Thu, 20 Sep 2012 16:49:17 +0200
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
Summary: use shorter instruction sequences for atomic add and atomic exchange when possible.
Reviewed-by: kvn, jrose
1 /*
2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
28 #include "gc_implementation/g1/heapRegion.hpp"
29 #include "gc_interface/collectedHeap.hpp"
30 #include "memory/barrierSet.hpp"
31 #include "memory/cardTableModRefBS.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/graphKit.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "runtime/deoptimization.hpp"
41 #include "runtime/sharedRuntime.hpp"
43 //----------------------------GraphKit-----------------------------------------
44 // Main utility constructor.
45 GraphKit::GraphKit(JVMState* jvms)
46 : Phase(Phase::Parser),
47 _env(C->env()),
48 _gvn(*C->initial_gvn())
49 {
50 _exceptions = jvms->map()->next_exception();
51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
52 set_jvms(jvms);
53 }
55 // Private constructor for parser.
56 GraphKit::GraphKit()
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn())
60 {
61 _exceptions = NULL;
62 set_map(NULL);
63 debug_only(_sp = -99);
64 debug_only(set_bci(-99));
65 }
69 //---------------------------clean_stack---------------------------------------
70 // Clear away rubbish from the stack area of the JVM state.
71 // This destroys any arguments that may be waiting on the stack.
72 void GraphKit::clean_stack(int from_sp) {
73 SafePointNode* map = this->map();
74 JVMState* jvms = this->jvms();
75 int stk_size = jvms->stk_size();
76 int stkoff = jvms->stkoff();
77 Node* top = this->top();
78 for (int i = from_sp; i < stk_size; i++) {
79 if (map->in(stkoff + i) != top) {
80 map->set_req(stkoff + i, top);
81 }
82 }
83 }
86 //--------------------------------sync_jvms-----------------------------------
87 // Make sure our current jvms agrees with our parse state.
88 JVMState* GraphKit::sync_jvms() const {
89 JVMState* jvms = this->jvms();
90 jvms->set_bci(bci()); // Record the new bci in the JVMState
91 jvms->set_sp(sp()); // Record the new sp in the JVMState
92 assert(jvms_in_sync(), "jvms is now in sync");
93 return jvms;
94 }
96 #ifdef ASSERT
97 bool GraphKit::jvms_in_sync() const {
98 Parse* parse = is_Parse();
99 if (parse == NULL) {
100 if (bci() != jvms()->bci()) return false;
101 if (sp() != (int)jvms()->sp()) return false;
102 return true;
103 }
104 if (jvms()->method() != parse->method()) return false;
105 if (jvms()->bci() != parse->bci()) return false;
106 int jvms_sp = jvms()->sp();
107 if (jvms_sp != parse->sp()) return false;
108 int jvms_depth = jvms()->depth();
109 if (jvms_depth != parse->depth()) return false;
110 return true;
111 }
113 // Local helper checks for special internal merge points
114 // used to accumulate and merge exception states.
115 // They are marked by the region's in(0) edge being the map itself.
116 // Such merge points must never "escape" into the parser at large,
117 // until they have been handed to gvn.transform.
118 static bool is_hidden_merge(Node* reg) {
119 if (reg == NULL) return false;
120 if (reg->is_Phi()) {
121 reg = reg->in(0);
122 if (reg == NULL) return false;
123 }
124 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
125 }
127 void GraphKit::verify_map() const {
128 if (map() == NULL) return; // null map is OK
129 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
130 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
131 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
132 }
134 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
135 assert(ex_map->next_exception() == NULL, "not already part of a chain");
136 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
137 }
138 #endif
140 //---------------------------stop_and_kill_map---------------------------------
141 // Set _map to NULL, signalling a stop to further bytecode execution.
142 // First smash the current map's control to a constant, to mark it dead.
143 void GraphKit::stop_and_kill_map() {
144 SafePointNode* dead_map = stop();
145 if (dead_map != NULL) {
146 dead_map->disconnect_inputs(NULL); // Mark the map as killed.
147 assert(dead_map->is_killed(), "must be so marked");
148 }
149 }
152 //--------------------------------stopped--------------------------------------
153 // Tell if _map is NULL, or control is top.
154 bool GraphKit::stopped() {
155 if (map() == NULL) return true;
156 else if (control() == top()) return true;
157 else return false;
158 }
161 //-----------------------------has_ex_handler----------------------------------
162 // Tell if this method or any caller method has exception handlers.
163 bool GraphKit::has_ex_handler() {
164 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
165 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
166 return true;
167 }
168 }
169 return false;
170 }
172 //------------------------------save_ex_oop------------------------------------
173 // Save an exception without blowing stack contents or other JVM state.
174 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
175 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
176 ex_map->add_req(ex_oop);
177 debug_only(verify_exception_state(ex_map));
178 }
180 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
181 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
182 Node* ex_oop = ex_map->in(ex_map->req()-1);
183 if (clear_it) ex_map->del_req(ex_map->req()-1);
184 return ex_oop;
185 }
187 //-----------------------------saved_ex_oop------------------------------------
188 // Recover a saved exception from its map.
189 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
190 return common_saved_ex_oop(ex_map, false);
191 }
193 //--------------------------clear_saved_ex_oop---------------------------------
194 // Erase a previously saved exception from its map.
195 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
196 return common_saved_ex_oop(ex_map, true);
197 }
199 #ifdef ASSERT
200 //---------------------------has_saved_ex_oop----------------------------------
201 // Erase a previously saved exception from its map.
202 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
203 return ex_map->req() == ex_map->jvms()->endoff()+1;
204 }
205 #endif
207 //-------------------------make_exception_state--------------------------------
208 // Turn the current JVM state into an exception state, appending the ex_oop.
209 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
210 sync_jvms();
211 SafePointNode* ex_map = stop(); // do not manipulate this map any more
212 set_saved_ex_oop(ex_map, ex_oop);
213 return ex_map;
214 }
217 //--------------------------add_exception_state--------------------------------
218 // Add an exception to my list of exceptions.
219 void GraphKit::add_exception_state(SafePointNode* ex_map) {
220 if (ex_map == NULL || ex_map->control() == top()) {
221 return;
222 }
223 #ifdef ASSERT
224 verify_exception_state(ex_map);
225 if (has_exceptions()) {
226 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
227 }
228 #endif
230 // If there is already an exception of exactly this type, merge with it.
231 // In particular, null-checks and other low-level exceptions common up here.
232 Node* ex_oop = saved_ex_oop(ex_map);
233 const Type* ex_type = _gvn.type(ex_oop);
234 if (ex_oop == top()) {
235 // No action needed.
236 return;
237 }
238 assert(ex_type->isa_instptr(), "exception must be an instance");
239 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
240 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
241 // We check sp also because call bytecodes can generate exceptions
242 // both before and after arguments are popped!
243 if (ex_type2 == ex_type
244 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
245 combine_exception_states(ex_map, e2);
246 return;
247 }
248 }
250 // No pre-existing exception of the same type. Chain it on the list.
251 push_exception_state(ex_map);
252 }
254 //-----------------------add_exception_states_from-----------------------------
255 void GraphKit::add_exception_states_from(JVMState* jvms) {
256 SafePointNode* ex_map = jvms->map()->next_exception();
257 if (ex_map != NULL) {
258 jvms->map()->set_next_exception(NULL);
259 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
260 next_map = ex_map->next_exception();
261 ex_map->set_next_exception(NULL);
262 add_exception_state(ex_map);
263 }
264 }
265 }
267 //-----------------------transfer_exceptions_into_jvms-------------------------
268 JVMState* GraphKit::transfer_exceptions_into_jvms() {
269 if (map() == NULL) {
270 // We need a JVMS to carry the exceptions, but the map has gone away.
271 // Create a scratch JVMS, cloned from any of the exception states...
272 if (has_exceptions()) {
273 _map = _exceptions;
274 _map = clone_map();
275 _map->set_next_exception(NULL);
276 clear_saved_ex_oop(_map);
277 debug_only(verify_map());
278 } else {
279 // ...or created from scratch
280 JVMState* jvms = new (C) JVMState(_method, NULL);
281 jvms->set_bci(_bci);
282 jvms->set_sp(_sp);
283 jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms));
284 set_jvms(jvms);
285 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
286 set_all_memory(top());
287 while (map()->req() < jvms->endoff()) map()->add_req(top());
288 }
289 // (This is a kludge, in case you didn't notice.)
290 set_control(top());
291 }
292 JVMState* jvms = sync_jvms();
293 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
294 jvms->map()->set_next_exception(_exceptions);
295 _exceptions = NULL; // done with this set of exceptions
296 return jvms;
297 }
299 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
300 assert(is_hidden_merge(dstphi), "must be a special merge node");
301 assert(is_hidden_merge(srcphi), "must be a special merge node");
302 uint limit = srcphi->req();
303 for (uint i = PhiNode::Input; i < limit; i++) {
304 dstphi->add_req(srcphi->in(i));
305 }
306 }
307 static inline void add_one_req(Node* dstphi, Node* src) {
308 assert(is_hidden_merge(dstphi), "must be a special merge node");
309 assert(!is_hidden_merge(src), "must not be a special merge node");
310 dstphi->add_req(src);
311 }
313 //-----------------------combine_exception_states------------------------------
314 // This helper function combines exception states by building phis on a
315 // specially marked state-merging region. These regions and phis are
316 // untransformed, and can build up gradually. The region is marked by
317 // having a control input of its exception map, rather than NULL. Such
318 // regions do not appear except in this function, and in use_exception_state.
319 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
320 if (failing()) return; // dying anyway...
321 JVMState* ex_jvms = ex_map->_jvms;
322 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
323 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
324 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
325 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
326 assert(ex_map->req() == phi_map->req(), "matching maps");
327 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
328 Node* hidden_merge_mark = root();
329 Node* region = phi_map->control();
330 MergeMemNode* phi_mem = phi_map->merged_memory();
331 MergeMemNode* ex_mem = ex_map->merged_memory();
332 if (region->in(0) != hidden_merge_mark) {
333 // The control input is not (yet) a specially-marked region in phi_map.
334 // Make it so, and build some phis.
335 region = new (C, 2) RegionNode(2);
336 _gvn.set_type(region, Type::CONTROL);
337 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
338 region->init_req(1, phi_map->control());
339 phi_map->set_control(region);
340 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
341 record_for_igvn(io_phi);
342 _gvn.set_type(io_phi, Type::ABIO);
343 phi_map->set_i_o(io_phi);
344 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
345 Node* m = mms.memory();
346 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
347 record_for_igvn(m_phi);
348 _gvn.set_type(m_phi, Type::MEMORY);
349 mms.set_memory(m_phi);
350 }
351 }
353 // Either or both of phi_map and ex_map might already be converted into phis.
354 Node* ex_control = ex_map->control();
355 // if there is special marking on ex_map also, we add multiple edges from src
356 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
357 // how wide was the destination phi_map, originally?
358 uint orig_width = region->req();
360 if (add_multiple) {
361 add_n_reqs(region, ex_control);
362 add_n_reqs(phi_map->i_o(), ex_map->i_o());
363 } else {
364 // ex_map has no merges, so we just add single edges everywhere
365 add_one_req(region, ex_control);
366 add_one_req(phi_map->i_o(), ex_map->i_o());
367 }
368 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
369 if (mms.is_empty()) {
370 // get a copy of the base memory, and patch some inputs into it
371 const TypePtr* adr_type = mms.adr_type(C);
372 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
373 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
374 mms.set_memory(phi);
375 // Prepare to append interesting stuff onto the newly sliced phi:
376 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
377 }
378 // Append stuff from ex_map:
379 if (add_multiple) {
380 add_n_reqs(mms.memory(), mms.memory2());
381 } else {
382 add_one_req(mms.memory(), mms.memory2());
383 }
384 }
385 uint limit = ex_map->req();
386 for (uint i = TypeFunc::Parms; i < limit; i++) {
387 // Skip everything in the JVMS after tos. (The ex_oop follows.)
388 if (i == tos) i = ex_jvms->monoff();
389 Node* src = ex_map->in(i);
390 Node* dst = phi_map->in(i);
391 if (src != dst) {
392 PhiNode* phi;
393 if (dst->in(0) != region) {
394 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
395 record_for_igvn(phi);
396 _gvn.set_type(phi, phi->type());
397 phi_map->set_req(i, dst);
398 // Prepare to append interesting stuff onto the new phi:
399 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
400 } else {
401 assert(dst->is_Phi(), "nobody else uses a hidden region");
402 phi = (PhiNode*)dst;
403 }
404 if (add_multiple && src->in(0) == ex_control) {
405 // Both are phis.
406 add_n_reqs(dst, src);
407 } else {
408 while (dst->req() < region->req()) add_one_req(dst, src);
409 }
410 const Type* srctype = _gvn.type(src);
411 if (phi->type() != srctype) {
412 const Type* dsttype = phi->type()->meet(srctype);
413 if (phi->type() != dsttype) {
414 phi->set_type(dsttype);
415 _gvn.set_type(phi, dsttype);
416 }
417 }
418 }
419 }
420 }
422 //--------------------------use_exception_state--------------------------------
423 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
424 if (failing()) { stop(); return top(); }
425 Node* region = phi_map->control();
426 Node* hidden_merge_mark = root();
427 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
428 Node* ex_oop = clear_saved_ex_oop(phi_map);
429 if (region->in(0) == hidden_merge_mark) {
430 // Special marking for internal ex-states. Process the phis now.
431 region->set_req(0, region); // now it's an ordinary region
432 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
433 // Note: Setting the jvms also sets the bci and sp.
434 set_control(_gvn.transform(region));
435 uint tos = jvms()->stkoff() + sp();
436 for (uint i = 1; i < tos; i++) {
437 Node* x = phi_map->in(i);
438 if (x->in(0) == region) {
439 assert(x->is_Phi(), "expected a special phi");
440 phi_map->set_req(i, _gvn.transform(x));
441 }
442 }
443 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
444 Node* x = mms.memory();
445 if (x->in(0) == region) {
446 assert(x->is_Phi(), "nobody else uses a hidden region");
447 mms.set_memory(_gvn.transform(x));
448 }
449 }
450 if (ex_oop->in(0) == region) {
451 assert(ex_oop->is_Phi(), "expected a special phi");
452 ex_oop = _gvn.transform(ex_oop);
453 }
454 } else {
455 set_jvms(phi_map->jvms());
456 }
458 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
459 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
460 return ex_oop;
461 }
463 //---------------------------------java_bc-------------------------------------
464 Bytecodes::Code GraphKit::java_bc() const {
465 ciMethod* method = this->method();
466 int bci = this->bci();
467 if (method != NULL && bci != InvocationEntryBci)
468 return method->java_code_at_bci(bci);
469 else
470 return Bytecodes::_illegal;
471 }
473 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
474 bool must_throw) {
475 // if the exception capability is set, then we will generate code
476 // to check the JavaThread.should_post_on_exceptions flag to see
477 // if we actually need to report exception events (for this
478 // thread). If we don't need to report exception events, we will
479 // take the normal fast path provided by add_exception_events. If
480 // exception event reporting is enabled for this thread, we will
481 // take the uncommon_trap in the BuildCutout below.
483 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
484 Node* jthread = _gvn.transform(new (C, 1) ThreadLocalNode());
485 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
486 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false);
488 // Test the should_post_on_exceptions_flag vs. 0
489 Node* chk = _gvn.transform( new (C, 3) CmpINode(should_post_flag, intcon(0)) );
490 Node* tst = _gvn.transform( new (C, 2) BoolNode(chk, BoolTest::eq) );
492 // Branch to slow_path if should_post_on_exceptions_flag was true
493 { BuildCutout unless(this, tst, PROB_MAX);
494 // Do not try anything fancy if we're notifying the VM on every throw.
