Tue, 08 Oct 2013 17:35:51 +0200
8014555: G1: Memory ordering problem with Conc refinement and card marking
Summary: Add a StoreLoad barrier in the G1 post-barrier to fix a race with concurrent refinement. Also-reviewed-by: martin.doerr@sap.com
Reviewed-by: iveresov, tschatzl, brutisso, roland, kvn
1 /*
2 * Copyright (c) 2001, 2012, 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 */
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"
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 }
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 }
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 }
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 }
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 }
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 }
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 }
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 }
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
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 }
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 }
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 }
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 }
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 }
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 }
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 }
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
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 }
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
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 }
260 // No pre-existing exception of the same type. Chain it on the list.
261 push_exception_state(ex_map);
262 }
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 }
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 }
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 }
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 }
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();
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(srctype);
424 if (phi->type() != dsttype) {
425 phi->set_type(dsttype);
426 _gvn.set_type(phi, dsttype);
427 }
428 }
429 }
430 }
431 }
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 }
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 }
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 }
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.
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);
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) );
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 }
511 }
513 //------------------------------builtin_throw----------------------------------
514 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
515 bool must_throw = true;
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 }
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();
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.)
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 }
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) );
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);
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);
601 add_exception_state(make_exception_state(ex_node));
602 return;
603 }
604 }
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.
611 // Usual case: Bail to interpreter.
612 // Reserve the right to recompile if we haven't seen anything yet.
614 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
615 if (treat_throw_as_hot
616 && (method()->method_data()->trap_recompiled_at(bci())
617 || C->too_many_traps(reason))) {
618 // We cannot afford to take more traps here. Suffer in the interpreter.
619 if (C->log() != NULL)
620 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
621 Deoptimization::trap_reason_name(reason),
622 C->trap_count(reason));
623 action = Deoptimization::Action_none;
624 }
626 // "must_throw" prunes the JVM state to include only the stack, if there
627 // are no local exception handlers. This should cut down on register
628 // allocation time and code size, by drastically reducing the number
629 // of in-edges on the call to the uncommon trap.
631 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
632 }
635 //----------------------------PreserveJVMState---------------------------------
636 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
637 debug_only(kit->verify_map());
638 _kit = kit;
639 _map = kit->map(); // preserve the map
640 _sp = kit->sp();
641 kit->set_map(clone_map ? kit->clone_map() : NULL);
642 #ifdef ASSERT
643 _bci = kit->bci();
644 Parse* parser = kit->is_Parse();
645 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
646 _block = block;
647 #endif
648 }
649 PreserveJVMState::~PreserveJVMState() {
650 GraphKit* kit = _kit;
651 #ifdef ASSERT
652 assert(kit->bci() == _bci, "bci must not shift");
653 Parse* parser = kit->is_Parse();
654 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
655 assert(block == _block, "block must not shift");
656 #endif
657 kit->set_map(_map);
658 kit->set_sp(_sp);
659 }
662 //-----------------------------BuildCutout-------------------------------------
663 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
664 : PreserveJVMState(kit)
665 {
666 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
667 SafePointNode* outer_map = _map; // preserved map is caller's
668 SafePointNode* inner_map = kit->map();
669 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
670 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
671 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
672 }
673 BuildCutout::~BuildCutout() {
674 GraphKit* kit = _kit;
675 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
676 }
678 //---------------------------PreserveReexecuteState----------------------------
679 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
680 assert(!kit->stopped(), "must call stopped() before");
681 _kit = kit;
682 _sp = kit->sp();
683 _reexecute = kit->jvms()->_reexecute;
684 }
685 PreserveReexecuteState::~PreserveReexecuteState() {
686 if (_kit->stopped()) return;
687 _kit->jvms()->_reexecute = _reexecute;
688 _kit->set_sp(_sp);
689 }
691 //------------------------------clone_map--------------------------------------
692 // Implementation of PreserveJVMState
693 //
694 // Only clone_map(...) here. If this function is only used in the
695 // PreserveJVMState class we may want to get rid of this extra
696 // function eventually and do it all there.
698 SafePointNode* GraphKit::clone_map() {
699 if (map() == NULL) return NULL;
701 // Clone the memory edge first
702 Node* mem = MergeMemNode::make(C, map()->memory());
703 gvn().set_type_bottom(mem);
705 SafePointNode *clonemap = (SafePointNode*)map()->clone();
706 JVMState* jvms = this->jvms();
707 JVMState* clonejvms = jvms->clone_shallow(C);
708 clonemap->set_memory(mem);
709 clonemap->set_jvms(clonejvms);
710 clonejvms->set_map(clonemap);
711 record_for_igvn(clonemap);
712 gvn().set_type_bottom(clonemap);
713 return clonemap;
714 }
717 //-----------------------------set_map_clone-----------------------------------
718 void GraphKit::set_map_clone(SafePointNode* m) {
719 _map = m;
720 _map = clone_map();
721 _map->set_next_exception(NULL);
722 debug_only(verify_map());
723 }
726 //----------------------------kill_dead_locals---------------------------------
727 // Detect any locals which are known to be dead, and force them to top.
728 void GraphKit::kill_dead_locals() {
729 // Consult the liveness information for the locals. If any
730 // of them are unused, then they can be replaced by top(). This
731 // should help register allocation time and cut down on the size
732 // of the deoptimization information.
734 // This call is made from many of the bytecode handling
735 // subroutines called from the Big Switch in do_one_bytecode.
736 // Every bytecode which might include a slow path is responsible
737 // for killing its dead locals. The more consistent we
738 // are about killing deads, the fewer useless phis will be
739 // constructed for them at various merge points.
741 // bci can be -1 (InvocationEntryBci). We return the entry
742 // liveness for the method.
744 if (method() == NULL || method()->code_size() == 0) {
745 // We are building a graph for a call to a native method.
746 // All locals are live.
747 return;
748 }
750 ResourceMark rm;
752 // Consult the liveness information for the locals. If any
753 // of them are unused, then they can be replaced by top(). This
754 // should help register allocation time and cut down on the size
755 // of the deoptimization information.
756 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
758 int len = (int)live_locals.size();
759 assert(len <= jvms()->loc_size(), "too many live locals");
760 for (int local = 0; local < len; local++) {
761 if (!live_locals.at(local)) {
762 set_local(local, top());
763 }
764 }
765 }
767 #ifdef ASSERT
768 //-------------------------dead_locals_are_killed------------------------------
769 // Return true if all dead locals are set to top in the map.
770 // Used to assert "clean" debug info at various points.
771 bool GraphKit::dead_locals_are_killed() {
772 if (method() == NULL || method()->code_size() == 0) {
773 // No locals need to be dead, so all is as it should be.
774 return true;
775 }
777 // Make sure somebody called kill_dead_locals upstream.
778 ResourceMark rm;
779 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
780 if (jvms->loc_size() == 0) continue; // no locals to consult
781 SafePointNode* map = jvms->map();
782 ciMethod* method = jvms->method();
783 int bci = jvms->bci();
784 if (jvms == this->jvms()) {
785 bci = this->bci(); // it might not yet be synched
786 }
787 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
788 int len = (int)live_locals.size();
789 if (!live_locals.is_valid() || len == 0)
790 // This method is trivial, or is poisoned by a breakpoint.
791 return true;
792 assert(len == jvms->loc_size(), "live map consistent with locals map");
793 for (int local = 0; local < len; local++) {
794 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
795 if (PrintMiscellaneous && (Verbose || WizardMode)) {
796 tty->print_cr("Zombie local %d: ", local);
797 jvms->dump();
798 }
799 return false;
800 }
801 }
802 }
803 return true;
804 }
806 #endif //ASSERT
808 // Helper function for enforcing certain bytecodes to reexecute if
809 // deoptimization happens
810 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
811 ciMethod* cur_method = jvms->method();
812 int cur_bci = jvms->bci();
813 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
814 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
815 return Interpreter::bytecode_should_reexecute(code) ||
816 is_anewarray && code == Bytecodes::_multianewarray;
817 // Reexecute _multianewarray bytecode which was replaced with
818 // sequence of [a]newarray. See Parse::do_multianewarray().
819 //
820 // Note: interpreter should not have it set since this optimization
821 // is limited by dimensions and guarded by flag so in some cases
822 // multianewarray() runtime calls will be generated and
823 // the bytecode should not be reexecutes (stack will not be reset).
824 } else
825 return false;
826 }
828 // Helper function for adding JVMState and debug information to node
829 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
830 // Add the safepoint edges to the call (or other safepoint).
832 // Make sure dead locals are set to top. This
833 // should help register allocation time and cut down on the size
834 // of the deoptimization information.
835 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
837 // Walk the inline list to fill in the correct set of JVMState's
838 // Also fill in the associated edges for each JVMState.
840 // If the bytecode needs to be reexecuted we need to put
841 // the arguments back on the stack.
842 const bool should_reexecute = jvms()->should_reexecute();
843 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
845 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
846 // undefined if the bci is different. This is normal for Parse but it
847 // should not happen for LibraryCallKit because only one bci is processed.
848 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
849 "in LibraryCallKit the reexecute bit should not change");
851 // If we are guaranteed to throw, we can prune everything but the
852 // input to the current bytecode.
853 bool can_prune_locals = false;
854 uint stack_slots_not_pruned = 0;
855 int inputs = 0, depth = 0;
856 if (must_throw) {
857 assert(method() == youngest_jvms->method(), "sanity");
858 if (compute_stack_effects(inputs, depth)) {
859 can_prune_locals = true;
860 stack_slots_not_pruned = inputs;
861 }
862 }
864 if (env()->jvmti_can_access_local_variables()) {
865 // At any safepoint, this method can get breakpointed, which would
866 // then require an immediate deoptimization.
867 can_prune_locals = false; // do not prune locals
868 stack_slots_not_pruned = 0;
869 }
871 // do not scribble on the input jvms
872 JVMState* out_jvms = youngest_jvms->clone_deep(C);
873 call->set_jvms(out_jvms); // Start jvms list for call node
875 // For a known set of bytecodes, the interpreter should reexecute them if
876 // deoptimization happens. We set the reexecute state for them here
877 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
878 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
879 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
880 }
882 // Presize the call:
883 DEBUG_ONLY(uint non_debug_edges = call->req());
884 call->add_req_batch(top(), youngest_jvms->debug_depth());
885 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
887 // Set up edges so that the call looks like this:
888 // Call [state:] ctl io mem fptr retadr
889 // [parms:] parm0 ... parmN
890 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
891 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
892 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
893 // Note that caller debug info precedes callee debug info.
895 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
896 uint debug_ptr = call->req();
898 // Loop over the map input edges associated with jvms, add them
899 // to the call node, & reset all offsets to match call node array.
900 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
901 uint debug_end = debug_ptr;
902 uint debug_start = debug_ptr - in_jvms->debug_size();
903 debug_ptr = debug_start; // back up the ptr
905 uint p = debug_start; // walks forward in [debug_start, debug_end)
906 uint j, k, l;
907 SafePointNode* in_map = in_jvms->map();
908 out_jvms->set_map(call);
910 if (can_prune_locals) {
911 assert(in_jvms->method() == out_jvms->method(), "sanity");
912 // If the current throw can reach an exception handler in this JVMS,
913 // then we must keep everything live that can reach that handler.
914 // As a quick and dirty approximation, we look for any handlers at all.
915 if (in_jvms->method()->has_exception_handlers()) {
916 can_prune_locals = false;
917 }
918 }
920 // Add the Locals
921 k = in_jvms->locoff();
922 l = in_jvms->loc_size();
923 out_jvms->set_locoff(p);
924 if (!can_prune_locals) {
925 for (j = 0; j < l; j++)
926 call->set_req(p++, in_map->in(k+j));
927 } else {
928 p += l; // already set to top above by add_req_batch
929 }
931 // Add the Expression Stack
932 k = in_jvms->stkoff();
933 l = in_jvms->sp();
934 out_jvms->set_stkoff(p);
935 if (!can_prune_locals) {
936 for (j = 0; j < l; j++)
937 call->set_req(p++, in_map->in(k+j));
938 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
939 // Divide stack into {S0,...,S1}, where S0 is set to top.
940 uint s1 = stack_slots_not_pruned;
941 stack_slots_not_pruned = 0; // for next iteration
942 if (s1 > l) s1 = l;
943 uint s0 = l - s1;
944 p += s0; // skip the tops preinstalled by add_req_batch
945 for (j = s0; j < l; j++)
946 call->set_req(p++, in_map->in(k+j));
947 } else {
948 p += l; // already set to top above by add_req_batch
949 }
951 // Add the Monitors
952 k = in_jvms->monoff();
953 l = in_jvms->mon_size();
954 out_jvms->set_monoff(p);
955 for (j = 0; j < l; j++)
956 call->set_req(p++, in_map->in(k+j));
958 // Copy any scalar object fields.
959 k = in_jvms->scloff();
960 l = in_jvms->scl_size();
961 out_jvms->set_scloff(p);
962 for (j = 0; j < l; j++)
963 call->set_req(p++, in_map->in(k+j));
965 // Finish the new jvms.
966 out_jvms->set_endoff(p);
968 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
969 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
970 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
971 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
972 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
973 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
975 // Update the two tail pointers in parallel.
