3098 ciKlass* spec_obj_type = obj_type->speculative_type();
3099 if (spec_obj_type != NULL || data != NULL) {
3100 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3101 if (cast_obj != NULL) {
3102 if (failure_control != NULL) // failure is now impossible
3103 (*failure_control) = top();
3104 // adjust the type of the phi to the exact klass:
3105 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3106 }
3107 }
3108 }
3109
3110 if (cast_obj == NULL) {
3111 // Load the object's klass
3112 Node* obj_klass = load_object_klass(not_null_obj);
3113
3114 // Generate the subtype check
3115 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3116
3117 // Plug in success path into the merge
3118 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3119 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3120 if (failure_control == NULL) {
3121 if (not_subtype_ctrl != top()) { // If failure is possible
3122 PreserveJVMState pjvms(this);
3123 set_control(not_subtype_ctrl);
3124 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3125 }
3126 } else {
3127 (*failure_control) = not_subtype_ctrl;
3128 }
3129 }
3130
3131 region->init_req(_obj_path, control());
3132 phi ->init_req(_obj_path, cast_obj);
3133
3134 // A merge of NULL or Casted-NotNull obj
3135 Node* res = _gvn.transform(phi);
3136
3137 // Note I do NOT always 'replace_in_map(obj,result)' here.
3138 // if( tk->klass()->can_be_primary_super() )
3139 // This means that if I successfully store an Object into an array-of-String
3140 // I 'forget' that the Object is really now known to be a String. I have to
3141 // do this because we don't have true union types for interfaces - if I store
3142 // a Baz into an array-of-Interface and then tell the optimizer it's an
3143 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3144 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3145 // replace_in_map( obj, res );
3146
3147 // Return final merged results
3148 set_control( _gvn.transform(region) );
3149 record_for_igvn(region);
3150 return res;
3151 }
3152
3153 //------------------------------next_monitor-----------------------------------
3154 // What number should be given to the next monitor?
3155 int GraphKit::next_monitor() {
3156 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3157 int next = current + C->sync_stack_slots();
3158 // Keep the toplevel high water mark current:
3159 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3160 return current;
3161 }
3162
3163 //------------------------------insert_mem_bar---------------------------------
3164 // Memory barrier to avoid floating things around
3165 // The membar serves as a pinch point between both control and all memory slices.
3166 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3167 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3168 mb->init_req(TypeFunc::Control, control());
3169 mb->init_req(TypeFunc::Memory, reset_memory());
3170 Node* membar = _gvn.transform(mb);
3171 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3172 set_all_memory_call(membar);
3173 return membar;
3174 }
3175
3176 //-------------------------insert_mem_bar_volatile----------------------------
3177 // Memory barrier to avoid floating things around
3178 // The membar serves as a pinch point between both control and memory(alias_idx).
3179 // If you want to make a pinch point on all memory slices, do not use this
3180 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3181 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3182 // When Parse::do_put_xxx updates a volatile field, it appends a series
3183 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3184 // The first membar is on the same memory slice as the field store opcode.
3185 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3186 // All the other membars (for other volatile slices, including AliasIdxBot,
3187 // which stands for all unknown volatile slices) are control-dependent
3188 // on the first membar. This prevents later volatile loads or stores
3189 // from sliding up past the just-emitted store.
3190
3191 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3192 mb->set_req(TypeFunc::Control,control());
3193 if (alias_idx == Compile::AliasIdxBot) {
3194 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3195 } else {
3196 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3197 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3198 }
3199 Node* membar = _gvn.transform(mb);
3200 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3201 if (alias_idx == Compile::AliasIdxBot) {
3202 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3203 } else {
3204 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3205 }
3206 return membar;
3207 }
3208
3209 void GraphKit::insert_store_load_for_barrier() {
3210 Node* mem = reset_memory();
3211 MemBarNode* mb = MemBarNode::make(C, Op_MemBarVolatile, Compile::AliasIdxBot);
3212 mb->init_req(TypeFunc::Control, control());
3213 mb->init_req(TypeFunc::Memory, mem);
3214 Node* membar = _gvn.transform(mb);
3215 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3216 Node* newmem = _gvn.transform(new ProjNode(membar, TypeFunc::Memory));
3217 set_all_memory(mem);
3218 set_memory(newmem, Compile::AliasIdxRaw);
3219 }
3220
3221
3222 //------------------------------shared_lock------------------------------------
3223 // Emit locking code.
3224 FastLockNode* GraphKit::shared_lock(Node* obj) {
3225 // bci is either a monitorenter bc or InvocationEntryBci
3226 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3227 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3228
3229 if( !GenerateSynchronizationCode )
3230 return NULL; // Not locking things?
3231 if (stopped()) // Dead monitor?
3232 return NULL;
3233
3234 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3235
3236 // Box the stack location
3237 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3238 Node* mem = reset_memory();
3239
3240 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3241 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3242 // Create the counters for this fast lock.
3243 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3244 }
3245
3246 // Create the rtm counters for this fast lock if needed.
3247 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3248
3249 // Add monitor to debug info for the slow path. If we block inside the
3250 // slow path and de-opt, we need the monitor hanging around
3251 map()->push_monitor( flock );
3252
3253 const TypeFunc *tf = LockNode::lock_type();
3254 LockNode *lock = new LockNode(C, tf);
3255
3256 lock->init_req( TypeFunc::Control, control() );
3257 lock->init_req( TypeFunc::Memory , mem );
3258 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3259 lock->init_req( TypeFunc::FramePtr, frameptr() );
3260 lock->init_req( TypeFunc::ReturnAdr, top() );
3261
3262 lock->init_req(TypeFunc::Parms + 0, obj);
3263 lock->init_req(TypeFunc::Parms + 1, box);
3264 lock->init_req(TypeFunc::Parms + 2, flock);
3265 add_safepoint_edges(lock);
3266
3267 lock = _gvn.transform( lock )->as_Lock();
3268
3269 // lock has no side-effects, sets few values
3270 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3271
3272 insert_mem_bar(Op_MemBarAcquireLock);
3273
3274 // Add this to the worklist so that the lock can be eliminated
3275 record_for_igvn(lock);
3276
3277 #ifndef PRODUCT
3278 if (PrintLockStatistics) {
3279 // Update the counter for this lock. Don't bother using an atomic
3280 // operation since we don't require absolute accuracy.