495 // Cf. case Bytecodes::_athrow in parse2.cpp.
496 uncommon_trap(reason, Deoptimization::Action_none,
497 (ciKlass*)NULL, (char*)NULL, must_throw);
498 }
500 }
502 //------------------------------builtin_throw----------------------------------
503 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
504 bool must_throw = true;
506 if (env()->jvmti_can_post_on_exceptions()) {
507 // check if we must post exception events, take uncommon trap if so
508 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
509 // here if should_post_on_exceptions is false
510 // continue on with the normal codegen
511 }
513 // If this particular condition has not yet happened at this
514 // bytecode, then use the uncommon trap mechanism, and allow for
515 // a future recompilation if several traps occur here.
516 // If the throw is hot, try to use a more complicated inline mechanism
517 // which keeps execution inside the compiled code.
518 bool treat_throw_as_hot = false;
519 ciMethodData* md = method()->method_data();
521 if (ProfileTraps) {
522 if (too_many_traps(reason)) {
523 treat_throw_as_hot = true;
524 }
525 // (If there is no MDO at all, assume it is early in
526 // execution, and that any deopts are part of the
527 // startup transient, and don't need to be remembered.)
529 // Also, if there is a local exception handler, treat all throws
530 // as hot if there has been at least one in this method.
531 if (C->trap_count(reason) != 0
532 && method()->method_data()->trap_count(reason) != 0
533 && has_ex_handler()) {
534 treat_throw_as_hot = true;
535 }
536 }
538 // If this throw happens frequently, an uncommon trap might cause
539 // a performance pothole. If there is a local exception handler,
540 // and if this particular bytecode appears to be deoptimizing often,
541 // let us handle the throw inline, with a preconstructed instance.
542 // Note: If the deopt count has blown up, the uncommon trap
543 // runtime is going to flush this nmethod, not matter what.
544 if (treat_throw_as_hot
545 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
546 // If the throw is local, we use a pre-existing instance and
547 // punt on the backtrace. This would lead to a missing backtrace
548 // (a repeat of 4292742) if the backtrace object is ever asked
549 // for its backtrace.
550 // Fixing this remaining case of 4292742 requires some flavor of
551 // escape analysis. Leave that for the future.
552 ciInstance* ex_obj = NULL;
553 switch (reason) {
554 case Deoptimization::Reason_null_check:
555 ex_obj = env()->NullPointerException_instance();
556 break;
557 case Deoptimization::Reason_div0_check:
558 ex_obj = env()->ArithmeticException_instance();
559 break;
560 case Deoptimization::Reason_range_check:
561 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
562 break;
563 case Deoptimization::Reason_class_check:
564 if (java_bc() == Bytecodes::_aastore) {
565 ex_obj = env()->ArrayStoreException_instance();
566 } else {
567 ex_obj = env()->ClassCastException_instance();
568 }
569 break;
570 }
571 if (failing()) { stop(); return; } // exception allocation might fail
572 if (ex_obj != NULL) {
573 // Cheat with a preallocated exception object.
574 if (C->log() != NULL)
575 C->log()->elem("hot_throw preallocated='1' reason='%s'",
576 Deoptimization::trap_reason_name(reason));
577 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
578 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
580 // Clear the detail message of the preallocated exception object.
581 // Weblogic sometimes mutates the detail message of exceptions
582 // using reflection.
583 int offset = java_lang_Throwable::get_detailMessage_offset();
584 const TypePtr* adr_typ = ex_con->add_offset(offset);
586 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
587 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
588 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT);
590 add_exception_state(make_exception_state(ex_node));
591 return;
592 }
593 }
595 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
596 // It won't be much cheaper than bailing to the interp., since we'll
597 // have to pass up all the debug-info, and the runtime will have to
598 // create the stack trace.
600 // Usual case: Bail to interpreter.
601 // Reserve the right to recompile if we haven't seen anything yet.
603 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
604 if (treat_throw_as_hot
605 && (method()->method_data()->trap_recompiled_at(bci())
606 || C->too_many_traps(reason))) {
607 // We cannot afford to take more traps here. Suffer in the interpreter.
608 if (C->log() != NULL)
609 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
610 Deoptimization::trap_reason_name(reason),
611 C->trap_count(reason));
612 action = Deoptimization::Action_none;
613 }
615 // "must_throw" prunes the JVM state to include only the stack, if there
616 // are no local exception handlers. This should cut down on register
617 // allocation time and code size, by drastically reducing the number
618 // of in-edges on the call to the uncommon trap.
620 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
621 }
624 //----------------------------PreserveJVMState---------------------------------
625 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
626 debug_only(kit->verify_map());
627 _kit = kit;
628 _map = kit->map(); // preserve the map
629 _sp = kit->sp();
630 kit->set_map(clone_map ? kit->clone_map() : NULL);
631 #ifdef ASSERT
632 _bci = kit->bci();
633 Parse* parser = kit->is_Parse();
634 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
635 _block = block;
636 #endif
637 }
638 PreserveJVMState::~PreserveJVMState() {
639 GraphKit* kit = _kit;
640 #ifdef ASSERT
641 assert(kit->bci() == _bci, "bci must not shift");
642 Parse* parser = kit->is_Parse();
643 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
644 assert(block == _block, "block must not shift");
645 #endif
646 kit->set_map(_map);
647 kit->set_sp(_sp);
648 }
651 //-----------------------------BuildCutout-------------------------------------
652 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
653 : PreserveJVMState(kit)
654 {
655 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
656 SafePointNode* outer_map = _map; // preserved map is caller's
657 SafePointNode* inner_map = kit->map();
658 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
659 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) ));
660 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) ));
661 }
662 BuildCutout::~BuildCutout() {
663 GraphKit* kit = _kit;
664 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
665 }
667 //---------------------------PreserveReexecuteState----------------------------
668 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
669 assert(!kit->stopped(), "must call stopped() before");
670 _kit = kit;
671 _sp = kit->sp();
672 _reexecute = kit->jvms()->_reexecute;
673 }
674 PreserveReexecuteState::~PreserveReexecuteState() {
675 if (_kit->stopped()) return;
676 _kit->jvms()->_reexecute = _reexecute;
677 _kit->set_sp(_sp);
678 }
680 //------------------------------clone_map--------------------------------------
681 // Implementation of PreserveJVMState
682 //
683 // Only clone_map(...) here. If this function is only used in the
684 // PreserveJVMState class we may want to get rid of this extra
685 // function eventually and do it all there.
687 SafePointNode* GraphKit::clone_map() {
688 if (map() == NULL) return NULL;
690 // Clone the memory edge first
691 Node* mem = MergeMemNode::make(C, map()->memory());
692 gvn().set_type_bottom(mem);
694 SafePointNode *clonemap = (SafePointNode*)map()->clone();
695 JVMState* jvms = this->jvms();
696 JVMState* clonejvms = jvms->clone_shallow(C);
697 clonemap->set_memory(mem);
698 clonemap->set_jvms(clonejvms);
699 clonejvms->set_map(clonemap);
700 record_for_igvn(clonemap);
701 gvn().set_type_bottom(clonemap);
702 return clonemap;
703 }
706 //-----------------------------set_map_clone-----------------------------------
707 void GraphKit::set_map_clone(SafePointNode* m) {
708 _map = m;
709 _map = clone_map();
710 _map->set_next_exception(NULL);
711 debug_only(verify_map());
712 }
715 //----------------------------kill_dead_locals---------------------------------
716 // Detect any locals which are known to be dead, and force them to top.
717 void GraphKit::kill_dead_locals() {
718 // Consult the liveness information for the locals. If any
719 // of them are unused, then they can be replaced by top(). This
720 // should help register allocation time and cut down on the size
721 // of the deoptimization information.
723 // This call is made from many of the bytecode handling
724 // subroutines called from the Big Switch in do_one_bytecode.
725 // Every bytecode which might include a slow path is responsible
726 // for killing its dead locals. The more consistent we
727 // are about killing deads, the fewer useless phis will be
728 // constructed for them at various merge points.
730 // bci can be -1 (InvocationEntryBci). We return the entry
731 // liveness for the method.
733 if (method() == NULL || method()->code_size() == 0) {
734 // We are building a graph for a call to a native method.
735 // All locals are live.
736 return;
737 }
739 ResourceMark rm;
741 // Consult the liveness information for the locals. If any
742 // of them are unused, then they can be replaced by top(). This
743 // should help register allocation time and cut down on the size
744 // of the deoptimization information.
745 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
747 int len = (int)live_locals.size();
748 assert(len <= jvms()->loc_size(), "too many live locals");
749 for (int local = 0; local < len; local++) {
750 if (!live_locals.at(local)) {
751 set_local(local, top());
752 }
753 }
754 }
756 #ifdef ASSERT
757 //-------------------------dead_locals_are_killed------------------------------
758 // Return true if all dead locals are set to top in the map.
759 // Used to assert "clean" debug info at various points.
760 bool GraphKit::dead_locals_are_killed() {
761 if (method() == NULL || method()->code_size() == 0) {
762 // No locals need to be dead, so all is as it should be.
763 return true;
764 }
766 // Make sure somebody called kill_dead_locals upstream.
767 ResourceMark rm;
768 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
769 if (jvms->loc_size() == 0) continue; // no locals to consult
770 SafePointNode* map = jvms->map();
771 ciMethod* method = jvms->method();
772 int bci = jvms->bci();
773 if (jvms == this->jvms()) {
774 bci = this->bci(); // it might not yet be synched
775 }
776 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
777 int len = (int)live_locals.size();
778 if (!live_locals.is_valid() || len == 0)
779 // This method is trivial, or is poisoned by a breakpoint.
780 return true;
781 assert(len == jvms->loc_size(), "live map consistent with locals map");
782 for (int local = 0; local < len; local++) {
783 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
784 if (PrintMiscellaneous && (Verbose || WizardMode)) {
785 tty->print_cr("Zombie local %d: ", local);
786 jvms->dump();
787 }
788 return false;
789 }
790 }
791 }
792 return true;
793 }
795 #endif //ASSERT
797 // Helper function for enforcing certain bytecodes to reexecute if
798 // deoptimization happens
799 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
800 ciMethod* cur_method = jvms->method();
801 int cur_bci = jvms->bci();
802 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
803 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
804 return Interpreter::bytecode_should_reexecute(code) ||
805 is_anewarray && code == Bytecodes::_multianewarray;
806 // Reexecute _multianewarray bytecode which was replaced with
807 // sequence of [a]newarray. See Parse::do_multianewarray().
808 //
809 // Note: interpreter should not have it set since this optimization
810 // is limited by dimensions and guarded by flag so in some cases
811 // multianewarray() runtime calls will be generated and
812 // the bytecode should not be reexecutes (stack will not be reset).
813 } else
814 return false;
815 }
817 // Helper function for adding JVMState and debug information to node
818 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
819 // Add the safepoint edges to the call (or other safepoint).
821 // Make sure dead locals are set to top. This
822 // should help register allocation time and cut down on the size
823 // of the deoptimization information.
824 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
826 // Walk the inline list to fill in the correct set of JVMState's
827 // Also fill in the associated edges for each JVMState.
829 JVMState* youngest_jvms = sync_jvms();
831 // If we are guaranteed to throw, we can prune everything but the
832 // input to the current bytecode.
833 bool can_prune_locals = false;
834 uint stack_slots_not_pruned = 0;
835 int inputs = 0, depth = 0;
836 if (must_throw) {
837 assert(method() == youngest_jvms->method(), "sanity");
838 if (compute_stack_effects(inputs, depth)) {
839 can_prune_locals = true;
840 stack_slots_not_pruned = inputs;
841 }
842 }
844 if (env()->jvmti_can_access_local_variables()) {
845 // At any safepoint, this method can get breakpointed, which would
846 // then require an immediate deoptimization.
847 can_prune_locals = false; // do not prune locals
848 stack_slots_not_pruned = 0;
849 }
851 // do not scribble on the input jvms
852 JVMState* out_jvms = youngest_jvms->clone_deep(C);
853 call->set_jvms(out_jvms); // Start jvms list for call node
855 // For a known set of bytecodes, the interpreter should reexecute them if
856 // deoptimization happens. We set the reexecute state for them here
857 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
858 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
859 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
860 }
862 // Presize the call:
863 debug_only(uint non_debug_edges = call->req());
864 call->add_req_batch(top(), youngest_jvms->debug_depth());
865 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
867 // Set up edges so that the call looks like this:
868 // Call [state:] ctl io mem fptr retadr
869 // [parms:] parm0 ... parmN
870 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
871 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
872 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
873 // Note that caller debug info precedes callee debug info.
875 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
876 uint debug_ptr = call->req();
878 // Loop over the map input edges associated with jvms, add them
879 // to the call node, & reset all offsets to match call node array.
880 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
881 uint debug_end = debug_ptr;
882 uint debug_start = debug_ptr - in_jvms->debug_size();
883 debug_ptr = debug_start; // back up the ptr
885 uint p = debug_start; // walks forward in [debug_start, debug_end)
886 uint j, k, l;
887 SafePointNode* in_map = in_jvms->map();
888 out_jvms->set_map(call);
890 if (can_prune_locals) {
891 assert(in_jvms->method() == out_jvms->method(), "sanity");
892 // If the current throw can reach an exception handler in this JVMS,
893 // then we must keep everything live that can reach that handler.
894 // As a quick and dirty approximation, we look for any handlers at all.
895 if (in_jvms->method()->has_exception_handlers()) {
896 can_prune_locals = false;
897 }
898 }
900 // Add the Locals
901 k = in_jvms->locoff();
902 l = in_jvms->loc_size();
903 out_jvms->set_locoff(p);
904 if (!can_prune_locals) {
905 for (j = 0; j < l; j++)
906 call->set_req(p++, in_map->in(k+j));
907 } else {
908 p += l; // already set to top above by add_req_batch
909 }
911 // Add the Expression Stack
912 k = in_jvms->stkoff();
913 l = in_jvms->sp();
914 out_jvms->set_stkoff(p);
915 if (!can_prune_locals) {
916 for (j = 0; j < l; j++)
917 call->set_req(p++, in_map->in(k+j));
918 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
919 // Divide stack into {S0,...,S1}, where S0 is set to top.
920 uint s1 = stack_slots_not_pruned;
921 stack_slots_not_pruned = 0; // for next iteration
922 if (s1 > l) s1 = l;
923 uint s0 = l - s1;
924 p += s0; // skip the tops preinstalled by add_req_batch
925 for (j = s0; j < l; j++)
926 call->set_req(p++, in_map->in(k+j));
927 } else {
928 p += l; // already set to top above by add_req_batch
929 }
931 // Add the Monitors
932 k = in_jvms->monoff();
933 l = in_jvms->mon_size();
934 out_jvms->set_monoff(p);
935 for (j = 0; j < l; j++)
936 call->set_req(p++, in_map->in(k+j));
938 // Copy any scalar object fields.
939 k = in_jvms->scloff();
940 l = in_jvms->scl_size();
941 out_jvms->set_scloff(p);
942 for (j = 0; j < l; j++)
943 call->set_req(p++, in_map->in(k+j));
945 // Finish the new jvms.
946 out_jvms->set_endoff(p);
948 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
949 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
950 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
951 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
952 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
953 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
955 // Update the two tail pointers in parallel.
