1 /*
2 * Copyright (c) 2001, 2013, 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 "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
28 #include "gc_implementation/g1/heapRegion.hpp"
29 #include "gc_interface/collectedHeap.hpp"
30 #include "memory/barrierSet.hpp"
31 #include "memory/cardTableModRefBS.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/graphKit.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "runtime/deoptimization.hpp"
41 #include "runtime/sharedRuntime.hpp"
42
43 //----------------------------GraphKit-----------------------------------------
44 // Main utility constructor.
45 GraphKit::GraphKit(JVMState* jvms)
46 : Phase(Phase::Parser),
47 _env(C->env()),
48 _gvn(*C->initial_gvn())
49 {
50 _exceptions = jvms->map()->next_exception();
51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
52 set_jvms(jvms);
53 }
54
55 // Private constructor for parser.
56 GraphKit::GraphKit()
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn())
60 {
61 _exceptions = NULL;
62 set_map(NULL);
63 debug_only(_sp = -99);
64 debug_only(set_bci(-99));
65 }
66
67
68
69 //---------------------------clean_stack---------------------------------------
70 // Clear away rubbish from the stack area of the JVM state.
71 // This destroys any arguments that may be waiting on the stack.
72 void GraphKit::clean_stack(int from_sp) {
73 SafePointNode* map = this->map();
74 JVMState* jvms = this->jvms();
75 int stk_size = jvms->stk_size();
76 int stkoff = jvms->stkoff();
77 Node* top = this->top();
78 for (int i = from_sp; i < stk_size; i++) {
79 if (map->in(stkoff + i) != top) {
80 map->set_req(stkoff + i, top);
81 }
82 }
83 }
84
85
86 //--------------------------------sync_jvms-----------------------------------
87 // Make sure our current jvms agrees with our parse state.
88 JVMState* GraphKit::sync_jvms() const {
89 JVMState* jvms = this->jvms();
90 jvms->set_bci(bci()); // Record the new bci in the JVMState
91 jvms->set_sp(sp()); // Record the new sp in the JVMState
92 assert(jvms_in_sync(), "jvms is now in sync");
93 return jvms;
94 }
95
96 //--------------------------------sync_jvms_for_reexecute---------------------
97 // Make sure our current jvms agrees with our parse state. This version
98 // uses the reexecute_sp for reexecuting bytecodes.
99 JVMState* GraphKit::sync_jvms_for_reexecute() {
100 JVMState* jvms = this->jvms();
101 jvms->set_bci(bci()); // Record the new bci in the JVMState
102 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
103 return jvms;
104 }
105
106 #ifdef ASSERT
107 bool GraphKit::jvms_in_sync() const {
108 Parse* parse = is_Parse();
109 if (parse == NULL) {
110 if (bci() != jvms()->bci()) return false;
111 if (sp() != (int)jvms()->sp()) return false;
112 return true;
113 }
114 if (jvms()->method() != parse->method()) return false;
115 if (jvms()->bci() != parse->bci()) return false;
116 int jvms_sp = jvms()->sp();
117 if (jvms_sp != parse->sp()) return false;
118 int jvms_depth = jvms()->depth();
119 if (jvms_depth != parse->depth()) return false;
120 return true;
121 }
122
123 // Local helper checks for special internal merge points
124 // used to accumulate and merge exception states.
125 // They are marked by the region's in(0) edge being the map itself.
126 // Such merge points must never "escape" into the parser at large,
127 // until they have been handed to gvn.transform.
128 static bool is_hidden_merge(Node* reg) {
129 if (reg == NULL) return false;
130 if (reg->is_Phi()) {
131 reg = reg->in(0);
132 if (reg == NULL) return false;
133 }
134 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
135 }
136
137 void GraphKit::verify_map() const {
138 if (map() == NULL) return; // null map is OK
139 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
140 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
141 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
142 }
143
144 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
145 assert(ex_map->next_exception() == NULL, "not already part of a chain");
146 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
147 }
148 #endif
149
150 //---------------------------stop_and_kill_map---------------------------------
151 // Set _map to NULL, signalling a stop to further bytecode execution.
152 // First smash the current map's control to a constant, to mark it dead.
153 void GraphKit::stop_and_kill_map() {
154 SafePointNode* dead_map = stop();
155 if (dead_map != NULL) {
156 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
157 assert(dead_map->is_killed(), "must be so marked");
158 }
159 }
160
161
162 //--------------------------------stopped--------------------------------------
163 // Tell if _map is NULL, or control is top.
164 bool GraphKit::stopped() {
165 if (map() == NULL) return true;
166 else if (control() == top()) return true;
167 else return false;
168 }
169
170
171 //-----------------------------has_ex_handler----------------------------------
172 // Tell if this method or any caller method has exception handlers.
173 bool GraphKit::has_ex_handler() {
174 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
175 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
176 return true;
177 }
178 }
179 return false;
180 }
181
182 //------------------------------save_ex_oop------------------------------------
183 // Save an exception without blowing stack contents or other JVM state.
184 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
185 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
186 ex_map->add_req(ex_oop);
187 debug_only(verify_exception_state(ex_map));
188 }
189
190 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
191 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
192 Node* ex_oop = ex_map->in(ex_map->req()-1);
193 if (clear_it) ex_map->del_req(ex_map->req()-1);
194 return ex_oop;
195 }
196
197 //-----------------------------saved_ex_oop------------------------------------
198 // Recover a saved exception from its map.
199 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
200 return common_saved_ex_oop(ex_map, false);
201 }
202
203 //--------------------------clear_saved_ex_oop---------------------------------
204 // Erase a previously saved exception from its map.
205 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
206 return common_saved_ex_oop(ex_map, true);
207 }
208
209 #ifdef ASSERT
210 //---------------------------has_saved_ex_oop----------------------------------
211 // Erase a previously saved exception from its map.
212 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
213 return ex_map->req() == ex_map->jvms()->endoff()+1;
214 }
215 #endif
216
217 //-------------------------make_exception_state--------------------------------
218 // Turn the current JVM state into an exception state, appending the ex_oop.
219 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
220 sync_jvms();
221 SafePointNode* ex_map = stop(); // do not manipulate this map any more
222 set_saved_ex_oop(ex_map, ex_oop);
223 return ex_map;
224 }
225
226
227 //--------------------------add_exception_state--------------------------------
228 // Add an exception to my list of exceptions.
229 void GraphKit::add_exception_state(SafePointNode* ex_map) {
230 if (ex_map == NULL || ex_map->control() == top()) {
231 return;
232 }
233 #ifdef ASSERT
234 verify_exception_state(ex_map);
235 if (has_exceptions()) {
236 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
237 }
238 #endif
239
240 // If there is already an exception of exactly this type, merge with it.
241 // In particular, null-checks and other low-level exceptions common up here.
242 Node* ex_oop = saved_ex_oop(ex_map);
243 const Type* ex_type = _gvn.type(ex_oop);
244 if (ex_oop == top()) {
245 // No action needed.
246 return;
247 }
248 assert(ex_type->isa_instptr(), "exception must be an instance");
249 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
250 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
251 // We check sp also because call bytecodes can generate exceptions
252 // both before and after arguments are popped!
253 if (ex_type2 == ex_type
254 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
255 combine_exception_states(ex_map, e2);
256 return;
257 }
258 }
259
260 // No pre-existing exception of the same type. Chain it on the list.
261 push_exception_state(ex_map);
262 }
263
264 //-----------------------add_exception_states_from-----------------------------
265 void GraphKit::add_exception_states_from(JVMState* jvms) {
266 SafePointNode* ex_map = jvms->map()->next_exception();
267 if (ex_map != NULL) {
268 jvms->map()->set_next_exception(NULL);
269 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
270 next_map = ex_map->next_exception();
271 ex_map->set_next_exception(NULL);
272 add_exception_state(ex_map);
273 }
274 }
275 }
276
277 //-----------------------transfer_exceptions_into_jvms-------------------------
278 JVMState* GraphKit::transfer_exceptions_into_jvms() {
279 if (map() == NULL) {
280 // We need a JVMS to carry the exceptions, but the map has gone away.
281 // Create a scratch JVMS, cloned from any of the exception states...
282 if (has_exceptions()) {
283 _map = _exceptions;
284 _map = clone_map();
285 _map->set_next_exception(NULL);
286 clear_saved_ex_oop(_map);
287 debug_only(verify_map());
288 } else {
289 // ...or created from scratch
290 JVMState* jvms = new (C) JVMState(_method, NULL);
291 jvms->set_bci(_bci);
292 jvms->set_sp(_sp);
293 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms));
294 set_jvms(jvms);
295 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
296 set_all_memory(top());
297 while (map()->req() < jvms->endoff()) map()->add_req(top());
298 }
299 // (This is a kludge, in case you didn't notice.)
300 set_control(top());
301 }
302 JVMState* jvms = sync_jvms();
303 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
304 jvms->map()->set_next_exception(_exceptions);
305 _exceptions = NULL; // done with this set of exceptions
306 return jvms;
307 }
308
309 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
310 assert(is_hidden_merge(dstphi), "must be a special merge node");
311 assert(is_hidden_merge(srcphi), "must be a special merge node");
312 uint limit = srcphi->req();
313 for (uint i = PhiNode::Input; i < limit; i++) {
314 dstphi->add_req(srcphi->in(i));
315 }
316 }
317 static inline void add_one_req(Node* dstphi, Node* src) {
318 assert(is_hidden_merge(dstphi), "must be a special merge node");
319 assert(!is_hidden_merge(src), "must not be a special merge node");
320 dstphi->add_req(src);
321 }
322
323 //-----------------------combine_exception_states------------------------------
324 // This helper function combines exception states by building phis on a
325 // specially marked state-merging region. These regions and phis are
326 // untransformed, and can build up gradually. The region is marked by
327 // having a control input of its exception map, rather than NULL. Such
328 // regions do not appear except in this function, and in use_exception_state.
329 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
330 if (failing()) return; // dying anyway...
331 JVMState* ex_jvms = ex_map->_jvms;
332 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
333 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
334 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
335 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
336 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
337 assert(ex_map->req() == phi_map->req(), "matching maps");
338 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
339 Node* hidden_merge_mark = root();
340 Node* region = phi_map->control();
341 MergeMemNode* phi_mem = phi_map->merged_memory();
342 MergeMemNode* ex_mem = ex_map->merged_memory();
343 if (region->in(0) != hidden_merge_mark) {
344 // The control input is not (yet) a specially-marked region in phi_map.
345 // Make it so, and build some phis.
346 region = new (C) RegionNode(2);
347 _gvn.set_type(region, Type::CONTROL);
348 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
349 region->init_req(1, phi_map->control());
350 phi_map->set_control(region);
351 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
352 record_for_igvn(io_phi);
353 _gvn.set_type(io_phi, Type::ABIO);
354 phi_map->set_i_o(io_phi);
355 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
356 Node* m = mms.memory();
357 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
358 record_for_igvn(m_phi);
359 _gvn.set_type(m_phi, Type::MEMORY);
360 mms.set_memory(m_phi);
361 }
362 }
363
364 // Either or both of phi_map and ex_map might already be converted into phis.
365 Node* ex_control = ex_map->control();
366 // if there is special marking on ex_map also, we add multiple edges from src
367 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
368 // how wide was the destination phi_map, originally?
369 uint orig_width = region->req();
370
371 if (add_multiple) {
372 add_n_reqs(region, ex_control);
373 add_n_reqs(phi_map->i_o(), ex_map->i_o());
374 } else {
375 // ex_map has no merges, so we just add single edges everywhere
376 add_one_req(region, ex_control);
377 add_one_req(phi_map->i_o(), ex_map->i_o());
378 }
379 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
380 if (mms.is_empty()) {
381 // get a copy of the base memory, and patch some inputs into it
382 const TypePtr* adr_type = mms.adr_type(C);
383 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
384 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
385 mms.set_memory(phi);
386 // Prepare to append interesting stuff onto the newly sliced phi:
387 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
388 }
389 // Append stuff from ex_map:
390 if (add_multiple) {
391 add_n_reqs(mms.memory(), mms.memory2());
392 } else {
393 add_one_req(mms.memory(), mms.memory2());
394 }
395 }
396 uint limit = ex_map->req();
397 for (uint i = TypeFunc::Parms; i < limit; i++) {
398 // Skip everything in the JVMS after tos. (The ex_oop follows.)
399 if (i == tos) i = ex_jvms->monoff();
400 Node* src = ex_map->in(i);
401 Node* dst = phi_map->in(i);
402 if (src != dst) {
403 PhiNode* phi;
404 if (dst->in(0) != region) {
405 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
406 record_for_igvn(phi);
407 _gvn.set_type(phi, phi->type());
408 phi_map->set_req(i, dst);
409 // Prepare to append interesting stuff onto the new phi:
410 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
411 } else {
412 assert(dst->is_Phi(), "nobody else uses a hidden region");
413 phi = dst->as_Phi();
414 }
415 if (add_multiple && src->in(0) == ex_control) {
416 // Both are phis.
417 add_n_reqs(dst, src);
418 } else {
419 while (dst->req() < region->req()) add_one_req(dst, src);
420 }
421 const Type* srctype = _gvn.type(src);
422 if (phi->type() != srctype) {
423 const Type* dsttype = phi->type()->meet_speculative(srctype);
424 if (phi->type() != dsttype) {
425 phi->set_type(dsttype);
426 _gvn.set_type(phi, dsttype);
427 }
428 }
429 }
430 }
431 }
432
433 //--------------------------use_exception_state--------------------------------
434 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
435 if (failing()) { stop(); return top(); }
436 Node* region = phi_map->control();
437 Node* hidden_merge_mark = root();
438 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
439 Node* ex_oop = clear_saved_ex_oop(phi_map);
440 if (region->in(0) == hidden_merge_mark) {
441 // Special marking for internal ex-states. Process the phis now.
442 region->set_req(0, region); // now it's an ordinary region
443 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
444 // Note: Setting the jvms also sets the bci and sp.
445 set_control(_gvn.transform(region));
446 uint tos = jvms()->stkoff() + sp();
447 for (uint i = 1; i < tos; i++) {
448 Node* x = phi_map->in(i);
449 if (x->in(0) == region) {
450 assert(x->is_Phi(), "expected a special phi");
451 phi_map->set_req(i, _gvn.transform(x));
452 }
453 }
454 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
455 Node* x = mms.memory();
456 if (x->in(0) == region) {
457 assert(x->is_Phi(), "nobody else uses a hidden region");
458 mms.set_memory(_gvn.transform(x));
459 }
460 }
461 if (ex_oop->in(0) == region) {
462 assert(ex_oop->is_Phi(), "expected a special phi");
463 ex_oop = _gvn.transform(ex_oop);
464 }
465 } else {
466 set_jvms(phi_map->jvms());
467 }
468
469 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
470 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
471 return ex_oop;
472 }
473
474 //---------------------------------java_bc-------------------------------------
475 Bytecodes::Code GraphKit::java_bc() const {
476 ciMethod* method = this->method();
477 int bci = this->bci();
478 if (method != NULL && bci != InvocationEntryBci)
479 return method->java_code_at_bci(bci);
480 else
481 return Bytecodes::_illegal;
482 }
483
484 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
485 bool must_throw) {
486 // if the exception capability is set, then we will generate code
487 // to check the JavaThread.should_post_on_exceptions flag to see
488 // if we actually need to report exception events (for this
489 // thread). If we don't need to report exception events, we will
490 // take the normal fast path provided by add_exception_events. If
491 // exception event reporting is enabled for this thread, we will
492 // take the uncommon_trap in the BuildCutout below.
493
494 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
495 Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
496 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
497 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false);
498
499 // Test the should_post_on_exceptions_flag vs. 0
500 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
501 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
502
503 // Branch to slow_path if should_post_on_exceptions_flag was true
504 { BuildCutout unless(this, tst, PROB_MAX);
505 // Do not try anything fancy if we're notifying the VM on every throw.
506 // Cf. case Bytecodes::_athrow in parse2.cpp.
507 uncommon_trap(reason, Deoptimization::Action_none,
508 (ciKlass*)NULL, (char*)NULL, must_throw);
509 }
510
511 }
512
513 //------------------------------builtin_throw----------------------------------
514 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
515 bool must_throw = true;
516
517 if (env()->jvmti_can_post_on_exceptions()) {
518 // check if we must post exception events, take uncommon trap if so
519 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
520 // here if should_post_on_exceptions is false
521 // continue on with the normal codegen
522 }
523
524 // If this particular condition has not yet happened at this
525 // bytecode, then use the uncommon trap mechanism, and allow for
526 // a future recompilation if several traps occur here.
527 // If the throw is hot, try to use a more complicated inline mechanism
528 // which keeps execution inside the compiled code.
529 bool treat_throw_as_hot = false;
530 ciMethodData* md = method()->method_data();
531
532 if (ProfileTraps) {
533 if (too_many_traps(reason)) {
534 treat_throw_as_hot = true;
535 }
536 // (If there is no MDO at all, assume it is early in
537 // execution, and that any deopts are part of the
538 // startup transient, and don't need to be remembered.)
