Wed, 24 Apr 2013 14:48:43 -0700
8012715: G1: GraphKit accesses PtrQueue::_index as int but is size_t
Summary: In graphKit INT operations were generated to access PtrQueue::_index which has type size_t. This is 64 bit on 64-bit machines. No problems occur on little endian machines as long as the index fits into 32 bit, but on big endian machines the upper part is read, which is zero. This leads to unnecessary branches to the slow path in the runtime.
Reviewed-by: twisti, johnc
Contributed-by: Martin Doerr <martin.doerr@sap.com>
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_map->req() == phi_map->req(), "matching maps");
337 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
338 Node* hidden_merge_mark = root();
339 Node* region = phi_map->control();
340 MergeMemNode* phi_mem = phi_map->merged_memory();
341 MergeMemNode* ex_mem = ex_map->merged_memory();
342 if (region->in(0) != hidden_merge_mark) {
343 // The control input is not (yet) a specially-marked region in phi_map.
344 // Make it so, and build some phis.
345 region = new (C) RegionNode(2);
346 _gvn.set_type(region, Type::CONTROL);
347 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
348 region->init_req(1, phi_map->control());
349 phi_map->set_control(region);
350 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
351 record_for_igvn(io_phi);
352 _gvn.set_type(io_phi, Type::ABIO);
353 phi_map->set_i_o(io_phi);
354 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
355 Node* m = mms.memory();
356 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
357 record_for_igvn(m_phi);
358 _gvn.set_type(m_phi, Type::MEMORY);
359 mms.set_memory(m_phi);
360 }
361 }
363 // Either or both of phi_map and ex_map might already be converted into phis.
364 Node* ex_control = ex_map->control();
365 // if there is special marking on ex_map also, we add multiple edges from src
366 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
367 // how wide was the destination phi_map, originally?
368 uint orig_width = region->req();
370 if (add_multiple) {
371 add_n_reqs(region, ex_control);
372 add_n_reqs(phi_map->i_o(), ex_map->i_o());
373 } else {
374 // ex_map has no merges, so we just add single edges everywhere
375 add_one_req(region, ex_control);
376 add_one_req(phi_map->i_o(), ex_map->i_o());
377 }
378 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
379 if (mms.is_empty()) {
380 // get a copy of the base memory, and patch some inputs into it
381 const TypePtr* adr_type = mms.adr_type(C);
382 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
383 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
384 mms.set_memory(phi);
385 // Prepare to append interesting stuff onto the newly sliced phi:
386 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
387 }
388 // Append stuff from ex_map:
389 if (add_multiple) {
390 add_n_reqs(mms.memory(), mms.memory2());
391 } else {
392 add_one_req(mms.memory(), mms.memory2());
393 }
394 }
395 uint limit = ex_map->req();
396 for (uint i = TypeFunc::Parms; i < limit; i++) {
397 // Skip everything in the JVMS after tos. (The ex_oop follows.)
398 if (i == tos) i = ex_jvms->monoff();
399 Node* src = ex_map->in(i);
400 Node* dst = phi_map->in(i);
401 if (src != dst) {
402 PhiNode* phi;
403 if (dst->in(0) != region) {
404 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
405 record_for_igvn(phi);
406 _gvn.set_type(phi, phi->type());
407 phi_map->set_req(i, dst);
408 // Prepare to append interesting stuff onto the new phi:
409 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
410 } else {
411 assert(dst->is_Phi(), "nobody else uses a hidden region");
412 phi = (PhiNode*)dst;
413 }
414 if (add_multiple && src->in(0) == ex_control) {
415 // Both are phis.
416 add_n_reqs(dst, src);
417 } else {
418 while (dst->req() < region->req()) add_one_req(dst, src);
419 }
420 const Type* srctype = _gvn.type(src);
421 if (phi->type() != srctype) {
422 const Type* dsttype = phi->type()->meet(srctype);
423 if (phi->type() != dsttype) {
424 phi->set_type(dsttype);
425 _gvn.set_type(phi, dsttype);
426 }
427 }
428 }
429 }
430 }
432 //--------------------------use_exception_state--------------------------------
433 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
434 if (failing()) { stop(); return top(); }
435 Node* region = phi_map->control();
436 Node* hidden_merge_mark = root();
437 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
438 Node* ex_oop = clear_saved_ex_oop(phi_map);
439 if (region->in(0) == hidden_merge_mark) {
440 // Special marking for internal ex-states. Process the phis now.
441 region->set_req(0, region); // now it's an ordinary region
442 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
443 // Note: Setting the jvms also sets the bci and sp.
444 set_control(_gvn.transform(region));
445 uint tos = jvms()->stkoff() + sp();
446 for (uint i = 1; i < tos; i++) {
447 Node* x = phi_map->in(i);
448 if (x->in(0) == region) {
449 assert(x->is_Phi(), "expected a special phi");
450 phi_map->set_req(i, _gvn.transform(x));
451 }
452 }
453 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
454 Node* x = mms.memory();
455 if (x->in(0) == region) {
456 assert(x->is_Phi(), "nobody else uses a hidden region");
457 mms.set_memory(_gvn.transform(x));
458 }
459 }
460 if (ex_oop->in(0) == region) {
461 assert(ex_oop->is_Phi(), "expected a special phi");
462 ex_oop = _gvn.transform(ex_oop);
463 }
464 } else {
465 set_jvms(phi_map->jvms());
466 }
468 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
469 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
470 return ex_oop;
471 }
473 //---------------------------------java_bc-------------------------------------
474 Bytecodes::Code GraphKit::java_bc() const {
475 ciMethod* method = this->method();
476 int bci = this->bci();
477 if (method != NULL && bci != InvocationEntryBci)
478 return method->java_code_at_bci(bci);
479 else
480 return Bytecodes::_illegal;
481 }
483 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
484 bool must_throw) {
485 // if the exception capability is set, then we will generate code
486 // to check the JavaThread.should_post_on_exceptions flag to see
487 // if we actually need to report exception events (for this
488 // thread). If we don't need to report exception events, we will
489 // take the normal fast path provided by add_exception_events. If
490 // exception event reporting is enabled for this thread, we will
491 // take the uncommon_trap in the BuildCutout below.
493 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
494 Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
495 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
496 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false);
498 // Test the should_post_on_exceptions_flag vs. 0
499 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
500 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
502 // Branch to slow_path if should_post_on_exceptions_flag was true
503 { BuildCutout unless(this, tst, PROB_MAX);
504 // Do not try anything fancy if we're notifying the VM on every throw.
505 // Cf. case Bytecodes::_athrow in parse2.cpp.
506 uncommon_trap(reason, Deoptimization::Action_none,
507 (ciKlass*)NULL, (char*)NULL, must_throw);
508 }
510 }
512 //------------------------------builtin_throw----------------------------------
513 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
514 bool must_throw = true;
516 if (env()->jvmti_can_post_on_exceptions()) {
517 // check if we must post exception events, take uncommon trap if so
518 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
519 // here if should_post_on_exceptions is false
520 // continue on with the normal codegen
521 }
523 // If this particular condition has not yet happened at this
524 // bytecode, then use the uncommon trap mechanism, and allow for
525 // a future recompilation if several traps occur here.
526 // If the throw is hot, try to use a more complicated inline mechanism
527 // which keeps execution inside the compiled code.
528 bool treat_throw_as_hot = false;
529 ciMethodData* md = method()->method_data();
531 if (ProfileTraps) {
532 if (too_many_traps(reason)) {
533 treat_throw_as_hot = true;
534 }
535 // (If there is no MDO at all, assume it is early in
536 // execution, and that any deopts are part of the
537 // startup transient, and don't need to be remembered.)
539 // Also, if there is a local exception handler, treat all throws
540 // as hot if there has been at least one in this method.
541 if (C->trap_count(reason) != 0
542 && method()->method_data()->trap_count(reason) != 0
543 && has_ex_handler()) {
544 treat_throw_as_hot = true;
545 }
546 }
548 // If this throw happens frequently, an uncommon trap might cause
549 // a performance pothole. If there is a local exception handler,
550 // and if this particular bytecode appears to be deoptimizing often,
551 // let us handle the throw inline, with a preconstructed instance.
552 // Note: If the deopt count has blown up, the uncommon trap
553 // runtime is going to flush this nmethod, not matter what.
554 if (treat_throw_as_hot
555 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
556 // If the throw is local, we use a pre-existing instance and
557 // punt on the backtrace. This would lead to a missing backtrace
558 // (a repeat of 4292742) if the backtrace object is ever asked
559 // for its backtrace.
560 // Fixing this remaining case of 4292742 requires some flavor of
561 // escape analysis. Leave that for the future.
562 ciInstance* ex_obj = NULL;
563 switch (reason) {
564 case Deoptimization::Reason_null_check:
565 ex_obj = env()->NullPointerException_instance();
566 break;
567 case Deoptimization::Reason_div0_check:
568 ex_obj = env()->ArithmeticException_instance();
569 break;
570 case Deoptimization::Reason_range_check:
571 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
572 break;
573 case Deoptimization::Reason_class_check:
574 if (java_bc() == Bytecodes::_aastore) {
575 ex_obj = env()->ArrayStoreException_instance();
576 } else {
577 ex_obj = env()->ClassCastException_instance();
578 }
579 break;
580 }
581 if (failing()) { stop(); return; } // exception allocation might fail
582 if (ex_obj != NULL) {
583 // Cheat with a preallocated exception object.
584 if (C->log() != NULL)
585 C->log()->elem("hot_throw preallocated='1' reason='%s'",
586 Deoptimization::trap_reason_name(reason));
587 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
588 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
590 // Clear the detail message of the preallocated exception object.
591 // Weblogic sometimes mutates the detail message of exceptions
592 // using reflection.
593 int offset = java_lang_Throwable::get_detailMessage_offset();
594 const TypePtr* adr_typ = ex_con->add_offset(offset);
596 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
597 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
598 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT);
600 add_exception_state(make_exception_state(ex_node));
601 return;
602 }
603 }
605 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
606 // It won't be much cheaper than bailing to the interp., since we'll
607 // have to pass up all the debug-info, and the runtime will have to
608 // create the stack trace.
610 // Usual case: Bail to interpreter.
611 // Reserve the right to recompile if we haven't seen anything yet.
613 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
614 if (treat_throw_as_hot
615 && (method()->method_data()->trap_recompiled_at(bci())
616 || C->too_many_traps(reason))) {
617 // We cannot afford to take more traps here. Suffer in the interpreter.
618 if (C->log() != NULL)
619 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
620 Deoptimization::trap_reason_name(reason),
621 C->trap_count(reason));
622 action = Deoptimization::Action_none;
623 }
625 // "must_throw" prunes the JVM state to include only the stack, if there
626 // are no local exception handlers. This should cut down on register
627 // allocation time and code size, by drastically reducing the number
628 // of in-edges on the call to the uncommon trap.
630 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
631 }
634 //----------------------------PreserveJVMState---------------------------------
635 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
636 debug_only(kit->verify_map());
637 _kit = kit;
638 _map = kit->map(); // preserve the map
639 _sp = kit->sp();
640 kit->set_map(clone_map ? kit->clone_map() : NULL);
641 #ifdef ASSERT
642 _bci = kit->bci();
643 Parse* parser = kit->is_Parse();
644 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
645 _block = block;
646 #endif
647 }
648 PreserveJVMState::~PreserveJVMState() {
649 GraphKit* kit = _kit;
650 #ifdef ASSERT
651 assert(kit->bci() == _bci, "bci must not shift");
652 Parse* parser = kit->is_Parse();
653 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
654 assert(block == _block, "block must not shift");
655 #endif
656 kit->set_map(_map);
657 kit->set_sp(_sp);
658 }
661 //-----------------------------BuildCutout-------------------------------------
662 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
663 : PreserveJVMState(kit)
664 {
665 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
666 SafePointNode* outer_map = _map; // preserved map is caller's
667 SafePointNode* inner_map = kit->map();
668 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
669 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
670 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
671 }
672 BuildCutout::~BuildCutout() {
673 GraphKit* kit = _kit;
674 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
675 }
677 //---------------------------PreserveReexecuteState----------------------------
678 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
679 assert(!kit->stopped(), "must call stopped() before");
680 _kit = kit;
681 _sp = kit->sp();
682 _reexecute = kit->jvms()->_reexecute;
683 }
684 PreserveReexecuteState::~PreserveReexecuteState() {
685 if (_kit->stopped()) return;
686 _kit->jvms()->_reexecute = _reexecute;
687 _kit->set_sp(_sp);
688 }
690 //------------------------------clone_map--------------------------------------
691 // Implementation of PreserveJVMState
692 //
693 // Only clone_map(...) here. If this function is only used in the
694 // PreserveJVMState class we may want to get rid of this extra
695 // function eventually and do it all there.
697 SafePointNode* GraphKit::clone_map() {
698 if (map() == NULL) return NULL;
700 // Clone the memory edge first
701 Node* mem = MergeMemNode::make(C, map()->memory());
702 gvn().set_type_bottom(mem);
704 SafePointNode *clonemap = (SafePointNode*)map()->clone();
705 JVMState* jvms = this->jvms();
706 JVMState* clonejvms = jvms->clone_shallow(C);
707 clonemap->set_memory(mem);
708 clonemap->set_jvms(clonejvms);
709 clonejvms->set_map(clonemap);
710 record_for_igvn(clonemap);
711 gvn().set_type_bottom(clonemap);
712 return clonemap;
713 }
716 //-----------------------------set_map_clone-----------------------------------
717 void GraphKit::set_map_clone(SafePointNode* m) {
718 _map = m;
719 _map = clone_map();
720 _map->set_next_exception(NULL);
721 debug_only(verify_map());
722 }
725 //----------------------------kill_dead_locals---------------------------------
726 // Detect any locals which are known to be dead, and force them to top.
727 void GraphKit::kill_dead_locals() {
728 // Consult the liveness information for the locals. If any
729 // of them are unused, then they can be replaced by top(). This
730 // should help register allocation time and cut down on the size
731 // of the deoptimization information.
733 // This call is made from many of the bytecode handling
734 // subroutines called from the Big Switch in do_one_bytecode.
735 // Every bytecode which might include a slow path is responsible
736 // for killing its dead locals. The more consistent we
737 // are about killing deads, the fewer useless phis will be
738 // constructed for them at various merge points.
740 // bci can be -1 (InvocationEntryBci). We return the entry
741 // liveness for the method.