956 out_jvms = out_jvms->caller();
957 in_jvms = in_jvms->caller();
958 }
960 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
962 // Test the correctness of JVMState::debug_xxx accessors:
963 assert(call->jvms()->debug_start() == non_debug_edges, "");
964 assert(call->jvms()->debug_end() == call->req(), "");
965 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
966 }
968 bool GraphKit::compute_stack_effects(int& inputs, int& depth, bool for_parse) {
969 Bytecodes::Code code = java_bc();
970 if (code == Bytecodes::_wide) {
971 code = method()->java_code_at_bci(bci() + 1);
972 }
974 BasicType rtype = T_ILLEGAL;
975 int rsize = 0;
977 if (code != Bytecodes::_illegal) {
978 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
979 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
980 if (rtype < T_CONFLICT)
981 rsize = type2size[rtype];
982 }
984 switch (code) {
985 case Bytecodes::_illegal:
986 return false;
988 case Bytecodes::_ldc:
989 case Bytecodes::_ldc_w:
990 case Bytecodes::_ldc2_w:
991 inputs = 0;
992 break;
994 case Bytecodes::_dup: inputs = 1; break;
995 case Bytecodes::_dup_x1: inputs = 2; break;
996 case Bytecodes::_dup_x2: inputs = 3; break;
997 case Bytecodes::_dup2: inputs = 2; break;
998 case Bytecodes::_dup2_x1: inputs = 3; break;
999 case Bytecodes::_dup2_x2: inputs = 4; break;
1000 case Bytecodes::_swap: inputs = 2; break;
1001 case Bytecodes::_arraylength: inputs = 1; break;
1003 case Bytecodes::_getstatic:
1004 case Bytecodes::_putstatic:
1005 case Bytecodes::_getfield:
1006 case Bytecodes::_putfield:
1007 {
1008 bool is_get = (depth >= 0), is_static = (depth & 1);
1009 ciBytecodeStream iter(method());
1010 iter.reset_to_bci(bci());
1011 iter.next();
1012 bool ignored_will_link;
1013 ciField* field = iter.get_field(ignored_will_link);
1014 int size = field->type()->size();
1015 inputs = (is_static ? 0 : 1);
1016 if (is_get) {
1017 depth = size - inputs;
1018 } else {
1019 inputs += size; // putxxx pops the value from the stack
1020 depth = - inputs;
1021 }
1022 }
1023 break;
1025 case Bytecodes::_invokevirtual:
1026 case Bytecodes::_invokespecial:
1027 case Bytecodes::_invokestatic:
1028 case Bytecodes::_invokedynamic:
1029 case Bytecodes::_invokeinterface:
1030 {
1031 ciBytecodeStream iter(method());
1032 iter.reset_to_bci(bci());
1033 iter.next();
1034 bool ignored_will_link;
1035 ciSignature* declared_signature = NULL;
1036 ciMethod* callee = iter.get_method(ignored_will_link, &declared_signature);
1037 assert(declared_signature != NULL, "cannot be null");
1038 // (Do not use ciMethod::arg_size(), because
1039 // it might be an unloaded method, which doesn't
1040 // know whether it is static or not.)
1041 if (for_parse) {
1042 // Case 1: When called from parse we are *before* the invoke (in the
1043 // caller) and need to to adjust the inputs by an appendix
1044 // argument that will be pushed implicitly.
1045 inputs = callee->invoke_arg_size(code) - (iter.has_appendix() ? 1 : 0);
1046 } else {
1047 // Case 2: Here we are *after* the invoke (in the callee) and need to
1048 // remove any appendix arguments that were popped.
1049 inputs = callee->invoke_arg_size(code) - (callee->has_member_arg() ? 1 : 0);
1050 }
1051 int size = declared_signature->return_type()->size();
1052 depth = size - inputs;
1053 }
1054 break;
1056 case Bytecodes::_multianewarray:
1057 {
1058 ciBytecodeStream iter(method());
1059 iter.reset_to_bci(bci());
1060 iter.next();
1061 inputs = iter.get_dimensions();
1062 assert(rsize == 1, "");
1063 depth = rsize - inputs;
1064 }
1065 break;
1067 case Bytecodes::_ireturn:
1068 case Bytecodes::_lreturn:
1069 case Bytecodes::_freturn:
1070 case Bytecodes::_dreturn:
1071 case Bytecodes::_areturn:
1072 assert(rsize = -depth, "");
1073 inputs = rsize;
1074 break;
1076 case Bytecodes::_jsr:
1077 case Bytecodes::_jsr_w:
1078 inputs = 0;
1079 depth = 1; // S.B. depth=1, not zero
1080 break;
1082 default:
1083 // bytecode produces a typed result
1084 inputs = rsize - depth;
1085 assert(inputs >= 0, "");
1086 break;
1087 }
1089 #ifdef ASSERT
1090 // spot check
1091 int outputs = depth + inputs;
1092 assert(outputs >= 0, "sanity");
1093 switch (code) {
1094 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1095 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1096 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1097 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1098 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1099 }
1100 #endif //ASSERT
1102 return true;
1103 }
1107 //------------------------------basic_plus_adr---------------------------------
1108 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1109 // short-circuit a common case
1110 if (offset == intcon(0)) return ptr;
1111 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
1112 }
1114 Node* GraphKit::ConvI2L(Node* offset) {
1115 // short-circuit a common case
1116 jint offset_con = find_int_con(offset, Type::OffsetBot);
1117 if (offset_con != Type::OffsetBot) {
1118 return longcon((long) offset_con);
1119 }
1120 return _gvn.transform( new (C, 2) ConvI2LNode(offset));
1121 }
1122 Node* GraphKit::ConvL2I(Node* offset) {
1123 // short-circuit a common case
1124 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1125 if (offset_con != (jlong)Type::OffsetBot) {
1126 return intcon((int) offset_con);
1127 }
1128 return _gvn.transform( new (C, 2) ConvL2INode(offset));
1129 }
1131 //-------------------------load_object_klass-----------------------------------
1132 Node* GraphKit::load_object_klass(Node* obj) {
1133 // Special-case a fresh allocation to avoid building nodes:
1134 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1135 if (akls != NULL) return akls;
1136 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1137 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1138 }
1140 //-------------------------load_array_length-----------------------------------
1141 Node* GraphKit::load_array_length(Node* array) {
1142 // Special-case a fresh allocation to avoid building nodes:
1143 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1144 Node *alen;
1145 if (alloc == NULL) {
1146 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1147 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1148 } else {
1149 alen = alloc->Ideal_length();
1150 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1151 if (ccast != alen) {
1152 alen = _gvn.transform(ccast);
1153 }
1154 }
1155 return alen;
1156 }
1158 //------------------------------do_null_check----------------------------------
1159 // Helper function to do a NULL pointer check. Returned value is
1160 // the incoming address with NULL casted away. You are allowed to use the
1161 // not-null value only if you are control dependent on the test.
1162 extern int explicit_null_checks_inserted,
1163 explicit_null_checks_elided;
1164 Node* GraphKit::null_check_common(Node* value, BasicType type,
1165 // optional arguments for variations:
1166 bool assert_null,
1167 Node* *null_control) {
1168 assert(!assert_null || null_control == NULL, "not both at once");
1169 if (stopped()) return top();
1170 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1171 // For some performance testing, we may wish to suppress null checking.
1172 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1173 return value;
1174 }
1175 explicit_null_checks_inserted++;
1177 // Construct NULL check
1178 Node *chk = NULL;
1179 switch(type) {
1180 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1181 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
1182 case T_ARRAY : // fall through
1183 type = T_OBJECT; // simplify further tests
1184 case T_OBJECT : {
1185 const Type *t = _gvn.type( value );
1187 const TypeOopPtr* tp = t->isa_oopptr();
1188 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1189 // Only for do_null_check, not any of its siblings:
1190 && !assert_null && null_control == NULL) {
1191 // Usually, any field access or invocation on an unloaded oop type
1192 // will simply fail to link, since the statically linked class is
1193 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1194 // the static class is loaded but the sharper oop type is not.
1195 // Rather than checking for this obscure case in lots of places,
1196 // we simply observe that a null check on an unloaded class
1197 // will always be followed by a nonsense operation, so we
1198 // can just issue the uncommon trap here.
1199 // Our access to the unloaded class will only be correct
1200 // after it has been loaded and initialized, which requires
1201 // a trip through the interpreter.
1202 #ifndef PRODUCT
1203 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1204 #endif
1205 uncommon_trap(Deoptimization::Reason_unloaded,
1206 Deoptimization::Action_reinterpret,
1207 tp->klass(), "!loaded");
1208 return top();
1209 }
1211 if (assert_null) {
1212 // See if the type is contained in NULL_PTR.
1213 // If so, then the value is already null.
1214 if (t->higher_equal(TypePtr::NULL_PTR)) {
1215 explicit_null_checks_elided++;
1216 return value; // Elided null assert quickly!
1217 }
1218 } else {
1219 // See if mixing in the NULL pointer changes type.
1220 // If so, then the NULL pointer was not allowed in the original
1221 // type. In other words, "value" was not-null.
1222 if (t->meet(TypePtr::NULL_PTR) != t) {
1223 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1224 explicit_null_checks_elided++;
1225 return value; // Elided null check quickly!
1226 }
1227 }
1228 chk = new (C, 3) CmpPNode( value, null() );
1229 break;
1230 }
1232 default : ShouldNotReachHere();
1233 }
1234 assert(chk != NULL, "sanity check");
1235 chk = _gvn.transform(chk);
1237 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1238 BoolNode *btst = new (C, 2) BoolNode( chk, btest);
1239 Node *tst = _gvn.transform( btst );
1241 //-----------
1242 // if peephole optimizations occurred, a prior test existed.
1243 // If a prior test existed, maybe it dominates as we can avoid this test.
1244 if (tst != btst && type == T_OBJECT) {
1245 // At this point we want to scan up the CFG to see if we can
1246 // find an identical test (and so avoid this test altogether).
1247 Node *cfg = control();
1248 int depth = 0;
1249 while( depth < 16 ) { // Limit search depth for speed
1250 if( cfg->Opcode() == Op_IfTrue &&
1251 cfg->in(0)->in(1) == tst ) {
1252 // Found prior test. Use "cast_not_null" to construct an identical
1253 // CastPP (and hence hash to) as already exists for the prior test.
1254 // Return that casted value.
1255 if (assert_null) {
1256 replace_in_map(value, null());
1257 return null(); // do not issue the redundant test
1258 }
1259 Node *oldcontrol = control();
1260 set_control(cfg);
1261 Node *res = cast_not_null(value);
1262 set_control(oldcontrol);
1263 explicit_null_checks_elided++;
1264 return res;
1265 }
1266 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1267 if (cfg == NULL) break; // Quit at region nodes
1268 depth++;
1269 }
1270 }
1272 //-----------
1273 // Branch to failure if null
1274 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1275 Deoptimization::DeoptReason reason;
1276 if (assert_null)
1277 reason = Deoptimization::Reason_null_assert;
1278 else if (type == T_OBJECT)
1279 reason = Deoptimization::Reason_null_check;
1280 else
1281 reason = Deoptimization::Reason_div0_check;
1283 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1284 // ciMethodData::has_trap_at will return a conservative -1 if any
1285 // must-be-null assertion has failed. This could cause performance
1286 // problems for a method after its first do_null_assert failure.
1287 // Consider using 'Reason_class_check' instead?
1289 // To cause an implicit null check, we set the not-null probability
1290 // to the maximum (PROB_MAX). For an explicit check the probability
1291 // is set to a smaller value.
1292 if (null_control != NULL || too_many_traps(reason)) {
1293 // probability is less likely
1294 ok_prob = PROB_LIKELY_MAG(3);
1295 } else if (!assert_null &&
1296 (ImplicitNullCheckThreshold > 0) &&
1297 method() != NULL &&
1298 (method()->method_data()->trap_count(reason)
1299 >= (uint)ImplicitNullCheckThreshold)) {
1300 ok_prob = PROB_LIKELY_MAG(3);
1301 }
1303 if (null_control != NULL) {
1304 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1305 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
1306 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff)));
1307 if (null_true == top())
1308 explicit_null_checks_elided++;
1309 (*null_control) = null_true;
1310 } else {
1311 BuildCutout unless(this, tst, ok_prob);
1312 // Check for optimizer eliding test at parse time
1313 if (stopped()) {
1314 // Failure not possible; do not bother making uncommon trap.
1315 explicit_null_checks_elided++;
1316 } else if (assert_null) {
1317 uncommon_trap(reason,
1318 Deoptimization::Action_make_not_entrant,
1319 NULL, "assert_null");
1320 } else {
1321 replace_in_map(value, zerocon(type));
1322 builtin_throw(reason);
1323 }
1324 }
1326 // Must throw exception, fall-thru not possible?
1327 if (stopped()) {
1328 return top(); // No result
1329 }
1331 if (assert_null) {
1332 // Cast obj to null on this path.
1333 replace_in_map(value, zerocon(type));
1334 return zerocon(type);
1335 }
1337 // Cast obj to not-null on this path, if there is no null_control.
1338 // (If there is a null_control, a non-null value may come back to haunt us.)
1339 if (type == T_OBJECT) {
1340 Node* cast = cast_not_null(value, false);
1341 if (null_control == NULL || (*null_control) == top())
1342 replace_in_map(value, cast);
1343 value = cast;
1344 }
1346 return value;
1347 }
1350 //------------------------------cast_not_null----------------------------------
1351 // Cast obj to not-null on this path
1352 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1353 const Type *t = _gvn.type(obj);
1354 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1355 // Object is already not-null?
1356 if( t == t_not_null ) return obj;
1358 Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
1359 cast->init_req(0, control());
1360 cast = _gvn.transform( cast );
1362 // Scan for instances of 'obj' in the current JVM mapping.
1363 // These instances are known to be not-null after the test.
1364 if (do_replace_in_map)
1365 replace_in_map(obj, cast);
1367 return cast; // Return casted value
1368 }
1371 //--------------------------replace_in_map-------------------------------------
1372 void GraphKit::replace_in_map(Node* old, Node* neww) {
1373 this->map()->replace_edge(old, neww);
1375 // Note: This operation potentially replaces any edge
1376 // on the map. This includes locals, stack, and monitors
1377 // of the current (innermost) JVM state.
1379 // We can consider replacing in caller maps.
1380 // The idea would be that an inlined function's null checks
1381 // can be shared with the entire inlining tree.
1382 // The expense of doing this is that the PreserveJVMState class
1383 // would have to preserve caller states too, with a deep copy.
1384 }
1387 //=============================================================================
1388 //--------------------------------memory---------------------------------------
1389 Node* GraphKit::memory(uint alias_idx) {
1390 MergeMemNode* mem = merged_memory();
1391 Node* p = mem->memory_at(alias_idx);
1392 _gvn.set_type(p, Type::MEMORY); // must be mapped
1393 return p;
1394 }
1396 //-----------------------------reset_memory------------------------------------
1397 Node* GraphKit::reset_memory() {
1398 Node* mem = map()->memory();
1399 // do not use this node for any more parsing!
1400 debug_only( map()->set_memory((Node*)NULL) );
1401 return _gvn.transform( mem );
1402 }
1404 //------------------------------set_all_memory---------------------------------
1405 void GraphKit::set_all_memory(Node* newmem) {
1406 Node* mergemem = MergeMemNode::make(C, newmem);
1407 gvn().set_type_bottom(mergemem);
1408 map()->set_memory(mergemem);
1409 }
1411 //------------------------------set_all_memory_call----------------------------
1412 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1413 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1414 set_all_memory(newmem);
1415 }
1417 //=============================================================================
1418 //
1419 // parser factory methods for MemNodes
1420 //
1421 // These are layered on top of the factory methods in LoadNode and StoreNode,
1422 // and integrate with the parser's memory state and _gvn engine.