976 out_jvms = out_jvms->caller();
977 in_jvms = in_jvms->caller();
978 }
980 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
982 // Test the correctness of JVMState::debug_xxx accessors:
983 assert(call->jvms()->debug_start() == non_debug_edges, "");
984 assert(call->jvms()->debug_end() == call->req(), "");
985 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
986 }
988 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
989 Bytecodes::Code code = java_bc();
990 if (code == Bytecodes::_wide) {
991 code = method()->java_code_at_bci(bci() + 1);
992 }
994 BasicType rtype = T_ILLEGAL;
995 int rsize = 0;
997 if (code != Bytecodes::_illegal) {
998 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
999 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1000 if (rtype < T_CONFLICT)
1001 rsize = type2size[rtype];
1002 }
1004 switch (code) {
1005 case Bytecodes::_illegal:
1006 return false;
1008 case Bytecodes::_ldc:
1009 case Bytecodes::_ldc_w:
1010 case Bytecodes::_ldc2_w:
1011 inputs = 0;
1012 break;
1014 case Bytecodes::_dup: inputs = 1; break;
1015 case Bytecodes::_dup_x1: inputs = 2; break;
1016 case Bytecodes::_dup_x2: inputs = 3; break;
1017 case Bytecodes::_dup2: inputs = 2; break;
1018 case Bytecodes::_dup2_x1: inputs = 3; break;
1019 case Bytecodes::_dup2_x2: inputs = 4; break;
1020 case Bytecodes::_swap: inputs = 2; break;
1021 case Bytecodes::_arraylength: inputs = 1; break;
1023 case Bytecodes::_getstatic:
1024 case Bytecodes::_putstatic:
1025 case Bytecodes::_getfield:
1026 case Bytecodes::_putfield:
1027 {
1028 bool ignored_will_link;
1029 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1030 int size = field->type()->size();
1031 bool is_get = (depth >= 0), is_static = (depth & 1);
1032 inputs = (is_static ? 0 : 1);
1033 if (is_get) {
1034 depth = size - inputs;
1035 } else {
1036 inputs += size; // putxxx pops the value from the stack
1037 depth = - inputs;
1038 }
1039 }
1040 break;
1042 case Bytecodes::_invokevirtual:
1043 case Bytecodes::_invokespecial:
1044 case Bytecodes::_invokestatic:
1045 case Bytecodes::_invokedynamic:
1046 case Bytecodes::_invokeinterface:
1047 {
1048 bool ignored_will_link;
1049 ciSignature* declared_signature = NULL;
1050 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1051 assert(declared_signature != NULL, "cannot be null");
1052 inputs = declared_signature->arg_size_for_bc(code);
1053 int size = declared_signature->return_type()->size();
1054 depth = size - inputs;
1055 }
1056 break;
1058 case Bytecodes::_multianewarray:
1059 {
1060 ciBytecodeStream iter(method());
1061 iter.reset_to_bci(bci());
1062 iter.next();
1063 inputs = iter.get_dimensions();
1064 assert(rsize == 1, "");
1065 depth = rsize - inputs;
1066 }
1067 break;
1069 case Bytecodes::_ireturn:
1070 case Bytecodes::_lreturn:
1071 case Bytecodes::_freturn:
1072 case Bytecodes::_dreturn:
1073 case Bytecodes::_areturn:
1074 assert(rsize = -depth, "");
1075 inputs = rsize;
1076 break;
1078 case Bytecodes::_jsr:
1079 case Bytecodes::_jsr_w:
1080 inputs = 0;
1081 depth = 1; // S.B. depth=1, not zero
1082 break;
1084 default:
1085 // bytecode produces a typed result
1086 inputs = rsize - depth;
1087 assert(inputs >= 0, "");
1088 break;
1089 }
1091 #ifdef ASSERT
1092 // spot check
1093 int outputs = depth + inputs;
1094 assert(outputs >= 0, "sanity");
1095 switch (code) {
1096 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1097 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1098 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1099 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1100 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1101 }
1102 #endif //ASSERT
1104 return true;
1105 }
1109 //------------------------------basic_plus_adr---------------------------------
1110 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1111 // short-circuit a common case
1112 if (offset == intcon(0)) return ptr;
1113 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1114 }
1116 Node* GraphKit::ConvI2L(Node* offset) {
1117 // short-circuit a common case
1118 jint offset_con = find_int_con(offset, Type::OffsetBot);
1119 if (offset_con != Type::OffsetBot) {
1120 return longcon((jlong) offset_con);
1121 }
1122 return _gvn.transform( new (C) ConvI2LNode(offset));
1123 }
1124 Node* GraphKit::ConvL2I(Node* offset) {
1125 // short-circuit a common case
1126 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1127 if (offset_con != (jlong)Type::OffsetBot) {
1128 return intcon((int) offset_con);
1129 }
1130 return _gvn.transform( new (C) ConvL2INode(offset));
1131 }
1133 //-------------------------load_object_klass-----------------------------------
1134 Node* GraphKit::load_object_klass(Node* obj) {
1135 // Special-case a fresh allocation to avoid building nodes:
1136 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1137 if (akls != NULL) return akls;
1138 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1139 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1140 }
1142 //-------------------------load_array_length-----------------------------------
1143 Node* GraphKit::load_array_length(Node* array) {
1144 // Special-case a fresh allocation to avoid building nodes:
1145 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1146 Node *alen;
1147 if (alloc == NULL) {
1148 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1149 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1150 } else {
1151 alen = alloc->Ideal_length();
1152 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1153 if (ccast != alen) {
1154 alen = _gvn.transform(ccast);
1155 }
1156 }
1157 return alen;
1158 }
1160 //------------------------------do_null_check----------------------------------
1161 // Helper function to do a NULL pointer check. Returned value is
1162 // the incoming address with NULL casted away. You are allowed to use the
1163 // not-null value only if you are control dependent on the test.
1164 extern int explicit_null_checks_inserted,
1165 explicit_null_checks_elided;
1166 Node* GraphKit::null_check_common(Node* value, BasicType type,
1167 // optional arguments for variations:
1168 bool assert_null,
1169 Node* *null_control) {
1170 assert(!assert_null || null_control == NULL, "not both at once");
1171 if (stopped()) return top();
1172 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1173 // For some performance testing, we may wish to suppress null checking.
1174 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1175 return value;
1176 }
1177 explicit_null_checks_inserted++;
1179 // Construct NULL check
1180 Node *chk = NULL;
1181 switch(type) {
1182 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1183 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1184 case T_ARRAY : // fall through
1185 type = T_OBJECT; // simplify further tests
1186 case T_OBJECT : {
1187 const Type *t = _gvn.type( value );
1189 const TypeOopPtr* tp = t->isa_oopptr();
1190 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1191 // Only for do_null_check, not any of its siblings:
1192 && !assert_null && null_control == NULL) {
1193 // Usually, any field access or invocation on an unloaded oop type
1194 // will simply fail to link, since the statically linked class is
1195 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1196 // the static class is loaded but the sharper oop type is not.
1197 // Rather than checking for this obscure case in lots of places,
1198 // we simply observe that a null check on an unloaded class
1199 // will always be followed by a nonsense operation, so we
1200 // can just issue the uncommon trap here.
1201 // Our access to the unloaded class will only be correct
1202 // after it has been loaded and initialized, which requires
1203 // a trip through the interpreter.
1204 #ifndef PRODUCT
1205 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1206 #endif
1207 uncommon_trap(Deoptimization::Reason_unloaded,
1208 Deoptimization::Action_reinterpret,
1209 tp->klass(), "!loaded");
1210 return top();
1211 }
1213 if (assert_null) {
1214 // See if the type is contained in NULL_PTR.
1215 // If so, then the value is already null.
1216 if (t->higher_equal(TypePtr::NULL_PTR)) {
1217 explicit_null_checks_elided++;
1218 return value; // Elided null assert quickly!
1219 }
1220 } else {
1221 // See if mixing in the NULL pointer changes type.
1222 // If so, then the NULL pointer was not allowed in the original
1223 // type. In other words, "value" was not-null.
1224 if (t->meet(TypePtr::NULL_PTR) != t) {
1225 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1226 explicit_null_checks_elided++;
1227 return value; // Elided null check quickly!
1228 }
1229 }
1230 chk = new (C) CmpPNode( value, null() );
1231 break;
1232 }
1234 default:
1235 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1236 }
1237 assert(chk != NULL, "sanity check");
1238 chk = _gvn.transform(chk);
1240 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1241 BoolNode *btst = new (C) BoolNode( chk, btest);
1242 Node *tst = _gvn.transform( btst );
1244 //-----------
1245 // if peephole optimizations occurred, a prior test existed.
1246 // If a prior test existed, maybe it dominates as we can avoid this test.
1247 if (tst != btst && type == T_OBJECT) {
1248 // At this point we want to scan up the CFG to see if we can
1249 // find an identical test (and so avoid this test altogether).
1250 Node *cfg = control();
1251 int depth = 0;
1252 while( depth < 16 ) { // Limit search depth for speed
1253 if( cfg->Opcode() == Op_IfTrue &&
1254 cfg->in(0)->in(1) == tst ) {
1255 // Found prior test. Use "cast_not_null" to construct an identical
1256 // CastPP (and hence hash to) as already exists for the prior test.
1257 // Return that casted value.
1258 if (assert_null) {
1259 replace_in_map(value, null());
1260 return null(); // do not issue the redundant test
1261 }
1262 Node *oldcontrol = control();
1263 set_control(cfg);
1264 Node *res = cast_not_null(value);
1265 set_control(oldcontrol);
1266 explicit_null_checks_elided++;
1267 return res;
1268 }
1269 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1270 if (cfg == NULL) break; // Quit at region nodes
1271 depth++;
1272 }
1273 }
1275 //-----------
1276 // Branch to failure if null
1277 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1278 Deoptimization::DeoptReason reason;
1279 if (assert_null)
1280 reason = Deoptimization::Reason_null_assert;
1281 else if (type == T_OBJECT)
1282 reason = Deoptimization::Reason_null_check;
1283 else
1284 reason = Deoptimization::Reason_div0_check;
1286 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1287 // ciMethodData::has_trap_at will return a conservative -1 if any
1288 // must-be-null assertion has failed. This could cause performance
1289 // problems for a method after its first do_null_assert failure.
1290 // Consider using 'Reason_class_check' instead?
1292 // To cause an implicit null check, we set the not-null probability
1293 // to the maximum (PROB_MAX). For an explicit check the probability
1294 // is set to a smaller value.
1295 if (null_control != NULL || too_many_traps(reason)) {
1296 // probability is less likely
1297 ok_prob = PROB_LIKELY_MAG(3);
1298 } else if (!assert_null &&
1299 (ImplicitNullCheckThreshold > 0) &&
1300 method() != NULL &&
1301 (method()->method_data()->trap_count(reason)
1302 >= (uint)ImplicitNullCheckThreshold)) {
1303 ok_prob = PROB_LIKELY_MAG(3);
1304 }
1306 if (null_control != NULL) {
1307 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1308 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1309 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1310 if (null_true == top())
1311 explicit_null_checks_elided++;
1312 (*null_control) = null_true;
1313 } else {
1314 BuildCutout unless(this, tst, ok_prob);
1315 // Check for optimizer eliding test at parse time
1316 if (stopped()) {
1317 // Failure not possible; do not bother making uncommon trap.
1318 explicit_null_checks_elided++;
1319 } else if (assert_null) {
1320 uncommon_trap(reason,
1321 Deoptimization::Action_make_not_entrant,
1322 NULL, "assert_null");
1323 } else {
1324 replace_in_map(value, zerocon(type));
1325 builtin_throw(reason);
1326 }
1327 }
1329 // Must throw exception, fall-thru not possible?
1330 if (stopped()) {
1331 return top(); // No result
1332 }
1334 if (assert_null) {
1335 // Cast obj to null on this path.
1336 replace_in_map(value, zerocon(type));
1337 return zerocon(type);
1338 }
1340 // Cast obj to not-null on this path, if there is no null_control.
1341 // (If there is a null_control, a non-null value may come back to haunt us.)
1342 if (type == T_OBJECT) {
1343 Node* cast = cast_not_null(value, false);
1344 if (null_control == NULL || (*null_control) == top())
1345 replace_in_map(value, cast);
1346 value = cast;
1347 }
1349 return value;
1350 }
1353 //------------------------------cast_not_null----------------------------------
1354 // Cast obj to not-null on this path
1355 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1356 const Type *t = _gvn.type(obj);
1357 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1358 // Object is already not-null?
1359 if( t == t_not_null ) return obj;
1361 Node *cast = new (C) CastPPNode(obj,t_not_null);
1362 cast->init_req(0, control());
1363 cast = _gvn.transform( cast );
1365 // Scan for instances of 'obj' in the current JVM mapping.
1366 // These instances are known to be not-null after the test.
1367 if (do_replace_in_map)
1368 replace_in_map(obj, cast);
1370 return cast; // Return casted value
1371 }
1374 //--------------------------replace_in_map-------------------------------------
1375 void GraphKit::replace_in_map(Node* old, Node* neww) {
1376 this->map()->replace_edge(old, neww);
1378 // Note: This operation potentially replaces any edge
1379 // on the map. This includes locals, stack, and monitors
1380 // of the current (innermost) JVM state.
1382 // We can consider replacing in caller maps.
1383 // The idea would be that an inlined function's null checks
1384 // can be shared with the entire inlining tree.
1385 // The expense of doing this is that the PreserveJVMState class
1386 // would have to preserve caller states too, with a deep copy.
