1 /*
2 * Copyright (c) 2001, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "compiler/compileLog.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/c2/barrierSetC2.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/idealKit.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/opaquenode.hpp"
42 #include "opto/parse.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/subtypenode.hpp"
46 #include "runtime/deoptimization.hpp"
47 #include "runtime/sharedRuntime.hpp"
48 #include "utilities/bitMap.inline.hpp"
49 #include "utilities/powerOfTwo.hpp"
50
51 //----------------------------GraphKit-----------------------------------------
52 // Main utility constructor.
53 GraphKit::GraphKit(JVMState* jvms)
54 : Phase(Phase::Parser),
55 _env(C->env()),
56 _gvn(*C->initial_gvn()),
57 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
58 {
59 _exceptions = jvms->map()->next_exception();
60 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
61 set_jvms(jvms);
62 }
63
64 // Private constructor for parser.
65 GraphKit::GraphKit()
66 : Phase(Phase::Parser),
67 _env(C->env()),
68 _gvn(*C->initial_gvn()),
69 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
70 {
71 _exceptions = NULL;
72 set_map(NULL);
73 debug_only(_sp = -99);
74 debug_only(set_bci(-99));
75 }
76
77
78
79 //---------------------------clean_stack---------------------------------------
80 // Clear away rubbish from the stack area of the JVM state.
81 // This destroys any arguments that may be waiting on the stack.
82 void GraphKit::clean_stack(int from_sp) {
83 SafePointNode* map = this->map();
84 JVMState* jvms = this->jvms();
85 int stk_size = jvms->stk_size();
86 int stkoff = jvms->stkoff();
87 Node* top = this->top();
88 for (int i = from_sp; i < stk_size; i++) {
89 if (map->in(stkoff + i) != top) {
90 map->set_req(stkoff + i, top);
91 }
92 }
93 }
94
95
96 //--------------------------------sync_jvms-----------------------------------
97 // Make sure our current jvms agrees with our parse state.
98 JVMState* GraphKit::sync_jvms() const {
99 JVMState* jvms = this->jvms();
100 jvms->set_bci(bci()); // Record the new bci in the JVMState
101 jvms->set_sp(sp()); // Record the new sp in the JVMState
102 assert(jvms_in_sync(), "jvms is now in sync");
103 return jvms;
104 }
105
106 //--------------------------------sync_jvms_for_reexecute---------------------
107 // Make sure our current jvms agrees with our parse state. This version
108 // uses the reexecute_sp for reexecuting bytecodes.
109 JVMState* GraphKit::sync_jvms_for_reexecute() {
110 JVMState* jvms = this->jvms();
111 jvms->set_bci(bci()); // Record the new bci in the JVMState
112 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
113 return jvms;
114 }
115
116 #ifdef ASSERT
117 bool GraphKit::jvms_in_sync() const {
118 Parse* parse = is_Parse();
119 if (parse == NULL) {
120 if (bci() != jvms()->bci()) return false;
121 if (sp() != (int)jvms()->sp()) return false;
122 return true;
123 }
124 if (jvms()->method() != parse->method()) return false;
125 if (jvms()->bci() != parse->bci()) return false;
126 int jvms_sp = jvms()->sp();
127 if (jvms_sp != parse->sp()) return false;
128 int jvms_depth = jvms()->depth();
129 if (jvms_depth != parse->depth()) return false;
130 return true;
131 }
132
133 // Local helper checks for special internal merge points
134 // used to accumulate and merge exception states.
135 // They are marked by the region's in(0) edge being the map itself.
136 // Such merge points must never "escape" into the parser at large,
137 // until they have been handed to gvn.transform.
138 static bool is_hidden_merge(Node* reg) {
139 if (reg == NULL) return false;
140 if (reg->is_Phi()) {
141 reg = reg->in(0);
142 if (reg == NULL) return false;
143 }
144 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
145 }
146
147 void GraphKit::verify_map() const {
148 if (map() == NULL) return; // null map is OK
149 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
150 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
151 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
152 }
153
154 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
155 assert(ex_map->next_exception() == NULL, "not already part of a chain");
156 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
157 }
158 #endif
159
160 //---------------------------stop_and_kill_map---------------------------------
161 // Set _map to NULL, signalling a stop to further bytecode execution.
162 // First smash the current map's control to a constant, to mark it dead.
163 void GraphKit::stop_and_kill_map() {
164 SafePointNode* dead_map = stop();
165 if (dead_map != NULL) {
166 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
167 assert(dead_map->is_killed(), "must be so marked");
168 }
169 }
170
171
172 //--------------------------------stopped--------------------------------------
173 // Tell if _map is NULL, or control is top.
174 bool GraphKit::stopped() {
175 if (map() == NULL) return true;
176 else if (control() == top()) return true;
177 else return false;
178 }
179
180
181 //-----------------------------has_ex_handler----------------------------------
182 // Tell if this method or any caller method has exception handlers.
183 bool GraphKit::has_ex_handler() {
184 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
185 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
186 return true;
187 }
188 }
189 return false;
190 }
191
192 //------------------------------save_ex_oop------------------------------------
193 // Save an exception without blowing stack contents or other JVM state.
194 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
195 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
196 ex_map->add_req(ex_oop);
197 debug_only(verify_exception_state(ex_map));
198 }
199
200 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
201 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
202 Node* ex_oop = ex_map->in(ex_map->req()-1);
203 if (clear_it) ex_map->del_req(ex_map->req()-1);
204 return ex_oop;
205 }
206
207 //-----------------------------saved_ex_oop------------------------------------
208 // Recover a saved exception from its map.
209 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
210 return common_saved_ex_oop(ex_map, false);
211 }
212
213 //--------------------------clear_saved_ex_oop---------------------------------
214 // Erase a previously saved exception from its map.
215 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
216 return common_saved_ex_oop(ex_map, true);
217 }
218
219 #ifdef ASSERT
220 //---------------------------has_saved_ex_oop----------------------------------
221 // Erase a previously saved exception from its map.
222 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
223 return ex_map->req() == ex_map->jvms()->endoff()+1;
224 }
225 #endif
226
227 //-------------------------make_exception_state--------------------------------
228 // Turn the current JVM state into an exception state, appending the ex_oop.
229 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
230 sync_jvms();
231 SafePointNode* ex_map = stop(); // do not manipulate this map any more
232 set_saved_ex_oop(ex_map, ex_oop);
233 return ex_map;
234 }
235
236
237 //--------------------------add_exception_state--------------------------------
238 // Add an exception to my list of exceptions.
239 void GraphKit::add_exception_state(SafePointNode* ex_map) {
240 if (ex_map == NULL || ex_map->control() == top()) {
241 return;
242 }
243 #ifdef ASSERT
244 verify_exception_state(ex_map);
245 if (has_exceptions()) {
246 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
247 }
248 #endif
249
250 // If there is already an exception of exactly this type, merge with it.
251 // In particular, null-checks and other low-level exceptions common up here.
252 Node* ex_oop = saved_ex_oop(ex_map);
253 const Type* ex_type = _gvn.type(ex_oop);
254 if (ex_oop == top()) {
255 // No action needed.
256 return;
257 }
258 assert(ex_type->isa_instptr(), "exception must be an instance");
259 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
260 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
261 // We check sp also because call bytecodes can generate exceptions
262 // both before and after arguments are popped!
263 if (ex_type2 == ex_type
264 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
265 combine_exception_states(ex_map, e2);
266 return;
267 }
268 }
269
270 // No pre-existing exception of the same type. Chain it on the list.
271 push_exception_state(ex_map);
272 }
273
274 //-----------------------add_exception_states_from-----------------------------
275 void GraphKit::add_exception_states_from(JVMState* jvms) {
276 SafePointNode* ex_map = jvms->map()->next_exception();
277 if (ex_map != NULL) {
278 jvms->map()->set_next_exception(NULL);
279 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
280 next_map = ex_map->next_exception();
281 ex_map->set_next_exception(NULL);
282 add_exception_state(ex_map);
283 }
284 }
285 }
286
287 //-----------------------transfer_exceptions_into_jvms-------------------------
288 JVMState* GraphKit::transfer_exceptions_into_jvms() {
289 if (map() == NULL) {
290 // We need a JVMS to carry the exceptions, but the map has gone away.
291 // Create a scratch JVMS, cloned from any of the exception states...
292 if (has_exceptions()) {
293 _map = _exceptions;
294 _map = clone_map();
295 _map->set_next_exception(NULL);
296 clear_saved_ex_oop(_map);
297 debug_only(verify_map());
298 } else {
299 // ...or created from scratch
300 JVMState* jvms = new (C) JVMState(_method, NULL);
301 jvms->set_bci(_bci);
302 jvms->set_sp(_sp);
303 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
304 set_jvms(jvms);
305 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
306 set_all_memory(top());
307 while (map()->req() < jvms->endoff()) map()->add_req(top());
308 }
309 // (This is a kludge, in case you didn't notice.)