539
540 // Also, if there is a local exception handler, treat all throws
541 // as hot if there has been at least one in this method.
542 if (C->trap_count(reason) != 0
543 && method()->method_data()->trap_count(reason) != 0
544 && has_ex_handler()) {
545 treat_throw_as_hot = true;
546 }
547 }
548
549 // If this throw happens frequently, an uncommon trap might cause
550 // a performance pothole. If there is a local exception handler,
551 // and if this particular bytecode appears to be deoptimizing often,
552 // let us handle the throw inline, with a preconstructed instance.
553 // Note: If the deopt count has blown up, the uncommon trap
554 // runtime is going to flush this nmethod, not matter what.
555 if (treat_throw_as_hot
556 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
557 // If the throw is local, we use a pre-existing instance and
558 // punt on the backtrace. This would lead to a missing backtrace
559 // (a repeat of 4292742) if the backtrace object is ever asked
560 // for its backtrace.
561 // Fixing this remaining case of 4292742 requires some flavor of
562 // escape analysis. Leave that for the future.
563 ciInstance* ex_obj = NULL;
564 switch (reason) {
565 case Deoptimization::Reason_null_check:
566 ex_obj = env()->NullPointerException_instance();
567 break;
568 case Deoptimization::Reason_div0_check:
569 ex_obj = env()->ArithmeticException_instance();
570 break;
571 case Deoptimization::Reason_range_check:
572 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
573 break;
574 case Deoptimization::Reason_class_check:
575 if (java_bc() == Bytecodes::_aastore) {
576 ex_obj = env()->ArrayStoreException_instance();
577 } else {
578 ex_obj = env()->ClassCastException_instance();
579 }
580 break;
581 }
582 if (failing()) { stop(); return; } // exception allocation might fail
583 if (ex_obj != NULL) {
584 // Cheat with a preallocated exception object.
585 if (C->log() != NULL)
586 C->log()->elem("hot_throw preallocated='1' reason='%s'",
587 Deoptimization::trap_reason_name(reason));
588 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
589 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
590
591 // Clear the detail message of the preallocated exception object.
592 // Weblogic sometimes mutates the detail message of exceptions
593 // using reflection.
594 int offset = java_lang_Throwable::get_detailMessage_offset();
595 const TypePtr* adr_typ = ex_con->add_offset(offset);
596
597 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
598 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
599 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT);
600
601 add_exception_state(make_exception_state(ex_node));
602 return;
603 }
604 }
605
606 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
607 // It won't be much cheaper than bailing to the interp., since we'll
608 // have to pass up all the debug-info, and the runtime will have to
609 // create the stack trace.
610
611 // Usual case: Bail to interpreter.
612 // Reserve the right to recompile if we haven't seen anything yet.
613
614 assert(!Deoptimization::reason_is_speculate(reason), "unsupported");
615 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
616 if (treat_throw_as_hot
617 && (method()->method_data()->trap_recompiled_at(bci(), NULL)
618 || C->too_many_traps(reason))) {
619 // We cannot afford to take more traps here. Suffer in the interpreter.
620 if (C->log() != NULL)
621 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
622 Deoptimization::trap_reason_name(reason),
623 C->trap_count(reason));
624 action = Deoptimization::Action_none;
625 }
626
627 // "must_throw" prunes the JVM state to include only the stack, if there
628 // are no local exception handlers. This should cut down on register
629 // allocation time and code size, by drastically reducing the number
630 // of in-edges on the call to the uncommon trap.
631
632 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
633 }
634
635
636 //----------------------------PreserveJVMState---------------------------------
637 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
638 debug_only(kit->verify_map());
639 _kit = kit;
640 _map = kit->map(); // preserve the map
641 _sp = kit->sp();
642 kit->set_map(clone_map ? kit->clone_map() : NULL);
643 Compile::current()->inc_preserve_jvm_state();
644 #ifdef ASSERT
645 _bci = kit->bci();
646 Parse* parser = kit->is_Parse();
647 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
648 _block = block;
649 #endif
650 }
651 PreserveJVMState::~PreserveJVMState() {
652 GraphKit* kit = _kit;
653 #ifdef ASSERT
654 assert(kit->bci() == _bci, "bci must not shift");
655 Parse* parser = kit->is_Parse();
656 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
657 assert(block == _block, "block must not shift");
658 #endif
659 kit->set_map(_map);
660 kit->set_sp(_sp);
661 Compile::current()->dec_preserve_jvm_state();
662 }
663
664
665 //-----------------------------BuildCutout-------------------------------------
666 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
667 : PreserveJVMState(kit)
668 {
669 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
670 SafePointNode* outer_map = _map; // preserved map is caller's
671 SafePointNode* inner_map = kit->map();
672 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
673 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
674 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
675 }
676 BuildCutout::~BuildCutout() {
677 GraphKit* kit = _kit;
678 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
679 }
680
681 //---------------------------PreserveReexecuteState----------------------------
682 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
683 assert(!kit->stopped(), "must call stopped() before");
684 _kit = kit;
685 _sp = kit->sp();
686 _reexecute = kit->jvms()->_reexecute;
687 }
688 PreserveReexecuteState::~PreserveReexecuteState() {
689 if (_kit->stopped()) return;
690 _kit->jvms()->_reexecute = _reexecute;
691 _kit->set_sp(_sp);
692 }
693
694 //------------------------------clone_map--------------------------------------
695 // Implementation of PreserveJVMState
696 //
697 // Only clone_map(...) here. If this function is only used in the
698 // PreserveJVMState class we may want to get rid of this extra
699 // function eventually and do it all there.
700
701 SafePointNode* GraphKit::clone_map() {
702 if (map() == NULL) return NULL;
703
704 // Clone the memory edge first
705 Node* mem = MergeMemNode::make(C, map()->memory());
706 gvn().set_type_bottom(mem);
707
708 SafePointNode *clonemap = (SafePointNode*)map()->clone();
709 JVMState* jvms = this->jvms();
710 JVMState* clonejvms = jvms->clone_shallow(C);
711 clonemap->set_memory(mem);
712 clonemap->set_jvms(clonejvms);
713 clonejvms->set_map(clonemap);
714 record_for_igvn(clonemap);
715 gvn().set_type_bottom(clonemap);
716 return clonemap;
717 }
718
719
720 //-----------------------------set_map_clone-----------------------------------
721 void GraphKit::set_map_clone(SafePointNode* m) {
722 _map = m;
723 _map = clone_map();
724 _map->set_next_exception(NULL);
725 debug_only(verify_map());
726 }
727
728
729 //----------------------------kill_dead_locals---------------------------------
730 // Detect any locals which are known to be dead, and force them to top.
731 void GraphKit::kill_dead_locals() {
732 // Consult the liveness information for the locals. If any
733 // of them are unused, then they can be replaced by top(). This
734 // should help register allocation time and cut down on the size
735 // of the deoptimization information.
736
737 // This call is made from many of the bytecode handling
738 // subroutines called from the Big Switch in do_one_bytecode.
739 // Every bytecode which might include a slow path is responsible
740 // for killing its dead locals. The more consistent we
741 // are about killing deads, the fewer useless phis will be
742 // constructed for them at various merge points.
743
744 // bci can be -1 (InvocationEntryBci). We return the entry
745 // liveness for the method.
746
747 if (method() == NULL || method()->code_size() == 0) {
748 // We are building a graph for a call to a native method.
749 // All locals are live.
750 return;
751 }
752
753 ResourceMark rm;
754
755 // Consult the liveness information for the locals. If any
756 // of them are unused, then they can be replaced by top(). This
757 // should help register allocation time and cut down on the size
758 // of the deoptimization information.
759 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
760
761 int len = (int)live_locals.size();
762 assert(len <= jvms()->loc_size(), "too many live locals");
763 for (int local = 0; local < len; local++) {
764 if (!live_locals.at(local)) {
765 set_local(local, top());
766 }
767 }
768 }
769
770 #ifdef ASSERT
771 //-------------------------dead_locals_are_killed------------------------------
772 // Return true if all dead locals are set to top in the map.
773 // Used to assert "clean" debug info at various points.
774 bool GraphKit::dead_locals_are_killed() {
775 if (method() == NULL || method()->code_size() == 0) {
776 // No locals need to be dead, so all is as it should be.
777 return true;
778 }
779
780 // Make sure somebody called kill_dead_locals upstream.
781 ResourceMark rm;
782 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
783 if (jvms->loc_size() == 0) continue; // no locals to consult
784 SafePointNode* map = jvms->map();
785 ciMethod* method = jvms->method();
786 int bci = jvms->bci();
787 if (jvms == this->jvms()) {
788 bci = this->bci(); // it might not yet be synched
789 }
790 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
791 int len = (int)live_locals.size();
792 if (!live_locals.is_valid() || len == 0)
793 // This method is trivial, or is poisoned by a breakpoint.
794 return true;
795 assert(len == jvms->loc_size(), "live map consistent with locals map");
796 for (int local = 0; local < len; local++) {
797 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
798 if (PrintMiscellaneous && (Verbose || WizardMode)) {
799 tty->print_cr("Zombie local %d: ", local);
800 jvms->dump();
801 }
802 return false;
803 }
804 }
805 }
806 return true;
807 }
808
809 #endif //ASSERT
810
811 // Helper function for enforcing certain bytecodes to reexecute if
812 // deoptimization happens
813 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
814 ciMethod* cur_method = jvms->method();
815 int cur_bci = jvms->bci();
816 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
817 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
818 return Interpreter::bytecode_should_reexecute(code) ||
819 is_anewarray && code == Bytecodes::_multianewarray;
820 // Reexecute _multianewarray bytecode which was replaced with
821 // sequence of [a]newarray. See Parse::do_multianewarray().
822 //
823 // Note: interpreter should not have it set since this optimization
824 // is limited by dimensions and guarded by flag so in some cases
825 // multianewarray() runtime calls will be generated and
826 // the bytecode should not be reexecutes (stack will not be reset).
827 } else
828 return false;
829 }
830
831 // Helper function for adding JVMState and debug information to node
832 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
833 // Add the safepoint edges to the call (or other safepoint).
834
835 // Make sure dead locals are set to top. This
836 // should help register allocation time and cut down on the size
837 // of the deoptimization information.
838 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
839
840 // Walk the inline list to fill in the correct set of JVMState's
841 // Also fill in the associated edges for each JVMState.
842
843 // If the bytecode needs to be reexecuted we need to put
844 // the arguments back on the stack.
845 const bool should_reexecute = jvms()->should_reexecute();
846 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
847
848 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
849 // undefined if the bci is different. This is normal for Parse but it
850 // should not happen for LibraryCallKit because only one bci is processed.
851 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
852 "in LibraryCallKit the reexecute bit should not change");
853
854 // If we are guaranteed to throw, we can prune everything but the
855 // input to the current bytecode.
856 bool can_prune_locals = false;
857 uint stack_slots_not_pruned = 0;
858 int inputs = 0, depth = 0;
859 if (must_throw) {
860 assert(method() == youngest_jvms->method(), "sanity");
861 if (compute_stack_effects(inputs, depth)) {
862 can_prune_locals = true;
863 stack_slots_not_pruned = inputs;
864 }
865 }
866
867 if (env()->jvmti_can_access_local_variables()) {
868 // At any safepoint, this method can get breakpointed, which would
869 // then require an immediate deoptimization.
870 can_prune_locals = false; // do not prune locals
871 stack_slots_not_pruned = 0;
872 }
873
874 // do not scribble on the input jvms
875 JVMState* out_jvms = youngest_jvms->clone_deep(C);
876 call->set_jvms(out_jvms); // Start jvms list for call node
877
878 // For a known set of bytecodes, the interpreter should reexecute them if
879 // deoptimization happens. We set the reexecute state for them here
880 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
881 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
882 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
883 }
884
885 // Presize the call:
886 DEBUG_ONLY(uint non_debug_edges = call->req());
887 call->add_req_batch(top(), youngest_jvms->debug_depth());
888 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
889
890 // Set up edges so that the call looks like this:
891 // Call [state:] ctl io mem fptr retadr
892 // [parms:] parm0 ... parmN
893 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
894 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
895 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
896 // Note that caller debug info precedes callee debug info.
897
898 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
899 uint debug_ptr = call->req();
900
901 // Loop over the map input edges associated with jvms, add them
902 // to the call node, & reset all offsets to match call node array.
903 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
904 uint debug_end = debug_ptr;
905 uint debug_start = debug_ptr - in_jvms->debug_size();
906 debug_ptr = debug_start; // back up the ptr
907
908 uint p = debug_start; // walks forward in [debug_start, debug_end)
909 uint j, k, l;
910 SafePointNode* in_map = in_jvms->map();
911 out_jvms->set_map(call);
912
913 if (can_prune_locals) {
914 assert(in_jvms->method() == out_jvms->method(), "sanity");
915 // If the current throw can reach an exception handler in this JVMS,
916 // then we must keep everything live that can reach that handler.
917 // As a quick and dirty approximation, we look for any handlers at all.
918 if (in_jvms->method()->has_exception_handlers()) {
919 can_prune_locals = false;
920 }
921 }
922
923 // Add the Locals
924 k = in_jvms->locoff();
925 l = in_jvms->loc_size();
926 out_jvms->set_locoff(p);
927 if (!can_prune_locals) {
928 for (j = 0; j < l; j++)
929 call->set_req(p++, in_map->in(k+j));
930 } else {
931 p += l; // already set to top above by add_req_batch
932 }
933
934 // Add the Expression Stack
935 k = in_jvms->stkoff();
936 l = in_jvms->sp();
937 out_jvms->set_stkoff(p);
938 if (!can_prune_locals) {
939 for (j = 0; j < l; j++)
940 call->set_req(p++, in_map->in(k+j));
941 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
942 // Divide stack into {S0,...,S1}, where S0 is set to top.
943 uint s1 = stack_slots_not_pruned;
944 stack_slots_not_pruned = 0; // for next iteration
945 if (s1 > l) s1 = l;
946 uint s0 = l - s1;
947 p += s0; // skip the tops preinstalled by add_req_batch
948 for (j = s0; j < l; j++)
949 call->set_req(p++, in_map->in(k+j));
950 } else {
951 p += l; // already set to top above by add_req_batch
952 }
953
954 // Add the Monitors
955 k = in_jvms->monoff();
956 l = in_jvms->mon_size();
957 out_jvms->set_monoff(p);
958 for (j = 0; j < l; j++)
959 call->set_req(p++, in_map->in(k+j));
960
961 // Copy any scalar object fields.
962 k = in_jvms->scloff();
963 l = in_jvms->scl_size();
964 out_jvms->set_scloff(p);
965 for (j = 0; j < l; j++)
966 call->set_req(p++, in_map->in(k+j));
967
968 // Finish the new jvms.
969 out_jvms->set_endoff(p);
970
971 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
972 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
973 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
974 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
975 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
976 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
977
978 // Update the two tail pointers in parallel.