3281 lock->create_lock_counter(map()->jvms());
3282 increment_counter(lock->counter()->addr());
3283 }
3284 #endif
3285
3286 return flock;
3287 }
3288
3289
3290 //------------------------------shared_unlock----------------------------------
3291 // Emit unlocking code.
3292 void GraphKit::shared_unlock(Node* box, Node* obj) {
3293 // bci is either a monitorenter bc or InvocationEntryBci
3294 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3295 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3296
3297 if( !GenerateSynchronizationCode )
3298 return;
3299 if (stopped()) { // Dead monitor?
3300 map()->pop_monitor(); // Kill monitor from debug info
3301 return;
3302 }
3303
3304 // Memory barrier to avoid floating things down past the locked region
3305 insert_mem_bar(Op_MemBarReleaseLock);
3306
3307 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3308 UnlockNode *unlock = new UnlockNode(C, tf);
3309 #ifdef ASSERT
3310 unlock->set_dbg_jvms(sync_jvms());
3311 #endif
3312 uint raw_idx = Compile::AliasIdxRaw;
3313 unlock->init_req( TypeFunc::Control, control() );
3314 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3315 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3316 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3317 unlock->init_req( TypeFunc::ReturnAdr, top() );
3318
3319 unlock->init_req(TypeFunc::Parms + 0, obj);
3320 unlock->init_req(TypeFunc::Parms + 1, box);
3321 unlock = _gvn.transform(unlock)->as_Unlock();
3322
3323 Node* mem = reset_memory();
3324
3325 // unlock has no side-effects, sets few values
3326 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3327
3328 // Kill monitor from debug info
3329 map()->pop_monitor( );
3330 }
3331
3332 //-------------------------------get_layout_helper-----------------------------
3333 // If the given klass is a constant or known to be an array,
3334 // fetch the constant layout helper value into constant_value
3335 // and return (Node*)NULL. Otherwise, load the non-constant
3336 // layout helper value, and return the node which represents it.
3337 // This two-faced routine is useful because allocation sites
3338 // almost always feature constant types.
3339 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3340 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3341 if (!StressReflectiveCode && inst_klass != NULL) {
3342 ciKlass* klass = inst_klass->klass();
3343 bool xklass = inst_klass->klass_is_exact();
3344 if (xklass || klass->is_array_klass()) {
3345 jint lhelper = klass->layout_helper();
3346 if (lhelper != Klass::_lh_neutral_value) {
3347 constant_value = lhelper;
3348 return (Node*) NULL;
3349 }
3350 }
3351 }
3352 constant_value = Klass::_lh_neutral_value; // put in a known value
3353 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3354 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3355 }
3356
3357 // We just put in an allocate/initialize with a big raw-memory effect.
3358 // Hook selected additional alias categories on the initialization.
3359 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3360 MergeMemNode* init_in_merge,
3361 Node* init_out_raw) {
3362 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3363 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3364
3365 Node* prevmem = kit.memory(alias_idx);
3366 init_in_merge->set_memory_at(alias_idx, prevmem);
3367 kit.set_memory(init_out_raw, alias_idx);
3368 }
3369
3370 //---------------------------set_output_for_allocation-------------------------
3371 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3372 const TypeOopPtr* oop_type,
3373 bool deoptimize_on_exception) {
3374 int rawidx = Compile::AliasIdxRaw;
3375 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3376 add_safepoint_edges(alloc);
3377 Node* allocx = _gvn.transform(alloc);
3378 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3379 // create memory projection for i_o
3380 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3381 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3382
3383 // create a memory projection as for the normal control path
3384 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3385 set_memory(malloc, rawidx);
3386
3387 // a normal slow-call doesn't change i_o, but an allocation does
3388 // we create a separate i_o projection for the normal control path
3389 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3390 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3391
3392 // put in an initialization barrier
3393 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3394 rawoop)->as_Initialize();
3395 assert(alloc->initialization() == init, "2-way macro link must work");
3396 assert(init ->allocation() == alloc, "2-way macro link must work");
3397 {
3398 // Extract memory strands which may participate in the new object's
3399 // initialization, and source them from the new InitializeNode.
3400 // This will allow us to observe initializations when they occur,
3401 // and link them properly (as a group) to the InitializeNode.
3402 assert(init->in(InitializeNode::Memory) == malloc, "");
3403 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3404 init->set_req(InitializeNode::Memory, minit_in);
3405 record_for_igvn(minit_in); // fold it up later, if possible
3406 Node* minit_out = memory(rawidx);
3407 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3408 if (oop_type->isa_aryptr()) {
3409 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3410 int elemidx = C->get_alias_index(telemref);
3411 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3412 } else if (oop_type->isa_instptr() || oop_type->isa_valuetypeptr()) {
3413 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3414 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3415 ciField* field = ik->nonstatic_field_at(i);
3416 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3417 continue; // do not bother to track really large numbers of fields
3418 // Find (or create) the alias category for this field:
3419 int fieldidx = C->alias_type(field)->index();
3420 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3421 }
3422 }
3423 }
3424
3425 // Cast raw oop to the real thing...
3426 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3427 javaoop = _gvn.transform(javaoop);
3428 C->set_recent_alloc(control(), javaoop);
3429 assert(just_allocated_object(control()) == javaoop, "just allocated");
3430
3431 #ifdef ASSERT
3432 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3433 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3434 "Ideal_allocation works");
3435 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3436 "Ideal_allocation works");
3437 if (alloc->is_AllocateArray()) {
3438 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3439 "Ideal_allocation works");
3440 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3441 "Ideal_allocation works");
3442 } else {
3443 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3444 }
3445 }
3446 #endif //ASSERT
3447
3448 return javaoop;
3449 }
3450
3451 //---------------------------new_instance--------------------------------------
3452 // This routine takes a klass_node which may be constant (for a static type)
3453 // or may be non-constant (for reflective code). It will work equally well
3454 // for either, and the graph will fold nicely if the optimizer later reduces
3455 // the type to a constant.
3456 // The optional arguments are for specialized use by intrinsics:
3457 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3458 // - If 'return_size_val', report the the total object size to the caller.
3459 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3460 Node* GraphKit::new_instance(Node* klass_node,
3461 Node* extra_slow_test,
3462 Node* *return_size_val,
3463 bool deoptimize_on_exception) {
3464 // Compute size in doublewords
3465 // The size is always an integral number of doublewords, represented
3466 // as a positive bytewise size stored in the klass's layout_helper.
3467 // The layout_helper also encodes (in a low bit) the need for a slow path.