743 if (method() == NULL || method()->code_size() == 0) {
744 // We are building a graph for a call to a native method.
745 // All locals are live.
746 return;
747 }
749 ResourceMark rm;
751 // Consult the liveness information for the locals. If any
752 // of them are unused, then they can be replaced by top(). This
753 // should help register allocation time and cut down on the size
754 // of the deoptimization information.
755 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
757 int len = (int)live_locals.size();
758 assert(len <= jvms()->loc_size(), "too many live locals");
759 for (int local = 0; local < len; local++) {
760 if (!live_locals.at(local)) {
761 set_local(local, top());
762 }
763 }
764 }
766 #ifdef ASSERT
767 //-------------------------dead_locals_are_killed------------------------------
768 // Return true if all dead locals are set to top in the map.
769 // Used to assert "clean" debug info at various points.
770 bool GraphKit::dead_locals_are_killed() {
771 if (method() == NULL || method()->code_size() == 0) {
772 // No locals need to be dead, so all is as it should be.
773 return true;
774 }
776 // Make sure somebody called kill_dead_locals upstream.
777 ResourceMark rm;
778 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
779 if (jvms->loc_size() == 0) continue; // no locals to consult
780 SafePointNode* map = jvms->map();
781 ciMethod* method = jvms->method();
782 int bci = jvms->bci();
783 if (jvms == this->jvms()) {
784 bci = this->bci(); // it might not yet be synched
785 }
786 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
787 int len = (int)live_locals.size();
788 if (!live_locals.is_valid() || len == 0)
789 // This method is trivial, or is poisoned by a breakpoint.
790 return true;
791 assert(len == jvms->loc_size(), "live map consistent with locals map");
792 for (int local = 0; local < len; local++) {
793 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
794 if (PrintMiscellaneous && (Verbose || WizardMode)) {
795 tty->print_cr("Zombie local %d: ", local);
796 jvms->dump();
797 }
798 return false;
799 }
800 }
801 }
802 return true;
803 }
805 #endif //ASSERT
807 // Helper function for enforcing certain bytecodes to reexecute if
808 // deoptimization happens
809 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
810 ciMethod* cur_method = jvms->method();
811 int cur_bci = jvms->bci();
812 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
813 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
814 return Interpreter::bytecode_should_reexecute(code) ||
815 is_anewarray && code == Bytecodes::_multianewarray;
816 // Reexecute _multianewarray bytecode which was replaced with
817 // sequence of [a]newarray. See Parse::do_multianewarray().
818 //
819 // Note: interpreter should not have it set since this optimization
820 // is limited by dimensions and guarded by flag so in some cases
821 // multianewarray() runtime calls will be generated and
822 // the bytecode should not be reexecutes (stack will not be reset).
823 } else
824 return false;
825 }
827 // Helper function for adding JVMState and debug information to node
828 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
829 // Add the safepoint edges to the call (or other safepoint).
831 // Make sure dead locals are set to top. This
832 // should help register allocation time and cut down on the size
833 // of the deoptimization information.
834 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
836 // Walk the inline list to fill in the correct set of JVMState's
837 // Also fill in the associated edges for each JVMState.
839 // If the bytecode needs to be reexecuted we need to put
840 // the arguments back on the stack.
841 const bool should_reexecute = jvms()->should_reexecute();
842 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
844 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
845 // undefined if the bci is different. This is normal for Parse but it
846 // should not happen for LibraryCallKit because only one bci is processed.
847 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
848 "in LibraryCallKit the reexecute bit should not change");
850 // If we are guaranteed to throw, we can prune everything but the
851 // input to the current bytecode.
852 bool can_prune_locals = false;
853 uint stack_slots_not_pruned = 0;
854 int inputs = 0, depth = 0;
855 if (must_throw) {
856 assert(method() == youngest_jvms->method(), "sanity");
857 if (compute_stack_effects(inputs, depth)) {
858 can_prune_locals = true;
859 stack_slots_not_pruned = inputs;
860 }
861 }
863 if (env()->jvmti_can_access_local_variables()) {
864 // At any safepoint, this method can get breakpointed, which would
865 // then require an immediate deoptimization.
866 can_prune_locals = false; // do not prune locals
867 stack_slots_not_pruned = 0;
868 }
870 // do not scribble on the input jvms
871 JVMState* out_jvms = youngest_jvms->clone_deep(C);
872 call->set_jvms(out_jvms); // Start jvms list for call node
874 // For a known set of bytecodes, the interpreter should reexecute them if
875 // deoptimization happens. We set the reexecute state for them here
876 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
877 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
878 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
879 }
881 // Presize the call:
882 DEBUG_ONLY(uint non_debug_edges = call->req());
883 call->add_req_batch(top(), youngest_jvms->debug_depth());
884 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
886 // Set up edges so that the call looks like this:
887 // Call [state:] ctl io mem fptr retadr
888 // [parms:] parm0 ... parmN
889 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
890 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
891 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
892 // Note that caller debug info precedes callee debug info.
894 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
895 uint debug_ptr = call->req();
897 // Loop over the map input edges associated with jvms, add them
898 // to the call node, & reset all offsets to match call node array.
899 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
900 uint debug_end = debug_ptr;
901 uint debug_start = debug_ptr - in_jvms->debug_size();
902 debug_ptr = debug_start; // back up the ptr
904 uint p = debug_start; // walks forward in [debug_start, debug_end)
905 uint j, k, l;
906 SafePointNode* in_map = in_jvms->map();
907 out_jvms->set_map(call);
909 if (can_prune_locals) {
910 assert(in_jvms->method() == out_jvms->method(), "sanity");
911 // If the current throw can reach an exception handler in this JVMS,
912 // then we must keep everything live that can reach that handler.
913 // As a quick and dirty approximation, we look for any handlers at all.
914 if (in_jvms->method()->has_exception_handlers()) {
915 can_prune_locals = false;
916 }
917 }
919 // Add the Locals
920 k = in_jvms->locoff();
921 l = in_jvms->loc_size();
922 out_jvms->set_locoff(p);
923 if (!can_prune_locals) {
924 for (j = 0; j < l; j++)
925 call->set_req(p++, in_map->in(k+j));
926 } else {
927 p += l; // already set to top above by add_req_batch
928 }
930 // Add the Expression Stack
931 k = in_jvms->stkoff();
932 l = in_jvms->sp();
933 out_jvms->set_stkoff(p);
934 if (!can_prune_locals) {
935 for (j = 0; j < l; j++)
936 call->set_req(p++, in_map->in(k+j));
937 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
938 // Divide stack into {S0,...,S1}, where S0 is set to top.
939 uint s1 = stack_slots_not_pruned;
940 stack_slots_not_pruned = 0; // for next iteration
941 if (s1 > l) s1 = l;
942 uint s0 = l - s1;
943 p += s0; // skip the tops preinstalled by add_req_batch
944 for (j = s0; j < l; j++)
945 call->set_req(p++, in_map->in(k+j));
946 } else {
947 p += l; // already set to top above by add_req_batch
948 }
950 // Add the Monitors
951 k = in_jvms->monoff();
952 l = in_jvms->mon_size();
953 out_jvms->set_monoff(p);
954 for (j = 0; j < l; j++)
955 call->set_req(p++, in_map->in(k+j));
957 // Copy any scalar object fields.
958 k = in_jvms->scloff();
959 l = in_jvms->scl_size();
960 out_jvms->set_scloff(p);
961 for (j = 0; j < l; j++)
962 call->set_req(p++, in_map->in(k+j));
964 // Finish the new jvms.
965 out_jvms->set_endoff(p);
967 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
968 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
969 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
970 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
971 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
972 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
974 // Update the two tail pointers in parallel.
975 out_jvms = out_jvms->caller();
976 in_jvms = in_jvms->caller();
977 }
979 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
981 // Test the correctness of JVMState::debug_xxx accessors:
982 assert(call->jvms()->debug_start() == non_debug_edges, "");
983 assert(call->jvms()->debug_end() == call->req(), "");
984 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
985 }
987 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
988 Bytecodes::Code code = java_bc();
989 if (code == Bytecodes::_wide) {
990 code = method()->java_code_at_bci(bci() + 1);
991 }
993 BasicType rtype = T_ILLEGAL;
994 int rsize = 0;
996 if (code != Bytecodes::_illegal) {
997 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
998 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
999 if (rtype < T_CONFLICT)
1000 rsize = type2size[rtype];
1001 }
1003 switch (code) {
1004 case Bytecodes::_illegal:
1005 return false;
1007 case Bytecodes::_ldc:
1008 case Bytecodes::_ldc_w:
1009 case Bytecodes::_ldc2_w:
1010 inputs = 0;
1011 break;
1013 case Bytecodes::_dup: inputs = 1; break;
1014 case Bytecodes::_dup_x1: inputs = 2; break;
1015 case Bytecodes::_dup_x2: inputs = 3; break;
1016 case Bytecodes::_dup2: inputs = 2; break;
1017 case Bytecodes::_dup2_x1: inputs = 3; break;
1018 case Bytecodes::_dup2_x2: inputs = 4; break;
1019 case Bytecodes::_swap: inputs = 2; break;
1020 case Bytecodes::_arraylength: inputs = 1; break;
1022 case Bytecodes::_getstatic:
1023 case Bytecodes::_putstatic:
1024 case Bytecodes::_getfield:
1025 case Bytecodes::_putfield:
1026 {
1027 bool ignored_will_link;
1028 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1029 int size = field->type()->size();
1030 bool is_get = (depth >= 0), is_static = (depth & 1);
1031 inputs = (is_static ? 0 : 1);
1032 if (is_get) {
1033 depth = size - inputs;
1034 } else {
1035 inputs += size; // putxxx pops the value from the stack
1036 depth = - inputs;
1037 }
1038 }
1039 break;
1041 case Bytecodes::_invokevirtual:
1042 case Bytecodes::_invokespecial:
1043 case Bytecodes::_invokestatic:
1044 case Bytecodes::_invokedynamic:
1045 case Bytecodes::_invokeinterface:
1046 {
1047 bool ignored_will_link;
1048 ciSignature* declared_signature = NULL;
1049 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1050 assert(declared_signature != NULL, "cannot be null");
1051 inputs = declared_signature->arg_size_for_bc(code);
1052 int size = declared_signature->return_type()->size();
1053 depth = size - inputs;
1054 }
1055 break;
1057 case Bytecodes::_multianewarray:
1058 {
1059 ciBytecodeStream iter(method());
1060 iter.reset_to_bci(bci());
1061 iter.next();
1062 inputs = iter.get_dimensions();
1063 assert(rsize == 1, "");
1064 depth = rsize - inputs;
1065 }
1066 break;
1068 case Bytecodes::_ireturn:
1069 case Bytecodes::_lreturn:
1070 case Bytecodes::_freturn:
1071 case Bytecodes::_dreturn:
1072 case Bytecodes::_areturn:
1073 assert(rsize = -depth, "");
1074 inputs = rsize;
1075 break;
1077 case Bytecodes::_jsr:
1078 case Bytecodes::_jsr_w:
1079 inputs = 0;
1080 depth = 1; // S.B. depth=1, not zero
1081 break;
1083 default:
1084 // bytecode produces a typed result
1085 inputs = rsize - depth;
1086 assert(inputs >= 0, "");
1087 break;
1088 }
1090 #ifdef ASSERT
1091 // spot check
1092 int outputs = depth + inputs;
1093 assert(outputs >= 0, "sanity");
1094 switch (code) {
1095 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1096 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1097 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1098 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1099 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1100 }
1101 #endif //ASSERT
1103 return true;
1104 }
1108 //------------------------------basic_plus_adr---------------------------------
1109 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1110 // short-circuit a common case
1111 if (offset == intcon(0)) return ptr;
1112 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1113 }
1115 Node* GraphKit::ConvI2L(Node* offset) {
1116 // short-circuit a common case
1117 jint offset_con = find_int_con(offset, Type::OffsetBot);
1118 if (offset_con != Type::OffsetBot) {
1119 return longcon((jlong) offset_con);
1120 }
1121 return _gvn.transform( new (C) ConvI2LNode(offset));
1122 }
1123 Node* GraphKit::ConvL2I(Node* offset) {
1124 // short-circuit a common case
1125 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1126 if (offset_con != (jlong)Type::OffsetBot) {
1127 return intcon((int) offset_con);
1128 }
1129 return _gvn.transform( new (C) ConvL2INode(offset));
1130 }
1132 //-------------------------load_object_klass-----------------------------------
1133 Node* GraphKit::load_object_klass(Node* obj) {
1134 // Special-case a fresh allocation to avoid building nodes:
1135 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1136 if (akls != NULL) return akls;
1137 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1138 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1139 }
1141 //-------------------------load_array_length-----------------------------------
1142 Node* GraphKit::load_array_length(Node* array) {
1143 // Special-case a fresh allocation to avoid building nodes:
1144 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1145 Node *alen;
1146 if (alloc == NULL) {
1147 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1148 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1149 } else {
1150 alen = alloc->Ideal_length();
1151 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1152 if (ccast != alen) {
1153 alen = _gvn.transform(ccast);
1154 }
1155 }
1156 return alen;
1157 }
1159 //------------------------------do_null_check----------------------------------
1160 // Helper function to do a NULL pointer check. Returned value is
1161 // the incoming address with NULL casted away. You are allowed to use the
1162 // not-null value only if you are control dependent on the test.
1163 extern int explicit_null_checks_inserted,
1164 explicit_null_checks_elided;
1165 Node* GraphKit::null_check_common(Node* value, BasicType type,
1166 // optional arguments for variations:
1167 bool assert_null,
1168 Node* *null_control) {
1169 assert(!assert_null || null_control == NULL, "not both at once");
1170 if (stopped()) return top();
1171 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1172 // For some performance testing, we may wish to suppress null checking.
1173 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1174 return value;
1175 }
1176 explicit_null_checks_inserted++;
1178 // Construct NULL check
1179 Node *chk = NULL;
1180 switch(type) {
1181 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1182 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1183 case T_ARRAY : // fall through
1184 type = T_OBJECT; // simplify further tests
1185 case T_OBJECT : {
1186 const Type *t = _gvn.type( value );
1188 const TypeOopPtr* tp = t->isa_oopptr();
1189 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1190 // Only for do_null_check, not any of its siblings:
1191 && !assert_null && null_control == NULL) {
1192 // Usually, any field access or invocation on an unloaded oop type
1193 // will simply fail to link, since the statically linked class is
1194 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1195 // the static class is loaded but the sharper oop type is not.