979 out_jvms = out_jvms->caller();
980 in_jvms = in_jvms->caller();
981 }
982
983 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
984
985 // Test the correctness of JVMState::debug_xxx accessors:
986 assert(call->jvms()->debug_start() == non_debug_edges, "");
987 assert(call->jvms()->debug_end() == call->req(), "");
988 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
989 }
990
991 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
992 Bytecodes::Code code = java_bc();
993 if (code == Bytecodes::_wide) {
994 code = method()->java_code_at_bci(bci() + 1);
995 }
996
997 BasicType rtype = T_ILLEGAL;
998 int rsize = 0;
999
1000 if (code != Bytecodes::_illegal) {
1001 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1002 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1003 if (rtype < T_CONFLICT)
1004 rsize = type2size[rtype];
1005 }
1006
1007 switch (code) {
1008 case Bytecodes::_illegal:
1009 return false;
1010
1011 case Bytecodes::_ldc:
1012 case Bytecodes::_ldc_w:
1013 case Bytecodes::_ldc2_w:
1014 inputs = 0;
1015 break;
1016
1017 case Bytecodes::_dup: inputs = 1; break;
1018 case Bytecodes::_dup_x1: inputs = 2; break;
1019 case Bytecodes::_dup_x2: inputs = 3; break;
1020 case Bytecodes::_dup2: inputs = 2; break;
1021 case Bytecodes::_dup2_x1: inputs = 3; break;
1022 case Bytecodes::_dup2_x2: inputs = 4; break;
1023 case Bytecodes::_swap: inputs = 2; break;
1024 case Bytecodes::_arraylength: inputs = 1; break;
1025
1026 case Bytecodes::_getstatic:
1027 case Bytecodes::_putstatic:
1028 case Bytecodes::_getfield:
1029 case Bytecodes::_putfield:
1030 {
1031 bool ignored_will_link;
1032 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1033 int size = field->type()->size();
1034 bool is_get = (depth >= 0), is_static = (depth & 1);
1035 inputs = (is_static ? 0 : 1);
1036 if (is_get) {
1037 depth = size - inputs;
1038 } else {
1039 inputs += size; // putxxx pops the value from the stack
1040 depth = - inputs;
1041 }
1042 }
1043 break;
1044
1045 case Bytecodes::_invokevirtual:
1046 case Bytecodes::_invokespecial:
1047 case Bytecodes::_invokestatic:
1048 case Bytecodes::_invokedynamic:
1049 case Bytecodes::_invokeinterface:
1050 {
1051 bool ignored_will_link;
1052 ciSignature* declared_signature = NULL;
1053 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1054 assert(declared_signature != NULL, "cannot be null");
1055 inputs = declared_signature->arg_size_for_bc(code);
1056 int size = declared_signature->return_type()->size();
1057 depth = size - inputs;
1058 }
1059 break;
1060
1061 case Bytecodes::_multianewarray:
1062 {
1063 ciBytecodeStream iter(method());
1064 iter.reset_to_bci(bci());
1065 iter.next();
1066 inputs = iter.get_dimensions();
1067 assert(rsize == 1, "");
1068 depth = rsize - inputs;
1069 }
1070 break;
1071
1072 case Bytecodes::_ireturn:
1073 case Bytecodes::_lreturn:
1074 case Bytecodes::_freturn:
1075 case Bytecodes::_dreturn:
1076 case Bytecodes::_areturn:
1077 assert(rsize = -depth, "");
1078 inputs = rsize;
1079 break;
1080
1081 case Bytecodes::_jsr:
1082 case Bytecodes::_jsr_w:
1083 inputs = 0;
1084 depth = 1; // S.B. depth=1, not zero
1085 break;
1086
1087 default:
1088 // bytecode produces a typed result
1089 inputs = rsize - depth;
1090 assert(inputs >= 0, "");
1091 break;
1092 }
1093
1094 #ifdef ASSERT
1095 // spot check
1096 int outputs = depth + inputs;
1097 assert(outputs >= 0, "sanity");
1098 switch (code) {
1099 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1100 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1101 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1102 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1103 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1104 }
1105 #endif //ASSERT
1106
1107 return true;
1108 }
1109
1110
1111
1112 //------------------------------basic_plus_adr---------------------------------
1113 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1114 // short-circuit a common case
1115 if (offset == intcon(0)) return ptr;
1116 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1117 }
1118
1119 Node* GraphKit::ConvI2L(Node* offset) {
1120 // short-circuit a common case
1121 jint offset_con = find_int_con(offset, Type::OffsetBot);
1122 if (offset_con != Type::OffsetBot) {
1123 return longcon((jlong) offset_con);
1124 }
1125 return _gvn.transform( new (C) ConvI2LNode(offset));
1126 }
1127 Node* GraphKit::ConvL2I(Node* offset) {
1128 // short-circuit a common case
1129 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1130 if (offset_con != (jlong)Type::OffsetBot) {
1131 return intcon((int) offset_con);
1132 }
1133 return _gvn.transform( new (C) ConvL2INode(offset));
1134 }
1135
1136 //-------------------------load_object_klass-----------------------------------
1137 Node* GraphKit::load_object_klass(Node* obj) {
1138 // Special-case a fresh allocation to avoid building nodes:
1139 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1140 if (akls != NULL) return akls;
1141 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1142 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1143 }
1144
1145 //-------------------------load_array_length-----------------------------------
1146 Node* GraphKit::load_array_length(Node* array) {
1147 // Special-case a fresh allocation to avoid building nodes:
1148 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1149 Node *alen;
1150 if (alloc == NULL) {
1151 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1152 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1153 } else {
1154 alen = alloc->Ideal_length();
1155 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1156 if (ccast != alen) {
1157 alen = _gvn.transform(ccast);
1158 }
1159 }
1160 return alen;
1161 }
1162
1163 //------------------------------do_null_check----------------------------------
1164 // Helper function to do a NULL pointer check. Returned value is
1165 // the incoming address with NULL casted away. You are allowed to use the
1166 // not-null value only if you are control dependent on the test.
1167 extern int explicit_null_checks_inserted,
1168 explicit_null_checks_elided;
1169 Node* GraphKit::null_check_common(Node* value, BasicType type,
1170 // optional arguments for variations:
1171 bool assert_null,
1172 Node* *null_control) {
1173 assert(!assert_null || null_control == NULL, "not both at once");
1174 if (stopped()) return top();
1175 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1176 // For some performance testing, we may wish to suppress null checking.
1177 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1178 return value;
1179 }
1180 explicit_null_checks_inserted++;
1181
1182 // Construct NULL check
1183 Node *chk = NULL;
1184 switch(type) {
1185 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1186 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1187 case T_ARRAY : // fall through
1188 type = T_OBJECT; // simplify further tests
1189 case T_OBJECT : {
1190 const Type *t = _gvn.type( value );
1191
1192 const TypeOopPtr* tp = t->isa_oopptr();
1193 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1194 // Only for do_null_check, not any of its siblings:
1195 && !assert_null && null_control == NULL) {
1196 // Usually, any field access or invocation on an unloaded oop type
1197 // will simply fail to link, since the statically linked class is
1198 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1199 // the static class is loaded but the sharper oop type is not.
1200 // Rather than checking for this obscure case in lots of places,
1201 // we simply observe that a null check on an unloaded class
1202 // will always be followed by a nonsense operation, so we
1203 // can just issue the uncommon trap here.
1204 // Our access to the unloaded class will only be correct
1205 // after it has been loaded and initialized, which requires
1206 // a trip through the interpreter.
1207 #ifndef PRODUCT
1208 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1209 #endif
1210 uncommon_trap(Deoptimization::Reason_unloaded,
1211 Deoptimization::Action_reinterpret,
1212 tp->klass(), "!loaded");
1213 return top();
1214 }
1215
1216 if (assert_null) {
1217 // See if the type is contained in NULL_PTR.
1218 // If so, then the value is already null.
1219 if (t->higher_equal(TypePtr::NULL_PTR)) {
1220 explicit_null_checks_elided++;
1221 return value; // Elided null assert quickly!
1222 }
1223 } else {
1224 // See if mixing in the NULL pointer changes type.
1225 // If so, then the NULL pointer was not allowed in the original
1226 // type. In other words, "value" was not-null.
1227 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1228 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1229 explicit_null_checks_elided++;
1230 return value; // Elided null check quickly!
1231 }
1232 }
1233 chk = new (C) CmpPNode( value, null() );
1234 break;
1235 }
1236
1237 default:
1238 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1239 }
1240 assert(chk != NULL, "sanity check");
1241 chk = _gvn.transform(chk);
1242
1243 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1244 BoolNode *btst = new (C) BoolNode( chk, btest);
1245 Node *tst = _gvn.transform( btst );
1246
1247 //-----------
1248 // if peephole optimizations occurred, a prior test existed.
1249 // If a prior test existed, maybe it dominates as we can avoid this test.
1250 if (tst != btst && type == T_OBJECT) {
1251 // At this point we want to scan up the CFG to see if we can
1252 // find an identical test (and so avoid this test altogether).
1253 Node *cfg = control();
1254 int depth = 0;
1255 while( depth < 16 ) { // Limit search depth for speed
1256 if( cfg->Opcode() == Op_IfTrue &&
1257 cfg->in(0)->in(1) == tst ) {
1258 // Found prior test. Use "cast_not_null" to construct an identical
1259 // CastPP (and hence hash to) as already exists for the prior test.
1260 // Return that casted value.
1261 if (assert_null) {
1262 replace_in_map(value, null());
1263 return null(); // do not issue the redundant test
1264 }
1265 Node *oldcontrol = control();
1266 set_control(cfg);
1267 Node *res = cast_not_null(value);
1268 set_control(oldcontrol);
1269 explicit_null_checks_elided++;
1270 return res;
1271 }
1272 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1273 if (cfg == NULL) break; // Quit at region nodes
1274 depth++;
1275 }
1276 }
1277
1278 //-----------
1279 // Branch to failure if null
1280 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1281 Deoptimization::DeoptReason reason;
1282 if (assert_null)
1283 reason = Deoptimization::Reason_null_assert;
1284 else if (type == T_OBJECT)
1285 reason = Deoptimization::Reason_null_check;
1286 else
1287 reason = Deoptimization::Reason_div0_check;
1288
1289 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1290 // ciMethodData::has_trap_at will return a conservative -1 if any
1291 // must-be-null assertion has failed. This could cause performance
1292 // problems for a method after its first do_null_assert failure.
1293 // Consider using 'Reason_class_check' instead?
1294
1295 // To cause an implicit null check, we set the not-null probability
1296 // to the maximum (PROB_MAX). For an explicit check the probability
1297 // is set to a smaller value.
1298 if (null_control != NULL || too_many_traps(reason)) {
1299 // probability is less likely
1300 ok_prob = PROB_LIKELY_MAG(3);
1301 } else if (!assert_null &&
1302 (ImplicitNullCheckThreshold > 0) &&
1303 method() != NULL &&
1304 (method()->method_data()->trap_count(reason)
1305 >= (uint)ImplicitNullCheckThreshold)) {
1306 ok_prob = PROB_LIKELY_MAG(3);
1307 }
1308
1309 if (null_control != NULL) {
1310 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1311 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1312 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1313 if (null_true == top())
1314 explicit_null_checks_elided++;
1315 (*null_control) = null_true;
1316 } else {
1317 BuildCutout unless(this, tst, ok_prob);
1318 // Check for optimizer eliding test at parse time
1319 if (stopped()) {
1320 // Failure not possible; do not bother making uncommon trap.
1321 explicit_null_checks_elided++;
1322 } else if (assert_null) {
1323 uncommon_trap(reason,
1324 Deoptimization::Action_make_not_entrant,
1325 NULL, "assert_null");
1326 } else {
1327 replace_in_map(value, zerocon(type));
1328 builtin_throw(reason);
1329 }
1330 }
1331
1332 // Must throw exception, fall-thru not possible?
1333 if (stopped()) {
1334 return top(); // No result
1335 }
1336
1337 if (assert_null) {
1338 // Cast obj to null on this path.
1339 replace_in_map(value, zerocon(type));
1340 return zerocon(type);
1341 }
1342
1343 // Cast obj to not-null on this path, if there is no null_control.
1344 // (If there is a null_control, a non-null value may come back to haunt us.)
1345 if (type == T_OBJECT) {
1346 Node* cast = cast_not_null(value, false);
1347 if (null_control == NULL || (*null_control) == top())
1348 replace_in_map(value, cast);
1349 value = cast;
1350 }
1351
1352 return value;
1353 }
1354
1355
1356 //------------------------------cast_not_null----------------------------------
1357 // Cast obj to not-null on this path
1358 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1359 const Type *t = _gvn.type(obj);
1360 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1361 // Object is already not-null?
1362 if( t == t_not_null ) return obj;
1363
1364 Node *cast = new (C) CastPPNode(obj,t_not_null);
1365 cast->init_req(0, control());
1366 cast = _gvn.transform( cast );
1367
1368 // Scan for instances of 'obj' in the current JVM mapping.
1369 // These instances are known to be not-null after the test.
1370 if (do_replace_in_map)
1371 replace_in_map(obj, cast);
1372
1373 return cast; // Return casted value
1374 }
1375
1376
1377 //--------------------------replace_in_map-------------------------------------
1378 void GraphKit::replace_in_map(Node* old, Node* neww) {
1379 if (old == neww) {
1380 return;
1381 }
1382
1383 map()->replace_edge(old, neww);
1384
1385 // Note: This operation potentially replaces any edge
1386 // on the map. This includes locals, stack, and monitors
1387 // of the current (innermost) JVM state.
1388
1389 if (!ReplaceInParentMaps) {
1390 return;
1391 }
1392
1393 // PreserveJVMState doesn't do a deep copy so we can't modify
1394 // parents
1395 if (Compile::current()->has_preserve_jvm_state()) {
1396 return;
1397 }
1398
1399 Parse* parser = is_Parse();
1400 bool progress = true;
1401 Node* ctrl = map()->in(0);
1402 // Follow the chain of parsers and see whether the update can be
1403 // done in the map of callers. We can do the replace for a caller if
1404 // the current control post dominates the control of a caller.
1405 while (parser != NULL && parser->caller() != NULL && progress) {
1406 progress = false;
1407 Node* parent_map = parser->caller()->map();
1408 assert(parser->exits().map()->jvms()->depth() == parser->caller()->depth(), "map mismatch");
1409
1410 Node* parent_ctrl = parent_map->in(0);
1411
1412 while (parent_ctrl->is_Region()) {
1413 Node* n = parent_ctrl->as_Region()->is_copy();
1414 if (n == NULL) {
1415 break;
1416 }
1417 parent_ctrl = n;
1418 }
1419
1420 for (;;) {
1421 if (ctrl == parent_ctrl) {
1422 // update the map of the exits which is the one that will be
1423 // used when compilation resume after inlining
1424 parser->exits().map()->replace_edge(old, neww);
1425 progress = true;
1426 break;
1427 }
1428 if (ctrl->is_Proj() && ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
1429 ctrl = ctrl->in(0)->in(0);
1430 } else if (ctrl->is_Region()) {
1431 Node* n = ctrl->as_Region()->is_copy();
1432 if (n == NULL) {
1433 break;
1434 }
1435 ctrl = n;
1436 } else {
1437 break;
1438 }
1439 }
1440
1441 parser = parser->parent_parser();
1442 }
1443 }
1444
1445
1446 //=============================================================================
1447 //--------------------------------memory---------------------------------------
1448 Node* GraphKit::memory(uint alias_idx) {
1449 MergeMemNode* mem = merged_memory();
1450 Node* p = mem->memory_at(alias_idx);
1451 _gvn.set_type(p, Type::MEMORY); // must be mapped
1452 return p;
1453 }
1454
1455 //-----------------------------reset_memory------------------------------------
1456 Node* GraphKit::reset_memory() {
1457 Node* mem = map()->memory();
1458 // do not use this node for any more parsing!
1459 debug_only( map()->set_memory((Node*)NULL) );
1460 return _gvn.transform( mem );
1461 }
1462
1463 //------------------------------set_all_memory---------------------------------
1464 void GraphKit::set_all_memory(Node* newmem) {
1465 Node* mergemem = MergeMemNode::make(C, newmem);
1466 gvn().set_type_bottom(mergemem);
1467 map()->set_memory(mergemem);
1468 }
1469
1470 //------------------------------set_all_memory_call----------------------------
1471 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1472 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1473 set_all_memory(newmem);
1474 }
1475
1476 //=============================================================================
1477 //
1478 // parser factory methods for MemNodes
1479 //
1480 // These are layered on top of the factory methods in LoadNode and StoreNode,
1481 // and integrate with the parser's memory state and _gvn engine.
1482 //
1483
1484 // factory methods in "int adr_idx"
1485 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1486 int adr_idx,
1487 bool require_atomic_access) {
1488 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1489 const TypePtr* adr_type = NULL; // debug-mode-only argument
1490 debug_only(adr_type = C->get_adr_type(adr_idx));
1491 Node* mem = memory(adr_idx);
1492 Node* ld;
1493 if (require_atomic_access && bt == T_LONG) {
1494 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1495 } else {
1496 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1497 }
1498 ld = _gvn.transform(ld);
1499 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1500 // Improve graph before escape analysis and boxing elimination.
1501 record_for_igvn(ld);
1502 }
1503 return ld;
1504 }
1505
1506 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1507 int adr_idx,
1508 bool require_atomic_access) {
1509 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1510 const TypePtr* adr_type = NULL;
1511 debug_only(adr_type = C->get_adr_type(adr_idx));
1512 Node *mem = memory(adr_idx);
1513 Node* st;
1514 if (require_atomic_access && bt == T_LONG) {
1515 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1516 } else {
1517 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1518 }
1519 st = _gvn.transform(st);
1520 set_memory(st, adr_idx);
1521 // Back-to-back stores can only remove intermediate store with DU info
1522 // so push on worklist for optimizer.
