Thu, 29 Jan 2015 10:26:02 -0800
8068915: uncommon trap w/ Reason_speculate_class_check causes performance regression due to continuous deoptimizations
Reviewed-by: kvn, roland, jrose
1 /*
2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
28 #include "gc_implementation/g1/heapRegion.hpp"
29 #include "gc_interface/collectedHeap.hpp"
30 #include "memory/barrierSet.hpp"
31 #include "memory/cardTableModRefBS.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/graphKit.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "runtime/deoptimization.hpp"
41 #include "runtime/sharedRuntime.hpp"
43 //----------------------------GraphKit-----------------------------------------
44 // Main utility constructor.
45 GraphKit::GraphKit(JVMState* jvms)
46 : Phase(Phase::Parser),
47 _env(C->env()),
48 _gvn(*C->initial_gvn())
49 {
50 _exceptions = jvms->map()->next_exception();
51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
52 set_jvms(jvms);
53 }
55 // Private constructor for parser.
56 GraphKit::GraphKit()
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn())
60 {
61 _exceptions = NULL;
62 set_map(NULL);
63 debug_only(_sp = -99);
64 debug_only(set_bci(-99));
65 }
69 //---------------------------clean_stack---------------------------------------
70 // Clear away rubbish from the stack area of the JVM state.
71 // This destroys any arguments that may be waiting on the stack.
72 void GraphKit::clean_stack(int from_sp) {
73 SafePointNode* map = this->map();
74 JVMState* jvms = this->jvms();
75 int stk_size = jvms->stk_size();
76 int stkoff = jvms->stkoff();
77 Node* top = this->top();
78 for (int i = from_sp; i < stk_size; i++) {
79 if (map->in(stkoff + i) != top) {
80 map->set_req(stkoff + i, top);
81 }
82 }
83 }
86 //--------------------------------sync_jvms-----------------------------------
87 // Make sure our current jvms agrees with our parse state.
88 JVMState* GraphKit::sync_jvms() const {
89 JVMState* jvms = this->jvms();
90 jvms->set_bci(bci()); // Record the new bci in the JVMState
91 jvms->set_sp(sp()); // Record the new sp in the JVMState
92 assert(jvms_in_sync(), "jvms is now in sync");
93 return jvms;
94 }
96 //--------------------------------sync_jvms_for_reexecute---------------------
97 // Make sure our current jvms agrees with our parse state. This version
98 // uses the reexecute_sp for reexecuting bytecodes.
99 JVMState* GraphKit::sync_jvms_for_reexecute() {
100 JVMState* jvms = this->jvms();
101 jvms->set_bci(bci()); // Record the new bci in the JVMState
102 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
103 return jvms;
104 }
106 #ifdef ASSERT
107 bool GraphKit::jvms_in_sync() const {
108 Parse* parse = is_Parse();
109 if (parse == NULL) {
110 if (bci() != jvms()->bci()) return false;
111 if (sp() != (int)jvms()->sp()) return false;
112 return true;
113 }
114 if (jvms()->method() != parse->method()) return false;
115 if (jvms()->bci() != parse->bci()) return false;
116 int jvms_sp = jvms()->sp();
117 if (jvms_sp != parse->sp()) return false;
118 int jvms_depth = jvms()->depth();
119 if (jvms_depth != parse->depth()) return false;
120 return true;
121 }
123 // Local helper checks for special internal merge points
124 // used to accumulate and merge exception states.
125 // They are marked by the region's in(0) edge being the map itself.
126 // Such merge points must never "escape" into the parser at large,
127 // until they have been handed to gvn.transform.
128 static bool is_hidden_merge(Node* reg) {
129 if (reg == NULL) return false;
130 if (reg->is_Phi()) {
131 reg = reg->in(0);
132 if (reg == NULL) return false;
133 }
134 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
135 }
137 void GraphKit::verify_map() const {
138 if (map() == NULL) return; // null map is OK
139 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
140 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
141 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
142 }
144 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
145 assert(ex_map->next_exception() == NULL, "not already part of a chain");
146 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
147 }
148 #endif
150 //---------------------------stop_and_kill_map---------------------------------
151 // Set _map to NULL, signalling a stop to further bytecode execution.
152 // First smash the current map's control to a constant, to mark it dead.
153 void GraphKit::stop_and_kill_map() {
154 SafePointNode* dead_map = stop();
155 if (dead_map != NULL) {
156 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
157 assert(dead_map->is_killed(), "must be so marked");
158 }
159 }
162 //--------------------------------stopped--------------------------------------
163 // Tell if _map is NULL, or control is top.
164 bool GraphKit::stopped() {
165 if (map() == NULL) return true;
166 else if (control() == top()) return true;
167 else return false;
168 }
171 //-----------------------------has_ex_handler----------------------------------
172 // Tell if this method or any caller method has exception handlers.
173 bool GraphKit::has_ex_handler() {
174 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
175 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
176 return true;
177 }
178 }
179 return false;
180 }
182 //------------------------------save_ex_oop------------------------------------
183 // Save an exception without blowing stack contents or other JVM state.
184 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
185 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
186 ex_map->add_req(ex_oop);
187 debug_only(verify_exception_state(ex_map));
188 }
190 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
191 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
192 Node* ex_oop = ex_map->in(ex_map->req()-1);
193 if (clear_it) ex_map->del_req(ex_map->req()-1);
194 return ex_oop;
195 }
197 //-----------------------------saved_ex_oop------------------------------------
198 // Recover a saved exception from its map.
199 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
200 return common_saved_ex_oop(ex_map, false);
201 }
203 //--------------------------clear_saved_ex_oop---------------------------------
204 // Erase a previously saved exception from its map.
205 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
206 return common_saved_ex_oop(ex_map, true);
207 }
209 #ifdef ASSERT
210 //---------------------------has_saved_ex_oop----------------------------------
211 // Erase a previously saved exception from its map.
212 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
213 return ex_map->req() == ex_map->jvms()->endoff()+1;
214 }
215 #endif
217 //-------------------------make_exception_state--------------------------------
218 // Turn the current JVM state into an exception state, appending the ex_oop.
219 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
220 sync_jvms();
221 SafePointNode* ex_map = stop(); // do not manipulate this map any more
222 set_saved_ex_oop(ex_map, ex_oop);
223 return ex_map;
224 }
227 //--------------------------add_exception_state--------------------------------
228 // Add an exception to my list of exceptions.
229 void GraphKit::add_exception_state(SafePointNode* ex_map) {
230 if (ex_map == NULL || ex_map->control() == top()) {
231 return;
232 }
233 #ifdef ASSERT
234 verify_exception_state(ex_map);
235 if (has_exceptions()) {
236 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
237 }
238 #endif
240 // If there is already an exception of exactly this type, merge with it.
241 // In particular, null-checks and other low-level exceptions common up here.
242 Node* ex_oop = saved_ex_oop(ex_map);
243 const Type* ex_type = _gvn.type(ex_oop);
244 if (ex_oop == top()) {
245 // No action needed.
246 return;
247 }
248 assert(ex_type->isa_instptr(), "exception must be an instance");
249 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
250 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
251 // We check sp also because call bytecodes can generate exceptions
252 // both before and after arguments are popped!
253 if (ex_type2 == ex_type
254 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
255 combine_exception_states(ex_map, e2);
256 return;
257 }
258 }
260 // No pre-existing exception of the same type. Chain it on the list.
261 push_exception_state(ex_map);
262 }
264 //-----------------------add_exception_states_from-----------------------------
265 void GraphKit::add_exception_states_from(JVMState* jvms) {
266 SafePointNode* ex_map = jvms->map()->next_exception();
267 if (ex_map != NULL) {
268 jvms->map()->set_next_exception(NULL);
269 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
270 next_map = ex_map->next_exception();
271 ex_map->set_next_exception(NULL);
272 add_exception_state(ex_map);
273 }
274 }
275 }
277 //-----------------------transfer_exceptions_into_jvms-------------------------
278 JVMState* GraphKit::transfer_exceptions_into_jvms() {
279 if (map() == NULL) {
280 // We need a JVMS to carry the exceptions, but the map has gone away.
281 // Create a scratch JVMS, cloned from any of the exception states...
282 if (has_exceptions()) {
283 _map = _exceptions;
284 _map = clone_map();
285 _map->set_next_exception(NULL);
286 clear_saved_ex_oop(_map);
287 debug_only(verify_map());
288 } else {
289 // ...or created from scratch
290 JVMState* jvms = new (C) JVMState(_method, NULL);
291 jvms->set_bci(_bci);
292 jvms->set_sp(_sp);
293 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms));
294 set_jvms(jvms);
295 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
296 set_all_memory(top());
297 while (map()->req() < jvms->endoff()) map()->add_req(top());
298 }
299 // (This is a kludge, in case you didn't notice.)
300 set_control(top());
301 }
302 JVMState* jvms = sync_jvms();
303 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
304 jvms->map()->set_next_exception(_exceptions);
305 _exceptions = NULL; // done with this set of exceptions
306 return jvms;
307 }
309 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
310 assert(is_hidden_merge(dstphi), "must be a special merge node");
311 assert(is_hidden_merge(srcphi), "must be a special merge node");
312 uint limit = srcphi->req();
313 for (uint i = PhiNode::Input; i < limit; i++) {
314 dstphi->add_req(srcphi->in(i));
315 }
316 }
317 static inline void add_one_req(Node* dstphi, Node* src) {
318 assert(is_hidden_merge(dstphi), "must be a special merge node");
319 assert(!is_hidden_merge(src), "must not be a special merge node");
320 dstphi->add_req(src);
321 }
323 //-----------------------combine_exception_states------------------------------
324 // This helper function combines exception states by building phis on a
325 // specially marked state-merging region. These regions and phis are
326 // untransformed, and can build up gradually. The region is marked by
327 // having a control input of its exception map, rather than NULL. Such
328 // regions do not appear except in this function, and in use_exception_state.
329 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
330 if (failing()) return; // dying anyway...
331 JVMState* ex_jvms = ex_map->_jvms;
332 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
333 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
334 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
335 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
336 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
337 assert(ex_map->req() == phi_map->req(), "matching maps");
338 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
339 Node* hidden_merge_mark = root();
340 Node* region = phi_map->control();
341 MergeMemNode* phi_mem = phi_map->merged_memory();
342 MergeMemNode* ex_mem = ex_map->merged_memory();
343 if (region->in(0) != hidden_merge_mark) {
344 // The control input is not (yet) a specially-marked region in phi_map.
345 // Make it so, and build some phis.
346 region = new (C) RegionNode(2);
347 _gvn.set_type(region, Type::CONTROL);
348 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
349 region->init_req(1, phi_map->control());
350 phi_map->set_control(region);
351 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
352 record_for_igvn(io_phi);
353 _gvn.set_type(io_phi, Type::ABIO);
354 phi_map->set_i_o(io_phi);
355 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
356 Node* m = mms.memory();
357 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
358 record_for_igvn(m_phi);
359 _gvn.set_type(m_phi, Type::MEMORY);
360 mms.set_memory(m_phi);
361 }
362 }
364 // Either or both of phi_map and ex_map might already be converted into phis.
365 Node* ex_control = ex_map->control();
366 // if there is special marking on ex_map also, we add multiple edges from src
367 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
368 // how wide was the destination phi_map, originally?
369 uint orig_width = region->req();
371 if (add_multiple) {
372 add_n_reqs(region, ex_control);
373 add_n_reqs(phi_map->i_o(), ex_map->i_o());
374 } else {
375 // ex_map has no merges, so we just add single edges everywhere
376 add_one_req(region, ex_control);
377 add_one_req(phi_map->i_o(), ex_map->i_o());
378 }
379 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
380 if (mms.is_empty()) {
381 // get a copy of the base memory, and patch some inputs into it
382 const TypePtr* adr_type = mms.adr_type(C);
383 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
384 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
385 mms.set_memory(phi);
386 // Prepare to append interesting stuff onto the newly sliced phi:
387 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
388 }
389 // Append stuff from ex_map:
390 if (add_multiple) {
391 add_n_reqs(mms.memory(), mms.memory2());
392 } else {
393 add_one_req(mms.memory(), mms.memory2());
394 }
395 }
396 uint limit = ex_map->req();
397 for (uint i = TypeFunc::Parms; i < limit; i++) {
398 // Skip everything in the JVMS after tos. (The ex_oop follows.)
399 if (i == tos) i = ex_jvms->monoff();
400 Node* src = ex_map->in(i);
401 Node* dst = phi_map->in(i);
402 if (src != dst) {
403 PhiNode* phi;
404 if (dst->in(0) != region) {
405 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
406 record_for_igvn(phi);
407 _gvn.set_type(phi, phi->type());
408 phi_map->set_req(i, dst);
409 // Prepare to append interesting stuff onto the new phi:
410 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
411 } else {
412 assert(dst->is_Phi(), "nobody else uses a hidden region");
413 phi = dst->as_Phi();
414 }
415 if (add_multiple && src->in(0) == ex_control) {
416 // Both are phis.
417 add_n_reqs(dst, src);
418 } else {
419 while (dst->req() < region->req()) add_one_req(dst, src);
420 }
421 const Type* srctype = _gvn.type(src);
422 if (phi->type() != srctype) {
423 const Type* dsttype = phi->type()->meet_speculative(srctype);
424 if (phi->type() != dsttype) {
425 phi->set_type(dsttype);
426 _gvn.set_type(phi, dsttype);
427 }
428 }
429 }
430 }
431 phi_map->merge_replaced_nodes_with(ex_map);
432 }
434 //--------------------------use_exception_state--------------------------------
435 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
436 if (failing()) { stop(); return top(); }
437 Node* region = phi_map->control();
438 Node* hidden_merge_mark = root();
439 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
440 Node* ex_oop = clear_saved_ex_oop(phi_map);
441 if (region->in(0) == hidden_merge_mark) {
442 // Special marking for internal ex-states. Process the phis now.
443 region->set_req(0, region); // now it's an ordinary region
444 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
445 // Note: Setting the jvms also sets the bci and sp.
446 set_control(_gvn.transform(region));
447 uint tos = jvms()->stkoff() + sp();
448 for (uint i = 1; i < tos; i++) {
449 Node* x = phi_map->in(i);
450 if (x->in(0) == region) {
451 assert(x->is_Phi(), "expected a special phi");
452 phi_map->set_req(i, _gvn.transform(x));
453 }
454 }
455 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
456 Node* x = mms.memory();
457 if (x->in(0) == region) {
458 assert(x->is_Phi(), "nobody else uses a hidden region");
459 mms.set_memory(_gvn.transform(x));
460 }
461 }
462 if (ex_oop->in(0) == region) {
463 assert(ex_oop->is_Phi(), "expected a special phi");
464 ex_oop = _gvn.transform(ex_oop);
465 }
466 } else {
467 set_jvms(phi_map->jvms());
468 }
470 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
471 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
472 return ex_oop;
473 }
475 //---------------------------------java_bc-------------------------------------
476 Bytecodes::Code GraphKit::java_bc() const {
477 ciMethod* method = this->method();
478 int bci = this->bci();
479 if (method != NULL && bci != InvocationEntryBci)
480 return method->java_code_at_bci(bci);
481 else
482 return Bytecodes::_illegal;
483 }
485 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
486 bool must_throw) {
487 // if the exception capability is set, then we will generate code
488 // to check the JavaThread.should_post_on_exceptions flag to see
489 // if we actually need to report exception events (for this
490 // thread). If we don't need to report exception events, we will
491 // take the normal fast path provided by add_exception_events. If
492 // exception event reporting is enabled for this thread, we will
493 // take the uncommon_trap in the BuildCutout below.
495 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
496 Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
497 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
498 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
500 // Test the should_post_on_exceptions_flag vs. 0
501 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
502 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
504 // Branch to slow_path if should_post_on_exceptions_flag was true
505 { BuildCutout unless(this, tst, PROB_MAX);
506 // Do not try anything fancy if we're notifying the VM on every throw.
507 // Cf. case Bytecodes::_athrow in parse2.cpp.
508 uncommon_trap(reason, Deoptimization::Action_none,
509 (ciKlass*)NULL, (char*)NULL, must_throw);
510 }
512 }
514 //------------------------------builtin_throw----------------------------------
515 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
516 bool must_throw = true;
518 if (env()->jvmti_can_post_on_exceptions()) {
519 // check if we must post exception events, take uncommon trap if so
520 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
521 // here if should_post_on_exceptions is false
522 // continue on with the normal codegen
523 }
525 // If this particular condition has not yet happened at this
526 // bytecode, then use the uncommon trap mechanism, and allow for
527 // a future recompilation if several traps occur here.