3468 jint layout_con = Klass::_lh_neutral_value;
3469 Node* layout_val = get_layout_helper(klass_node, layout_con);
3470 int layout_is_con = (layout_val == NULL);
3471
3472 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3473 // Generate the initial go-slow test. It's either ALWAYS (return a
3474 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3475 // case) a computed value derived from the layout_helper.
3476 Node* initial_slow_test = NULL;
3477 if (layout_is_con) {
3478 assert(!StressReflectiveCode, "stress mode does not use these paths");
3479 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3480 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3481 } else { // reflective case
3482 // This reflective path is used by Unsafe.allocateInstance.
3483 // (It may be stress-tested by specifying StressReflectiveCode.)
3484 // Basically, we want to get into the VM is there's an illegal argument.
3485 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3486 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3487 if (extra_slow_test != intcon(0)) {
3488 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3489 }
3490 // (Macro-expander will further convert this to a Bool, if necessary.)
3491 }
3492
3493 // Find the size in bytes. This is easy; it's the layout_helper.
3494 // The size value must be valid even if the slow path is taken.
3495 Node* size = NULL;
3496 if (layout_is_con) {
3497 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3498 } else { // reflective case
3499 // This reflective path is used by clone and Unsafe.allocateInstance.
3500 size = ConvI2X(layout_val);
3501
3502 // Clear the low bits to extract layout_helper_size_in_bytes:
3503 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3504 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3505 size = _gvn.transform( new AndXNode(size, mask) );
3506 }
3507 if (return_size_val != NULL) {
3508 (*return_size_val) = size;
3509 }
3510
3511 // This is a precise notnull oop of the klass.
3512 // (Actually, it need not be precise if this is a reflective allocation.)
3513 // It's what we cast the result to.
3514 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3515 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3516 const TypeOopPtr* oop_type = tklass->as_instance_type();
3517
3518 // Now generate allocation code
3519
3520 // The entire memory state is needed for slow path of the allocation
3521 // since GC and deoptimization can happen.
3522 Node *mem = reset_memory();
3523 set_all_memory(mem); // Create new memory state
3524
3525 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3526 control(), mem, i_o(),
3527 size, klass_node,
3528 initial_slow_test);
3529
3530 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3531 }
3532
3533 //-------------------------------new_array-------------------------------------
3534 // helper for newarray, anewarray and vnewarray
3535 // The 'length' parameter is (obviously) the length of the array.
3536 // See comments on new_instance for the meaning of the other arguments.
3537 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3538 Node* length, // number of array elements
3539 int nargs, // number of arguments to push back for uncommon trap
3540 Node* *return_size_val,
3541 bool deoptimize_on_exception) {
3542 jint layout_con = Klass::_lh_neutral_value;
3543 Node* layout_val = get_layout_helper(klass_node, layout_con);
3544 int layout_is_con = (layout_val == NULL);
3545
3546 if (!layout_is_con && !StressReflectiveCode &&
3547 !too_many_traps(Deoptimization::Reason_class_check)) {
3548 // This is a reflective array creation site.
3549 // Optimistically assume that it is a subtype of Object[],
3550 // so that we can fold up all the address arithmetic.
3551 layout_con = Klass::array_layout_helper(T_OBJECT);
3552 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3553 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3554 { BuildCutout unless(this, bol_lh, PROB_MAX);
3555 inc_sp(nargs);
3556 uncommon_trap(Deoptimization::Reason_class_check,
3557 Deoptimization::Action_maybe_recompile);
3558 }
3559 layout_val = NULL;
3560 layout_is_con = true;
3561 }
3562
3563 // Generate the initial go-slow test. Make sure we do not overflow
3564 // if length is huge (near 2Gig) or negative! We do not need
3565 // exact double-words here, just a close approximation of needed
3566 // double-words. We can't add any offset or rounding bits, lest we
3567 // take a size -1 of bytes and make it positive. Use an unsigned
3568 // compare, so negative sizes look hugely positive.
3569 int fast_size_limit = FastAllocateSizeLimit;
3570 if (layout_is_con) {
3571 assert(!StressReflectiveCode, "stress mode does not use these paths");
3572 // Increase the size limit if we have exact knowledge of array type.
3573 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3574 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3575 }
3576
3577 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3578 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3579
3580 // --- Size Computation ---
3581 // array_size = round_to_heap(array_header + (length << elem_shift));
3582 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3583 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3584 // The rounding mask is strength-reduced, if possible.
3585 int round_mask = MinObjAlignmentInBytes - 1;
3586 Node* header_size = NULL;
3587 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3588 // (T_BYTE has the weakest alignment and size restrictions...)
3589 if (layout_is_con) {
3590 int hsize = Klass::layout_helper_header_size(layout_con);
3591 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3592 BasicType etype = Klass::layout_helper_element_type(layout_con);
3593 if ((round_mask & ~right_n_bits(eshift)) == 0)
3594 round_mask = 0; // strength-reduce it if it goes away completely
3595 // TODO re-enabled assert
3596 // assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3597 assert(header_size_min <= hsize, "generic minimum is smallest");
3598 header_size_min = hsize;
3599 header_size = intcon(hsize + round_mask);
3600 } else {
3601 Node* hss = intcon(Klass::_lh_header_size_shift);
3602 Node* hsm = intcon(Klass::_lh_header_size_mask);
3603 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3604 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3605 Node* mask = intcon(round_mask);
3606 header_size = _gvn.transform( new AddINode(hsize, mask) );
3607 }
3608
3609 Node* elem_shift = NULL;
3610 if (layout_is_con) {
3611 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3612 if (eshift != 0)
3613 elem_shift = intcon(eshift);
3614 } else {
3615 // There is no need to mask or shift this value.
3616 // The semantics of LShiftINode include an implicit mask to 0x1F.
3617 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3618 elem_shift = layout_val;
3619 }
3620
3621 // Transition to native address size for all offset calculations:
3622 Node* lengthx = ConvI2X(length);
3623 Node* headerx = ConvI2X(header_size);
3624 #ifdef _LP64
3625 { const TypeInt* tilen = _gvn.find_int_type(length);
3626 if (tilen != NULL && tilen->_lo < 0) {
3627 // Add a manual constraint to a positive range. Cf. array_element_address.
3628 jint size_max = fast_size_limit;
3629 if (size_max > tilen->_hi) size_max = tilen->_hi;
3630 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3631
3632 // Only do a narrow I2L conversion if the range check passed.