1523 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1524 record_for_igvn(st);
1525
1526 return st;
1527 }
1528
1529
1530 void GraphKit::pre_barrier(bool do_load,
1531 Node* ctl,
1532 Node* obj,
1533 Node* adr,
1534 uint adr_idx,
1535 Node* val,
1536 const TypeOopPtr* val_type,
1537 Node* pre_val,
1538 BasicType bt) {
1539
1540 BarrierSet* bs = Universe::heap()->barrier_set();
1541 set_control(ctl);
1542 switch (bs->kind()) {
1543 case BarrierSet::G1SATBCT:
1544 case BarrierSet::G1SATBCTLogging:
1545 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1546 break;
1547
1548 case BarrierSet::CardTableModRef:
1549 case BarrierSet::CardTableExtension:
1550 case BarrierSet::ModRef:
1551 break;
1552
1553 case BarrierSet::Other:
1554 default :
1555 ShouldNotReachHere();
1556
1557 }
1558 }
1559
1560 bool GraphKit::can_move_pre_barrier() const {
1561 BarrierSet* bs = Universe::heap()->barrier_set();
1562 switch (bs->kind()) {
1563 case BarrierSet::G1SATBCT:
1564 case BarrierSet::G1SATBCTLogging:
1565 return true; // Can move it if no safepoint
1566
1567 case BarrierSet::CardTableModRef:
1568 case BarrierSet::CardTableExtension:
1569 case BarrierSet::ModRef:
1570 return true; // There is no pre-barrier
1571
1572 case BarrierSet::Other:
1573 default :
1574 ShouldNotReachHere();
1575 }
1576 return false;
1577 }
1578
1579 void GraphKit::post_barrier(Node* ctl,
1580 Node* store,
1581 Node* obj,
1582 Node* adr,
1583 uint adr_idx,
1584 Node* val,
1585 BasicType bt,
1586 bool use_precise) {
1587 BarrierSet* bs = Universe::heap()->barrier_set();
1588 set_control(ctl);
1589 switch (bs->kind()) {
1590 case BarrierSet::G1SATBCT:
1591 case BarrierSet::G1SATBCTLogging:
1592 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1593 break;
1594
1595 case BarrierSet::CardTableModRef:
1596 case BarrierSet::CardTableExtension:
1597 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1598 break;
1599
1600 case BarrierSet::ModRef:
1601 break;
1602
1603 case BarrierSet::Other:
1604 default :
1605 ShouldNotReachHere();
1606
1607 }
1608 }
1609
1610 Node* GraphKit::store_oop(Node* ctl,
1611 Node* obj,
1612 Node* adr,
1613 const TypePtr* adr_type,
1614 Node* val,
1615 const TypeOopPtr* val_type,
1616 BasicType bt,
1617 bool use_precise) {
1618 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1619 // could be delayed during Parse (for example, in adjust_map_after_if()).
1620 // Execute transformation here to avoid barrier generation in such case.
1621 if (_gvn.type(val) == TypePtr::NULL_PTR)
1622 val = _gvn.makecon(TypePtr::NULL_PTR);
1623
1624 set_control(ctl);
1625 if (stopped()) return top(); // Dead path ?
1626
1627 assert(bt == T_OBJECT, "sanity");
1628 assert(val != NULL, "not dead path");
1629 uint adr_idx = C->get_alias_index(adr_type);
1630 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1631
1632 pre_barrier(true /* do_load */,
1633 control(), obj, adr, adr_idx, val, val_type,
1634 NULL /* pre_val */,
1635 bt);
1636
1637 Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1638 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1639 return store;
1640 }
1641
1642 // Could be an array or object we don't know at compile time (unsafe ref.)
1643 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1644 Node* obj, // containing obj
1645 Node* adr, // actual adress to store val at
1646 const TypePtr* adr_type,
1647 Node* val,
1648 BasicType bt) {
1649 Compile::AliasType* at = C->alias_type(adr_type);
1650 const TypeOopPtr* val_type = NULL;
1651 if (adr_type->isa_instptr()) {
1652 if (at->field() != NULL) {
1653 // known field. This code is a copy of the do_put_xxx logic.
1654 ciField* field = at->field();
1655 if (!field->type()->is_loaded()) {
1656 val_type = TypeInstPtr::BOTTOM;
1657 } else {
1658 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1659 }
1660 }
1661 } else if (adr_type->isa_aryptr()) {
1662 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1663 }
1664 if (val_type == NULL) {
1665 val_type = TypeInstPtr::BOTTOM;
1666 }
1667 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true);
1668 }
1669
1670
1671 //-------------------------array_element_address-------------------------
1672 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1673 const TypeInt* sizetype) {
1674 uint shift = exact_log2(type2aelembytes(elembt));
1675 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1676
1677 // short-circuit a common case (saves lots of confusing waste motion)
1678 jint idx_con = find_int_con(idx, -1);
1679 if (idx_con >= 0) {
1680 intptr_t offset = header + ((intptr_t)idx_con << shift);
1681 return basic_plus_adr(ary, offset);
1682 }
1683
1684 // must be correct type for alignment purposes
1685 Node* base = basic_plus_adr(ary, header);
1686 #ifdef _LP64
1687 // The scaled index operand to AddP must be a clean 64-bit value.
1688 // Java allows a 32-bit int to be incremented to a negative
1689 // value, which appears in a 64-bit register as a large
1690 // positive number. Using that large positive number as an
1691 // operand in pointer arithmetic has bad consequences.
1692 // On the other hand, 32-bit overflow is rare, and the possibility
1693 // can often be excluded, if we annotate the ConvI2L node with
1694 // a type assertion that its value is known to be a small positive
1695 // number. (The prior range check has ensured this.)
1696 // This assertion is used by ConvI2LNode::Ideal.
1697 int index_max = max_jint - 1; // array size is max_jint, index is one less
1698 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1699 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1700 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) );
1701 #endif
1702 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1703 return basic_plus_adr(ary, base, scale);
1704 }
1705
1706 //-------------------------load_array_element-------------------------
1707 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1708 const Type* elemtype = arytype->elem();
1709 BasicType elembt = elemtype->array_element_basic_type();
1710 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1711 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1712 return ld;
1713 }
1714
1715 //-------------------------set_arguments_for_java_call-------------------------
1716 // Arguments (pre-popped from the stack) are taken from the JVMS.
1717 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1718 // Add the call arguments:
1719 uint nargs = call->method()->arg_size();
1720 for (uint i = 0; i < nargs; i++) {
1721 Node* arg = argument(i);
1722 call->init_req(i + TypeFunc::Parms, arg);
1723 }
1724 }
1725
1726 //---------------------------set_edges_for_java_call---------------------------
1727 // Connect a newly created call into the current JVMS.
1728 // A return value node (if any) is returned from set_edges_for_java_call.
1729 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1730
1731 // Add the predefined inputs:
1732 call->init_req( TypeFunc::Control, control() );
1733 call->init_req( TypeFunc::I_O , i_o() );
1734 call->init_req( TypeFunc::Memory , reset_memory() );
1735 call->init_req( TypeFunc::FramePtr, frameptr() );
1736 call->init_req( TypeFunc::ReturnAdr, top() );
1737
1738 add_safepoint_edges(call, must_throw);
1739
1740 Node* xcall = _gvn.transform(call);
1741
1742 if (xcall == top()) {
1743 set_control(top());
1744 return;
1745 }
1746 assert(xcall == call, "call identity is stable");
1747
1748 // Re-use the current map to produce the result.
1749
1750 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1751 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1752 set_all_memory_call(xcall, separate_io_proj);
1753
1754 //return xcall; // no need, caller already has it
1755 }
1756
1757 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1758 if (stopped()) return top(); // maybe the call folded up?
1759
1760 // Capture the return value, if any.
1761 Node* ret;
1762 if (call->method() == NULL ||
1763 call->method()->return_type()->basic_type() == T_VOID)
1764 ret = top();
1765 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1766
1767 // Note: Since any out-of-line call can produce an exception,
1768 // we always insert an I_O projection from the call into the result.
1769
1770 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1771
1772 if (separate_io_proj) {
1773 // The caller requested separate projections be used by the fall
1774 // through and exceptional paths, so replace the projections for
1775 // the fall through path.
1776 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1777 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1778 }
1779 return ret;
1780 }
1781
1782 //--------------------set_predefined_input_for_runtime_call--------------------
1783 // Reading and setting the memory state is way conservative here.
1784 // The real problem is that I am not doing real Type analysis on memory,
1785 // so I cannot distinguish card mark stores from other stores. Across a GC
1786 // point the Store Barrier and the card mark memory has to agree. I cannot
1787 // have a card mark store and its barrier split across the GC point from
1788 // either above or below. Here I get that to happen by reading ALL of memory.
1789 // A better answer would be to separate out card marks from other memory.
1790 // For now, return the input memory state, so that it can be reused
1791 // after the call, if this call has restricted memory effects.
1792 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1793 // Set fixed predefined input arguments
1794 Node* memory = reset_memory();
1795 call->init_req( TypeFunc::Control, control() );
1796 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1797 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1798 call->init_req( TypeFunc::FramePtr, frameptr() );
1799 call->init_req( TypeFunc::ReturnAdr, top() );
1800 return memory;
1801 }
1802
1803 //-------------------set_predefined_output_for_runtime_call--------------------
1804 // Set control and memory (not i_o) from the call.
1805 // If keep_mem is not NULL, use it for the output state,
1806 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1807 // If hook_mem is NULL, this call produces no memory effects at all.
1808 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1809 // then only that memory slice is taken from the call.
1810 // In the last case, we must put an appropriate memory barrier before
1811 // the call, so as to create the correct anti-dependencies on loads
1812 // preceding the call.
1813 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1814 Node* keep_mem,
1815 const TypePtr* hook_mem) {
1816 // no i/o
1817 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1818 if (keep_mem) {
1819 // First clone the existing memory state
1820 set_all_memory(keep_mem);
1821 if (hook_mem != NULL) {
1822 // Make memory for the call
1823 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1824 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1825 // We also use hook_mem to extract specific effects from arraycopy stubs.
1826 set_memory(mem, hook_mem);
1827 }
1828 // ...else the call has NO memory effects.
1829
1830 // Make sure the call advertises its memory effects precisely.
1831 // This lets us build accurate anti-dependences in gcm.cpp.
1832 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1833 "call node must be constructed correctly");
1834 } else {
1835 assert(hook_mem == NULL, "");
1836 // This is not a "slow path" call; all memory comes from the call.
1837 set_all_memory_call(call);
1838 }
1839 }
1840
1841
1842 // Replace the call with the current state of the kit.
1843 void GraphKit::replace_call(CallNode* call, Node* result) {
1844 JVMState* ejvms = NULL;
1845 if (has_exceptions()) {
1846 ejvms = transfer_exceptions_into_jvms();
1847 }
1848
1849 SafePointNode* final_state = stop();
1850
1851 // Find all the needed outputs of this call
1852 CallProjections callprojs;
1853 call->extract_projections(&callprojs, true);
1854
1855 Node* init_mem = call->in(TypeFunc::Memory);
1856 Node* final_mem = final_state->in(TypeFunc::Memory);
1857 Node* final_ctl = final_state->in(TypeFunc::Control);
1858 Node* final_io = final_state->in(TypeFunc::I_O);
1859
1860 // Replace all the old call edges with the edges from the inlining result
1861 if (callprojs.fallthrough_catchproj != NULL) {
1862 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1863 }
1864 if (callprojs.fallthrough_memproj != NULL) {
1865 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1866 }
1867 if (callprojs.fallthrough_ioproj != NULL) {
1868 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1869 }
1870
1871 // Replace the result with the new result if it exists and is used
1872 if (callprojs.resproj != NULL && result != NULL) {
1873 C->gvn_replace_by(callprojs.resproj, result);
1874 }
1875
1876 if (ejvms == NULL) {
1877 // No exception edges to simply kill off those paths
1878 if (callprojs.catchall_catchproj != NULL) {
1879 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1880 }
1881 if (callprojs.catchall_memproj != NULL) {
1882 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1883 }
1884 if (callprojs.catchall_ioproj != NULL) {
1885 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1886 }
1887 // Replace the old exception object with top
1888 if (callprojs.exobj != NULL) {
1889 C->gvn_replace_by(callprojs.exobj, C->top());
1890 }
1891 } else {
1892 GraphKit ekit(ejvms);
1893
1894 // Load my combined exception state into the kit, with all phis transformed:
1895 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1896
1897 Node* ex_oop = ekit.use_exception_state(ex_map);
1898 if (callprojs.catchall_catchproj != NULL) {
1899 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1900 }
1901 if (callprojs.catchall_memproj != NULL) {
1902 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1903 }
1904 if (callprojs.catchall_ioproj != NULL) {
1905 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1906 }
1907
1908 // Replace the old exception object with the newly created one
1909 if (callprojs.exobj != NULL) {
1910 C->gvn_replace_by(callprojs.exobj, ex_oop);
1911 }
1912 }
1913
1914 // Disconnect the call from the graph
1915 call->disconnect_inputs(NULL, C);
1916 C->gvn_replace_by(call, C->top());
1917
1918 // Clean up any MergeMems that feed other MergeMems since the
1919 // optimizer doesn't like that.
1920 if (final_mem->is_MergeMem()) {
1921 Node_List wl;
1922 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1923 Node* m = i.get();
1924 if (m->is_MergeMem() && !wl.contains(m)) {
1925 wl.push(m);
1926 }
1927 }
1928 while (wl.size() > 0) {
1929 _gvn.transform(wl.pop());
1930 }
1931 }
1932 }
1933
1934
1935 //------------------------------increment_counter------------------------------
1936 // for statistics: increment a VM counter by 1
1937
1938 void GraphKit::increment_counter(address counter_addr) {
1939 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1940 increment_counter(adr1);
1941 }
1942
1943 void GraphKit::increment_counter(Node* counter_addr) {
1944 int adr_type = Compile::AliasIdxRaw;
1945 Node* ctrl = control();
1946 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type);
1947 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1948 store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type );
1949 }
1950
1951
1952 //------------------------------uncommon_trap----------------------------------
1953 // Bail out to the interpreter in mid-method. Implemented by calling the
1954 // uncommon_trap blob. This helper function inserts a runtime call with the
1955 // right debug info.
1956 void GraphKit::uncommon_trap(int trap_request,
1957 ciKlass* klass, const char* comment,
1958 bool must_throw,
1959 bool keep_exact_action) {
1960 if (failing()) stop();
1961 if (stopped()) return; // trap reachable?
1962
1963 // Note: If ProfileTraps is true, and if a deopt. actually
1964 // occurs here, the runtime will make sure an MDO exists. There is
1965 // no need to call method()->ensure_method_data() at this point.
1966
1967 // Set the stack pointer to the right value for reexecution:
1968 set_sp(reexecute_sp());
1969
1970 #ifdef ASSERT
1971 if (!must_throw) {
1972 // Make sure the stack has at least enough depth to execute
1973 // the current bytecode.
1974 int inputs, ignored_depth;
1975 if (compute_stack_effects(inputs, ignored_depth)) {
1976 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1977 Bytecodes::name(java_bc()), sp(), inputs));
1978 }
1979 }
1980 #endif
1981
1982 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1983 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1984
1985 switch (action) {
1986 case Deoptimization::Action_maybe_recompile:
1987 case Deoptimization::Action_reinterpret:
1988 // Temporary fix for 6529811 to allow virtual calls to be sure they
1989 // get the chance to go from mono->bi->mega
1990 if (!keep_exact_action &&
1991 Deoptimization::trap_request_index(trap_request) < 0 &&
1992 too_many_recompiles(reason)) {
1993 // This BCI is causing too many recompilations.
1994 action = Deoptimization::Action_none;
1995 trap_request = Deoptimization::make_trap_request(reason, action);
1996 } else {
1997 C->set_trap_can_recompile(true);
1998 }
1999 break;
2000 case Deoptimization::Action_make_not_entrant:
2001 C->set_trap_can_recompile(true);
2002 break;
2003 #ifdef ASSERT
2004 case Deoptimization::Action_none:
2005 case Deoptimization::Action_make_not_compilable:
2006 break;
2007 default:
2008 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
2009 break;
2010 #endif
2011 }
2012
2013 if (TraceOptoParse) {
2014 char buf[100];
2015 tty->print_cr("Uncommon trap %s at bci:%d",
2016 Deoptimization::format_trap_request(buf, sizeof(buf),
2017 trap_request), bci());
2018 }
2019
2020 CompileLog* log = C->log();
2021 if (log != NULL) {
2022 int kid = (klass == NULL)? -1: log->identify(klass);
2023 log->begin_elem("uncommon_trap bci='%d'", bci());
2024 char buf[100];
2025 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2026 trap_request));
2027 if (kid >= 0) log->print(" klass='%d'", kid);
2028 if (comment != NULL) log->print(" comment='%s'", comment);
2029 log->end_elem();
2030 }
2031
2032 // Make sure any guarding test views this path as very unlikely
2033 Node *i0 = control()->in(0);
2034 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2035 IfNode *iff = i0->as_If();
2036 float f = iff->_prob; // Get prob
2037 if (control()->Opcode() == Op_IfTrue) {
2038 if (f > PROB_UNLIKELY_MAG(4))
2039 iff->_prob = PROB_MIN;
2040 } else {
2041 if (f < PROB_LIKELY_MAG(4))
2042 iff->_prob = PROB_MAX;
2043 }
2044 }
2045
2046 // Clear out dead values from the debug info.
2047 kill_dead_locals();
2048
2049 // Now insert the uncommon trap subroutine call
2050 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2051 const TypePtr* no_memory_effects = NULL;
2052 // Pass the index of the class to be loaded
2053 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2054 (must_throw ? RC_MUST_THROW : 0),
2055 OptoRuntime::uncommon_trap_Type(),
2056 call_addr, "uncommon_trap", no_memory_effects,
2057 intcon(trap_request));
2058 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2059 "must extract request correctly from the graph");
2060 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2061
2062 call->set_req(TypeFunc::ReturnAdr, returnadr());
2063 // The debug info is the only real input to this call.