310 set_control(top());
311 }
312 JVMState* jvms = sync_jvms();
313 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
314 jvms->map()->set_next_exception(_exceptions);
315 _exceptions = NULL; // done with this set of exceptions
316 return jvms;
317 }
318
319 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
320 assert(is_hidden_merge(dstphi), "must be a special merge node");
321 assert(is_hidden_merge(srcphi), "must be a special merge node");
322 uint limit = srcphi->req();
323 for (uint i = PhiNode::Input; i < limit; i++) {
324 dstphi->add_req(srcphi->in(i));
325 }
326 }
327 static inline void add_one_req(Node* dstphi, Node* src) {
328 assert(is_hidden_merge(dstphi), "must be a special merge node");
329 assert(!is_hidden_merge(src), "must not be a special merge node");
330 dstphi->add_req(src);
331 }
332
333 //-----------------------combine_exception_states------------------------------
334 // This helper function combines exception states by building phis on a
335 // specially marked state-merging region. These regions and phis are
336 // untransformed, and can build up gradually. The region is marked by
337 // having a control input of its exception map, rather than NULL. Such
338 // regions do not appear except in this function, and in use_exception_state.
339 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
340 if (failing()) return; // dying anyway...
341 JVMState* ex_jvms = ex_map->_jvms;
342 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
343 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
344 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
345 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
346 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
347 assert(ex_map->req() == phi_map->req(), "matching maps");
348 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
349 Node* hidden_merge_mark = root();
350 Node* region = phi_map->control();
351 MergeMemNode* phi_mem = phi_map->merged_memory();
352 MergeMemNode* ex_mem = ex_map->merged_memory();
353 if (region->in(0) != hidden_merge_mark) {
354 // The control input is not (yet) a specially-marked region in phi_map.
355 // Make it so, and build some phis.
356 region = new RegionNode(2);
357 _gvn.set_type(region, Type::CONTROL);
358 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
359 region->init_req(1, phi_map->control());
360 phi_map->set_control(region);
361 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
362 record_for_igvn(io_phi);
363 _gvn.set_type(io_phi, Type::ABIO);
364 phi_map->set_i_o(io_phi);
365 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
366 Node* m = mms.memory();
367 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
368 record_for_igvn(m_phi);
369 _gvn.set_type(m_phi, Type::MEMORY);
370 mms.set_memory(m_phi);
371 }
372 }
373
374 // Either or both of phi_map and ex_map might already be converted into phis.
375 Node* ex_control = ex_map->control();
376 // if there is special marking on ex_map also, we add multiple edges from src
377 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
378 // how wide was the destination phi_map, originally?
379 uint orig_width = region->req();
380
381 if (add_multiple) {
382 add_n_reqs(region, ex_control);
383 add_n_reqs(phi_map->i_o(), ex_map->i_o());
384 } else {
385 // ex_map has no merges, so we just add single edges everywhere
386 add_one_req(region, ex_control);
387 add_one_req(phi_map->i_o(), ex_map->i_o());
388 }
389 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
390 if (mms.is_empty()) {
391 // get a copy of the base memory, and patch some inputs into it
392 const TypePtr* adr_type = mms.adr_type(C);
393 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
394 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
395 mms.set_memory(phi);
396 // Prepare to append interesting stuff onto the newly sliced phi:
397 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
398 }
399 // Append stuff from ex_map:
400 if (add_multiple) {
401 add_n_reqs(mms.memory(), mms.memory2());
402 } else {
403 add_one_req(mms.memory(), mms.memory2());
404 }
405 }
406 uint limit = ex_map->req();
407 for (uint i = TypeFunc::Parms; i < limit; i++) {
408 // Skip everything in the JVMS after tos. (The ex_oop follows.)
409 if (i == tos) i = ex_jvms->monoff();
410 Node* src = ex_map->in(i);
411 Node* dst = phi_map->in(i);
412 if (src != dst) {
413 PhiNode* phi;
414 if (dst->in(0) != region) {
415 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
416 record_for_igvn(phi);
417 _gvn.set_type(phi, phi->type());
418 phi_map->set_req(i, dst);
419 // Prepare to append interesting stuff onto the new phi:
420 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
421 } else {
422 assert(dst->is_Phi(), "nobody else uses a hidden region");
423 phi = dst->as_Phi();
424 }
425 if (add_multiple && src->in(0) == ex_control) {
426 // Both are phis.
427 add_n_reqs(dst, src);
428 } else {
429 while (dst->req() < region->req()) add_one_req(dst, src);
430 }
431 const Type* srctype = _gvn.type(src);
432 if (phi->type() != srctype) {
433 const Type* dsttype = phi->type()->meet_speculative(srctype);
434 if (phi->type() != dsttype) {
435 phi->set_type(dsttype);
436 _gvn.set_type(phi, dsttype);
437 }
438 }
439 }
440 }
441 phi_map->merge_replaced_nodes_with(ex_map);
442 }
443
444 //--------------------------use_exception_state--------------------------------
445 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
446 if (failing()) { stop(); return top(); }
447 Node* region = phi_map->control();
448 Node* hidden_merge_mark = root();
449 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
450 Node* ex_oop = clear_saved_ex_oop(phi_map);
451 if (region->in(0) == hidden_merge_mark) {
452 // Special marking for internal ex-states. Process the phis now.
453 region->set_req(0, region); // now it's an ordinary region
454 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
455 // Note: Setting the jvms also sets the bci and sp.
456 set_control(_gvn.transform(region));
457 uint tos = jvms()->stkoff() + sp();
458 for (uint i = 1; i < tos; i++) {
459 Node* x = phi_map->in(i);
460 if (x->in(0) == region) {
461 assert(x->is_Phi(), "expected a special phi");
462 phi_map->set_req(i, _gvn.transform(x));
463 }
464 }
465 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
466 Node* x = mms.memory();
467 if (x->in(0) == region) {
468 assert(x->is_Phi(), "nobody else uses a hidden region");
469 mms.set_memory(_gvn.transform(x));
470 }
471 }
472 if (ex_oop->in(0) == region) {
473 assert(ex_oop->is_Phi(), "expected a special phi");
474 ex_oop = _gvn.transform(ex_oop);
475 }
476 } else {
477 set_jvms(phi_map->jvms());
478 }
479
480 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
481 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
482 return ex_oop;
483 }
484
485 //---------------------------------java_bc-------------------------------------
486 Bytecodes::Code GraphKit::java_bc() const {
487 ciMethod* method = this->method();
488 int bci = this->bci();
489 if (method != NULL && bci != InvocationEntryBci)
490 return method->java_code_at_bci(bci);
491 else
492 return Bytecodes::_illegal;
493 }
494
495 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
496 bool must_throw) {
497 // if the exception capability is set, then we will generate code
498 // to check the JavaThread.should_post_on_exceptions flag to see
499 // if we actually need to report exception events (for this
500 // thread). If we don't need to report exception events, we will
501 // take the normal fast path provided by add_exception_events. If
502 // exception event reporting is enabled for this thread, we will
503 // take the uncommon_trap in the BuildCutout below.
504
505 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
506 Node* jthread = _gvn.transform(new ThreadLocalNode());
507 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
508 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
509
510 // Test the should_post_on_exceptions_flag vs. 0
511 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
512 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
513
514 // Branch to slow_path if should_post_on_exceptions_flag was true
515 { BuildCutout unless(this, tst, PROB_MAX);
516 // Do not try anything fancy if we're notifying the VM on every throw.
517 // Cf. case Bytecodes::_athrow in parse2.cpp.
518 uncommon_trap(reason, Deoptimization::Action_none,
519 (ciKlass*)NULL, (char*)NULL, must_throw);
520 }
521
522 }
523
524 //------------------------------builtin_throw----------------------------------
525 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
526 bool must_throw = true;
527
528 if (env()->jvmti_can_post_on_exceptions()) {
529 // check if we must post exception events, take uncommon trap if so
530 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
531 // here if should_post_on_exceptions is false
532 // continue on with the normal codegen
533 }
534
535 // If this particular condition has not yet happened at this
536 // bytecode, then use the uncommon trap mechanism, and allow for
537 // a future recompilation if several traps occur here.
538 // If the throw is hot, try to use a more complicated inline mechanism
539 // which keeps execution inside the compiled code.
540 bool treat_throw_as_hot = false;
541 ciMethodData* md = method()->method_data();
542
543 if (ProfileTraps) {
544 if (too_many_traps(reason)) {
545 treat_throw_as_hot = true;
546 }
547 // (If there is no MDO at all, assume it is early in
548 // execution, and that any deopts are part of the
549 // startup transient, and don't need to be remembered.)
550
551 // Also, if there is a local exception handler, treat all throws
552 // as hot if there has been at least one in this method.