528 // If the throw is hot, try to use a more complicated inline mechanism
529 // which keeps execution inside the compiled code.
530 bool treat_throw_as_hot = false;
531 ciMethodData* md = method()->method_data();
533 if (ProfileTraps) {
534 if (too_many_traps(reason)) {
535 treat_throw_as_hot = true;
536 }
537 // (If there is no MDO at all, assume it is early in
538 // execution, and that any deopts are part of the
539 // startup transient, and don't need to be remembered.)
541 // Also, if there is a local exception handler, treat all throws
542 // as hot if there has been at least one in this method.
543 if (C->trap_count(reason) != 0
544 && method()->method_data()->trap_count(reason) != 0
545 && has_ex_handler()) {
546 treat_throw_as_hot = true;
547 }
548 }
550 // If this throw happens frequently, an uncommon trap might cause
551 // a performance pothole. If there is a local exception handler,
552 // and if this particular bytecode appears to be deoptimizing often,
553 // let us handle the throw inline, with a preconstructed instance.
554 // Note: If the deopt count has blown up, the uncommon trap
555 // runtime is going to flush this nmethod, not matter what.
556 if (treat_throw_as_hot
557 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
558 // If the throw is local, we use a pre-existing instance and
559 // punt on the backtrace. This would lead to a missing backtrace
560 // (a repeat of 4292742) if the backtrace object is ever asked
561 // for its backtrace.
562 // Fixing this remaining case of 4292742 requires some flavor of
563 // escape analysis. Leave that for the future.
564 ciInstance* ex_obj = NULL;
565 switch (reason) {
566 case Deoptimization::Reason_null_check:
567 ex_obj = env()->NullPointerException_instance();
568 break;
569 case Deoptimization::Reason_div0_check:
570 ex_obj = env()->ArithmeticException_instance();
571 break;
572 case Deoptimization::Reason_range_check:
573 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
574 break;
575 case Deoptimization::Reason_class_check:
576 if (java_bc() == Bytecodes::_aastore) {
577 ex_obj = env()->ArrayStoreException_instance();
578 } else {
579 ex_obj = env()->ClassCastException_instance();
580 }
581 break;
582 }
583 if (failing()) { stop(); return; } // exception allocation might fail
584 if (ex_obj != NULL) {
585 // Cheat with a preallocated exception object.
586 if (C->log() != NULL)
587 C->log()->elem("hot_throw preallocated='1' reason='%s'",
588 Deoptimization::trap_reason_name(reason));
589 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
590 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
592 // Clear the detail message of the preallocated exception object.
593 // Weblogic sometimes mutates the detail message of exceptions
594 // using reflection.
595 int offset = java_lang_Throwable::get_detailMessage_offset();
596 const TypePtr* adr_typ = ex_con->add_offset(offset);
598 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
599 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
600 // Conservatively release stores of object references.
601 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
603 add_exception_state(make_exception_state(ex_node));
604 return;
605 }
606 }
608 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
609 // It won't be much cheaper than bailing to the interp., since we'll
610 // have to pass up all the debug-info, and the runtime will have to
611 // create the stack trace.
613 // Usual case: Bail to interpreter.
614 // Reserve the right to recompile if we haven't seen anything yet.
616 assert(!Deoptimization::reason_is_speculate(reason), "unsupported");
617 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
618 if (treat_throw_as_hot
619 && (method()->method_data()->trap_recompiled_at(bci(), NULL)
620 || C->too_many_traps(reason))) {
621 // We cannot afford to take more traps here. Suffer in the interpreter.
622 if (C->log() != NULL)
623 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
624 Deoptimization::trap_reason_name(reason),
625 C->trap_count(reason));
626 action = Deoptimization::Action_none;
627 }
629 // "must_throw" prunes the JVM state to include only the stack, if there
630 // are no local exception handlers. This should cut down on register
631 // allocation time and code size, by drastically reducing the number
632 // of in-edges on the call to the uncommon trap.
634 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
635 }
638 //----------------------------PreserveJVMState---------------------------------
639 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
640 debug_only(kit->verify_map());
641 _kit = kit;
642 _map = kit->map(); // preserve the map
643 _sp = kit->sp();
644 kit->set_map(clone_map ? kit->clone_map() : NULL);
645 #ifdef ASSERT
646 _bci = kit->bci();
647 Parse* parser = kit->is_Parse();
648 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
649 _block = block;
650 #endif
651 }
652 PreserveJVMState::~PreserveJVMState() {
653 GraphKit* kit = _kit;
654 #ifdef ASSERT
655 assert(kit->bci() == _bci, "bci must not shift");
656 Parse* parser = kit->is_Parse();
657 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
658 assert(block == _block, "block must not shift");
659 #endif
660 kit->set_map(_map);
661 kit->set_sp(_sp);
662 }
665 //-----------------------------BuildCutout-------------------------------------
666 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
667 : PreserveJVMState(kit)
668 {
669 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
670 SafePointNode* outer_map = _map; // preserved map is caller's
671 SafePointNode* inner_map = kit->map();
672 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
673 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
674 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
675 }
676 BuildCutout::~BuildCutout() {
677 GraphKit* kit = _kit;
678 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
679 }
681 //---------------------------PreserveReexecuteState----------------------------
682 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
683 assert(!kit->stopped(), "must call stopped() before");
684 _kit = kit;
685 _sp = kit->sp();
686 _reexecute = kit->jvms()->_reexecute;
687 }
688 PreserveReexecuteState::~PreserveReexecuteState() {
689 if (_kit->stopped()) return;
690 _kit->jvms()->_reexecute = _reexecute;
691 _kit->set_sp(_sp);
692 }
694 //------------------------------clone_map--------------------------------------
695 // Implementation of PreserveJVMState
696 //
697 // Only clone_map(...) here. If this function is only used in the
698 // PreserveJVMState class we may want to get rid of this extra
699 // function eventually and do it all there.
701 SafePointNode* GraphKit::clone_map() {
702 if (map() == NULL) return NULL;
704 // Clone the memory edge first
705 Node* mem = MergeMemNode::make(C, map()->memory());
706 gvn().set_type_bottom(mem);
708 SafePointNode *clonemap = (SafePointNode*)map()->clone();
709 JVMState* jvms = this->jvms();
710 JVMState* clonejvms = jvms->clone_shallow(C);
711 clonemap->set_memory(mem);
712 clonemap->set_jvms(clonejvms);
713 clonejvms->set_map(clonemap);
714 record_for_igvn(clonemap);
715 gvn().set_type_bottom(clonemap);
716 return clonemap;
717 }
720 //-----------------------------set_map_clone-----------------------------------
721 void GraphKit::set_map_clone(SafePointNode* m) {
722 _map = m;
723 _map = clone_map();
724 _map->set_next_exception(NULL);
725 debug_only(verify_map());
726 }
729 //----------------------------kill_dead_locals---------------------------------
730 // Detect any locals which are known to be dead, and force them to top.
731 void GraphKit::kill_dead_locals() {
732 // Consult the liveness information for the locals. If any
733 // of them are unused, then they can be replaced by top(). This
734 // should help register allocation time and cut down on the size
735 // of the deoptimization information.
737 // This call is made from many of the bytecode handling
738 // subroutines called from the Big Switch in do_one_bytecode.
739 // Every bytecode which might include a slow path is responsible
740 // for killing its dead locals. The more consistent we
741 // are about killing deads, the fewer useless phis will be
742 // constructed for them at various merge points.
744 // bci can be -1 (InvocationEntryBci). We return the entry
745 // liveness for the method.
747 if (method() == NULL || method()->code_size() == 0) {
748 // We are building a graph for a call to a native method.
749 // All locals are live.
750 return;
751 }
753 ResourceMark rm;
755 // Consult the liveness information for the locals. If any
756 // of them are unused, then they can be replaced by top(). This
757 // should help register allocation time and cut down on the size
758 // of the deoptimization information.
759 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
761 int len = (int)live_locals.size();
762 assert(len <= jvms()->loc_size(), "too many live locals");
763 for (int local = 0; local < len; local++) {
764 if (!live_locals.at(local)) {
765 set_local(local, top());
766 }
767 }
768 }
770 #ifdef ASSERT
771 //-------------------------dead_locals_are_killed------------------------------
772 // Return true if all dead locals are set to top in the map.
773 // Used to assert "clean" debug info at various points.
774 bool GraphKit::dead_locals_are_killed() {
775 if (method() == NULL || method()->code_size() == 0) {
776 // No locals need to be dead, so all is as it should be.
777 return true;
778 }
780 // Make sure somebody called kill_dead_locals upstream.
781 ResourceMark rm;
782 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
783 if (jvms->loc_size() == 0) continue; // no locals to consult
784 SafePointNode* map = jvms->map();
785 ciMethod* method = jvms->method();
786 int bci = jvms->bci();
787 if (jvms == this->jvms()) {
788 bci = this->bci(); // it might not yet be synched
789 }
790 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
791 int len = (int)live_locals.size();
792 if (!live_locals.is_valid() || len == 0)
793 // This method is trivial, or is poisoned by a breakpoint.
794 return true;
795 assert(len == jvms->loc_size(), "live map consistent with locals map");
796 for (int local = 0; local < len; local++) {
797 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
798 if (PrintMiscellaneous && (Verbose || WizardMode)) {
799 tty->print_cr("Zombie local %d: ", local);
800 jvms->dump();
801 }
802 return false;
803 }
804 }
805 }
806 return true;
807 }
809 #endif //ASSERT
811 // Helper function for enforcing certain bytecodes to reexecute if
812 // deoptimization happens
813 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
814 ciMethod* cur_method = jvms->method();
815 int cur_bci = jvms->bci();
816 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
817 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
818 return Interpreter::bytecode_should_reexecute(code) ||
819 is_anewarray && code == Bytecodes::_multianewarray;
820 // Reexecute _multianewarray bytecode which was replaced with
821 // sequence of [a]newarray. See Parse::do_multianewarray().
822 //
823 // Note: interpreter should not have it set since this optimization
824 // is limited by dimensions and guarded by flag so in some cases
825 // multianewarray() runtime calls will be generated and
826 // the bytecode should not be reexecutes (stack will not be reset).
827 } else
828 return false;
829 }
831 // Helper function for adding JVMState and debug information to node
832 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
833 // Add the safepoint edges to the call (or other safepoint).
835 // Make sure dead locals are set to top. This
836 // should help register allocation time and cut down on the size
837 // of the deoptimization information.
838 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
840 // Walk the inline list to fill in the correct set of JVMState's
841 // Also fill in the associated edges for each JVMState.
843 // If the bytecode needs to be reexecuted we need to put
844 // the arguments back on the stack.
845 const bool should_reexecute = jvms()->should_reexecute();
846 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
848 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
849 // undefined if the bci is different. This is normal for Parse but it
850 // should not happen for LibraryCallKit because only one bci is processed.
851 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
852 "in LibraryCallKit the reexecute bit should not change");
854 // If we are guaranteed to throw, we can prune everything but the
855 // input to the current bytecode.
856 bool can_prune_locals = false;
857 uint stack_slots_not_pruned = 0;
858 int inputs = 0, depth = 0;
859 if (must_throw) {
860 assert(method() == youngest_jvms->method(), "sanity");
861 if (compute_stack_effects(inputs, depth)) {
862 can_prune_locals = true;
863 stack_slots_not_pruned = inputs;
864 }
865 }
867 if (env()->jvmti_can_access_local_variables()) {
868 // At any safepoint, this method can get breakpointed, which would
869 // then require an immediate deoptimization.
870 can_prune_locals = false; // do not prune locals
871 stack_slots_not_pruned = 0;
872 }
874 // do not scribble on the input jvms
875 JVMState* out_jvms = youngest_jvms->clone_deep(C);
876 call->set_jvms(out_jvms); // Start jvms list for call node
878 // For a known set of bytecodes, the interpreter should reexecute them if
879 // deoptimization happens. We set the reexecute state for them here
880 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
881 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
882 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
883 }
885 // Presize the call:
886 DEBUG_ONLY(uint non_debug_edges = call->req());
887 call->add_req_batch(top(), youngest_jvms->debug_depth());
888 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
890 // Set up edges so that the call looks like this:
891 // Call [state:] ctl io mem fptr retadr
892 // [parms:] parm0 ... parmN
893 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
894 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
895 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
896 // Note that caller debug info precedes callee debug info.
898 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
899 uint debug_ptr = call->req();
901 // Loop over the map input edges associated with jvms, add them
902 // to the call node, & reset all offsets to match call node array.
903 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
904 uint debug_end = debug_ptr;
905 uint debug_start = debug_ptr - in_jvms->debug_size();
906 debug_ptr = debug_start; // back up the ptr
908 uint p = debug_start; // walks forward in [debug_start, debug_end)
909 uint j, k, l;
910 SafePointNode* in_map = in_jvms->map();
911 out_jvms->set_map(call);
913 if (can_prune_locals) {
914 assert(in_jvms->method() == out_jvms->method(), "sanity");
915 // If the current throw can reach an exception handler in this JVMS,
916 // then we must keep everything live that can reach that handler.
917 // As a quick and dirty approximation, we look for any handlers at all.
918 if (in_jvms->method()->has_exception_handlers()) {
919 can_prune_locals = false;
920 }
921 }
923 // Add the Locals
924 k = in_jvms->locoff();
925 l = in_jvms->loc_size();
926 out_jvms->set_locoff(p);
927 if (!can_prune_locals) {
928 for (j = 0; j < l; j++)
929 call->set_req(p++, in_map->in(k+j));
930 } else {
931 p += l; // already set to top above by add_req_batch
932 }
934 // Add the Expression Stack
935 k = in_jvms->stkoff();
936 l = in_jvms->sp();
937 out_jvms->set_stkoff(p);
938 if (!can_prune_locals) {
939 for (j = 0; j < l; j++)
940 call->set_req(p++, in_map->in(k+j));
941 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
942 // Divide stack into {S0,...,S1}, where S0 is set to top.
943 uint s1 = stack_slots_not_pruned;
944 stack_slots_not_pruned = 0; // for next iteration
945 if (s1 > l) s1 = l;
946 uint s0 = l - s1;
947 p += s0; // skip the tops preinstalled by add_req_batch
948 for (j = s0; j < l; j++)
949 call->set_req(p++, in_map->in(k+j));
950 } else {
951 p += l; // already set to top above by add_req_batch
952 }
954 // Add the Monitors
955 k = in_jvms->monoff();
956 l = in_jvms->mon_size();
957 out_jvms->set_monoff(p);
958 for (j = 0; j < l; j++)
959 call->set_req(p++, in_map->in(k+j));
961 // Copy any scalar object fields.
962 k = in_jvms->scloff();
963 l = in_jvms->scl_size();
964 out_jvms->set_scloff(p);
965 for (j = 0; j < l; j++)
966 call->set_req(p++, in_map->in(k+j));
968 // Finish the new jvms.
969 out_jvms->set_endoff(p);
971 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
972 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
973 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
974 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
975 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
976 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
978 // Update the two tail pointers in parallel.