3633 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3634 _gvn.transform(iff);
3635 RegionNode* region = new RegionNode(3);
3636 _gvn.set_type(region, Type::CONTROL);
3637 lengthx = new PhiNode(region, TypeLong::LONG);
3638 _gvn.set_type(lengthx, TypeLong::LONG);
3639
3640 // Range check passed. Use ConvI2L node with narrow type.
3641 Node* passed = IfFalse(iff);
3642 region->init_req(1, passed);
3643 // Make I2L conversion control dependent to prevent it from
3644 // floating above the range check during loop optimizations.
3645 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3646
3647 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3648 region->init_req(2, IfTrue(iff));
3649 lengthx->init_req(2, ConvI2X(length));
3650
3651 set_control(region);
3652 record_for_igvn(region);
3653 record_for_igvn(lengthx);
3654 }
3655 }
3656 #endif
3657
3658 // Combine header size (plus rounding) and body size. Then round down.
3659 // This computation cannot overflow, because it is used only in two
3660 // places, one where the length is sharply limited, and the other
3661 // after a successful allocation.
3662 Node* abody = lengthx;
3663 if (elem_shift != NULL)
3664 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3665 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3666 if (round_mask != 0) {
3667 Node* mask = MakeConX(~round_mask);
3668 size = _gvn.transform( new AndXNode(size, mask) );
3669 }
3670 // else if round_mask == 0, the size computation is self-rounding
3671
3672 if (return_size_val != NULL) {
3673 // This is the size
3674 (*return_size_val) = size;
3675 }
3676
3677 // Now generate allocation code
3678
3679 // The entire memory state is needed for slow path of the allocation
3680 // since GC and deoptimization can happen.
3681 Node *mem = reset_memory();
3682 set_all_memory(mem); // Create new memory state
3683
3684 if (initial_slow_test->is_Bool()) {
3685 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3686 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3687 }
3688
3689 // Create the AllocateArrayNode and its result projections
3690 AllocateArrayNode* alloc
3691 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3692 control(), mem, i_o(),
3693 size, klass_node,
3694 initial_slow_test,
3695 length);
3696
3697 // Cast to correct type. Note that the klass_node may be constant or not,
3698 // and in the latter case the actual array type will be inexact also.
3699 // (This happens via a non-constant argument to inline_native_newArray.)
3700 // In any case, the value of klass_node provides the desired array type.
3701 const TypeInt* length_type = _gvn.find_int_type(length);
3702 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3703 if (ary_type->isa_aryptr() && length_type != NULL) {
3704 // Try to get a better type than POS for the size
3705 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3706 }
3707
3708 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3709
3710 // Cast length on remaining path to be as narrow as possible
3711 if (map()->find_edge(length) >= 0) {
3712 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3713 if (ccast != length) {
3714 _gvn.set_type_bottom(ccast);
3715 record_for_igvn(ccast);
3716 replace_in_map(length, ccast);
3717 }
3718 }
3719
3720 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
3721 ciKlass* elem_klass = ary_ptr != NULL ? ary_ptr->klass()->as_array_klass()->element_klass() : NULL;
3722 //if (layout_is_con && Klass::layout_helper_element_type(layout_con) == T_VALUETYPE) {
3723 if (elem_klass != NULL && elem_klass->is_valuetype()) {
3724 ciValueKlass* vk = elem_klass->as_value_klass();
3725 if (vk->flatten_array()) {
3726 // TODO
3727 } else {
3728 // TODO explain this and add asserts
3729 initialize_value_type_array(javaoop, length, elem_klass->as_value_klass(), nargs);
3730 InitializeNode* init = alloc->initialization();
3731 init->set_complete_with_arraycopy();
3732 }
3733 }
3734
3735 return javaoop;
3736 }
3737
3738 void GraphKit::initialize_value_type_array(Node* array, Node* length, ciValueKlass* vk, int nargs) {
3739 // Check for zero length
3740 Node* null_ctl = top();
3741 null_check_common(length, T_INT, false, &null_ctl, false);
3742 if (stopped()) {
3743 set_control(null_ctl); // Always zero
3744 return;
3745 }
3746
3747 // Prepare for merging control and IO
3748 RegionNode* res_ctl = new RegionNode(3);
3749 res_ctl->init_req(1, null_ctl);
3750 gvn().set_type(res_ctl, Type::CONTROL);
3751 record_for_igvn(res_ctl);
3752 Node* res_io = PhiNode::make(res_ctl, i_o(), Type::ABIO);
3753 gvn().set_type(res_io, Type::ABIO);
3754 record_for_igvn(res_io);
3755
3756 // TODO comment
3757 SafePointNode* loop_map = NULL;
3758 {
3759 PreserveJVMState pjvms(this);
3760 // Create default value type and store it to memory
3761 Node* oop = ValueTypeNode::make_default(gvn(), vk);
3762 oop = oop->as_ValueType()->store_to_memory(this);
3763
3764 length = SubI(length, intcon(1));
3765 add_predicate(nargs);
3766 RegionNode* loop = new RegionNode(3);
3767 loop->init_req(1, control());
3768 gvn().set_type(loop, Type::CONTROL);
3769 record_for_igvn(loop);
3770
3771 Node* index = new PhiNode(loop, TypeInt::INT);
3772 index->init_req(1, intcon(0));
3773 gvn().set_type(index, TypeInt::INT);
3774 record_for_igvn(index);
3775
3776 // TODO explain why we need to capture all memory
3777 PhiNode* mem = new PhiNode(loop, Type::MEMORY, TypePtr::BOTTOM);
3778 mem->init_req(1, reset_memory());
3779 gvn().set_type(mem, Type::MEMORY);
3780 record_for_igvn(mem);
3781 set_control(loop);
3782 set_all_memory(mem);
3783 // Initialize array element
3784 Node* adr = array_element_address(array, index, T_OBJECT);
3785 const TypeOopPtr* elemtype = TypeValueTypePtr::make(TypePtr::NotNull, vk); // ary_type->is_aryptr()->elem()->make_oopptr();
3786 Node* store = store_oop_to_array(control(), array, adr, TypeAryPtr::OOPS, oop, elemtype, T_OBJECT, MemNode::release);
3787
3788 IfNode* iff = create_and_map_if(control(), Bool(CmpI(index, length), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
3789 loop->init_req(2, IfTrue(iff));
3790 mem->init_req(2, merged_memory());
3791 index->init_req(2, AddI(index, intcon(1)));
3792
3793 res_ctl->init_req(2, IfFalse(iff));
3794 res_io->set_req(2, i_o());
3795 loop_map = stop();
3796 }
3797 // Set merged control, IO and memory
3798 set_control(res_ctl);
3799 set_i_o(res_io);
3800 merge_memory(loop_map->merged_memory(), res_ctl, 2);
3801 }
3802
3803 // The following "Ideal_foo" functions are placed here because they recognize
3804 // the graph shapes created by the functions immediately above.