2064
2065 // Halt-and-catch fire here. The above call should never return!
2066 HaltNode* halt = new(C) HaltNode(control(), frameptr());
2067 _gvn.set_type_bottom(halt);
2068 root()->add_req(halt);
2069
2070 stop_and_kill_map();
2071 }
2072
2073
2074 //--------------------------just_allocated_object------------------------------
2075 // Report the object that was just allocated.
2076 // It must be the case that there are no intervening safepoints.
2077 // We use this to determine if an object is so "fresh" that
2078 // it does not require card marks.
2079 Node* GraphKit::just_allocated_object(Node* current_control) {
2080 if (C->recent_alloc_ctl() == current_control)
2081 return C->recent_alloc_obj();
2082 return NULL;
2083 }
2084
2085
2086 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2087 // (Note: TypeFunc::make has a cache that makes this fast.)
2088 const TypeFunc* tf = TypeFunc::make(dest_method);
2089 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2090 for (int j = 0; j < nargs; j++) {
2091 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2092 if( targ->basic_type() == T_DOUBLE ) {
2093 // If any parameters are doubles, they must be rounded before
2094 // the call, dstore_rounding does gvn.transform
2095 Node *arg = argument(j);
2096 arg = dstore_rounding(arg);
2097 set_argument(j, arg);
2098 }
2099 }
2100 }
2101
2102 /**
2103 * Record profiling data exact_kls for Node n with the type system so
2104 * that it can propagate it (speculation)
2105 *
2106 * @param n node that the type applies to
2107 * @param exact_kls type from profiling
2108 *
2109 * @return node with improved type
2110 */
2111 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) {
2112 const TypeOopPtr* current_type = _gvn.type(n)->isa_oopptr();
2113 assert(UseTypeSpeculation, "type speculation must be on");
2114 if (exact_kls != NULL &&
2115 // nothing to improve if type is already exact
2116 (current_type == NULL ||
2117 (!current_type->klass_is_exact() &&
2118 (current_type->speculative() == NULL ||
2119 !current_type->speculative()->klass_is_exact())))) {
2120 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2121 const TypeOopPtr* xtype = tklass->as_instance_type();
2122 assert(xtype->klass_is_exact(), "Should be exact");
2123
2124 // Build a type with a speculative type (what we think we know
2125 // about the type but will need a guard when we use it)
2126 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, xtype);
2127 // We're changing the type, we need a new cast node to carry the
2128 // new type. The new type depends on the control: what profiling
2129 // tells us is only valid from here as far as we can tell.
2130 Node* cast = new(C) CastPPNode(n, spec_type);
2131 cast->init_req(0, control());
2132 cast = _gvn.transform(cast);
2133 replace_in_map(n, cast);
2134 n = cast;
2135 }
2136 return n;
2137 }
2138
2139 /**
2140 * Record profiling data from receiver profiling at an invoke with the
2141 * type system so that it can propagate it (speculation)
2142 *
2143 * @param n receiver node
2144 *
2145 * @return node with improved type
2146 */
2147 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2148 if (!UseTypeSpeculation) {
2149 return n;
2150 }
2151 ciKlass* exact_kls = profile_has_unique_klass();
2152 return record_profile_for_speculation(n, exact_kls);
2153 }
2154
2155 /**
2156 * Record profiling data from argument profiling at an invoke with the
2157 * type system so that it can propagate it (speculation)
2158 *
2159 * @param dest_method target method for the call
2160 * @param bc what invoke bytecode is this?
2161 */
2162 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2163 if (!UseTypeSpeculation) {
2164 return;
2165 }
2166 const TypeFunc* tf = TypeFunc::make(dest_method);
2167 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2168 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2169 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2170 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2171 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2172 ciKlass* better_type = method()->argument_profiled_type(bci(), i);
2173 if (better_type != NULL) {
2174 record_profile_for_speculation(argument(j), better_type);
2175 }
2176 i++;
2177 }
2178 }
2179 }
2180
2181 /**
2182 * Record profiling data from parameter profiling at an invoke with
2183 * the type system so that it can propagate it (speculation)
2184 */
2185 void GraphKit::record_profiled_parameters_for_speculation() {
2186 if (!UseTypeSpeculation) {
2187 return;
2188 }
2189 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2190 if (_gvn.type(local(i))->isa_oopptr()) {
2191 ciKlass* better_type = method()->parameter_profiled_type(j);
2192 if (better_type != NULL) {
2193 record_profile_for_speculation(local(i), better_type);
2194 }
2195 j++;
2196 }
2197 }
2198 }
2199
2200 void GraphKit::round_double_result(ciMethod* dest_method) {
2201 // A non-strict method may return a double value which has an extended
2202 // exponent, but this must not be visible in a caller which is 'strict'
2203 // If a strict caller invokes a non-strict callee, round a double result
2204
2205 BasicType result_type = dest_method->return_type()->basic_type();
2206 assert( method() != NULL, "must have caller context");
2207 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2208 // Destination method's return value is on top of stack
2209 // dstore_rounding() does gvn.transform
2210 Node *result = pop_pair();
2211 result = dstore_rounding(result);
2212 push_pair(result);
2213 }
2214 }
2215
2216 // rounding for strict float precision conformance
2217 Node* GraphKit::precision_rounding(Node* n) {
2218 return UseStrictFP && _method->flags().is_strict()
2219 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2220 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2221 : n;
2222 }
2223
2224 // rounding for strict double precision conformance
2225 Node* GraphKit::dprecision_rounding(Node *n) {
2226 return UseStrictFP && _method->flags().is_strict()
2227 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2228 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2229 : n;
2230 }
2231
2232 // rounding for non-strict double stores
2233 Node* GraphKit::dstore_rounding(Node* n) {
2234 return Matcher::strict_fp_requires_explicit_rounding
2235 && UseSSE <= 1
2236 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2237 : n;
2238 }
2239
2240 //=============================================================================
2241 // Generate a fast path/slow path idiom. Graph looks like:
2242 // [foo] indicates that 'foo' is a parameter
2243 //
2244 // [in] NULL
2245 // \ /
2246 // CmpP
2247 // Bool ne
2248 // If
2249 // / \
2250 // True False-<2>
2251 // / |
2252 // / cast_not_null
2253 // Load | | ^
2254 // [fast_test] | |
2255 // gvn to opt_test | |
2256 // / \ | <1>
2257 // True False |
2258 // | \\ |
2259 // [slow_call] \[fast_result]
2260 // Ctl Val \ \
2261 // | \ \
2262 // Catch <1> \ \
2263 // / \ ^ \ \
2264 // Ex No_Ex | \ \
2265 // | \ \ | \ <2> \
2266 // ... \ [slow_res] | | \ [null_result]
2267 // \ \--+--+--- | |
2268 // \ | / \ | /
2269 // --------Region Phi
2270 //
2271 //=============================================================================
2272 // Code is structured as a series of driver functions all called 'do_XXX' that
2273 // call a set of helper functions. Helper functions first, then drivers.
2274
2275 //------------------------------null_check_oop---------------------------------
2276 // Null check oop. Set null-path control into Region in slot 3.
2277 // Make a cast-not-nullness use the other not-null control. Return cast.
2278 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2279 bool never_see_null, bool safe_for_replace) {
2280 // Initial NULL check taken path
2281 (*null_control) = top();
2282 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2283
2284 // Generate uncommon_trap:
2285 if (never_see_null && (*null_control) != top()) {
2286 // If we see an unexpected null at a check-cast we record it and force a
2287 // recompile; the offending check-cast will be compiled to handle NULLs.
2288 // If we see more than one offending BCI, then all checkcasts in the
2289 // method will be compiled to handle NULLs.
2290 PreserveJVMState pjvms(this);
2291 set_control(*null_control);
2292 replace_in_map(value, null());
2293 uncommon_trap(Deoptimization::Reason_null_check,
2294 Deoptimization::Action_make_not_entrant);
2295 (*null_control) = top(); // NULL path is dead
2296 }
2297 if ((*null_control) == top() && safe_for_replace) {
2298 replace_in_map(value, cast);
2299 }
2300
2301 // Cast away null-ness on the result
2302 return cast;
2303 }
2304
2305 //------------------------------opt_iff----------------------------------------
2306 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2307 // Return slow-path control.
2308 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2309 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2310
2311 // Fast path taken; set region slot 2
2312 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2313 region->init_req(2,fast_taken); // Capture fast-control
2314
2315 // Fast path not-taken, i.e. slow path
2316 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2317 return slow_taken;
2318 }
2319
2320 //-----------------------------make_runtime_call-------------------------------
2321 Node* GraphKit::make_runtime_call(int flags,
2322 const TypeFunc* call_type, address call_addr,
2323 const char* call_name,
2324 const TypePtr* adr_type,
2325 // The following parms are all optional.
2326 // The first NULL ends the list.
2327 Node* parm0, Node* parm1,
2328 Node* parm2, Node* parm3,
2329 Node* parm4, Node* parm5,
2330 Node* parm6, Node* parm7) {
2331 // Slow-path call
2332 bool is_leaf = !(flags & RC_NO_LEAF);
2333 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2334 if (call_name == NULL) {
2335 assert(!is_leaf, "must supply name for leaf");
2336 call_name = OptoRuntime::stub_name(call_addr);
2337 }
2338 CallNode* call;
2339 if (!is_leaf) {
2340 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2341 bci(), adr_type);
2342 } else if (flags & RC_NO_FP) {
2343 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2344 } else {
2345 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2346 }
2347
2348 // The following is similar to set_edges_for_java_call,
2349 // except that the memory effects of the call are restricted to AliasIdxRaw.
2350
2351 // Slow path call has no side-effects, uses few values
2352 bool wide_in = !(flags & RC_NARROW_MEM);
2353 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2354
2355 Node* prev_mem = NULL;
2356 if (wide_in) {
2357 prev_mem = set_predefined_input_for_runtime_call(call);
2358 } else {
2359 assert(!wide_out, "narrow in => narrow out");
2360 Node* narrow_mem = memory(adr_type);
2361 prev_mem = reset_memory();
2362 map()->set_memory(narrow_mem);
2363 set_predefined_input_for_runtime_call(call);
2364 }
2365
2366 // Hook each parm in order. Stop looking at the first NULL.
2367 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2368 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2369 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2370 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2371 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2372 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2373 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2374 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2375 /* close each nested if ===> */ } } } } } } } }
2376 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2377
2378 if (!is_leaf) {
2379 // Non-leaves can block and take safepoints:
2380 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2381 }
2382 // Non-leaves can throw exceptions:
2383 if (has_io) {
2384 call->set_req(TypeFunc::I_O, i_o());
2385 }
2386
2387 if (flags & RC_UNCOMMON) {
2388 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2389 // (An "if" probability corresponds roughly to an unconditional count.
2390 // Sort of.)
2391 call->set_cnt(PROB_UNLIKELY_MAG(4));
2392 }
2393
2394 Node* c = _gvn.transform(call);
2395 assert(c == call, "cannot disappear");
2396
2397 if (wide_out) {
2398 // Slow path call has full side-effects.
2399 set_predefined_output_for_runtime_call(call);
2400 } else {
2401 // Slow path call has few side-effects, and/or sets few values.
2402 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2403 }
2404
2405 if (has_io) {
2406 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2407 }
2408 return call;
2409
2410 }
2411
2412 //------------------------------merge_memory-----------------------------------
2413 // Merge memory from one path into the current memory state.
2414 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2415 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2416 Node* old_slice = mms.force_memory();
2417 Node* new_slice = mms.memory2();
2418 if (old_slice != new_slice) {
2419 PhiNode* phi;
2420 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2421 phi = new_slice->as_Phi();
2422 #ifdef ASSERT
2423 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2424 old_slice = old_slice->in(new_path);
2425 // Caller is responsible for ensuring that any pre-existing
2426 // phis are already aware of old memory.
2427 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2428 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2429 #endif
2430 mms.set_memory(phi);
2431 } else {
2432 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2433 _gvn.set_type(phi, Type::MEMORY);
2434 phi->set_req(new_path, new_slice);
2435 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2436 }
2437 }
2438 }
2439 }
2440
2441 //------------------------------make_slow_call_ex------------------------------
2442 // Make the exception handler hookups for the slow call
2443 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2444 if (stopped()) return;
2445
2446 // Make a catch node with just two handlers: fall-through and catch-all
2447 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2448 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2449 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2450 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2451
2452 { PreserveJVMState pjvms(this);
2453 set_control(excp);
2454 set_i_o(i_o);
2455
2456 if (excp != top()) {
2457 // Create an exception state also.
2458 // Use an exact type if the caller has specified a specific exception.
2459 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2460 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2461 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2462 }
2463 }
2464
2465 // Get the no-exception control from the CatchNode.
2466 set_control(norm);
2467 }
2468
2469
2470 //-------------------------------gen_subtype_check-----------------------------
2471 // Generate a subtyping check. Takes as input the subtype and supertype.
2472 // Returns 2 values: sets the default control() to the true path and returns
2473 // the false path. Only reads invariant memory; sets no (visible) memory.
2474 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2475 // but that's not exposed to the optimizer. This call also doesn't take in an
2476 // Object; if you wish to check an Object you need to load the Object's class
2477 // prior to coming here.
2478 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2479 // Fast check for identical types, perhaps identical constants.
2480 // The types can even be identical non-constants, in cases
2481 // involving Array.newInstance, Object.clone, etc.
2482 if (subklass == superklass)
2483 return top(); // false path is dead; no test needed.
2484
2485 if (_gvn.type(superklass)->singleton()) {
2486 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2487 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2488
2489 // In the common case of an exact superklass, try to fold up the
2490 // test before generating code. You may ask, why not just generate
2491 // the code and then let it fold up? The answer is that the generated
2492 // code will necessarily include null checks, which do not always
2493 // completely fold away. If they are also needless, then they turn
2494 // into a performance loss. Example:
2495 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2496 // Here, the type of 'fa' is often exact, so the store check
2497 // of fa[1]=x will fold up, without testing the nullness of x.
2498 switch (static_subtype_check(superk, subk)) {
2499 case SSC_always_false:
2500 {
2501 Node* always_fail = control();
2502 set_control(top());
2503 return always_fail;
2504 }
2505 case SSC_always_true:
2506 return top();
2507 case SSC_easy_test:
2508 {
2509 // Just do a direct pointer compare and be done.
2510 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2511 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2512 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2513 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2514 return _gvn.transform( new(C) IfFalseNode(iff) );
2515 }
2516 case SSC_full_test:
2517 break;
2518 default:
2519 ShouldNotReachHere();
2520 }
2521 }
2522
2523 // %%% Possible further optimization: Even if the superklass is not exact,
2524 // if the subklass is the unique subtype of the superklass, the check
2525 // will always succeed. We could leave a dependency behind to ensure this.
2526
2527 // First load the super-klass's check-offset
2528 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2529 Node *chk_off = _gvn.transform( new (C) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2530 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2531 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2532
2533 // Load from the sub-klass's super-class display list, or a 1-word cache of
2534 // the secondary superclass list, or a failing value with a sentinel offset
2535 // if the super-klass is an interface or exceptionally deep in the Java
2536 // hierarchy and we have to scan the secondary superclass list the hard way.
2537 // Worst-case type is a little odd: NULL is allowed as a result (usually
2538 // klass loads can never produce a NULL).
2539 Node *chk_off_X = ConvI2X(chk_off);
2540 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2541 // For some types like interfaces the following loadKlass is from a 1-word
2542 // cache which is mutable so can't use immutable memory. Other
2543 // types load from the super-class display table which is immutable.
2544 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2545 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2546
2547 // Compile speed common case: ARE a subtype and we canNOT fail
2548 if( superklass == nkls )
2549 return top(); // false path is dead; no test needed.
2550
2551 // See if we get an immediate positive hit. Happens roughly 83% of the
2552 // time. Test to see if the value loaded just previously from the subklass
2553 // is exactly the superklass.
2554 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2555 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2556 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2557 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2558 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2559
2560 // Compile speed common case: Check for being deterministic right now. If
2561 // chk_off is a constant and not equal to cacheoff then we are NOT a
2562 // subklass. In this case we need exactly the 1 test above and we can
2563 // return those results immediately.
2564 if (!might_be_cache) {
2565 Node* not_subtype_ctrl = control();
2566 set_control(iftrue1); // We need exactly the 1 test above
2567 return not_subtype_ctrl;
2568 }
2569
2570 // Gather the various success & failures here
2571 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2572 record_for_igvn(r_ok_subtype);
2573 RegionNode *r_not_subtype = new (C) RegionNode(3);
2574 record_for_igvn(r_not_subtype);
2575
2576 r_ok_subtype->init_req(1, iftrue1);
2577
2578 // Check for immediate negative hit. Happens roughly 11% of the time (which
2579 // is roughly 63% of the remaining cases). Test to see if the loaded
2580 // check-offset points into the subklass display list or the 1-element
2581 // cache. If it points to the display (and NOT the cache) and the display
2582 // missed then it's not a subtype.