979 out_jvms = out_jvms->caller();
980 in_jvms = in_jvms->caller();
981 }
983 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
985 // Test the correctness of JVMState::debug_xxx accessors:
986 assert(call->jvms()->debug_start() == non_debug_edges, "");
987 assert(call->jvms()->debug_end() == call->req(), "");
988 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
989 }
991 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
992 Bytecodes::Code code = java_bc();
993 if (code == Bytecodes::_wide) {
994 code = method()->java_code_at_bci(bci() + 1);
995 }
997 BasicType rtype = T_ILLEGAL;
998 int rsize = 0;
1000 if (code != Bytecodes::_illegal) {
1001 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1002 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1003 if (rtype < T_CONFLICT)
1004 rsize = type2size[rtype];
1005 }
1007 switch (code) {
1008 case Bytecodes::_illegal:
1009 return false;
1011 case Bytecodes::_ldc:
1012 case Bytecodes::_ldc_w:
1013 case Bytecodes::_ldc2_w:
1014 inputs = 0;
1015 break;
1017 case Bytecodes::_dup: inputs = 1; break;
1018 case Bytecodes::_dup_x1: inputs = 2; break;
1019 case Bytecodes::_dup_x2: inputs = 3; break;
1020 case Bytecodes::_dup2: inputs = 2; break;
1021 case Bytecodes::_dup2_x1: inputs = 3; break;
1022 case Bytecodes::_dup2_x2: inputs = 4; break;
1023 case Bytecodes::_swap: inputs = 2; break;
1024 case Bytecodes::_arraylength: inputs = 1; break;
1026 case Bytecodes::_getstatic:
1027 case Bytecodes::_putstatic:
1028 case Bytecodes::_getfield:
1029 case Bytecodes::_putfield:
1030 {
1031 bool ignored_will_link;
1032 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1033 int size = field->type()->size();
1034 bool is_get = (depth >= 0), is_static = (depth & 1);
1035 inputs = (is_static ? 0 : 1);
1036 if (is_get) {
1037 depth = size - inputs;
1038 } else {
1039 inputs += size; // putxxx pops the value from the stack
1040 depth = - inputs;
1041 }
1042 }
1043 break;
1045 case Bytecodes::_invokevirtual:
1046 case Bytecodes::_invokespecial:
1047 case Bytecodes::_invokestatic:
1048 case Bytecodes::_invokedynamic:
1049 case Bytecodes::_invokeinterface:
1050 {
1051 bool ignored_will_link;
1052 ciSignature* declared_signature = NULL;
1053 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1054 assert(declared_signature != NULL, "cannot be null");
1055 inputs = declared_signature->arg_size_for_bc(code);
1056 int size = declared_signature->return_type()->size();
1057 depth = size - inputs;
1058 }
1059 break;
1061 case Bytecodes::_multianewarray:
1062 {
1063 ciBytecodeStream iter(method());
1064 iter.reset_to_bci(bci());
1065 iter.next();
1066 inputs = iter.get_dimensions();
1067 assert(rsize == 1, "");
1068 depth = rsize - inputs;
1069 }
1070 break;
1072 case Bytecodes::_ireturn:
1073 case Bytecodes::_lreturn:
1074 case Bytecodes::_freturn:
1075 case Bytecodes::_dreturn:
1076 case Bytecodes::_areturn:
1077 assert(rsize = -depth, "");
1078 inputs = rsize;
1079 break;
1081 case Bytecodes::_jsr:
1082 case Bytecodes::_jsr_w:
1083 inputs = 0;
1084 depth = 1; // S.B. depth=1, not zero
1085 break;
1087 default:
1088 // bytecode produces a typed result
1089 inputs = rsize - depth;
1090 assert(inputs >= 0, "");
1091 break;
1092 }
1094 #ifdef ASSERT
1095 // spot check
1096 int outputs = depth + inputs;
1097 assert(outputs >= 0, "sanity");
1098 switch (code) {
1099 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1100 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1101 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1102 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1103 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1104 }
1105 #endif //ASSERT
1107 return true;
1108 }
1112 //------------------------------basic_plus_adr---------------------------------
1113 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1114 // short-circuit a common case
1115 if (offset == intcon(0)) return ptr;
1116 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1117 }
1119 Node* GraphKit::ConvI2L(Node* offset) {
1120 // short-circuit a common case
1121 jint offset_con = find_int_con(offset, Type::OffsetBot);
1122 if (offset_con != Type::OffsetBot) {
1123 return longcon((jlong) offset_con);
1124 }
1125 return _gvn.transform( new (C) ConvI2LNode(offset));
1126 }
1128 Node* GraphKit::ConvI2UL(Node* offset) {
1129 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1130 if (offset_con != (juint) Type::OffsetBot) {
1131 return longcon((julong) offset_con);
1132 }
1133 Node* conv = _gvn.transform( new (C) ConvI2LNode(offset));
1134 Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) );
1135 return _gvn.transform( new (C) AndLNode(conv, mask) );
1136 }
1138 Node* GraphKit::ConvL2I(Node* offset) {
1139 // short-circuit a common case
1140 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1141 if (offset_con != (jlong)Type::OffsetBot) {
1142 return intcon((int) offset_con);
1143 }
1144 return _gvn.transform( new (C) ConvL2INode(offset));
1145 }
1147 //-------------------------load_object_klass-----------------------------------
1148 Node* GraphKit::load_object_klass(Node* obj) {
1149 // Special-case a fresh allocation to avoid building nodes:
1150 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1151 if (akls != NULL) return akls;
1152 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1153 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1154 }
1156 //-------------------------load_array_length-----------------------------------
1157 Node* GraphKit::load_array_length(Node* array) {
1158 // Special-case a fresh allocation to avoid building nodes:
1159 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1160 Node *alen;
1161 if (alloc == NULL) {
1162 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1163 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1164 } else {
1165 alen = alloc->Ideal_length();
1166 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1167 if (ccast != alen) {
1168 alen = _gvn.transform(ccast);
1169 }
1170 }
1171 return alen;
1172 }
1174 //------------------------------do_null_check----------------------------------
1175 // Helper function to do a NULL pointer check. Returned value is
1176 // the incoming address with NULL casted away. You are allowed to use the
1177 // not-null value only if you are control dependent on the test.
1178 extern int explicit_null_checks_inserted,
1179 explicit_null_checks_elided;
1180 Node* GraphKit::null_check_common(Node* value, BasicType type,
1181 // optional arguments for variations:
1182 bool assert_null,
1183 Node* *null_control) {
1184 assert(!assert_null || null_control == NULL, "not both at once");
1185 if (stopped()) return top();
1186 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1187 // For some performance testing, we may wish to suppress null checking.
1188 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1189 return value;
1190 }
1191 explicit_null_checks_inserted++;
1193 // Construct NULL check
1194 Node *chk = NULL;
1195 switch(type) {
1196 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1197 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1198 case T_ARRAY : // fall through
1199 type = T_OBJECT; // simplify further tests
1200 case T_OBJECT : {
1201 const Type *t = _gvn.type( value );
1203 const TypeOopPtr* tp = t->isa_oopptr();
1204 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1205 // Only for do_null_check, not any of its siblings:
1206 && !assert_null && null_control == NULL) {
1207 // Usually, any field access or invocation on an unloaded oop type
1208 // will simply fail to link, since the statically linked class is
1209 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1210 // the static class is loaded but the sharper oop type is not.
1211 // Rather than checking for this obscure case in lots of places,
1212 // we simply observe that a null check on an unloaded class
1213 // will always be followed by a nonsense operation, so we
1214 // can just issue the uncommon trap here.
1215 // Our access to the unloaded class will only be correct
1216 // after it has been loaded and initialized, which requires
1217 // a trip through the interpreter.
1218 #ifndef PRODUCT
1219 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1220 #endif
1221 uncommon_trap(Deoptimization::Reason_unloaded,
1222 Deoptimization::Action_reinterpret,
1223 tp->klass(), "!loaded");
1224 return top();
1225 }
1227 if (assert_null) {
1228 // See if the type is contained in NULL_PTR.
1229 // If so, then the value is already null.
1230 if (t->higher_equal(TypePtr::NULL_PTR)) {
1231 explicit_null_checks_elided++;
1232 return value; // Elided null assert quickly!
1233 }
1234 } else {
1235 // See if mixing in the NULL pointer changes type.
1236 // If so, then the NULL pointer was not allowed in the original
1237 // type. In other words, "value" was not-null.
1238 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1239 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1240 explicit_null_checks_elided++;
1241 return value; // Elided null check quickly!
1242 }
1243 }
1244 chk = new (C) CmpPNode( value, null() );
1245 break;
1246 }
1248 default:
1249 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1250 }
1251 assert(chk != NULL, "sanity check");
1252 chk = _gvn.transform(chk);
1254 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1255 BoolNode *btst = new (C) BoolNode( chk, btest);
1256 Node *tst = _gvn.transform( btst );
1258 //-----------
1259 // if peephole optimizations occurred, a prior test existed.
1260 // If a prior test existed, maybe it dominates as we can avoid this test.
1261 if (tst != btst && type == T_OBJECT) {
1262 // At this point we want to scan up the CFG to see if we can
1263 // find an identical test (and so avoid this test altogether).
1264 Node *cfg = control();
1265 int depth = 0;
1266 while( depth < 16 ) { // Limit search depth for speed
1267 if( cfg->Opcode() == Op_IfTrue &&
1268 cfg->in(0)->in(1) == tst ) {
1269 // Found prior test. Use "cast_not_null" to construct an identical
1270 // CastPP (and hence hash to) as already exists for the prior test.
1271 // Return that casted value.
1272 if (assert_null) {
1273 replace_in_map(value, null());
1274 return null(); // do not issue the redundant test
1275 }
1276 Node *oldcontrol = control();
1277 set_control(cfg);
1278 Node *res = cast_not_null(value);
1279 set_control(oldcontrol);
1280 explicit_null_checks_elided++;
1281 return res;
1282 }
1283 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1284 if (cfg == NULL) break; // Quit at region nodes
1285 depth++;
1286 }
1287 }
1289 //-----------
1290 // Branch to failure if null
1291 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1292 Deoptimization::DeoptReason reason;
1293 if (assert_null)
1294 reason = Deoptimization::Reason_null_assert;
1295 else if (type == T_OBJECT)
1296 reason = Deoptimization::Reason_null_check;
1297 else
1298 reason = Deoptimization::Reason_div0_check;
1300 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1301 // ciMethodData::has_trap_at will return a conservative -1 if any
1302 // must-be-null assertion has failed. This could cause performance
1303 // problems for a method after its first do_null_assert failure.
1304 // Consider using 'Reason_class_check' instead?
1306 // To cause an implicit null check, we set the not-null probability
1307 // to the maximum (PROB_MAX). For an explicit check the probability
1308 // is set to a smaller value.
1309 if (null_control != NULL || too_many_traps(reason)) {
1310 // probability is less likely
1311 ok_prob = PROB_LIKELY_MAG(3);
1312 } else if (!assert_null &&
1313 (ImplicitNullCheckThreshold > 0) &&
1314 method() != NULL &&
1315 (method()->method_data()->trap_count(reason)
1316 >= (uint)ImplicitNullCheckThreshold)) {
1317 ok_prob = PROB_LIKELY_MAG(3);
1318 }
1320 if (null_control != NULL) {
1321 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1322 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1323 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1324 if (null_true == top())
1325 explicit_null_checks_elided++;
1326 (*null_control) = null_true;
1327 } else {
1328 BuildCutout unless(this, tst, ok_prob);
1329 // Check for optimizer eliding test at parse time
1330 if (stopped()) {
1331 // Failure not possible; do not bother making uncommon trap.
1332 explicit_null_checks_elided++;
1333 } else if (assert_null) {
1334 uncommon_trap(reason,
1335 Deoptimization::Action_make_not_entrant,
1336 NULL, "assert_null");
1337 } else {
1338 replace_in_map(value, zerocon(type));
1339 builtin_throw(reason);
1340 }
1341 }
1343 // Must throw exception, fall-thru not possible?
1344 if (stopped()) {
1345 return top(); // No result
1346 }
1348 if (assert_null) {
1349 // Cast obj to null on this path.
1350 replace_in_map(value, zerocon(type));
1351 return zerocon(type);
1352 }
1354 // Cast obj to not-null on this path, if there is no null_control.
1355 // (If there is a null_control, a non-null value may come back to haunt us.)
1356 if (type == T_OBJECT) {
1357 Node* cast = cast_not_null(value, false);
1358 if (null_control == NULL || (*null_control) == top())
1359 replace_in_map(value, cast);
1360 value = cast;
1361 }
1363 return value;
1364 }
1367 //------------------------------cast_not_null----------------------------------
1368 // Cast obj to not-null on this path
1369 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1370 const Type *t = _gvn.type(obj);
1371 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1372 // Object is already not-null?
1373 if( t == t_not_null ) return obj;
1375 Node *cast = new (C) CastPPNode(obj,t_not_null);
1376 cast->init_req(0, control());
1377 cast = _gvn.transform( cast );
1379 // Scan for instances of 'obj' in the current JVM mapping.
1380 // These instances are known to be not-null after the test.
1381 if (do_replace_in_map)
1382 replace_in_map(obj, cast);
1384 return cast; // Return casted value
1385 }
1388 //--------------------------replace_in_map-------------------------------------
1389 void GraphKit::replace_in_map(Node* old, Node* neww) {
1390 if (old == neww) {
1391 return;
1392 }
1394 map()->replace_edge(old, neww);
1396 // Note: This operation potentially replaces any edge
1397 // on the map. This includes locals, stack, and monitors
1398 // of the current (innermost) JVM state.
1400 // don't let inconsistent types from profiling escape this
1401 // method
1403 const Type* told = _gvn.type(old);
1404 const Type* tnew = _gvn.type(neww);
1406 if (!tnew->higher_equal(told)) {
1407 return;
1408 }
1410 map()->record_replaced_node(old, neww);
1411 }
1414 //=============================================================================
1415 //--------------------------------memory---------------------------------------
1416 Node* GraphKit::memory(uint alias_idx) {
1417 MergeMemNode* mem = merged_memory();
1418 Node* p = mem->memory_at(alias_idx);
1419 _gvn.set_type(p, Type::MEMORY); // must be mapped
1420 return p;
1421 }
1423 //-----------------------------reset_memory------------------------------------
1424 Node* GraphKit::reset_memory() {
1425 Node* mem = map()->memory();
1426 // do not use this node for any more parsing!
1427 debug_only( map()->set_memory((Node*)NULL) );
1428 return _gvn.transform( mem );
1429 }
1431 //------------------------------set_all_memory---------------------------------
1432 void GraphKit::set_all_memory(Node* newmem) {
1433 Node* mergemem = MergeMemNode::make(C, newmem);
1434 gvn().set_type_bottom(mergemem);
1435 map()->set_memory(mergemem);
1436 }
1438 //------------------------------set_all_memory_call----------------------------
1439 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1440 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1441 set_all_memory(newmem);
1442 }
1444 //=============================================================================
1445 //
1446 // parser factory methods for MemNodes
1447 //
1448 // These are layered on top of the factory methods in LoadNode and StoreNode,
1449 // and integrate with the parser's memory state and _gvn engine.
1450 //
1452 // factory methods in "int adr_idx"
1453 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1454 int adr_idx,
1455 MemNode::MemOrd mo, bool require_atomic_access) {
1456 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1457 const TypePtr* adr_type = NULL; // debug-mode-only argument
1458 debug_only(adr_type = C->get_adr_type(adr_idx));
1459 Node* mem = memory(adr_idx);
1460 Node* ld;
1461 if (require_atomic_access && bt == T_LONG) {
1462 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo);
1463 } else {
1464 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo);
1465 }
1466 ld = _gvn.transform(ld);
1467 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1468 // Improve graph before escape analysis and boxing elimination.
1469 record_for_igvn(ld);
1470 }
1471 return ld;
1472 }
1474 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1475 int adr_idx,
1476 MemNode::MemOrd mo,
1477 bool require_atomic_access) {
1478 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1479 const TypePtr* adr_type = NULL;
1480 debug_only(adr_type = C->get_adr_type(adr_idx));
1481 Node *mem = memory(adr_idx);
1482 Node* st;
1483 if (require_atomic_access && bt == T_LONG) {
1484 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1485 } else {
1486 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1487 }
1488 st = _gvn.transform(st);
1489 set_memory(st, adr_idx);
1490 // Back-to-back stores can only remove intermediate store with DU info
1491 // so push on worklist for optimizer.
1492 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1493 record_for_igvn(st);
1495 return st;
1496 }
1499 void GraphKit::pre_barrier(bool do_load,
1500 Node* ctl,
1501 Node* obj,
1502 Node* adr,
1503 uint adr_idx,
1504 Node* val,
1505 const TypeOopPtr* val_type,
1506 Node* pre_val,
1507 BasicType bt) {
1509 BarrierSet* bs = Universe::heap()->barrier_set();
1510 set_control(ctl);
1511 switch (bs->kind()) {
1512 case BarrierSet::G1SATBCT:
1513 case BarrierSet::G1SATBCTLogging:
1514 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1515 break;
1517 case BarrierSet::CardTableModRef:
1518 case BarrierSet::CardTableExtension:
1519 case BarrierSet::ModRef:
1520 break;
1522 case BarrierSet::Other:
1523 default :
1524 ShouldNotReachHere();
1526 }
1527 }
1529 bool GraphKit::can_move_pre_barrier() const {
1530 BarrierSet* bs = Universe::heap()->barrier_set();
1531 switch (bs->kind()) {
1532 case BarrierSet::G1SATBCT:
1533 case BarrierSet::G1SATBCTLogging:
1534 return true; // Can move it if no safepoint
1536 case BarrierSet::CardTableModRef:
1537 case BarrierSet::CardTableExtension:
1538 case BarrierSet::ModRef:
1539 return true; // There is no pre-barrier
1541 case BarrierSet::Other:
1542 default :
1543 ShouldNotReachHere();
1544 }
1545 return false;
1546 }
1548 void GraphKit::post_barrier(Node* ctl,
1549 Node* store,
1550 Node* obj,
1551 Node* adr,
1552 uint adr_idx,
1553 Node* val,
1554 BasicType bt,
1555 bool use_precise) {
1556 BarrierSet* bs = Universe::heap()->barrier_set();
1557 set_control(ctl);
1558 switch (bs->kind()) {
1559 case BarrierSet::G1SATBCT:
1560 case BarrierSet::G1SATBCTLogging:
1561 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1562 break;
1564 case BarrierSet::CardTableModRef:
1565 case BarrierSet::CardTableExtension:
1566 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1567 break;
1569 case BarrierSet::ModRef:
1570 break;
1572 case BarrierSet::Other:
1573 default :
1574 ShouldNotReachHere();
1576 }
1577 }
1579 Node* GraphKit::store_oop(Node* ctl,
1580 Node* obj,
1581 Node* adr,
1582 const TypePtr* adr_type,
1583 Node* val,
1584 const TypeOopPtr* val_type,
1585 BasicType bt,
1586 bool use_precise,
1587 MemNode::MemOrd mo) {
1588 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1589 // could be delayed during Parse (for example, in adjust_map_after_if()).