1196 // Rather than checking for this obscure case in lots of places,
1197 // we simply observe that a null check on an unloaded class
1198 // will always be followed by a nonsense operation, so we
1199 // can just issue the uncommon trap here.
1200 // Our access to the unloaded class will only be correct
1201 // after it has been loaded and initialized, which requires
1202 // a trip through the interpreter.
1203 #ifndef PRODUCT
1204 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1205 #endif
1206 uncommon_trap(Deoptimization::Reason_unloaded,
1207 Deoptimization::Action_reinterpret,
1208 tp->klass(), "!loaded");
1209 return top();
1210 }
1212 if (assert_null) {
1213 // See if the type is contained in NULL_PTR.
1214 // If so, then the value is already null.
1215 if (t->higher_equal(TypePtr::NULL_PTR)) {
1216 explicit_null_checks_elided++;
1217 return value; // Elided null assert quickly!
1218 }
1219 } else {
1220 // See if mixing in the NULL pointer changes type.
1221 // If so, then the NULL pointer was not allowed in the original
1222 // type. In other words, "value" was not-null.
1223 if (t->meet(TypePtr::NULL_PTR) != t) {
1224 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1225 explicit_null_checks_elided++;
1226 return value; // Elided null check quickly!
1227 }
1228 }
1229 chk = new (C) CmpPNode( value, null() );
1230 break;
1231 }
1233 default:
1234 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1235 }
1236 assert(chk != NULL, "sanity check");
1237 chk = _gvn.transform(chk);
1239 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1240 BoolNode *btst = new (C) BoolNode( chk, btest);
1241 Node *tst = _gvn.transform( btst );
1243 //-----------
1244 // if peephole optimizations occurred, a prior test existed.
1245 // If a prior test existed, maybe it dominates as we can avoid this test.
1246 if (tst != btst && type == T_OBJECT) {
1247 // At this point we want to scan up the CFG to see if we can
1248 // find an identical test (and so avoid this test altogether).
1249 Node *cfg = control();
1250 int depth = 0;
1251 while( depth < 16 ) { // Limit search depth for speed
1252 if( cfg->Opcode() == Op_IfTrue &&
1253 cfg->in(0)->in(1) == tst ) {
1254 // Found prior test. Use "cast_not_null" to construct an identical
1255 // CastPP (and hence hash to) as already exists for the prior test.
1256 // Return that casted value.
1257 if (assert_null) {
1258 replace_in_map(value, null());
1259 return null(); // do not issue the redundant test
1260 }
1261 Node *oldcontrol = control();
1262 set_control(cfg);
1263 Node *res = cast_not_null(value);
1264 set_control(oldcontrol);
1265 explicit_null_checks_elided++;
1266 return res;
1267 }
1268 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1269 if (cfg == NULL) break; // Quit at region nodes
1270 depth++;
1271 }
1272 }
1274 //-----------
1275 // Branch to failure if null
1276 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1277 Deoptimization::DeoptReason reason;
1278 if (assert_null)
1279 reason = Deoptimization::Reason_null_assert;
1280 else if (type == T_OBJECT)
1281 reason = Deoptimization::Reason_null_check;
1282 else
1283 reason = Deoptimization::Reason_div0_check;
1285 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1286 // ciMethodData::has_trap_at will return a conservative -1 if any
1287 // must-be-null assertion has failed. This could cause performance
1288 // problems for a method after its first do_null_assert failure.
1289 // Consider using 'Reason_class_check' instead?
1291 // To cause an implicit null check, we set the not-null probability
1292 // to the maximum (PROB_MAX). For an explicit check the probability
1293 // is set to a smaller value.
1294 if (null_control != NULL || too_many_traps(reason)) {
1295 // probability is less likely
1296 ok_prob = PROB_LIKELY_MAG(3);
1297 } else if (!assert_null &&
1298 (ImplicitNullCheckThreshold > 0) &&
1299 method() != NULL &&
1300 (method()->method_data()->trap_count(reason)
1301 >= (uint)ImplicitNullCheckThreshold)) {
1302 ok_prob = PROB_LIKELY_MAG(3);
1303 }
1305 if (null_control != NULL) {
1306 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1307 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1308 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1309 if (null_true == top())
1310 explicit_null_checks_elided++;
1311 (*null_control) = null_true;
1312 } else {
1313 BuildCutout unless(this, tst, ok_prob);
1314 // Check for optimizer eliding test at parse time
1315 if (stopped()) {
1316 // Failure not possible; do not bother making uncommon trap.
1317 explicit_null_checks_elided++;
1318 } else if (assert_null) {
1319 uncommon_trap(reason,
1320 Deoptimization::Action_make_not_entrant,
1321 NULL, "assert_null");
1322 } else {
1323 replace_in_map(value, zerocon(type));
1324 builtin_throw(reason);
1325 }
1326 }
1328 // Must throw exception, fall-thru not possible?
1329 if (stopped()) {
1330 return top(); // No result
1331 }
1333 if (assert_null) {
1334 // Cast obj to null on this path.
1335 replace_in_map(value, zerocon(type));
1336 return zerocon(type);
1337 }
1339 // Cast obj to not-null on this path, if there is no null_control.
1340 // (If there is a null_control, a non-null value may come back to haunt us.)
1341 if (type == T_OBJECT) {
1342 Node* cast = cast_not_null(value, false);
1343 if (null_control == NULL || (*null_control) == top())
1344 replace_in_map(value, cast);
1345 value = cast;
1346 }
1348 return value;
1349 }
1352 //------------------------------cast_not_null----------------------------------
1353 // Cast obj to not-null on this path
1354 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1355 const Type *t = _gvn.type(obj);
1356 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1357 // Object is already not-null?
1358 if( t == t_not_null ) return obj;
1360 Node *cast = new (C) CastPPNode(obj,t_not_null);
1361 cast->init_req(0, control());
1362 cast = _gvn.transform( cast );
1364 // Scan for instances of 'obj' in the current JVM mapping.
1365 // These instances are known to be not-null after the test.
1366 if (do_replace_in_map)
1367 replace_in_map(obj, cast);
1369 return cast; // Return casted value
1370 }
1373 //--------------------------replace_in_map-------------------------------------
1374 void GraphKit::replace_in_map(Node* old, Node* neww) {
1375 this->map()->replace_edge(old, neww);
1377 // Note: This operation potentially replaces any edge
1378 // on the map. This includes locals, stack, and monitors
1379 // of the current (innermost) JVM state.
1381 // We can consider replacing in caller maps.
1382 // The idea would be that an inlined function's null checks
1383 // can be shared with the entire inlining tree.
1384 // The expense of doing this is that the PreserveJVMState class
1385 // would have to preserve caller states too, with a deep copy.
1386 }
1389 //=============================================================================
1390 //--------------------------------memory---------------------------------------
1391 Node* GraphKit::memory(uint alias_idx) {
1392 MergeMemNode* mem = merged_memory();
1393 Node* p = mem->memory_at(alias_idx);
1394 _gvn.set_type(p, Type::MEMORY); // must be mapped
1395 return p;
1396 }
1398 //-----------------------------reset_memory------------------------------------
1399 Node* GraphKit::reset_memory() {
1400 Node* mem = map()->memory();
1401 // do not use this node for any more parsing!
1402 debug_only( map()->set_memory((Node*)NULL) );
1403 return _gvn.transform( mem );
1404 }
1406 //------------------------------set_all_memory---------------------------------
1407 void GraphKit::set_all_memory(Node* newmem) {
1408 Node* mergemem = MergeMemNode::make(C, newmem);
1409 gvn().set_type_bottom(mergemem);
1410 map()->set_memory(mergemem);
1411 }
1413 //------------------------------set_all_memory_call----------------------------
1414 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1415 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1416 set_all_memory(newmem);
1417 }
1419 //=============================================================================
1420 //
1421 // parser factory methods for MemNodes
1422 //
1423 // These are layered on top of the factory methods in LoadNode and StoreNode,
1424 // and integrate with the parser's memory state and _gvn engine.
1425 //
1427 // factory methods in "int adr_idx"
1428 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1429 int adr_idx,
1430 bool require_atomic_access) {
1431 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1432 const TypePtr* adr_type = NULL; // debug-mode-only argument
1433 debug_only(adr_type = C->get_adr_type(adr_idx));
1434 Node* mem = memory(adr_idx);
1435 Node* ld;
1436 if (require_atomic_access && bt == T_LONG) {
1437 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1438 } else {
1439 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1440 }
1441 return _gvn.transform(ld);
1442 }
1444 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1445 int adr_idx,
1446 bool require_atomic_access) {
1447 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1448 const TypePtr* adr_type = NULL;
1449 debug_only(adr_type = C->get_adr_type(adr_idx));
1450 Node *mem = memory(adr_idx);
1451 Node* st;
1452 if (require_atomic_access && bt == T_LONG) {
1453 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1454 } else {
1455 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1456 }
1457 st = _gvn.transform(st);
1458 set_memory(st, adr_idx);
1459 // Back-to-back stores can only remove intermediate store with DU info
1460 // so push on worklist for optimizer.
1461 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1462 record_for_igvn(st);
1464 return st;
1465 }
1468 void GraphKit::pre_barrier(bool do_load,
1469 Node* ctl,
1470 Node* obj,
1471 Node* adr,
1472 uint adr_idx,
1473 Node* val,
1474 const TypeOopPtr* val_type,
1475 Node* pre_val,
1476 BasicType bt) {
1478 BarrierSet* bs = Universe::heap()->barrier_set();
1479 set_control(ctl);
1480 switch (bs->kind()) {
1481 case BarrierSet::G1SATBCT:
1482 case BarrierSet::G1SATBCTLogging:
1483 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1484 break;
1486 case BarrierSet::CardTableModRef:
1487 case BarrierSet::CardTableExtension:
1488 case BarrierSet::ModRef:
1489 break;
1491 case BarrierSet::Other:
1492 default :
1493 ShouldNotReachHere();
1495 }
1496 }
1498 void GraphKit::post_barrier(Node* ctl,
1499 Node* store,
1500 Node* obj,
1501 Node* adr,
1502 uint adr_idx,
1503 Node* val,
1504 BasicType bt,
1505 bool use_precise) {
1506 BarrierSet* bs = Universe::heap()->barrier_set();
1507 set_control(ctl);
1508 switch (bs->kind()) {
1509 case BarrierSet::G1SATBCT:
1510 case BarrierSet::G1SATBCTLogging:
1511 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1512 break;
1514 case BarrierSet::CardTableModRef:
1515 case BarrierSet::CardTableExtension:
1516 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1517 break;
1519 case BarrierSet::ModRef:
1520 break;
1522 case BarrierSet::Other:
1523 default :
1524 ShouldNotReachHere();
1526 }
1527 }
1529 Node* GraphKit::store_oop(Node* ctl,
1530 Node* obj,
1531 Node* adr,
1532 const TypePtr* adr_type,
1533 Node* val,
1534 const TypeOopPtr* val_type,
1535 BasicType bt,
1536 bool use_precise) {
1537 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1538 // could be delayed during Parse (for example, in adjust_map_after_if()).
1539 // Execute transformation here to avoid barrier generation in such case.
1540 if (_gvn.type(val) == TypePtr::NULL_PTR)
1541 val = _gvn.makecon(TypePtr::NULL_PTR);
1543 set_control(ctl);
1544 if (stopped()) return top(); // Dead path ?
1546 assert(bt == T_OBJECT, "sanity");
1547 assert(val != NULL, "not dead path");
1548 uint adr_idx = C->get_alias_index(adr_type);
1549 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1551 pre_barrier(true /* do_load */,
1552 control(), obj, adr, adr_idx, val, val_type,
1553 NULL /* pre_val */,
1554 bt);
1556 Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1557 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1558 return store;
1559 }
1561 // Could be an array or object we don't know at compile time (unsafe ref.)
1562 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1563 Node* obj, // containing obj
1564 Node* adr, // actual adress to store val at
1565 const TypePtr* adr_type,
1566 Node* val,
1567 BasicType bt) {
1568 Compile::AliasType* at = C->alias_type(adr_type);
1569 const TypeOopPtr* val_type = NULL;
1570 if (adr_type->isa_instptr()) {
1571 if (at->field() != NULL) {
1572 // known field. This code is a copy of the do_put_xxx logic.
1573 ciField* field = at->field();
1574 if (!field->type()->is_loaded()) {
1575 val_type = TypeInstPtr::BOTTOM;
1576 } else {
1577 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1578 }
1579 }
1580 } else if (adr_type->isa_aryptr()) {
1581 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1582 }
1583 if (val_type == NULL) {
1584 val_type = TypeInstPtr::BOTTOM;
1585 }
1586 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true);
1587 }
1590 //-------------------------array_element_address-------------------------
1591 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1592 const TypeInt* sizetype) {
1593 uint shift = exact_log2(type2aelembytes(elembt));
1594 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1596 // short-circuit a common case (saves lots of confusing waste motion)
1597 jint idx_con = find_int_con(idx, -1);
1598 if (idx_con >= 0) {
1599 intptr_t offset = header + ((intptr_t)idx_con << shift);
1600 return basic_plus_adr(ary, offset);
1601 }
1603 // must be correct type for alignment purposes
1604 Node* base = basic_plus_adr(ary, header);
1605 #ifdef _LP64
1606 // The scaled index operand to AddP must be a clean 64-bit value.
1607 // Java allows a 32-bit int to be incremented to a negative
1608 // value, which appears in a 64-bit register as a large
1609 // positive number. Using that large positive number as an
1610 // operand in pointer arithmetic has bad consequences.
1611 // On the other hand, 32-bit overflow is rare, and the possibility
1612 // can often be excluded, if we annotate the ConvI2L node with
1613 // a type assertion that its value is known to be a small positive
1614 // number. (The prior range check has ensured this.)
1615 // This assertion is used by ConvI2LNode::Ideal.
1616 int index_max = max_jint - 1; // array size is max_jint, index is one less
1617 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1618 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1619 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) );
1620 #endif
1621 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1622 return basic_plus_adr(ary, base, scale);
1623 }
1625 //-------------------------load_array_element-------------------------
1626 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1627 const Type* elemtype = arytype->elem();
1628 BasicType elembt = elemtype->array_element_basic_type();
1629 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1630 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1631 return ld;
1632 }
1634 //-------------------------set_arguments_for_java_call-------------------------
1635 // Arguments (pre-popped from the stack) are taken from the JVMS.