553 if (C->trap_count(reason) != 0
554 && method()->method_data()->trap_count(reason) != 0
555 && has_ex_handler()) {
556 treat_throw_as_hot = true;
557 }
558 }
559
560 // If this throw happens frequently, an uncommon trap might cause
561 // a performance pothole. If there is a local exception handler,
562 // and if this particular bytecode appears to be deoptimizing often,
563 // let us handle the throw inline, with a preconstructed instance.
564 // Note: If the deopt count has blown up, the uncommon trap
565 // runtime is going to flush this nmethod, not matter what.
566 if (treat_throw_as_hot
567 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
568 // If the throw is local, we use a pre-existing instance and
569 // punt on the backtrace. This would lead to a missing backtrace
570 // (a repeat of 4292742) if the backtrace object is ever asked
571 // for its backtrace.
572 // Fixing this remaining case of 4292742 requires some flavor of
573 // escape analysis. Leave that for the future.
574 ciInstance* ex_obj = NULL;
575 switch (reason) {
576 case Deoptimization::Reason_null_check:
577 ex_obj = env()->NullPointerException_instance();
578 break;
579 case Deoptimization::Reason_div0_check:
580 ex_obj = env()->ArithmeticException_instance();
581 break;
582 case Deoptimization::Reason_range_check:
583 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
584 break;
585 case Deoptimization::Reason_class_check:
586 if (java_bc() == Bytecodes::_aastore) {
587 ex_obj = env()->ArrayStoreException_instance();
588 } else {
589 ex_obj = env()->ClassCastException_instance();
590 }
591 break;
592 default:
593 break;
594 }
595 if (failing()) { stop(); return; } // exception allocation might fail
596 if (ex_obj != NULL) {
597 // Cheat with a preallocated exception object.
598 if (C->log() != NULL)
599 C->log()->elem("hot_throw preallocated='1' reason='%s'",
600 Deoptimization::trap_reason_name(reason));
601 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
602 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
603
604 // Clear the detail message of the preallocated exception object.
605 // Weblogic sometimes mutates the detail message of exceptions
606 // using reflection.
607 int offset = java_lang_Throwable::get_detailMessage_offset();
608 const TypePtr* adr_typ = ex_con->add_offset(offset);
609
610 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
611 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
612 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
613
614 add_exception_state(make_exception_state(ex_node));
615 return;
616 }
617 }
618
619 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
620 // It won't be much cheaper than bailing to the interp., since we'll
621 // have to pass up all the debug-info, and the runtime will have to
622 // create the stack trace.
623
624 // Usual case: Bail to interpreter.
625 // Reserve the right to recompile if we haven't seen anything yet.
626
627 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
628 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
629 if (treat_throw_as_hot
630 && (method()->method_data()->trap_recompiled_at(bci(), m)
631 || C->too_many_traps(reason))) {
632 // We cannot afford to take more traps here. Suffer in the interpreter.
633 if (C->log() != NULL)
634 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
635 Deoptimization::trap_reason_name(reason),
636 C->trap_count(reason));
637 action = Deoptimization::Action_none;
638 }
639
640 // "must_throw" prunes the JVM state to include only the stack, if there
641 // are no local exception handlers. This should cut down on register
642 // allocation time and code size, by drastically reducing the number
643 // of in-edges on the call to the uncommon trap.
644
645 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
646 }
647
648
649 //----------------------------PreserveJVMState---------------------------------
650 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
651 debug_only(kit->verify_map());
652 _kit = kit;
653 _map = kit->map(); // preserve the map
654 _sp = kit->sp();
655 kit->set_map(clone_map ? kit->clone_map() : NULL);
656 #ifdef ASSERT
657 _bci = kit->bci();
658 Parse* parser = kit->is_Parse();
659 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
660 _block = block;
661 #endif
662 }
663 PreserveJVMState::~PreserveJVMState() {
664 GraphKit* kit = _kit;
665 #ifdef ASSERT
666 assert(kit->bci() == _bci, "bci must not shift");
667 Parse* parser = kit->is_Parse();
668 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
669 assert(block == _block, "block must not shift");
670 #endif
671 kit->set_map(_map);
672 kit->set_sp(_sp);
673 }
674
675
676 //-----------------------------BuildCutout-------------------------------------
677 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
678 : PreserveJVMState(kit)
679 {
680 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
681 SafePointNode* outer_map = _map; // preserved map is caller's
682 SafePointNode* inner_map = kit->map();
683 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
684 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
685 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
686 }
687 BuildCutout::~BuildCutout() {
688 GraphKit* kit = _kit;
689 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
690 }
691
692 //---------------------------PreserveReexecuteState----------------------------
693 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
694 assert(!kit->stopped(), "must call stopped() before");
695 _kit = kit;
696 _sp = kit->sp();
697 _reexecute = kit->jvms()->_reexecute;
698 }
699 PreserveReexecuteState::~PreserveReexecuteState() {
700 if (_kit->stopped()) return;
701 _kit->jvms()->_reexecute = _reexecute;
702 _kit->set_sp(_sp);
703 }
704
705 //------------------------------clone_map--------------------------------------
706 // Implementation of PreserveJVMState
707 //
708 // Only clone_map(...) here. If this function is only used in the
709 // PreserveJVMState class we may want to get rid of this extra
710 // function eventually and do it all there.
711
712 SafePointNode* GraphKit::clone_map() {
713 if (map() == NULL) return NULL;
714
715 // Clone the memory edge first
716 Node* mem = MergeMemNode::make(map()->memory());
717 gvn().set_type_bottom(mem);
718
719 SafePointNode *clonemap = (SafePointNode*)map()->clone();
720 JVMState* jvms = this->jvms();
721 JVMState* clonejvms = jvms->clone_shallow(C);
722 clonemap->set_memory(mem);
723 clonemap->set_jvms(clonejvms);
724 clonejvms->set_map(clonemap);
725 record_for_igvn(clonemap);
726 gvn().set_type_bottom(clonemap);
727 return clonemap;
728 }
729
730
731 //-----------------------------set_map_clone-----------------------------------
732 void GraphKit::set_map_clone(SafePointNode* m) {
733 _map = m;
734 _map = clone_map();
735 _map->set_next_exception(NULL);
736 debug_only(verify_map());
737 }
738
739
740 //----------------------------kill_dead_locals---------------------------------
741 // Detect any locals which are known to be dead, and force them to top.
742 void GraphKit::kill_dead_locals() {
743 // Consult the liveness information for the locals. If any
744 // of them are unused, then they can be replaced by top(). This
745 // should help register allocation time and cut down on the size
746 // of the deoptimization information.
747
748 // This call is made from many of the bytecode handling
749 // subroutines called from the Big Switch in do_one_bytecode.
750 // Every bytecode which might include a slow path is responsible
751 // for killing its dead locals. The more consistent we
752 // are about killing deads, the fewer useless phis will be
753 // constructed for them at various merge points.
754
755 // bci can be -1 (InvocationEntryBci). We return the entry
756 // liveness for the method.
757
758 if (method() == NULL || method()->code_size() == 0) {
759 // We are building a graph for a call to a native method.
760 // All locals are live.
761 return;
762 }
763
764 ResourceMark rm;
765
766 // Consult the liveness information for the locals. If any
767 // of them are unused, then they can be replaced by top(). This
768 // should help register allocation time and cut down on the size
769 // of the deoptimization information.
770 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
771
772 int len = (int)live_locals.size();
773 assert(len <= jvms()->loc_size(), "too many live locals");
774 for (int local = 0; local < len; local++) {
775 if (!live_locals.at(local)) {
776 set_local(local, top());
777 }
778 }
779 }
780
781 #ifdef ASSERT
782 //-------------------------dead_locals_are_killed------------------------------
783 // Return true if all dead locals are set to top in the map.
784 // Used to assert "clean" debug info at various points.
785 bool GraphKit::dead_locals_are_killed() {
786 if (method() == NULL || method()->code_size() == 0) {
787 // No locals need to be dead, so all is as it should be.
788 return true;
789 }
790
791 // Make sure somebody called kill_dead_locals upstream.
792 ResourceMark rm;
793 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
794 if (jvms->loc_size() == 0) continue; // no locals to consult
795 SafePointNode* map = jvms->map();
796 ciMethod* method = jvms->method();
797 int bci = jvms->bci();
798 if (jvms == this->jvms()) {
799 bci = this->bci(); // it might not yet be synched
800 }
801 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
802 int len = (int)live_locals.size();
803 if (!live_locals.is_valid() || len == 0)
804 // This method is trivial, or is poisoned by a breakpoint.
805 return true;
806 assert(len == jvms->loc_size(), "live map consistent with locals map");
807 for (int local = 0; local < len; local++) {
808 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
809 if (PrintMiscellaneous && (Verbose || WizardMode)) {
810 tty->print_cr("Zombie local %d: ", local);
811 jvms->dump();
812 }
813 return false;
814 }
815 }
816 }
817 return true;
818 }
819
820 #endif //ASSERT
821
822 // Helper function for enforcing certain bytecodes to reexecute if
823 // deoptimization happens
824 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
825 ciMethod* cur_method = jvms->method();
826 int cur_bci = jvms->bci();
827 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
828 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
829 return Interpreter::bytecode_should_reexecute(code) ||
830 (is_anewarray && code == Bytecodes::_multianewarray);
831 // Reexecute _multianewarray bytecode which was replaced with
832 // sequence of [a]newarray. See Parse::do_multianewarray().