1590 // Execute transformation here to avoid barrier generation in such case.
1591 if (_gvn.type(val) == TypePtr::NULL_PTR)
1592 val = _gvn.makecon(TypePtr::NULL_PTR);
1594 set_control(ctl);
1595 if (stopped()) return top(); // Dead path ?
1597 assert(bt == T_OBJECT, "sanity");
1598 assert(val != NULL, "not dead path");
1599 uint adr_idx = C->get_alias_index(adr_type);
1600 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1602 pre_barrier(true /* do_load */,
1603 control(), obj, adr, adr_idx, val, val_type,
1604 NULL /* pre_val */,
1605 bt);
1607 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo);
1608 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1609 return store;
1610 }
1612 // Could be an array or object we don't know at compile time (unsafe ref.)
1613 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1614 Node* obj, // containing obj
1615 Node* adr, // actual adress to store val at
1616 const TypePtr* adr_type,
1617 Node* val,
1618 BasicType bt,
1619 MemNode::MemOrd mo) {
1620 Compile::AliasType* at = C->alias_type(adr_type);
1621 const TypeOopPtr* val_type = NULL;
1622 if (adr_type->isa_instptr()) {
1623 if (at->field() != NULL) {
1624 // known field. This code is a copy of the do_put_xxx logic.
1625 ciField* field = at->field();
1626 if (!field->type()->is_loaded()) {
1627 val_type = TypeInstPtr::BOTTOM;
1628 } else {
1629 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1630 }
1631 }
1632 } else if (adr_type->isa_aryptr()) {
1633 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1634 }
1635 if (val_type == NULL) {
1636 val_type = TypeInstPtr::BOTTOM;
1637 }
1638 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo);
1639 }
1642 //-------------------------array_element_address-------------------------
1643 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1644 const TypeInt* sizetype) {
1645 uint shift = exact_log2(type2aelembytes(elembt));
1646 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1648 // short-circuit a common case (saves lots of confusing waste motion)
1649 jint idx_con = find_int_con(idx, -1);
1650 if (idx_con >= 0) {
1651 intptr_t offset = header + ((intptr_t)idx_con << shift);
1652 return basic_plus_adr(ary, offset);
1653 }
1655 // must be correct type for alignment purposes
1656 Node* base = basic_plus_adr(ary, header);
1657 #ifdef _LP64
1658 // The scaled index operand to AddP must be a clean 64-bit value.
1659 // Java allows a 32-bit int to be incremented to a negative
1660 // value, which appears in a 64-bit register as a large
1661 // positive number. Using that large positive number as an
1662 // operand in pointer arithmetic has bad consequences.
1663 // On the other hand, 32-bit overflow is rare, and the possibility
1664 // can often be excluded, if we annotate the ConvI2L node with
1665 // a type assertion that its value is known to be a small positive
1666 // number. (The prior range check has ensured this.)
1667 // This assertion is used by ConvI2LNode::Ideal.
1668 int index_max = max_jint - 1; // array size is max_jint, index is one less
1669 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1670 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1671 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) );
1672 #endif
1673 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1674 return basic_plus_adr(ary, base, scale);
1675 }
1677 //-------------------------load_array_element-------------------------
1678 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1679 const Type* elemtype = arytype->elem();
1680 BasicType elembt = elemtype->array_element_basic_type();
1681 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1682 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1683 return ld;
1684 }
1686 //-------------------------set_arguments_for_java_call-------------------------
1687 // Arguments (pre-popped from the stack) are taken from the JVMS.
1688 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1689 // Add the call arguments:
1690 uint nargs = call->method()->arg_size();
1691 for (uint i = 0; i < nargs; i++) {
1692 Node* arg = argument(i);
1693 call->init_req(i + TypeFunc::Parms, arg);
1694 }
1695 }
1697 //---------------------------set_edges_for_java_call---------------------------
1698 // Connect a newly created call into the current JVMS.
1699 // A return value node (if any) is returned from set_edges_for_java_call.
1700 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1702 // Add the predefined inputs:
1703 call->init_req( TypeFunc::Control, control() );
1704 call->init_req( TypeFunc::I_O , i_o() );
1705 call->init_req( TypeFunc::Memory , reset_memory() );
1706 call->init_req( TypeFunc::FramePtr, frameptr() );
1707 call->init_req( TypeFunc::ReturnAdr, top() );
1709 add_safepoint_edges(call, must_throw);
1711 Node* xcall = _gvn.transform(call);
1713 if (xcall == top()) {
1714 set_control(top());
1715 return;
1716 }
1717 assert(xcall == call, "call identity is stable");
1719 // Re-use the current map to produce the result.
1721 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1722 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1723 set_all_memory_call(xcall, separate_io_proj);
1725 //return xcall; // no need, caller already has it
1726 }
1728 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1729 if (stopped()) return top(); // maybe the call folded up?
1731 // Capture the return value, if any.
1732 Node* ret;
1733 if (call->method() == NULL ||
1734 call->method()->return_type()->basic_type() == T_VOID)
1735 ret = top();
1736 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1738 // Note: Since any out-of-line call can produce an exception,
1739 // we always insert an I_O projection from the call into the result.
1741 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1743 if (separate_io_proj) {
1744 // The caller requested separate projections be used by the fall
1745 // through and exceptional paths, so replace the projections for
1746 // the fall through path.
1747 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1748 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1749 }
1750 return ret;
1751 }
1753 //--------------------set_predefined_input_for_runtime_call--------------------
1754 // Reading and setting the memory state is way conservative here.
1755 // The real problem is that I am not doing real Type analysis on memory,
1756 // so I cannot distinguish card mark stores from other stores. Across a GC
1757 // point the Store Barrier and the card mark memory has to agree. I cannot
1758 // have a card mark store and its barrier split across the GC point from
1759 // either above or below. Here I get that to happen by reading ALL of memory.
1760 // A better answer would be to separate out card marks from other memory.
1761 // For now, return the input memory state, so that it can be reused
1762 // after the call, if this call has restricted memory effects.
1763 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1764 // Set fixed predefined input arguments
1765 Node* memory = reset_memory();
1766 call->init_req( TypeFunc::Control, control() );
1767 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1768 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1769 call->init_req( TypeFunc::FramePtr, frameptr() );
1770 call->init_req( TypeFunc::ReturnAdr, top() );
1771 return memory;
1772 }
1774 //-------------------set_predefined_output_for_runtime_call--------------------
1775 // Set control and memory (not i_o) from the call.
1776 // If keep_mem is not NULL, use it for the output state,
1777 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1778 // If hook_mem is NULL, this call produces no memory effects at all.
1779 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1780 // then only that memory slice is taken from the call.
1781 // In the last case, we must put an appropriate memory barrier before
1782 // the call, so as to create the correct anti-dependencies on loads
1783 // preceding the call.
1784 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1785 Node* keep_mem,
1786 const TypePtr* hook_mem) {
1787 // no i/o
1788 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1789 if (keep_mem) {
1790 // First clone the existing memory state
1791 set_all_memory(keep_mem);
1792 if (hook_mem != NULL) {
1793 // Make memory for the call
1794 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1795 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1796 // We also use hook_mem to extract specific effects from arraycopy stubs.
1797 set_memory(mem, hook_mem);
1798 }
1799 // ...else the call has NO memory effects.
1801 // Make sure the call advertises its memory effects precisely.
1802 // This lets us build accurate anti-dependences in gcm.cpp.
1803 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1804 "call node must be constructed correctly");
1805 } else {
1806 assert(hook_mem == NULL, "");
1807 // This is not a "slow path" call; all memory comes from the call.
1808 set_all_memory_call(call);
1809 }
1810 }
1813 // Replace the call with the current state of the kit.
1814 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1815 JVMState* ejvms = NULL;
1816 if (has_exceptions()) {
1817 ejvms = transfer_exceptions_into_jvms();
1818 }
1820 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1821 ReplacedNodes replaced_nodes_exception;
1822 Node* ex_ctl = top();
1824 SafePointNode* final_state = stop();
1826 // Find all the needed outputs of this call
1827 CallProjections callprojs;
1828 call->extract_projections(&callprojs, true);
1830 Node* init_mem = call->in(TypeFunc::Memory);
1831 Node* final_mem = final_state->in(TypeFunc::Memory);
1832 Node* final_ctl = final_state->in(TypeFunc::Control);
1833 Node* final_io = final_state->in(TypeFunc::I_O);
1835 // Replace all the old call edges with the edges from the inlining result
1836 if (callprojs.fallthrough_catchproj != NULL) {
1837 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1838 }
1839 if (callprojs.fallthrough_memproj != NULL) {
1840 if (final_mem->is_MergeMem()) {
1841 // Parser's exits MergeMem was not transformed but may be optimized
1842 final_mem = _gvn.transform(final_mem);
1843 }
1844 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1845 }
1846 if (callprojs.fallthrough_ioproj != NULL) {
1847 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1848 }
1850 // Replace the result with the new result if it exists and is used
1851 if (callprojs.resproj != NULL && result != NULL) {
1852 C->gvn_replace_by(callprojs.resproj, result);
1853 }
1855 if (ejvms == NULL) {
1856 // No exception edges to simply kill off those paths
1857 if (callprojs.catchall_catchproj != NULL) {
1858 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1859 }
1860 if (callprojs.catchall_memproj != NULL) {
1861 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1862 }
1863 if (callprojs.catchall_ioproj != NULL) {
1864 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1865 }
1866 // Replace the old exception object with top
1867 if (callprojs.exobj != NULL) {
1868 C->gvn_replace_by(callprojs.exobj, C->top());
1869 }
1870 } else {
1871 GraphKit ekit(ejvms);
1873 // Load my combined exception state into the kit, with all phis transformed:
1874 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1875 replaced_nodes_exception = ex_map->replaced_nodes();
1877 Node* ex_oop = ekit.use_exception_state(ex_map);
1879 if (callprojs.catchall_catchproj != NULL) {
1880 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1881 ex_ctl = ekit.control();
1882 }
1883 if (callprojs.catchall_memproj != NULL) {
1884 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1885 }
1886 if (callprojs.catchall_ioproj != NULL) {
1887 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1888 }
1890 // Replace the old exception object with the newly created one
1891 if (callprojs.exobj != NULL) {
1892 C->gvn_replace_by(callprojs.exobj, ex_oop);
1893 }
1894 }
1896 // Disconnect the call from the graph
1897 call->disconnect_inputs(NULL, C);
1898 C->gvn_replace_by(call, C->top());
1900 // Clean up any MergeMems that feed other MergeMems since the
1901 // optimizer doesn't like that.
1902 if (final_mem->is_MergeMem()) {
1903 Node_List wl;
1904 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1905 Node* m = i.get();
1906 if (m->is_MergeMem() && !wl.contains(m)) {
1907 wl.push(m);
1908 }
1909 }
1910 while (wl.size() > 0) {
1911 _gvn.transform(wl.pop());
1912 }
1913 }
1915 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1916 replaced_nodes.apply(C, final_ctl);
1917 }
1918 if (!ex_ctl->is_top() && do_replaced_nodes) {
1919 replaced_nodes_exception.apply(C, ex_ctl);
1920 }
1921 }
1924 //------------------------------increment_counter------------------------------
1925 // for statistics: increment a VM counter by 1
1927 void GraphKit::increment_counter(address counter_addr) {
1928 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1929 increment_counter(adr1);
1930 }
1932 void GraphKit::increment_counter(Node* counter_addr) {
1933 int adr_type = Compile::AliasIdxRaw;
1934 Node* ctrl = control();
1935 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1936 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1937 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1938 }
1941 //------------------------------uncommon_trap----------------------------------
1942 // Bail out to the interpreter in mid-method. Implemented by calling the
1943 // uncommon_trap blob. This helper function inserts a runtime call with the
1944 // right debug info.
1945 void GraphKit::uncommon_trap(int trap_request,
1946 ciKlass* klass, const char* comment,
1947 bool must_throw,
1948 bool keep_exact_action) {
1949 if (failing()) stop();
1950 if (stopped()) return; // trap reachable?
1952 // Note: If ProfileTraps is true, and if a deopt. actually
1953 // occurs here, the runtime will make sure an MDO exists. There is
1954 // no need to call method()->ensure_method_data() at this point.
1956 // Set the stack pointer to the right value for reexecution:
1957 set_sp(reexecute_sp());
1959 #ifdef ASSERT
1960 if (!must_throw) {
1961 // Make sure the stack has at least enough depth to execute
1962 // the current bytecode.
1963 int inputs, ignored_depth;
1964 if (compute_stack_effects(inputs, ignored_depth)) {
1965 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1966 Bytecodes::name(java_bc()), sp(), inputs));
1967 }
1968 }
1969 #endif
1971 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1972 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1974 switch (action) {
1975 case Deoptimization::Action_maybe_recompile:
1976 case Deoptimization::Action_reinterpret:
1977 // Temporary fix for 6529811 to allow virtual calls to be sure they
1978 // get the chance to go from mono->bi->mega
1979 if (!keep_exact_action &&
1980 Deoptimization::trap_request_index(trap_request) < 0 &&
1981 too_many_recompiles(reason)) {
1982 // This BCI is causing too many recompilations.
1983 if (C->log() != NULL) {
1984 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
1985 Deoptimization::trap_reason_name(reason),
1986 Deoptimization::trap_action_name(action));
1987 }
1988 action = Deoptimization::Action_none;
1989 trap_request = Deoptimization::make_trap_request(reason, action);
1990 } else {
1991 C->set_trap_can_recompile(true);
1992 }
1993 break;
1994 case Deoptimization::Action_make_not_entrant:
1995 C->set_trap_can_recompile(true);
1996 break;
1997 #ifdef ASSERT
1998 case Deoptimization::Action_none:
1999 case Deoptimization::Action_make_not_compilable:
2000 break;
2001 default:
2002 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
2003 break;
2004 #endif
2005 }
2007 if (TraceOptoParse) {
2008 char buf[100];
2009 tty->print_cr("Uncommon trap %s at bci:%d",
2010 Deoptimization::format_trap_request(buf, sizeof(buf),
2011 trap_request), bci());
2012 }
2014 CompileLog* log = C->log();
2015 if (log != NULL) {
2016 int kid = (klass == NULL)? -1: log->identify(klass);
2017 log->begin_elem("uncommon_trap bci='%d'", bci());
2018 char buf[100];
2019 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2020 trap_request));
2021 if (kid >= 0) log->print(" klass='%d'", kid);
2022 if (comment != NULL) log->print(" comment='%s'", comment);
2023 log->end_elem();
2024 }
2026 // Make sure any guarding test views this path as very unlikely
2027 Node *i0 = control()->in(0);
2028 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2029 IfNode *iff = i0->as_If();
2030 float f = iff->_prob; // Get prob
2031 if (control()->Opcode() == Op_IfTrue) {
2032 if (f > PROB_UNLIKELY_MAG(4))
2033 iff->_prob = PROB_MIN;
2034 } else {
2035 if (f < PROB_LIKELY_MAG(4))
2036 iff->_prob = PROB_MAX;
2037 }
2038 }
2040 // Clear out dead values from the debug info.
2041 kill_dead_locals();
2043 // Now insert the uncommon trap subroutine call
2044 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2045 const TypePtr* no_memory_effects = NULL;
2046 // Pass the index of the class to be loaded
2047 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2048 (must_throw ? RC_MUST_THROW : 0),
2049 OptoRuntime::uncommon_trap_Type(),
2050 call_addr, "uncommon_trap", no_memory_effects,
2051 intcon(trap_request));
2052 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2053 "must extract request correctly from the graph");
2054 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2056 call->set_req(TypeFunc::ReturnAdr, returnadr());
2057 // The debug info is the only real input to this call.