2583 Node *cacheoff = _gvn.intcon(cacheoff_con);
2584 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2585 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2586 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2587 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2588 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2589
2590 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2591 // No performance impact (too rare) but allows sharing of secondary arrays
2592 // which has some footprint reduction.
2593 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2594 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2595 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2596 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2597 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2598
2599 // -- Roads not taken here: --
2600 // We could also have chosen to perform the self-check at the beginning
2601 // of this code sequence, as the assembler does. This would not pay off
2602 // the same way, since the optimizer, unlike the assembler, can perform
2603 // static type analysis to fold away many successful self-checks.
2604 // Non-foldable self checks work better here in second position, because
2605 // the initial primary superclass check subsumes a self-check for most
2606 // types. An exception would be a secondary type like array-of-interface,
2607 // which does not appear in its own primary supertype display.
2608 // Finally, we could have chosen to move the self-check into the
2609 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2610 // dependent manner. But it is worthwhile to have the check here,
2611 // where it can be perhaps be optimized. The cost in code space is
2612 // small (register compare, branch).
2613
2614 // Now do a linear scan of the secondary super-klass array. Again, no real
2615 // performance impact (too rare) but it's gotta be done.
2616 // Since the code is rarely used, there is no penalty for moving it
2617 // out of line, and it can only improve I-cache density.
2618 // The decision to inline or out-of-line this final check is platform
2619 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2620 Node* psc = _gvn.transform(
2621 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2622
2623 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2624 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2625 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2626 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2627 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2628
2629 // Return false path; set default control to true path.
2630 set_control( _gvn.transform(r_ok_subtype) );
2631 return _gvn.transform(r_not_subtype);
2632 }
2633
2634 //----------------------------static_subtype_check-----------------------------
2635 // Shortcut important common cases when superklass is exact:
2636 // (0) superklass is java.lang.Object (can occur in reflective code)
2637 // (1) subklass is already limited to a subtype of superklass => always ok
2638 // (2) subklass does not overlap with superklass => always fail
2639 // (3) superklass has NO subtypes and we can check with a simple compare.
2640 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2641 if (StressReflectiveCode) {
2642 return SSC_full_test; // Let caller generate the general case.
2643 }
2644
2645 if (superk == env()->Object_klass()) {
2646 return SSC_always_true; // (0) this test cannot fail
2647 }
2648
2649 ciType* superelem = superk;
2650 if (superelem->is_array_klass())
2651 superelem = superelem->as_array_klass()->base_element_type();
2652
2653 if (!subk->is_interface()) { // cannot trust static interface types yet
2654 if (subk->is_subtype_of(superk)) {
2655 return SSC_always_true; // (1) false path dead; no dynamic test needed
2656 }
2657 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2658 !superk->is_subtype_of(subk)) {
2659 return SSC_always_false;
2660 }
2661 }
2662
2663 // If casting to an instance klass, it must have no subtypes
2664 if (superk->is_interface()) {
2665 // Cannot trust interfaces yet.
2666 // %%% S.B. superk->nof_implementors() == 1
2667 } else if (superelem->is_instance_klass()) {
2668 ciInstanceKlass* ik = superelem->as_instance_klass();
2669 if (!ik->has_subklass() && !ik->is_interface()) {
2670 if (!ik->is_final()) {
2671 // Add a dependency if there is a chance of a later subclass.
2672 C->dependencies()->assert_leaf_type(ik);
2673 }
2674 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2675 }
2676 } else {
2677 // A primitive array type has no subtypes.
2678 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2679 }
2680
2681 return SSC_full_test;
2682 }
2683
2684 // Profile-driven exact type check:
2685 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2686 float prob,
2687 Node* *casted_receiver) {
2688 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2689 Node* recv_klass = load_object_klass(receiver);
2690 Node* want_klass = makecon(tklass);
2691 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2692 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2693 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2694 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2695 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2696
2697 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2698 assert(recv_xtype->klass_is_exact(), "");
2699
2700 // Subsume downstream occurrences of receiver with a cast to
2701 // recv_xtype, since now we know what the type will be.
2702 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2703 (*casted_receiver) = _gvn.transform(cast);
2704 // (User must make the replace_in_map call.)
2705
2706 return fail;
2707 }
2708
2709
2710 //------------------------------seems_never_null-------------------------------
2711 // Use null_seen information if it is available from the profile.
2712 // If we see an unexpected null at a type check we record it and force a
2713 // recompile; the offending check will be recompiled to handle NULLs.
2714 // If we see several offending BCIs, then all checks in the
2715 // method will be recompiled.
2716 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2717 if (UncommonNullCast // Cutout for this technique
2718 && obj != null() // And not the -Xcomp stupid case?
2719 && !too_many_traps(Deoptimization::Reason_null_check)
2720 ) {
2721 if (data == NULL)
2722 // Edge case: no mature data. Be optimistic here.
2723 return true;
2724 // If the profile has not seen a null, assume it won't happen.
2725 assert(java_bc() == Bytecodes::_checkcast ||
2726 java_bc() == Bytecodes::_instanceof ||
2727 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2728 return !data->as_BitData()->null_seen();
2729 }
2730 return false;
2731 }
2732
2733 //------------------------maybe_cast_profiled_receiver-------------------------
2734 // If the profile has seen exactly one type, narrow to exactly that type.
2735 // Subsequent type checks will always fold up.
2736 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2737 ciKlass* require_klass,
2738 ciKlass* spec_klass,
2739 bool safe_for_replace) {
2740 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2741
2742 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check;
2743
2744 // Make sure we haven't already deoptimized from this tactic.
2745 if (too_many_traps(reason))
2746 return NULL;
2747
2748 // (No, this isn't a call, but it's enough like a virtual call
2749 // to use the same ciMethod accessor to get the profile info...)
2750 // If we have a speculative type use it instead of profiling (which
2751 // may not help us)
2752 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2753 if (exact_kls != NULL) {// no cast failures here
2754 if (require_klass == NULL ||
2755 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2756 // If we narrow the type to match what the type profile sees or
2757 // the speculative type, we can then remove the rest of the
2758 // cast.
2759 // This is a win, even if the exact_kls is very specific,
2760 // because downstream operations, such as method calls,
2761 // will often benefit from the sharper type.
2762 Node* exact_obj = not_null_obj; // will get updated in place...
2763 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2764 &exact_obj);
2765 { PreserveJVMState pjvms(this);
2766 set_control(slow_ctl);
2767 uncommon_trap(reason,
2768 Deoptimization::Action_maybe_recompile);
2769 }
2770 if (safe_for_replace) {
2771 replace_in_map(not_null_obj, exact_obj);
2772 }
2773 return exact_obj;
2774 }
2775 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2776 }
2777
2778 return NULL;
2779 }
2780
2781 /**
2782 * Cast obj to type and emit guard unless we had too many traps here
2783 * already
2784 *
2785 * @param obj node being casted
2786 * @param type type to cast the node to
2787 * @param not_null true if we know node cannot be null
2788 */
2789 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2790 ciKlass* type,
2791 bool not_null) {
2792 // type == NULL if profiling tells us this object is always null
2793 if (type != NULL) {
2794 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2795 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check;
2796 if (!too_many_traps(null_reason) &&
2797 !too_many_traps(class_reason)) {
2798 Node* not_null_obj = NULL;
2799 // not_null is true if we know the object is not null and
2800 // there's no need for a null check
2801 if (!not_null) {
2802 Node* null_ctl = top();
2803 not_null_obj = null_check_oop(obj, &null_ctl, true, true);
2804 assert(null_ctl->is_top(), "no null control here");
2805 } else {
2806 not_null_obj = obj;
2807 }
2808
2809 Node* exact_obj = not_null_obj;
2810 ciKlass* exact_kls = type;
2811 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2812 &exact_obj);
2813 {
2814 PreserveJVMState pjvms(this);
2815 set_control(slow_ctl);
2816 uncommon_trap(class_reason,
2817 Deoptimization::Action_maybe_recompile);
2818 }
2819 replace_in_map(not_null_obj, exact_obj);
2820 obj = exact_obj;
2821 }
2822 } else {
2823 if (!too_many_traps(Deoptimization::Reason_null_assert)) {
2824 Node* exact_obj = null_assert(obj);
2825 replace_in_map(obj, exact_obj);
2826 obj = exact_obj;
2827 }
2828 }
2829 return obj;
2830 }
2831
2832 //-------------------------------gen_instanceof--------------------------------
2833 // Generate an instance-of idiom. Used by both the instance-of bytecode
2834 // and the reflective instance-of call.
2835 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2836 kill_dead_locals(); // Benefit all the uncommon traps
2837 assert( !stopped(), "dead parse path should be checked in callers" );
2838 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2839 "must check for not-null not-dead klass in callers");
2840
2841 // Make the merge point
2842 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2843 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2844 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2845 C->set_has_split_ifs(true); // Has chance for split-if optimization
2846
2847 ciProfileData* data = NULL;
2848 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2849 data = method()->method_data()->bci_to_data(bci());
2850 }
2851 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2852 && seems_never_null(obj, data));
2853
2854 // Null check; get casted pointer; set region slot 3
2855 Node* null_ctl = top();
2856 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2857
2858 // If not_null_obj is dead, only null-path is taken
2859 if (stopped()) { // Doing instance-of on a NULL?
2860 set_control(null_ctl);
2861 return intcon(0);
2862 }
2863 region->init_req(_null_path, null_ctl);
2864 phi ->init_req(_null_path, intcon(0)); // Set null path value
2865 if (null_ctl == top()) {
2866 // Do this eagerly, so that pattern matches like is_diamond_phi
2867 // will work even during parsing.
2868 assert(_null_path == PATH_LIMIT-1, "delete last");
2869 region->del_req(_null_path);
2870 phi ->del_req(_null_path);
2871 }
2872
2873 // Do we know the type check always succeed?
2874 bool known_statically = false;
2875 if (_gvn.type(superklass)->singleton()) {
2876 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2877 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2878 if (subk != NULL && subk->is_loaded()) {
2879 int static_res = static_subtype_check(superk, subk);
2880 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false);
2881 }
2882 }
2883
2884 if (known_statically && UseTypeSpeculation) {
2885 // If we know the type check always succeeds then we don't use the
2886 // profiling data at this bytecode. Don't lose it, feed it to the
2887 // type system as a speculative type.
2888 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2889 } else {
2890 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2891 // We may not have profiling here or it may not help us. If we
2892 // have a speculative type use it to perform an exact cast.
2893 ciKlass* spec_obj_type = obj_type->speculative_type();
2894 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2895 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2896 if (stopped()) { // Profile disagrees with this path.
2897 set_control(null_ctl); // Null is the only remaining possibility.
2898 return intcon(0);
2899 }
2900 if (cast_obj != NULL) {
2901 not_null_obj = cast_obj;
2902 }
2903 }
2904 }
2905
2906 // Load the object's klass
2907 Node* obj_klass = load_object_klass(not_null_obj);
2908
2909 // Generate the subtype check
2910 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2911
2912 // Plug in the success path to the general merge in slot 1.
2913 region->init_req(_obj_path, control());
2914 phi ->init_req(_obj_path, intcon(1));
2915
2916 // Plug in the failing path to the general merge in slot 2.
2917 region->init_req(_fail_path, not_subtype_ctrl);
2918 phi ->init_req(_fail_path, intcon(0));
2919
2920 // Return final merged results
2921 set_control( _gvn.transform(region) );
2922 record_for_igvn(region);
2923 return _gvn.transform(phi);
2924 }
2925
2926 //-------------------------------gen_checkcast---------------------------------
2927 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2928 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2929 // uncommon-trap paths work. Adjust stack after this call.
2930 // If failure_control is supplied and not null, it is filled in with
2931 // the control edge for the cast failure. Otherwise, an appropriate
2932 // uncommon trap or exception is thrown.
2933 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2934 Node* *failure_control) {
2935 kill_dead_locals(); // Benefit all the uncommon traps
2936 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2937 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2938
2939 // Fast cutout: Check the case that the cast is vacuously true.
2940 // This detects the common cases where the test will short-circuit
2941 // away completely. We do this before we perform the null check,
2942 // because if the test is going to turn into zero code, we don't
2943 // want a residual null check left around. (Causes a slowdown,
2944 // for example, in some objArray manipulations, such as a[i]=a[j].)
2945 if (tk->singleton()) {
2946 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2947 if (objtp != NULL && objtp->klass() != NULL) {
2948 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2949 case SSC_always_true:
2950 // If we know the type check always succeed then we don't use
2951 // the profiling data at this bytecode. Don't lose it, feed it
2952 // to the type system as a speculative type.
2953 return record_profiled_receiver_for_speculation(obj);
2954 case SSC_always_false:
2955 // It needs a null check because a null will *pass* the cast check.
2956 // A non-null value will always produce an exception.
2957 return null_assert(obj);
2958 }
2959 }
2960 }
2961
2962 ciProfileData* data = NULL;
2963 bool safe_for_replace = false;
2964 if (failure_control == NULL) { // use MDO in regular case only
2965 assert(java_bc() == Bytecodes::_aastore ||
2966 java_bc() == Bytecodes::_checkcast,
2967 "interpreter profiles type checks only for these BCs");
2968 data = method()->method_data()->bci_to_data(bci());
2969 safe_for_replace = true;
2970 }
2971
2972 // Make the merge point
2973 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2974 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
2975 Node* phi = new (C) PhiNode(region, toop);
2976 C->set_has_split_ifs(true); // Has chance for split-if optimization
2977
2978 // Use null-cast information if it is available
2979 bool never_see_null = ((failure_control == NULL) // regular case only
2980 && seems_never_null(obj, data));
2981
2982 // Null check; get casted pointer; set region slot 3
2983 Node* null_ctl = top();
2984 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2985
2986 // If not_null_obj is dead, only null-path is taken
2987 if (stopped()) { // Doing instance-of on a NULL?
2988 set_control(null_ctl);
2989 return null();
2990 }
2991 region->init_req(_null_path, null_ctl);
2992 phi ->init_req(_null_path, null()); // Set null path value
2993 if (null_ctl == top()) {
2994 // Do this eagerly, so that pattern matches like is_diamond_phi
2995 // will work even during parsing.
2996 assert(_null_path == PATH_LIMIT-1, "delete last");
2997 region->del_req(_null_path);
2998 phi ->del_req(_null_path);
2999 }
3000
3001 Node* cast_obj = NULL;
3002 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3003 // We may not have profiling here or it may not help us. If we have
3004 // a speculative type use it to perform an exact cast.
3005 ciKlass* spec_obj_type = obj_type->speculative_type();
3006 if (spec_obj_type != NULL ||
3007 (data != NULL &&
3008 // Counter has never been decremented (due to cast failure).
3009 // ...This is a reasonable thing to expect. It is true of
3010 // all casts inserted by javac to implement generic types.
3011 data->as_CounterData()->count() >= 0)) {
3012 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3013 if (cast_obj != NULL) {
3014 if (failure_control != NULL) // failure is now impossible
3015 (*failure_control) = top();
3016 // adjust the type of the phi to the exact klass:
3017 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3018 }
3019 }
3020
3021 if (cast_obj == NULL) {
3022 // Load the object's klass
3023 Node* obj_klass = load_object_klass(not_null_obj);
3024
3025 // Generate the subtype check
3026 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3027
3028 // Plug in success path into the merge
3029 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
3030 not_null_obj, toop));
3031 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3032 if (failure_control == NULL) {
3033 if (not_subtype_ctrl != top()) { // If failure is possible
3034 PreserveJVMState pjvms(this);
3035 set_control(not_subtype_ctrl);
3036 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3037 }
3038 } else {
3039 (*failure_control) = not_subtype_ctrl;
3040 }
3041 }
3042
3043 region->init_req(_obj_path, control());
3044 phi ->init_req(_obj_path, cast_obj);
3045
3046 // A merge of NULL or Casted-NotNull obj
3047 Node* res = _gvn.transform(phi);
3048
3049 // Note I do NOT always 'replace_in_map(obj,result)' here.
3050 // if( tk->klass()->can_be_primary_super() )
3051 // This means that if I successfully store an Object into an array-of-String
3052 // I 'forget' that the Object is really now known to be a String. I have to
3053 // do this because we don't have true union types for interfaces - if I store
3054 // a Baz into an array-of-Interface and then tell the optimizer it's an
3055 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3056 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3057 // replace_in_map( obj, res );
3058
3059 // Return final merged results
3060 set_control( _gvn.transform(region) );
3061 record_for_igvn(region);
3062 return res;
3063 }
3064
3065 //------------------------------next_monitor-----------------------------------
3066 // What number should be given to the next monitor?
3067 int GraphKit::next_monitor() {
3068 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3069 int next = current + C->sync_stack_slots();
3070 // Keep the toplevel high water mark current:
3071 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3072 return current;
3073 }
3074
3075 //------------------------------insert_mem_bar---------------------------------
3076 // Memory barrier to avoid floating things around
3077 // The membar serves as a pinch point between both control and all memory slices.