1387 }
1390 //=============================================================================
1391 //--------------------------------memory---------------------------------------
1392 Node* GraphKit::memory(uint alias_idx) {
1393 MergeMemNode* mem = merged_memory();
1394 Node* p = mem->memory_at(alias_idx);
1395 _gvn.set_type(p, Type::MEMORY); // must be mapped
1396 return p;
1397 }
1399 //-----------------------------reset_memory------------------------------------
1400 Node* GraphKit::reset_memory() {
1401 Node* mem = map()->memory();
1402 // do not use this node for any more parsing!
1403 debug_only( map()->set_memory((Node*)NULL) );
1404 return _gvn.transform( mem );
1405 }
1407 //------------------------------set_all_memory---------------------------------
1408 void GraphKit::set_all_memory(Node* newmem) {
1409 Node* mergemem = MergeMemNode::make(C, newmem);
1410 gvn().set_type_bottom(mergemem);
1411 map()->set_memory(mergemem);
1412 }
1414 //------------------------------set_all_memory_call----------------------------
1415 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1416 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1417 set_all_memory(newmem);
1418 }
1420 //=============================================================================
1421 //
1422 // parser factory methods for MemNodes
1423 //
1424 // These are layered on top of the factory methods in LoadNode and StoreNode,
1425 // and integrate with the parser's memory state and _gvn engine.
1426 //
1428 // factory methods in "int adr_idx"
1429 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1430 int adr_idx,
1431 bool require_atomic_access) {
1432 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1433 const TypePtr* adr_type = NULL; // debug-mode-only argument
1434 debug_only(adr_type = C->get_adr_type(adr_idx));
1435 Node* mem = memory(adr_idx);
1436 Node* ld;
1437 if (require_atomic_access && bt == T_LONG) {
1438 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1439 } else {
1440 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1441 }
1442 ld = _gvn.transform(ld);
1443 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1444 // Improve graph before escape analysis and boxing elimination.
1445 record_for_igvn(ld);
1446 }
1447 return ld;
1448 }
1450 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1451 int adr_idx,
1452 bool require_atomic_access) {
1453 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1454 const TypePtr* adr_type = NULL;
1455 debug_only(adr_type = C->get_adr_type(adr_idx));
1456 Node *mem = memory(adr_idx);
1457 Node* st;
1458 if (require_atomic_access && bt == T_LONG) {
1459 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1460 } else {
1461 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1462 }
1463 st = _gvn.transform(st);
1464 set_memory(st, adr_idx);
1465 // Back-to-back stores can only remove intermediate store with DU info
1466 // so push on worklist for optimizer.
1467 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1468 record_for_igvn(st);
1470 return st;
1471 }
1474 void GraphKit::pre_barrier(bool do_load,
1475 Node* ctl,
1476 Node* obj,
1477 Node* adr,
1478 uint adr_idx,
1479 Node* val,
1480 const TypeOopPtr* val_type,
1481 Node* pre_val,
1482 BasicType bt) {
1484 BarrierSet* bs = Universe::heap()->barrier_set();
1485 set_control(ctl);
1486 switch (bs->kind()) {
1487 case BarrierSet::G1SATBCT:
1488 case BarrierSet::G1SATBCTLogging:
1489 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1490 break;
1492 case BarrierSet::CardTableModRef:
1493 case BarrierSet::CardTableExtension:
1494 case BarrierSet::ModRef:
1495 break;
1497 case BarrierSet::Other:
1498 default :
1499 ShouldNotReachHere();
1501 }
1502 }
1504 bool GraphKit::can_move_pre_barrier() const {
1505 BarrierSet* bs = Universe::heap()->barrier_set();
1506 switch (bs->kind()) {
1507 case BarrierSet::G1SATBCT:
1508 case BarrierSet::G1SATBCTLogging:
1509 return true; // Can move it if no safepoint
1511 case BarrierSet::CardTableModRef:
1512 case BarrierSet::CardTableExtension:
1513 case BarrierSet::ModRef:
1514 return true; // There is no pre-barrier
1516 case BarrierSet::Other:
1517 default :
1518 ShouldNotReachHere();
1519 }
1520 return false;
1521 }
1523 void GraphKit::post_barrier(Node* ctl,
1524 Node* store,
1525 Node* obj,
1526 Node* adr,
1527 uint adr_idx,
1528 Node* val,
1529 BasicType bt,
1530 bool use_precise) {
1531 BarrierSet* bs = Universe::heap()->barrier_set();
1532 set_control(ctl);
1533 switch (bs->kind()) {
1534 case BarrierSet::G1SATBCT:
1535 case BarrierSet::G1SATBCTLogging:
1536 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1537 break;
1539 case BarrierSet::CardTableModRef:
1540 case BarrierSet::CardTableExtension:
1541 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1542 break;
1544 case BarrierSet::ModRef:
1545 break;
1547 case BarrierSet::Other:
1548 default :
1549 ShouldNotReachHere();
1551 }
1552 }
1554 Node* GraphKit::store_oop(Node* ctl,
1555 Node* obj,
1556 Node* adr,
1557 const TypePtr* adr_type,
1558 Node* val,
1559 const TypeOopPtr* val_type,
1560 BasicType bt,
1561 bool use_precise) {
1562 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1563 // could be delayed during Parse (for example, in adjust_map_after_if()).
1564 // Execute transformation here to avoid barrier generation in such case.
1565 if (_gvn.type(val) == TypePtr::NULL_PTR)
1566 val = _gvn.makecon(TypePtr::NULL_PTR);
1568 set_control(ctl);
1569 if (stopped()) return top(); // Dead path ?
1571 assert(bt == T_OBJECT, "sanity");
1572 assert(val != NULL, "not dead path");
1573 uint adr_idx = C->get_alias_index(adr_type);
1574 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1576 pre_barrier(true /* do_load */,
1577 control(), obj, adr, adr_idx, val, val_type,
1578 NULL /* pre_val */,
1579 bt);
1581 Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1582 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1583 return store;
1584 }
1586 // Could be an array or object we don't know at compile time (unsafe ref.)
1587 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1588 Node* obj, // containing obj
1589 Node* adr, // actual adress to store val at
1590 const TypePtr* adr_type,
1591 Node* val,
1592 BasicType bt) {
1593 Compile::AliasType* at = C->alias_type(adr_type);
1594 const TypeOopPtr* val_type = NULL;
1595 if (adr_type->isa_instptr()) {
1596 if (at->field() != NULL) {
1597 // known field. This code is a copy of the do_put_xxx logic.
1598 ciField* field = at->field();
1599 if (!field->type()->is_loaded()) {
1600 val_type = TypeInstPtr::BOTTOM;
1601 } else {
1602 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1603 }
1604 }
1605 } else if (adr_type->isa_aryptr()) {
1606 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1607 }
1608 if (val_type == NULL) {
1609 val_type = TypeInstPtr::BOTTOM;
1610 }
1611 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true);
1612 }
1615 //-------------------------array_element_address-------------------------
1616 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1617 const TypeInt* sizetype) {
1618 uint shift = exact_log2(type2aelembytes(elembt));
1619 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1621 // short-circuit a common case (saves lots of confusing waste motion)
1622 jint idx_con = find_int_con(idx, -1);
1623 if (idx_con >= 0) {
1624 intptr_t offset = header + ((intptr_t)idx_con << shift);
1625 return basic_plus_adr(ary, offset);
1626 }
1628 // must be correct type for alignment purposes
1629 Node* base = basic_plus_adr(ary, header);
1630 #ifdef _LP64
1631 // The scaled index operand to AddP must be a clean 64-bit value.
1632 // Java allows a 32-bit int to be incremented to a negative
1633 // value, which appears in a 64-bit register as a large
1634 // positive number. Using that large positive number as an
1635 // operand in pointer arithmetic has bad consequences.
1636 // On the other hand, 32-bit overflow is rare, and the possibility
1637 // can often be excluded, if we annotate the ConvI2L node with
1638 // a type assertion that its value is known to be a small positive
1639 // number. (The prior range check has ensured this.)
1640 // This assertion is used by ConvI2LNode::Ideal.
1641 int index_max = max_jint - 1; // array size is max_jint, index is one less
1642 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1643 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1644 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) );
1645 #endif
1646 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1647 return basic_plus_adr(ary, base, scale);
1648 }
1650 //-------------------------load_array_element-------------------------
1651 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1652 const Type* elemtype = arytype->elem();
1653 BasicType elembt = elemtype->array_element_basic_type();
1654 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1655 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1656 return ld;
1657 }
1659 //-------------------------set_arguments_for_java_call-------------------------
1660 // Arguments (pre-popped from the stack) are taken from the JVMS.
1661 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1662 // Add the call arguments:
1663 uint nargs = call->method()->arg_size();
1664 for (uint i = 0; i < nargs; i++) {
1665 Node* arg = argument(i);
1666 call->init_req(i + TypeFunc::Parms, arg);
1667 }
1668 }
1670 //---------------------------set_edges_for_java_call---------------------------
1671 // Connect a newly created call into the current JVMS.
1672 // A return value node (if any) is returned from set_edges_for_java_call.
1673 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1675 // Add the predefined inputs:
1676 call->init_req( TypeFunc::Control, control() );
1677 call->init_req( TypeFunc::I_O , i_o() );
1678 call->init_req( TypeFunc::Memory , reset_memory() );
1679 call->init_req( TypeFunc::FramePtr, frameptr() );
1680 call->init_req( TypeFunc::ReturnAdr, top() );
1682 add_safepoint_edges(call, must_throw);
1684 Node* xcall = _gvn.transform(call);
1686 if (xcall == top()) {
1687 set_control(top());
1688 return;
1689 }
1690 assert(xcall == call, "call identity is stable");
1692 // Re-use the current map to produce the result.
1694 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1695 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1696 set_all_memory_call(xcall, separate_io_proj);
1698 //return xcall; // no need, caller already has it
1699 }
1701 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1702 if (stopped()) return top(); // maybe the call folded up?
1704 // Capture the return value, if any.
1705 Node* ret;
1706 if (call->method() == NULL ||
1707 call->method()->return_type()->basic_type() == T_VOID)
1708 ret = top();
1709 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1711 // Note: Since any out-of-line call can produce an exception,
1712 // we always insert an I_O projection from the call into the result.
1714 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1716 if (separate_io_proj) {
1717 // The caller requested separate projections be used by the fall
1718 // through and exceptional paths, so replace the projections for
1719 // the fall through path.
1720 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1721 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1722 }
1723 return ret;
1724 }
1726 //--------------------set_predefined_input_for_runtime_call--------------------
1727 // Reading and setting the memory state is way conservative here.
1728 // The real problem is that I am not doing real Type analysis on memory,
1729 // so I cannot distinguish card mark stores from other stores. Across a GC
1730 // point the Store Barrier and the card mark memory has to agree. I cannot
1731 // have a card mark store and its barrier split across the GC point from
1732 // either above or below. Here I get that to happen by reading ALL of memory.
1733 // A better answer would be to separate out card marks from other memory.
1734 // For now, return the input memory state, so that it can be reused
1735 // after the call, if this call has restricted memory effects.
1736 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1737 // Set fixed predefined input arguments
1738 Node* memory = reset_memory();
1739 call->init_req( TypeFunc::Control, control() );
1740 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1741 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1742 call->init_req( TypeFunc::FramePtr, frameptr() );
1743 call->init_req( TypeFunc::ReturnAdr, top() );
1744 return memory;
1745 }
1747 //-------------------set_predefined_output_for_runtime_call--------------------
1748 // Set control and memory (not i_o) from the call.
1749 // If keep_mem is not NULL, use it for the output state,
1750 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1751 // If hook_mem is NULL, this call produces no memory effects at all.
1752 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1753 // then only that memory slice is taken from the call.
1754 // In the last case, we must put an appropriate memory barrier before
1755 // the call, so as to create the correct anti-dependencies on loads
1756 // preceding the call.
1757 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1758 Node* keep_mem,
1759 const TypePtr* hook_mem) {
1760 // no i/o
1761 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1762 if (keep_mem) {
1763 // First clone the existing memory state
1764 set_all_memory(keep_mem);
1765 if (hook_mem != NULL) {
1766 // Make memory for the call
1767 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1768 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1769 // We also use hook_mem to extract specific effects from arraycopy stubs.
1770 set_memory(mem, hook_mem);
1771 }
1772 // ...else the call has NO memory effects.
1774 // Make sure the call advertises its memory effects precisely.
1775 // This lets us build accurate anti-dependences in gcm.cpp.
1776 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1777 "call node must be constructed correctly");
1778 } else {
1779 assert(hook_mem == NULL, "");
1780 // This is not a "slow path" call; all memory comes from the call.
1781 set_all_memory_call(call);
1782 }
1783 }
1786 // Replace the call with the current state of the kit.