3805
3806 //---------------------------Ideal_allocation----------------------------------
3807 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3808 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3809 if (ptr == NULL) { // reduce dumb test in callers
3810 return NULL;
3811 }
3812 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3813 ptr = ptr->in(1);
3814 if (ptr == NULL) return NULL;
3815 }
3816 // Return NULL for allocations with several casts:
3817 // j.l.reflect.Array.newInstance(jobject, jint)
3818 // Object.clone()
3819 // to keep more precise type from last cast.
3820 if (ptr->is_Proj()) {
3821 Node* allo = ptr->in(0);
3822 if (allo != NULL && allo->is_Allocate()) {
3823 return allo->as_Allocate();
3824 }
3825 }
3826 // Report failure to match.
3827 return NULL;
3828 }
3829
3830 // Fancy version which also strips off an offset (and reports it to caller).
3831 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3832 intptr_t& offset) {
3833 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3834 if (base == NULL) return NULL;
3835 return Ideal_allocation(base, phase);
3836 }
3837
3838 // Trace Initialize <- Proj[Parm] <- Allocate
3839 AllocateNode* InitializeNode::allocation() {
3840 Node* rawoop = in(InitializeNode::RawAddress);
3841 if (rawoop->is_Proj()) {
3842 Node* alloc = rawoop->in(0);
3843 if (alloc->is_Allocate()) {
3844 return alloc->as_Allocate();
3845 }
3846 }
3847 return NULL;
3848 }
3849
3850 // Trace Allocate -> Proj[Parm] -> Initialize
3851 InitializeNode* AllocateNode::initialization() {
3852 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3853 if (rawoop == NULL) return NULL;
3854 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3855 Node* init = rawoop->fast_out(i);
3856 if (init->is_Initialize()) {
3857 assert(init->as_Initialize()->allocation() == this, "2-way link");
3858 return init->as_Initialize();
3859 }
3860 }
3861 return NULL;
3862 }
3863
3864 //----------------------------- loop predicates ---------------------------
3865
3866 //------------------------------add_predicate_impl----------------------------
3867 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3868 // Too many traps seen?
3869 if (too_many_traps(reason)) {
3870 #ifdef ASSERT
3871 if (TraceLoopPredicate) {
3872 int tc = C->trap_count(reason);
3873 tty->print("too many traps=%s tcount=%d in ",
3874 Deoptimization::trap_reason_name(reason), tc);
3875 method()->print(); // which method has too many predicate traps
3876 tty->cr();
3877 }
3878 #endif
3879 // We cannot afford to take more traps here,
3880 // do not generate predicate.
3881 return;
3882 }
3883
3884 Node *cont = _gvn.intcon(1);
3885 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
3886 Node *bol = _gvn.transform(new Conv2BNode(opq));
3887 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3888 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3889 C->add_predicate_opaq(opq);
3890 {
3891 PreserveJVMState pjvms(this);
3892 set_control(iffalse);
3893 inc_sp(nargs);
3894 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3895 }
3896 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3897 set_control(iftrue);
3898 }
3899
3900 //------------------------------add_predicate---------------------------------
3901 void GraphKit::add_predicate(int nargs) {
3902 if (UseLoopPredicate) {
3903 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3904 }
3905 // loop's limit check predicate should be near the loop.
3906 if (LoopLimitCheck) {
3907 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3908 }
3909 }
3910
3911 //----------------------------- store barriers ----------------------------
3912 #define __ ideal.
3913
3914 void GraphKit::sync_kit(IdealKit& ideal) {
3915 set_all_memory(__ merged_memory());
3916 set_i_o(__ i_o());
3917 set_control(__ ctrl());
3918 }
3919
3920 void GraphKit::final_sync(IdealKit& ideal) {
3921 // Final sync IdealKit and graphKit.
3922 sync_kit(ideal);
3923 }
3924
3925 Node* GraphKit::byte_map_base_node() {
3926 // Get base of card map
3927 CardTableModRefBS* ct =
3928 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
3929 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code");
3930 if (ct->byte_map_base != NULL) {
3931 return makecon(TypeRawPtr::make((address)ct->byte_map_base));
3932 } else {
3933 return null();
3934 }
3935 }
3936
3937 // vanilla/CMS post barrier
3938 // Insert a write-barrier store. This is to let generational GC work; we have
3939 // to flag all oop-stores before the next GC point.
3940 void GraphKit::write_barrier_post(Node* oop_store,
3941 Node* obj,
3942 Node* adr,
3943 uint adr_idx,
3944 Node* val,
3945 bool use_precise) {
3946 // No store check needed if we're storing a NULL or an old object
3947 // (latter case is probably a string constant). The concurrent
3948 // mark sweep garbage collector, however, needs to have all nonNull
3949 // oop updates flagged via card-marks.
3950 if (val != NULL && val->is_Con()) {
3951 // must be either an oop or NULL
3952 const Type* t = val->bottom_type();
3953 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3954 // stores of null never (?) need barriers
3955 return;
3956 }
3957
3958 if (use_ReduceInitialCardMarks()
3959 && obj == just_allocated_object(control())) {
3960 // We can skip marks on a freshly-allocated object in Eden.
3961 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3962 // That routine informs GC to take appropriate compensating steps,
3963 // upon a slow-path allocation, so as to make this card-mark
3964 // elision safe.
3965 return;
3966 }
3967
3968 if (!use_precise) {
3969 // All card marks for a (non-array) instance are in one place:
3970 adr = obj;
3971 }
3972 // (Else it's an array (or unknown), and we want more precise card marks.)