3078 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3079 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3080 mb->init_req(TypeFunc::Control, control());
3081 mb->init_req(TypeFunc::Memory, reset_memory());
3082 Node* membar = _gvn.transform(mb);
3083 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3084 set_all_memory_call(membar);
3085 return membar;
3086 }
3087
3088 //-------------------------insert_mem_bar_volatile----------------------------
3089 // Memory barrier to avoid floating things around
3090 // The membar serves as a pinch point between both control and memory(alias_idx).
3091 // If you want to make a pinch point on all memory slices, do not use this
3092 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3093 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3094 // When Parse::do_put_xxx updates a volatile field, it appends a series
3095 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3096 // The first membar is on the same memory slice as the field store opcode.
3097 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3098 // All the other membars (for other volatile slices, including AliasIdxBot,
3099 // which stands for all unknown volatile slices) are control-dependent
3100 // on the first membar. This prevents later volatile loads or stores
3101 // from sliding up past the just-emitted store.
3102
3103 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3104 mb->set_req(TypeFunc::Control,control());
3105 if (alias_idx == Compile::AliasIdxBot) {
3106 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3107 } else {
3108 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3109 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3110 }
3111 Node* membar = _gvn.transform(mb);
3112 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3113 if (alias_idx == Compile::AliasIdxBot) {
3114 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
3115 } else {
3116 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
3117 }
3118 return membar;
3119 }
3120
3121 //------------------------------shared_lock------------------------------------
3122 // Emit locking code.
3123 FastLockNode* GraphKit::shared_lock(Node* obj) {
3124 // bci is either a monitorenter bc or InvocationEntryBci
3125 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3126 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3127
3128 if( !GenerateSynchronizationCode )
3129 return NULL; // Not locking things?
3130 if (stopped()) // Dead monitor?
3131 return NULL;
3132
3133 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3134
3135 // Box the stack location
3136 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
3137 Node* mem = reset_memory();
3138
3139 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
3140 if (PrintPreciseBiasedLockingStatistics) {
3141 // Create the counters for this fast lock.
3142 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3143 }
3144 // Add monitor to debug info for the slow path. If we block inside the
3145 // slow path and de-opt, we need the monitor hanging around
3146 map()->push_monitor( flock );
3147
3148 const TypeFunc *tf = LockNode::lock_type();
3149 LockNode *lock = new (C) LockNode(C, tf);
3150
3151 lock->init_req( TypeFunc::Control, control() );
3152 lock->init_req( TypeFunc::Memory , mem );
3153 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3154 lock->init_req( TypeFunc::FramePtr, frameptr() );
3155 lock->init_req( TypeFunc::ReturnAdr, top() );
3156
3157 lock->init_req(TypeFunc::Parms + 0, obj);
3158 lock->init_req(TypeFunc::Parms + 1, box);
3159 lock->init_req(TypeFunc::Parms + 2, flock);
3160 add_safepoint_edges(lock);
3161
3162 lock = _gvn.transform( lock )->as_Lock();
3163
3164 // lock has no side-effects, sets few values
3165 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3166
3167 insert_mem_bar(Op_MemBarAcquireLock);
3168
3169 // Add this to the worklist so that the lock can be eliminated
3170 record_for_igvn(lock);
3171
3172 #ifndef PRODUCT
3173 if (PrintLockStatistics) {
3174 // Update the counter for this lock. Don't bother using an atomic
3175 // operation since we don't require absolute accuracy.
3176 lock->create_lock_counter(map()->jvms());
3177 increment_counter(lock->counter()->addr());
3178 }
3179 #endif
3180
3181 return flock;
3182 }
3183
3184
3185 //------------------------------shared_unlock----------------------------------
3186 // Emit unlocking code.
3187 void GraphKit::shared_unlock(Node* box, Node* obj) {
3188 // bci is either a monitorenter bc or InvocationEntryBci
3189 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3190 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3191
3192 if( !GenerateSynchronizationCode )
3193 return;
3194 if (stopped()) { // Dead monitor?
3195 map()->pop_monitor(); // Kill monitor from debug info
3196 return;
3197 }
3198
3199 // Memory barrier to avoid floating things down past the locked region
3200 insert_mem_bar(Op_MemBarReleaseLock);
3201
3202 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3203 UnlockNode *unlock = new (C) UnlockNode(C, tf);
3204 uint raw_idx = Compile::AliasIdxRaw;
3205 unlock->init_req( TypeFunc::Control, control() );
3206 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3207 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3208 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3209 unlock->init_req( TypeFunc::ReturnAdr, top() );
3210
3211 unlock->init_req(TypeFunc::Parms + 0, obj);
3212 unlock->init_req(TypeFunc::Parms + 1, box);
3213 unlock = _gvn.transform(unlock)->as_Unlock();
3214
3215 Node* mem = reset_memory();
3216
3217 // unlock has no side-effects, sets few values
3218 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3219
3220 // Kill monitor from debug info
3221 map()->pop_monitor( );
3222 }
3223
3224 //-------------------------------get_layout_helper-----------------------------
3225 // If the given klass is a constant or known to be an array,
3226 // fetch the constant layout helper value into constant_value
3227 // and return (Node*)NULL. Otherwise, load the non-constant
3228 // layout helper value, and return the node which represents it.
3229 // This two-faced routine is useful because allocation sites
3230 // almost always feature constant types.
3231 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3232 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3233 if (!StressReflectiveCode && inst_klass != NULL) {
3234 ciKlass* klass = inst_klass->klass();
3235 bool xklass = inst_klass->klass_is_exact();
3236 if (xklass || klass->is_array_klass()) {
3237 jint lhelper = klass->layout_helper();
3238 if (lhelper != Klass::_lh_neutral_value) {
3239 constant_value = lhelper;
3240 return (Node*) NULL;
3241 }
3242 }
3243 }
3244 constant_value = Klass::_lh_neutral_value; // put in a known value
3245 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3246 return make_load(NULL, lhp, TypeInt::INT, T_INT);
3247 }
3248
3249 // We just put in an allocate/initialize with a big raw-memory effect.
3250 // Hook selected additional alias categories on the initialization.
3251 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3252 MergeMemNode* init_in_merge,
3253 Node* init_out_raw) {
3254 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3255 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3256
3257 Node* prevmem = kit.memory(alias_idx);
3258 init_in_merge->set_memory_at(alias_idx, prevmem);
3259 kit.set_memory(init_out_raw, alias_idx);
3260 }
3261
3262 //---------------------------set_output_for_allocation-------------------------
3263 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3264 const TypeOopPtr* oop_type) {
3265 int rawidx = Compile::AliasIdxRaw;
3266 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3267 add_safepoint_edges(alloc);
3268 Node* allocx = _gvn.transform(alloc);
3269 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3270 // create memory projection for i_o
3271 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3272 make_slow_call_ex(allocx, env()->Throwable_klass(), true);
3273
3274 // create a memory projection as for the normal control path
3275 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3276 set_memory(malloc, rawidx);
3277
3278 // a normal slow-call doesn't change i_o, but an allocation does
3279 // we create a separate i_o projection for the normal control path
3280 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3281 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3282
3283 // put in an initialization barrier
3284 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3285 rawoop)->as_Initialize();
3286 assert(alloc->initialization() == init, "2-way macro link must work");
3287 assert(init ->allocation() == alloc, "2-way macro link must work");
3288 {
3289 // Extract memory strands which may participate in the new object's
3290 // initialization, and source them from the new InitializeNode.
3291 // This will allow us to observe initializations when they occur,
3292 // and link them properly (as a group) to the InitializeNode.
3293 assert(init->in(InitializeNode::Memory) == malloc, "");
3294 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3295 init->set_req(InitializeNode::Memory, minit_in);
3296 record_for_igvn(minit_in); // fold it up later, if possible
3297 Node* minit_out = memory(rawidx);
3298 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3299 if (oop_type->isa_aryptr()) {
3300 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3301 int elemidx = C->get_alias_index(telemref);
3302 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3303 } else if (oop_type->isa_instptr()) {
3304 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3305 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3306 ciField* field = ik->nonstatic_field_at(i);
3307 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3308 continue; // do not bother to track really large numbers of fields
3309 // Find (or create) the alias category for this field:
3310 int fieldidx = C->alias_type(field)->index();
3311 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3312 }
3313 }
3314 }
3315
3316 // Cast raw oop to the real thing...
3317 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3318 javaoop = _gvn.transform(javaoop);
3319 C->set_recent_alloc(control(), javaoop);
3320 assert(just_allocated_object(control()) == javaoop, "just allocated");
3321
3322 #ifdef ASSERT
3323 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3324 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3325 "Ideal_allocation works");
3326 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3327 "Ideal_allocation works");
3328 if (alloc->is_AllocateArray()) {
3329 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3330 "Ideal_allocation works");
3331 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3332 "Ideal_allocation works");
3333 } else {
3334 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3335 }
3336 }
3337 #endif //ASSERT
3338
3339 return javaoop;
3340 }
3341
3342 //---------------------------new_instance--------------------------------------
3343 // This routine takes a klass_node which may be constant (for a static type)
3344 // or may be non-constant (for reflective code). It will work equally well
3345 // for either, and the graph will fold nicely if the optimizer later reduces
3346 // the type to a constant.
3347 // The optional arguments are for specialized use by intrinsics:
3348 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3349 // - If 'return_size_val', report the the total object size to the caller.
3350 Node* GraphKit::new_instance(Node* klass_node,
3351 Node* extra_slow_test,
3352 Node* *return_size_val) {
3353 // Compute size in doublewords
3354 // The size is always an integral number of doublewords, represented
3355 // as a positive bytewise size stored in the klass's layout_helper.
3356 // The layout_helper also encodes (in a low bit) the need for a slow path.
3357 jint layout_con = Klass::_lh_neutral_value;
3358 Node* layout_val = get_layout_helper(klass_node, layout_con);
3359 int layout_is_con = (layout_val == NULL);
3360
3361 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3362 // Generate the initial go-slow test. It's either ALWAYS (return a
3363 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3364 // case) a computed value derived from the layout_helper.
3365 Node* initial_slow_test = NULL;
3366 if (layout_is_con) {
3367 assert(!StressReflectiveCode, "stress mode does not use these paths");
3368 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3369 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3370
3371 } else { // reflective case
3372 // This reflective path is used by Unsafe.allocateInstance.
3373 // (It may be stress-tested by specifying StressReflectiveCode.)
3374 // Basically, we want to get into the VM is there's an illegal argument.
3375 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3376 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3377 if (extra_slow_test != intcon(0)) {
3378 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3379 }
3380 // (Macro-expander will further convert this to a Bool, if necessary.)
3381 }
3382
3383 // Find the size in bytes. This is easy; it's the layout_helper.
3384 // The size value must be valid even if the slow path is taken.
3385 Node* size = NULL;
3386 if (layout_is_con) {
3387 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3388 } else { // reflective case
3389 // This reflective path is used by clone and Unsafe.allocateInstance.
3390 size = ConvI2X(layout_val);
3391
3392 // Clear the low bits to extract layout_helper_size_in_bytes:
3393 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3394 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3395 size = _gvn.transform( new (C) AndXNode(size, mask) );
3396 }
3397 if (return_size_val != NULL) {
3398 (*return_size_val) = size;
3399 }
3400
3401 // This is a precise notnull oop of the klass.
3402 // (Actually, it need not be precise if this is a reflective allocation.)
3403 // It's what we cast the result to.
3404 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3405 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3406 const TypeOopPtr* oop_type = tklass->as_instance_type();
3407
3408 // Now generate allocation code
3409
3410 // The entire memory state is needed for slow path of the allocation
3411 // since GC and deoptimization can happened.
3412 Node *mem = reset_memory();
3413 set_all_memory(mem); // Create new memory state
3414
3415 AllocateNode* alloc
3416 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3417 control(), mem, i_o(),
3418 size, klass_node,
3419 initial_slow_test);
3420
3421 return set_output_for_allocation(alloc, oop_type);
3422 }
3423
3424 //-------------------------------new_array-------------------------------------
3425 // helper for both newarray and anewarray
3426 // The 'length' parameter is (obviously) the length of the array.
3427 // See comments on new_instance for the meaning of the other arguments.
3428 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3429 Node* length, // number of array elements
3430 int nargs, // number of arguments to push back for uncommon trap
3431 Node* *return_size_val) {
3432 jint layout_con = Klass::_lh_neutral_value;
3433 Node* layout_val = get_layout_helper(klass_node, layout_con);
3434 int layout_is_con = (layout_val == NULL);
3435
3436 if (!layout_is_con && !StressReflectiveCode &&
3437 !too_many_traps(Deoptimization::Reason_class_check)) {
3438 // This is a reflective array creation site.
3439 // Optimistically assume that it is a subtype of Object[],
3440 // so that we can fold up all the address arithmetic.
3441 layout_con = Klass::array_layout_helper(T_OBJECT);
3442 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3443 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3444 { BuildCutout unless(this, bol_lh, PROB_MAX);
3445 inc_sp(nargs);
3446 uncommon_trap(Deoptimization::Reason_class_check,
3447 Deoptimization::Action_maybe_recompile);
3448 }
3449 layout_val = NULL;
3450 layout_is_con = true;
3451 }
3452
3453 // Generate the initial go-slow test. Make sure we do not overflow
3454 // if length is huge (near 2Gig) or negative! We do not need
3455 // exact double-words here, just a close approximation of needed
3456 // double-words. We can't add any offset or rounding bits, lest we
3457 // take a size -1 of bytes and make it positive. Use an unsigned
3458 // compare, so negative sizes look hugely positive.
3459 int fast_size_limit = FastAllocateSizeLimit;
3460 if (layout_is_con) {
3461 assert(!StressReflectiveCode, "stress mode does not use these paths");
3462 // Increase the size limit if we have exact knowledge of array type.
3463 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3464 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3465 }
3466
3467 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3468 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3469 if (initial_slow_test->is_Bool()) {
3470 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3471 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3472 }
3473
3474 // --- Size Computation ---
3475 // array_size = round_to_heap(array_header + (length << elem_shift));
3476 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3477 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3478 // The rounding mask is strength-reduced, if possible.
3479 int round_mask = MinObjAlignmentInBytes - 1;
3480 Node* header_size = NULL;
3481 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3482 // (T_BYTE has the weakest alignment and size restrictions...)
3483 if (layout_is_con) {
3484 int hsize = Klass::layout_helper_header_size(layout_con);
3485 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3486 BasicType etype = Klass::layout_helper_element_type(layout_con);
3487 if ((round_mask & ~right_n_bits(eshift)) == 0)
3488 round_mask = 0; // strength-reduce it if it goes away completely
3489 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3490 assert(header_size_min <= hsize, "generic minimum is smallest");
3491 header_size_min = hsize;
3492 header_size = intcon(hsize + round_mask);
3493 } else {
3494 Node* hss = intcon(Klass::_lh_header_size_shift);
3495 Node* hsm = intcon(Klass::_lh_header_size_mask);
3496 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3497 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3498 Node* mask = intcon(round_mask);
3499 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3500 }
3501
3502 Node* elem_shift = NULL;
3503 if (layout_is_con) {
3504 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3505 if (eshift != 0)
3506 elem_shift = intcon(eshift);
3507 } else {
3508 // There is no need to mask or shift this value.
3509 // The semantics of LShiftINode include an implicit mask to 0x1F.
3510 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3511 elem_shift = layout_val;
3512 }
3513
3514 // Transition to native address size for all offset calculations:
3515 Node* lengthx = ConvI2X(length);
3516 Node* headerx = ConvI2X(header_size);
3517 #ifdef _LP64
3518 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3519 if (tllen != NULL && tllen->_lo < 0) {
3520 // Add a manual constraint to a positive range. Cf. array_element_address.
3521 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3522 if (size_max > tllen->_hi) size_max = tllen->_hi;
3523 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3524 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon));
3525 }
3526 }
3527 #endif
3528
3529 // Combine header size (plus rounding) and body size. Then round down.
3530 // This computation cannot overflow, because it is used only in two
3531 // places, one where the length is sharply limited, and the other
3532 // after a successful allocation.
3533 Node* abody = lengthx;
3534 if (elem_shift != NULL)
3535 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3536 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3537 if (round_mask != 0) {
3538 Node* mask = MakeConX(~round_mask);
3539 size = _gvn.transform( new(C) AndXNode(size, mask) );
3540 }
3541 // else if round_mask == 0, the size computation is self-rounding
3542
3543 if (return_size_val != NULL) {
3544 // This is the size
3545 (*return_size_val) = size;
3546 }
3547
3548 // Now generate allocation code
3549
3550 // The entire memory state is needed for slow path of the allocation
3551 // since GC and deoptimization can happened.