1787 void GraphKit::replace_call(CallNode* call, Node* result) {
1788 JVMState* ejvms = NULL;
1789 if (has_exceptions()) {
1790 ejvms = transfer_exceptions_into_jvms();
1791 }
1793 SafePointNode* final_state = stop();
1795 // Find all the needed outputs of this call
1796 CallProjections callprojs;
1797 call->extract_projections(&callprojs, true);
1799 Node* init_mem = call->in(TypeFunc::Memory);
1800 Node* final_mem = final_state->in(TypeFunc::Memory);
1801 Node* final_ctl = final_state->in(TypeFunc::Control);
1802 Node* final_io = final_state->in(TypeFunc::I_O);
1804 // Replace all the old call edges with the edges from the inlining result
1805 if (callprojs.fallthrough_catchproj != NULL) {
1806 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1807 }
1808 if (callprojs.fallthrough_memproj != NULL) {
1809 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1810 }
1811 if (callprojs.fallthrough_ioproj != NULL) {
1812 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1813 }
1815 // Replace the result with the new result if it exists and is used
1816 if (callprojs.resproj != NULL && result != NULL) {
1817 C->gvn_replace_by(callprojs.resproj, result);
1818 }
1820 if (ejvms == NULL) {
1821 // No exception edges to simply kill off those paths
1822 if (callprojs.catchall_catchproj != NULL) {
1823 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1824 }
1825 if (callprojs.catchall_memproj != NULL) {
1826 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1827 }
1828 if (callprojs.catchall_ioproj != NULL) {
1829 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1830 }
1831 // Replace the old exception object with top
1832 if (callprojs.exobj != NULL) {
1833 C->gvn_replace_by(callprojs.exobj, C->top());
1834 }
1835 } else {
1836 GraphKit ekit(ejvms);
1838 // Load my combined exception state into the kit, with all phis transformed:
1839 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1841 Node* ex_oop = ekit.use_exception_state(ex_map);
1842 if (callprojs.catchall_catchproj != NULL) {
1843 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1844 }
1845 if (callprojs.catchall_memproj != NULL) {
1846 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1847 }
1848 if (callprojs.catchall_ioproj != NULL) {
1849 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1850 }
1852 // Replace the old exception object with the newly created one
1853 if (callprojs.exobj != NULL) {
1854 C->gvn_replace_by(callprojs.exobj, ex_oop);
1855 }
1856 }
1858 // Disconnect the call from the graph
1859 call->disconnect_inputs(NULL, C);
1860 C->gvn_replace_by(call, C->top());
1862 // Clean up any MergeMems that feed other MergeMems since the
1863 // optimizer doesn't like that.
1864 if (final_mem->is_MergeMem()) {
1865 Node_List wl;
1866 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1867 Node* m = i.get();
1868 if (m->is_MergeMem() && !wl.contains(m)) {
1869 wl.push(m);
1870 }
1871 }
1872 while (wl.size() > 0) {
1873 _gvn.transform(wl.pop());
1874 }
1875 }
1876 }
1879 //------------------------------increment_counter------------------------------
1880 // for statistics: increment a VM counter by 1
1882 void GraphKit::increment_counter(address counter_addr) {
1883 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1884 increment_counter(adr1);
1885 }
1887 void GraphKit::increment_counter(Node* counter_addr) {
1888 int adr_type = Compile::AliasIdxRaw;
1889 Node* ctrl = control();
1890 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type);
1891 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1892 store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type );
1893 }
1896 //------------------------------uncommon_trap----------------------------------
1897 // Bail out to the interpreter in mid-method. Implemented by calling the
1898 // uncommon_trap blob. This helper function inserts a runtime call with the
1899 // right debug info.
1900 void GraphKit::uncommon_trap(int trap_request,
1901 ciKlass* klass, const char* comment,
1902 bool must_throw,
1903 bool keep_exact_action) {
1904 if (failing()) stop();
1905 if (stopped()) return; // trap reachable?
1907 // Note: If ProfileTraps is true, and if a deopt. actually
1908 // occurs here, the runtime will make sure an MDO exists. There is
1909 // no need to call method()->ensure_method_data() at this point.
1911 // Set the stack pointer to the right value for reexecution:
1912 set_sp(reexecute_sp());
1914 #ifdef ASSERT
1915 if (!must_throw) {
1916 // Make sure the stack has at least enough depth to execute
1917 // the current bytecode.
1918 int inputs, ignored_depth;
1919 if (compute_stack_effects(inputs, ignored_depth)) {
1920 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1921 Bytecodes::name(java_bc()), sp(), inputs));
1922 }
1923 }
1924 #endif
1926 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1927 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1929 switch (action) {
1930 case Deoptimization::Action_maybe_recompile:
1931 case Deoptimization::Action_reinterpret:
1932 // Temporary fix for 6529811 to allow virtual calls to be sure they
1933 // get the chance to go from mono->bi->mega
1934 if (!keep_exact_action &&
1935 Deoptimization::trap_request_index(trap_request) < 0 &&
1936 too_many_recompiles(reason)) {
1937 // This BCI is causing too many recompilations.
1938 action = Deoptimization::Action_none;
1939 trap_request = Deoptimization::make_trap_request(reason, action);
1940 } else {
1941 C->set_trap_can_recompile(true);
1942 }
1943 break;
1944 case Deoptimization::Action_make_not_entrant:
1945 C->set_trap_can_recompile(true);
1946 break;
1947 #ifdef ASSERT
1948 case Deoptimization::Action_none:
1949 case Deoptimization::Action_make_not_compilable:
1950 break;
1951 default:
1952 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
1953 break;
1954 #endif
1955 }
1957 if (TraceOptoParse) {
1958 char buf[100];
1959 tty->print_cr("Uncommon trap %s at bci:%d",
1960 Deoptimization::format_trap_request(buf, sizeof(buf),
1961 trap_request), bci());
1962 }
1964 CompileLog* log = C->log();
1965 if (log != NULL) {
1966 int kid = (klass == NULL)? -1: log->identify(klass);
1967 log->begin_elem("uncommon_trap bci='%d'", bci());
1968 char buf[100];
1969 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1970 trap_request));
1971 if (kid >= 0) log->print(" klass='%d'", kid);
1972 if (comment != NULL) log->print(" comment='%s'", comment);
1973 log->end_elem();
1974 }
1976 // Make sure any guarding test views this path as very unlikely
1977 Node *i0 = control()->in(0);
1978 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1979 IfNode *iff = i0->as_If();
1980 float f = iff->_prob; // Get prob
1981 if (control()->Opcode() == Op_IfTrue) {
1982 if (f > PROB_UNLIKELY_MAG(4))
1983 iff->_prob = PROB_MIN;
1984 } else {
1985 if (f < PROB_LIKELY_MAG(4))
1986 iff->_prob = PROB_MAX;
1987 }
1988 }
1990 // Clear out dead values from the debug info.
1991 kill_dead_locals();
1993 // Now insert the uncommon trap subroutine call
1994 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
1995 const TypePtr* no_memory_effects = NULL;
1996 // Pass the index of the class to be loaded
1997 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1998 (must_throw ? RC_MUST_THROW : 0),
1999 OptoRuntime::uncommon_trap_Type(),
2000 call_addr, "uncommon_trap", no_memory_effects,
2001 intcon(trap_request));
2002 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2003 "must extract request correctly from the graph");
2004 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2006 call->set_req(TypeFunc::ReturnAdr, returnadr());
2007 // The debug info is the only real input to this call.
2009 // Halt-and-catch fire here. The above call should never return!
2010 HaltNode* halt = new(C) HaltNode(control(), frameptr());
2011 _gvn.set_type_bottom(halt);
2012 root()->add_req(halt);
2014 stop_and_kill_map();
2015 }
2018 //--------------------------just_allocated_object------------------------------
2019 // Report the object that was just allocated.
2020 // It must be the case that there are no intervening safepoints.
2021 // We use this to determine if an object is so "fresh" that
2022 // it does not require card marks.
2023 Node* GraphKit::just_allocated_object(Node* current_control) {
2024 if (C->recent_alloc_ctl() == current_control)
2025 return C->recent_alloc_obj();
2026 return NULL;
2027 }
2030 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2031 // (Note: TypeFunc::make has a cache that makes this fast.)
2032 const TypeFunc* tf = TypeFunc::make(dest_method);
2033 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2034 for (int j = 0; j < nargs; j++) {
2035 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2036 if( targ->basic_type() == T_DOUBLE ) {
2037 // If any parameters are doubles, they must be rounded before
2038 // the call, dstore_rounding does gvn.transform
2039 Node *arg = argument(j);
2040 arg = dstore_rounding(arg);
2041 set_argument(j, arg);
2042 }
2043 }
2044 }
2046 void GraphKit::round_double_result(ciMethod* dest_method) {
2047 // A non-strict method may return a double value which has an extended
2048 // exponent, but this must not be visible in a caller which is 'strict'
2049 // If a strict caller invokes a non-strict callee, round a double result
2051 BasicType result_type = dest_method->return_type()->basic_type();
2052 assert( method() != NULL, "must have caller context");
2053 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2054 // Destination method's return value is on top of stack
2055 // dstore_rounding() does gvn.transform
2056 Node *result = pop_pair();
2057 result = dstore_rounding(result);
2058 push_pair(result);
2059 }
2060 }
2062 // rounding for strict float precision conformance
2063 Node* GraphKit::precision_rounding(Node* n) {
2064 return UseStrictFP && _method->flags().is_strict()
2065 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2066 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2067 : n;
2068 }
2070 // rounding for strict double precision conformance
2071 Node* GraphKit::dprecision_rounding(Node *n) {
2072 return UseStrictFP && _method->flags().is_strict()
2073 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2074 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2075 : n;
2076 }
2078 // rounding for non-strict double stores
2079 Node* GraphKit::dstore_rounding(Node* n) {
2080 return Matcher::strict_fp_requires_explicit_rounding
2081 && UseSSE <= 1
2082 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2083 : n;
2084 }
2086 //=============================================================================
2087 // Generate a fast path/slow path idiom. Graph looks like:
2088 // [foo] indicates that 'foo' is a parameter
2089 //
2090 // [in] NULL
2091 // \ /
2092 // CmpP
2093 // Bool ne
2094 // If
2095 // / \
2096 // True False-<2>
2097 // / |
2098 // / cast_not_null
2099 // Load | | ^
2100 // [fast_test] | |
2101 // gvn to opt_test | |
2102 // / \ | <1>
2103 // True False |
2104 // | \\ |
2105 // [slow_call] \[fast_result]
2106 // Ctl Val \ \
2107 // | \ \
2108 // Catch <1> \ \
2109 // / \ ^ \ \
2110 // Ex No_Ex | \ \
2111 // | \ \ | \ <2> \
2112 // ... \ [slow_res] | | \ [null_result]
2113 // \ \--+--+--- | |
2114 // \ | / \ | /
2115 // --------Region Phi
2116 //
2117 //=============================================================================
2118 // Code is structured as a series of driver functions all called 'do_XXX' that
2119 // call a set of helper functions. Helper functions first, then drivers.
2121 //------------------------------null_check_oop---------------------------------
2122 // Null check oop. Set null-path control into Region in slot 3.
2123 // Make a cast-not-nullness use the other not-null control. Return cast.
2124 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2125 bool never_see_null) {
2126 // Initial NULL check taken path
2127 (*null_control) = top();
2128 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2130 // Generate uncommon_trap:
2131 if (never_see_null && (*null_control) != top()) {
2132 // If we see an unexpected null at a check-cast we record it and force a
2133 // recompile; the offending check-cast will be compiled to handle NULLs.
2134 // If we see more than one offending BCI, then all checkcasts in the
2135 // method will be compiled to handle NULLs.
2136 PreserveJVMState pjvms(this);
2137 set_control(*null_control);
2138 replace_in_map(value, null());
2139 uncommon_trap(Deoptimization::Reason_null_check,
2140 Deoptimization::Action_make_not_entrant);
2141 (*null_control) = top(); // NULL path is dead
2142 }
2144 // Cast away null-ness on the result
2145 return cast;
2146 }
2148 //------------------------------opt_iff----------------------------------------
2149 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2150 // Return slow-path control.
2151 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2152 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2154 // Fast path taken; set region slot 2
2155 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2156 region->init_req(2,fast_taken); // Capture fast-control
2158 // Fast path not-taken, i.e. slow path
2159 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2160 return slow_taken;
2161 }
2163 //-----------------------------make_runtime_call-------------------------------
2164 Node* GraphKit::make_runtime_call(int flags,
2165 const TypeFunc* call_type, address call_addr,
2166 const char* call_name,
2167 const TypePtr* adr_type,
2168 // The following parms are all optional.
2169 // The first NULL ends the list.
2170 Node* parm0, Node* parm1,
2171 Node* parm2, Node* parm3,
2172 Node* parm4, Node* parm5,
2173 Node* parm6, Node* parm7) {
2174 // Slow-path call
2175 bool is_leaf = !(flags & RC_NO_LEAF);
2176 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2177 if (call_name == NULL) {
2178 assert(!is_leaf, "must supply name for leaf");
2179 call_name = OptoRuntime::stub_name(call_addr);
2180 }
2181 CallNode* call;
2182 if (!is_leaf) {
2183 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2184 bci(), adr_type);
2185 } else if (flags & RC_NO_FP) {
2186 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2187 } else {
2188 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2189 }
2191 // The following is similar to set_edges_for_java_call,
2192 // except that the memory effects of the call are restricted to AliasIdxRaw.
2194 // Slow path call has no side-effects, uses few values
2195 bool wide_in = !(flags & RC_NARROW_MEM);
2196 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2198 Node* prev_mem = NULL;
2199 if (wide_in) {
2200 prev_mem = set_predefined_input_for_runtime_call(call);
2201 } else {
2202 assert(!wide_out, "narrow in => narrow out");
2203 Node* narrow_mem = memory(adr_type);
2204 prev_mem = reset_memory();
2205 map()->set_memory(narrow_mem);
2206 set_predefined_input_for_runtime_call(call);
2207 }
2209 // Hook each parm in order. Stop looking at the first NULL.