833 //
834 // Note: interpreter should not have it set since this optimization
835 // is limited by dimensions and guarded by flag so in some cases
836 // multianewarray() runtime calls will be generated and
837 // the bytecode should not be reexecutes (stack will not be reset).
838 } else
839 return false;
840 }
841
842 // Helper function for adding JVMState and debug information to node
843 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
844 // Add the safepoint edges to the call (or other safepoint).
845
846 // Make sure dead locals are set to top. This
847 // should help register allocation time and cut down on the size
848 // of the deoptimization information.
849 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
850
851 // Walk the inline list to fill in the correct set of JVMState's
852 // Also fill in the associated edges for each JVMState.
853
854 // If the bytecode needs to be reexecuted we need to put
855 // the arguments back on the stack.
856 const bool should_reexecute = jvms()->should_reexecute();
857 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
858
859 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
860 // undefined if the bci is different. This is normal for Parse but it
861 // should not happen for LibraryCallKit because only one bci is processed.
862 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
863 "in LibraryCallKit the reexecute bit should not change");
864
865 // If we are guaranteed to throw, we can prune everything but the
866 // input to the current bytecode.
867 bool can_prune_locals = false;
868 uint stack_slots_not_pruned = 0;
869 int inputs = 0, depth = 0;
870 if (must_throw) {
871 assert(method() == youngest_jvms->method(), "sanity");
872 if (compute_stack_effects(inputs, depth)) {
873 can_prune_locals = true;
874 stack_slots_not_pruned = inputs;
875 }
876 }
877
878 if (env()->should_retain_local_variables()) {
879 // At any safepoint, this method can get breakpointed, which would
880 // then require an immediate deoptimization.
881 can_prune_locals = false; // do not prune locals
882 stack_slots_not_pruned = 0;
883 }
884
885 // do not scribble on the input jvms
886 JVMState* out_jvms = youngest_jvms->clone_deep(C);
887 call->set_jvms(out_jvms); // Start jvms list for call node
888
889 // For a known set of bytecodes, the interpreter should reexecute them if
890 // deoptimization happens. We set the reexecute state for them here
891 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
892 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
893 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
894 }
895
896 // Presize the call:
897 DEBUG_ONLY(uint non_debug_edges = call->req());
898 call->add_req_batch(top(), youngest_jvms->debug_depth());
899 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
900
901 // Set up edges so that the call looks like this:
902 // Call [state:] ctl io mem fptr retadr
903 // [parms:] parm0 ... parmN
904 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
905 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
906 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
907 // Note that caller debug info precedes callee debug info.
908
909 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
910 uint debug_ptr = call->req();
911
912 // Loop over the map input edges associated with jvms, add them
913 // to the call node, & reset all offsets to match call node array.
914 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
915 uint debug_end = debug_ptr;
916 uint debug_start = debug_ptr - in_jvms->debug_size();
917 debug_ptr = debug_start; // back up the ptr
918
919 uint p = debug_start; // walks forward in [debug_start, debug_end)
920 uint j, k, l;
921 SafePointNode* in_map = in_jvms->map();
922 out_jvms->set_map(call);
923
924 if (can_prune_locals) {
925 assert(in_jvms->method() == out_jvms->method(), "sanity");
926 // If the current throw can reach an exception handler in this JVMS,
927 // then we must keep everything live that can reach that handler.
928 // As a quick and dirty approximation, we look for any handlers at all.
929 if (in_jvms->method()->has_exception_handlers()) {
930 can_prune_locals = false;
931 }
932 }
933
934 // Add the Locals
935 k = in_jvms->locoff();
936 l = in_jvms->loc_size();
937 out_jvms->set_locoff(p);
938 if (!can_prune_locals) {
939 for (j = 0; j < l; j++)
940 call->set_req(p++, in_map->in(k+j));
941 } else {
942 p += l; // already set to top above by add_req_batch
943 }
944
945 // Add the Expression Stack
946 k = in_jvms->stkoff();
947 l = in_jvms->sp();
948 out_jvms->set_stkoff(p);
949 if (!can_prune_locals) {
950 for (j = 0; j < l; j++)
951 call->set_req(p++, in_map->in(k+j));
952 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
953 // Divide stack into {S0,...,S1}, where S0 is set to top.
954 uint s1 = stack_slots_not_pruned;
955 stack_slots_not_pruned = 0; // for next iteration
956 if (s1 > l) s1 = l;
957 uint s0 = l - s1;
958 p += s0; // skip the tops preinstalled by add_req_batch
959 for (j = s0; j < l; j++)
960 call->set_req(p++, in_map->in(k+j));
961 } else {
962 p += l; // already set to top above by add_req_batch
963 }
964
965 // Add the Monitors
966 k = in_jvms->monoff();
967 l = in_jvms->mon_size();
968 out_jvms->set_monoff(p);
969 for (j = 0; j < l; j++)
970 call->set_req(p++, in_map->in(k+j));
971
972 // Copy any scalar object fields.
973 k = in_jvms->scloff();
974 l = in_jvms->scl_size();
975 out_jvms->set_scloff(p);
976 for (j = 0; j < l; j++)
977 call->set_req(p++, in_map->in(k+j));
978
979 // Finish the new jvms.
980 out_jvms->set_endoff(p);
981
982 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
983 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
984 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
985 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
986 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
987 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
988
989 // Update the two tail pointers in parallel.
990 out_jvms = out_jvms->caller();
991 in_jvms = in_jvms->caller();
992 }
993
994 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
995
996 // Test the correctness of JVMState::debug_xxx accessors:
997 assert(call->jvms()->debug_start() == non_debug_edges, "");
998 assert(call->jvms()->debug_end() == call->req(), "");
999 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1000 }
1001
1002 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1003 Bytecodes::Code code = java_bc();
1004 if (code == Bytecodes::_wide) {
1005 code = method()->java_code_at_bci(bci() + 1);
1006 }
1007
1008 BasicType rtype = T_ILLEGAL;
1009 int rsize = 0;
1010
1011 if (code != Bytecodes::_illegal) {
1012 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1013 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1014 if (rtype < T_CONFLICT)
1015 rsize = type2size[rtype];
1016 }
1017
1018 switch (code) {
1019 case Bytecodes::_illegal:
1020 return false;
1021
1022 case Bytecodes::_ldc:
1023 case Bytecodes::_ldc_w:
1024 case Bytecodes::_ldc2_w:
1025 inputs = 0;
1026 break;
1027
1028 case Bytecodes::_dup: inputs = 1; break;
1029 case Bytecodes::_dup_x1: inputs = 2; break;
1030 case Bytecodes::_dup_x2: inputs = 3; break;
1031 case Bytecodes::_dup2: inputs = 2; break;
1032 case Bytecodes::_dup2_x1: inputs = 3; break;
1033 case Bytecodes::_dup2_x2: inputs = 4; break;
1034 case Bytecodes::_swap: inputs = 2; break;
1035 case Bytecodes::_arraylength: inputs = 1; break;
1036
1037 case Bytecodes::_getstatic:
1038 case Bytecodes::_putstatic:
1039 case Bytecodes::_getfield:
1040 case Bytecodes::_putfield:
1041 {
1042 bool ignored_will_link;
1043 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1044 int size = field->type()->size();
1045 bool is_get = (depth >= 0), is_static = (depth & 1);
1046 inputs = (is_static ? 0 : 1);
1047 if (is_get) {
1048 depth = size - inputs;
1049 } else {
1050 inputs += size; // putxxx pops the value from the stack
1051 depth = - inputs;
1052 }
1053 }
1054 break;
1055
1056 case Bytecodes::_invokevirtual:
1057 case Bytecodes::_invokespecial:
1058 case Bytecodes::_invokestatic:
1059 case Bytecodes::_invokedynamic:
1060 case Bytecodes::_invokeinterface:
1061 {
1062 bool ignored_will_link;
1063 ciSignature* declared_signature = NULL;
1064 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1065 assert(declared_signature != NULL, "cannot be null");
1066 inputs = declared_signature->arg_size_for_bc(code);
1067 int size = declared_signature->return_type()->size();
1068 depth = size - inputs;
1069 }
1070 break;
1071
1072 case Bytecodes::_multianewarray:
1073 {
1074 ciBytecodeStream iter(method());
1075 iter.reset_to_bci(bci());
1076 iter.next();
1077 inputs = iter.get_dimensions();
1078 assert(rsize == 1, "");
1079 depth = rsize - inputs;
1080 }
1081 break;
1082
1083 case Bytecodes::_ireturn:
1084 case Bytecodes::_lreturn:
1085 case Bytecodes::_freturn:
1086 case Bytecodes::_dreturn:
1087 case Bytecodes::_areturn:
1088 assert(rsize == -depth, "");
1089 inputs = rsize;
1090 break;
1091
1092 case Bytecodes::_jsr:
1093 case Bytecodes::_jsr_w:
1094 inputs = 0;
1095 depth = 1; // S.B. depth=1, not zero
1096 break;
1097
1098 default:
1099 // bytecode produces a typed result