2059 // Halt-and-catch fire here. The above call should never return!
2060 HaltNode* halt = new(C) HaltNode(control(), frameptr());
2061 _gvn.set_type_bottom(halt);
2062 root()->add_req(halt);
2064 stop_and_kill_map();
2065 }
2068 //--------------------------just_allocated_object------------------------------
2069 // Report the object that was just allocated.
2070 // It must be the case that there are no intervening safepoints.
2071 // We use this to determine if an object is so "fresh" that
2072 // it does not require card marks.
2073 Node* GraphKit::just_allocated_object(Node* current_control) {
2074 if (C->recent_alloc_ctl() == current_control)
2075 return C->recent_alloc_obj();
2076 return NULL;
2077 }
2080 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2081 // (Note: TypeFunc::make has a cache that makes this fast.)
2082 const TypeFunc* tf = TypeFunc::make(dest_method);
2083 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2084 for (int j = 0; j < nargs; j++) {
2085 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2086 if( targ->basic_type() == T_DOUBLE ) {
2087 // If any parameters are doubles, they must be rounded before
2088 // the call, dstore_rounding does gvn.transform
2089 Node *arg = argument(j);
2090 arg = dstore_rounding(arg);
2091 set_argument(j, arg);
2092 }
2093 }
2094 }
2096 /**
2097 * Record profiling data exact_kls for Node n with the type system so
2098 * that it can propagate it (speculation)
2099 *
2100 * @param n node that the type applies to
2101 * @param exact_kls type from profiling
2102 *
2103 * @return node with improved type
2104 */
2105 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) {
2106 const Type* current_type = _gvn.type(n);
2107 assert(UseTypeSpeculation, "type speculation must be on");
2109 const TypeOopPtr* speculative = current_type->speculative();
2111 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2112 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2113 const TypeOopPtr* xtype = tklass->as_instance_type();
2114 assert(xtype->klass_is_exact(), "Should be exact");
2115 // record the new speculative type's depth
2116 speculative = xtype->with_inline_depth(jvms()->depth());
2117 }
2119 if (speculative != current_type->speculative()) {
2120 // Build a type with a speculative type (what we think we know
2121 // about the type but will need a guard when we use it)
2122 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2123 // We're changing the type, we need a new CheckCast node to carry
2124 // the new type. The new type depends on the control: what
2125 // profiling tells us is only valid from here as far as we can
2126 // tell.
2127 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2128 cast = _gvn.transform(cast);
2129 replace_in_map(n, cast);
2130 n = cast;
2131 }
2133 return n;
2134 }
2136 /**
2137 * Record profiling data from receiver profiling at an invoke with the
2138 * type system so that it can propagate it (speculation)
2139 *
2140 * @param n receiver node
2141 *
2142 * @return node with improved type
2143 */
2144 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2145 if (!UseTypeSpeculation) {
2146 return n;
2147 }
2148 ciKlass* exact_kls = profile_has_unique_klass();
2149 return record_profile_for_speculation(n, exact_kls);
2150 }
2152 /**
2153 * Record profiling data from argument profiling at an invoke with the
2154 * type system so that it can propagate it (speculation)
2155 *
2156 * @param dest_method target method for the call
2157 * @param bc what invoke bytecode is this?
2158 */
2159 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2160 if (!UseTypeSpeculation) {
2161 return;
2162 }
2163 const TypeFunc* tf = TypeFunc::make(dest_method);
2164 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2165 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2166 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2167 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2168 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2169 ciKlass* better_type = method()->argument_profiled_type(bci(), i);
2170 if (better_type != NULL) {
2171 record_profile_for_speculation(argument(j), better_type);
2172 }
2173 i++;
2174 }
2175 }
2176 }
2178 /**
2179 * Record profiling data from parameter profiling at an invoke with
2180 * the type system so that it can propagate it (speculation)
2181 */
2182 void GraphKit::record_profiled_parameters_for_speculation() {
2183 if (!UseTypeSpeculation) {
2184 return;
2185 }
2186 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2187 if (_gvn.type(local(i))->isa_oopptr()) {
2188 ciKlass* better_type = method()->parameter_profiled_type(j);
2189 if (better_type != NULL) {
2190 record_profile_for_speculation(local(i), better_type);
2191 }
2192 j++;
2193 }
2194 }
2195 }
2197 void GraphKit::round_double_result(ciMethod* dest_method) {
2198 // A non-strict method may return a double value which has an extended
2199 // exponent, but this must not be visible in a caller which is 'strict'
2200 // If a strict caller invokes a non-strict callee, round a double result
2202 BasicType result_type = dest_method->return_type()->basic_type();
2203 assert( method() != NULL, "must have caller context");
2204 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2205 // Destination method's return value is on top of stack
2206 // dstore_rounding() does gvn.transform
2207 Node *result = pop_pair();
2208 result = dstore_rounding(result);
2209 push_pair(result);
2210 }
2211 }
2213 // rounding for strict float precision conformance
2214 Node* GraphKit::precision_rounding(Node* n) {
2215 return UseStrictFP && _method->flags().is_strict()
2216 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2217 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2218 : n;
2219 }
2221 // rounding for strict double precision conformance
2222 Node* GraphKit::dprecision_rounding(Node *n) {
2223 return UseStrictFP && _method->flags().is_strict()
2224 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2225 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2226 : n;
2227 }
2229 // rounding for non-strict double stores
2230 Node* GraphKit::dstore_rounding(Node* n) {
2231 return Matcher::strict_fp_requires_explicit_rounding
2232 && UseSSE <= 1
2233 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2234 : n;
2235 }
2237 //=============================================================================
2238 // Generate a fast path/slow path idiom. Graph looks like:
2239 // [foo] indicates that 'foo' is a parameter
2240 //
2241 // [in] NULL
2242 // \ /
2243 // CmpP
2244 // Bool ne
2245 // If
2246 // / \
2247 // True False-<2>
2248 // / |
2249 // / cast_not_null
2250 // Load | | ^
2251 // [fast_test] | |
2252 // gvn to opt_test | |
2253 // / \ | <1>
2254 // True False |
2255 // | \\ |
2256 // [slow_call] \[fast_result]
2257 // Ctl Val \ \
2258 // | \ \
2259 // Catch <1> \ \
2260 // / \ ^ \ \
2261 // Ex No_Ex | \ \
2262 // | \ \ | \ <2> \
2263 // ... \ [slow_res] | | \ [null_result]
2264 // \ \--+--+--- | |
2265 // \ | / \ | /
2266 // --------Region Phi
2267 //
2268 //=============================================================================
2269 // Code is structured as a series of driver functions all called 'do_XXX' that
2270 // call a set of helper functions. Helper functions first, then drivers.
2272 //------------------------------null_check_oop---------------------------------
2273 // Null check oop. Set null-path control into Region in slot 3.
2274 // Make a cast-not-nullness use the other not-null control. Return cast.
2275 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2276 bool never_see_null, bool safe_for_replace) {
2277 // Initial NULL check taken path
2278 (*null_control) = top();
2279 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2281 // Generate uncommon_trap:
2282 if (never_see_null && (*null_control) != top()) {
2283 // If we see an unexpected null at a check-cast we record it and force a
2284 // recompile; the offending check-cast will be compiled to handle NULLs.
2285 // If we see more than one offending BCI, then all checkcasts in the
2286 // method will be compiled to handle NULLs.
2287 PreserveJVMState pjvms(this);
2288 set_control(*null_control);
2289 replace_in_map(value, null());
2290 uncommon_trap(Deoptimization::Reason_null_check,
2291 Deoptimization::Action_make_not_entrant);
2292 (*null_control) = top(); // NULL path is dead
2293 }
2294 if ((*null_control) == top() && safe_for_replace) {
2295 replace_in_map(value, cast);
2296 }
2298 // Cast away null-ness on the result
2299 return cast;
2300 }
2302 //------------------------------opt_iff----------------------------------------
2303 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2304 // Return slow-path control.
2305 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2306 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2308 // Fast path taken; set region slot 2
2309 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2310 region->init_req(2,fast_taken); // Capture fast-control
2312 // Fast path not-taken, i.e. slow path
2313 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2314 return slow_taken;
2315 }
2317 //-----------------------------make_runtime_call-------------------------------
2318 Node* GraphKit::make_runtime_call(int flags,
2319 const TypeFunc* call_type, address call_addr,
2320 const char* call_name,
2321 const TypePtr* adr_type,
2322 // The following parms are all optional.
2323 // The first NULL ends the list.
2324 Node* parm0, Node* parm1,
2325 Node* parm2, Node* parm3,
2326 Node* parm4, Node* parm5,
2327 Node* parm6, Node* parm7) {
2328 // Slow-path call
2329 bool is_leaf = !(flags & RC_NO_LEAF);
2330 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2331 if (call_name == NULL) {
2332 assert(!is_leaf, "must supply name for leaf");
2333 call_name = OptoRuntime::stub_name(call_addr);
2334 }
2335 CallNode* call;
2336 if (!is_leaf) {
2337 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2338 bci(), adr_type);
2339 } else if (flags & RC_NO_FP) {
2340 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2341 } else {
2342 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2343 }
2345 // The following is similar to set_edges_for_java_call,
2346 // except that the memory effects of the call are restricted to AliasIdxRaw.
2348 // Slow path call has no side-effects, uses few values
2349 bool wide_in = !(flags & RC_NARROW_MEM);
2350 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2352 Node* prev_mem = NULL;
2353 if (wide_in) {
2354 prev_mem = set_predefined_input_for_runtime_call(call);
2355 } else {
2356 assert(!wide_out, "narrow in => narrow out");
2357 Node* narrow_mem = memory(adr_type);
2358 prev_mem = reset_memory();
2359 map()->set_memory(narrow_mem);
2360 set_predefined_input_for_runtime_call(call);
2361 }
2363 // Hook each parm in order. Stop looking at the first NULL.
2364 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2365 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2366 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2367 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2368 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2369 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2370 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2371 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2372 /* close each nested if ===> */ } } } } } } } }
2373 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2375 if (!is_leaf) {
2376 // Non-leaves can block and take safepoints:
2377 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2378 }
2379 // Non-leaves can throw exceptions:
2380 if (has_io) {
2381 call->set_req(TypeFunc::I_O, i_o());
2382 }
2384 if (flags & RC_UNCOMMON) {
2385 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2386 // (An "if" probability corresponds roughly to an unconditional count.
2387 // Sort of.)
2388 call->set_cnt(PROB_UNLIKELY_MAG(4));
2389 }
2391 Node* c = _gvn.transform(call);
2392 assert(c == call, "cannot disappear");
2394 if (wide_out) {
2395 // Slow path call has full side-effects.
2396 set_predefined_output_for_runtime_call(call);
2397 } else {
2398 // Slow path call has few side-effects, and/or sets few values.
2399 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2400 }
2402 if (has_io) {
2403 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2404 }
2405 return call;
2407 }
2409 //------------------------------merge_memory-----------------------------------
2410 // Merge memory from one path into the current memory state.
2411 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2412 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2413 Node* old_slice = mms.force_memory();
2414 Node* new_slice = mms.memory2();
2415 if (old_slice != new_slice) {
2416 PhiNode* phi;
2417 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2418 if (mms.is_empty()) {
2419 // clone base memory Phi's inputs for this memory slice
2420 assert(old_slice == mms.base_memory(), "sanity");
2421 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2422 _gvn.set_type(phi, Type::MEMORY);
2423 for (uint i = 1; i < phi->req(); i++) {
2424 phi->init_req(i, old_slice->in(i));
2425 }
2426 } else {
2427 phi = old_slice->as_Phi(); // Phi was generated already
2428 }
2429 } else {
2430 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2431 _gvn.set_type(phi, Type::MEMORY);
2432 }
2433 phi->set_req(new_path, new_slice);
2434 mms.set_memory(phi);
2435 }
2436 }
2437 }
2439 //------------------------------make_slow_call_ex------------------------------
2440 // Make the exception handler hookups for the slow call
2441 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2442 if (stopped()) return;
2444 // Make a catch node with just two handlers: fall-through and catch-all
2445 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2446 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2447 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2448 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2450 { PreserveJVMState pjvms(this);
2451 set_control(excp);
2452 set_i_o(i_o);
2454 if (excp != top()) {
2455 if (deoptimize) {
2456 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2457 uncommon_trap(Deoptimization::Reason_unhandled,
2458 Deoptimization::Action_none);
2459 } else {
2460 // Create an exception state also.
2461 // Use an exact type if the caller has specified a specific exception.
2462 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2463 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2464 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2465 }
2466 }
2467 }
2469 // Get the no-exception control from the CatchNode.
2470 set_control(norm);
2471 }
2474 //-------------------------------gen_subtype_check-----------------------------
2475 // Generate a subtyping check. Takes as input the subtype and supertype.
2476 // Returns 2 values: sets the default control() to the true path and returns
2477 // the false path. Only reads invariant memory; sets no (visible) memory.
2478 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2479 // but that's not exposed to the optimizer. This call also doesn't take in an
2480 // Object; if you wish to check an Object you need to load the Object's class
2481 // prior to coming here.
2482 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2483 // Fast check for identical types, perhaps identical constants.
2484 // The types can even be identical non-constants, in cases
2485 // involving Array.newInstance, Object.clone, etc.
2486 if (subklass == superklass)
2487 return top(); // false path is dead; no test needed.
2489 if (_gvn.type(superklass)->singleton()) {
2490 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2491 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2493 // In the common case of an exact superklass, try to fold up the
2494 // test before generating code. You may ask, why not just generate
2495 // the code and then let it fold up? The answer is that the generated
2496 // code will necessarily include null checks, which do not always
2497 // completely fold away. If they are also needless, then they turn
2498 // into a performance loss. Example:
2499 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2500 // Here, the type of 'fa' is often exact, so the store check
2501 // of fa[1]=x will fold up, without testing the nullness of x.
2502 switch (static_subtype_check(superk, subk)) {
2503 case SSC_always_false:
2504 {
2505 Node* always_fail = control();
2506 set_control(top());
2507 return always_fail;
2508 }
2509 case SSC_always_true:
2510 return top();
2511 case SSC_easy_test:
2512 {
2513 // Just do a direct pointer compare and be done.
2514 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2515 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2516 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2517 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2518 return _gvn.transform( new(C) IfFalseNode(iff) );
2519 }
2520 case SSC_full_test:
2521 break;
2522 default:
2523 ShouldNotReachHere();
2524 }
2525 }
2527 // %%% Possible further optimization: Even if the superklass is not exact,
2528 // if the subklass is the unique subtype of the superklass, the check
2529 // will always succeed. We could leave a dependency behind to ensure this.
2531 // First load the super-klass's check-offset
2532 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2533 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(),
2534 TypeInt::INT, MemNode::unordered));
2535 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2536 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2538 // Load from the sub-klass's super-class display list, or a 1-word cache of
2539 // the secondary superclass list, or a failing value with a sentinel offset
2540 // if the super-klass is an interface or exceptionally deep in the Java
2541 // hierarchy and we have to scan the secondary superclass list the hard way.
2542 // Worst-case type is a little odd: NULL is allowed as a result (usually
2543 // klass loads can never produce a NULL).
2544 Node *chk_off_X = ConvI2X(chk_off);
2545 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2546 // For some types like interfaces the following loadKlass is from a 1-word
2547 // cache which is mutable so can't use immutable memory. Other
2548 // types load from the super-class display table which is immutable.
2549 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2550 Node* nkls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2552 // Compile speed common case: ARE a subtype and we canNOT fail
2553 if( superklass == nkls )
2554 return top(); // false path is dead; no test needed.
2556 // See if we get an immediate positive hit. Happens roughly 83% of the
2557 // time. Test to see if the value loaded just previously from the subklass
2558 // is exactly the superklass.
2559 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2560 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2561 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2562 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2563 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2565 // Compile speed common case: Check for being deterministic right now. If
2566 // chk_off is a constant and not equal to cacheoff then we are NOT a
2567 // subklass. In this case we need exactly the 1 test above and we can
2568 // return those results immediately.
2569 if (!might_be_cache) {
2570 Node* not_subtype_ctrl = control();
2571 set_control(iftrue1); // We need exactly the 1 test above
2572 return not_subtype_ctrl;
2573 }
2575 // Gather the various success & failures here
2576 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2577 record_for_igvn(r_ok_subtype);
2578 RegionNode *r_not_subtype = new (C) RegionNode(3);
2579 record_for_igvn(r_not_subtype);
2581 r_ok_subtype->init_req(1, iftrue1);
2583 // Check for immediate negative hit. Happens roughly 11% of the time (which
2584 // is roughly 63% of the remaining cases). Test to see if the loaded
2585 // check-offset points into the subklass display list or the 1-element
2586 // cache. If it points to the display (and NOT the cache) and the display
2587 // missed then it's not a subtype.