2210 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2211 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2212 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2213 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2214 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2215 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2216 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2217 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2218 /* close each nested if ===> */ } } } } } } } }
2219 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2221 if (!is_leaf) {
2222 // Non-leaves can block and take safepoints:
2223 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2224 }
2225 // Non-leaves can throw exceptions:
2226 if (has_io) {
2227 call->set_req(TypeFunc::I_O, i_o());
2228 }
2230 if (flags & RC_UNCOMMON) {
2231 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2232 // (An "if" probability corresponds roughly to an unconditional count.
2233 // Sort of.)
2234 call->set_cnt(PROB_UNLIKELY_MAG(4));
2235 }
2237 Node* c = _gvn.transform(call);
2238 assert(c == call, "cannot disappear");
2240 if (wide_out) {
2241 // Slow path call has full side-effects.
2242 set_predefined_output_for_runtime_call(call);
2243 } else {
2244 // Slow path call has few side-effects, and/or sets few values.
2245 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2246 }
2248 if (has_io) {
2249 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2250 }
2251 return call;
2253 }
2255 //------------------------------merge_memory-----------------------------------
2256 // Merge memory from one path into the current memory state.
2257 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2258 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2259 Node* old_slice = mms.force_memory();
2260 Node* new_slice = mms.memory2();
2261 if (old_slice != new_slice) {
2262 PhiNode* phi;
2263 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2264 phi = new_slice->as_Phi();
2265 #ifdef ASSERT
2266 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2267 old_slice = old_slice->in(new_path);
2268 // Caller is responsible for ensuring that any pre-existing
2269 // phis are already aware of old memory.
2270 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2271 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2272 #endif
2273 mms.set_memory(phi);
2274 } else {
2275 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2276 _gvn.set_type(phi, Type::MEMORY);
2277 phi->set_req(new_path, new_slice);
2278 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2279 }
2280 }
2281 }
2282 }
2284 //------------------------------make_slow_call_ex------------------------------
2285 // Make the exception handler hookups for the slow call
2286 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2287 if (stopped()) return;
2289 // Make a catch node with just two handlers: fall-through and catch-all
2290 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2291 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2292 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2293 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2295 { PreserveJVMState pjvms(this);
2296 set_control(excp);
2297 set_i_o(i_o);
2299 if (excp != top()) {
2300 // Create an exception state also.
2301 // Use an exact type if the caller has specified a specific exception.
2302 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2303 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2304 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2305 }
2306 }
2308 // Get the no-exception control from the CatchNode.
2309 set_control(norm);
2310 }
2313 //-------------------------------gen_subtype_check-----------------------------
2314 // Generate a subtyping check. Takes as input the subtype and supertype.
2315 // Returns 2 values: sets the default control() to the true path and returns
2316 // the false path. Only reads invariant memory; sets no (visible) memory.
2317 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2318 // but that's not exposed to the optimizer. This call also doesn't take in an
2319 // Object; if you wish to check an Object you need to load the Object's class
2320 // prior to coming here.
2321 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2322 // Fast check for identical types, perhaps identical constants.
2323 // The types can even be identical non-constants, in cases
2324 // involving Array.newInstance, Object.clone, etc.
2325 if (subklass == superklass)
2326 return top(); // false path is dead; no test needed.
2328 if (_gvn.type(superklass)->singleton()) {
2329 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2330 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2332 // In the common case of an exact superklass, try to fold up the
2333 // test before generating code. You may ask, why not just generate
2334 // the code and then let it fold up? The answer is that the generated
2335 // code will necessarily include null checks, which do not always
2336 // completely fold away. If they are also needless, then they turn
2337 // into a performance loss. Example:
2338 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2339 // Here, the type of 'fa' is often exact, so the store check
2340 // of fa[1]=x will fold up, without testing the nullness of x.
2341 switch (static_subtype_check(superk, subk)) {
2342 case SSC_always_false:
2343 {
2344 Node* always_fail = control();
2345 set_control(top());
2346 return always_fail;
2347 }
2348 case SSC_always_true:
2349 return top();
2350 case SSC_easy_test:
2351 {
2352 // Just do a direct pointer compare and be done.
2353 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2354 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2355 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2356 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2357 return _gvn.transform( new(C) IfFalseNode(iff) );
2358 }
2359 case SSC_full_test:
2360 break;
2361 default:
2362 ShouldNotReachHere();
2363 }
2364 }
2366 // %%% Possible further optimization: Even if the superklass is not exact,
2367 // if the subklass is the unique subtype of the superklass, the check
2368 // will always succeed. We could leave a dependency behind to ensure this.
2370 // First load the super-klass's check-offset
2371 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2372 Node *chk_off = _gvn.transform( new (C) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2373 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2374 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2376 // Load from the sub-klass's super-class display list, or a 1-word cache of
2377 // the secondary superclass list, or a failing value with a sentinel offset
2378 // if the super-klass is an interface or exceptionally deep in the Java
2379 // hierarchy and we have to scan the secondary superclass list the hard way.
2380 // Worst-case type is a little odd: NULL is allowed as a result (usually
2381 // klass loads can never produce a NULL).
2382 Node *chk_off_X = ConvI2X(chk_off);
2383 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2384 // For some types like interfaces the following loadKlass is from a 1-word
2385 // cache which is mutable so can't use immutable memory. Other
2386 // types load from the super-class display table which is immutable.
2387 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2388 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2390 // Compile speed common case: ARE a subtype and we canNOT fail
2391 if( superklass == nkls )
2392 return top(); // false path is dead; no test needed.
2394 // See if we get an immediate positive hit. Happens roughly 83% of the
2395 // time. Test to see if the value loaded just previously from the subklass
2396 // is exactly the superklass.
2397 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2398 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2399 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2400 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2401 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2403 // Compile speed common case: Check for being deterministic right now. If
2404 // chk_off is a constant and not equal to cacheoff then we are NOT a
2405 // subklass. In this case we need exactly the 1 test above and we can
2406 // return those results immediately.
2407 if (!might_be_cache) {
2408 Node* not_subtype_ctrl = control();
2409 set_control(iftrue1); // We need exactly the 1 test above
2410 return not_subtype_ctrl;
2411 }
2413 // Gather the various success & failures here
2414 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2415 record_for_igvn(r_ok_subtype);
2416 RegionNode *r_not_subtype = new (C) RegionNode(3);
2417 record_for_igvn(r_not_subtype);
2419 r_ok_subtype->init_req(1, iftrue1);
2421 // Check for immediate negative hit. Happens roughly 11% of the time (which
2422 // is roughly 63% of the remaining cases). Test to see if the loaded
2423 // check-offset points into the subklass display list or the 1-element
2424 // cache. If it points to the display (and NOT the cache) and the display
2425 // missed then it's not a subtype.
2426 Node *cacheoff = _gvn.intcon(cacheoff_con);
2427 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2428 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2429 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2430 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2431 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2433 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2434 // No performance impact (too rare) but allows sharing of secondary arrays
2435 // which has some footprint reduction.
2436 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2437 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2438 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2439 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2440 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2442 // -- Roads not taken here: --
2443 // We could also have chosen to perform the self-check at the beginning
2444 // of this code sequence, as the assembler does. This would not pay off
2445 // the same way, since the optimizer, unlike the assembler, can perform
2446 // static type analysis to fold away many successful self-checks.
2447 // Non-foldable self checks work better here in second position, because
2448 // the initial primary superclass check subsumes a self-check for most
2449 // types. An exception would be a secondary type like array-of-interface,
2450 // which does not appear in its own primary supertype display.
2451 // Finally, we could have chosen to move the self-check into the
2452 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2453 // dependent manner. But it is worthwhile to have the check here,
2454 // where it can be perhaps be optimized. The cost in code space is
2455 // small (register compare, branch).
2457 // Now do a linear scan of the secondary super-klass array. Again, no real
2458 // performance impact (too rare) but it's gotta be done.
2459 // Since the code is rarely used, there is no penalty for moving it
2460 // out of line, and it can only improve I-cache density.
2461 // The decision to inline or out-of-line this final check is platform
2462 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2463 Node* psc = _gvn.transform(
2464 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2466 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2467 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2468 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2469 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2470 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2472 // Return false path; set default control to true path.
2473 set_control( _gvn.transform(r_ok_subtype) );
2474 return _gvn.transform(r_not_subtype);
2475 }
2477 //----------------------------static_subtype_check-----------------------------
2478 // Shortcut important common cases when superklass is exact:
2479 // (0) superklass is java.lang.Object (can occur in reflective code)
2480 // (1) subklass is already limited to a subtype of superklass => always ok
2481 // (2) subklass does not overlap with superklass => always fail
2482 // (3) superklass has NO subtypes and we can check with a simple compare.
2483 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2484 if (StressReflectiveCode) {
2485 return SSC_full_test; // Let caller generate the general case.
2486 }
2488 if (superk == env()->Object_klass()) {
2489 return SSC_always_true; // (0) this test cannot fail
2490 }
2492 ciType* superelem = superk;
2493 if (superelem->is_array_klass())
2494 superelem = superelem->as_array_klass()->base_element_type();
2496 if (!subk->is_interface()) { // cannot trust static interface types yet
2497 if (subk->is_subtype_of(superk)) {
2498 return SSC_always_true; // (1) false path dead; no dynamic test needed
2499 }
2500 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2501 !superk->is_subtype_of(subk)) {
2502 return SSC_always_false;
2503 }
2504 }
2506 // If casting to an instance klass, it must have no subtypes
2507 if (superk->is_interface()) {
2508 // Cannot trust interfaces yet.
2509 // %%% S.B. superk->nof_implementors() == 1
2510 } else if (superelem->is_instance_klass()) {
2511 ciInstanceKlass* ik = superelem->as_instance_klass();
2512 if (!ik->has_subklass() && !ik->is_interface()) {
2513 if (!ik->is_final()) {
2514 // Add a dependency if there is a chance of a later subclass.
2515 C->dependencies()->assert_leaf_type(ik);
2516 }
2517 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2518 }
2519 } else {
2520 // A primitive array type has no subtypes.
2521 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2522 }
2524 return SSC_full_test;
2525 }
2527 // Profile-driven exact type check:
2528 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2529 float prob,
2530 Node* *casted_receiver) {
2531 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2532 Node* recv_klass = load_object_klass(receiver);
2533 Node* want_klass = makecon(tklass);
2534 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2535 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2536 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2537 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2538 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2540 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2541 assert(recv_xtype->klass_is_exact(), "");
2543 // Subsume downstream occurrences of receiver with a cast to
2544 // recv_xtype, since now we know what the type will be.
2545 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2546 (*casted_receiver) = _gvn.transform(cast);
2547 // (User must make the replace_in_map call.)
2549 return fail;
2550 }
2553 //------------------------------seems_never_null-------------------------------
2554 // Use null_seen information if it is available from the profile.
2555 // If we see an unexpected null at a type check we record it and force a
2556 // recompile; the offending check will be recompiled to handle NULLs.
2557 // If we see several offending BCIs, then all checks in the
2558 // method will be recompiled.
2559 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2560 if (UncommonNullCast // Cutout for this technique
2561 && obj != null() // And not the -Xcomp stupid case?
2562 && !too_many_traps(Deoptimization::Reason_null_check)
2563 ) {
2564 if (data == NULL)
2565 // Edge case: no mature data. Be optimistic here.
2566 return true;
2567 // If the profile has not seen a null, assume it won't happen.
2568 assert(java_bc() == Bytecodes::_checkcast ||
2569 java_bc() == Bytecodes::_instanceof ||
2570 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2571 return !data->as_BitData()->null_seen();
2572 }
2573 return false;
2574 }
2576 //------------------------maybe_cast_profiled_receiver-------------------------
2577 // If the profile has seen exactly one type, narrow to exactly that type.
2578 // Subsequent type checks will always fold up.
2579 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2580 ciProfileData* data,
2581 ciKlass* require_klass) {
2582 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2583 if (data == NULL) return NULL;
2585 // Make sure we haven't already deoptimized from this tactic.
2586 if (too_many_traps(Deoptimization::Reason_class_check))
2587 return NULL;
2589 // (No, this isn't a call, but it's enough like a virtual call
2590 // to use the same ciMethod accessor to get the profile info...)
2591 ciCallProfile profile = method()->call_profile_at_bci(bci());
2592 if (profile.count() >= 0 && // no cast failures here
2593 profile.has_receiver(0) &&
2594 profile.morphism() == 1) {
2595 ciKlass* exact_kls = profile.receiver(0);
2596 if (require_klass == NULL ||
2597 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2598 // If we narrow the type to match what the type profile sees,
2599 // we can then remove the rest of the cast.
2600 // This is a win, even if the exact_kls is very specific,
2601 // because downstream operations, such as method calls,
2602 // will often benefit from the sharper type.
2603 Node* exact_obj = not_null_obj; // will get updated in place...
2604 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2605 &exact_obj);
2606 { PreserveJVMState pjvms(this);
2607 set_control(slow_ctl);
2608 uncommon_trap(Deoptimization::Reason_class_check,
2609 Deoptimization::Action_maybe_recompile);
2610 }
2611 replace_in_map(not_null_obj, exact_obj);
2612 return exact_obj;
2613 }
2614 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2615 }
2617 return NULL;
2618 }
2621 //-------------------------------gen_instanceof--------------------------------
2622 // Generate an instance-of idiom. Used by both the instance-of bytecode
2623 // and the reflective instance-of call.