3973 assert(adr != NULL, "");
3974
3975 IdealKit ideal(this, true);
3976
3977 // Convert the pointer to an int prior to doing math on it
3978 Node* cast = __ CastPX(__ ctrl(), adr);
3979
3980 // Divide by card size
3981 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef),
3982 "Only one we handle so far.");
3983 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3984
3985 // Combine card table base and card offset
3986 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3987
3988 // Get the alias_index for raw card-mark memory
3989 int adr_type = Compile::AliasIdxRaw;
3990 Node* zero = __ ConI(0); // Dirty card value
3991 BasicType bt = T_BYTE;
3992
3993 if (UseConcMarkSweepGC && UseCondCardMark) {
3994 insert_store_load_for_barrier();
3995 __ sync_kit(this);
3996 }
3997
3998 if (UseCondCardMark) {
3999 // The classic GC reference write barrier is typically implemented
4000 // as a store into the global card mark table. Unfortunately
4001 // unconditional stores can result in false sharing and excessive
4002 // coherence traffic as well as false transactional aborts.
4003 // UseCondCardMark enables MP "polite" conditional card mark
4004 // stores. In theory we could relax the load from ctrl() to
4005 // no_ctrl, but that doesn't buy much latitude.
4006 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
4007 __ if_then(card_val, BoolTest::ne, zero);
4008 }
4009
4010 // Smash zero into card
4011 if( !UseConcMarkSweepGC ) {
4012 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
4013 } else {
4014 // Specialized path for CM store barrier
4015 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
4016 }
4017
4018 if (UseCondCardMark) {
4019 __ end_if();
4020 }
4021
4022 // Final sync IdealKit and GraphKit.
4023 final_sync(ideal);
4024 }
4025 /*
4026 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
4027 * required by SATB to make sure all objects live at the start of the
4028 * marking are kept alive, all reference updates need to any previous
4029 * reference stored before writing.
4030 *
4031 * If the previous value is NULL there is no need to save the old value.
4032 * References that are NULL are filtered during runtime by the barrier
4033 * code to avoid unnecessary queuing.
4034 *
4035 * However in the case of newly allocated objects it might be possible to
4036 * prove that the reference about to be overwritten is NULL during compile
4037 * time and avoid adding the barrier code completely.
4038 *
4039 * The compiler needs to determine that the object in which a field is about
4040 * to be written is newly allocated, and that no prior store to the same field
4041 * has happened since the allocation.
4042 *
4043 * Returns true if the pre-barrier can be removed
4044 */
4045 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
4046 BasicType bt, uint adr_idx) {
4047 intptr_t offset = 0;
4048 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4049 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4050
4051 if (offset == Type::OffsetBot) {
4052 return false; // cannot unalias unless there are precise offsets
4053 }
4054
4055 if (alloc == NULL) {
4056 return false; // No allocation found
4057 }
4058
4059 intptr_t size_in_bytes = type2aelembytes(bt);
4060
4061 Node* mem = memory(adr_idx); // start searching here...
4062
4063 for (int cnt = 0; cnt < 50; cnt++) {
4064
4065 if (mem->is_Store()) {
4066
4067 Node* st_adr = mem->in(MemNode::Address);
4068 intptr_t st_offset = 0;
4069 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
4070
4071 if (st_base == NULL) {
4072 break; // inscrutable pointer
4073 }
4074
4075 // Break we have found a store with same base and offset as ours so break
4076 if (st_base == base && st_offset == offset) {
4077 break;
4078 }
4079
4080 if (st_offset != offset && st_offset != Type::OffsetBot) {
4081 const int MAX_STORE = BytesPerLong;
4082 if (st_offset >= offset + size_in_bytes ||
4083 st_offset <= offset - MAX_STORE ||
4084 st_offset <= offset - mem->as_Store()->memory_size()) {
4085 // Success: The offsets are provably independent.
4086 // (You may ask, why not just test st_offset != offset and be done?
4087 // The answer is that stores of different sizes can co-exist
4088 // in the same sequence of RawMem effects. We sometimes initialize
4089 // a whole 'tile' of array elements with a single jint or jlong.)
4090 mem = mem->in(MemNode::Memory);
4091 continue; // advance through independent store memory
4092 }
4093 }
4094
4095 if (st_base != base
4096 && MemNode::detect_ptr_independence(base, alloc, st_base,
4097 AllocateNode::Ideal_allocation(st_base, phase),
4098 phase)) {
4099 // Success: The bases are provably independent.
4100 mem = mem->in(MemNode::Memory);
4101 continue; // advance through independent store memory
4102 }
4103 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4104
4105 InitializeNode* st_init = mem->in(0)->as_Initialize();
4106 AllocateNode* st_alloc = st_init->allocation();
4107
4108 // Make sure that we are looking at the same allocation site.
4109 // The alloc variable is guaranteed to not be null here from earlier check.
4110 if (alloc == st_alloc) {
4111 // Check that the initialization is storing NULL so that no previous store
4112 // has been moved up and directly write a reference
4113 Node* captured_store = st_init->find_captured_store(offset,
4114 type2aelembytes(T_OBJECT),
4115 phase);
4116 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4117 return true;
4118 }
4119 }
4120 }
4121
4122 // Unless there is an explicit 'continue', we must bail out here,
4123 // because 'mem' is an inscrutable memory state (e.g., a call).
4124 break;
4125 }
4126
4127 return false;
4128 }
4129
4130 // G1 pre/post barriers
4131 void GraphKit::g1_write_barrier_pre(bool do_load,
4132 Node* obj,
4133 Node* adr,
4134 uint alias_idx,
4135 Node* val,
4136 const TypeOopPtr* val_type,
4137 Node* pre_val,
4138 BasicType bt) {
4139
4140 // Some sanity checks
4141 // Note: val is unused in this routine.
4142
4143 if (do_load) {
4144 // We need to generate the load of the previous value
4145 assert(obj != NULL, "must have a base");
4146 assert(adr != NULL, "where are loading from?");
4147 assert(pre_val == NULL, "loaded already?");
4148 assert(val_type != NULL, "need a type");
4149
4150 if (use_ReduceInitialCardMarks()
4151 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4152 return;
4153 }
4154
4155 } else {
4156 // In this case both val_type and alias_idx are unused.
4157 assert(pre_val != NULL, "must be loaded already");
4158 // Nothing to be done if pre_val is null.