1636 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1637 // Add the call arguments:
1638 uint nargs = call->method()->arg_size();
1639 for (uint i = 0; i < nargs; i++) {
1640 Node* arg = argument(i);
1641 call->init_req(i + TypeFunc::Parms, arg);
1642 }
1643 }
1645 //---------------------------set_edges_for_java_call---------------------------
1646 // Connect a newly created call into the current JVMS.
1647 // A return value node (if any) is returned from set_edges_for_java_call.
1648 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1650 // Add the predefined inputs:
1651 call->init_req( TypeFunc::Control, control() );
1652 call->init_req( TypeFunc::I_O , i_o() );
1653 call->init_req( TypeFunc::Memory , reset_memory() );
1654 call->init_req( TypeFunc::FramePtr, frameptr() );
1655 call->init_req( TypeFunc::ReturnAdr, top() );
1657 add_safepoint_edges(call, must_throw);
1659 Node* xcall = _gvn.transform(call);
1661 if (xcall == top()) {
1662 set_control(top());
1663 return;
1664 }
1665 assert(xcall == call, "call identity is stable");
1667 // Re-use the current map to produce the result.
1669 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1670 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1671 set_all_memory_call(xcall, separate_io_proj);
1673 //return xcall; // no need, caller already has it
1674 }
1676 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1677 if (stopped()) return top(); // maybe the call folded up?
1679 // Capture the return value, if any.
1680 Node* ret;
1681 if (call->method() == NULL ||
1682 call->method()->return_type()->basic_type() == T_VOID)
1683 ret = top();
1684 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1686 // Note: Since any out-of-line call can produce an exception,
1687 // we always insert an I_O projection from the call into the result.
1689 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1691 if (separate_io_proj) {
1692 // The caller requested separate projections be used by the fall
1693 // through and exceptional paths, so replace the projections for
1694 // the fall through path.
1695 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1696 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1697 }
1698 return ret;
1699 }
1701 //--------------------set_predefined_input_for_runtime_call--------------------
1702 // Reading and setting the memory state is way conservative here.
1703 // The real problem is that I am not doing real Type analysis on memory,
1704 // so I cannot distinguish card mark stores from other stores. Across a GC
1705 // point the Store Barrier and the card mark memory has to agree. I cannot
1706 // have a card mark store and its barrier split across the GC point from
1707 // either above or below. Here I get that to happen by reading ALL of memory.
1708 // A better answer would be to separate out card marks from other memory.
1709 // For now, return the input memory state, so that it can be reused
1710 // after the call, if this call has restricted memory effects.
1711 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1712 // Set fixed predefined input arguments
1713 Node* memory = reset_memory();
1714 call->init_req( TypeFunc::Control, control() );
1715 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1716 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1717 call->init_req( TypeFunc::FramePtr, frameptr() );
1718 call->init_req( TypeFunc::ReturnAdr, top() );
1719 return memory;
1720 }
1722 //-------------------set_predefined_output_for_runtime_call--------------------
1723 // Set control and memory (not i_o) from the call.
1724 // If keep_mem is not NULL, use it for the output state,
1725 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1726 // If hook_mem is NULL, this call produces no memory effects at all.
1727 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1728 // then only that memory slice is taken from the call.
1729 // In the last case, we must put an appropriate memory barrier before
1730 // the call, so as to create the correct anti-dependencies on loads
1731 // preceding the call.
1732 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1733 Node* keep_mem,
1734 const TypePtr* hook_mem) {
1735 // no i/o
1736 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1737 if (keep_mem) {
1738 // First clone the existing memory state
1739 set_all_memory(keep_mem);
1740 if (hook_mem != NULL) {
1741 // Make memory for the call
1742 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1743 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1744 // We also use hook_mem to extract specific effects from arraycopy stubs.
1745 set_memory(mem, hook_mem);
1746 }
1747 // ...else the call has NO memory effects.
1749 // Make sure the call advertises its memory effects precisely.
1750 // This lets us build accurate anti-dependences in gcm.cpp.
1751 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1752 "call node must be constructed correctly");
1753 } else {
1754 assert(hook_mem == NULL, "");
1755 // This is not a "slow path" call; all memory comes from the call.
1756 set_all_memory_call(call);
1757 }
1758 }
1761 // Replace the call with the current state of the kit.
1762 void GraphKit::replace_call(CallNode* call, Node* result) {
1763 JVMState* ejvms = NULL;
1764 if (has_exceptions()) {
1765 ejvms = transfer_exceptions_into_jvms();
1766 }
1768 SafePointNode* final_state = stop();
1770 // Find all the needed outputs of this call
1771 CallProjections callprojs;
1772 call->extract_projections(&callprojs, true);
1774 Node* init_mem = call->in(TypeFunc::Memory);
1775 Node* final_mem = final_state->in(TypeFunc::Memory);
1776 Node* final_ctl = final_state->in(TypeFunc::Control);
1777 Node* final_io = final_state->in(TypeFunc::I_O);
1779 // Replace all the old call edges with the edges from the inlining result
1780 if (callprojs.fallthrough_catchproj != NULL) {
1781 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1782 }
1783 if (callprojs.fallthrough_memproj != NULL) {
1784 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1785 }
1786 if (callprojs.fallthrough_ioproj != NULL) {
1787 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1788 }
1790 // Replace the result with the new result if it exists and is used
1791 if (callprojs.resproj != NULL && result != NULL) {
1792 C->gvn_replace_by(callprojs.resproj, result);
1793 }
1795 if (ejvms == NULL) {
1796 // No exception edges to simply kill off those paths
1797 if (callprojs.catchall_catchproj != NULL) {
1798 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1799 }
1800 if (callprojs.catchall_memproj != NULL) {
1801 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1802 }
1803 if (callprojs.catchall_ioproj != NULL) {
1804 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1805 }
1806 // Replace the old exception object with top
1807 if (callprojs.exobj != NULL) {
1808 C->gvn_replace_by(callprojs.exobj, C->top());
1809 }
1810 } else {
1811 GraphKit ekit(ejvms);
1813 // Load my combined exception state into the kit, with all phis transformed:
1814 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1816 Node* ex_oop = ekit.use_exception_state(ex_map);
1817 if (callprojs.catchall_catchproj != NULL) {
1818 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1819 }
1820 if (callprojs.catchall_memproj != NULL) {
1821 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1822 }
1823 if (callprojs.catchall_ioproj != NULL) {
1824 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1825 }
1827 // Replace the old exception object with the newly created one
1828 if (callprojs.exobj != NULL) {
1829 C->gvn_replace_by(callprojs.exobj, ex_oop);
1830 }
1831 }
1833 // Disconnect the call from the graph
1834 call->disconnect_inputs(NULL, C);
1835 C->gvn_replace_by(call, C->top());
1837 // Clean up any MergeMems that feed other MergeMems since the
1838 // optimizer doesn't like that.
1839 if (final_mem->is_MergeMem()) {
1840 Node_List wl;
1841 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1842 Node* m = i.get();
1843 if (m->is_MergeMem() && !wl.contains(m)) {
1844 wl.push(m);
1845 }
1846 }
1847 while (wl.size() > 0) {
1848 _gvn.transform(wl.pop());
1849 }
1850 }
1851 }
1854 //------------------------------increment_counter------------------------------
1855 // for statistics: increment a VM counter by 1
1857 void GraphKit::increment_counter(address counter_addr) {
1858 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1859 increment_counter(adr1);
1860 }
1862 void GraphKit::increment_counter(Node* counter_addr) {
1863 int adr_type = Compile::AliasIdxRaw;
1864 Node* ctrl = control();
1865 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type);
1866 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1867 store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type );
1868 }
1871 //------------------------------uncommon_trap----------------------------------
1872 // Bail out to the interpreter in mid-method. Implemented by calling the
1873 // uncommon_trap blob. This helper function inserts a runtime call with the
1874 // right debug info.
1875 void GraphKit::uncommon_trap(int trap_request,
1876 ciKlass* klass, const char* comment,
1877 bool must_throw,
1878 bool keep_exact_action) {
1879 if (failing()) stop();
1880 if (stopped()) return; // trap reachable?
1882 // Note: If ProfileTraps is true, and if a deopt. actually
1883 // occurs here, the runtime will make sure an MDO exists. There is
1884 // no need to call method()->ensure_method_data() at this point.
1886 // Set the stack pointer to the right value for reexecution:
1887 set_sp(reexecute_sp());
1889 #ifdef ASSERT
1890 if (!must_throw) {
1891 // Make sure the stack has at least enough depth to execute
1892 // the current bytecode.
1893 int inputs, ignored_depth;
1894 if (compute_stack_effects(inputs, ignored_depth)) {
1895 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1896 Bytecodes::name(java_bc()), sp(), inputs));
1897 }
1898 }
1899 #endif
1901 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1902 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1904 switch (action) {
1905 case Deoptimization::Action_maybe_recompile:
1906 case Deoptimization::Action_reinterpret:
1907 // Temporary fix for 6529811 to allow virtual calls to be sure they
1908 // get the chance to go from mono->bi->mega
1909 if (!keep_exact_action &&
1910 Deoptimization::trap_request_index(trap_request) < 0 &&
1911 too_many_recompiles(reason)) {
1912 // This BCI is causing too many recompilations.
1913 action = Deoptimization::Action_none;
1914 trap_request = Deoptimization::make_trap_request(reason, action);
1915 } else {
1916 C->set_trap_can_recompile(true);
1917 }
1918 break;
1919 case Deoptimization::Action_make_not_entrant:
1920 C->set_trap_can_recompile(true);
1921 break;
1922 #ifdef ASSERT
1923 case Deoptimization::Action_none:
1924 case Deoptimization::Action_make_not_compilable:
1925 break;
1926 default:
1927 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
1928 break;
1929 #endif
1930 }
1932 if (TraceOptoParse) {
1933 char buf[100];
1934 tty->print_cr("Uncommon trap %s at bci:%d",
1935 Deoptimization::format_trap_request(buf, sizeof(buf),
1936 trap_request), bci());
1937 }
1939 CompileLog* log = C->log();
1940 if (log != NULL) {
1941 int kid = (klass == NULL)? -1: log->identify(klass);
1942 log->begin_elem("uncommon_trap bci='%d'", bci());
1943 char buf[100];
1944 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1945 trap_request));
1946 if (kid >= 0) log->print(" klass='%d'", kid);
1947 if (comment != NULL) log->print(" comment='%s'", comment);
1948 log->end_elem();
1949 }
1951 // Make sure any guarding test views this path as very unlikely
1952 Node *i0 = control()->in(0);
1953 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1954 IfNode *iff = i0->as_If();
1955 float f = iff->_prob; // Get prob
1956 if (control()->Opcode() == Op_IfTrue) {
1957 if (f > PROB_UNLIKELY_MAG(4))
1958 iff->_prob = PROB_MIN;
1959 } else {
1960 if (f < PROB_LIKELY_MAG(4))
1961 iff->_prob = PROB_MAX;
1962 }
1963 }
1965 // Clear out dead values from the debug info.
1966 kill_dead_locals();
1968 // Now insert the uncommon trap subroutine call
1969 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
1970 const TypePtr* no_memory_effects = NULL;
1971 // Pass the index of the class to be loaded
1972 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1973 (must_throw ? RC_MUST_THROW : 0),
1974 OptoRuntime::uncommon_trap_Type(),
1975 call_addr, "uncommon_trap", no_memory_effects,
1976 intcon(trap_request));
1977 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1978 "must extract request correctly from the graph");
1979 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1981 call->set_req(TypeFunc::ReturnAdr, returnadr());
1982 // The debug info is the only real input to this call.
1984 // Halt-and-catch fire here. The above call should never return!
1985 HaltNode* halt = new(C) HaltNode(control(), frameptr());
1986 _gvn.set_type_bottom(halt);
1987 root()->add_req(halt);
1989 stop_and_kill_map();
1990 }
1993 //--------------------------just_allocated_object------------------------------
1994 // Report the object that was just allocated.
1995 // It must be the case that there are no intervening safepoints.
1996 // We use this to determine if an object is so "fresh" that
1997 // it does not require card marks.
1998 Node* GraphKit::just_allocated_object(Node* current_control) {
1999 if (C->recent_alloc_ctl() == current_control)
2000 return C->recent_alloc_obj();
2001 return NULL;
2002 }
2005 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2006 // (Note: TypeFunc::make has a cache that makes this fast.)
2007 const TypeFunc* tf = TypeFunc::make(dest_method);
2008 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2009 for (int j = 0; j < nargs; j++) {
2010 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2011 if( targ->basic_type() == T_DOUBLE ) {
2012 // If any parameters are doubles, they must be rounded before
2013 // the call, dstore_rounding does gvn.transform
2014 Node *arg = argument(j);
2015 arg = dstore_rounding(arg);
2016 set_argument(j, arg);
2017 }
2018 }
2019 }
2021 void GraphKit::round_double_result(ciMethod* dest_method) {
2022 // A non-strict method may return a double value which has an extended
2023 // exponent, but this must not be visible in a caller which is 'strict'
2024 // If a strict caller invokes a non-strict callee, round a double result
2026 BasicType result_type = dest_method->return_type()->basic_type();
2027 assert( method() != NULL, "must have caller context");
2028 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2029 // Destination method's return value is on top of stack
2030 // dstore_rounding() does gvn.transform
2031 Node *result = pop_pair();
2032 result = dstore_rounding(result);
2033 push_pair(result);
2034 }
2035 }
2037 // rounding for strict float precision conformance
2038 Node* GraphKit::precision_rounding(Node* n) {
2039 return UseStrictFP && _method->flags().is_strict()
2040 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2041 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2042 : n;
2043 }
2045 // rounding for strict double precision conformance
2046 Node* GraphKit::dprecision_rounding(Node *n) {
2047 return UseStrictFP && _method->flags().is_strict()
2048 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2049 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2050 : n;
2051 }
2053 // rounding for non-strict double stores
2054 Node* GraphKit::dstore_rounding(Node* n) {
2055 return Matcher::strict_fp_requires_explicit_rounding
2056 && UseSSE <= 1
2057 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2058 : n;
2059 }
2061 //=============================================================================
2062 // Generate a fast path/slow path idiom. Graph looks like:
2063 // [foo] indicates that 'foo' is a parameter
2064 //
2065 // [in] NULL
2066 // \ /
2067 // CmpP
2068 // Bool ne
2069 // If
2070 // / \
2071 // True False-<2>
2072 // / |
2073 // / cast_not_null
2074 // Load | | ^
2075 // [fast_test] | |
2076 // gvn to opt_test | |
2077 // / \ | <1>
2078 // True False |
2079 // | \\ |
2080 // [slow_call] \[fast_result]
2081 // Ctl Val \ \
2082 // | \ \
2083 // Catch <1> \ \
2084 // / \ ^ \ \
2085 // Ex No_Ex | \ \
2086 // | \ \ | \ <2> \
2087 // ... \ [slow_res] | | \ [null_result]
2088 // \ \--+--+--- | |
2089 // \ | / \ | /
2090 // --------Region Phi
2091 //
2092 //=============================================================================
2093 // Code is structured as a series of driver functions all called 'do_XXX' that
2094 // call a set of helper functions. Helper functions first, then drivers.