1100 inputs = rsize - depth;
1101 assert(inputs >= 0, "");
1102 break;
1103 }
1104
1105 #ifdef ASSERT
1106 // spot check
1107 int outputs = depth + inputs;
1108 assert(outputs >= 0, "sanity");
1109 switch (code) {
1110 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1111 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1112 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1113 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1114 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1115 default: break;
1116 }
1117 #endif //ASSERT
1118
1119 return true;
1120 }
1121
1122
1123
1124 //------------------------------basic_plus_adr---------------------------------
1125 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1126 // short-circuit a common case
1127 if (offset == intcon(0)) return ptr;
1128 return _gvn.transform( new AddPNode(base, ptr, offset) );
1129 }
1130
1131 Node* GraphKit::ConvI2L(Node* offset) {
1132 // short-circuit a common case
1133 jint offset_con = find_int_con(offset, Type::OffsetBot);
1134 if (offset_con != Type::OffsetBot) {
1135 return longcon((jlong) offset_con);
1136 }
1137 return _gvn.transform( new ConvI2LNode(offset));
1138 }
1139
1140 Node* GraphKit::ConvI2UL(Node* offset) {
1141 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1142 if (offset_con != (juint) Type::OffsetBot) {
1143 return longcon((julong) offset_con);
1144 }
1145 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1146 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1147 return _gvn.transform( new AndLNode(conv, mask) );
1148 }
1149
1150 Node* GraphKit::ConvL2I(Node* offset) {
1151 // short-circuit a common case
1152 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1153 if (offset_con != (jlong)Type::OffsetBot) {
1154 return intcon((int) offset_con);
1155 }
1156 return _gvn.transform( new ConvL2INode(offset));
1157 }
1158
1159 //-------------------------load_object_klass-----------------------------------
1160 Node* GraphKit::load_object_klass(Node* obj) {
1161 // Special-case a fresh allocation to avoid building nodes:
1162 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1163 if (akls != NULL) return akls;
1164 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1165 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1166 }
1167
1168 //-------------------------load_array_length-----------------------------------
1169 Node* GraphKit::load_array_length(Node* array) {
1170 // Special-case a fresh allocation to avoid building nodes:
1171 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1172 Node *alen;
1173 if (alloc == NULL) {
1174 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1175 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1176 } else {
1177 alen = alloc->Ideal_length();
1178 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1179 if (ccast != alen) {
1180 alen = _gvn.transform(ccast);
1181 }
1182 }
1183 return alen;
1184 }
1185
1186 //------------------------------do_null_check----------------------------------
1187 // Helper function to do a NULL pointer check. Returned value is
1188 // the incoming address with NULL casted away. You are allowed to use the
1189 // not-null value only if you are control dependent on the test.
1190 #ifndef PRODUCT
1191 extern int explicit_null_checks_inserted,
1192 explicit_null_checks_elided;
1193 #endif
1194 Node* GraphKit::null_check_common(Node* value, BasicType type,
1195 // optional arguments for variations:
1196 bool assert_null,
1197 Node* *null_control,
1198 bool speculative) {
1199 assert(!assert_null || null_control == NULL, "not both at once");
1200 if (stopped()) return top();
1201 NOT_PRODUCT(explicit_null_checks_inserted++);
1202
1203 // Construct NULL check
1204 Node *chk = NULL;
1205 switch(type) {
1206 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1207 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1208 case T_ARRAY : // fall through
1209 type = T_OBJECT; // simplify further tests
1210 case T_OBJECT : {
1211 const Type *t = _gvn.type( value );
1212
1213 const TypeOopPtr* tp = t->isa_oopptr();
1214 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1215 // Only for do_null_check, not any of its siblings:
1216 && !assert_null && null_control == NULL) {
1217 // Usually, any field access or invocation on an unloaded oop type
1218 // will simply fail to link, since the statically linked class is
1219 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1220 // the static class is loaded but the sharper oop type is not.
1221 // Rather than checking for this obscure case in lots of places,
1222 // we simply observe that a null check on an unloaded class
1223 // will always be followed by a nonsense operation, so we
1224 // can just issue the uncommon trap here.
1225 // Our access to the unloaded class will only be correct
1226 // after it has been loaded and initialized, which requires
1227 // a trip through the interpreter.
1228 #ifndef PRODUCT
1229 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1230 #endif
1231 uncommon_trap(Deoptimization::Reason_unloaded,
1232 Deoptimization::Action_reinterpret,
1233 tp->klass(), "!loaded");
1234 return top();
1235 }
1236
1237 if (assert_null) {
1238 // See if the type is contained in NULL_PTR.
1239 // If so, then the value is already null.
1240 if (t->higher_equal(TypePtr::NULL_PTR)) {
1241 NOT_PRODUCT(explicit_null_checks_elided++);
1242 return value; // Elided null assert quickly!
1243 }
1244 } else {
1245 // See if mixing in the NULL pointer changes type.
1246 // If so, then the NULL pointer was not allowed in the original
1247 // type. In other words, "value" was not-null.
1248 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1249 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1250 NOT_PRODUCT(explicit_null_checks_elided++);
1251 return value; // Elided null check quickly!
1252 }
1253 }
1254 chk = new CmpPNode( value, null() );
1255 break;
1256 }
1257
1258 default:
1259 fatal("unexpected type: %s", type2name(type));
1260 }
1261 assert(chk != NULL, "sanity check");
1262 chk = _gvn.transform(chk);
1263
1264 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1265 BoolNode *btst = new BoolNode( chk, btest);
1266 Node *tst = _gvn.transform( btst );
1267
1268 //-----------
1269 // if peephole optimizations occurred, a prior test existed.
1270 // If a prior test existed, maybe it dominates as we can avoid this test.
1271 if (tst != btst && type == T_OBJECT) {
1272 // At this point we want to scan up the CFG to see if we can
1273 // find an identical test (and so avoid this test altogether).
1274 Node *cfg = control();
1275 int depth = 0;
1276 while( depth < 16 ) { // Limit search depth for speed
1277 if( cfg->Opcode() == Op_IfTrue &&
1278 cfg->in(0)->in(1) == tst ) {
1279 // Found prior test. Use "cast_not_null" to construct an identical
1280 // CastPP (and hence hash to) as already exists for the prior test.
1281 // Return that casted value.
1282 if (assert_null) {
1283 replace_in_map(value, null());
1284 return null(); // do not issue the redundant test
1285 }
1286 Node *oldcontrol = control();
1287 set_control(cfg);
1288 Node *res = cast_not_null(value);
1289 set_control(oldcontrol);
1290 NOT_PRODUCT(explicit_null_checks_elided++);
1291 return res;
1292 }
1293 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1294 if (cfg == NULL) break; // Quit at region nodes
1295 depth++;
1296 }
1297 }
1298
1299 //-----------
1300 // Branch to failure if null
1301 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1302 Deoptimization::DeoptReason reason;
1303 if (assert_null) {
1304 reason = Deoptimization::reason_null_assert(speculative);
1305 } else if (type == T_OBJECT) {
1306 reason = Deoptimization::reason_null_check(speculative);
1307 } else {
1308 reason = Deoptimization::Reason_div0_check;
1309 }
1310 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1311 // ciMethodData::has_trap_at will return a conservative -1 if any
1312 // must-be-null assertion has failed. This could cause performance
1313 // problems for a method after its first do_null_assert failure.
1314 // Consider using 'Reason_class_check' instead?
1315
1316 // To cause an implicit null check, we set the not-null probability
1317 // to the maximum (PROB_MAX). For an explicit check the probability
1318 // is set to a smaller value.
1319 if (null_control != NULL || too_many_traps(reason)) {
1320 // probability is less likely
1321 ok_prob = PROB_LIKELY_MAG(3);
1322 } else if (!assert_null &&
1323 (ImplicitNullCheckThreshold > 0) &&
1324 method() != NULL &&
1325 (method()->method_data()->trap_count(reason)
1326 >= (uint)ImplicitNullCheckThreshold)) {
1327 ok_prob = PROB_LIKELY_MAG(3);
1328 }
1329
1330 if (null_control != NULL) {
1331 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1332 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1333 set_control( _gvn.transform( new IfTrueNode(iff)));
1334 #ifndef PRODUCT
1335 if (null_true == top()) {
1336 explicit_null_checks_elided++;
1337 }
1338 #endif
1339 (*null_control) = null_true;
1340 } else {
1341 BuildCutout unless(this, tst, ok_prob);
1342 // Check for optimizer eliding test at parse time
1343 if (stopped()) {
1344 // Failure not possible; do not bother making uncommon trap.