1423 //
1425 // factory methods in "int adr_idx"
1426 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1427 int adr_idx,
1428 bool require_atomic_access) {
1429 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1430 const TypePtr* adr_type = NULL; // debug-mode-only argument
1431 debug_only(adr_type = C->get_adr_type(adr_idx));
1432 Node* mem = memory(adr_idx);
1433 Node* ld;
1434 if (require_atomic_access && bt == T_LONG) {
1435 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1436 } else {
1437 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1438 }
1439 return _gvn.transform(ld);
1440 }
1442 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1443 int adr_idx,
1444 bool require_atomic_access) {
1445 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1446 const TypePtr* adr_type = NULL;
1447 debug_only(adr_type = C->get_adr_type(adr_idx));
1448 Node *mem = memory(adr_idx);
1449 Node* st;
1450 if (require_atomic_access && bt == T_LONG) {
1451 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1452 } else {
1453 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1454 }
1455 st = _gvn.transform(st);
1456 set_memory(st, adr_idx);
1457 // Back-to-back stores can only remove intermediate store with DU info
1458 // so push on worklist for optimizer.
1459 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1460 record_for_igvn(st);
1462 return st;
1463 }
1466 void GraphKit::pre_barrier(bool do_load,
1467 Node* ctl,
1468 Node* obj,
1469 Node* adr,
1470 uint adr_idx,
1471 Node* val,
1472 const TypeOopPtr* val_type,
1473 Node* pre_val,
1474 BasicType bt) {
1476 BarrierSet* bs = Universe::heap()->barrier_set();
1477 set_control(ctl);
1478 switch (bs->kind()) {
1479 case BarrierSet::G1SATBCT:
1480 case BarrierSet::G1SATBCTLogging:
1481 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1482 break;
1484 case BarrierSet::CardTableModRef:
1485 case BarrierSet::CardTableExtension:
1486 case BarrierSet::ModRef:
1487 break;
1489 case BarrierSet::Other:
1490 default :
1491 ShouldNotReachHere();
1493 }
1494 }
1496 void GraphKit::post_barrier(Node* ctl,
1497 Node* store,
1498 Node* obj,
1499 Node* adr,
1500 uint adr_idx,
1501 Node* val,
1502 BasicType bt,
1503 bool use_precise) {
1504 BarrierSet* bs = Universe::heap()->barrier_set();
1505 set_control(ctl);
1506 switch (bs->kind()) {
1507 case BarrierSet::G1SATBCT:
1508 case BarrierSet::G1SATBCTLogging:
1509 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1510 break;
1512 case BarrierSet::CardTableModRef:
1513 case BarrierSet::CardTableExtension:
1514 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1515 break;
1517 case BarrierSet::ModRef:
1518 break;
1520 case BarrierSet::Other:
1521 default :
1522 ShouldNotReachHere();
1524 }
1525 }
1527 Node* GraphKit::store_oop(Node* ctl,
1528 Node* obj,
1529 Node* adr,
1530 const TypePtr* adr_type,
1531 Node* val,
1532 const TypeOopPtr* val_type,
1533 BasicType bt,
1534 bool use_precise) {
1535 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1536 // could be delayed during Parse (for example, in adjust_map_after_if()).
1537 // Execute transformation here to avoid barrier generation in such case.
1538 if (_gvn.type(val) == TypePtr::NULL_PTR)
1539 val = _gvn.makecon(TypePtr::NULL_PTR);
1541 set_control(ctl);
1542 if (stopped()) return top(); // Dead path ?
1544 assert(bt == T_OBJECT, "sanity");
1545 assert(val != NULL, "not dead path");
1546 uint adr_idx = C->get_alias_index(adr_type);
1547 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1549 pre_barrier(true /* do_load */,
1550 control(), obj, adr, adr_idx, val, val_type,
1551 NULL /* pre_val */,
1552 bt);
1554 Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1555 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1556 return store;
1557 }
1559 // Could be an array or object we don't know at compile time (unsafe ref.)
1560 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1561 Node* obj, // containing obj
1562 Node* adr, // actual adress to store val at
1563 const TypePtr* adr_type,
1564 Node* val,
1565 BasicType bt) {
1566 Compile::AliasType* at = C->alias_type(adr_type);
1567 const TypeOopPtr* val_type = NULL;
1568 if (adr_type->isa_instptr()) {
1569 if (at->field() != NULL) {
1570 // known field. This code is a copy of the do_put_xxx logic.
1571 ciField* field = at->field();
1572 if (!field->type()->is_loaded()) {
1573 val_type = TypeInstPtr::BOTTOM;
1574 } else {
1575 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1576 }
1577 }
1578 } else if (adr_type->isa_aryptr()) {
1579 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1580 }
1581 if (val_type == NULL) {
1582 val_type = TypeInstPtr::BOTTOM;
1583 }
1584 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true);
1585 }
1588 //-------------------------array_element_address-------------------------
1589 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1590 const TypeInt* sizetype) {
1591 uint shift = exact_log2(type2aelembytes(elembt));
1592 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1594 // short-circuit a common case (saves lots of confusing waste motion)
1595 jint idx_con = find_int_con(idx, -1);
1596 if (idx_con >= 0) {
1597 intptr_t offset = header + ((intptr_t)idx_con << shift);
1598 return basic_plus_adr(ary, offset);
1599 }
1601 // must be correct type for alignment purposes
1602 Node* base = basic_plus_adr(ary, header);
1603 #ifdef _LP64
1604 // The scaled index operand to AddP must be a clean 64-bit value.
1605 // Java allows a 32-bit int to be incremented to a negative
1606 // value, which appears in a 64-bit register as a large
1607 // positive number. Using that large positive number as an
1608 // operand in pointer arithmetic has bad consequences.
1609 // On the other hand, 32-bit overflow is rare, and the possibility
1610 // can often be excluded, if we annotate the ConvI2L node with
1611 // a type assertion that its value is known to be a small positive
1612 // number. (The prior range check has ensured this.)
1613 // This assertion is used by ConvI2LNode::Ideal.
1614 int index_max = max_jint - 1; // array size is max_jint, index is one less
1615 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1616 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1617 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
1618 #endif
1619 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
1620 return basic_plus_adr(ary, base, scale);
1621 }
1623 //-------------------------load_array_element-------------------------
1624 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1625 const Type* elemtype = arytype->elem();
1626 BasicType elembt = elemtype->array_element_basic_type();
1627 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1628 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1629 return ld;
1630 }
1632 //-------------------------set_arguments_for_java_call-------------------------
1633 // Arguments (pre-popped from the stack) are taken from the JVMS.
1634 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1635 // Add the call arguments:
1636 uint nargs = call->method()->arg_size();
1637 for (uint i = 0; i < nargs; i++) {
1638 Node* arg = argument(i);
1639 call->init_req(i + TypeFunc::Parms, arg);
1640 }
1641 }
1643 //---------------------------set_edges_for_java_call---------------------------
1644 // Connect a newly created call into the current JVMS.
1645 // A return value node (if any) is returned from set_edges_for_java_call.
1646 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1648 // Add the predefined inputs:
1649 call->init_req( TypeFunc::Control, control() );
1650 call->init_req( TypeFunc::I_O , i_o() );
1651 call->init_req( TypeFunc::Memory , reset_memory() );
1652 call->init_req( TypeFunc::FramePtr, frameptr() );
1653 call->init_req( TypeFunc::ReturnAdr, top() );
1655 add_safepoint_edges(call, must_throw);
1657 Node* xcall = _gvn.transform(call);
1659 if (xcall == top()) {
1660 set_control(top());
1661 return;
1662 }
1663 assert(xcall == call, "call identity is stable");
1665 // Re-use the current map to produce the result.
1667 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
1668 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1669 set_all_memory_call(xcall, separate_io_proj);
1671 //return xcall; // no need, caller already has it
1672 }
1674 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1675 if (stopped()) return top(); // maybe the call folded up?
1677 // Capture the return value, if any.
1678 Node* ret;
1679 if (call->method() == NULL ||
1680 call->method()->return_type()->basic_type() == T_VOID)
1681 ret = top();
1682 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1684 // Note: Since any out-of-line call can produce an exception,
1685 // we always insert an I_O projection from the call into the result.
1687 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1689 if (separate_io_proj) {
1690 // The caller requested separate projections be used by the fall
1691 // through and exceptional paths, so replace the projections for
1692 // the fall through path.
1693 set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) ));
1694 set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ));
1695 }
1696 return ret;
1697 }
1699 //--------------------set_predefined_input_for_runtime_call--------------------
1700 // Reading and setting the memory state is way conservative here.
1701 // The real problem is that I am not doing real Type analysis on memory,
1702 // so I cannot distinguish card mark stores from other stores. Across a GC
1703 // point the Store Barrier and the card mark memory has to agree. I cannot
1704 // have a card mark store and its barrier split across the GC point from
1705 // either above or below. Here I get that to happen by reading ALL of memory.
1706 // A better answer would be to separate out card marks from other memory.
1707 // For now, return the input memory state, so that it can be reused
1708 // after the call, if this call has restricted memory effects.
1709 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1710 // Set fixed predefined input arguments
1711 Node* memory = reset_memory();
1712 call->init_req( TypeFunc::Control, control() );
1713 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1714 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1715 call->init_req( TypeFunc::FramePtr, frameptr() );
1716 call->init_req( TypeFunc::ReturnAdr, top() );
1717 return memory;
1718 }
1720 //-------------------set_predefined_output_for_runtime_call--------------------
1721 // Set control and memory (not i_o) from the call.
1722 // If keep_mem is not NULL, use it for the output state,
1723 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1724 // If hook_mem is NULL, this call produces no memory effects at all.
1725 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1726 // then only that memory slice is taken from the call.
1727 // In the last case, we must put an appropriate memory barrier before
1728 // the call, so as to create the correct anti-dependencies on loads
1729 // preceding the call.
1730 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1731 Node* keep_mem,
1732 const TypePtr* hook_mem) {
1733 // no i/o
1734 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
1735 if (keep_mem) {
1736 // First clone the existing memory state
1737 set_all_memory(keep_mem);
1738 if (hook_mem != NULL) {
1739 // Make memory for the call
1740 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1741 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1742 // We also use hook_mem to extract specific effects from arraycopy stubs.
1743 set_memory(mem, hook_mem);
1744 }
1745 // ...else the call has NO memory effects.
1747 // Make sure the call advertises its memory effects precisely.
1748 // This lets us build accurate anti-dependences in gcm.cpp.
1749 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1750 "call node must be constructed correctly");
1751 } else {
1752 assert(hook_mem == NULL, "");
1753 // This is not a "slow path" call; all memory comes from the call.
1754 set_all_memory_call(call);
1755 }
1756 }
1759 // Replace the call with the current state of the kit.
1760 void GraphKit::replace_call(CallNode* call, Node* result) {
1761 JVMState* ejvms = NULL;
1762 if (has_exceptions()) {
1763 ejvms = transfer_exceptions_into_jvms();
1764 }
1766 SafePointNode* final_state = stop();
1768 // Find all the needed outputs of this call
1769 CallProjections callprojs;
1770 call->extract_projections(&callprojs, true);
1772 // Replace all the old call edges with the edges from the inlining result
1773 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control));
1774 C->gvn_replace_by(callprojs.fallthrough_memproj, final_state->in(TypeFunc::Memory));
1775 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_state->in(TypeFunc::I_O));
1776 Node* final_mem = final_state->in(TypeFunc::Memory);
1778 // Replace the result with the new result if it exists and is used
1779 if (callprojs.resproj != NULL && result != NULL) {
1780 C->gvn_replace_by(callprojs.resproj, result);
1781 }
1783 if (ejvms == NULL) {
1784 // No exception edges to simply kill off those paths
1785 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1786 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1787 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1789 // Replace the old exception object with top
1790 if (callprojs.exobj != NULL) {
1791 C->gvn_replace_by(callprojs.exobj, C->top());
1792 }
1793 } else {
1794 GraphKit ekit(ejvms);
1796 // Load my combined exception state into the kit, with all phis transformed:
1797 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1799 Node* ex_oop = ekit.use_exception_state(ex_map);
1801 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1802 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1803 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1805 // Replace the old exception object with the newly created one
1806 if (callprojs.exobj != NULL) {
1807 C->gvn_replace_by(callprojs.exobj, ex_oop);
1808 }
1809 }
1811 // Disconnect the call from the graph
1812 call->disconnect_inputs(NULL);
1813 C->gvn_replace_by(call, C->top());
1815 // Clean up any MergeMems that feed other MergeMems since the
1816 // optimizer doesn't like that.
1817 if (final_mem->is_MergeMem()) {
1818 Node_List wl;
1819 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1820 Node* m = i.get();
1821 if (m->is_MergeMem() && !wl.contains(m)) {
1822 wl.push(m);
1823 }
1824 }
1825 while (wl.size() > 0) {
1826 _gvn.transform(wl.pop());
1827 }
1828 }
1829 }
1832 //------------------------------increment_counter------------------------------
1833 // for statistics: increment a VM counter by 1
1835 void GraphKit::increment_counter(address counter_addr) {
1836 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1837 increment_counter(adr1);
1838 }
1840 void GraphKit::increment_counter(Node* counter_addr) {
1841 int adr_type = Compile::AliasIdxRaw;
1842 Node* ctrl = control();
1843 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type);
1844 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
1845 store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type );
1846 }
1849 //------------------------------uncommon_trap----------------------------------
1850 // Bail out to the interpreter in mid-method. Implemented by calling the
1851 // uncommon_trap blob. This helper function inserts a runtime call with the
1852 // right debug info.
1853 void GraphKit::uncommon_trap(int trap_request,
1854 ciKlass* klass, const char* comment,
1855 bool must_throw,
1856 bool keep_exact_action) {
1857 if (failing()) stop();
1858 if (stopped()) return; // trap reachable?
1860 // Note: If ProfileTraps is true, and if a deopt. actually
1861 // occurs here, the runtime will make sure an MDO exists. There is
1862 // no need to call method()->ensure_method_data() at this point.
1864 #ifdef ASSERT
1865 if (!must_throw) {
1866 // Make sure the stack has at least enough depth to execute
1867 // the current bytecode.
1868 int inputs, ignore;
1869 if (compute_stack_effects(inputs, ignore)) {
1870 assert(sp() >= inputs, "must have enough JVMS stack to execute");
1871 // It is a frequent error in library_call.cpp to issue an
1872 // uncommon trap with the _sp value already popped.
1873 }
1874 }
1875 #endif
1877 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1878 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1880 switch (action) {
1881 case Deoptimization::Action_maybe_recompile:
1882 case Deoptimization::Action_reinterpret:
1883 // Temporary fix for 6529811 to allow virtual calls to be sure they
1884 // get the chance to go from mono->bi->mega
1885 if (!keep_exact_action &&
1886 Deoptimization::trap_request_index(trap_request) < 0 &&
1887 too_many_recompiles(reason)) {
1888 // This BCI is causing too many recompilations.