2096 //------------------------------null_check_oop---------------------------------
2097 // Null check oop. Set null-path control into Region in slot 3.
2098 // Make a cast-not-nullness use the other not-null control. Return cast.
2099 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2100 bool never_see_null) {
2101 // Initial NULL check taken path
2102 (*null_control) = top();
2103 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2105 // Generate uncommon_trap:
2106 if (never_see_null && (*null_control) != top()) {
2107 // If we see an unexpected null at a check-cast we record it and force a
2108 // recompile; the offending check-cast will be compiled to handle NULLs.
2109 // If we see more than one offending BCI, then all checkcasts in the
2110 // method will be compiled to handle NULLs.
2111 PreserveJVMState pjvms(this);
2112 set_control(*null_control);
2113 replace_in_map(value, null());
2114 uncommon_trap(Deoptimization::Reason_null_check,
2115 Deoptimization::Action_make_not_entrant);
2116 (*null_control) = top(); // NULL path is dead
2117 }
2119 // Cast away null-ness on the result
2120 return cast;
2121 }
2123 //------------------------------opt_iff----------------------------------------
2124 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2125 // Return slow-path control.
2126 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2127 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2129 // Fast path taken; set region slot 2
2130 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2131 region->init_req(2,fast_taken); // Capture fast-control
2133 // Fast path not-taken, i.e. slow path
2134 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2135 return slow_taken;
2136 }
2138 //-----------------------------make_runtime_call-------------------------------
2139 Node* GraphKit::make_runtime_call(int flags,
2140 const TypeFunc* call_type, address call_addr,
2141 const char* call_name,
2142 const TypePtr* adr_type,
2143 // The following parms are all optional.
2144 // The first NULL ends the list.
2145 Node* parm0, Node* parm1,
2146 Node* parm2, Node* parm3,
2147 Node* parm4, Node* parm5,
2148 Node* parm6, Node* parm7) {
2149 // Slow-path call
2150 bool is_leaf = !(flags & RC_NO_LEAF);
2151 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2152 if (call_name == NULL) {
2153 assert(!is_leaf, "must supply name for leaf");
2154 call_name = OptoRuntime::stub_name(call_addr);
2155 }
2156 CallNode* call;
2157 if (!is_leaf) {
2158 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2159 bci(), adr_type);
2160 } else if (flags & RC_NO_FP) {
2161 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2162 } else {
2163 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2164 }
2166 // The following is similar to set_edges_for_java_call,
2167 // except that the memory effects of the call are restricted to AliasIdxRaw.
2169 // Slow path call has no side-effects, uses few values
2170 bool wide_in = !(flags & RC_NARROW_MEM);
2171 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2173 Node* prev_mem = NULL;
2174 if (wide_in) {
2175 prev_mem = set_predefined_input_for_runtime_call(call);
2176 } else {
2177 assert(!wide_out, "narrow in => narrow out");
2178 Node* narrow_mem = memory(adr_type);
2179 prev_mem = reset_memory();
2180 map()->set_memory(narrow_mem);
2181 set_predefined_input_for_runtime_call(call);
2182 }
2184 // Hook each parm in order. Stop looking at the first NULL.
2185 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2186 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2187 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2188 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2189 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2190 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2191 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2192 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2193 /* close each nested if ===> */ } } } } } } } }
2194 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2196 if (!is_leaf) {
2197 // Non-leaves can block and take safepoints:
2198 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2199 }
2200 // Non-leaves can throw exceptions:
2201 if (has_io) {
2202 call->set_req(TypeFunc::I_O, i_o());
2203 }
2205 if (flags & RC_UNCOMMON) {
2206 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2207 // (An "if" probability corresponds roughly to an unconditional count.
2208 // Sort of.)
2209 call->set_cnt(PROB_UNLIKELY_MAG(4));
2210 }
2212 Node* c = _gvn.transform(call);
2213 assert(c == call, "cannot disappear");
2215 if (wide_out) {
2216 // Slow path call has full side-effects.
2217 set_predefined_output_for_runtime_call(call);
2218 } else {
2219 // Slow path call has few side-effects, and/or sets few values.
2220 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2221 }
2223 if (has_io) {
2224 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2225 }
2226 return call;
2228 }
2230 //------------------------------merge_memory-----------------------------------
2231 // Merge memory from one path into the current memory state.
2232 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2233 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2234 Node* old_slice = mms.force_memory();
2235 Node* new_slice = mms.memory2();
2236 if (old_slice != new_slice) {
2237 PhiNode* phi;
2238 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2239 phi = new_slice->as_Phi();
2240 #ifdef ASSERT
2241 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2242 old_slice = old_slice->in(new_path);
2243 // Caller is responsible for ensuring that any pre-existing
2244 // phis are already aware of old memory.
2245 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2246 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2247 #endif
2248 mms.set_memory(phi);
2249 } else {
2250 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2251 _gvn.set_type(phi, Type::MEMORY);
2252 phi->set_req(new_path, new_slice);
2253 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2254 }
2255 }
2256 }
2257 }
2259 //------------------------------make_slow_call_ex------------------------------
2260 // Make the exception handler hookups for the slow call
2261 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2262 if (stopped()) return;
2264 // Make a catch node with just two handlers: fall-through and catch-all
2265 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2266 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2267 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2268 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2270 { PreserveJVMState pjvms(this);
2271 set_control(excp);
2272 set_i_o(i_o);
2274 if (excp != top()) {
2275 // Create an exception state also.
2276 // Use an exact type if the caller has specified a specific exception.
2277 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2278 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2279 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2280 }
2281 }
2283 // Get the no-exception control from the CatchNode.
2284 set_control(norm);
2285 }
2288 //-------------------------------gen_subtype_check-----------------------------
2289 // Generate a subtyping check. Takes as input the subtype and supertype.
2290 // Returns 2 values: sets the default control() to the true path and returns
2291 // the false path. Only reads invariant memory; sets no (visible) memory.
2292 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2293 // but that's not exposed to the optimizer. This call also doesn't take in an
2294 // Object; if you wish to check an Object you need to load the Object's class
2295 // prior to coming here.
2296 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2297 // Fast check for identical types, perhaps identical constants.
2298 // The types can even be identical non-constants, in cases
2299 // involving Array.newInstance, Object.clone, etc.
2300 if (subklass == superklass)
2301 return top(); // false path is dead; no test needed.
2303 if (_gvn.type(superklass)->singleton()) {
2304 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2305 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2307 // In the common case of an exact superklass, try to fold up the
2308 // test before generating code. You may ask, why not just generate
2309 // the code and then let it fold up? The answer is that the generated
2310 // code will necessarily include null checks, which do not always
2311 // completely fold away. If they are also needless, then they turn
2312 // into a performance loss. Example:
2313 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2314 // Here, the type of 'fa' is often exact, so the store check
2315 // of fa[1]=x will fold up, without testing the nullness of x.
2316 switch (static_subtype_check(superk, subk)) {
2317 case SSC_always_false:
2318 {
2319 Node* always_fail = control();
2320 set_control(top());
2321 return always_fail;
2322 }
2323 case SSC_always_true:
2324 return top();
2325 case SSC_easy_test:
2326 {
2327 // Just do a direct pointer compare and be done.
2328 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2329 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2330 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2331 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2332 return _gvn.transform( new(C) IfFalseNode(iff) );
2333 }
2334 case SSC_full_test:
2335 break;
2336 default:
2337 ShouldNotReachHere();
2338 }
2339 }
2341 // %%% Possible further optimization: Even if the superklass is not exact,
2342 // if the subklass is the unique subtype of the superklass, the check
2343 // will always succeed. We could leave a dependency behind to ensure this.
2345 // First load the super-klass's check-offset
2346 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2347 Node *chk_off = _gvn.transform( new (C) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2348 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2349 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2351 // Load from the sub-klass's super-class display list, or a 1-word cache of
2352 // the secondary superclass list, or a failing value with a sentinel offset
2353 // if the super-klass is an interface or exceptionally deep in the Java
2354 // hierarchy and we have to scan the secondary superclass list the hard way.
2355 // Worst-case type is a little odd: NULL is allowed as a result (usually
2356 // klass loads can never produce a NULL).
2357 Node *chk_off_X = ConvI2X(chk_off);
2358 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2359 // For some types like interfaces the following loadKlass is from a 1-word
2360 // cache which is mutable so can't use immutable memory. Other
2361 // types load from the super-class display table which is immutable.
2362 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2363 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2365 // Compile speed common case: ARE a subtype and we canNOT fail
2366 if( superklass == nkls )
2367 return top(); // false path is dead; no test needed.
2369 // See if we get an immediate positive hit. Happens roughly 83% of the
2370 // time. Test to see if the value loaded just previously from the subklass
2371 // is exactly the superklass.
2372 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2373 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2374 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2375 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2376 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2378 // Compile speed common case: Check for being deterministic right now. If
2379 // chk_off is a constant and not equal to cacheoff then we are NOT a
2380 // subklass. In this case we need exactly the 1 test above and we can
2381 // return those results immediately.
2382 if (!might_be_cache) {
2383 Node* not_subtype_ctrl = control();
2384 set_control(iftrue1); // We need exactly the 1 test above
2385 return not_subtype_ctrl;
2386 }
2388 // Gather the various success & failures here
2389 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2390 record_for_igvn(r_ok_subtype);
2391 RegionNode *r_not_subtype = new (C) RegionNode(3);
2392 record_for_igvn(r_not_subtype);
2394 r_ok_subtype->init_req(1, iftrue1);
2396 // Check for immediate negative hit. Happens roughly 11% of the time (which
2397 // is roughly 63% of the remaining cases). Test to see if the loaded
2398 // check-offset points into the subklass display list or the 1-element
2399 // cache. If it points to the display (and NOT the cache) and the display
2400 // missed then it's not a subtype.
2401 Node *cacheoff = _gvn.intcon(cacheoff_con);
2402 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2403 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2404 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2405 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2406 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2408 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2409 // No performance impact (too rare) but allows sharing of secondary arrays
2410 // which has some footprint reduction.
2411 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2412 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2413 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2414 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2415 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2417 // -- Roads not taken here: --
2418 // We could also have chosen to perform the self-check at the beginning
2419 // of this code sequence, as the assembler does. This would not pay off
2420 // the same way, since the optimizer, unlike the assembler, can perform
2421 // static type analysis to fold away many successful self-checks.
2422 // Non-foldable self checks work better here in second position, because
2423 // the initial primary superclass check subsumes a self-check for most
2424 // types. An exception would be a secondary type like array-of-interface,
2425 // which does not appear in its own primary supertype display.
2426 // Finally, we could have chosen to move the self-check into the
2427 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2428 // dependent manner. But it is worthwhile to have the check here,
2429 // where it can be perhaps be optimized. The cost in code space is
2430 // small (register compare, branch).
2432 // Now do a linear scan of the secondary super-klass array. Again, no real
2433 // performance impact (too rare) but it's gotta be done.
2434 // Since the code is rarely used, there is no penalty for moving it
2435 // out of line, and it can only improve I-cache density.
2436 // The decision to inline or out-of-line this final check is platform
2437 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2438 Node* psc = _gvn.transform(
2439 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2441 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2442 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2443 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2444 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2445 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2447 // Return false path; set default control to true path.
2448 set_control( _gvn.transform(r_ok_subtype) );
2449 return _gvn.transform(r_not_subtype);
2450 }
2452 //----------------------------static_subtype_check-----------------------------
2453 // Shortcut important common cases when superklass is exact:
2454 // (0) superklass is java.lang.Object (can occur in reflective code)
2455 // (1) subklass is already limited to a subtype of superklass => always ok
2456 // (2) subklass does not overlap with superklass => always fail
2457 // (3) superklass has NO subtypes and we can check with a simple compare.
2458 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2459 if (StressReflectiveCode) {
2460 return SSC_full_test; // Let caller generate the general case.
2461 }
2463 if (superk == env()->Object_klass()) {
2464 return SSC_always_true; // (0) this test cannot fail
2465 }
2467 ciType* superelem = superk;
2468 if (superelem->is_array_klass())
2469 superelem = superelem->as_array_klass()->base_element_type();
2471 if (!subk->is_interface()) { // cannot trust static interface types yet
2472 if (subk->is_subtype_of(superk)) {
2473 return SSC_always_true; // (1) false path dead; no dynamic test needed
2474 }
2475 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2476 !superk->is_subtype_of(subk)) {
2477 return SSC_always_false;
2478 }
2479 }
2481 // If casting to an instance klass, it must have no subtypes
2482 if (superk->is_interface()) {
2483 // Cannot trust interfaces yet.
2484 // %%% S.B. superk->nof_implementors() == 1
2485 } else if (superelem->is_instance_klass()) {
2486 ciInstanceKlass* ik = superelem->as_instance_klass();
2487 if (!ik->has_subklass() && !ik->is_interface()) {
2488 if (!ik->is_final()) {
2489 // Add a dependency if there is a chance of a later subclass.
2490 C->dependencies()->assert_leaf_type(ik);
2491 }
2492 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2493 }
2494 } else {
2495 // A primitive array type has no subtypes.
2496 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2497 }
2499 return SSC_full_test;
2500 }
2502 // Profile-driven exact type check:
2503 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2504 float prob,
2505 Node* *casted_receiver) {
2506 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2507 Node* recv_klass = load_object_klass(receiver);
2508 Node* want_klass = makecon(tklass);
2509 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2510 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2511 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2512 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2513 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2515 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2516 assert(recv_xtype->klass_is_exact(), "");
2518 // Subsume downstream occurrences of receiver with a cast to
2519 // recv_xtype, since now we know what the type will be.
2520 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2521 (*casted_receiver) = _gvn.transform(cast);
2522 // (User must make the replace_in_map call.)