1345 NOT_PRODUCT(explicit_null_checks_elided++);
1346 } else if (assert_null) {
1347 uncommon_trap(reason,
1348 Deoptimization::Action_make_not_entrant,
1349 NULL, "assert_null");
1350 } else {
1351 replace_in_map(value, zerocon(type));
1352 builtin_throw(reason);
1353 }
1354 }
1355
1356 // Must throw exception, fall-thru not possible?
1357 if (stopped()) {
1358 return top(); // No result
1359 }
1360
1361 if (assert_null) {
1362 // Cast obj to null on this path.
1363 replace_in_map(value, zerocon(type));
1364 return zerocon(type);
1365 }
1366
1367 // Cast obj to not-null on this path, if there is no null_control.
1368 // (If there is a null_control, a non-null value may come back to haunt us.)
1369 if (type == T_OBJECT) {
1370 Node* cast = cast_not_null(value, false);
1371 if (null_control == NULL || (*null_control) == top())
1372 replace_in_map(value, cast);
1373 value = cast;
1374 }
1375
1376 return value;
1377 }
1378
1379
1380 //------------------------------cast_not_null----------------------------------
1381 // Cast obj to not-null on this path
1382 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1383 const Type *t = _gvn.type(obj);
1384 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1385 // Object is already not-null?
1386 if( t == t_not_null ) return obj;
1387
1388 Node *cast = new CastPPNode(obj,t_not_null);
1389 cast->init_req(0, control());
1390 cast = _gvn.transform( cast );
1391
1392 // Scan for instances of 'obj' in the current JVM mapping.
1393 // These instances are known to be not-null after the test.
1394 if (do_replace_in_map)
1395 replace_in_map(obj, cast);
1396
1397 return cast; // Return casted value
1398 }
1399
1400 // Sometimes in intrinsics, we implicitly know an object is not null
1401 // (there's no actual null check) so we can cast it to not null. In
1402 // the course of optimizations, the input to the cast can become null.
1403 // In that case that data path will die and we need the control path
1404 // to become dead as well to keep the graph consistent. So we have to
1405 // add a check for null for which one branch can't be taken. It uses
1406 // an Opaque4 node that will cause the check to be removed after loop
1407 // opts so the test goes away and the compiled code doesn't execute a
1408 // useless check.
1409 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1410 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1411 return value;
1412 }
1413 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1414 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1415 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1416 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1417 _gvn.set_type(iff, iff->Value(&_gvn));
1418 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1419 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1420 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1421 C->root()->add_req(halt);
1422 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1423 set_control(if_t);
1424 return cast_not_null(value, do_replace_in_map);
1425 }
1426
1427
1428 //--------------------------replace_in_map-------------------------------------
1429 void GraphKit::replace_in_map(Node* old, Node* neww) {
1430 if (old == neww) {
1431 return;
1432 }
1433
1434 map()->replace_edge(old, neww);
1435
1436 // Note: This operation potentially replaces any edge
1437 // on the map. This includes locals, stack, and monitors
1438 // of the current (innermost) JVM state.
1439
1440 // don't let inconsistent types from profiling escape this
1441 // method
1442
1443 const Type* told = _gvn.type(old);
1444 const Type* tnew = _gvn.type(neww);
1445
1446 if (!tnew->higher_equal(told)) {
1447 return;
1448 }
1449
1450 map()->record_replaced_node(old, neww);
1451 }
1452
1453
1454 //=============================================================================
1455 //--------------------------------memory---------------------------------------
1456 Node* GraphKit::memory(uint alias_idx) {
1457 MergeMemNode* mem = merged_memory();
1458 Node* p = mem->memory_at(alias_idx);
1459 _gvn.set_type(p, Type::MEMORY); // must be mapped
1460 return p;
1461 }
1462
1463 //-----------------------------reset_memory------------------------------------
1464 Node* GraphKit::reset_memory() {
1465 Node* mem = map()->memory();
1466 // do not use this node for any more parsing!
1467 debug_only( map()->set_memory((Node*)NULL) );
1468 return _gvn.transform( mem );
1469 }
1470
1471 //------------------------------set_all_memory---------------------------------
1472 void GraphKit::set_all_memory(Node* newmem) {
1473 Node* mergemem = MergeMemNode::make(newmem);
1474 gvn().set_type_bottom(mergemem);
1475 map()->set_memory(mergemem);
1476 }
1477
1478 //------------------------------set_all_memory_call----------------------------
1479 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1480 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1481 set_all_memory(newmem);
1482 }
1483
1484 //=============================================================================
1485 //
1486 // parser factory methods for MemNodes
1487 //
1488 // These are layered on top of the factory methods in LoadNode and StoreNode,
1489 // and integrate with the parser's memory state and _gvn engine.
1490 //
1491
1492 // factory methods in "int adr_idx"
1493 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1494 int adr_idx,
1495 MemNode::MemOrd mo,
1496 LoadNode::ControlDependency control_dependency,
1497 bool require_atomic_access,
1498 bool unaligned,
1499 bool mismatched,
1500 bool unsafe,
1501 uint8_t barrier_data) {
1502 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1503 const TypePtr* adr_type = NULL; // debug-mode-only argument
1504 debug_only(adr_type = C->get_adr_type(adr_idx));
1505 Node* mem = memory(adr_idx);
1506 Node* ld;
1507 if (require_atomic_access && bt == T_LONG) {
1508 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1509 } else if (require_atomic_access && bt == T_DOUBLE) {
1510 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1511 } else {
1512 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1513 }
1514 ld = _gvn.transform(ld);
1515 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1516 // Improve graph before escape analysis and boxing elimination.
1517 record_for_igvn(ld);
1518 }
1519 return ld;
1520 }
1521
1522 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1523 int adr_idx,
1524 MemNode::MemOrd mo,
1525 bool require_atomic_access,
1526 bool unaligned,
1527 bool mismatched,
1528 bool unsafe) {
1529 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1530 const TypePtr* adr_type = NULL;
1531 debug_only(adr_type = C->get_adr_type(adr_idx));
1532 Node *mem = memory(adr_idx);
1533 Node* st;
1534 if (require_atomic_access && bt == T_LONG) {
1535 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1536 } else if (require_atomic_access && bt == T_DOUBLE) {
1537 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1538 } else {
1539 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1540 }
1541 if (unaligned) {
1542 st->as_Store()->set_unaligned_access();
1543 }
1544 if (mismatched) {
1545 st->as_Store()->set_mismatched_access();
1546 }
1547 if (unsafe) {
1548 st->as_Store()->set_unsafe_access();
1549 }
1550 st = _gvn.transform(st);
1551 set_memory(st, adr_idx);
1552 // Back-to-back stores can only remove intermediate store with DU info
1553 // so push on worklist for optimizer.
1554 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1555 record_for_igvn(st);
1556
1557 return st;
1558 }
1559
1560 Node* GraphKit::access_store_at(Node* obj,
1561 Node* adr,
1562 const TypePtr* adr_type,
1563 Node* val,
1564 const Type* val_type,
1565 BasicType bt,
1566 DecoratorSet decorators) {
1567 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1568 // could be delayed during Parse (for example, in adjust_map_after_if()).
1569 // Execute transformation here to avoid barrier generation in such case.
1570 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1571 val = _gvn.makecon(TypePtr::NULL_PTR);
1572 }
1573
1574 if (stopped()) {
1575 return top(); // Dead path ?
1576 }
1577
1578 assert(val != NULL, "not dead path");
1579
1580 C2AccessValuePtr addr(adr, adr_type);
1581 C2AccessValue value(val, val_type);
1582 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1583 if (access.is_raw()) {
1584 return _barrier_set->BarrierSetC2::store_at(access, value);
1585 } else {
1586 return _barrier_set->store_at(access, value);
1587 }
1588 }
1589
1590 Node* GraphKit::access_load_at(Node* obj, // containing obj
1591 Node* adr, // actual adress to store val at
1592 const TypePtr* adr_type,
1593 const Type* val_type,
1594 BasicType bt,
1595 DecoratorSet decorators) {
1596 if (stopped()) {
1597 return top(); // Dead path ?
1598 }
1599
1600 C2AccessValuePtr addr(adr, adr_type);
1601 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1602 if (access.is_raw()) {
1603 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1604 } else {
1605 return _barrier_set->load_at(access, val_type);
1606 }
1607 }
1608
1609 Node* GraphKit::access_load(Node* adr, // actual adress to load val at
1610 const Type* val_type,
1611 BasicType bt,
1612 DecoratorSet decorators) {
1613 if (stopped()) {
1614 return top(); // Dead path ?