1889 action = Deoptimization::Action_none;
1890 trap_request = Deoptimization::make_trap_request(reason, action);
1891 } else {
1892 C->set_trap_can_recompile(true);
1893 }
1894 break;
1895 case Deoptimization::Action_make_not_entrant:
1896 C->set_trap_can_recompile(true);
1897 break;
1898 #ifdef ASSERT
1899 case Deoptimization::Action_none:
1900 case Deoptimization::Action_make_not_compilable:
1901 break;
1902 default:
1903 assert(false, "bad action");
1904 #endif
1905 }
1907 if (TraceOptoParse) {
1908 char buf[100];
1909 tty->print_cr("Uncommon trap %s at bci:%d",
1910 Deoptimization::format_trap_request(buf, sizeof(buf),
1911 trap_request), bci());
1912 }
1914 CompileLog* log = C->log();
1915 if (log != NULL) {
1916 int kid = (klass == NULL)? -1: log->identify(klass);
1917 log->begin_elem("uncommon_trap bci='%d'", bci());
1918 char buf[100];
1919 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1920 trap_request));
1921 if (kid >= 0) log->print(" klass='%d'", kid);
1922 if (comment != NULL) log->print(" comment='%s'", comment);
1923 log->end_elem();
1924 }
1926 // Make sure any guarding test views this path as very unlikely
1927 Node *i0 = control()->in(0);
1928 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1929 IfNode *iff = i0->as_If();
1930 float f = iff->_prob; // Get prob
1931 if (control()->Opcode() == Op_IfTrue) {
1932 if (f > PROB_UNLIKELY_MAG(4))
1933 iff->_prob = PROB_MIN;
1934 } else {
1935 if (f < PROB_LIKELY_MAG(4))
1936 iff->_prob = PROB_MAX;
1937 }
1938 }
1940 // Clear out dead values from the debug info.
1941 kill_dead_locals();
1943 // Now insert the uncommon trap subroutine call
1944 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
1945 const TypePtr* no_memory_effects = NULL;
1946 // Pass the index of the class to be loaded
1947 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1948 (must_throw ? RC_MUST_THROW : 0),
1949 OptoRuntime::uncommon_trap_Type(),
1950 call_addr, "uncommon_trap", no_memory_effects,
1951 intcon(trap_request));
1952 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1953 "must extract request correctly from the graph");
1954 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1956 call->set_req(TypeFunc::ReturnAdr, returnadr());
1957 // The debug info is the only real input to this call.
1959 // Halt-and-catch fire here. The above call should never return!
1960 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
1961 _gvn.set_type_bottom(halt);
1962 root()->add_req(halt);
1964 stop_and_kill_map();
1965 }
1968 //--------------------------just_allocated_object------------------------------
1969 // Report the object that was just allocated.
1970 // It must be the case that there are no intervening safepoints.
1971 // We use this to determine if an object is so "fresh" that
1972 // it does not require card marks.
1973 Node* GraphKit::just_allocated_object(Node* current_control) {
1974 if (C->recent_alloc_ctl() == current_control)
1975 return C->recent_alloc_obj();
1976 return NULL;
1977 }
1980 void GraphKit::round_double_arguments(ciMethod* dest_method) {
1981 // (Note: TypeFunc::make has a cache that makes this fast.)
1982 const TypeFunc* tf = TypeFunc::make(dest_method);
1983 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
1984 for (int j = 0; j < nargs; j++) {
1985 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
1986 if( targ->basic_type() == T_DOUBLE ) {
1987 // If any parameters are doubles, they must be rounded before
1988 // the call, dstore_rounding does gvn.transform
1989 Node *arg = argument(j);
1990 arg = dstore_rounding(arg);
1991 set_argument(j, arg);
1992 }
1993 }
1994 }
1996 void GraphKit::round_double_result(ciMethod* dest_method) {
1997 // A non-strict method may return a double value which has an extended
1998 // exponent, but this must not be visible in a caller which is 'strict'
1999 // If a strict caller invokes a non-strict callee, round a double result
2001 BasicType result_type = dest_method->return_type()->basic_type();
2002 assert( method() != NULL, "must have caller context");
2003 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2004 // Destination method's return value is on top of stack
2005 // dstore_rounding() does gvn.transform
2006 Node *result = pop_pair();
2007 result = dstore_rounding(result);
2008 push_pair(result);
2009 }
2010 }
2012 // rounding for strict float precision conformance
2013 Node* GraphKit::precision_rounding(Node* n) {
2014 return UseStrictFP && _method->flags().is_strict()
2015 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2016 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
2017 : n;
2018 }
2020 // rounding for strict double precision conformance
2021 Node* GraphKit::dprecision_rounding(Node *n) {
2022 return UseStrictFP && _method->flags().is_strict()
2023 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2024 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
2025 : n;
2026 }
2028 // rounding for non-strict double stores
2029 Node* GraphKit::dstore_rounding(Node* n) {
2030 return Matcher::strict_fp_requires_explicit_rounding
2031 && UseSSE <= 1
2032 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
2033 : n;
2034 }
2036 //=============================================================================
2037 // Generate a fast path/slow path idiom. Graph looks like:
2038 // [foo] indicates that 'foo' is a parameter
2039 //
2040 // [in] NULL
2041 // \ /
2042 // CmpP
2043 // Bool ne
2044 // If
2045 // / \
2046 // True False-<2>
2047 // / |
2048 // / cast_not_null
2049 // Load | | ^
2050 // [fast_test] | |
2051 // gvn to opt_test | |
2052 // / \ | <1>
2053 // True False |
2054 // | \\ |
2055 // [slow_call] \[fast_result]
2056 // Ctl Val \ \
2057 // | \ \
2058 // Catch <1> \ \
2059 // / \ ^ \ \
2060 // Ex No_Ex | \ \
2061 // | \ \ | \ <2> \
2062 // ... \ [slow_res] | | \ [null_result]
2063 // \ \--+--+--- | |
2064 // \ | / \ | /
2065 // --------Region Phi
2066 //
2067 //=============================================================================
2068 // Code is structured as a series of driver functions all called 'do_XXX' that
2069 // call a set of helper functions. Helper functions first, then drivers.
2071 //------------------------------null_check_oop---------------------------------
2072 // Null check oop. Set null-path control into Region in slot 3.
2073 // Make a cast-not-nullness use the other not-null control. Return cast.
2074 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2075 bool never_see_null) {
2076 // Initial NULL check taken path
2077 (*null_control) = top();
2078 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2080 // Generate uncommon_trap:
2081 if (never_see_null && (*null_control) != top()) {
2082 // If we see an unexpected null at a check-cast we record it and force a
2083 // recompile; the offending check-cast will be compiled to handle NULLs.
2084 // If we see more than one offending BCI, then all checkcasts in the
2085 // method will be compiled to handle NULLs.
2086 PreserveJVMState pjvms(this);
2087 set_control(*null_control);
2088 replace_in_map(value, null());
2089 uncommon_trap(Deoptimization::Reason_null_check,
2090 Deoptimization::Action_make_not_entrant);
2091 (*null_control) = top(); // NULL path is dead
2092 }
2094 // Cast away null-ness on the result
2095 return cast;
2096 }
2098 //------------------------------opt_iff----------------------------------------
2099 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2100 // Return slow-path control.
2101 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2102 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2104 // Fast path taken; set region slot 2
2105 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
2106 region->init_req(2,fast_taken); // Capture fast-control
2108 // Fast path not-taken, i.e. slow path
2109 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
2110 return slow_taken;
2111 }
2113 //-----------------------------make_runtime_call-------------------------------
2114 Node* GraphKit::make_runtime_call(int flags,
2115 const TypeFunc* call_type, address call_addr,
2116 const char* call_name,
2117 const TypePtr* adr_type,
2118 // The following parms are all optional.
2119 // The first NULL ends the list.
2120 Node* parm0, Node* parm1,
2121 Node* parm2, Node* parm3,
2122 Node* parm4, Node* parm5,
2123 Node* parm6, Node* parm7) {
2124 // Slow-path call
2125 int size = call_type->domain()->cnt();
2126 bool is_leaf = !(flags & RC_NO_LEAF);
2127 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2128 if (call_name == NULL) {
2129 assert(!is_leaf, "must supply name for leaf");
2130 call_name = OptoRuntime::stub_name(call_addr);
2131 }
2132 CallNode* call;
2133 if (!is_leaf) {
2134 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
2135 bci(), adr_type);
2136 } else if (flags & RC_NO_FP) {
2137 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2138 } else {
2139 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
2140 }
2142 // The following is similar to set_edges_for_java_call,
2143 // except that the memory effects of the call are restricted to AliasIdxRaw.
2145 // Slow path call has no side-effects, uses few values
2146 bool wide_in = !(flags & RC_NARROW_MEM);
2147 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2149 Node* prev_mem = NULL;
2150 if (wide_in) {
2151 prev_mem = set_predefined_input_for_runtime_call(call);
2152 } else {
2153 assert(!wide_out, "narrow in => narrow out");
2154 Node* narrow_mem = memory(adr_type);
2155 prev_mem = reset_memory();
2156 map()->set_memory(narrow_mem);
2157 set_predefined_input_for_runtime_call(call);
2158 }
2160 // Hook each parm in order. Stop looking at the first NULL.
2161 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2162 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2163 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2164 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2165 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2166 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2167 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2168 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2169 /* close each nested if ===> */ } } } } } } } }
2170 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2172 if (!is_leaf) {
2173 // Non-leaves can block and take safepoints:
2174 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2175 }
2176 // Non-leaves can throw exceptions:
2177 if (has_io) {
2178 call->set_req(TypeFunc::I_O, i_o());
2179 }
2181 if (flags & RC_UNCOMMON) {
2182 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2183 // (An "if" probability corresponds roughly to an unconditional count.
2184 // Sort of.)
2185 call->set_cnt(PROB_UNLIKELY_MAG(4));
2186 }
2188 Node* c = _gvn.transform(call);
2189 assert(c == call, "cannot disappear");
2191 if (wide_out) {
2192 // Slow path call has full side-effects.
2193 set_predefined_output_for_runtime_call(call);
2194 } else {
2195 // Slow path call has few side-effects, and/or sets few values.
2196 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2197 }
2199 if (has_io) {
2200 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
2201 }
2202 return call;
2204 }
2206 //------------------------------merge_memory-----------------------------------
2207 // Merge memory from one path into the current memory state.
2208 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2209 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2210 Node* old_slice = mms.force_memory();
2211 Node* new_slice = mms.memory2();
2212 if (old_slice != new_slice) {
2213 PhiNode* phi;
2214 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2215 phi = new_slice->as_Phi();
2216 #ifdef ASSERT
2217 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2218 old_slice = old_slice->in(new_path);
2219 // Caller is responsible for ensuring that any pre-existing
2220 // phis are already aware of old memory.
2221 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2222 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2223 #endif
2224 mms.set_memory(phi);
2225 } else {
2226 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2227 _gvn.set_type(phi, Type::MEMORY);
2228 phi->set_req(new_path, new_slice);
2229 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2230 }
2231 }
2232 }
2233 }
2235 //------------------------------make_slow_call_ex------------------------------
2236 // Make the exception handler hookups for the slow call
2237 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2238 if (stopped()) return;
2240 // Make a catch node with just two handlers: fall-through and catch-all
2241 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2242 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
2243 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2244 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2246 { PreserveJVMState pjvms(this);
2247 set_control(excp);
2248 set_i_o(i_o);
2250 if (excp != top()) {
2251 // Create an exception state also.
2252 // Use an exact type if the caller has specified a specific exception.
2253 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2254 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o);
2255 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2256 }
2257 }
2259 // Get the no-exception control from the CatchNode.
2260 set_control(norm);
2261 }
2264 //-------------------------------gen_subtype_check-----------------------------
2265 // Generate a subtyping check. Takes as input the subtype and supertype.
2266 // Returns 2 values: sets the default control() to the true path and returns
2267 // the false path. Only reads invariant memory; sets no (visible) memory.
2268 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2269 // but that's not exposed to the optimizer. This call also doesn't take in an
2270 // Object; if you wish to check an Object you need to load the Object's class
2271 // prior to coming here.
2272 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2273 // Fast check for identical types, perhaps identical constants.
2274 // The types can even be identical non-constants, in cases
2275 // involving Array.newInstance, Object.clone, etc.
2276 if (subklass == superklass)
2277 return top(); // false path is dead; no test needed.
2279 if (_gvn.type(superklass)->singleton()) {
2280 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2281 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2283 // In the common case of an exact superklass, try to fold up the
2284 // test before generating code. You may ask, why not just generate
2285 // the code and then let it fold up? The answer is that the generated
2286 // code will necessarily include null checks, which do not always
2287 // completely fold away. If they are also needless, then they turn
2288 // into a performance loss. Example:
2289 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2290 // Here, the type of 'fa' is often exact, so the store check
2291 // of fa[1]=x will fold up, without testing the nullness of x.
2292 switch (static_subtype_check(superk, subk)) {
2293 case SSC_always_false:
2294 {
2295 Node* always_fail = control();
2296 set_control(top());
2297 return always_fail;
2298 }
2299 case SSC_always_true:
2300 return top();
2301 case SSC_easy_test:
2302 {
2303 // Just do a direct pointer compare and be done.
2304 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
2305 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2306 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2307 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
2308 return _gvn.transform( new(C, 1) IfFalseNode(iff) );
2309 }
2310 case SSC_full_test:
2311 break;
2312 default:
2313 ShouldNotReachHere();
2314 }
2315 }
2317 // %%% Possible further optimization: Even if the superklass is not exact,
2318 // if the subklass is the unique subtype of the superklass, the check
2319 // will always succeed. We could leave a dependency behind to ensure this.
2321 // First load the super-klass's check-offset
2322 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2323 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2324 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2325 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2327 // Load from the sub-klass's super-class display list, or a 1-word cache of
2328 // the secondary superclass list, or a failing value with a sentinel offset
2329 // if the super-klass is an interface or exceptionally deep in the Java
2330 // hierarchy and we have to scan the secondary superclass list the hard way.
2331 // Worst-case type is a little odd: NULL is allowed as a result (usually
2332 // klass loads can never produce a NULL).
2333 Node *chk_off_X = ConvI2X(chk_off);
2334 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
2335 // For some types like interfaces the following loadKlass is from a 1-word
2336 // cache which is mutable so can't use immutable memory. Other
2337 // types load from the super-class display table which is immutable.
2338 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2339 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2341 // Compile speed common case: ARE a subtype and we canNOT fail
2342 if( superklass == nkls )
2343 return top(); // false path is dead; no test needed.
2345 // See if we get an immediate positive hit. Happens roughly 83% of the
2346 // time. Test to see if the value loaded just previously from the subklass
2347 // is exactly the superklass.
2348 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
2349 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
2350 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2351 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
2352 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
2354 // Compile speed common case: Check for being deterministic right now. If
2355 // chk_off is a constant and not equal to cacheoff then we are NOT a
2356 // subklass. In this case we need exactly the 1 test above and we can
2357 // return those results immediately.
2358 if (!might_be_cache) {
2359 Node* not_subtype_ctrl = control();
2360 set_control(iftrue1); // We need exactly the 1 test above
2361 return not_subtype_ctrl;
2362 }
2364 // Gather the various success & failures here
2365 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
2366 record_for_igvn(r_ok_subtype);
2367 RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
2368 record_for_igvn(r_not_subtype);
2370 r_ok_subtype->init_req(1, iftrue1);
2372 // Check for immediate negative hit. Happens roughly 11% of the time (which
2373 // is roughly 63% of the remaining cases). Test to see if the loaded
2374 // check-offset points into the subklass display list or the 1-element
2375 // cache. If it points to the display (and NOT the cache) and the display
2376 // missed then it's not a subtype.
2377 Node *cacheoff = _gvn.intcon(cacheoff_con);
2378 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
2379 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
2380 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2381 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
2382 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
2384 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2385 // No performance impact (too rare) but allows sharing of secondary arrays
2386 // which has some footprint reduction.
2387 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
2388 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
2389 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2390 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
2391 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
2393 // -- Roads not taken here: --
2394 // We could also have chosen to perform the self-check at the beginning
2395 // of this code sequence, as the assembler does. This would not pay off
2396 // the same way, since the optimizer, unlike the assembler, can perform
2397 // static type analysis to fold away many successful self-checks.