2588 Node *cacheoff = _gvn.intcon(cacheoff_con);
2589 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2590 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2591 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2592 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2593 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2595 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2596 // No performance impact (too rare) but allows sharing of secondary arrays
2597 // which has some footprint reduction.
2598 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2599 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2600 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2601 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2602 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2604 // -- Roads not taken here: --
2605 // We could also have chosen to perform the self-check at the beginning
2606 // of this code sequence, as the assembler does. This would not pay off
2607 // the same way, since the optimizer, unlike the assembler, can perform
2608 // static type analysis to fold away many successful self-checks.
2609 // Non-foldable self checks work better here in second position, because
2610 // the initial primary superclass check subsumes a self-check for most
2611 // types. An exception would be a secondary type like array-of-interface,
2612 // which does not appear in its own primary supertype display.
2613 // Finally, we could have chosen to move the self-check into the
2614 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2615 // dependent manner. But it is worthwhile to have the check here,
2616 // where it can be perhaps be optimized. The cost in code space is
2617 // small (register compare, branch).
2619 // Now do a linear scan of the secondary super-klass array. Again, no real
2620 // performance impact (too rare) but it's gotta be done.
2621 // Since the code is rarely used, there is no penalty for moving it
2622 // out of line, and it can only improve I-cache density.
2623 // The decision to inline or out-of-line this final check is platform
2624 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2625 Node* psc = _gvn.transform(
2626 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2628 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2629 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2630 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2631 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2632 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2634 // Return false path; set default control to true path.
2635 set_control( _gvn.transform(r_ok_subtype) );
2636 return _gvn.transform(r_not_subtype);
2637 }
2639 //----------------------------static_subtype_check-----------------------------
2640 // Shortcut important common cases when superklass is exact:
2641 // (0) superklass is java.lang.Object (can occur in reflective code)
2642 // (1) subklass is already limited to a subtype of superklass => always ok
2643 // (2) subklass does not overlap with superklass => always fail
2644 // (3) superklass has NO subtypes and we can check with a simple compare.
2645 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2646 if (StressReflectiveCode) {
2647 return SSC_full_test; // Let caller generate the general case.
2648 }
2650 if (superk == env()->Object_klass()) {
2651 return SSC_always_true; // (0) this test cannot fail
2652 }
2654 ciType* superelem = superk;
2655 if (superelem->is_array_klass())
2656 superelem = superelem->as_array_klass()->base_element_type();
2658 if (!subk->is_interface()) { // cannot trust static interface types yet
2659 if (subk->is_subtype_of(superk)) {
2660 return SSC_always_true; // (1) false path dead; no dynamic test needed
2661 }
2662 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2663 !superk->is_subtype_of(subk)) {
2664 return SSC_always_false;
2665 }
2666 }
2668 // If casting to an instance klass, it must have no subtypes
2669 if (superk->is_interface()) {
2670 // Cannot trust interfaces yet.
2671 // %%% S.B. superk->nof_implementors() == 1
2672 } else if (superelem->is_instance_klass()) {
2673 ciInstanceKlass* ik = superelem->as_instance_klass();
2674 if (!ik->has_subklass() && !ik->is_interface()) {
2675 if (!ik->is_final()) {
2676 // Add a dependency if there is a chance of a later subclass.
2677 C->dependencies()->assert_leaf_type(ik);
2678 }
2679 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2680 }
2681 } else {
2682 // A primitive array type has no subtypes.
2683 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2684 }
2686 return SSC_full_test;
2687 }
2689 // Profile-driven exact type check:
2690 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2691 float prob,
2692 Node* *casted_receiver) {
2693 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2694 Node* recv_klass = load_object_klass(receiver);
2695 Node* want_klass = makecon(tklass);
2696 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2697 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2698 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2699 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2700 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2702 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2703 assert(recv_xtype->klass_is_exact(), "");
2705 // Subsume downstream occurrences of receiver with a cast to
2706 // recv_xtype, since now we know what the type will be.
2707 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2708 (*casted_receiver) = _gvn.transform(cast);
2709 // (User must make the replace_in_map call.)
2711 return fail;
2712 }
2715 //------------------------------seems_never_null-------------------------------
2716 // Use null_seen information if it is available from the profile.
2717 // If we see an unexpected null at a type check we record it and force a
2718 // recompile; the offending check will be recompiled to handle NULLs.
2719 // If we see several offending BCIs, then all checks in the
2720 // method will be recompiled.
2721 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2722 if (UncommonNullCast // Cutout for this technique
2723 && obj != null() // And not the -Xcomp stupid case?
2724 && !too_many_traps(Deoptimization::Reason_null_check)
2725 ) {
2726 if (data == NULL)
2727 // Edge case: no mature data. Be optimistic here.
2728 return true;
2729 // If the profile has not seen a null, assume it won't happen.
2730 assert(java_bc() == Bytecodes::_checkcast ||
2731 java_bc() == Bytecodes::_instanceof ||
2732 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2733 return !data->as_BitData()->null_seen();
2734 }
2735 return false;
2736 }
2738 //------------------------maybe_cast_profiled_receiver-------------------------
2739 // If the profile has seen exactly one type, narrow to exactly that type.
2740 // Subsequent type checks will always fold up.
2741 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2742 ciKlass* require_klass,
2743 ciKlass* spec_klass,
2744 bool safe_for_replace) {
2745 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2747 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check;
2749 // Make sure we haven't already deoptimized from this tactic.
2750 if (too_many_traps(reason) || too_many_recompiles(reason))
2751 return NULL;
2753 // (No, this isn't a call, but it's enough like a virtual call
2754 // to use the same ciMethod accessor to get the profile info...)
2755 // If we have a speculative type use it instead of profiling (which
2756 // may not help us)
2757 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2758 if (exact_kls != NULL) {// no cast failures here
2759 if (require_klass == NULL ||
2760 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2761 // If we narrow the type to match what the type profile sees or
2762 // the speculative type, we can then remove the rest of the
2763 // cast.
2764 // This is a win, even if the exact_kls is very specific,
2765 // because downstream operations, such as method calls,
2766 // will often benefit from the sharper type.
2767 Node* exact_obj = not_null_obj; // will get updated in place...
2768 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2769 &exact_obj);
2770 { PreserveJVMState pjvms(this);
2771 set_control(slow_ctl);
2772 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2773 }
2774 if (safe_for_replace) {
2775 replace_in_map(not_null_obj, exact_obj);
2776 }
2777 return exact_obj;
2778 }
2779 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2780 }
2782 return NULL;
2783 }
2785 /**
2786 * Cast obj to type and emit guard unless we had too many traps here
2787 * already
2788 *
2789 * @param obj node being casted
2790 * @param type type to cast the node to
2791 * @param not_null true if we know node cannot be null
2792 */
2793 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2794 ciKlass* type,
2795 bool not_null) {
2796 // type == NULL if profiling tells us this object is always null
2797 if (type != NULL) {
2798 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2799 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check;
2800 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2801 !too_many_traps(class_reason) && !too_many_recompiles(class_reason)) {
2802 Node* not_null_obj = NULL;
2803 // not_null is true if we know the object is not null and
2804 // there's no need for a null check
2805 if (!not_null) {
2806 Node* null_ctl = top();
2807 not_null_obj = null_check_oop(obj, &null_ctl, true, true);
2808 assert(null_ctl->is_top(), "no null control here");
2809 } else {
2810 not_null_obj = obj;
2811 }
2813 Node* exact_obj = not_null_obj;
2814 ciKlass* exact_kls = type;
2815 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2816 &exact_obj);
2817 {
2818 PreserveJVMState pjvms(this);
2819 set_control(slow_ctl);
2820 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2821 }
2822 replace_in_map(not_null_obj, exact_obj);
2823 obj = exact_obj;
2824 }
2825 } else {
2826 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2827 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2828 Node* exact_obj = null_assert(obj);
2829 replace_in_map(obj, exact_obj);
2830 obj = exact_obj;
2831 }
2832 }
2833 return obj;
2834 }
2836 //-------------------------------gen_instanceof--------------------------------
2837 // Generate an instance-of idiom. Used by both the instance-of bytecode
2838 // and the reflective instance-of call.
2839 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2840 kill_dead_locals(); // Benefit all the uncommon traps
2841 assert( !stopped(), "dead parse path should be checked in callers" );
2842 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2843 "must check for not-null not-dead klass in callers");
2845 // Make the merge point
2846 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2847 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2848 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2849 C->set_has_split_ifs(true); // Has chance for split-if optimization
2851 ciProfileData* data = NULL;
2852 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2853 data = method()->method_data()->bci_to_data(bci());
2854 }
2855 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2856 && seems_never_null(obj, data));
2858 // Null check; get casted pointer; set region slot 3
2859 Node* null_ctl = top();
2860 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2862 // If not_null_obj is dead, only null-path is taken
2863 if (stopped()) { // Doing instance-of on a NULL?
2864 set_control(null_ctl);
2865 return intcon(0);
2866 }
2867 region->init_req(_null_path, null_ctl);
2868 phi ->init_req(_null_path, intcon(0)); // Set null path value
2869 if (null_ctl == top()) {
2870 // Do this eagerly, so that pattern matches like is_diamond_phi
2871 // will work even during parsing.
2872 assert(_null_path == PATH_LIMIT-1, "delete last");
2873 region->del_req(_null_path);
2874 phi ->del_req(_null_path);
2875 }
2877 // Do we know the type check always succeed?
2878 bool known_statically = false;
2879 if (_gvn.type(superklass)->singleton()) {
2880 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2881 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2882 if (subk != NULL && subk->is_loaded()) {
2883 int static_res = static_subtype_check(superk, subk);
2884 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false);
2885 }
2886 }
2888 if (known_statically && UseTypeSpeculation) {
2889 // If we know the type check always succeeds then we don't use the
2890 // profiling data at this bytecode. Don't lose it, feed it to the
2891 // type system as a speculative type.
2892 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2893 } else {
2894 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2895 // We may not have profiling here or it may not help us. If we
2896 // have a speculative type use it to perform an exact cast.
2897 ciKlass* spec_obj_type = obj_type->speculative_type();
2898 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2899 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2900 if (stopped()) { // Profile disagrees with this path.
2901 set_control(null_ctl); // Null is the only remaining possibility.
2902 return intcon(0);
2903 }
2904 if (cast_obj != NULL) {
2905 not_null_obj = cast_obj;
2906 }
2907 }
2908 }
2910 // Load the object's klass
2911 Node* obj_klass = load_object_klass(not_null_obj);
2913 // Generate the subtype check
2914 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2916 // Plug in the success path to the general merge in slot 1.
2917 region->init_req(_obj_path, control());
2918 phi ->init_req(_obj_path, intcon(1));
2920 // Plug in the failing path to the general merge in slot 2.
2921 region->init_req(_fail_path, not_subtype_ctrl);
2922 phi ->init_req(_fail_path, intcon(0));
2924 // Return final merged results
2925 set_control( _gvn.transform(region) );
2926 record_for_igvn(region);
2927 return _gvn.transform(phi);
2928 }
2930 //-------------------------------gen_checkcast---------------------------------
2931 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2932 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2933 // uncommon-trap paths work. Adjust stack after this call.
2934 // If failure_control is supplied and not null, it is filled in with
2935 // the control edge for the cast failure. Otherwise, an appropriate
2936 // uncommon trap or exception is thrown.
2937 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2938 Node* *failure_control) {
2939 kill_dead_locals(); // Benefit all the uncommon traps
2940 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2941 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2943 // Fast cutout: Check the case that the cast is vacuously true.
2944 // This detects the common cases where the test will short-circuit
2945 // away completely. We do this before we perform the null check,
2946 // because if the test is going to turn into zero code, we don't
2947 // want a residual null check left around. (Causes a slowdown,
2948 // for example, in some objArray manipulations, such as a[i]=a[j].)
2949 if (tk->singleton()) {
2950 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2951 if (objtp != NULL && objtp->klass() != NULL) {
2952 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2953 case SSC_always_true:
2954 // If we know the type check always succeed then we don't use
2955 // the profiling data at this bytecode. Don't lose it, feed it
2956 // to the type system as a speculative type.
2957 return record_profiled_receiver_for_speculation(obj);
2958 case SSC_always_false:
2959 // It needs a null check because a null will *pass* the cast check.
2960 // A non-null value will always produce an exception.
2961 return null_assert(obj);
2962 }
2963 }
2964 }
2966 ciProfileData* data = NULL;
2967 bool safe_for_replace = false;
2968 if (failure_control == NULL) { // use MDO in regular case only
2969 assert(java_bc() == Bytecodes::_aastore ||
2970 java_bc() == Bytecodes::_checkcast,
2971 "interpreter profiles type checks only for these BCs");
2972 data = method()->method_data()->bci_to_data(bci());
2973 safe_for_replace = true;
2974 }
2976 // Make the merge point
2977 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2978 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
2979 Node* phi = new (C) PhiNode(region, toop);
2980 C->set_has_split_ifs(true); // Has chance for split-if optimization
2982 // Use null-cast information if it is available
2983 bool never_see_null = ((failure_control == NULL) // regular case only
2984 && seems_never_null(obj, data));
2986 // Null check; get casted pointer; set region slot 3
2987 Node* null_ctl = top();
2988 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2990 // If not_null_obj is dead, only null-path is taken
2991 if (stopped()) { // Doing instance-of on a NULL?
2992 set_control(null_ctl);
2993 return null();
2994 }
2995 region->init_req(_null_path, null_ctl);
2996 phi ->init_req(_null_path, null()); // Set null path value
2997 if (null_ctl == top()) {
2998 // Do this eagerly, so that pattern matches like is_diamond_phi
2999 // will work even during parsing.
3000 assert(_null_path == PATH_LIMIT-1, "delete last");
3001 region->del_req(_null_path);
3002 phi ->del_req(_null_path);
3003 }
3005 Node* cast_obj = NULL;
3006 if (tk->klass_is_exact()) {
3007 // The following optimization tries to statically cast the speculative type of the object
3008 // (for example obtained during profiling) to the type of the superklass and then do a
3009 // dynamic check that the type of the object is what we expect. To work correctly
3010 // for checkcast and aastore the type of superklass should be exact.
3011 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3012 // We may not have profiling here or it may not help us. If we have
3013 // a speculative type use it to perform an exact cast.
3014 ciKlass* spec_obj_type = obj_type->speculative_type();
3015 if (spec_obj_type != NULL ||
3016 (data != NULL &&
3017 // Counter has never been decremented (due to cast failure).
3018 // ...This is a reasonable thing to expect. It is true of
3019 // all casts inserted by javac to implement generic types.
3020 data->as_CounterData()->count() >= 0)) {
3021 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3022 if (cast_obj != NULL) {
3023 if (failure_control != NULL) // failure is now impossible
3024 (*failure_control) = top();
3025 // adjust the type of the phi to the exact klass:
3026 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3027 }
3028 }
3029 }
3031 if (cast_obj == NULL) {
3032 // Load the object's klass
3033 Node* obj_klass = load_object_klass(not_null_obj);
3035 // Generate the subtype check
3036 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3038 // Plug in success path into the merge
3039 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
3040 not_null_obj, toop));
3041 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3042 if (failure_control == NULL) {
3043 if (not_subtype_ctrl != top()) { // If failure is possible
3044 PreserveJVMState pjvms(this);
3045 set_control(not_subtype_ctrl);
3046 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3047 }
3048 } else {
3049 (*failure_control) = not_subtype_ctrl;
3050 }
3051 }
3053 region->init_req(_obj_path, control());
3054 phi ->init_req(_obj_path, cast_obj);
3056 // A merge of NULL or Casted-NotNull obj
3057 Node* res = _gvn.transform(phi);
3059 // Note I do NOT always 'replace_in_map(obj,result)' here.
3060 // if( tk->klass()->can_be_primary_super() )
3061 // This means that if I successfully store an Object into an array-of-String
3062 // I 'forget' that the Object is really now known to be a String. I have to
3063 // do this because we don't have true union types for interfaces - if I store
3064 // a Baz into an array-of-Interface and then tell the optimizer it's an
3065 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3066 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3067 // replace_in_map( obj, res );
3069 // Return final merged results
3070 set_control( _gvn.transform(region) );
3071 record_for_igvn(region);
3072 return res;
3073 }
3075 //------------------------------next_monitor-----------------------------------
3076 // What number should be given to the next monitor?