2524 return fail;
2525 }
2528 //------------------------------seems_never_null-------------------------------
2529 // Use null_seen information if it is available from the profile.
2530 // If we see an unexpected null at a type check we record it and force a
2531 // recompile; the offending check will be recompiled to handle NULLs.
2532 // If we see several offending BCIs, then all checks in the
2533 // method will be recompiled.
2534 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2535 if (UncommonNullCast // Cutout for this technique
2536 && obj != null() // And not the -Xcomp stupid case?
2537 && !too_many_traps(Deoptimization::Reason_null_check)
2538 ) {
2539 if (data == NULL)
2540 // Edge case: no mature data. Be optimistic here.
2541 return true;
2542 // If the profile has not seen a null, assume it won't happen.
2543 assert(java_bc() == Bytecodes::_checkcast ||
2544 java_bc() == Bytecodes::_instanceof ||
2545 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2546 return !data->as_BitData()->null_seen();
2547 }
2548 return false;
2549 }
2551 //------------------------maybe_cast_profiled_receiver-------------------------
2552 // If the profile has seen exactly one type, narrow to exactly that type.
2553 // Subsequent type checks will always fold up.
2554 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2555 ciProfileData* data,
2556 ciKlass* require_klass) {
2557 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2558 if (data == NULL) return NULL;
2560 // Make sure we haven't already deoptimized from this tactic.
2561 if (too_many_traps(Deoptimization::Reason_class_check))
2562 return NULL;
2564 // (No, this isn't a call, but it's enough like a virtual call
2565 // to use the same ciMethod accessor to get the profile info...)
2566 ciCallProfile profile = method()->call_profile_at_bci(bci());
2567 if (profile.count() >= 0 && // no cast failures here
2568 profile.has_receiver(0) &&
2569 profile.morphism() == 1) {
2570 ciKlass* exact_kls = profile.receiver(0);
2571 if (require_klass == NULL ||
2572 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2573 // If we narrow the type to match what the type profile sees,
2574 // we can then remove the rest of the cast.
2575 // This is a win, even if the exact_kls is very specific,
2576 // because downstream operations, such as method calls,
2577 // will often benefit from the sharper type.
2578 Node* exact_obj = not_null_obj; // will get updated in place...
2579 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2580 &exact_obj);
2581 { PreserveJVMState pjvms(this);
2582 set_control(slow_ctl);
2583 uncommon_trap(Deoptimization::Reason_class_check,
2584 Deoptimization::Action_maybe_recompile);
2585 }
2586 replace_in_map(not_null_obj, exact_obj);
2587 return exact_obj;
2588 }
2589 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2590 }
2592 return NULL;
2593 }
2596 //-------------------------------gen_instanceof--------------------------------
2597 // Generate an instance-of idiom. Used by both the instance-of bytecode
2598 // and the reflective instance-of call.
2599 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass) {
2600 kill_dead_locals(); // Benefit all the uncommon traps
2601 assert( !stopped(), "dead parse path should be checked in callers" );
2602 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2603 "must check for not-null not-dead klass in callers");
2605 // Make the merge point
2606 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2607 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2608 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2609 C->set_has_split_ifs(true); // Has chance for split-if optimization
2611 ciProfileData* data = NULL;
2612 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2613 data = method()->method_data()->bci_to_data(bci());
2614 }
2615 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2616 && seems_never_null(obj, data));
2618 // Null check; get casted pointer; set region slot 3
2619 Node* null_ctl = top();
2620 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2622 // If not_null_obj is dead, only null-path is taken
2623 if (stopped()) { // Doing instance-of on a NULL?
2624 set_control(null_ctl);
2625 return intcon(0);
2626 }
2627 region->init_req(_null_path, null_ctl);
2628 phi ->init_req(_null_path, intcon(0)); // Set null path value
2629 if (null_ctl == top()) {
2630 // Do this eagerly, so that pattern matches like is_diamond_phi
2631 // will work even during parsing.
2632 assert(_null_path == PATH_LIMIT-1, "delete last");
2633 region->del_req(_null_path);
2634 phi ->del_req(_null_path);
2635 }
2637 if (ProfileDynamicTypes && data != NULL) {
2638 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, NULL);
2639 if (stopped()) { // Profile disagrees with this path.
2640 set_control(null_ctl); // Null is the only remaining possibility.
2641 return intcon(0);
2642 }
2643 if (cast_obj != NULL)
2644 not_null_obj = cast_obj;
2645 }
2647 // Load the object's klass
2648 Node* obj_klass = load_object_klass(not_null_obj);
2650 // Generate the subtype check
2651 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2653 // Plug in the success path to the general merge in slot 1.
2654 region->init_req(_obj_path, control());
2655 phi ->init_req(_obj_path, intcon(1));
2657 // Plug in the failing path to the general merge in slot 2.
2658 region->init_req(_fail_path, not_subtype_ctrl);
2659 phi ->init_req(_fail_path, intcon(0));
2661 // Return final merged results
2662 set_control( _gvn.transform(region) );
2663 record_for_igvn(region);
2664 return _gvn.transform(phi);
2665 }
2667 //-------------------------------gen_checkcast---------------------------------
2668 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2669 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2670 // uncommon-trap paths work. Adjust stack after this call.
2671 // If failure_control is supplied and not null, it is filled in with
2672 // the control edge for the cast failure. Otherwise, an appropriate
2673 // uncommon trap or exception is thrown.
2674 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2675 Node* *failure_control) {
2676 kill_dead_locals(); // Benefit all the uncommon traps
2677 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2678 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2680 // Fast cutout: Check the case that the cast is vacuously true.
2681 // This detects the common cases where the test will short-circuit
2682 // away completely. We do this before we perform the null check,
2683 // because if the test is going to turn into zero code, we don't
2684 // want a residual null check left around. (Causes a slowdown,
2685 // for example, in some objArray manipulations, such as a[i]=a[j].)
2686 if (tk->singleton()) {
2687 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2688 if (objtp != NULL && objtp->klass() != NULL) {
2689 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2690 case SSC_always_true:
2691 return obj;
2692 case SSC_always_false:
2693 // It needs a null check because a null will *pass* the cast check.
2694 // A non-null value will always produce an exception.
2695 return null_assert(obj);
2696 }
2697 }
2698 }
2700 ciProfileData* data = NULL;
2701 if (failure_control == NULL) { // use MDO in regular case only
2702 assert(java_bc() == Bytecodes::_aastore ||
2703 java_bc() == Bytecodes::_checkcast,
2704 "interpreter profiles type checks only for these BCs");
2705 data = method()->method_data()->bci_to_data(bci());
2706 }
2708 // Make the merge point
2709 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2710 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
2711 Node* phi = new (C) PhiNode(region, toop);
2712 C->set_has_split_ifs(true); // Has chance for split-if optimization
2714 // Use null-cast information if it is available
2715 bool never_see_null = ((failure_control == NULL) // regular case only
2716 && seems_never_null(obj, data));
2718 // Null check; get casted pointer; set region slot 3
2719 Node* null_ctl = top();
2720 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2722 // If not_null_obj is dead, only null-path is taken
2723 if (stopped()) { // Doing instance-of on a NULL?
2724 set_control(null_ctl);
2725 return null();
2726 }
2727 region->init_req(_null_path, null_ctl);
2728 phi ->init_req(_null_path, null()); // Set null path value
2729 if (null_ctl == top()) {
2730 // Do this eagerly, so that pattern matches like is_diamond_phi
2731 // will work even during parsing.
2732 assert(_null_path == PATH_LIMIT-1, "delete last");
2733 region->del_req(_null_path);
2734 phi ->del_req(_null_path);
2735 }
2737 Node* cast_obj = NULL;
2738 if (data != NULL &&
2739 // Counter has never been decremented (due to cast failure).
2740 // ...This is a reasonable thing to expect. It is true of
2741 // all casts inserted by javac to implement generic types.
2742 data->as_CounterData()->count() >= 0) {
2743 cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, tk->klass());
2744 if (cast_obj != NULL) {
2745 if (failure_control != NULL) // failure is now impossible
2746 (*failure_control) = top();
2747 // adjust the type of the phi to the exact klass:
2748 phi->raise_bottom_type(_gvn.type(cast_obj)->meet(TypePtr::NULL_PTR));
2749 }
2750 }
2752 if (cast_obj == NULL) {
2753 // Load the object's klass
2754 Node* obj_klass = load_object_klass(not_null_obj);
2756 // Generate the subtype check
2757 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2759 // Plug in success path into the merge
2760 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
2761 not_null_obj, toop));
2762 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2763 if (failure_control == NULL) {
2764 if (not_subtype_ctrl != top()) { // If failure is possible
2765 PreserveJVMState pjvms(this);
2766 set_control(not_subtype_ctrl);
2767 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2768 }
2769 } else {
2770 (*failure_control) = not_subtype_ctrl;
2771 }
2772 }
2774 region->init_req(_obj_path, control());
2775 phi ->init_req(_obj_path, cast_obj);
2777 // A merge of NULL or Casted-NotNull obj
2778 Node* res = _gvn.transform(phi);
2780 // Note I do NOT always 'replace_in_map(obj,result)' here.
2781 // if( tk->klass()->can_be_primary_super() )
2782 // This means that if I successfully store an Object into an array-of-String
2783 // I 'forget' that the Object is really now known to be a String. I have to
2784 // do this because we don't have true union types for interfaces - if I store
2785 // a Baz into an array-of-Interface and then tell the optimizer it's an
2786 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2787 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2788 // replace_in_map( obj, res );
2790 // Return final merged results
2791 set_control( _gvn.transform(region) );
2792 record_for_igvn(region);
2793 return res;
2794 }
2796 //------------------------------next_monitor-----------------------------------
2797 // What number should be given to the next monitor?
2798 int GraphKit::next_monitor() {
2799 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2800 int next = current + C->sync_stack_slots();
2801 // Keep the toplevel high water mark current:
2802 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2803 return current;
2804 }
2806 //------------------------------insert_mem_bar---------------------------------
2807 // Memory barrier to avoid floating things around
2808 // The membar serves as a pinch point between both control and all memory slices.
2809 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2810 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2811 mb->init_req(TypeFunc::Control, control());
2812 mb->init_req(TypeFunc::Memory, reset_memory());
2813 Node* membar = _gvn.transform(mb);
2814 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
2815 set_all_memory_call(membar);
2816 return membar;
2817 }
2819 //-------------------------insert_mem_bar_volatile----------------------------
2820 // Memory barrier to avoid floating things around
2821 // The membar serves as a pinch point between both control and memory(alias_idx).
2822 // If you want to make a pinch point on all memory slices, do not use this
2823 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2824 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2825 // When Parse::do_put_xxx updates a volatile field, it appends a series
2826 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2827 // The first membar is on the same memory slice as the field store opcode.
2828 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2829 // All the other membars (for other volatile slices, including AliasIdxBot,
2830 // which stands for all unknown volatile slices) are control-dependent
2831 // on the first membar. This prevents later volatile loads or stores
2832 // from sliding up past the just-emitted store.
2834 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2835 mb->set_req(TypeFunc::Control,control());
2836 if (alias_idx == Compile::AliasIdxBot) {
2837 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2838 } else {
2839 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2840 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2841 }
2842 Node* membar = _gvn.transform(mb);
2843 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
2844 if (alias_idx == Compile::AliasIdxBot) {
2845 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
2846 } else {
2847 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2848 }
2849 return membar;
2850 }
2852 //------------------------------shared_lock------------------------------------
2853 // Emit locking code.
2854 FastLockNode* GraphKit::shared_lock(Node* obj) {
2855 // bci is either a monitorenter bc or InvocationEntryBci
2856 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2857 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2859 if( !GenerateSynchronizationCode )
2860 return NULL; // Not locking things?
2861 if (stopped()) // Dead monitor?
2862 return NULL;
2864 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2866 // Box the stack location
2867 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
2868 Node* mem = reset_memory();
2870 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
2871 if (PrintPreciseBiasedLockingStatistics) {
2872 // Create the counters for this fast lock.
2873 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2874 }
2875 // Add monitor to debug info for the slow path. If we block inside the
2876 // slow path and de-opt, we need the monitor hanging around
2877 map()->push_monitor( flock );
2879 const TypeFunc *tf = LockNode::lock_type();
2880 LockNode *lock = new (C) LockNode(C, tf);
2882 lock->init_req( TypeFunc::Control, control() );
2883 lock->init_req( TypeFunc::Memory , mem );
2884 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2885 lock->init_req( TypeFunc::FramePtr, frameptr() );
2886 lock->init_req( TypeFunc::ReturnAdr, top() );
2888 lock->init_req(TypeFunc::Parms + 0, obj);
2889 lock->init_req(TypeFunc::Parms + 1, box);
2890 lock->init_req(TypeFunc::Parms + 2, flock);
2891 add_safepoint_edges(lock);
2893 lock = _gvn.transform( lock )->as_Lock();
2895 // lock has no side-effects, sets few values
2896 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2898 insert_mem_bar(Op_MemBarAcquireLock);
2900 // Add this to the worklist so that the lock can be eliminated
2901 record_for_igvn(lock);
2903 #ifndef PRODUCT
2904 if (PrintLockStatistics) {
2905 // Update the counter for this lock. Don't bother using an atomic
2906 // operation since we don't require absolute accuracy.
2907 lock->create_lock_counter(map()->jvms());
2908 increment_counter(lock->counter()->addr());
2909 }
2910 #endif
2912 return flock;
2913 }
2916 //------------------------------shared_unlock----------------------------------
2917 // Emit unlocking code.
2918 void GraphKit::shared_unlock(Node* box, Node* obj) {
2919 // bci is either a monitorenter bc or InvocationEntryBci
2920 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2921 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2923 if( !GenerateSynchronizationCode )
2924 return;
2925 if (stopped()) { // Dead monitor?
2926 map()->pop_monitor(); // Kill monitor from debug info
2927 return;
2928 }
2930 // Memory barrier to avoid floating things down past the locked region
2931 insert_mem_bar(Op_MemBarReleaseLock);
2933 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2934 UnlockNode *unlock = new (C) UnlockNode(C, tf);
2935 uint raw_idx = Compile::AliasIdxRaw;
2936 unlock->init_req( TypeFunc::Control, control() );
2937 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2938 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2939 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2940 unlock->init_req( TypeFunc::ReturnAdr, top() );
2942 unlock->init_req(TypeFunc::Parms + 0, obj);
2943 unlock->init_req(TypeFunc::Parms + 1, box);
2944 unlock = _gvn.transform(unlock)->as_Unlock();
2946 Node* mem = reset_memory();
2948 // unlock has no side-effects, sets few values
2949 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2951 // Kill monitor from debug info
2952 map()->pop_monitor( );
2953 }
2955 //-------------------------------get_layout_helper-----------------------------
2956 // If the given klass is a constant or known to be an array,
2957 // fetch the constant layout helper value into constant_value
2958 // and return (Node*)NULL. Otherwise, load the non-constant
2959 // layout helper value, and return the node which represents it.