2398 // Non-foldable self checks work better here in second position, because
2399 // the initial primary superclass check subsumes a self-check for most
2400 // types. An exception would be a secondary type like array-of-interface,
2401 // which does not appear in its own primary supertype display.
2402 // Finally, we could have chosen to move the self-check into the
2403 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2404 // dependent manner. But it is worthwhile to have the check here,
2405 // where it can be perhaps be optimized. The cost in code space is
2406 // small (register compare, branch).
2408 // Now do a linear scan of the secondary super-klass array. Again, no real
2409 // performance impact (too rare) but it's gotta be done.
2410 // Since the code is rarely used, there is no penalty for moving it
2411 // out of line, and it can only improve I-cache density.
2412 // The decision to inline or out-of-line this final check is platform
2413 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2414 Node* psc = _gvn.transform(
2415 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
2417 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
2418 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
2419 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2420 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
2421 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
2423 // Return false path; set default control to true path.
2424 set_control( _gvn.transform(r_ok_subtype) );
2425 return _gvn.transform(r_not_subtype);
2426 }
2428 //----------------------------static_subtype_check-----------------------------
2429 // Shortcut important common cases when superklass is exact:
2430 // (0) superklass is java.lang.Object (can occur in reflective code)
2431 // (1) subklass is already limited to a subtype of superklass => always ok
2432 // (2) subklass does not overlap with superklass => always fail
2433 // (3) superklass has NO subtypes and we can check with a simple compare.
2434 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2435 if (StressReflectiveCode) {
2436 return SSC_full_test; // Let caller generate the general case.
2437 }
2439 if (superk == env()->Object_klass()) {
2440 return SSC_always_true; // (0) this test cannot fail
2441 }
2443 ciType* superelem = superk;
2444 if (superelem->is_array_klass())
2445 superelem = superelem->as_array_klass()->base_element_type();
2447 if (!subk->is_interface()) { // cannot trust static interface types yet
2448 if (subk->is_subtype_of(superk)) {
2449 return SSC_always_true; // (1) false path dead; no dynamic test needed
2450 }
2451 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2452 !superk->is_subtype_of(subk)) {
2453 return SSC_always_false;
2454 }
2455 }
2457 // If casting to an instance klass, it must have no subtypes
2458 if (superk->is_interface()) {
2459 // Cannot trust interfaces yet.
2460 // %%% S.B. superk->nof_implementors() == 1
2461 } else if (superelem->is_instance_klass()) {
2462 ciInstanceKlass* ik = superelem->as_instance_klass();
2463 if (!ik->has_subklass() && !ik->is_interface()) {
2464 if (!ik->is_final()) {
2465 // Add a dependency if there is a chance of a later subclass.
2466 C->dependencies()->assert_leaf_type(ik);
2467 }
2468 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2469 }
2470 } else {
2471 // A primitive array type has no subtypes.
2472 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2473 }
2475 return SSC_full_test;
2476 }
2478 // Profile-driven exact type check:
2479 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2480 float prob,
2481 Node* *casted_receiver) {
2482 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2483 Node* recv_klass = load_object_klass(receiver);
2484 Node* want_klass = makecon(tklass);
2485 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
2486 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2487 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2488 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
2489 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
2491 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2492 assert(recv_xtype->klass_is_exact(), "");
2494 // Subsume downstream occurrences of receiver with a cast to
2495 // recv_xtype, since now we know what the type will be.
2496 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
2497 (*casted_receiver) = _gvn.transform(cast);
2498 // (User must make the replace_in_map call.)
2500 return fail;
2501 }
2504 //------------------------------seems_never_null-------------------------------
2505 // Use null_seen information if it is available from the profile.
2506 // If we see an unexpected null at a type check we record it and force a
2507 // recompile; the offending check will be recompiled to handle NULLs.
2508 // If we see several offending BCIs, then all checks in the
2509 // method will be recompiled.
2510 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2511 if (UncommonNullCast // Cutout for this technique
2512 && obj != null() // And not the -Xcomp stupid case?
2513 && !too_many_traps(Deoptimization::Reason_null_check)
2514 ) {
2515 if (data == NULL)
2516 // Edge case: no mature data. Be optimistic here.
2517 return true;
2518 // If the profile has not seen a null, assume it won't happen.
2519 assert(java_bc() == Bytecodes::_checkcast ||
2520 java_bc() == Bytecodes::_instanceof ||
2521 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2522 return !data->as_BitData()->null_seen();
2523 }
2524 return false;
2525 }
2527 //------------------------maybe_cast_profiled_receiver-------------------------
2528 // If the profile has seen exactly one type, narrow to exactly that type.
2529 // Subsequent type checks will always fold up.
2530 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2531 ciProfileData* data,
2532 ciKlass* require_klass) {
2533 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2534 if (data == NULL) return NULL;
2536 // Make sure we haven't already deoptimized from this tactic.
2537 if (too_many_traps(Deoptimization::Reason_class_check))
2538 return NULL;
2540 // (No, this isn't a call, but it's enough like a virtual call
2541 // to use the same ciMethod accessor to get the profile info...)
2542 ciCallProfile profile = method()->call_profile_at_bci(bci());
2543 if (profile.count() >= 0 && // no cast failures here
2544 profile.has_receiver(0) &&
2545 profile.morphism() == 1) {
2546 ciKlass* exact_kls = profile.receiver(0);
2547 if (require_klass == NULL ||
2548 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2549 // If we narrow the type to match what the type profile sees,
2550 // we can then remove the rest of the cast.
2551 // This is a win, even if the exact_kls is very specific,
2552 // because downstream operations, such as method calls,
2553 // will often benefit from the sharper type.
2554 Node* exact_obj = not_null_obj; // will get updated in place...
2555 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2556 &exact_obj);
2557 { PreserveJVMState pjvms(this);
2558 set_control(slow_ctl);
2559 uncommon_trap(Deoptimization::Reason_class_check,
2560 Deoptimization::Action_maybe_recompile);
2561 }
2562 replace_in_map(not_null_obj, exact_obj);
2563 return exact_obj;
2564 }
2565 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2566 }
2568 return NULL;
2569 }
2572 //-------------------------------gen_instanceof--------------------------------
2573 // Generate an instance-of idiom. Used by both the instance-of bytecode
2574 // and the reflective instance-of call.
2575 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass) {
2576 kill_dead_locals(); // Benefit all the uncommon traps
2577 assert( !stopped(), "dead parse path should be checked in callers" );
2578 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2579 "must check for not-null not-dead klass in callers");
2581 // Make the merge point
2582 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2583 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2584 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2585 C->set_has_split_ifs(true); // Has chance for split-if optimization
2587 ciProfileData* data = NULL;
2588 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2589 data = method()->method_data()->bci_to_data(bci());
2590 }
2591 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2592 && seems_never_null(obj, data));
2594 // Null check; get casted pointer; set region slot 3
2595 Node* null_ctl = top();
2596 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2598 // If not_null_obj is dead, only null-path is taken
2599 if (stopped()) { // Doing instance-of on a NULL?
2600 set_control(null_ctl);
2601 return intcon(0);
2602 }
2603 region->init_req(_null_path, null_ctl);
2604 phi ->init_req(_null_path, intcon(0)); // Set null path value
2605 if (null_ctl == top()) {
2606 // Do this eagerly, so that pattern matches like is_diamond_phi
2607 // will work even during parsing.
2608 assert(_null_path == PATH_LIMIT-1, "delete last");
2609 region->del_req(_null_path);
2610 phi ->del_req(_null_path);
2611 }
2613 if (ProfileDynamicTypes && data != NULL) {
2614 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, NULL);
2615 if (stopped()) { // Profile disagrees with this path.
2616 set_control(null_ctl); // Null is the only remaining possibility.
2617 return intcon(0);
2618 }
2619 if (cast_obj != NULL)
2620 not_null_obj = cast_obj;
2621 }
2623 // Load the object's klass
2624 Node* obj_klass = load_object_klass(not_null_obj);
2626 // Generate the subtype check
2627 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2629 // Plug in the success path to the general merge in slot 1.
2630 region->init_req(_obj_path, control());
2631 phi ->init_req(_obj_path, intcon(1));
2633 // Plug in the failing path to the general merge in slot 2.
2634 region->init_req(_fail_path, not_subtype_ctrl);
2635 phi ->init_req(_fail_path, intcon(0));
2637 // Return final merged results
2638 set_control( _gvn.transform(region) );
2639 record_for_igvn(region);
2640 return _gvn.transform(phi);
2641 }
2643 //-------------------------------gen_checkcast---------------------------------
2644 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2645 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2646 // uncommon-trap paths work. Adjust stack after this call.
2647 // If failure_control is supplied and not null, it is filled in with
2648 // the control edge for the cast failure. Otherwise, an appropriate
2649 // uncommon trap or exception is thrown.
2650 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2651 Node* *failure_control) {
2652 kill_dead_locals(); // Benefit all the uncommon traps
2653 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2654 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2656 // Fast cutout: Check the case that the cast is vacuously true.
2657 // This detects the common cases where the test will short-circuit
2658 // away completely. We do this before we perform the null check,
2659 // because if the test is going to turn into zero code, we don't
2660 // want a residual null check left around. (Causes a slowdown,
2661 // for example, in some objArray manipulations, such as a[i]=a[j].)
2662 if (tk->singleton()) {
2663 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2664 if (objtp != NULL && objtp->klass() != NULL) {
2665 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2666 case SSC_always_true:
2667 return obj;
2668 case SSC_always_false:
2669 // It needs a null check because a null will *pass* the cast check.
2670 // A non-null value will always produce an exception.
2671 return do_null_assert(obj, T_OBJECT);
2672 }
2673 }
2674 }
2676 ciProfileData* data = NULL;
2677 if (failure_control == NULL) { // use MDO in regular case only
2678 assert(java_bc() == Bytecodes::_aastore ||
2679 java_bc() == Bytecodes::_checkcast,
2680 "interpreter profiles type checks only for these BCs");
2681 data = method()->method_data()->bci_to_data(bci());
2682 }
2684 // Make the merge point
2685 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2686 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2687 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop);
2688 C->set_has_split_ifs(true); // Has chance for split-if optimization
2690 // Use null-cast information if it is available
2691 bool never_see_null = ((failure_control == NULL) // regular case only
2692 && seems_never_null(obj, data));
2694 // Null check; get casted pointer; set region slot 3
2695 Node* null_ctl = top();
2696 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2698 // If not_null_obj is dead, only null-path is taken
2699 if (stopped()) { // Doing instance-of on a NULL?
2700 set_control(null_ctl);
2701 return null();
2702 }
2703 region->init_req(_null_path, null_ctl);
2704 phi ->init_req(_null_path, null()); // Set null path value
2705 if (null_ctl == top()) {
2706 // Do this eagerly, so that pattern matches like is_diamond_phi
2707 // will work even during parsing.
2708 assert(_null_path == PATH_LIMIT-1, "delete last");
2709 region->del_req(_null_path);
2710 phi ->del_req(_null_path);
2711 }
2713 Node* cast_obj = NULL;
2714 if (data != NULL &&
2715 // Counter has never been decremented (due to cast failure).
2716 // ...This is a reasonable thing to expect. It is true of
2717 // all casts inserted by javac to implement generic types.
2718 data->as_CounterData()->count() >= 0) {
2719 cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, tk->klass());
2720 if (cast_obj != NULL) {
2721 if (failure_control != NULL) // failure is now impossible
2722 (*failure_control) = top();
2723 // adjust the type of the phi to the exact klass:
2724 phi->raise_bottom_type(_gvn.type(cast_obj)->meet(TypePtr::NULL_PTR));
2725 }
2726 }
2728 if (cast_obj == NULL) {
2729 // Load the object's klass
2730 Node* obj_klass = load_object_klass(not_null_obj);
2732 // Generate the subtype check
2733 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2735 // Plug in success path into the merge
2736 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
2737 not_null_obj, toop));
2738 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2739 if (failure_control == NULL) {
2740 if (not_subtype_ctrl != top()) { // If failure is possible
2741 PreserveJVMState pjvms(this);
2742 set_control(not_subtype_ctrl);
2743 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2744 }
2745 } else {
2746 (*failure_control) = not_subtype_ctrl;
2747 }
2748 }
2750 region->init_req(_obj_path, control());
2751 phi ->init_req(_obj_path, cast_obj);
2753 // A merge of NULL or Casted-NotNull obj
2754 Node* res = _gvn.transform(phi);
2756 // Note I do NOT always 'replace_in_map(obj,result)' here.
2757 // if( tk->klass()->can_be_primary_super() )
2758 // This means that if I successfully store an Object into an array-of-String
2759 // I 'forget' that the Object is really now known to be a String. I have to
2760 // do this because we don't have true union types for interfaces - if I store
2761 // a Baz into an array-of-Interface and then tell the optimizer it's an
2762 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2763 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2764 // replace_in_map( obj, res );
2766 // Return final merged results
2767 set_control( _gvn.transform(region) );
2768 record_for_igvn(region);
2769 return res;
2770 }
2772 //------------------------------next_monitor-----------------------------------
2773 // What number should be given to the next monitor?
2774 int GraphKit::next_monitor() {
2775 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2776 int next = current + C->sync_stack_slots();
2777 // Keep the toplevel high water mark current:
2778 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2779 return current;
2780 }
2782 //------------------------------insert_mem_bar---------------------------------
2783 // Memory barrier to avoid floating things around
2784 // The membar serves as a pinch point between both control and all memory slices.
2785 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2786 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2787 mb->init_req(TypeFunc::Control, control());
2788 mb->init_req(TypeFunc::Memory, reset_memory());
2789 Node* membar = _gvn.transform(mb);
2790 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
2791 set_all_memory_call(membar);
2792 return membar;
2793 }
2795 //-------------------------insert_mem_bar_volatile----------------------------
2796 // Memory barrier to avoid floating things around
2797 // The membar serves as a pinch point between both control and memory(alias_idx).
2798 // If you want to make a pinch point on all memory slices, do not use this
2799 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2800 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2801 // When Parse::do_put_xxx updates a volatile field, it appends a series
2802 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2803 // The first membar is on the same memory slice as the field store opcode.
2804 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2805 // All the other membars (for other volatile slices, including AliasIdxBot,
2806 // which stands for all unknown volatile slices) are control-dependent
2807 // on the first membar. This prevents later volatile loads or stores
2808 // from sliding up past the just-emitted store.
2810 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2811 mb->set_req(TypeFunc::Control,control());
2812 if (alias_idx == Compile::AliasIdxBot) {
2813 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2814 } else {
2815 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2816 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2817 }
2818 Node* membar = _gvn.transform(mb);
2819 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
2820 if (alias_idx == Compile::AliasIdxBot) {
2821 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
2822 } else {
2823 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2824 }
2825 return membar;
2826 }
2828 //------------------------------shared_lock------------------------------------
2829 // Emit locking code.
2830 FastLockNode* GraphKit::shared_lock(Node* obj) {
2831 // bci is either a monitorenter bc or InvocationEntryBci
2832 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2833 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2835 if( !GenerateSynchronizationCode )
2836 return NULL; // Not locking things?
2837 if (stopped()) // Dead monitor?
2838 return NULL;
2840 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2842 // Box the stack location
2843 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
2844 Node* mem = reset_memory();
2846 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
2847 if (PrintPreciseBiasedLockingStatistics) {
2848 // Create the counters for this fast lock.