4159 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4160 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4161 }
4162 assert(bt == T_OBJECT || bt == T_VALUETYPE, "or we shouldn't be here");
4163
4164 IdealKit ideal(this, true);
4165
4166 Node* tls = __ thread(); // ThreadLocalStorage
4167
4168 Node* no_ctrl = NULL;
4169 Node* no_base = __ top();
4170 Node* zero = __ ConI(0);
4171 Node* zeroX = __ ConX(0);
4172
4173 float likely = PROB_LIKELY(0.999);
4174 float unlikely = PROB_UNLIKELY(0.999);
4175
4176 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4177 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4178
4179 // Offsets into the thread
4180 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
4181 SATBMarkQueue::byte_offset_of_active());
4182 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
4183 SATBMarkQueue::byte_offset_of_index());
4184 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
4185 SATBMarkQueue::byte_offset_of_buf());
4186
4187 // Now the actual pointers into the thread
4188 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4189 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4190 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4191
4192 // Now some of the values
4193 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4194
4195 // if (!marking)
4196 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4197 BasicType index_bt = TypeX_X->basic_type();
4198 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4199 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4200
4201 if (do_load) {
4202 // load original value
4203 // alias_idx correct??
4204 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4205 }
4206
4207 // if (pre_val != NULL)
4208 __ if_then(pre_val, BoolTest::ne, null()); {
4209 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4210
4211 // is the queue for this thread full?
4212 __ if_then(index, BoolTest::ne, zeroX, likely); {
4213
4214 // decrement the index
4215 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4216
4217 // Now get the buffer location we will log the previous value into and store it
4218 Node *log_addr = __ AddP(no_base, buffer, next_index);
4219 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4220 // update the index
4221 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4222
4223 } __ else_(); {
4224
4225 // logging buffer is full, call the runtime
4226 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4227 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4228 } __ end_if(); // (!index)
4229 } __ end_if(); // (pre_val != NULL)
4230 } __ end_if(); // (!marking)
4231
4232 // Final sync IdealKit and GraphKit.
4233 final_sync(ideal);
4234 }
4235
4236 /*
4237 * G1 similar to any GC with a Young Generation requires a way to keep track of
4238 * references from Old Generation to Young Generation to make sure all live
4239 * objects are found. G1 also requires to keep track of object references
4240 * between different regions to enable evacuation of old regions, which is done
4241 * as part of mixed collections. References are tracked in remembered sets and
4242 * is continuously updated as reference are written to with the help of the
4243 * post-barrier.
4244 *
4245 * To reduce the number of updates to the remembered set the post-barrier
4246 * filters updates to fields in objects located in the Young Generation,
4247 * the same region as the reference, when the NULL is being written or
4248 * if the card is already marked as dirty by an earlier write.
4249 *
4250 * Under certain circumstances it is possible to avoid generating the
4251 * post-barrier completely if it is possible during compile time to prove
4252 * the object is newly allocated and that no safepoint exists between the
4253 * allocation and the store.
4254 *
4255 * In the case of slow allocation the allocation code must handle the barrier
4256 * as part of the allocation in the case the allocated object is not located
4257 * in the nursery, this would happen for humongous objects. This is similar to
4258 * how CMS is required to handle this case, see the comments for the method
4259 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier.
4260 * A deferred card mark is required for these objects and handled in the above
4261 * mentioned methods.
4262 *
4263 * Returns true if the post barrier can be removed
4264 */
4265 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4266 Node* adr) {
4267 intptr_t offset = 0;
4268 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4269 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4270
4271 if (offset == Type::OffsetBot) {
4272 return false; // cannot unalias unless there are precise offsets
4273 }
4274
4275 if (alloc == NULL) {
4276 return false; // No allocation found
4277 }
4278
4279 // Start search from Store node
4280 Node* mem = store->in(MemNode::Control);
4281 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4282
4283 InitializeNode* st_init = mem->in(0)->as_Initialize();
4284 AllocateNode* st_alloc = st_init->allocation();
4285
4286 // Make sure we are looking at the same allocation
4287 if (alloc == st_alloc) {
4288 return true;
4289 }
4290 }
4291
4292 return false;
4293 }
4294
4295 //
4296 // Update the card table and add card address to the queue
4297 //
4298 void GraphKit::g1_mark_card(IdealKit& ideal,
4299 Node* card_adr,
4300 Node* oop_store,
4301 uint oop_alias_idx,
4302 Node* index,
4303 Node* index_adr,
4304 Node* buffer,
4305 const TypeFunc* tf) {
4306
4307 Node* zero = __ ConI(0);
4308 Node* zeroX = __ ConX(0);
4309 Node* no_base = __ top();
4310 BasicType card_bt = T_BYTE;
4311 // Smash zero into card. MUST BE ORDERED WRT TO STORE
4312 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4313
4314 // Now do the queue work
4315 __ if_then(index, BoolTest::ne, zeroX); {
4316
4317 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4318 Node* log_addr = __ AddP(no_base, buffer, next_index);
4319
4320 // Order, see storeCM.
4321 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4322 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4323
4324 } __ else_(); {
4325 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4326 } __ end_if();
4327
4328 }
4329
4330 void GraphKit::g1_write_barrier_post(Node* oop_store,
4331 Node* obj,
4332 Node* adr,
4333 uint alias_idx,
4334 Node* val,
4335 BasicType bt,
4336 bool use_precise) {
4337 // If we are writing a NULL then we need no post barrier
4338
4339 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4340 // Must be NULL
4341 const Type* t = val->bottom_type();
4342 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4343 // No post barrier if writing NULLx
4344 return;
4345 }
4346
4347 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4348 // We can skip marks on a freshly-allocated object in Eden.
4349 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
4350 // That routine informs GC to take appropriate compensating steps,
4351 // upon a slow-path allocation, so as to make this card-mark
4352 // elision safe.
4353 return;
4354 }
4355
4356 if (use_ReduceInitialCardMarks()
4357 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4358 return;
4359 }
4360
4361 if (!use_precise) {
4362 // All card marks for a (non-array) instance are in one place:
4363 adr = obj;
4364 }
4365 // (Else it's an array (or unknown), and we want more precise card marks.)
4366 assert(adr != NULL, "");
4367
4368 IdealKit ideal(this, true);
4369
4370 Node* tls = __ thread(); // ThreadLocalStorage
4371
4372 Node* no_base = __ top();
4373 float likely = PROB_LIKELY(0.999);
4374 float unlikely = PROB_UNLIKELY(0.999);
4375 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
4376 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
4377 Node* zeroX = __ ConX(0);
4378
4379 // Get the alias_index for raw card-mark memory
4380 const TypePtr* card_type = TypeRawPtr::BOTTOM;
4381
4382 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4383
4384 // Offsets into the thread
4385 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4386 DirtyCardQueue::byte_offset_of_index());
4387 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4388 DirtyCardQueue::byte_offset_of_buf());
4389
4390 // Pointers into the thread
4391
4392 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4393 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4394
4395 // Now some values
4396 // Use ctrl to avoid hoisting these values past a safepoint, which could
4397 // potentially reset these fields in the JavaThread.