2624 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass) {
2625 kill_dead_locals(); // Benefit all the uncommon traps
2626 assert( !stopped(), "dead parse path should be checked in callers" );
2627 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2628 "must check for not-null not-dead klass in callers");
2630 // Make the merge point
2631 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2632 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2633 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2634 C->set_has_split_ifs(true); // Has chance for split-if optimization
2636 ciProfileData* data = NULL;
2637 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2638 data = method()->method_data()->bci_to_data(bci());
2639 }
2640 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2641 && seems_never_null(obj, data));
2643 // Null check; get casted pointer; set region slot 3
2644 Node* null_ctl = top();
2645 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2647 // If not_null_obj is dead, only null-path is taken
2648 if (stopped()) { // Doing instance-of on a NULL?
2649 set_control(null_ctl);
2650 return intcon(0);
2651 }
2652 region->init_req(_null_path, null_ctl);
2653 phi ->init_req(_null_path, intcon(0)); // Set null path value
2654 if (null_ctl == top()) {
2655 // Do this eagerly, so that pattern matches like is_diamond_phi
2656 // will work even during parsing.
2657 assert(_null_path == PATH_LIMIT-1, "delete last");
2658 region->del_req(_null_path);
2659 phi ->del_req(_null_path);
2660 }
2662 if (ProfileDynamicTypes && data != NULL) {
2663 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, NULL);
2664 if (stopped()) { // Profile disagrees with this path.
2665 set_control(null_ctl); // Null is the only remaining possibility.
2666 return intcon(0);
2667 }
2668 if (cast_obj != NULL)
2669 not_null_obj = cast_obj;
2670 }
2672 // Load the object's klass
2673 Node* obj_klass = load_object_klass(not_null_obj);
2675 // Generate the subtype check
2676 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2678 // Plug in the success path to the general merge in slot 1.
2679 region->init_req(_obj_path, control());
2680 phi ->init_req(_obj_path, intcon(1));
2682 // Plug in the failing path to the general merge in slot 2.
2683 region->init_req(_fail_path, not_subtype_ctrl);
2684 phi ->init_req(_fail_path, intcon(0));
2686 // Return final merged results
2687 set_control( _gvn.transform(region) );
2688 record_for_igvn(region);
2689 return _gvn.transform(phi);
2690 }
2692 //-------------------------------gen_checkcast---------------------------------
2693 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2694 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2695 // uncommon-trap paths work. Adjust stack after this call.
2696 // If failure_control is supplied and not null, it is filled in with
2697 // the control edge for the cast failure. Otherwise, an appropriate
2698 // uncommon trap or exception is thrown.
2699 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2700 Node* *failure_control) {
2701 kill_dead_locals(); // Benefit all the uncommon traps
2702 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2703 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2705 // Fast cutout: Check the case that the cast is vacuously true.
2706 // This detects the common cases where the test will short-circuit
2707 // away completely. We do this before we perform the null check,
2708 // because if the test is going to turn into zero code, we don't
2709 // want a residual null check left around. (Causes a slowdown,
2710 // for example, in some objArray manipulations, such as a[i]=a[j].)
2711 if (tk->singleton()) {
2712 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2713 if (objtp != NULL && objtp->klass() != NULL) {
2714 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2715 case SSC_always_true:
2716 return obj;
2717 case SSC_always_false:
2718 // It needs a null check because a null will *pass* the cast check.
2719 // A non-null value will always produce an exception.
2720 return null_assert(obj);
2721 }
2722 }
2723 }
2725 ciProfileData* data = NULL;
2726 if (failure_control == NULL) { // use MDO in regular case only
2727 assert(java_bc() == Bytecodes::_aastore ||
2728 java_bc() == Bytecodes::_checkcast,
2729 "interpreter profiles type checks only for these BCs");
2730 data = method()->method_data()->bci_to_data(bci());
2731 }
2733 // Make the merge point
2734 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2735 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
2736 Node* phi = new (C) PhiNode(region, toop);
2737 C->set_has_split_ifs(true); // Has chance for split-if optimization
2739 // Use null-cast information if it is available
2740 bool never_see_null = ((failure_control == NULL) // regular case only
2741 && seems_never_null(obj, data));
2743 // Null check; get casted pointer; set region slot 3
2744 Node* null_ctl = top();
2745 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2747 // If not_null_obj is dead, only null-path is taken
2748 if (stopped()) { // Doing instance-of on a NULL?
2749 set_control(null_ctl);
2750 return null();
2751 }
2752 region->init_req(_null_path, null_ctl);
2753 phi ->init_req(_null_path, null()); // Set null path value
2754 if (null_ctl == top()) {
2755 // Do this eagerly, so that pattern matches like is_diamond_phi
2756 // will work even during parsing.
2757 assert(_null_path == PATH_LIMIT-1, "delete last");
2758 region->del_req(_null_path);
2759 phi ->del_req(_null_path);
2760 }
2762 Node* cast_obj = NULL;
2763 if (data != NULL &&
2764 // Counter has never been decremented (due to cast failure).
2765 // ...This is a reasonable thing to expect. It is true of
2766 // all casts inserted by javac to implement generic types.
2767 data->as_CounterData()->count() >= 0) {
2768 cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, tk->klass());
2769 if (cast_obj != NULL) {
2770 if (failure_control != NULL) // failure is now impossible
2771 (*failure_control) = top();
2772 // adjust the type of the phi to the exact klass:
2773 phi->raise_bottom_type(_gvn.type(cast_obj)->meet(TypePtr::NULL_PTR));
2774 }
2775 }
2777 if (cast_obj == NULL) {
2778 // Load the object's klass
2779 Node* obj_klass = load_object_klass(not_null_obj);
2781 // Generate the subtype check
2782 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2784 // Plug in success path into the merge
2785 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
2786 not_null_obj, toop));
2787 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2788 if (failure_control == NULL) {
2789 if (not_subtype_ctrl != top()) { // If failure is possible
2790 PreserveJVMState pjvms(this);
2791 set_control(not_subtype_ctrl);
2792 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2793 }
2794 } else {
2795 (*failure_control) = not_subtype_ctrl;
2796 }
2797 }
2799 region->init_req(_obj_path, control());
2800 phi ->init_req(_obj_path, cast_obj);
2802 // A merge of NULL or Casted-NotNull obj
2803 Node* res = _gvn.transform(phi);
2805 // Note I do NOT always 'replace_in_map(obj,result)' here.
2806 // if( tk->klass()->can_be_primary_super() )
2807 // This means that if I successfully store an Object into an array-of-String
2808 // I 'forget' that the Object is really now known to be a String. I have to
2809 // do this because we don't have true union types for interfaces - if I store
2810 // a Baz into an array-of-Interface and then tell the optimizer it's an
2811 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2812 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2813 // replace_in_map( obj, res );
2815 // Return final merged results
2816 set_control( _gvn.transform(region) );
2817 record_for_igvn(region);
2818 return res;
2819 }
2821 //------------------------------next_monitor-----------------------------------
2822 // What number should be given to the next monitor?
2823 int GraphKit::next_monitor() {
2824 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2825 int next = current + C->sync_stack_slots();
2826 // Keep the toplevel high water mark current:
2827 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2828 return current;
2829 }
2831 //------------------------------insert_mem_bar---------------------------------
2832 // Memory barrier to avoid floating things around
2833 // The membar serves as a pinch point between both control and all memory slices.
2834 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2835 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2836 mb->init_req(TypeFunc::Control, control());
2837 mb->init_req(TypeFunc::Memory, reset_memory());
2838 Node* membar = _gvn.transform(mb);
2839 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
2840 set_all_memory_call(membar);
2841 return membar;
2842 }
2844 //-------------------------insert_mem_bar_volatile----------------------------
2845 // Memory barrier to avoid floating things around
2846 // The membar serves as a pinch point between both control and memory(alias_idx).
2847 // If you want to make a pinch point on all memory slices, do not use this
2848 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2849 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2850 // When Parse::do_put_xxx updates a volatile field, it appends a series
2851 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2852 // The first membar is on the same memory slice as the field store opcode.
2853 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2854 // All the other membars (for other volatile slices, including AliasIdxBot,
2855 // which stands for all unknown volatile slices) are control-dependent
2856 // on the first membar. This prevents later volatile loads or stores
2857 // from sliding up past the just-emitted store.
2859 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2860 mb->set_req(TypeFunc::Control,control());
2861 if (alias_idx == Compile::AliasIdxBot) {
2862 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2863 } else {
2864 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2865 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2866 }
2867 Node* membar = _gvn.transform(mb);
2868 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
2869 if (alias_idx == Compile::AliasIdxBot) {
2870 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
2871 } else {
2872 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2873 }
2874 return membar;
2875 }
2877 //------------------------------shared_lock------------------------------------
2878 // Emit locking code.
2879 FastLockNode* GraphKit::shared_lock(Node* obj) {
2880 // bci is either a monitorenter bc or InvocationEntryBci
2881 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2882 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2884 if( !GenerateSynchronizationCode )
2885 return NULL; // Not locking things?
2886 if (stopped()) // Dead monitor?
2887 return NULL;
2889 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2891 // Box the stack location
2892 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
2893 Node* mem = reset_memory();
2895 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
2896 if (PrintPreciseBiasedLockingStatistics) {
2897 // Create the counters for this fast lock.
2898 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2899 }
2900 // Add monitor to debug info for the slow path. If we block inside the
2901 // slow path and de-opt, we need the monitor hanging around
2902 map()->push_monitor( flock );
2904 const TypeFunc *tf = LockNode::lock_type();
2905 LockNode *lock = new (C) LockNode(C, tf);
2907 lock->init_req( TypeFunc::Control, control() );
2908 lock->init_req( TypeFunc::Memory , mem );
2909 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2910 lock->init_req( TypeFunc::FramePtr, frameptr() );
2911 lock->init_req( TypeFunc::ReturnAdr, top() );
2913 lock->init_req(TypeFunc::Parms + 0, obj);
2914 lock->init_req(TypeFunc::Parms + 1, box);
2915 lock->init_req(TypeFunc::Parms + 2, flock);
2916 add_safepoint_edges(lock);
2918 lock = _gvn.transform( lock )->as_Lock();
2920 // lock has no side-effects, sets few values
2921 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2923 insert_mem_bar(Op_MemBarAcquireLock);
2925 // Add this to the worklist so that the lock can be eliminated
2926 record_for_igvn(lock);
2928 #ifndef PRODUCT
2929 if (PrintLockStatistics) {
2930 // Update the counter for this lock. Don't bother using an atomic
2931 // operation since we don't require absolute accuracy.
2932 lock->create_lock_counter(map()->jvms());
2933 increment_counter(lock->counter()->addr());
2934 }
2935 #endif
2937 return flock;
2938 }
2941 //------------------------------shared_unlock----------------------------------
2942 // Emit unlocking code.
2943 void GraphKit::shared_unlock(Node* box, Node* obj) {
2944 // bci is either a monitorenter bc or InvocationEntryBci
2945 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2946 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2948 if( !GenerateSynchronizationCode )
2949 return;
2950 if (stopped()) { // Dead monitor?
2951 map()->pop_monitor(); // Kill monitor from debug info
2952 return;
2953 }
2955 // Memory barrier to avoid floating things down past the locked region
2956 insert_mem_bar(Op_MemBarReleaseLock);
2958 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2959 UnlockNode *unlock = new (C) UnlockNode(C, tf);
2960 uint raw_idx = Compile::AliasIdxRaw;
2961 unlock->init_req( TypeFunc::Control, control() );
2962 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2963 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2964 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2965 unlock->init_req( TypeFunc::ReturnAdr, top() );
2967 unlock->init_req(TypeFunc::Parms + 0, obj);
2968 unlock->init_req(TypeFunc::Parms + 1, box);
2969 unlock = _gvn.transform(unlock)->as_Unlock();
2971 Node* mem = reset_memory();
2973 // unlock has no side-effects, sets few values
2974 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2976 // Kill monitor from debug info
2977 map()->pop_monitor( );
2978 }
2980 //-------------------------------get_layout_helper-----------------------------
2981 // If the given klass is a constant or known to be an array,
2982 // fetch the constant layout helper value into constant_value
2983 // and return (Node*)NULL. Otherwise, load the non-constant
2984 // layout helper value, and return the node which represents it.
2985 // This two-faced routine is useful because allocation sites
2986 // almost always feature constant types.
2987 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2988 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2989 if (!StressReflectiveCode && inst_klass != NULL) {
2990 ciKlass* klass = inst_klass->klass();
2991 bool xklass = inst_klass->klass_is_exact();
2992 if (xklass || klass->is_array_klass()) {
2993 jint lhelper = klass->layout_helper();
2994 if (lhelper != Klass::_lh_neutral_value) {
2995 constant_value = lhelper;
2996 return (Node*) NULL;
2997 }
2998 }
2999 }
3000 constant_value = Klass::_lh_neutral_value; // put in a known value
3001 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3002 return make_load(NULL, lhp, TypeInt::INT, T_INT);
3003 }
3005 // We just put in an allocate/initialize with a big raw-memory effect.
3006 // Hook selected additional alias categories on the initialization.