1615 }
1616
1617 C2AccessValuePtr addr(adr, NULL);
1618 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1619 if (access.is_raw()) {
1620 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1621 } else {
1622 return _barrier_set->load_at(access, val_type);
1623 }
1624 }
1625
1626 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1627 Node* adr,
1628 const TypePtr* adr_type,
1629 int alias_idx,
1630 Node* expected_val,
1631 Node* new_val,
1632 const Type* value_type,
1633 BasicType bt,
1634 DecoratorSet decorators) {
1635 C2AccessValuePtr addr(adr, adr_type);
1636 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1637 bt, obj, addr, alias_idx);
1638 if (access.is_raw()) {
1639 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1640 } else {
1641 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1642 }
1643 }
1644
1645 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1646 Node* adr,
1647 const TypePtr* adr_type,
1648 int alias_idx,
1649 Node* expected_val,
1650 Node* new_val,
1651 const Type* value_type,
1652 BasicType bt,
1653 DecoratorSet decorators) {
1654 C2AccessValuePtr addr(adr, adr_type);
1655 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1656 bt, obj, addr, alias_idx);
1657 if (access.is_raw()) {
1658 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1659 } else {
1660 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1661 }
1662 }
1663
1664 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1665 Node* adr,
1666 const TypePtr* adr_type,
1667 int alias_idx,
1668 Node* new_val,
1669 const Type* value_type,
1670 BasicType bt,
1671 DecoratorSet decorators) {
1672 C2AccessValuePtr addr(adr, adr_type);
1673 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1674 bt, obj, addr, alias_idx);
1675 if (access.is_raw()) {
1676 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1677 } else {
1678 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1679 }
1680 }
1681
1682 Node* GraphKit::access_atomic_add_at(Node* obj,
1683 Node* adr,
1684 const TypePtr* adr_type,
1685 int alias_idx,
1686 Node* new_val,
1687 const Type* value_type,
1688 BasicType bt,
1689 DecoratorSet decorators) {
1690 C2AccessValuePtr addr(adr, adr_type);
1691 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1692 if (access.is_raw()) {
1693 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1694 } else {
1695 return _barrier_set->atomic_add_at(access, new_val, value_type);
1696 }
1697 }
1698
1699 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1700 return _barrier_set->clone(this, src, dst, size, is_array);
1701 }
1702
1703 //-------------------------array_element_address-------------------------
1704 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1705 const TypeInt* sizetype, Node* ctrl) {
1706 uint shift = exact_log2(type2aelembytes(elembt));
1707 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1708
1709 // short-circuit a common case (saves lots of confusing waste motion)
1710 jint idx_con = find_int_con(idx, -1);
1711 if (idx_con >= 0) {
1712 intptr_t offset = header + ((intptr_t)idx_con << shift);
1713 return basic_plus_adr(ary, offset);
1714 }
1715
1716 // must be correct type for alignment purposes
1717 Node* base = basic_plus_adr(ary, header);
1718 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1719 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1720 return basic_plus_adr(ary, base, scale);
1721 }
1722
1723 //-------------------------load_array_element-------------------------
1724 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1725 const Type* elemtype = arytype->elem();
1726 BasicType elembt = elemtype->array_element_basic_type();
1727 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1728 if (elembt == T_NARROWOOP) {
1729 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1730 }
1731 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1732 return ld;
1733 }
1734
1735 //-------------------------set_arguments_for_java_call-------------------------
1736 // Arguments (pre-popped from the stack) are taken from the JVMS.
1737 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1738 // Add the call arguments:
1739 uint nargs = call->method()->arg_size();
1740 for (uint i = 0; i < nargs; i++) {
1741 Node* arg = argument(i);
1742 call->init_req(i + TypeFunc::Parms, arg);
1743 }
1744 }
1745
1746 //---------------------------set_edges_for_java_call---------------------------
1747 // Connect a newly created call into the current JVMS.
1748 // A return value node (if any) is returned from set_edges_for_java_call.
1749 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1750
1751 // Add the predefined inputs:
1752 call->init_req( TypeFunc::Control, control() );
1753 call->init_req( TypeFunc::I_O , i_o() );
1754 call->init_req( TypeFunc::Memory , reset_memory() );
1755 call->init_req( TypeFunc::FramePtr, frameptr() );
1756 call->init_req( TypeFunc::ReturnAdr, top() );
1757
1758 add_safepoint_edges(call, must_throw);
1759
1760 Node* xcall = _gvn.transform(call);
1761
1762 if (xcall == top()) {
1763 set_control(top());
1764 return;
1765 }
1766 assert(xcall == call, "call identity is stable");
1767
1768 // Re-use the current map to produce the result.
1769
1770 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1771 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1772 set_all_memory_call(xcall, separate_io_proj);
1773
1774 //return xcall; // no need, caller already has it
1775 }
1776
1777 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1778 if (stopped()) return top(); // maybe the call folded up?
1779
1780 // Capture the return value, if any.
1781 Node* ret;
1782 if (call->method() == NULL ||
1783 call->method()->return_type()->basic_type() == T_VOID)
1784 ret = top();
1785 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1786
1787 // Note: Since any out-of-line call can produce an exception,
1788 // we always insert an I_O projection from the call into the result.
1789
1790 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1791
1792 if (separate_io_proj) {
1793 // The caller requested separate projections be used by the fall
1794 // through and exceptional paths, so replace the projections for
1795 // the fall through path.
1796 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1797 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1798 }
1799 return ret;
1800 }
1801
1802 //--------------------set_predefined_input_for_runtime_call--------------------
1803 // Reading and setting the memory state is way conservative here.
1804 // The real problem is that I am not doing real Type analysis on memory,
1805 // so I cannot distinguish card mark stores from other stores. Across a GC
1806 // point the Store Barrier and the card mark memory has to agree. I cannot
1807 // have a card mark store and its barrier split across the GC point from
1808 // either above or below. Here I get that to happen by reading ALL of memory.
1809 // A better answer would be to separate out card marks from other memory.
1810 // For now, return the input memory state, so that it can be reused
1811 // after the call, if this call has restricted memory effects.
1812 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1813 // Set fixed predefined input arguments
1814 Node* memory = reset_memory();
1815 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1816 call->init_req( TypeFunc::Control, control() );
1817 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1818 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1819 call->init_req( TypeFunc::FramePtr, frameptr() );
1820 call->init_req( TypeFunc::ReturnAdr, top() );
1821 return memory;
1822 }
1823
1824 //-------------------set_predefined_output_for_runtime_call--------------------
1825 // Set control and memory (not i_o) from the call.
1826 // If keep_mem is not NULL, use it for the output state,
1827 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1828 // If hook_mem is NULL, this call produces no memory effects at all.
1829 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1830 // then only that memory slice is taken from the call.
1831 // In the last case, we must put an appropriate memory barrier before
1832 // the call, so as to create the correct anti-dependencies on loads
1833 // preceding the call.
1834 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1835 Node* keep_mem,
1836 const TypePtr* hook_mem) {
1837 // no i/o
1838 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1839 if (keep_mem) {
1840 // First clone the existing memory state
1841 set_all_memory(keep_mem);
1842 if (hook_mem != NULL) {
1843 // Make memory for the call
1844 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1845 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1846 // We also use hook_mem to extract specific effects from arraycopy stubs.
1847 set_memory(mem, hook_mem);
1848 }
1849 // ...else the call has NO memory effects.
1850
1851 // Make sure the call advertises its memory effects precisely.
1852 // This lets us build accurate anti-dependences in gcm.cpp.
1853 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1854 "call node must be constructed correctly");
1855 } else {
1856 assert(hook_mem == NULL, "");
1857 // This is not a "slow path" call; all memory comes from the call.
1858 set_all_memory_call(call);
1859 }
1860 }
1861
1862 // Keep track of MergeMems feeding into other MergeMems
1863 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
1864 if (!mem->is_MergeMem()) {
1865 return;
1866 }
1867 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
1868 Node* use = i.get();
1869 if (use->is_MergeMem()) {
1870 wl.push(use);
1871 }
1872 }
1873 }
1874
1875 // Replace the call with the current state of the kit.