4398 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4399 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4400
4401 // Convert the store obj pointer to an int prior to doing math on it
4402 // Must use ctrl to prevent "integerized oop" existing across safepoint
4403 Node* cast = __ CastPX(__ ctrl(), adr);
4404
4405 // Divide pointer by card size
4406 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4407
4408 // Combine card table base and card offset
4409 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4410
4411 // If we know the value being stored does it cross regions?
4412
4413 if (val != NULL) {
4414 // Does the store cause us to cross regions?
4415
4416 // Should be able to do an unsigned compare of region_size instead of
4417 // and extra shift. Do we have an unsigned compare??
4418 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4419 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4420
4421 // if (xor_res == 0) same region so skip
4422 __ if_then(xor_res, BoolTest::ne, zeroX); {
4423
4424 // No barrier if we are storing a NULL
4425 __ if_then(val, BoolTest::ne, null(), unlikely); {
4426
4427 // Ok must mark the card if not already dirty
4428
4429 // load the original value of the card
4430 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4431
4432 __ if_then(card_val, BoolTest::ne, young_card); {
4433 sync_kit(ideal);
4434 insert_store_load_for_barrier();
4435 __ sync_kit(this);
4436
4437 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4438 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4439 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4440 } __ end_if();
4441 } __ end_if();
4442 } __ end_if();
4443 } __ end_if();
4444 } else {
4445 // Object.clone() instrinsic uses this path.
4446 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4447 }
4448
4449 // Final sync IdealKit and GraphKit.
4450 final_sync(ideal);
4451 }
4452 #undef __
4453
4454
4455 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4456 Node* len = load_array_length(load_String_value(ctrl, str));
4457 Node* coder = load_String_coder(ctrl, str);
4458 // Divide length by 2 if coder is UTF16
4459 return _gvn.transform(new RShiftINode(len, coder));
4460 }
4461
4462 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4463 int value_offset = java_lang_String::value_offset_in_bytes();
4464 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4465 false, NULL, Type::Offset(0));
4466 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4467 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4468 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4469 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4470 int value_field_idx = C->get_alias_index(value_field_type);
4471 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4472 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4473 // String.value field is known to be @Stable.
4474 if (UseImplicitStableValues) {
4475 load = cast_array_to_stable(load, value_type);
4476 }
4477 return load;
4478 }
4479
4480 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4481 if (java_lang_String::has_coder_field()) {
4482 if (!CompactStrings) {
4483 return intcon(java_lang_String::CODER_UTF16);
4484 }
4485 int coder_offset = java_lang_String::coder_offset_in_bytes();
4486 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4487 false, NULL, Type::Offset(0));
4488 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4489 int coder_field_idx = C->get_alias_index(coder_field_type);
4490 return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4491 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4492 } else {
4493 return intcon(0); // false
4494 }
4495 }
4496
4497 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4498 int value_offset = java_lang_String::value_offset_in_bytes();
4499 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4500 false, NULL, Type::Offset(0));
4501 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4502 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4503 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4504 }
4505
4506 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4507 int coder_offset = java_lang_String::coder_offset_in_bytes();
4508 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4509 false, NULL, Type::Offset(0));
4510 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4511 int coder_field_idx = C->get_alias_index(coder_field_type);
4512 store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4513 value, T_BYTE, coder_field_idx, MemNode::unordered);
4514 }
4515
4516 // Capture src and dst memory state with a MergeMemNode
4517 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4518 if (src_type == dst_type) {
4519 // Types are equal, we don't need a MergeMemNode
4520 return memory(src_type);
4521 }
4522 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4523 record_for_igvn(merge); // fold it up later, if possible
4524 int src_idx = C->get_alias_index(src_type);
4525 int dst_idx = C->get_alias_index(dst_type);
4526 merge->set_memory_at(src_idx, memory(src_idx));
4527 merge->set_memory_at(dst_idx, memory(dst_idx));
4528 return merge;
4529 }
4530
4531 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4532 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4533 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4534 // If input and output memory types differ, capture both states to preserve
4535 // the dependency between preceding and subsequent loads/stores.
4536 // For example, the following program:
4537 // StoreB
4538 // compress_string
4539 // LoadB
4540 // has this memory graph (use->def):
4541 // LoadB -> compress_string -> CharMem
4542 // ... -> StoreB -> ByteMem
4543 // The intrinsic hides the dependency between LoadB and StoreB, causing
4544 // the load to read from memory not containing the result of the StoreB.
4545 // The correct memory graph should look like this:
4546 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4547 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4548 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4549 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4550 set_memory(res_mem, TypeAryPtr::BYTES);
4551 return str;
4552 }
4553
4554 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4555 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4556 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4557 // Capture src and dst memory (see comment in 'compress_string').
4558 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4559 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4560 set_memory(_gvn.transform(str), dst_type);
4561 }
4562
4563 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4564 /**
4565 * int i_char = start;
4566 * for (int i_byte = 0; i_byte < count; i_byte++) {
4567 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4568 * }
4569 */
4570 add_predicate();
4571 RegionNode* head = new RegionNode(3);
4572 head->init_req(1, control());
4573 gvn().set_type(head, Type::CONTROL);
4574 record_for_igvn(head);
4575
4576 Node* i_byte = new PhiNode(head, TypeInt::INT);
4577 i_byte->init_req(1, intcon(0));
4578 gvn().set_type(i_byte, TypeInt::INT);
4579 record_for_igvn(i_byte);
4580
4581 Node* i_char = new PhiNode(head, TypeInt::INT);
4582 i_char->init_req(1, start);
4583 gvn().set_type(i_char, TypeInt::INT);
4584 record_for_igvn(i_char);
4585
4586 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4587 gvn().set_type(mem, Type::MEMORY);
4588 record_for_igvn(mem);
4589 set_control(head);
4590 set_memory(mem, TypeAryPtr::BYTES);
4591 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4592 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4593 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4594 false, false, true /* mismatched */);
4595
4596 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4597 head->init_req(2, IfTrue(iff));
4598 mem->init_req(2, st);
4599 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4600 i_char->init_req(2, AddI(i_char, intcon(2)));
4601
4602 set_control(IfFalse(iff));
4603 set_memory(st, TypeAryPtr::BYTES);
4604 }
4605
4606 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4607 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4608 // assumption of CCP analysis.
4609 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4610 }