3077 int GraphKit::next_monitor() {
3078 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3079 int next = current + C->sync_stack_slots();
3080 // Keep the toplevel high water mark current:
3081 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3082 return current;
3083 }
3085 //------------------------------insert_mem_bar---------------------------------
3086 // Memory barrier to avoid floating things around
3087 // The membar serves as a pinch point between both control and all memory slices.
3088 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3089 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3090 mb->init_req(TypeFunc::Control, control());
3091 mb->init_req(TypeFunc::Memory, reset_memory());
3092 Node* membar = _gvn.transform(mb);
3093 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3094 set_all_memory_call(membar);
3095 return membar;
3096 }
3098 //-------------------------insert_mem_bar_volatile----------------------------
3099 // Memory barrier to avoid floating things around
3100 // The membar serves as a pinch point between both control and memory(alias_idx).
3101 // If you want to make a pinch point on all memory slices, do not use this
3102 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3103 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3104 // When Parse::do_put_xxx updates a volatile field, it appends a series
3105 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3106 // The first membar is on the same memory slice as the field store opcode.
3107 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3108 // All the other membars (for other volatile slices, including AliasIdxBot,
3109 // which stands for all unknown volatile slices) are control-dependent
3110 // on the first membar. This prevents later volatile loads or stores
3111 // from sliding up past the just-emitted store.
3113 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3114 mb->set_req(TypeFunc::Control,control());
3115 if (alias_idx == Compile::AliasIdxBot) {
3116 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3117 } else {
3118 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3119 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3120 }
3121 Node* membar = _gvn.transform(mb);
3122 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3123 if (alias_idx == Compile::AliasIdxBot) {
3124 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
3125 } else {
3126 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
3127 }
3128 return membar;
3129 }
3131 //------------------------------shared_lock------------------------------------
3132 // Emit locking code.
3133 FastLockNode* GraphKit::shared_lock(Node* obj) {
3134 // bci is either a monitorenter bc or InvocationEntryBci
3135 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3136 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3138 if( !GenerateSynchronizationCode )
3139 return NULL; // Not locking things?
3140 if (stopped()) // Dead monitor?
3141 return NULL;
3143 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3145 // Box the stack location
3146 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
3147 Node* mem = reset_memory();
3149 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
3150 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3151 // Create the counters for this fast lock.
3152 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3153 }
3155 // Create the rtm counters for this fast lock if needed.
3156 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3158 // Add monitor to debug info for the slow path. If we block inside the
3159 // slow path and de-opt, we need the monitor hanging around
3160 map()->push_monitor( flock );
3162 const TypeFunc *tf = LockNode::lock_type();
3163 LockNode *lock = new (C) LockNode(C, tf);
3165 lock->init_req( TypeFunc::Control, control() );
3166 lock->init_req( TypeFunc::Memory , mem );
3167 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3168 lock->init_req( TypeFunc::FramePtr, frameptr() );
3169 lock->init_req( TypeFunc::ReturnAdr, top() );
3171 lock->init_req(TypeFunc::Parms + 0, obj);
3172 lock->init_req(TypeFunc::Parms + 1, box);
3173 lock->init_req(TypeFunc::Parms + 2, flock);
3174 add_safepoint_edges(lock);
3176 lock = _gvn.transform( lock )->as_Lock();
3178 // lock has no side-effects, sets few values
3179 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3181 insert_mem_bar(Op_MemBarAcquireLock);
3183 // Add this to the worklist so that the lock can be eliminated
3184 record_for_igvn(lock);
3186 #ifndef PRODUCT
3187 if (PrintLockStatistics) {
3188 // Update the counter for this lock. Don't bother using an atomic
3189 // operation since we don't require absolute accuracy.
3190 lock->create_lock_counter(map()->jvms());
3191 increment_counter(lock->counter()->addr());
3192 }
3193 #endif
3195 return flock;
3196 }
3199 //------------------------------shared_unlock----------------------------------
3200 // Emit unlocking code.
3201 void GraphKit::shared_unlock(Node* box, Node* obj) {
3202 // bci is either a monitorenter bc or InvocationEntryBci
3203 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3204 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3206 if( !GenerateSynchronizationCode )
3207 return;
3208 if (stopped()) { // Dead monitor?
3209 map()->pop_monitor(); // Kill monitor from debug info
3210 return;
3211 }
3213 // Memory barrier to avoid floating things down past the locked region
3214 insert_mem_bar(Op_MemBarReleaseLock);
3216 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3217 UnlockNode *unlock = new (C) UnlockNode(C, tf);
3218 #ifdef ASSERT
3219 unlock->set_dbg_jvms(sync_jvms());
3220 #endif
3221 uint raw_idx = Compile::AliasIdxRaw;
3222 unlock->init_req( TypeFunc::Control, control() );
3223 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3224 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3225 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3226 unlock->init_req( TypeFunc::ReturnAdr, top() );
3228 unlock->init_req(TypeFunc::Parms + 0, obj);
3229 unlock->init_req(TypeFunc::Parms + 1, box);
3230 unlock = _gvn.transform(unlock)->as_Unlock();
3232 Node* mem = reset_memory();
3234 // unlock has no side-effects, sets few values
3235 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3237 // Kill monitor from debug info
3238 map()->pop_monitor( );
3239 }
3241 //-------------------------------get_layout_helper-----------------------------
3242 // If the given klass is a constant or known to be an array,
3243 // fetch the constant layout helper value into constant_value
3244 // and return (Node*)NULL. Otherwise, load the non-constant
3245 // layout helper value, and return the node which represents it.
3246 // This two-faced routine is useful because allocation sites
3247 // almost always feature constant types.
3248 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3249 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3250 if (!StressReflectiveCode && inst_klass != NULL) {
3251 ciKlass* klass = inst_klass->klass();
3252 bool xklass = inst_klass->klass_is_exact();
3253 if (xklass || klass->is_array_klass()) {
3254 jint lhelper = klass->layout_helper();
3255 if (lhelper != Klass::_lh_neutral_value) {
3256 constant_value = lhelper;
3257 return (Node*) NULL;
3258 }
3259 }
3260 }
3261 constant_value = Klass::_lh_neutral_value; // put in a known value
3262 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3263 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3264 }
3266 // We just put in an allocate/initialize with a big raw-memory effect.
3267 // Hook selected additional alias categories on the initialization.
3268 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3269 MergeMemNode* init_in_merge,
3270 Node* init_out_raw) {
3271 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3272 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3274 Node* prevmem = kit.memory(alias_idx);
3275 init_in_merge->set_memory_at(alias_idx, prevmem);
3276 kit.set_memory(init_out_raw, alias_idx);
3277 }
3279 //---------------------------set_output_for_allocation-------------------------
3280 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3281 const TypeOopPtr* oop_type,
3282 bool deoptimize_on_exception) {
3283 int rawidx = Compile::AliasIdxRaw;
3284 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3285 add_safepoint_edges(alloc);
3286 Node* allocx = _gvn.transform(alloc);
3287 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3288 // create memory projection for i_o
3289 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3290 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3292 // create a memory projection as for the normal control path
3293 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3294 set_memory(malloc, rawidx);
3296 // a normal slow-call doesn't change i_o, but an allocation does
3297 // we create a separate i_o projection for the normal control path
3298 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3299 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3301 // put in an initialization barrier
3302 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3303 rawoop)->as_Initialize();
3304 assert(alloc->initialization() == init, "2-way macro link must work");
3305 assert(init ->allocation() == alloc, "2-way macro link must work");
3306 {
3307 // Extract memory strands which may participate in the new object's
3308 // initialization, and source them from the new InitializeNode.
3309 // This will allow us to observe initializations when they occur,
3310 // and link them properly (as a group) to the InitializeNode.
3311 assert(init->in(InitializeNode::Memory) == malloc, "");
3312 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3313 init->set_req(InitializeNode::Memory, minit_in);
3314 record_for_igvn(minit_in); // fold it up later, if possible
3315 Node* minit_out = memory(rawidx);
3316 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3317 if (oop_type->isa_aryptr()) {
3318 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3319 int elemidx = C->get_alias_index(telemref);
3320 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3321 } else if (oop_type->isa_instptr()) {
3322 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3323 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3324 ciField* field = ik->nonstatic_field_at(i);
3325 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3326 continue; // do not bother to track really large numbers of fields
3327 // Find (or create) the alias category for this field:
3328 int fieldidx = C->alias_type(field)->index();
3329 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3330 }
3331 }
3332 }
3334 // Cast raw oop to the real thing...
3335 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3336 javaoop = _gvn.transform(javaoop);
3337 C->set_recent_alloc(control(), javaoop);
3338 assert(just_allocated_object(control()) == javaoop, "just allocated");
3340 #ifdef ASSERT
3341 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3342 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3343 "Ideal_allocation works");
3344 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3345 "Ideal_allocation works");
3346 if (alloc->is_AllocateArray()) {
3347 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3348 "Ideal_allocation works");
3349 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3350 "Ideal_allocation works");
3351 } else {
3352 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3353 }
3354 }
3355 #endif //ASSERT
3357 return javaoop;
3358 }
3360 //---------------------------new_instance--------------------------------------
3361 // This routine takes a klass_node which may be constant (for a static type)
3362 // or may be non-constant (for reflective code). It will work equally well
3363 // for either, and the graph will fold nicely if the optimizer later reduces
3364 // the type to a constant.
3365 // The optional arguments are for specialized use by intrinsics:
3366 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3367 // - If 'return_size_val', report the the total object size to the caller.
3368 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3369 Node* GraphKit::new_instance(Node* klass_node,
3370 Node* extra_slow_test,
3371 Node* *return_size_val,
3372 bool deoptimize_on_exception) {
3373 // Compute size in doublewords
3374 // The size is always an integral number of doublewords, represented
3375 // as a positive bytewise size stored in the klass's layout_helper.
3376 // The layout_helper also encodes (in a low bit) the need for a slow path.
3377 jint layout_con = Klass::_lh_neutral_value;
3378 Node* layout_val = get_layout_helper(klass_node, layout_con);
3379 int layout_is_con = (layout_val == NULL);
3381 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3382 // Generate the initial go-slow test. It's either ALWAYS (return a
3383 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3384 // case) a computed value derived from the layout_helper.
3385 Node* initial_slow_test = NULL;
3386 if (layout_is_con) {
3387 assert(!StressReflectiveCode, "stress mode does not use these paths");
3388 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3389 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3391 } else { // reflective case
3392 // This reflective path is used by Unsafe.allocateInstance.
3393 // (It may be stress-tested by specifying StressReflectiveCode.)
3394 // Basically, we want to get into the VM is there's an illegal argument.
3395 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3396 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3397 if (extra_slow_test != intcon(0)) {
3398 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3399 }
3400 // (Macro-expander will further convert this to a Bool, if necessary.)
3401 }
3403 // Find the size in bytes. This is easy; it's the layout_helper.
3404 // The size value must be valid even if the slow path is taken.
3405 Node* size = NULL;
3406 if (layout_is_con) {
3407 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3408 } else { // reflective case
3409 // This reflective path is used by clone and Unsafe.allocateInstance.
3410 size = ConvI2X(layout_val);
3412 // Clear the low bits to extract layout_helper_size_in_bytes:
3413 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3414 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3415 size = _gvn.transform( new (C) AndXNode(size, mask) );
3416 }
3417 if (return_size_val != NULL) {
3418 (*return_size_val) = size;
3419 }
3421 // This is a precise notnull oop of the klass.
3422 // (Actually, it need not be precise if this is a reflective allocation.)
3423 // It's what we cast the result to.
3424 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3425 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3426 const TypeOopPtr* oop_type = tklass->as_instance_type();
3428 // Now generate allocation code
3430 // The entire memory state is needed for slow path of the allocation
3431 // since GC and deoptimization can happened.
3432 Node *mem = reset_memory();
3433 set_all_memory(mem); // Create new memory state
3435 AllocateNode* alloc
3436 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3437 control(), mem, i_o(),
3438 size, klass_node,
3439 initial_slow_test);
3441 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3442 }
3444 //-------------------------------new_array-------------------------------------
3445 // helper for both newarray and anewarray
3446 // The 'length' parameter is (obviously) the length of the array.
3447 // See comments on new_instance for the meaning of the other arguments.
3448 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3449 Node* length, // number of array elements
3450 int nargs, // number of arguments to push back for uncommon trap
3451 Node* *return_size_val,
3452 bool deoptimize_on_exception) {
3453 jint layout_con = Klass::_lh_neutral_value;
3454 Node* layout_val = get_layout_helper(klass_node, layout_con);
3455 int layout_is_con = (layout_val == NULL);
3457 if (!layout_is_con && !StressReflectiveCode &&
3458 !too_many_traps(Deoptimization::Reason_class_check)) {
3459 // This is a reflective array creation site.
3460 // Optimistically assume that it is a subtype of Object[],
3461 // so that we can fold up all the address arithmetic.
3462 layout_con = Klass::array_layout_helper(T_OBJECT);
3463 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3464 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3465 { BuildCutout unless(this, bol_lh, PROB_MAX);
3466 inc_sp(nargs);
3467 uncommon_trap(Deoptimization::Reason_class_check,
3468 Deoptimization::Action_maybe_recompile);
3469 }
3470 layout_val = NULL;
3471 layout_is_con = true;
3472 }
3474 // Generate the initial go-slow test. Make sure we do not overflow
3475 // if length is huge (near 2Gig) or negative! We do not need
3476 // exact double-words here, just a close approximation of needed
3477 // double-words. We can't add any offset or rounding bits, lest we
3478 // take a size -1 of bytes and make it positive. Use an unsigned
3479 // compare, so negative sizes look hugely positive.
3480 int fast_size_limit = FastAllocateSizeLimit;
3481 if (layout_is_con) {
3482 assert(!StressReflectiveCode, "stress mode does not use these paths");
3483 // Increase the size limit if we have exact knowledge of array type.
3484 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3485 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3486 }
3488 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3489 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3490 if (initial_slow_test->is_Bool()) {
3491 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3492 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3493 }
3495 // --- Size Computation ---
3496 // array_size = round_to_heap(array_header + (length << elem_shift));
3497 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3498 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3499 // The rounding mask is strength-reduced, if possible.
3500 int round_mask = MinObjAlignmentInBytes - 1;
3501 Node* header_size = NULL;
3502 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3503 // (T_BYTE has the weakest alignment and size restrictions...)
3504 if (layout_is_con) {
3505 int hsize = Klass::layout_helper_header_size(layout_con);
3506 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3507 BasicType etype = Klass::layout_helper_element_type(layout_con);
3508 if ((round_mask & ~right_n_bits(eshift)) == 0)
3509 round_mask = 0; // strength-reduce it if it goes away completely
3510 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3511 assert(header_size_min <= hsize, "generic minimum is smallest");
3512 header_size_min = hsize;
3513 header_size = intcon(hsize + round_mask);
3514 } else {
3515 Node* hss = intcon(Klass::_lh_header_size_shift);
3516 Node* hsm = intcon(Klass::_lh_header_size_mask);
3517 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3518 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3519 Node* mask = intcon(round_mask);
3520 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3521 }
3523 Node* elem_shift = NULL;
3524 if (layout_is_con) {
3525 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3526 if (eshift != 0)
3527 elem_shift = intcon(eshift);
3528 } else {
3529 // There is no need to mask or shift this value.
3530 // The semantics of LShiftINode include an implicit mask to 0x1F.
3531 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3532 elem_shift = layout_val;
3533 }
3535 // Transition to native address size for all offset calculations:
3536 Node* lengthx = ConvI2X(length);
3537 Node* headerx = ConvI2X(header_size);
3538 #ifdef _LP64
3539 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3540 if (tllen != NULL && tllen->_lo < 0) {
3541 // Add a manual constraint to a positive range. Cf. array_element_address.
3542 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3543 if (size_max > tllen->_hi) size_max = tllen->_hi;
3544 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3545 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon));
3546 }
3547 }
3548 #endif
3550 // Combine header size (plus rounding) and body size. Then round down.
3551 // This computation cannot overflow, because it is used only in two
3552 // places, one where the length is sharply limited, and the other
3553 // after a successful allocation.
3554 Node* abody = lengthx;
3555 if (elem_shift != NULL)
3556 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3557 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3558 if (round_mask != 0) {
3559 Node* mask = MakeConX(~round_mask);
3560 size = _gvn.transform( new(C) AndXNode(size, mask) );
3561 }
3562 // else if round_mask == 0, the size computation is self-rounding
3564 if (return_size_val != NULL) {
3565 // This is the size
3566 (*return_size_val) = size;
3567 }
3569 // Now generate allocation code
3571 // The entire memory state is needed for slow path of the allocation
3572 // since GC and deoptimization can happened.