3552 Node *mem = reset_memory();
3553 set_all_memory(mem); // Create new memory state
3554
3555 // Create the AllocateArrayNode and its result projections
3556 AllocateArrayNode* alloc
3557 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3558 control(), mem, i_o(),
3559 size, klass_node,
3560 initial_slow_test,
3561 length);
3562
3563 // Cast to correct type. Note that the klass_node may be constant or not,
3564 // and in the latter case the actual array type will be inexact also.
3565 // (This happens via a non-constant argument to inline_native_newArray.)
3566 // In any case, the value of klass_node provides the desired array type.
3567 const TypeInt* length_type = _gvn.find_int_type(length);
3568 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3569 if (ary_type->isa_aryptr() && length_type != NULL) {
3570 // Try to get a better type than POS for the size
3571 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3572 }
3573
3574 Node* javaoop = set_output_for_allocation(alloc, ary_type);
3575
3576 // Cast length on remaining path to be as narrow as possible
3577 if (map()->find_edge(length) >= 0) {
3578 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3579 if (ccast != length) {
3580 _gvn.set_type_bottom(ccast);
3581 record_for_igvn(ccast);
3582 replace_in_map(length, ccast);
3583 }
3584 }
3585
3586 return javaoop;
3587 }
3588
3589 // The following "Ideal_foo" functions are placed here because they recognize
3590 // the graph shapes created by the functions immediately above.
3591
3592 //---------------------------Ideal_allocation----------------------------------
3593 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3594 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3595 if (ptr == NULL) { // reduce dumb test in callers
3596 return NULL;
3597 }
3598 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3599 ptr = ptr->in(1);
3600 if (ptr == NULL) return NULL;
3601 }
3602 // Return NULL for allocations with several casts:
3603 // j.l.reflect.Array.newInstance(jobject, jint)
3604 // Object.clone()
3605 // to keep more precise type from last cast.
3606 if (ptr->is_Proj()) {
3607 Node* allo = ptr->in(0);
3608 if (allo != NULL && allo->is_Allocate()) {
3609 return allo->as_Allocate();
3610 }
3611 }
3612 // Report failure to match.
3613 return NULL;
3614 }
3615
3616 // Fancy version which also strips off an offset (and reports it to caller).
3617 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3618 intptr_t& offset) {
3619 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3620 if (base == NULL) return NULL;
3621 return Ideal_allocation(base, phase);
3622 }
3623
3624 // Trace Initialize <- Proj[Parm] <- Allocate
3625 AllocateNode* InitializeNode::allocation() {
3626 Node* rawoop = in(InitializeNode::RawAddress);
3627 if (rawoop->is_Proj()) {
3628 Node* alloc = rawoop->in(0);
3629 if (alloc->is_Allocate()) {
3630 return alloc->as_Allocate();
3631 }
3632 }
3633 return NULL;
3634 }
3635
3636 // Trace Allocate -> Proj[Parm] -> Initialize
3637 InitializeNode* AllocateNode::initialization() {
3638 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3639 if (rawoop == NULL) return NULL;
3640 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3641 Node* init = rawoop->fast_out(i);
3642 if (init->is_Initialize()) {
3643 assert(init->as_Initialize()->allocation() == this, "2-way link");
3644 return init->as_Initialize();
3645 }
3646 }
3647 return NULL;
3648 }
3649
3650 //----------------------------- loop predicates ---------------------------
3651
3652 //------------------------------add_predicate_impl----------------------------
3653 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3654 // Too many traps seen?
3655 if (too_many_traps(reason)) {
3656 #ifdef ASSERT
3657 if (TraceLoopPredicate) {
3658 int tc = C->trap_count(reason);
3659 tty->print("too many traps=%s tcount=%d in ",
3660 Deoptimization::trap_reason_name(reason), tc);
3661 method()->print(); // which method has too many predicate traps
3662 tty->cr();
3663 }
3664 #endif
3665 // We cannot afford to take more traps here,
3666 // do not generate predicate.
3667 return;
3668 }
3669
3670 Node *cont = _gvn.intcon(1);
3671 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3672 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3673 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3674 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3675 C->add_predicate_opaq(opq);
3676 {
3677 PreserveJVMState pjvms(this);
3678 set_control(iffalse);
3679 inc_sp(nargs);
3680 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3681 }
3682 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3683 set_control(iftrue);
3684 }
3685
3686 //------------------------------add_predicate---------------------------------
3687 void GraphKit::add_predicate(int nargs) {
3688 if (UseLoopPredicate) {
3689 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3690 }
3691 // loop's limit check predicate should be near the loop.
3692 if (LoopLimitCheck) {
3693 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3694 }
3695 }
3696
3697 //----------------------------- store barriers ----------------------------
3698 #define __ ideal.
3699
3700 void GraphKit::sync_kit(IdealKit& ideal) {
3701 set_all_memory(__ merged_memory());
3702 set_i_o(__ i_o());
3703 set_control(__ ctrl());
3704 }
3705
3706 void GraphKit::final_sync(IdealKit& ideal) {
3707 // Final sync IdealKit and graphKit.
3708 sync_kit(ideal);
3709 }
3710
3711 // vanilla/CMS post barrier
3712 // Insert a write-barrier store. This is to let generational GC work; we have
3713 // to flag all oop-stores before the next GC point.
3714 void GraphKit::write_barrier_post(Node* oop_store,
3715 Node* obj,
3716 Node* adr,
3717 uint adr_idx,
3718 Node* val,
3719 bool use_precise) {
3720 // No store check needed if we're storing a NULL or an old object
3721 // (latter case is probably a string constant). The concurrent
3722 // mark sweep garbage collector, however, needs to have all nonNull
3723 // oop updates flagged via card-marks.
3724 if (val != NULL && val->is_Con()) {
3725 // must be either an oop or NULL
3726 const Type* t = val->bottom_type();
3727 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3728 // stores of null never (?) need barriers
3729 return;
3730 }
3731
3732 if (use_ReduceInitialCardMarks()
3733 && obj == just_allocated_object(control())) {
3734 // We can skip marks on a freshly-allocated object in Eden.
3735 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3736 // That routine informs GC to take appropriate compensating steps,
3737 // upon a slow-path allocation, so as to make this card-mark
3738 // elision safe.
3739 return;
3740 }
3741
3742 if (!use_precise) {
3743 // All card marks for a (non-array) instance are in one place:
3744 adr = obj;
3745 }
3746 // (Else it's an array (or unknown), and we want more precise card marks.)
3747 assert(adr != NULL, "");
3748
3749 IdealKit ideal(this, true);
3750
3751 // Convert the pointer to an int prior to doing math on it
3752 Node* cast = __ CastPX(__ ctrl(), adr);
3753
3754 // Divide by card size
3755 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3756 "Only one we handle so far.");
3757 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3758
3759 // Combine card table base and card offset
3760 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3761
3762 // Get the alias_index for raw card-mark memory
3763 int adr_type = Compile::AliasIdxRaw;
3764 Node* zero = __ ConI(0); // Dirty card value
3765 BasicType bt = T_BYTE;
3766
3767 if (UseCondCardMark) {
3768 // The classic GC reference write barrier is typically implemented
3769 // as a store into the global card mark table. Unfortunately
3770 // unconditional stores can result in false sharing and excessive
3771 // coherence traffic as well as false transactional aborts.
3772 // UseCondCardMark enables MP "polite" conditional card mark
3773 // stores. In theory we could relax the load from ctrl() to
3774 // no_ctrl, but that doesn't buy much latitude.
3775 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3776 __ if_then(card_val, BoolTest::ne, zero);
3777 }
3778
3779 // Smash zero into card
3780 if( !UseConcMarkSweepGC ) {
3781 __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3782 } else {
3783 // Specialized path for CM store barrier
3784 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3785 }
3786
3787 if (UseCondCardMark) {
3788 __ end_if();
3789 }
3790
3791 // Final sync IdealKit and GraphKit.
3792 final_sync(ideal);
3793 }
3794
3795 // G1 pre/post barriers
3796 void GraphKit::g1_write_barrier_pre(bool do_load,
3797 Node* obj,
3798 Node* adr,
3799 uint alias_idx,
3800 Node* val,
3801 const TypeOopPtr* val_type,
3802 Node* pre_val,
3803 BasicType bt) {
3804
3805 // Some sanity checks
3806 // Note: val is unused in this routine.
3807
3808 if (do_load) {
3809 // We need to generate the load of the previous value
3810 assert(obj != NULL, "must have a base");
3811 assert(adr != NULL, "where are loading from?");
3812 assert(pre_val == NULL, "loaded already?");
3813 assert(val_type != NULL, "need a type");
3814 } else {
3815 // In this case both val_type and alias_idx are unused.
3816 assert(pre_val != NULL, "must be loaded already");
3817 // Nothing to be done if pre_val is null.
3818 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3819 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3820 }
3821 assert(bt == T_OBJECT, "or we shouldn't be here");
3822
3823 IdealKit ideal(this, true);
3824
3825 Node* tls = __ thread(); // ThreadLocalStorage
3826
3827 Node* no_ctrl = NULL;
3828 Node* no_base = __ top();
3829 Node* zero = __ ConI(0);
3830 Node* zeroX = __ ConX(0);
3831
3832 float likely = PROB_LIKELY(0.999);
3833 float unlikely = PROB_UNLIKELY(0.999);
3834
3835 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3836 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3837
3838 // Offsets into the thread
3839 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3840 PtrQueue::byte_offset_of_active());
3841 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3842 PtrQueue::byte_offset_of_index());
3843 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3844 PtrQueue::byte_offset_of_buf());
3845
3846 // Now the actual pointers into the thread
3847 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3848 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3849 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3850
3851 // Now some of the values
3852 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3853
3854 // if (!marking)
3855 __ if_then(marking, BoolTest::ne, zero, unlikely); {
3856 BasicType index_bt = TypeX_X->basic_type();
3857 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3858 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3859
3860 if (do_load) {
3861 // load original value
3862 // alias_idx correct??
3863 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3864 }
3865
3866 // if (pre_val != NULL)
3867 __ if_then(pre_val, BoolTest::ne, null()); {
3868 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3869
3870 // is the queue for this thread full?
3871 __ if_then(index, BoolTest::ne, zeroX, likely); {
3872
3873 // decrement the index
3874 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3875
3876 // Now get the buffer location we will log the previous value into and store it
3877 Node *log_addr = __ AddP(no_base, buffer, next_index);
3878 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw);
3879 // update the index
3880 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw);
3881
3882 } __ else_(); {
3883
3884 // logging buffer is full, call the runtime
3885 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3886 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3887 } __ end_if(); // (!index)
3888 } __ end_if(); // (pre_val != NULL)
3889 } __ end_if(); // (!marking)
3890
3891 // Final sync IdealKit and GraphKit.
3892 final_sync(ideal);
3893 }
3894
3895 //
3896 // Update the card table and add card address to the queue
3897 //
3898 void GraphKit::g1_mark_card(IdealKit& ideal,
3899 Node* card_adr,
3900 Node* oop_store,
3901 uint oop_alias_idx,
3902 Node* index,
3903 Node* index_adr,
3904 Node* buffer,
3905 const TypeFunc* tf) {
3906
3907 Node* zero = __ ConI(0);
3908 Node* zeroX = __ ConX(0);
3909 Node* no_base = __ top();
3910 BasicType card_bt = T_BYTE;
3911 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3912 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3913
3914 // Now do the queue work
3915 __ if_then(index, BoolTest::ne, zeroX); {
3916
3917 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3918 Node* log_addr = __ AddP(no_base, buffer, next_index);
3919
3920 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3921 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw);
3922
3923 } __ else_(); {
3924 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3925 } __ end_if();
3926
3927 }
3928
3929 void GraphKit::g1_write_barrier_post(Node* oop_store,
3930 Node* obj,
3931 Node* adr,
3932 uint alias_idx,
3933 Node* val,
3934 BasicType bt,
3935 bool use_precise) {
3936 // If we are writing a NULL then we need no post barrier
3937
3938 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3939 // Must be NULL
3940 const Type* t = val->bottom_type();
3941 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3942 // No post barrier if writing NULLx
3943 return;
3944 }
3945
3946 if (!use_precise) {
3947 // All card marks for a (non-array) instance are in one place:
3948 adr = obj;
3949 }
3950 // (Else it's an array (or unknown), and we want more precise card marks.)
3951 assert(adr != NULL, "");
3952
3953 IdealKit ideal(this, true);
3954
3955 Node* tls = __ thread(); // ThreadLocalStorage
3956
3957 Node* no_base = __ top();
3958 float likely = PROB_LIKELY(0.999);
3959 float unlikely = PROB_UNLIKELY(0.999);
3960 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
3961 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
3962 Node* zeroX = __ ConX(0);
3963
3964 // Get the alias_index for raw card-mark memory
3965 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3966
3967 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3968
3969 // Offsets into the thread
3970 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3971 PtrQueue::byte_offset_of_index());
3972 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3973 PtrQueue::byte_offset_of_buf());
3974
3975 // Pointers into the thread
3976
3977 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3978 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3979
3980 // Now some values
3981 // Use ctrl to avoid hoisting these values past a safepoint, which could
3982 // potentially reset these fields in the JavaThread.
3983 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
3984 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3985
3986 // Convert the store obj pointer to an int prior to doing math on it
3987 // Must use ctrl to prevent "integerized oop" existing across safepoint
3988 Node* cast = __ CastPX(__ ctrl(), adr);
3989
3990 // Divide pointer by card size
3991 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3992
3993 // Combine card table base and card offset
3994 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3995
3996 // If we know the value being stored does it cross regions?
3997
3998 if (val != NULL) {
3999 // Does the store cause us to cross regions?
4000
4001 // Should be able to do an unsigned compare of region_size instead of
4002 // and extra shift. Do we have an unsigned compare??
4003 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4004 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4005
4006 // if (xor_res == 0) same region so skip
4007 __ if_then(xor_res, BoolTest::ne, zeroX); {
4008
4009 // No barrier if we are storing a NULL
4010 __ if_then(val, BoolTest::ne, null(), unlikely); {
4011
4012 // Ok must mark the card if not already dirty
4013
4014 // load the original value of the card
4015 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4016
4017 __ if_then(card_val, BoolTest::ne, young_card); {
4018 sync_kit(ideal);
4019 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4020 insert_mem_bar(Op_MemBarVolatile, oop_store);
4021 __ sync_kit(this);
4022
4023 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4024 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4025 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4026 } __ end_if();
4027 } __ end_if();
4028 } __ end_if();
4029 } __ end_if();
4030 } else {
4031 // Object.clone() instrinsic uses this path.
4032 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4033 }
4034
4035 // Final sync IdealKit and GraphKit.
4036 final_sync(ideal);
4037 }
4038 #undef __
4039
4040
4041
4042 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
4043 if (java_lang_String::has_offset_field()) {
4044 int offset_offset = java_lang_String::offset_offset_in_bytes();
4045 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4046 false, NULL, 0);
4047 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4048 int offset_field_idx = C->get_alias_index(offset_field_type);
4049 return make_load(ctrl,
4050 basic_plus_adr(str, str, offset_offset),
4051 TypeInt::INT, T_INT, offset_field_idx);
4052 } else {
4053 return intcon(0);
4054 }
4055 }
4056
4057 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4058 if (java_lang_String::has_count_field()) {
4059 int count_offset = java_lang_String::count_offset_in_bytes();
4060 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4061 false, NULL, 0);
4062 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4063 int count_field_idx = C->get_alias_index(count_field_type);
4064 return make_load(ctrl,
4065 basic_plus_adr(str, str, count_offset),
4066 TypeInt::INT, T_INT, count_field_idx);
4067 } else {
4068 return load_array_length(load_String_value(ctrl, str));
4069 }
4070 }
4071
4072 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4073 int value_offset = java_lang_String::value_offset_in_bytes();
4074 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4075 false, NULL, 0);
4076 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4077 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4078 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
4079 ciTypeArrayKlass::make(T_CHAR), true, 0);
4080 int value_field_idx = C->get_alias_index(value_field_type);
4081 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4082 value_type, T_OBJECT, value_field_idx);
4083 // String.value field is known to be @Stable.
4084 if (UseImplicitStableValues) {
4085 load = cast_array_to_stable(load, value_type);
4086 }
4087 return load;
4088 }
4089
4090 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
4091 int offset_offset = java_lang_String::offset_offset_in_bytes();
4092 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4093 false, NULL, 0);
4094 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4095 int offset_field_idx = C->get_alias_index(offset_field_type);
4096 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
4097 value, T_INT, offset_field_idx);
4098 }
4099
4100 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4101 int value_offset = java_lang_String::value_offset_in_bytes();
4102 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4103 false, NULL, 0);
4104 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4105
4106 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4107 value, TypeAryPtr::CHARS, T_OBJECT);
4108 }
4109
4110 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
4111 int count_offset = java_lang_String::count_offset_in_bytes();
4112 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4113 false, NULL, 0);
4114 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4115 int count_field_idx = C->get_alias_index(count_field_type);
4116 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
4117 value, T_INT, count_field_idx);
4118 }
4119
4120 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4121 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4122 // assumption of CCP analysis.
4123 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
4124 }