2849 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2850 }
2851 // Add monitor to debug info for the slow path. If we block inside the
2852 // slow path and de-opt, we need the monitor hanging around
2853 map()->push_monitor( flock );
2855 const TypeFunc *tf = LockNode::lock_type();
2856 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
2858 lock->init_req( TypeFunc::Control, control() );
2859 lock->init_req( TypeFunc::Memory , mem );
2860 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2861 lock->init_req( TypeFunc::FramePtr, frameptr() );
2862 lock->init_req( TypeFunc::ReturnAdr, top() );
2864 lock->init_req(TypeFunc::Parms + 0, obj);
2865 lock->init_req(TypeFunc::Parms + 1, box);
2866 lock->init_req(TypeFunc::Parms + 2, flock);
2867 add_safepoint_edges(lock);
2869 lock = _gvn.transform( lock )->as_Lock();
2871 // lock has no side-effects, sets few values
2872 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2874 insert_mem_bar(Op_MemBarAcquireLock);
2876 // Add this to the worklist so that the lock can be eliminated
2877 record_for_igvn(lock);
2879 #ifndef PRODUCT
2880 if (PrintLockStatistics) {
2881 // Update the counter for this lock. Don't bother using an atomic
2882 // operation since we don't require absolute accuracy.
2883 lock->create_lock_counter(map()->jvms());
2884 increment_counter(lock->counter()->addr());
2885 }
2886 #endif
2888 return flock;
2889 }
2892 //------------------------------shared_unlock----------------------------------
2893 // Emit unlocking code.
2894 void GraphKit::shared_unlock(Node* box, Node* obj) {
2895 // bci is either a monitorenter bc or InvocationEntryBci
2896 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2897 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2899 if( !GenerateSynchronizationCode )
2900 return;
2901 if (stopped()) { // Dead monitor?
2902 map()->pop_monitor(); // Kill monitor from debug info
2903 return;
2904 }
2906 // Memory barrier to avoid floating things down past the locked region
2907 insert_mem_bar(Op_MemBarReleaseLock);
2909 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2910 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
2911 uint raw_idx = Compile::AliasIdxRaw;
2912 unlock->init_req( TypeFunc::Control, control() );
2913 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2914 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2915 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2916 unlock->init_req( TypeFunc::ReturnAdr, top() );
2918 unlock->init_req(TypeFunc::Parms + 0, obj);
2919 unlock->init_req(TypeFunc::Parms + 1, box);
2920 unlock = _gvn.transform(unlock)->as_Unlock();
2922 Node* mem = reset_memory();
2924 // unlock has no side-effects, sets few values
2925 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2927 // Kill monitor from debug info
2928 map()->pop_monitor( );
2929 }
2931 //-------------------------------get_layout_helper-----------------------------
2932 // If the given klass is a constant or known to be an array,
2933 // fetch the constant layout helper value into constant_value
2934 // and return (Node*)NULL. Otherwise, load the non-constant
2935 // layout helper value, and return the node which represents it.
2936 // This two-faced routine is useful because allocation sites
2937 // almost always feature constant types.
2938 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2939 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2940 if (!StressReflectiveCode && inst_klass != NULL) {
2941 ciKlass* klass = inst_klass->klass();
2942 bool xklass = inst_klass->klass_is_exact();
2943 if (xklass || klass->is_array_klass()) {
2944 jint lhelper = klass->layout_helper();
2945 if (lhelper != Klass::_lh_neutral_value) {
2946 constant_value = lhelper;
2947 return (Node*) NULL;
2948 }
2949 }
2950 }
2951 constant_value = Klass::_lh_neutral_value; // put in a known value
2952 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
2953 return make_load(NULL, lhp, TypeInt::INT, T_INT);
2954 }
2956 // We just put in an allocate/initialize with a big raw-memory effect.
2957 // Hook selected additional alias categories on the initialization.
2958 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
2959 MergeMemNode* init_in_merge,
2960 Node* init_out_raw) {
2961 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
2962 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
2964 Node* prevmem = kit.memory(alias_idx);
2965 init_in_merge->set_memory_at(alias_idx, prevmem);
2966 kit.set_memory(init_out_raw, alias_idx);
2967 }
2969 //---------------------------set_output_for_allocation-------------------------
2970 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
2971 const TypeOopPtr* oop_type) {
2972 int rawidx = Compile::AliasIdxRaw;
2973 alloc->set_req( TypeFunc::FramePtr, frameptr() );
2974 add_safepoint_edges(alloc);
2975 Node* allocx = _gvn.transform(alloc);
2976 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
2977 // create memory projection for i_o
2978 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
2979 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
2981 // create a memory projection as for the normal control path
2982 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
2983 set_memory(malloc, rawidx);
2985 // a normal slow-call doesn't change i_o, but an allocation does
2986 // we create a separate i_o projection for the normal control path
2987 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
2988 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
2990 // put in an initialization barrier
2991 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
2992 rawoop)->as_Initialize();
2993 assert(alloc->initialization() == init, "2-way macro link must work");
2994 assert(init ->allocation() == alloc, "2-way macro link must work");
2995 {
2996 // Extract memory strands which may participate in the new object's
2997 // initialization, and source them from the new InitializeNode.
2998 // This will allow us to observe initializations when they occur,
2999 // and link them properly (as a group) to the InitializeNode.
3000 assert(init->in(InitializeNode::Memory) == malloc, "");
3001 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3002 init->set_req(InitializeNode::Memory, minit_in);
3003 record_for_igvn(minit_in); // fold it up later, if possible
3004 Node* minit_out = memory(rawidx);
3005 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3006 if (oop_type->isa_aryptr()) {
3007 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3008 int elemidx = C->get_alias_index(telemref);
3009 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3010 } else if (oop_type->isa_instptr()) {
3011 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3012 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3013 ciField* field = ik->nonstatic_field_at(i);
3014 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3015 continue; // do not bother to track really large numbers of fields
3016 // Find (or create) the alias category for this field:
3017 int fieldidx = C->alias_type(field)->index();
3018 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3019 }
3020 }
3021 }
3023 // Cast raw oop to the real thing...
3024 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
3025 javaoop = _gvn.transform(javaoop);
3026 C->set_recent_alloc(control(), javaoop);
3027 assert(just_allocated_object(control()) == javaoop, "just allocated");
3029 #ifdef ASSERT
3030 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3031 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3032 "Ideal_allocation works");
3033 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3034 "Ideal_allocation works");
3035 if (alloc->is_AllocateArray()) {
3036 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3037 "Ideal_allocation works");
3038 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3039 "Ideal_allocation works");
3040 } else {
3041 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3042 }
3043 }
3044 #endif //ASSERT
3046 return javaoop;
3047 }
3049 //---------------------------new_instance--------------------------------------
3050 // This routine takes a klass_node which may be constant (for a static type)
3051 // or may be non-constant (for reflective code). It will work equally well
3052 // for either, and the graph will fold nicely if the optimizer later reduces
3053 // the type to a constant.
3054 // The optional arguments are for specialized use by intrinsics:
3055 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3056 // - If 'return_size_val', report the the total object size to the caller.
3057 Node* GraphKit::new_instance(Node* klass_node,
3058 Node* extra_slow_test,
3059 Node* *return_size_val) {
3060 // Compute size in doublewords
3061 // The size is always an integral number of doublewords, represented
3062 // as a positive bytewise size stored in the klass's layout_helper.
3063 // The layout_helper also encodes (in a low bit) the need for a slow path.
3064 jint layout_con = Klass::_lh_neutral_value;
3065 Node* layout_val = get_layout_helper(klass_node, layout_con);
3066 int layout_is_con = (layout_val == NULL);
3068 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3069 // Generate the initial go-slow test. It's either ALWAYS (return a
3070 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3071 // case) a computed value derived from the layout_helper.
3072 Node* initial_slow_test = NULL;
3073 if (layout_is_con) {
3074 assert(!StressReflectiveCode, "stress mode does not use these paths");
3075 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3076 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3078 } else { // reflective case
3079 // This reflective path is used by Unsafe.allocateInstance.
3080 // (It may be stress-tested by specifying StressReflectiveCode.)
3081 // Basically, we want to get into the VM is there's an illegal argument.
3082 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3083 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
3084 if (extra_slow_test != intcon(0)) {
3085 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
3086 }
3087 // (Macro-expander will further convert this to a Bool, if necessary.)
3088 }
3090 // Find the size in bytes. This is easy; it's the layout_helper.
3091 // The size value must be valid even if the slow path is taken.
3092 Node* size = NULL;
3093 if (layout_is_con) {
3094 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3095 } else { // reflective case
3096 // This reflective path is used by clone and Unsafe.allocateInstance.
3097 size = ConvI2X(layout_val);
3099 // Clear the low bits to extract layout_helper_size_in_bytes:
3100 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3101 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3102 size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
3103 }
3104 if (return_size_val != NULL) {
3105 (*return_size_val) = size;
3106 }
3108 // This is a precise notnull oop of the klass.
3109 // (Actually, it need not be precise if this is a reflective allocation.)
3110 // It's what we cast the result to.
3111 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3112 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3113 const TypeOopPtr* oop_type = tklass->as_instance_type();
3115 // Now generate allocation code
3117 // The entire memory state is needed for slow path of the allocation
3118 // since GC and deoptimization can happened.
3119 Node *mem = reset_memory();
3120 set_all_memory(mem); // Create new memory state
3122 AllocateNode* alloc
3123 = new (C, AllocateNode::ParmLimit)
3124 AllocateNode(C, AllocateNode::alloc_type(),
3125 control(), mem, i_o(),
3126 size, klass_node,
3127 initial_slow_test);
3129 return set_output_for_allocation(alloc, oop_type);
3130 }
3132 //-------------------------------new_array-------------------------------------
3133 // helper for both newarray and anewarray
3134 // The 'length' parameter is (obviously) the length of the array.
3135 // See comments on new_instance for the meaning of the other arguments.
3136 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3137 Node* length, // number of array elements
3138 int nargs, // number of arguments to push back for uncommon trap
3139 Node* *return_size_val) {
3140 jint layout_con = Klass::_lh_neutral_value;
3141 Node* layout_val = get_layout_helper(klass_node, layout_con);
3142 int layout_is_con = (layout_val == NULL);
3144 if (!layout_is_con && !StressReflectiveCode &&
3145 !too_many_traps(Deoptimization::Reason_class_check)) {
3146 // This is a reflective array creation site.
3147 // Optimistically assume that it is a subtype of Object[],
3148 // so that we can fold up all the address arithmetic.
3149 layout_con = Klass::array_layout_helper(T_OBJECT);
3150 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
3151 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
3152 { BuildCutout unless(this, bol_lh, PROB_MAX);
3153 _sp += nargs;
3154 uncommon_trap(Deoptimization::Reason_class_check,
3155 Deoptimization::Action_maybe_recompile);
3156 }
3157 layout_val = NULL;
3158 layout_is_con = true;
3159 }
3161 // Generate the initial go-slow test. Make sure we do not overflow
3162 // if length is huge (near 2Gig) or negative! We do not need
3163 // exact double-words here, just a close approximation of needed
3164 // double-words. We can't add any offset or rounding bits, lest we
3165 // take a size -1 of bytes and make it positive. Use an unsigned
3166 // compare, so negative sizes look hugely positive.
3167 int fast_size_limit = FastAllocateSizeLimit;
3168 if (layout_is_con) {
3169 assert(!StressReflectiveCode, "stress mode does not use these paths");
3170 // Increase the size limit if we have exact knowledge of array type.
3171 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3172 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3173 }
3175 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
3176 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3177 if (initial_slow_test->is_Bool()) {
3178 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3179 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3180 }
3182 // --- Size Computation ---
3183 // array_size = round_to_heap(array_header + (length << elem_shift));
3184 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3185 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3186 // The rounding mask is strength-reduced, if possible.
3187 int round_mask = MinObjAlignmentInBytes - 1;
3188 Node* header_size = NULL;
3189 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3190 // (T_BYTE has the weakest alignment and size restrictions...)
3191 if (layout_is_con) {
3192 int hsize = Klass::layout_helper_header_size(layout_con);
3193 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3194 BasicType etype = Klass::layout_helper_element_type(layout_con);
3195 if ((round_mask & ~right_n_bits(eshift)) == 0)
3196 round_mask = 0; // strength-reduce it if it goes away completely
3197 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3198 assert(header_size_min <= hsize, "generic minimum is smallest");
3199 header_size_min = hsize;
3200 header_size = intcon(hsize + round_mask);
3201 } else {
3202 Node* hss = intcon(Klass::_lh_header_size_shift);
3203 Node* hsm = intcon(Klass::_lh_header_size_mask);
3204 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
3205 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
3206 Node* mask = intcon(round_mask);
3207 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
3208 }
3210 Node* elem_shift = NULL;
3211 if (layout_is_con) {
3212 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3213 if (eshift != 0)
3214 elem_shift = intcon(eshift);
3215 } else {
3216 // There is no need to mask or shift this value.
3217 // The semantics of LShiftINode include an implicit mask to 0x1F.
3218 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3219 elem_shift = layout_val;
3220 }
3222 // Transition to native address size for all offset calculations:
3223 Node* lengthx = ConvI2X(length);
3224 Node* headerx = ConvI2X(header_size);
3225 #ifdef _LP64
3226 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3227 if (tllen != NULL && tllen->_lo < 0) {
3228 // Add a manual constraint to a positive range. Cf. array_element_address.
3229 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3230 if (size_max > tllen->_hi) size_max = tllen->_hi;
3231 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3232 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
3233 }
3234 }
3235 #endif
3237 // Combine header size (plus rounding) and body size. Then round down.
3238 // This computation cannot overflow, because it is used only in two
3239 // places, one where the length is sharply limited, and the other
3240 // after a successful allocation.
3241 Node* abody = lengthx;
3242 if (elem_shift != NULL)
3243 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
3244 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
3245 if (round_mask != 0) {
3246 Node* mask = MakeConX(~round_mask);
3247 size = _gvn.transform( new(C, 3) AndXNode(size, mask) );
3248 }
3249 // else if round_mask == 0, the size computation is self-rounding
3251 if (return_size_val != NULL) {
3252 // This is the size
3253 (*return_size_val) = size;
3254 }
3256 // Now generate allocation code
3258 // The entire memory state is needed for slow path of the allocation
3259 // since GC and deoptimization can happened.
3260 Node *mem = reset_memory();
3261 set_all_memory(mem); // Create new memory state
3263 // Create the AllocateArrayNode and its result projections
3264 AllocateArrayNode* alloc
3265 = new (C, AllocateArrayNode::ParmLimit)
3266 AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3267 control(), mem, i_o(),
3268 size, klass_node,
3269 initial_slow_test,
3270 length);
3272 // Cast to correct type. Note that the klass_node may be constant or not,
3273 // and in the latter case the actual array type will be inexact also.
3274 // (This happens via a non-constant argument to inline_native_newArray.)
3275 // In any case, the value of klass_node provides the desired array type.
3276 const TypeInt* length_type = _gvn.find_int_type(length);
3277 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3278 if (ary_type->isa_aryptr() && length_type != NULL) {
3279 // Try to get a better type than POS for the size
3280 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3281 }
3283 Node* javaoop = set_output_for_allocation(alloc, ary_type);
3285 // Cast length on remaining path to be as narrow as possible
3286 if (map()->find_edge(length) >= 0) {
3287 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3288 if (ccast != length) {
3289 _gvn.set_type_bottom(ccast);
3290 record_for_igvn(ccast);
3291 replace_in_map(length, ccast);
3292 }
3293 }
3295 return javaoop;
3296 }
3298 // The following "Ideal_foo" functions are placed here because they recognize
3299 // the graph shapes created by the functions immediately above.
3301 //---------------------------Ideal_allocation----------------------------------
3302 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3303 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3304 if (ptr == NULL) { // reduce dumb test in callers
3305 return NULL;
3306 }
3307 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
3308 ptr = ptr->in(1);
3309 if (ptr == NULL) return NULL;
3310 }
3311 if (ptr->is_Proj()) {
3312 Node* allo = ptr->in(0);
3313 if (allo != NULL && allo->is_Allocate()) {
3314 return allo->as_Allocate();
3315 }
3316 }
3317 // Report failure to match.