2960 // This two-faced routine is useful because allocation sites
2961 // almost always feature constant types.
2962 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2963 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2964 if (!StressReflectiveCode && inst_klass != NULL) {
2965 ciKlass* klass = inst_klass->klass();
2966 bool xklass = inst_klass->klass_is_exact();
2967 if (xklass || klass->is_array_klass()) {
2968 jint lhelper = klass->layout_helper();
2969 if (lhelper != Klass::_lh_neutral_value) {
2970 constant_value = lhelper;
2971 return (Node*) NULL;
2972 }
2973 }
2974 }
2975 constant_value = Klass::_lh_neutral_value; // put in a known value
2976 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
2977 return make_load(NULL, lhp, TypeInt::INT, T_INT);
2978 }
2980 // We just put in an allocate/initialize with a big raw-memory effect.
2981 // Hook selected additional alias categories on the initialization.
2982 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
2983 MergeMemNode* init_in_merge,
2984 Node* init_out_raw) {
2985 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
2986 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
2988 Node* prevmem = kit.memory(alias_idx);
2989 init_in_merge->set_memory_at(alias_idx, prevmem);
2990 kit.set_memory(init_out_raw, alias_idx);
2991 }
2993 //---------------------------set_output_for_allocation-------------------------
2994 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
2995 const TypeOopPtr* oop_type) {
2996 int rawidx = Compile::AliasIdxRaw;
2997 alloc->set_req( TypeFunc::FramePtr, frameptr() );
2998 add_safepoint_edges(alloc);
2999 Node* allocx = _gvn.transform(alloc);
3000 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3001 // create memory projection for i_o
3002 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3003 make_slow_call_ex(allocx, env()->Throwable_klass(), true);
3005 // create a memory projection as for the normal control path
3006 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3007 set_memory(malloc, rawidx);
3009 // a normal slow-call doesn't change i_o, but an allocation does
3010 // we create a separate i_o projection for the normal control path
3011 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3012 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3014 // put in an initialization barrier
3015 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3016 rawoop)->as_Initialize();
3017 assert(alloc->initialization() == init, "2-way macro link must work");
3018 assert(init ->allocation() == alloc, "2-way macro link must work");
3019 {
3020 // Extract memory strands which may participate in the new object's
3021 // initialization, and source them from the new InitializeNode.
3022 // This will allow us to observe initializations when they occur,
3023 // and link them properly (as a group) to the InitializeNode.
3024 assert(init->in(InitializeNode::Memory) == malloc, "");
3025 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3026 init->set_req(InitializeNode::Memory, minit_in);
3027 record_for_igvn(minit_in); // fold it up later, if possible
3028 Node* minit_out = memory(rawidx);
3029 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3030 if (oop_type->isa_aryptr()) {
3031 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3032 int elemidx = C->get_alias_index(telemref);
3033 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3034 } else if (oop_type->isa_instptr()) {
3035 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3036 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3037 ciField* field = ik->nonstatic_field_at(i);
3038 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3039 continue; // do not bother to track really large numbers of fields
3040 // Find (or create) the alias category for this field:
3041 int fieldidx = C->alias_type(field)->index();
3042 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3043 }
3044 }
3045 }
3047 // Cast raw oop to the real thing...
3048 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3049 javaoop = _gvn.transform(javaoop);
3050 C->set_recent_alloc(control(), javaoop);
3051 assert(just_allocated_object(control()) == javaoop, "just allocated");
3053 #ifdef ASSERT
3054 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3055 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3056 "Ideal_allocation works");
3057 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3058 "Ideal_allocation works");
3059 if (alloc->is_AllocateArray()) {
3060 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3061 "Ideal_allocation works");
3062 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3063 "Ideal_allocation works");
3064 } else {
3065 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3066 }
3067 }
3068 #endif //ASSERT
3070 return javaoop;
3071 }
3073 //---------------------------new_instance--------------------------------------
3074 // This routine takes a klass_node which may be constant (for a static type)
3075 // or may be non-constant (for reflective code). It will work equally well
3076 // for either, and the graph will fold nicely if the optimizer later reduces
3077 // the type to a constant.
3078 // The optional arguments are for specialized use by intrinsics:
3079 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3080 // - If 'return_size_val', report the the total object size to the caller.
3081 Node* GraphKit::new_instance(Node* klass_node,
3082 Node* extra_slow_test,
3083 Node* *return_size_val) {
3084 // Compute size in doublewords
3085 // The size is always an integral number of doublewords, represented
3086 // as a positive bytewise size stored in the klass's layout_helper.
3087 // The layout_helper also encodes (in a low bit) the need for a slow path.
3088 jint layout_con = Klass::_lh_neutral_value;
3089 Node* layout_val = get_layout_helper(klass_node, layout_con);
3090 int layout_is_con = (layout_val == NULL);
3092 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3093 // Generate the initial go-slow test. It's either ALWAYS (return a
3094 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3095 // case) a computed value derived from the layout_helper.
3096 Node* initial_slow_test = NULL;
3097 if (layout_is_con) {
3098 assert(!StressReflectiveCode, "stress mode does not use these paths");
3099 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3100 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3102 } else { // reflective case
3103 // This reflective path is used by Unsafe.allocateInstance.
3104 // (It may be stress-tested by specifying StressReflectiveCode.)
3105 // Basically, we want to get into the VM is there's an illegal argument.
3106 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3107 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3108 if (extra_slow_test != intcon(0)) {
3109 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3110 }
3111 // (Macro-expander will further convert this to a Bool, if necessary.)
3112 }
3114 // Find the size in bytes. This is easy; it's the layout_helper.
3115 // The size value must be valid even if the slow path is taken.
3116 Node* size = NULL;
3117 if (layout_is_con) {
3118 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3119 } else { // reflective case
3120 // This reflective path is used by clone and Unsafe.allocateInstance.
3121 size = ConvI2X(layout_val);
3123 // Clear the low bits to extract layout_helper_size_in_bytes:
3124 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3125 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3126 size = _gvn.transform( new (C) AndXNode(size, mask) );
3127 }
3128 if (return_size_val != NULL) {
3129 (*return_size_val) = size;
3130 }
3132 // This is a precise notnull oop of the klass.
3133 // (Actually, it need not be precise if this is a reflective allocation.)
3134 // It's what we cast the result to.
3135 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3136 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3137 const TypeOopPtr* oop_type = tklass->as_instance_type();
3139 // Now generate allocation code
3141 // The entire memory state is needed for slow path of the allocation
3142 // since GC and deoptimization can happened.
3143 Node *mem = reset_memory();
3144 set_all_memory(mem); // Create new memory state
3146 AllocateNode* alloc
3147 = new (C) AllocateNode(C, AllocateNode::alloc_type(),
3148 control(), mem, i_o(),
3149 size, klass_node,
3150 initial_slow_test);
3152 return set_output_for_allocation(alloc, oop_type);
3153 }
3155 //-------------------------------new_array-------------------------------------
3156 // helper for both newarray and anewarray
3157 // The 'length' parameter is (obviously) the length of the array.
3158 // See comments on new_instance for the meaning of the other arguments.
3159 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3160 Node* length, // number of array elements
3161 int nargs, // number of arguments to push back for uncommon trap
3162 Node* *return_size_val) {
3163 jint layout_con = Klass::_lh_neutral_value;
3164 Node* layout_val = get_layout_helper(klass_node, layout_con);
3165 int layout_is_con = (layout_val == NULL);
3167 if (!layout_is_con && !StressReflectiveCode &&
3168 !too_many_traps(Deoptimization::Reason_class_check)) {
3169 // This is a reflective array creation site.
3170 // Optimistically assume that it is a subtype of Object[],
3171 // so that we can fold up all the address arithmetic.
3172 layout_con = Klass::array_layout_helper(T_OBJECT);
3173 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3174 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3175 { BuildCutout unless(this, bol_lh, PROB_MAX);
3176 inc_sp(nargs);
3177 uncommon_trap(Deoptimization::Reason_class_check,
3178 Deoptimization::Action_maybe_recompile);
3179 }
3180 layout_val = NULL;
3181 layout_is_con = true;
3182 }
3184 // Generate the initial go-slow test. Make sure we do not overflow
3185 // if length is huge (near 2Gig) or negative! We do not need
3186 // exact double-words here, just a close approximation of needed
3187 // double-words. We can't add any offset or rounding bits, lest we
3188 // take a size -1 of bytes and make it positive. Use an unsigned
3189 // compare, so negative sizes look hugely positive.
3190 int fast_size_limit = FastAllocateSizeLimit;
3191 if (layout_is_con) {
3192 assert(!StressReflectiveCode, "stress mode does not use these paths");
3193 // Increase the size limit if we have exact knowledge of array type.
3194 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3195 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3196 }
3198 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3199 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3200 if (initial_slow_test->is_Bool()) {
3201 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3202 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3203 }
3205 // --- Size Computation ---
3206 // array_size = round_to_heap(array_header + (length << elem_shift));
3207 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3208 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3209 // The rounding mask is strength-reduced, if possible.
3210 int round_mask = MinObjAlignmentInBytes - 1;
3211 Node* header_size = NULL;
3212 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3213 // (T_BYTE has the weakest alignment and size restrictions...)
3214 if (layout_is_con) {
3215 int hsize = Klass::layout_helper_header_size(layout_con);
3216 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3217 BasicType etype = Klass::layout_helper_element_type(layout_con);
3218 if ((round_mask & ~right_n_bits(eshift)) == 0)
3219 round_mask = 0; // strength-reduce it if it goes away completely
3220 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3221 assert(header_size_min <= hsize, "generic minimum is smallest");
3222 header_size_min = hsize;
3223 header_size = intcon(hsize + round_mask);
3224 } else {
3225 Node* hss = intcon(Klass::_lh_header_size_shift);
3226 Node* hsm = intcon(Klass::_lh_header_size_mask);
3227 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3228 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3229 Node* mask = intcon(round_mask);
3230 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3231 }
3233 Node* elem_shift = NULL;
3234 if (layout_is_con) {
3235 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3236 if (eshift != 0)
3237 elem_shift = intcon(eshift);
3238 } else {
3239 // There is no need to mask or shift this value.
3240 // The semantics of LShiftINode include an implicit mask to 0x1F.
3241 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3242 elem_shift = layout_val;
3243 }
3245 // Transition to native address size for all offset calculations:
3246 Node* lengthx = ConvI2X(length);
3247 Node* headerx = ConvI2X(header_size);
3248 #ifdef _LP64
3249 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3250 if (tllen != NULL && tllen->_lo < 0) {
3251 // Add a manual constraint to a positive range. Cf. array_element_address.
3252 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3253 if (size_max > tllen->_hi) size_max = tllen->_hi;
3254 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3255 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon));
3256 }
3257 }
3258 #endif
3260 // Combine header size (plus rounding) and body size. Then round down.
3261 // This computation cannot overflow, because it is used only in two
3262 // places, one where the length is sharply limited, and the other
3263 // after a successful allocation.
3264 Node* abody = lengthx;
3265 if (elem_shift != NULL)
3266 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3267 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3268 if (round_mask != 0) {
3269 Node* mask = MakeConX(~round_mask);
3270 size = _gvn.transform( new(C) AndXNode(size, mask) );
3271 }
3272 // else if round_mask == 0, the size computation is self-rounding
3274 if (return_size_val != NULL) {
3275 // This is the size
3276 (*return_size_val) = size;
3277 }
3279 // Now generate allocation code
3281 // The entire memory state is needed for slow path of the allocation
3282 // since GC and deoptimization can happened.
3283 Node *mem = reset_memory();
3284 set_all_memory(mem); // Create new memory state
3286 // Create the AllocateArrayNode and its result projections
3287 AllocateArrayNode* alloc
3288 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3289 control(), mem, i_o(),
3290 size, klass_node,
3291 initial_slow_test,
3292 length);
3294 // Cast to correct type. Note that the klass_node may be constant or not,
3295 // and in the latter case the actual array type will be inexact also.
3296 // (This happens via a non-constant argument to inline_native_newArray.)
3297 // In any case, the value of klass_node provides the desired array type.
3298 const TypeInt* length_type = _gvn.find_int_type(length);
3299 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3300 if (ary_type->isa_aryptr() && length_type != NULL) {
3301 // Try to get a better type than POS for the size
3302 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3303 }
3305 Node* javaoop = set_output_for_allocation(alloc, ary_type);
3307 // Cast length on remaining path to be as narrow as possible
3308 if (map()->find_edge(length) >= 0) {
3309 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3310 if (ccast != length) {
3311 _gvn.set_type_bottom(ccast);
3312 record_for_igvn(ccast);
3313 replace_in_map(length, ccast);
3314 }
3315 }
3317 return javaoop;
3318 }
3320 // The following "Ideal_foo" functions are placed here because they recognize
3321 // the graph shapes created by the functions immediately above.
3323 //---------------------------Ideal_allocation----------------------------------
3324 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3325 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3326 if (ptr == NULL) { // reduce dumb test in callers
3327 return NULL;
3328 }
3329 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
3330 ptr = ptr->in(1);
3331 if (ptr == NULL) return NULL;
3332 }
3333 if (ptr->is_Proj()) {
3334 Node* allo = ptr->in(0);
3335 if (allo != NULL && allo->is_Allocate()) {
3336 return allo->as_Allocate();
3337 }
3338 }
3339 // Report failure to match.
3340 return NULL;
3341 }
3343 // Fancy version which also strips off an offset (and reports it to caller).