3007 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3008 MergeMemNode* init_in_merge,
3009 Node* init_out_raw) {
3010 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3011 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3013 Node* prevmem = kit.memory(alias_idx);
3014 init_in_merge->set_memory_at(alias_idx, prevmem);
3015 kit.set_memory(init_out_raw, alias_idx);
3016 }
3018 //---------------------------set_output_for_allocation-------------------------
3019 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3020 const TypeOopPtr* oop_type) {
3021 int rawidx = Compile::AliasIdxRaw;
3022 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3023 add_safepoint_edges(alloc);
3024 Node* allocx = _gvn.transform(alloc);
3025 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3026 // create memory projection for i_o
3027 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3028 make_slow_call_ex(allocx, env()->Throwable_klass(), true);
3030 // create a memory projection as for the normal control path
3031 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3032 set_memory(malloc, rawidx);
3034 // a normal slow-call doesn't change i_o, but an allocation does
3035 // we create a separate i_o projection for the normal control path
3036 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3037 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3039 // put in an initialization barrier
3040 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3041 rawoop)->as_Initialize();
3042 assert(alloc->initialization() == init, "2-way macro link must work");
3043 assert(init ->allocation() == alloc, "2-way macro link must work");
3044 {
3045 // Extract memory strands which may participate in the new object's
3046 // initialization, and source them from the new InitializeNode.
3047 // This will allow us to observe initializations when they occur,
3048 // and link them properly (as a group) to the InitializeNode.
3049 assert(init->in(InitializeNode::Memory) == malloc, "");
3050 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3051 init->set_req(InitializeNode::Memory, minit_in);
3052 record_for_igvn(minit_in); // fold it up later, if possible
3053 Node* minit_out = memory(rawidx);
3054 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3055 if (oop_type->isa_aryptr()) {
3056 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3057 int elemidx = C->get_alias_index(telemref);
3058 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3059 } else if (oop_type->isa_instptr()) {
3060 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3061 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3062 ciField* field = ik->nonstatic_field_at(i);
3063 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3064 continue; // do not bother to track really large numbers of fields
3065 // Find (or create) the alias category for this field:
3066 int fieldidx = C->alias_type(field)->index();
3067 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3068 }
3069 }
3070 }
3072 // Cast raw oop to the real thing...
3073 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3074 javaoop = _gvn.transform(javaoop);
3075 C->set_recent_alloc(control(), javaoop);
3076 assert(just_allocated_object(control()) == javaoop, "just allocated");
3078 #ifdef ASSERT
3079 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3080 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3081 "Ideal_allocation works");
3082 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3083 "Ideal_allocation works");
3084 if (alloc->is_AllocateArray()) {
3085 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3086 "Ideal_allocation works");
3087 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3088 "Ideal_allocation works");
3089 } else {
3090 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3091 }
3092 }
3093 #endif //ASSERT
3095 return javaoop;
3096 }
3098 //---------------------------new_instance--------------------------------------
3099 // This routine takes a klass_node which may be constant (for a static type)
3100 // or may be non-constant (for reflective code). It will work equally well
3101 // for either, and the graph will fold nicely if the optimizer later reduces
3102 // the type to a constant.
3103 // The optional arguments are for specialized use by intrinsics:
3104 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3105 // - If 'return_size_val', report the the total object size to the caller.
3106 Node* GraphKit::new_instance(Node* klass_node,
3107 Node* extra_slow_test,
3108 Node* *return_size_val) {
3109 // Compute size in doublewords
3110 // The size is always an integral number of doublewords, represented
3111 // as a positive bytewise size stored in the klass's layout_helper.
3112 // The layout_helper also encodes (in a low bit) the need for a slow path.
3113 jint layout_con = Klass::_lh_neutral_value;
3114 Node* layout_val = get_layout_helper(klass_node, layout_con);
3115 int layout_is_con = (layout_val == NULL);
3117 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3118 // Generate the initial go-slow test. It's either ALWAYS (return a
3119 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3120 // case) a computed value derived from the layout_helper.
3121 Node* initial_slow_test = NULL;
3122 if (layout_is_con) {
3123 assert(!StressReflectiveCode, "stress mode does not use these paths");
3124 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3125 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3127 } else { // reflective case
3128 // This reflective path is used by Unsafe.allocateInstance.
3129 // (It may be stress-tested by specifying StressReflectiveCode.)
3130 // Basically, we want to get into the VM is there's an illegal argument.
3131 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3132 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3133 if (extra_slow_test != intcon(0)) {
3134 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3135 }
3136 // (Macro-expander will further convert this to a Bool, if necessary.)
3137 }
3139 // Find the size in bytes. This is easy; it's the layout_helper.
3140 // The size value must be valid even if the slow path is taken.
3141 Node* size = NULL;
3142 if (layout_is_con) {
3143 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3144 } else { // reflective case
3145 // This reflective path is used by clone and Unsafe.allocateInstance.
3146 size = ConvI2X(layout_val);
3148 // Clear the low bits to extract layout_helper_size_in_bytes:
3149 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3150 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3151 size = _gvn.transform( new (C) AndXNode(size, mask) );
3152 }
3153 if (return_size_val != NULL) {
3154 (*return_size_val) = size;
3155 }
3157 // This is a precise notnull oop of the klass.
3158 // (Actually, it need not be precise if this is a reflective allocation.)
3159 // It's what we cast the result to.
3160 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3161 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3162 const TypeOopPtr* oop_type = tklass->as_instance_type();
3164 // Now generate allocation code
3166 // The entire memory state is needed for slow path of the allocation
3167 // since GC and deoptimization can happened.
3168 Node *mem = reset_memory();
3169 set_all_memory(mem); // Create new memory state
3171 AllocateNode* alloc
3172 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3173 control(), mem, i_o(),
3174 size, klass_node,
3175 initial_slow_test);
3177 return set_output_for_allocation(alloc, oop_type);
3178 }
3180 //-------------------------------new_array-------------------------------------
3181 // helper for both newarray and anewarray
3182 // The 'length' parameter is (obviously) the length of the array.
3183 // See comments on new_instance for the meaning of the other arguments.
3184 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3185 Node* length, // number of array elements
3186 int nargs, // number of arguments to push back for uncommon trap
3187 Node* *return_size_val) {
3188 jint layout_con = Klass::_lh_neutral_value;
3189 Node* layout_val = get_layout_helper(klass_node, layout_con);
3190 int layout_is_con = (layout_val == NULL);
3192 if (!layout_is_con && !StressReflectiveCode &&
3193 !too_many_traps(Deoptimization::Reason_class_check)) {
3194 // This is a reflective array creation site.
3195 // Optimistically assume that it is a subtype of Object[],
3196 // so that we can fold up all the address arithmetic.
3197 layout_con = Klass::array_layout_helper(T_OBJECT);
3198 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3199 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3200 { BuildCutout unless(this, bol_lh, PROB_MAX);
3201 inc_sp(nargs);
3202 uncommon_trap(Deoptimization::Reason_class_check,
3203 Deoptimization::Action_maybe_recompile);
3204 }
3205 layout_val = NULL;
3206 layout_is_con = true;
3207 }
3209 // Generate the initial go-slow test. Make sure we do not overflow
3210 // if length is huge (near 2Gig) or negative! We do not need
3211 // exact double-words here, just a close approximation of needed
3212 // double-words. We can't add any offset or rounding bits, lest we
3213 // take a size -1 of bytes and make it positive. Use an unsigned
3214 // compare, so negative sizes look hugely positive.
3215 int fast_size_limit = FastAllocateSizeLimit;
3216 if (layout_is_con) {
3217 assert(!StressReflectiveCode, "stress mode does not use these paths");
3218 // Increase the size limit if we have exact knowledge of array type.
3219 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3220 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3221 }
3223 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3224 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3225 if (initial_slow_test->is_Bool()) {
3226 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3227 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3228 }
3230 // --- Size Computation ---
3231 // array_size = round_to_heap(array_header + (length << elem_shift));
3232 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3233 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3234 // The rounding mask is strength-reduced, if possible.
3235 int round_mask = MinObjAlignmentInBytes - 1;
3236 Node* header_size = NULL;
3237 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3238 // (T_BYTE has the weakest alignment and size restrictions...)
3239 if (layout_is_con) {
3240 int hsize = Klass::layout_helper_header_size(layout_con);
3241 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3242 BasicType etype = Klass::layout_helper_element_type(layout_con);
3243 if ((round_mask & ~right_n_bits(eshift)) == 0)
3244 round_mask = 0; // strength-reduce it if it goes away completely
3245 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3246 assert(header_size_min <= hsize, "generic minimum is smallest");
3247 header_size_min = hsize;
3248 header_size = intcon(hsize + round_mask);
3249 } else {
3250 Node* hss = intcon(Klass::_lh_header_size_shift);
3251 Node* hsm = intcon(Klass::_lh_header_size_mask);
3252 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3253 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3254 Node* mask = intcon(round_mask);
3255 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3256 }
3258 Node* elem_shift = NULL;
3259 if (layout_is_con) {
3260 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3261 if (eshift != 0)
3262 elem_shift = intcon(eshift);
3263 } else {
3264 // There is no need to mask or shift this value.
3265 // The semantics of LShiftINode include an implicit mask to 0x1F.
3266 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3267 elem_shift = layout_val;
3268 }
3270 // Transition to native address size for all offset calculations:
3271 Node* lengthx = ConvI2X(length);
3272 Node* headerx = ConvI2X(header_size);
3273 #ifdef _LP64
3274 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3275 if (tllen != NULL && tllen->_lo < 0) {
3276 // Add a manual constraint to a positive range. Cf. array_element_address.
3277 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3278 if (size_max > tllen->_hi) size_max = tllen->_hi;
3279 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3280 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon));
3281 }
3282 }
3283 #endif
3285 // Combine header size (plus rounding) and body size. Then round down.
3286 // This computation cannot overflow, because it is used only in two
3287 // places, one where the length is sharply limited, and the other
3288 // after a successful allocation.
3289 Node* abody = lengthx;
3290 if (elem_shift != NULL)
3291 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3292 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3293 if (round_mask != 0) {
3294 Node* mask = MakeConX(~round_mask);
3295 size = _gvn.transform( new(C) AndXNode(size, mask) );
3296 }
3297 // else if round_mask == 0, the size computation is self-rounding
3299 if (return_size_val != NULL) {
3300 // This is the size
3301 (*return_size_val) = size;
3302 }
3304 // Now generate allocation code
3306 // The entire memory state is needed for slow path of the allocation
3307 // since GC and deoptimization can happened.
3308 Node *mem = reset_memory();
3309 set_all_memory(mem); // Create new memory state
3311 // Create the AllocateArrayNode and its result projections
3312 AllocateArrayNode* alloc
3313 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3314 control(), mem, i_o(),
3315 size, klass_node,
3316 initial_slow_test,
3317 length);
3319 // Cast to correct type. Note that the klass_node may be constant or not,
3320 // and in the latter case the actual array type will be inexact also.
3321 // (This happens via a non-constant argument to inline_native_newArray.)
3322 // In any case, the value of klass_node provides the desired array type.
3323 const TypeInt* length_type = _gvn.find_int_type(length);
3324 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3325 if (ary_type->isa_aryptr() && length_type != NULL) {
3326 // Try to get a better type than POS for the size
3327 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3328 }
3330 Node* javaoop = set_output_for_allocation(alloc, ary_type);
3332 // Cast length on remaining path to be as narrow as possible
3333 if (map()->find_edge(length) >= 0) {
3334 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3335 if (ccast != length) {
3336 _gvn.set_type_bottom(ccast);
3337 record_for_igvn(ccast);
3338 replace_in_map(length, ccast);
3339 }
3340 }
3342 return javaoop;
3343 }
3345 // The following "Ideal_foo" functions are placed here because they recognize
3346 // the graph shapes created by the functions immediately above.
3348 //---------------------------Ideal_allocation----------------------------------
3349 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3350 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3351 if (ptr == NULL) { // reduce dumb test in callers
3352 return NULL;
3353 }
3354 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3355 ptr = ptr->in(1);
3356 if (ptr == NULL) return NULL;
3357 }
3358 // Return NULL for allocations with several casts:
3359 // j.l.reflect.Array.newInstance(jobject, jint)
3360 // Object.clone()
3361 // to keep more precise type from last cast.
3362 if (ptr->is_Proj()) {
3363 Node* allo = ptr->in(0);
3364 if (allo != NULL && allo->is_Allocate()) {
3365 return allo->as_Allocate();
3366 }
3367 }
3368 // Report failure to match.
3369 return NULL;
3370 }
3372 // Fancy version which also strips off an offset (and reports it to caller).
3373 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3374 intptr_t& offset) {
3375 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3376 if (base == NULL) return NULL;
3377 return Ideal_allocation(base, phase);
3378 }
3380 // Trace Initialize <- Proj[Parm] <- Allocate
3381 AllocateNode* InitializeNode::allocation() {
3382 Node* rawoop = in(InitializeNode::RawAddress);
3383 if (rawoop->is_Proj()) {
3384 Node* alloc = rawoop->in(0);
3385 if (alloc->is_Allocate()) {
3386 return alloc->as_Allocate();
3387 }
3388 }
3389 return NULL;
3390 }
3392 // Trace Allocate -> Proj[Parm] -> Initialize
3393 InitializeNode* AllocateNode::initialization() {
3394 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3395 if (rawoop == NULL) return NULL;
3396 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3397 Node* init = rawoop->fast_out(i);
3398 if (init->is_Initialize()) {
3399 assert(init->as_Initialize()->allocation() == this, "2-way link");
3400 return init->as_Initialize();
3401 }
3402 }
3403 return NULL;
3404 }
3406 //----------------------------- loop predicates ---------------------------
3408 //------------------------------add_predicate_impl----------------------------
3409 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3410 // Too many traps seen?