3573 Node *mem = reset_memory();
3574 set_all_memory(mem); // Create new memory state
3576 // Create the AllocateArrayNode and its result projections
3577 AllocateArrayNode* alloc
3578 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3579 control(), mem, i_o(),
3580 size, klass_node,
3581 initial_slow_test,
3582 length);
3584 // Cast to correct type. Note that the klass_node may be constant or not,
3585 // and in the latter case the actual array type will be inexact also.
3586 // (This happens via a non-constant argument to inline_native_newArray.)
3587 // In any case, the value of klass_node provides the desired array type.
3588 const TypeInt* length_type = _gvn.find_int_type(length);
3589 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3590 if (ary_type->isa_aryptr() && length_type != NULL) {
3591 // Try to get a better type than POS for the size
3592 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3593 }
3595 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3597 // Cast length on remaining path to be as narrow as possible
3598 if (map()->find_edge(length) >= 0) {
3599 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3600 if (ccast != length) {
3601 _gvn.set_type_bottom(ccast);
3602 record_for_igvn(ccast);
3603 replace_in_map(length, ccast);
3604 }
3605 }
3607 return javaoop;
3608 }
3610 // The following "Ideal_foo" functions are placed here because they recognize
3611 // the graph shapes created by the functions immediately above.
3613 //---------------------------Ideal_allocation----------------------------------
3614 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3615 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3616 if (ptr == NULL) { // reduce dumb test in callers
3617 return NULL;
3618 }
3619 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3620 ptr = ptr->in(1);
3621 if (ptr == NULL) return NULL;
3622 }
3623 // Return NULL for allocations with several casts:
3624 // j.l.reflect.Array.newInstance(jobject, jint)
3625 // Object.clone()
3626 // to keep more precise type from last cast.
3627 if (ptr->is_Proj()) {
3628 Node* allo = ptr->in(0);
3629 if (allo != NULL && allo->is_Allocate()) {
3630 return allo->as_Allocate();
3631 }
3632 }
3633 // Report failure to match.
3634 return NULL;
3635 }
3637 // Fancy version which also strips off an offset (and reports it to caller).
3638 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3639 intptr_t& offset) {
3640 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3641 if (base == NULL) return NULL;
3642 return Ideal_allocation(base, phase);
3643 }
3645 // Trace Initialize <- Proj[Parm] <- Allocate
3646 AllocateNode* InitializeNode::allocation() {
3647 Node* rawoop = in(InitializeNode::RawAddress);
3648 if (rawoop->is_Proj()) {
3649 Node* alloc = rawoop->in(0);
3650 if (alloc->is_Allocate()) {
3651 return alloc->as_Allocate();
3652 }
3653 }
3654 return NULL;
3655 }
3657 // Trace Allocate -> Proj[Parm] -> Initialize
3658 InitializeNode* AllocateNode::initialization() {
3659 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3660 if (rawoop == NULL) return NULL;
3661 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3662 Node* init = rawoop->fast_out(i);
3663 if (init->is_Initialize()) {
3664 assert(init->as_Initialize()->allocation() == this, "2-way link");
3665 return init->as_Initialize();
3666 }
3667 }
3668 return NULL;
3669 }
3671 //----------------------------- loop predicates ---------------------------
3673 //------------------------------add_predicate_impl----------------------------
3674 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3675 // Too many traps seen?
3676 if (too_many_traps(reason)) {
3677 #ifdef ASSERT
3678 if (TraceLoopPredicate) {
3679 int tc = C->trap_count(reason);
3680 tty->print("too many traps=%s tcount=%d in ",
3681 Deoptimization::trap_reason_name(reason), tc);
3682 method()->print(); // which method has too many predicate traps
3683 tty->cr();
3684 }
3685 #endif
3686 // We cannot afford to take more traps here,
3687 // do not generate predicate.
3688 return;
3689 }
3691 Node *cont = _gvn.intcon(1);
3692 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3693 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3694 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3695 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3696 C->add_predicate_opaq(opq);
3697 {
3698 PreserveJVMState pjvms(this);
3699 set_control(iffalse);
3700 inc_sp(nargs);
3701 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3702 }
3703 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3704 set_control(iftrue);
3705 }
3707 //------------------------------add_predicate---------------------------------
3708 void GraphKit::add_predicate(int nargs) {
3709 if (UseLoopPredicate) {
3710 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3711 }
3712 // loop's limit check predicate should be near the loop.
3713 if (LoopLimitCheck) {
3714 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3715 }
3716 }
3718 //----------------------------- store barriers ----------------------------
3719 #define __ ideal.
3721 void GraphKit::sync_kit(IdealKit& ideal) {
3722 set_all_memory(__ merged_memory());
3723 set_i_o(__ i_o());
3724 set_control(__ ctrl());
3725 }
3727 void GraphKit::final_sync(IdealKit& ideal) {
3728 // Final sync IdealKit and graphKit.
3729 sync_kit(ideal);
3730 }
3732 // vanilla/CMS post barrier
3733 // Insert a write-barrier store. This is to let generational GC work; we have
3734 // to flag all oop-stores before the next GC point.
3735 void GraphKit::write_barrier_post(Node* oop_store,
3736 Node* obj,
3737 Node* adr,
3738 uint adr_idx,
3739 Node* val,
3740 bool use_precise) {
3741 // No store check needed if we're storing a NULL or an old object
3742 // (latter case is probably a string constant). The concurrent
3743 // mark sweep garbage collector, however, needs to have all nonNull
3744 // oop updates flagged via card-marks.
3745 if (val != NULL && val->is_Con()) {
3746 // must be either an oop or NULL
3747 const Type* t = val->bottom_type();
3748 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3749 // stores of null never (?) need barriers
3750 return;
3751 }
3753 if (use_ReduceInitialCardMarks()
3754 && obj == just_allocated_object(control())) {
3755 // We can skip marks on a freshly-allocated object in Eden.
3756 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3757 // That routine informs GC to take appropriate compensating steps,
3758 // upon a slow-path allocation, so as to make this card-mark
3759 // elision safe.
3760 return;
3761 }
3763 if (!use_precise) {
3764 // All card marks for a (non-array) instance are in one place:
3765 adr = obj;
3766 }
3767 // (Else it's an array (or unknown), and we want more precise card marks.)
3768 assert(adr != NULL, "");
3770 IdealKit ideal(this, true);
3772 // Convert the pointer to an int prior to doing math on it
3773 Node* cast = __ CastPX(__ ctrl(), adr);
3775 // Divide by card size
3776 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3777 "Only one we handle so far.");
3778 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3780 // Combine card table base and card offset
3781 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3783 // Get the alias_index for raw card-mark memory
3784 int adr_type = Compile::AliasIdxRaw;
3785 Node* zero = __ ConI(0); // Dirty card value
3786 BasicType bt = T_BYTE;
3788 if (UseCondCardMark) {
3789 // The classic GC reference write barrier is typically implemented
3790 // as a store into the global card mark table. Unfortunately
3791 // unconditional stores can result in false sharing and excessive
3792 // coherence traffic as well as false transactional aborts.
3793 // UseCondCardMark enables MP "polite" conditional card mark
3794 // stores. In theory we could relax the load from ctrl() to
3795 // no_ctrl, but that doesn't buy much latitude.
3796 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3797 __ if_then(card_val, BoolTest::ne, zero);
3798 }
3800 // Smash zero into card
3801 if( !UseConcMarkSweepGC ) {
3802 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release);
3803 } else {
3804 // Specialized path for CM store barrier
3805 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3806 }
3808 if (UseCondCardMark) {
3809 __ end_if();
3810 }
3812 // Final sync IdealKit and GraphKit.
3813 final_sync(ideal);
3814 }
3816 // G1 pre/post barriers
3817 void GraphKit::g1_write_barrier_pre(bool do_load,
3818 Node* obj,
3819 Node* adr,
3820 uint alias_idx,
3821 Node* val,
3822 const TypeOopPtr* val_type,
3823 Node* pre_val,
3824 BasicType bt) {
3826 // Some sanity checks
3827 // Note: val is unused in this routine.
3829 if (do_load) {
3830 // We need to generate the load of the previous value
3831 assert(obj != NULL, "must have a base");
3832 assert(adr != NULL, "where are loading from?");
3833 assert(pre_val == NULL, "loaded already?");
3834 assert(val_type != NULL, "need a type");
3835 } else {
3836 // In this case both val_type and alias_idx are unused.
3837 assert(pre_val != NULL, "must be loaded already");
3838 // Nothing to be done if pre_val is null.
3839 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3840 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3841 }
3842 assert(bt == T_OBJECT, "or we shouldn't be here");
3844 IdealKit ideal(this, true);
3846 Node* tls = __ thread(); // ThreadLocalStorage
3848 Node* no_ctrl = NULL;
3849 Node* no_base = __ top();
3850 Node* zero = __ ConI(0);
3851 Node* zeroX = __ ConX(0);
3853 float likely = PROB_LIKELY(0.999);
3854 float unlikely = PROB_UNLIKELY(0.999);
3856 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3857 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3859 // Offsets into the thread
3860 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3861 PtrQueue::byte_offset_of_active());
3862 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3863 PtrQueue::byte_offset_of_index());
3864 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3865 PtrQueue::byte_offset_of_buf());
3867 // Now the actual pointers into the thread
3868 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3869 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3870 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3872 // Now some of the values
3873 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3875 // if (!marking)
3876 __ if_then(marking, BoolTest::ne, zero, unlikely); {
3877 BasicType index_bt = TypeX_X->basic_type();
3878 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3879 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3881 if (do_load) {
3882 // load original value
3883 // alias_idx correct??
3884 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3885 }
3887 // if (pre_val != NULL)
3888 __ if_then(pre_val, BoolTest::ne, null()); {
3889 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3891 // is the queue for this thread full?
3892 __ if_then(index, BoolTest::ne, zeroX, likely); {
3894 // decrement the index
3895 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3897 // Now get the buffer location we will log the previous value into and store it
3898 Node *log_addr = __ AddP(no_base, buffer, next_index);
3899 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
3900 // update the index
3901 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
3903 } __ else_(); {
3905 // logging buffer is full, call the runtime
3906 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3907 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3908 } __ end_if(); // (!index)
3909 } __ end_if(); // (pre_val != NULL)
3910 } __ end_if(); // (!marking)
3912 // Final sync IdealKit and GraphKit.
3913 final_sync(ideal);
3914 }
3916 //
3917 // Update the card table and add card address to the queue
3918 //
3919 void GraphKit::g1_mark_card(IdealKit& ideal,
3920 Node* card_adr,
3921 Node* oop_store,
3922 uint oop_alias_idx,
3923 Node* index,
3924 Node* index_adr,
3925 Node* buffer,
3926 const TypeFunc* tf) {
3928 Node* zero = __ ConI(0);
3929 Node* zeroX = __ ConX(0);
3930 Node* no_base = __ top();
3931 BasicType card_bt = T_BYTE;
3932 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3933 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3935 // Now do the queue work
3936 __ if_then(index, BoolTest::ne, zeroX); {
3938 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3939 Node* log_addr = __ AddP(no_base, buffer, next_index);
3941 // Order, see storeCM.
3942 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
3943 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
3945 } __ else_(); {
3946 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3947 } __ end_if();
3949 }
3951 void GraphKit::g1_write_barrier_post(Node* oop_store,
3952 Node* obj,
3953 Node* adr,
3954 uint alias_idx,
3955 Node* val,
3956 BasicType bt,
3957 bool use_precise) {
3958 // If we are writing a NULL then we need no post barrier
3960 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3961 // Must be NULL
3962 const Type* t = val->bottom_type();
3963 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3964 // No post barrier if writing NULLx
3965 return;
3966 }
3968 if (!use_precise) {
3969 // All card marks for a (non-array) instance are in one place:
3970 adr = obj;
3971 }
3972 // (Else it's an array (or unknown), and we want more precise card marks.)
3973 assert(adr != NULL, "");
3975 IdealKit ideal(this, true);
3977 Node* tls = __ thread(); // ThreadLocalStorage
3979 Node* no_base = __ top();
3980 float likely = PROB_LIKELY(0.999);
3981 float unlikely = PROB_UNLIKELY(0.999);
3982 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
3983 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
3984 Node* zeroX = __ ConX(0);
3986 // Get the alias_index for raw card-mark memory
3987 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3989 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3991 // Offsets into the thread
3992 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3993 PtrQueue::byte_offset_of_index());
3994 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3995 PtrQueue::byte_offset_of_buf());
3997 // Pointers into the thread
3999 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4000 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4002 // Now some values
4003 // Use ctrl to avoid hoisting these values past a safepoint, which could
4004 // potentially reset these fields in the JavaThread.
4005 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4006 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4008 // Convert the store obj pointer to an int prior to doing math on it
4009 // Must use ctrl to prevent "integerized oop" existing across safepoint
4010 Node* cast = __ CastPX(__ ctrl(), adr);
4012 // Divide pointer by card size
4013 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4015 // Combine card table base and card offset
4016 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4018 // If we know the value being stored does it cross regions?
4020 if (val != NULL) {
4021 // Does the store cause us to cross regions?
4023 // Should be able to do an unsigned compare of region_size instead of
4024 // and extra shift. Do we have an unsigned compare??
4025 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4026 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4028 // if (xor_res == 0) same region so skip
4029 __ if_then(xor_res, BoolTest::ne, zeroX); {
4031 // No barrier if we are storing a NULL
4032 __ if_then(val, BoolTest::ne, null(), unlikely); {
4034 // Ok must mark the card if not already dirty
4036 // load the original value of the card
4037 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4039 __ if_then(card_val, BoolTest::ne, young_card); {
4040 sync_kit(ideal);
4041 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4042 insert_mem_bar(Op_MemBarVolatile, oop_store);
4043 __ sync_kit(this);
4045 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4046 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4047 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4048 } __ end_if();
4049 } __ end_if();
4050 } __ end_if();
4051 } __ end_if();
4052 } else {
4053 // Object.clone() instrinsic uses this path.
4054 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4055 }
4057 // Final sync IdealKit and GraphKit.
4058 final_sync(ideal);
4059 }
4060 #undef __
4064 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
4065 if (java_lang_String::has_offset_field()) {
4066 int offset_offset = java_lang_String::offset_offset_in_bytes();
4067 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4068 false, NULL, 0);
4069 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4070 int offset_field_idx = C->get_alias_index(offset_field_type);
4071 return make_load(ctrl,
4072 basic_plus_adr(str, str, offset_offset),
4073 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered);
4074 } else {
4075 return intcon(0);
4076 }
4077 }
4079 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4080 if (java_lang_String::has_count_field()) {
4081 int count_offset = java_lang_String::count_offset_in_bytes();
4082 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4083 false, NULL, 0);
4084 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4085 int count_field_idx = C->get_alias_index(count_field_type);
4086 return make_load(ctrl,
4087 basic_plus_adr(str, str, count_offset),
4088 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered);
4089 } else {
4090 return load_array_length(load_String_value(ctrl, str));
4091 }
4092 }
4094 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4095 int value_offset = java_lang_String::value_offset_in_bytes();
4096 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4097 false, NULL, 0);
4098 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4099 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4100 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
4101 ciTypeArrayKlass::make(T_CHAR), true, 0);
4102 int value_field_idx = C->get_alias_index(value_field_type);
4103 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4104 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4105 // String.value field is known to be @Stable.
4106 if (UseImplicitStableValues) {
4107 load = cast_array_to_stable(load, value_type);
4108 }
4109 return load;
4110 }
4112 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
4113 int offset_offset = java_lang_String::offset_offset_in_bytes();
4114 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4115 false, NULL, 0);
4116 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4117 int offset_field_idx = C->get_alias_index(offset_field_type);
4118 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
4119 value, T_INT, offset_field_idx, MemNode::unordered);
4120 }
4122 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4123 int value_offset = java_lang_String::value_offset_in_bytes();
4124 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4125 false, NULL, 0);
4126 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4128 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4129 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered);
4130 }
4132 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
4133 int count_offset = java_lang_String::count_offset_in_bytes();
4134 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4135 false, NULL, 0);
4136 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4137 int count_field_idx = C->get_alias_index(count_field_type);
4138 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
4139 value, T_INT, count_field_idx, MemNode::unordered);
4140 }
4142 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4143 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4144 // assumption of CCP analysis.
4145 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
4146 }