3344 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3345 intptr_t& offset) {
3346 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3347 if (base == NULL) return NULL;
3348 return Ideal_allocation(base, phase);
3349 }
3351 // Trace Initialize <- Proj[Parm] <- Allocate
3352 AllocateNode* InitializeNode::allocation() {
3353 Node* rawoop = in(InitializeNode::RawAddress);
3354 if (rawoop->is_Proj()) {
3355 Node* alloc = rawoop->in(0);
3356 if (alloc->is_Allocate()) {
3357 return alloc->as_Allocate();
3358 }
3359 }
3360 return NULL;
3361 }
3363 // Trace Allocate -> Proj[Parm] -> Initialize
3364 InitializeNode* AllocateNode::initialization() {
3365 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3366 if (rawoop == NULL) return NULL;
3367 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3368 Node* init = rawoop->fast_out(i);
3369 if (init->is_Initialize()) {
3370 assert(init->as_Initialize()->allocation() == this, "2-way link");
3371 return init->as_Initialize();
3372 }
3373 }
3374 return NULL;
3375 }
3377 // Trace Allocate -> Proj[Parm] -> MemBarStoreStore
3378 MemBarStoreStoreNode* AllocateNode::storestore() {
3379 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3380 if (rawoop == NULL) return NULL;
3381 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3382 Node* storestore = rawoop->fast_out(i);
3383 if (storestore->is_MemBarStoreStore()) {
3384 return storestore->as_MemBarStoreStore();
3385 }
3386 }
3387 return NULL;
3388 }
3390 //----------------------------- loop predicates ---------------------------
3392 //------------------------------add_predicate_impl----------------------------
3393 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3394 // Too many traps seen?
3395 if (too_many_traps(reason)) {
3396 #ifdef ASSERT
3397 if (TraceLoopPredicate) {
3398 int tc = C->trap_count(reason);
3399 tty->print("too many traps=%s tcount=%d in ",
3400 Deoptimization::trap_reason_name(reason), tc);
3401 method()->print(); // which method has too many predicate traps
3402 tty->cr();
3403 }
3404 #endif
3405 // We cannot afford to take more traps here,
3406 // do not generate predicate.
3407 return;
3408 }
3410 Node *cont = _gvn.intcon(1);
3411 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3412 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3413 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3414 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3415 C->add_predicate_opaq(opq);
3416 {
3417 PreserveJVMState pjvms(this);
3418 set_control(iffalse);
3419 inc_sp(nargs);
3420 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3421 }
3422 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3423 set_control(iftrue);
3424 }
3426 //------------------------------add_predicate---------------------------------
3427 void GraphKit::add_predicate(int nargs) {
3428 if (UseLoopPredicate) {
3429 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3430 }
3431 // loop's limit check predicate should be near the loop.
3432 if (LoopLimitCheck) {
3433 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3434 }
3435 }
3437 //----------------------------- store barriers ----------------------------
3438 #define __ ideal.
3440 void GraphKit::sync_kit(IdealKit& ideal) {
3441 set_all_memory(__ merged_memory());
3442 set_i_o(__ i_o());
3443 set_control(__ ctrl());
3444 }
3446 void GraphKit::final_sync(IdealKit& ideal) {
3447 // Final sync IdealKit and graphKit.
3448 sync_kit(ideal);
3449 }
3451 // vanilla/CMS post barrier
3452 // Insert a write-barrier store. This is to let generational GC work; we have
3453 // to flag all oop-stores before the next GC point.
3454 void GraphKit::write_barrier_post(Node* oop_store,
3455 Node* obj,
3456 Node* adr,
3457 uint adr_idx,
3458 Node* val,
3459 bool use_precise) {
3460 // No store check needed if we're storing a NULL or an old object
3461 // (latter case is probably a string constant). The concurrent
3462 // mark sweep garbage collector, however, needs to have all nonNull
3463 // oop updates flagged via card-marks.
3464 if (val != NULL && val->is_Con()) {
3465 // must be either an oop or NULL
3466 const Type* t = val->bottom_type();
3467 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3468 // stores of null never (?) need barriers
3469 return;
3470 }
3472 if (use_ReduceInitialCardMarks()
3473 && obj == just_allocated_object(control())) {
3474 // We can skip marks on a freshly-allocated object in Eden.
3475 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3476 // That routine informs GC to take appropriate compensating steps,
3477 // upon a slow-path allocation, so as to make this card-mark
3478 // elision safe.
3479 return;
3480 }
3482 if (!use_precise) {
3483 // All card marks for a (non-array) instance are in one place:
3484 adr = obj;
3485 }
3486 // (Else it's an array (or unknown), and we want more precise card marks.)
3487 assert(adr != NULL, "");
3489 IdealKit ideal(this, true);
3491 // Convert the pointer to an int prior to doing math on it
3492 Node* cast = __ CastPX(__ ctrl(), adr);
3494 // Divide by card size
3495 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3496 "Only one we handle so far.");
3497 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3499 // Combine card table base and card offset
3500 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3502 // Get the alias_index for raw card-mark memory
3503 int adr_type = Compile::AliasIdxRaw;
3504 Node* zero = __ ConI(0); // Dirty card value
3505 BasicType bt = T_BYTE;
3507 if (UseCondCardMark) {
3508 // The classic GC reference write barrier is typically implemented
3509 // as a store into the global card mark table. Unfortunately
3510 // unconditional stores can result in false sharing and excessive
3511 // coherence traffic as well as false transactional aborts.
3512 // UseCondCardMark enables MP "polite" conditional card mark
3513 // stores. In theory we could relax the load from ctrl() to
3514 // no_ctrl, but that doesn't buy much latitude.
3515 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3516 __ if_then(card_val, BoolTest::ne, zero);
3517 }
3519 // Smash zero into card
3520 if( !UseConcMarkSweepGC ) {
3521 __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3522 } else {
3523 // Specialized path for CM store barrier
3524 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3525 }
3527 if (UseCondCardMark) {
3528 __ end_if();
3529 }
3531 // Final sync IdealKit and GraphKit.
3532 final_sync(ideal);
3533 }
3535 // G1 pre/post barriers
3536 void GraphKit::g1_write_barrier_pre(bool do_load,
3537 Node* obj,
3538 Node* adr,
3539 uint alias_idx,
3540 Node* val,
3541 const TypeOopPtr* val_type,
3542 Node* pre_val,
3543 BasicType bt) {
3545 // Some sanity checks
3546 // Note: val is unused in this routine.
3548 if (do_load) {
3549 // We need to generate the load of the previous value
3550 assert(obj != NULL, "must have a base");
3551 assert(adr != NULL, "where are loading from?");
3552 assert(pre_val == NULL, "loaded already?");
3553 assert(val_type != NULL, "need a type");
3554 } else {
3555 // In this case both val_type and alias_idx are unused.
3556 assert(pre_val != NULL, "must be loaded already");
3557 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3558 }
3559 assert(bt == T_OBJECT, "or we shouldn't be here");
3561 IdealKit ideal(this, true);
3563 Node* tls = __ thread(); // ThreadLocalStorage
3565 Node* no_ctrl = NULL;
3566 Node* no_base = __ top();
3567 Node* zero = __ ConI(0);
3568 Node* zeroX = __ ConX(0);
3570 float likely = PROB_LIKELY(0.999);
3571 float unlikely = PROB_UNLIKELY(0.999);
3573 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3574 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3576 // Offsets into the thread
3577 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3578 PtrQueue::byte_offset_of_active());
3579 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3580 PtrQueue::byte_offset_of_index());
3581 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3582 PtrQueue::byte_offset_of_buf());
3584 // Now the actual pointers into the thread
3585 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3586 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3587 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3589 // Now some of the values
3590 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3592 // if (!marking)
3593 __ if_then(marking, BoolTest::ne, zero); {
3594 BasicType index_bt = TypeX_X->basic_type();
3595 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3596 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3598 if (do_load) {
3599 // load original value
3600 // alias_idx correct??
3601 pre_val = __ load(no_ctrl, adr, val_type, bt, alias_idx);
3602 }
3604 // if (pre_val != NULL)
3605 __ if_then(pre_val, BoolTest::ne, null()); {
3606 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3608 // is the queue for this thread full?
3609 __ if_then(index, BoolTest::ne, zeroX, likely); {
3611 // decrement the index
3612 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3614 // Now get the buffer location we will log the previous value into and store it
3615 Node *log_addr = __ AddP(no_base, buffer, next_index);
3616 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw);
3617 // update the index
3618 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw);
3620 } __ else_(); {
3622 // logging buffer is full, call the runtime
3623 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3624 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3625 } __ end_if(); // (!index)
3626 } __ end_if(); // (pre_val != NULL)
3627 } __ end_if(); // (!marking)
3629 // Final sync IdealKit and GraphKit.
3630 final_sync(ideal);
3631 }
3633 //
3634 // Update the card table and add card address to the queue
3635 //
3636 void GraphKit::g1_mark_card(IdealKit& ideal,
3637 Node* card_adr,
3638 Node* oop_store,
3639 uint oop_alias_idx,
3640 Node* index,
3641 Node* index_adr,
3642 Node* buffer,
3643 const TypeFunc* tf) {
3645 Node* zero = __ ConI(0);
3646 Node* zeroX = __ ConX(0);
3647 Node* no_base = __ top();
3648 BasicType card_bt = T_BYTE;
3649 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3650 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3652 // Now do the queue work
3653 __ if_then(index, BoolTest::ne, zeroX); {
3655 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3656 Node* log_addr = __ AddP(no_base, buffer, next_index);
3658 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3659 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw);
3661 } __ else_(); {
3662 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3663 } __ end_if();
3665 }
3667 void GraphKit::g1_write_barrier_post(Node* oop_store,
3668 Node* obj,
3669 Node* adr,
3670 uint alias_idx,
3671 Node* val,
3672 BasicType bt,
3673 bool use_precise) {
3674 // If we are writing a NULL then we need no post barrier
3676 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3677 // Must be NULL
3678 const Type* t = val->bottom_type();
3679 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3680 // No post barrier if writing NULLx
3681 return;
3682 }
3684 if (!use_precise) {
3685 // All card marks for a (non-array) instance are in one place:
3686 adr = obj;
3687 }
3688 // (Else it's an array (or unknown), and we want more precise card marks.)
3689 assert(adr != NULL, "");
3691 IdealKit ideal(this, true);
3693 Node* tls = __ thread(); // ThreadLocalStorage
3695 Node* no_base = __ top();
3696 float likely = PROB_LIKELY(0.999);
3697 float unlikely = PROB_UNLIKELY(0.999);
3698 Node* zero = __ ConI(0);
3699 Node* zeroX = __ ConX(0);
3701 // Get the alias_index for raw card-mark memory
3702 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3704 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3706 // Offsets into the thread
3707 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3708 PtrQueue::byte_offset_of_index());
3709 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3710 PtrQueue::byte_offset_of_buf());
3712 // Pointers into the thread
3714 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3715 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3717 // Now some values
3718 // Use ctrl to avoid hoisting these values past a safepoint, which could
3719 // potentially reset these fields in the JavaThread.
3720 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
3721 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3723 // Convert the store obj pointer to an int prior to doing math on it
3724 // Must use ctrl to prevent "integerized oop" existing across safepoint
3725 Node* cast = __ CastPX(__ ctrl(), adr);
3727 // Divide pointer by card size
3728 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3730 // Combine card table base and card offset
3731 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3733 // If we know the value being stored does it cross regions?
3735 if (val != NULL) {
3736 // Does the store cause us to cross regions?
3738 // Should be able to do an unsigned compare of region_size instead of
3739 // and extra shift. Do we have an unsigned compare??
3740 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3741 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3743 // if (xor_res == 0) same region so skip
3744 __ if_then(xor_res, BoolTest::ne, zeroX); {
3746 // No barrier if we are storing a NULL
3747 __ if_then(val, BoolTest::ne, null(), unlikely); {
3749 // Ok must mark the card if not already dirty
3751 // load the original value of the card
3752 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3754 __ if_then(card_val, BoolTest::ne, zero); {
3755 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3756 } __ end_if();
3757 } __ end_if();
3758 } __ end_if();
3759 } else {
3760 // Object.clone() instrinsic uses this path.
3761 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3762 }
3764 // Final sync IdealKit and GraphKit.
3765 final_sync(ideal);
3766 }
3767 #undef __
3771 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
3772 if (java_lang_String::has_offset_field()) {
3773 int offset_offset = java_lang_String::offset_offset_in_bytes();
3774 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3775 false, NULL, 0);
3776 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
3777 int offset_field_idx = C->get_alias_index(offset_field_type);
3778 return make_load(ctrl,
3779 basic_plus_adr(str, str, offset_offset),
3780 TypeInt::INT, T_INT, offset_field_idx);
3781 } else {
3782 return intcon(0);
3783 }
3784 }
3786 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
3787 if (java_lang_String::has_count_field()) {
3788 int count_offset = java_lang_String::count_offset_in_bytes();
3789 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3790 false, NULL, 0);
3791 const TypePtr* count_field_type = string_type->add_offset(count_offset);
3792 int count_field_idx = C->get_alias_index(count_field_type);
3793 return make_load(ctrl,
3794 basic_plus_adr(str, str, count_offset),
3795 TypeInt::INT, T_INT, count_field_idx);
3796 } else {
3797 return load_array_length(load_String_value(ctrl, str));
3798 }
3799 }
3801 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
3802 int value_offset = java_lang_String::value_offset_in_bytes();
3803 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3804 false, NULL, 0);
3805 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3806 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3807 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
3808 ciTypeArrayKlass::make(T_CHAR), true, 0);
3809 int value_field_idx = C->get_alias_index(value_field_type);
3810 return make_load(ctrl, basic_plus_adr(str, str, value_offset),
3811 value_type, T_OBJECT, value_field_idx);
3812 }
3814 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
3815 int offset_offset = java_lang_String::offset_offset_in_bytes();
3816 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3817 false, NULL, 0);
3818 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
3819 int offset_field_idx = C->get_alias_index(offset_field_type);
3820 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
3821 value, T_INT, offset_field_idx);
3822 }
3824 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
3825 int value_offset = java_lang_String::value_offset_in_bytes();
3826 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3827 false, NULL, 0);
3828 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3829 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3830 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
3831 ciTypeArrayKlass::make(T_CHAR), true, 0);
3832 int value_field_idx = C->get_alias_index(value_field_type);
3833 store_to_memory(ctrl, basic_plus_adr(str, value_offset),
3834 value, T_OBJECT, value_field_idx);
3835 }
3837 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
3838 int count_offset = java_lang_String::count_offset_in_bytes();
3839 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3840 false, NULL, 0);
3841 const TypePtr* count_field_type = string_type->add_offset(count_offset);
3842 int count_field_idx = C->get_alias_index(count_field_type);
3843 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
3844 value, T_INT, count_field_idx);
3845 }