3318 return NULL;
3319 }
3321 // Fancy version which also strips off an offset (and reports it to caller).
3322 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3323 intptr_t& offset) {
3324 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3325 if (base == NULL) return NULL;
3326 return Ideal_allocation(base, phase);
3327 }
3329 // Trace Initialize <- Proj[Parm] <- Allocate
3330 AllocateNode* InitializeNode::allocation() {
3331 Node* rawoop = in(InitializeNode::RawAddress);
3332 if (rawoop->is_Proj()) {
3333 Node* alloc = rawoop->in(0);
3334 if (alloc->is_Allocate()) {
3335 return alloc->as_Allocate();
3336 }
3337 }
3338 return NULL;
3339 }
3341 // Trace Allocate -> Proj[Parm] -> Initialize
3342 InitializeNode* AllocateNode::initialization() {
3343 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3344 if (rawoop == NULL) return NULL;
3345 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3346 Node* init = rawoop->fast_out(i);
3347 if (init->is_Initialize()) {
3348 assert(init->as_Initialize()->allocation() == this, "2-way link");
3349 return init->as_Initialize();
3350 }
3351 }
3352 return NULL;
3353 }
3355 // Trace Allocate -> Proj[Parm] -> MemBarStoreStore
3356 MemBarStoreStoreNode* AllocateNode::storestore() {
3357 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3358 if (rawoop == NULL) return NULL;
3359 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3360 Node* storestore = rawoop->fast_out(i);
3361 if (storestore->is_MemBarStoreStore()) {
3362 return storestore->as_MemBarStoreStore();
3363 }
3364 }
3365 return NULL;
3366 }
3368 //----------------------------- loop predicates ---------------------------
3370 //------------------------------add_predicate_impl----------------------------
3371 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3372 // Too many traps seen?
3373 if (too_many_traps(reason)) {
3374 #ifdef ASSERT
3375 if (TraceLoopPredicate) {
3376 int tc = C->trap_count(reason);
3377 tty->print("too many traps=%s tcount=%d in ",
3378 Deoptimization::trap_reason_name(reason), tc);
3379 method()->print(); // which method has too many predicate traps
3380 tty->cr();
3381 }
3382 #endif
3383 // We cannot afford to take more traps here,
3384 // do not generate predicate.
3385 return;
3386 }
3388 Node *cont = _gvn.intcon(1);
3389 Node* opq = _gvn.transform(new (C, 2) Opaque1Node(C, cont));
3390 Node *bol = _gvn.transform(new (C, 2) Conv2BNode(opq));
3391 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3392 Node* iffalse = _gvn.transform(new (C, 1) IfFalseNode(iff));
3393 C->add_predicate_opaq(opq);
3394 {
3395 PreserveJVMState pjvms(this);
3396 set_control(iffalse);
3397 _sp += nargs;
3398 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3399 }
3400 Node* iftrue = _gvn.transform(new (C, 1) IfTrueNode(iff));
3401 set_control(iftrue);
3402 }
3404 //------------------------------add_predicate---------------------------------
3405 void GraphKit::add_predicate(int nargs) {
3406 if (UseLoopPredicate) {
3407 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3408 }
3409 // loop's limit check predicate should be near the loop.
3410 if (LoopLimitCheck) {
3411 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3412 }
3413 }
3415 //----------------------------- store barriers ----------------------------
3416 #define __ ideal.
3418 void GraphKit::sync_kit(IdealKit& ideal) {
3419 set_all_memory(__ merged_memory());
3420 set_i_o(__ i_o());
3421 set_control(__ ctrl());
3422 }
3424 void GraphKit::final_sync(IdealKit& ideal) {
3425 // Final sync IdealKit and graphKit.
3426 __ drain_delay_transform();
3427 sync_kit(ideal);
3428 }
3430 // vanilla/CMS post barrier
3431 // Insert a write-barrier store. This is to let generational GC work; we have
3432 // to flag all oop-stores before the next GC point.
3433 void GraphKit::write_barrier_post(Node* oop_store,
3434 Node* obj,
3435 Node* adr,
3436 uint adr_idx,
3437 Node* val,
3438 bool use_precise) {
3439 // No store check needed if we're storing a NULL or an old object
3440 // (latter case is probably a string constant). The concurrent
3441 // mark sweep garbage collector, however, needs to have all nonNull
3442 // oop updates flagged via card-marks.
3443 if (val != NULL && val->is_Con()) {
3444 // must be either an oop or NULL
3445 const Type* t = val->bottom_type();
3446 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3447 // stores of null never (?) need barriers
3448 return;
3449 }
3451 if (use_ReduceInitialCardMarks()
3452 && obj == just_allocated_object(control())) {
3453 // We can skip marks on a freshly-allocated object in Eden.
3454 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3455 // That routine informs GC to take appropriate compensating steps,
3456 // upon a slow-path allocation, so as to make this card-mark
3457 // elision safe.
3458 return;
3459 }
3461 if (!use_precise) {
3462 // All card marks for a (non-array) instance are in one place:
3463 adr = obj;
3464 }
3465 // (Else it's an array (or unknown), and we want more precise card marks.)
3466 assert(adr != NULL, "");
3468 IdealKit ideal(this, true);
3470 // Convert the pointer to an int prior to doing math on it
3471 Node* cast = __ CastPX(__ ctrl(), adr);
3473 // Divide by card size
3474 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3475 "Only one we handle so far.");
3476 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3478 // Combine card table base and card offset
3479 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3481 // Get the alias_index for raw card-mark memory
3482 int adr_type = Compile::AliasIdxRaw;
3483 Node* zero = __ ConI(0); // Dirty card value
3484 BasicType bt = T_BYTE;
3486 if (UseCondCardMark) {
3487 // The classic GC reference write barrier is typically implemented
3488 // as a store into the global card mark table. Unfortunately
3489 // unconditional stores can result in false sharing and excessive
3490 // coherence traffic as well as false transactional aborts.
3491 // UseCondCardMark enables MP "polite" conditional card mark
3492 // stores. In theory we could relax the load from ctrl() to
3493 // no_ctrl, but that doesn't buy much latitude.
3494 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3495 __ if_then(card_val, BoolTest::ne, zero);
3496 }
3498 // Smash zero into card
3499 if( !UseConcMarkSweepGC ) {
3500 __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3501 } else {
3502 // Specialized path for CM store barrier
3503 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3504 }
3506 if (UseCondCardMark) {
3507 __ end_if();
3508 }
3510 // Final sync IdealKit and GraphKit.
3511 final_sync(ideal);
3512 }
3514 // G1 pre/post barriers
3515 void GraphKit::g1_write_barrier_pre(bool do_load,
3516 Node* obj,
3517 Node* adr,
3518 uint alias_idx,
3519 Node* val,
3520 const TypeOopPtr* val_type,
3521 Node* pre_val,
3522 BasicType bt) {
3524 // Some sanity checks
3525 // Note: val is unused in this routine.
3527 if (do_load) {
3528 // We need to generate the load of the previous value
3529 assert(obj != NULL, "must have a base");
3530 assert(adr != NULL, "where are loading from?");
3531 assert(pre_val == NULL, "loaded already?");
3532 assert(val_type != NULL, "need a type");
3533 } else {
3534 // In this case both val_type and alias_idx are unused.
3535 assert(pre_val != NULL, "must be loaded already");
3536 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3537 }
3538 assert(bt == T_OBJECT, "or we shouldn't be here");
3540 IdealKit ideal(this, true);
3542 Node* tls = __ thread(); // ThreadLocalStorage
3544 Node* no_ctrl = NULL;
3545 Node* no_base = __ top();
3546 Node* zero = __ ConI(0);
3548 float likely = PROB_LIKELY(0.999);
3549 float unlikely = PROB_UNLIKELY(0.999);
3551 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3552 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3554 // Offsets into the thread
3555 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3556 PtrQueue::byte_offset_of_active());
3557 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3558 PtrQueue::byte_offset_of_index());
3559 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3560 PtrQueue::byte_offset_of_buf());
3562 // Now the actual pointers into the thread
3563 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3564 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3565 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3567 // Now some of the values
3568 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3570 // if (!marking)
3571 __ if_then(marking, BoolTest::ne, zero); {
3572 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3574 if (do_load) {
3575 // load original value
3576 // alias_idx correct??
3577 pre_val = __ load(no_ctrl, adr, val_type, bt, alias_idx);
3578 }
3580 // if (pre_val != NULL)
3581 __ if_then(pre_val, BoolTest::ne, null()); {
3582 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3584 // is the queue for this thread full?
3585 __ if_then(index, BoolTest::ne, zero, likely); {
3587 // decrement the index
3588 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3589 Node* next_indexX = next_index;
3590 #ifdef _LP64
3591 // We could refine the type for what it's worth
3592 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3593 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3594 #endif
3596 // Now get the buffer location we will log the previous value into and store it
3597 Node *log_addr = __ AddP(no_base, buffer, next_indexX);
3598 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw);
3599 // update the index
3600 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3602 } __ else_(); {
3604 // logging buffer is full, call the runtime
3605 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3606 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3607 } __ end_if(); // (!index)
3608 } __ end_if(); // (pre_val != NULL)
3609 } __ end_if(); // (!marking)
3611 // Final sync IdealKit and GraphKit.
3612 final_sync(ideal);
3613 }
3615 //
3616 // Update the card table and add card address to the queue
3617 //
3618 void GraphKit::g1_mark_card(IdealKit& ideal,
3619 Node* card_adr,
3620 Node* oop_store,
3621 uint oop_alias_idx,
3622 Node* index,
3623 Node* index_adr,
3624 Node* buffer,
3625 const TypeFunc* tf) {
3627 Node* zero = __ ConI(0);
3628 Node* no_base = __ top();
3629 BasicType card_bt = T_BYTE;
3630 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3631 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3633 // Now do the queue work
3634 __ if_then(index, BoolTest::ne, zero); {
3636 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3637 Node* next_indexX = next_index;
3638 #ifdef _LP64
3639 // We could refine the type for what it's worth
3640 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3641 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3642 #endif // _LP64
3643 Node* log_addr = __ AddP(no_base, buffer, next_indexX);
3645 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3646 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3648 } __ else_(); {
3649 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3650 } __ end_if();
3652 }
3654 void GraphKit::g1_write_barrier_post(Node* oop_store,
3655 Node* obj,
3656 Node* adr,
3657 uint alias_idx,
3658 Node* val,
3659 BasicType bt,
3660 bool use_precise) {
3661 // If we are writing a NULL then we need no post barrier
3663 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3664 // Must be NULL
3665 const Type* t = val->bottom_type();
3666 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3667 // No post barrier if writing NULLx
3668 return;
3669 }
3671 if (!use_precise) {
3672 // All card marks for a (non-array) instance are in one place:
3673 adr = obj;
3674 }
3675 // (Else it's an array (or unknown), and we want more precise card marks.)
3676 assert(adr != NULL, "");
3678 IdealKit ideal(this, true);
3680 Node* tls = __ thread(); // ThreadLocalStorage
3682 Node* no_base = __ top();
3683 float likely = PROB_LIKELY(0.999);
3684 float unlikely = PROB_UNLIKELY(0.999);
3685 Node* zero = __ ConI(0);
3686 Node* zeroX = __ ConX(0);
3688 // Get the alias_index for raw card-mark memory
3689 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3691 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3693 // Offsets into the thread
3694 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3695 PtrQueue::byte_offset_of_index());
3696 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3697 PtrQueue::byte_offset_of_buf());
3699 // Pointers into the thread
3701 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3702 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3704 // Now some values
3705 // Use ctrl to avoid hoisting these values past a safepoint, which could
3706 // potentially reset these fields in the JavaThread.
3707 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3708 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3710 // Convert the store obj pointer to an int prior to doing math on it
3711 // Must use ctrl to prevent "integerized oop" existing across safepoint
3712 Node* cast = __ CastPX(__ ctrl(), adr);
3714 // Divide pointer by card size
3715 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3717 // Combine card table base and card offset
3718 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3720 // If we know the value being stored does it cross regions?
3722 if (val != NULL) {
3723 // Does the store cause us to cross regions?
3725 // Should be able to do an unsigned compare of region_size instead of
3726 // and extra shift. Do we have an unsigned compare??
3727 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3728 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3730 // if (xor_res == 0) same region so skip
3731 __ if_then(xor_res, BoolTest::ne, zeroX); {
3733 // No barrier if we are storing a NULL
3734 __ if_then(val, BoolTest::ne, null(), unlikely); {
3736 // Ok must mark the card if not already dirty
3738 // load the original value of the card
3739 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3741 __ if_then(card_val, BoolTest::ne, zero); {
3742 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3743 } __ end_if();
3744 } __ end_if();
3745 } __ end_if();
3746 } else {
3747 // Object.clone() instrinsic uses this path.
3748 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3749 }
3751 // Final sync IdealKit and GraphKit.
3752 final_sync(ideal);
3753 }
3754 #undef __
3758 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
3759 if (java_lang_String::has_offset_field()) {
3760 int offset_offset = java_lang_String::offset_offset_in_bytes();
3761 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3762 false, NULL, 0);
3763 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
3764 int offset_field_idx = C->get_alias_index(offset_field_type);
3765 return make_load(ctrl,
3766 basic_plus_adr(str, str, offset_offset),
3767 TypeInt::INT, T_INT, offset_field_idx);
3768 } else {
3769 return intcon(0);
3770 }
3771 }
3773 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
3774 if (java_lang_String::has_count_field()) {
3775 int count_offset = java_lang_String::count_offset_in_bytes();
3776 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3777 false, NULL, 0);
3778 const TypePtr* count_field_type = string_type->add_offset(count_offset);
3779 int count_field_idx = C->get_alias_index(count_field_type);
3780 return make_load(ctrl,
3781 basic_plus_adr(str, str, count_offset),
3782 TypeInt::INT, T_INT, count_field_idx);
3783 } else {
3784 return load_array_length(load_String_value(ctrl, str));
3785 }
3786 }
3788 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
3789 int value_offset = java_lang_String::value_offset_in_bytes();
3790 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3791 false, NULL, 0);
3792 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3793 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3794 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
3795 ciTypeArrayKlass::make(T_CHAR), true, 0);
3796 int value_field_idx = C->get_alias_index(value_field_type);
3797 return make_load(ctrl, basic_plus_adr(str, str, value_offset),
3798 value_type, T_OBJECT, value_field_idx);
3799 }
3801 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
3802 int offset_offset = java_lang_String::offset_offset_in_bytes();
3803 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3804 false, NULL, 0);
3805 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
3806 int offset_field_idx = C->get_alias_index(offset_field_type);
3807 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
3808 value, T_INT, offset_field_idx);
3809 }
3811 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
3812 int value_offset = java_lang_String::value_offset_in_bytes();
3813 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3814 false, NULL, 0);
3815 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3816 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3817 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
3818 ciTypeArrayKlass::make(T_CHAR), true, 0);
3819 int value_field_idx = C->get_alias_index(value_field_type);
3820 store_to_memory(ctrl, basic_plus_adr(str, value_offset),
3821 value, T_OBJECT, value_field_idx);
3822 }
3824 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
3825 int count_offset = java_lang_String::count_offset_in_bytes();
3826 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3827 false, NULL, 0);
3828 const TypePtr* count_field_type = string_type->add_offset(count_offset);
3829 int count_field_idx = C->get_alias_index(count_field_type);
3830 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
3831 value, T_INT, count_field_idx);
3832 }