3411 if (too_many_traps(reason)) {
3412 #ifdef ASSERT
3413 if (TraceLoopPredicate) {
3414 int tc = C->trap_count(reason);
3415 tty->print("too many traps=%s tcount=%d in ",
3416 Deoptimization::trap_reason_name(reason), tc);
3417 method()->print(); // which method has too many predicate traps
3418 tty->cr();
3419 }
3420 #endif
3421 // We cannot afford to take more traps here,
3422 // do not generate predicate.
3423 return;
3424 }
3426 Node *cont = _gvn.intcon(1);
3427 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3428 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3429 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3430 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3431 C->add_predicate_opaq(opq);
3432 {
3433 PreserveJVMState pjvms(this);
3434 set_control(iffalse);
3435 inc_sp(nargs);
3436 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3437 }
3438 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3439 set_control(iftrue);
3440 }
3442 //------------------------------add_predicate---------------------------------
3443 void GraphKit::add_predicate(int nargs) {
3444 if (UseLoopPredicate) {
3445 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3446 }
3447 // loop's limit check predicate should be near the loop.
3448 if (LoopLimitCheck) {
3449 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3450 }
3451 }
3453 //----------------------------- store barriers ----------------------------
3454 #define __ ideal.
3456 void GraphKit::sync_kit(IdealKit& ideal) {
3457 set_all_memory(__ merged_memory());
3458 set_i_o(__ i_o());
3459 set_control(__ ctrl());
3460 }
3462 void GraphKit::final_sync(IdealKit& ideal) {
3463 // Final sync IdealKit and graphKit.
3464 sync_kit(ideal);
3465 }
3467 // vanilla/CMS post barrier
3468 // Insert a write-barrier store. This is to let generational GC work; we have
3469 // to flag all oop-stores before the next GC point.
3470 void GraphKit::write_barrier_post(Node* oop_store,
3471 Node* obj,
3472 Node* adr,
3473 uint adr_idx,
3474 Node* val,
3475 bool use_precise) {
3476 // No store check needed if we're storing a NULL or an old object
3477 // (latter case is probably a string constant). The concurrent
3478 // mark sweep garbage collector, however, needs to have all nonNull
3479 // oop updates flagged via card-marks.
3480 if (val != NULL && val->is_Con()) {
3481 // must be either an oop or NULL
3482 const Type* t = val->bottom_type();
3483 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3484 // stores of null never (?) need barriers
3485 return;
3486 }
3488 if (use_ReduceInitialCardMarks()
3489 && obj == just_allocated_object(control())) {
3490 // We can skip marks on a freshly-allocated object in Eden.
3491 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3492 // That routine informs GC to take appropriate compensating steps,
3493 // upon a slow-path allocation, so as to make this card-mark
3494 // elision safe.
3495 return;
3496 }
3498 if (!use_precise) {
3499 // All card marks for a (non-array) instance are in one place:
3500 adr = obj;
3501 }
3502 // (Else it's an array (or unknown), and we want more precise card marks.)
3503 assert(adr != NULL, "");
3505 IdealKit ideal(this, true);
3507 // Convert the pointer to an int prior to doing math on it
3508 Node* cast = __ CastPX(__ ctrl(), adr);
3510 // Divide by card size
3511 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3512 "Only one we handle so far.");
3513 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3515 // Combine card table base and card offset
3516 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3518 // Get the alias_index for raw card-mark memory
3519 int adr_type = Compile::AliasIdxRaw;
3520 Node* zero = __ ConI(0); // Dirty card value
3521 BasicType bt = T_BYTE;
3523 if (UseCondCardMark) {
3524 // The classic GC reference write barrier is typically implemented
3525 // as a store into the global card mark table. Unfortunately
3526 // unconditional stores can result in false sharing and excessive
3527 // coherence traffic as well as false transactional aborts.
3528 // UseCondCardMark enables MP "polite" conditional card mark
3529 // stores. In theory we could relax the load from ctrl() to
3530 // no_ctrl, but that doesn't buy much latitude.
3531 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3532 __ if_then(card_val, BoolTest::ne, zero);
3533 }
3535 // Smash zero into card
3536 if( !UseConcMarkSweepGC ) {
3537 __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3538 } else {
3539 // Specialized path for CM store barrier
3540 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3541 }
3543 if (UseCondCardMark) {
3544 __ end_if();
3545 }
3547 // Final sync IdealKit and GraphKit.
3548 final_sync(ideal);
3549 }
3551 // G1 pre/post barriers
3552 void GraphKit::g1_write_barrier_pre(bool do_load,
3553 Node* obj,
3554 Node* adr,
3555 uint alias_idx,
3556 Node* val,
3557 const TypeOopPtr* val_type,
3558 Node* pre_val,
3559 BasicType bt) {
3561 // Some sanity checks
3562 // Note: val is unused in this routine.
3564 if (do_load) {
3565 // We need to generate the load of the previous value
3566 assert(obj != NULL, "must have a base");
3567 assert(adr != NULL, "where are loading from?");
3568 assert(pre_val == NULL, "loaded already?");
3569 assert(val_type != NULL, "need a type");
3570 } else {
3571 // In this case both val_type and alias_idx are unused.
3572 assert(pre_val != NULL, "must be loaded already");
3573 // Nothing to be done if pre_val is null.
3574 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3575 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3576 }
3577 assert(bt == T_OBJECT, "or we shouldn't be here");
3579 IdealKit ideal(this, true);
3581 Node* tls = __ thread(); // ThreadLocalStorage
3583 Node* no_ctrl = NULL;
3584 Node* no_base = __ top();
3585 Node* zero = __ ConI(0);
3586 Node* zeroX = __ ConX(0);
3588 float likely = PROB_LIKELY(0.999);
3589 float unlikely = PROB_UNLIKELY(0.999);
3591 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3592 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3594 // Offsets into the thread
3595 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3596 PtrQueue::byte_offset_of_active());
3597 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3598 PtrQueue::byte_offset_of_index());
3599 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3600 PtrQueue::byte_offset_of_buf());
3602 // Now the actual pointers into the thread
3603 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3604 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3605 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3607 // Now some of the values
3608 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3610 // if (!marking)
3611 __ if_then(marking, BoolTest::ne, zero); {
3612 BasicType index_bt = TypeX_X->basic_type();
3613 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3614 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3616 if (do_load) {
3617 // load original value
3618 // alias_idx correct??
3619 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3620 }
3622 // if (pre_val != NULL)
3623 __ if_then(pre_val, BoolTest::ne, null()); {
3624 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3626 // is the queue for this thread full?
3627 __ if_then(index, BoolTest::ne, zeroX, likely); {
3629 // decrement the index
3630 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3632 // Now get the buffer location we will log the previous value into and store it
3633 Node *log_addr = __ AddP(no_base, buffer, next_index);
3634 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw);
3635 // update the index
3636 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw);
3638 } __ else_(); {
3640 // logging buffer is full, call the runtime
3641 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3642 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3643 } __ end_if(); // (!index)
3644 } __ end_if(); // (pre_val != NULL)
3645 } __ end_if(); // (!marking)
3647 // Final sync IdealKit and GraphKit.
3648 final_sync(ideal);
3649 }
3651 //
3652 // Update the card table and add card address to the queue
3653 //
3654 void GraphKit::g1_mark_card(IdealKit& ideal,
3655 Node* card_adr,
3656 Node* oop_store,
3657 uint oop_alias_idx,
3658 Node* index,
3659 Node* index_adr,
3660 Node* buffer,
3661 const TypeFunc* tf) {
3663 Node* zero = __ ConI(0);
3664 Node* zeroX = __ ConX(0);
3665 Node* no_base = __ top();
3666 BasicType card_bt = T_BYTE;
3667 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3668 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3670 // Now do the queue work
3671 __ if_then(index, BoolTest::ne, zeroX); {
3673 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3674 Node* log_addr = __ AddP(no_base, buffer, next_index);
3676 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3677 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw);
3679 } __ else_(); {
3680 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3681 } __ end_if();
3683 }
3685 void GraphKit::g1_write_barrier_post(Node* oop_store,
3686 Node* obj,
3687 Node* adr,
3688 uint alias_idx,
3689 Node* val,
3690 BasicType bt,
3691 bool use_precise) {
3692 // If we are writing a NULL then we need no post barrier
3694 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3695 // Must be NULL
3696 const Type* t = val->bottom_type();
3697 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3698 // No post barrier if writing NULLx
3699 return;
3700 }
3702 if (!use_precise) {
3703 // All card marks for a (non-array) instance are in one place:
3704 adr = obj;
3705 }
3706 // (Else it's an array (or unknown), and we want more precise card marks.)
3707 assert(adr != NULL, "");
3709 IdealKit ideal(this, true);
3711 Node* tls = __ thread(); // ThreadLocalStorage
3713 Node* no_base = __ top();
3714 float likely = PROB_LIKELY(0.999);
3715 float unlikely = PROB_UNLIKELY(0.999);
3716 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
3717 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
3718 Node* zeroX = __ ConX(0);
3720 // Get the alias_index for raw card-mark memory
3721 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3723 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3725 // Offsets into the thread
3726 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3727 PtrQueue::byte_offset_of_index());
3728 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3729 PtrQueue::byte_offset_of_buf());
3731 // Pointers into the thread
3733 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3734 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3736 // Now some values
3737 // Use ctrl to avoid hoisting these values past a safepoint, which could
3738 // potentially reset these fields in the JavaThread.
3739 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
3740 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3742 // Convert the store obj pointer to an int prior to doing math on it
3743 // Must use ctrl to prevent "integerized oop" existing across safepoint
3744 Node* cast = __ CastPX(__ ctrl(), adr);
3746 // Divide pointer by card size
3747 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3749 // Combine card table base and card offset
3750 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3752 // If we know the value being stored does it cross regions?
3754 if (val != NULL) {
3755 // Does the store cause us to cross regions?
3757 // Should be able to do an unsigned compare of region_size instead of
3758 // and extra shift. Do we have an unsigned compare??
3759 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3760 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3762 // if (xor_res == 0) same region so skip
3763 __ if_then(xor_res, BoolTest::ne, zeroX); {
3765 // No barrier if we are storing a NULL
3766 __ if_then(val, BoolTest::ne, null(), unlikely); {
3768 // Ok must mark the card if not already dirty
3770 // load the original value of the card
3771 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3773 __ if_then(card_val, BoolTest::ne, young_card); {
3774 sync_kit(ideal);
3775 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
3776 insert_mem_bar(Op_MemBarVolatile, oop_store);
3777 __ sync_kit(this);
3779 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3780 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
3781 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3782 } __ end_if();
3783 } __ end_if();
3784 } __ end_if();
3785 } __ end_if();
3786 } else {
3787 // Object.clone() instrinsic uses this path.
3788 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3789 }
3791 // Final sync IdealKit and GraphKit.
3792 final_sync(ideal);
3793 }
3794 #undef __
3798 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
3799 if (java_lang_String::has_offset_field()) {
3800 int offset_offset = java_lang_String::offset_offset_in_bytes();
3801 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3802 false, NULL, 0);
3803 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
3804 int offset_field_idx = C->get_alias_index(offset_field_type);
3805 return make_load(ctrl,
3806 basic_plus_adr(str, str, offset_offset),
3807 TypeInt::INT, T_INT, offset_field_idx);
3808 } else {
3809 return intcon(0);
3810 }
3811 }
3813 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
3814 if (java_lang_String::has_count_field()) {
3815 int count_offset = java_lang_String::count_offset_in_bytes();
3816 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3817 false, NULL, 0);
3818 const TypePtr* count_field_type = string_type->add_offset(count_offset);
3819 int count_field_idx = C->get_alias_index(count_field_type);
3820 return make_load(ctrl,
3821 basic_plus_adr(str, str, count_offset),
3822 TypeInt::INT, T_INT, count_field_idx);
3823 } else {
3824 return load_array_length(load_String_value(ctrl, str));
3825 }
3826 }
3828 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
3829 int value_offset = java_lang_String::value_offset_in_bytes();
3830 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3831 false, NULL, 0);
3832 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3833 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3834 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
3835 ciTypeArrayKlass::make(T_CHAR), true, 0);
3836 int value_field_idx = C->get_alias_index(value_field_type);
3837 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
3838 value_type, T_OBJECT, value_field_idx);
3839 // String.value field is known to be @Stable.
3840 if (UseImplicitStableValues) {
3841 load = cast_array_to_stable(load, value_type);
3842 }
3843 return load;
3844 }
3846 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
3847 int offset_offset = java_lang_String::offset_offset_in_bytes();
3848 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3849 false, NULL, 0);
3850 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
3851 int offset_field_idx = C->get_alias_index(offset_field_type);
3852 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
3853 value, T_INT, offset_field_idx);
3854 }
3856 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
3857 int value_offset = java_lang_String::value_offset_in_bytes();
3858 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3859 false, NULL, 0);
3860 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3861 int value_field_idx = C->get_alias_index(value_field_type);
3862 store_to_memory(ctrl, basic_plus_adr(str, value_offset),
3863 value, T_OBJECT, value_field_idx);
3864 }
3866 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
3867 int count_offset = java_lang_String::count_offset_in_bytes();
3868 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3869 false, NULL, 0);
3870 const TypePtr* count_field_type = string_type->add_offset(count_offset);
3871 int count_field_idx = C->get_alias_index(count_field_type);
3872 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
3873 value, T_INT, count_field_idx);
3874 }
3876 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
3877 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
3878 // assumption of CCP analysis.
3879 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
3880 }