1876 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1877 JVMState* ejvms = NULL;
1878 if (has_exceptions()) {
1879 ejvms = transfer_exceptions_into_jvms();
1880 }
1881
1882 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1883 ReplacedNodes replaced_nodes_exception;
1884 Node* ex_ctl = top();
1885
1886 SafePointNode* final_state = stop();
1887
1888 // Find all the needed outputs of this call
1889 CallProjections callprojs;
1890 call->extract_projections(&callprojs, true);
1891
1892 Unique_Node_List wl;
1893 Node* init_mem = call->in(TypeFunc::Memory);
1894 Node* final_mem = final_state->in(TypeFunc::Memory);
1895 Node* final_ctl = final_state->in(TypeFunc::Control);
1896 Node* final_io = final_state->in(TypeFunc::I_O);
1897
1898 // Replace all the old call edges with the edges from the inlining result
1899 if (callprojs.fallthrough_catchproj != NULL) {
1900 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1901 }
1902 if (callprojs.fallthrough_memproj != NULL) {
1903 if (final_mem->is_MergeMem()) {
1904 // Parser's exits MergeMem was not transformed but may be optimized
1905 final_mem = _gvn.transform(final_mem);
1906 }
1907 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1908 add_mergemem_users_to_worklist(wl, final_mem);
1909 }
1910 if (callprojs.fallthrough_ioproj != NULL) {
1911 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1912 }
1913
1914 // Replace the result with the new result if it exists and is used
1915 if (callprojs.resproj != NULL && result != NULL) {
1916 C->gvn_replace_by(callprojs.resproj, result);
1917 }
1918
1919 if (ejvms == NULL) {
1920 // No exception edges to simply kill off those paths
1921 if (callprojs.catchall_catchproj != NULL) {
1922 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1923 }
1924 if (callprojs.catchall_memproj != NULL) {
1925 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1926 }
1927 if (callprojs.catchall_ioproj != NULL) {
1928 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1929 }
1930 // Replace the old exception object with top
1931 if (callprojs.exobj != NULL) {
1932 C->gvn_replace_by(callprojs.exobj, C->top());
1933 }
1934 } else {
1935 GraphKit ekit(ejvms);
1936
1937 // Load my combined exception state into the kit, with all phis transformed:
1938 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1939 replaced_nodes_exception = ex_map->replaced_nodes();
1940
1941 Node* ex_oop = ekit.use_exception_state(ex_map);
1942
1943 if (callprojs.catchall_catchproj != NULL) {
1944 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1945 ex_ctl = ekit.control();
1946 }
1947 if (callprojs.catchall_memproj != NULL) {
1948 Node* ex_mem = ekit.reset_memory();
1949 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
1950 add_mergemem_users_to_worklist(wl, ex_mem);
1951 }
1952 if (callprojs.catchall_ioproj != NULL) {
1953 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1954 }
1955
1956 // Replace the old exception object with the newly created one
1957 if (callprojs.exobj != NULL) {
1958 C->gvn_replace_by(callprojs.exobj, ex_oop);
1959 }
1960 }
1961
1962 // Disconnect the call from the graph
1963 call->disconnect_inputs(NULL, C);
1964 C->gvn_replace_by(call, C->top());
1965
1966 // Clean up any MergeMems that feed other MergeMems since the
1967 // optimizer doesn't like that.
1968 while (wl.size() > 0) {
1969 _gvn.transform(wl.pop());
1970 }
1971
1972 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1973 replaced_nodes.apply(C, final_ctl);
1974 }
1975 if (!ex_ctl->is_top() && do_replaced_nodes) {
1976 replaced_nodes_exception.apply(C, ex_ctl);
1977 }
1978 }
1979
1980
1981 //------------------------------increment_counter------------------------------
1982 // for statistics: increment a VM counter by 1
1983
1984 void GraphKit::increment_counter(address counter_addr) {
1985 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1986 increment_counter(adr1);
1987 }
1988
1989 void GraphKit::increment_counter(Node* counter_addr) {
1990 int adr_type = Compile::AliasIdxRaw;
1991 Node* ctrl = control();
1992 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1993 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1994 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1995 }
1996
1997
1998 //------------------------------uncommon_trap----------------------------------
1999 // Bail out to the interpreter in mid-method. Implemented by calling the
2000 // uncommon_trap blob. This helper function inserts a runtime call with the
2001 // right debug info.
2002 void GraphKit::uncommon_trap(int trap_request,
2003 ciKlass* klass, const char* comment,
2004 bool must_throw,
2005 bool keep_exact_action) {
2006 if (failing()) stop();
2007 if (stopped()) return; // trap reachable?
2008
2009 // Note: If ProfileTraps is true, and if a deopt. actually
2010 // occurs here, the runtime will make sure an MDO exists. There is
2011 // no need to call method()->ensure_method_data() at this point.
2012
2013 // Set the stack pointer to the right value for reexecution:
2014 set_sp(reexecute_sp());
2015
2016 #ifdef ASSERT
2017 if (!must_throw) {
2018 // Make sure the stack has at least enough depth to execute
2019 // the current bytecode.
2020 int inputs, ignored_depth;
2021 if (compute_stack_effects(inputs, ignored_depth)) {
2022 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2023 Bytecodes::name(java_bc()), sp(), inputs);
2024 }
2025 }
2026 #endif
2027
2028 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2029 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2030
2031 switch (action) {
2032 case Deoptimization::Action_maybe_recompile:
2033 case Deoptimization::Action_reinterpret:
2034 // Temporary fix for 6529811 to allow virtual calls to be sure they
2035 // get the chance to go from mono->bi->mega
2036 if (!keep_exact_action &&
2037 Deoptimization::trap_request_index(trap_request) < 0 &&
2038 too_many_recompiles(reason)) {
2039 // This BCI is causing too many recompilations.
2040 if (C->log() != NULL) {
2041 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2042 Deoptimization::trap_reason_name(reason),
2043 Deoptimization::trap_action_name(action));
2044 }
2045 action = Deoptimization::Action_none;
2046 trap_request = Deoptimization::make_trap_request(reason, action);
2047 } else {
2048 C->set_trap_can_recompile(true);
2049 }
2050 break;
2051 case Deoptimization::Action_make_not_entrant:
2052 C->set_trap_can_recompile(true);
2053 break;
2054 case Deoptimization::Action_none:
2055 case Deoptimization::Action_make_not_compilable:
2056 break;
2057 default:
2058 #ifdef ASSERT
2059 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2060 #endif
2061 break;
2062 }
2063
2064 if (TraceOptoParse) {
2065 char buf[100];
2066 tty->print_cr("Uncommon trap %s at bci:%d",
2067 Deoptimization::format_trap_request(buf, sizeof(buf),
2068 trap_request), bci());
2069 }
2070
2071 CompileLog* log = C->log();
2072 if (log != NULL) {
2073 int kid = (klass == NULL)? -1: log->identify(klass);
2074 log->begin_elem("uncommon_trap bci='%d'", bci());
2075 char buf[100];
2076 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2077 trap_request));
2078 if (kid >= 0) log->print(" klass='%d'", kid);
2079 if (comment != NULL) log->print(" comment='%s'", comment);
2080 log->end_elem();
2081 }
2082
2083 // Make sure any guarding test views this path as very unlikely
2084 Node *i0 = control()->in(0);
2085 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2086 IfNode *iff = i0->as_If();
2087 float f = iff->_prob; // Get prob
2088 if (control()->Opcode() == Op_IfTrue) {
2089 if (f > PROB_UNLIKELY_MAG(4))
2090 iff->_prob = PROB_MIN;
2091 } else {
2092 if (f < PROB_LIKELY_MAG(4))
2093 iff->_prob = PROB_MAX;
2094 }
2095 }
2096
2097 // Clear out dead values from the debug info.
2098 kill_dead_locals();
2099
2100 // Now insert the uncommon trap subroutine call
2101 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2102 const TypePtr* no_memory_effects = NULL;
2103 // Pass the index of the class to be loaded
2104 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2105 (must_throw ? RC_MUST_THROW : 0),
2106 OptoRuntime::uncommon_trap_Type(),
2107 call_addr, "uncommon_trap", no_memory_effects,
2108 intcon(trap_request));
2109 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2110 "must extract request correctly from the graph");
2111 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2112
2113 call->set_req(TypeFunc::ReturnAdr, returnadr());
2114 // The debug info is the only real input to this call.
2115
2116 // Halt-and-catch fire here. The above call should never return!
2117 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2118 PRODUCT_ONLY(COMMA /*reachable*/false));
2119 _gvn.set_type_bottom(halt);
2120 root()->add_req(halt);
2121
2122 stop_and_kill_map();
2123 }
2124
2125
2126 //--------------------------just_allocated_object------------------------------
2127 // Report the object that was just allocated.
2128 // It must be the case that there are no intervening safepoints.
2129 // We use this to determine if an object is so "fresh" that
2130 // it does not require card marks.
2131 Node* GraphKit::just_allocated_object(Node* current_control) {
2132 Node* ctrl = current_control;
2133 // Object::<init> is invoked after allocation, most of invoke nodes
2134 // will be reduced, but a region node is kept in parse time, we check
2135 // the pattern and skip the region node if it degraded to a copy.
2136 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 &&
2137 ctrl->as_Region()->is_copy()) {
2138 ctrl = ctrl->as_Region()->is_copy();
2139 }
2140 if (C->recent_alloc_ctl() == ctrl) {
2141 return C->recent_alloc_obj();
2142 }
2143 return NULL;
2144 }
2145
2146
2147 /**
2148 * Record profiling data exact_kls for Node n with the type system so
2149 * that it can propagate it (speculation)
2150 *
2151 * @param n node that the type applies to
2152 * @param exact_kls type from profiling
2153 * @param maybe_null did profiling see null?
2154 *
2155 * @return node with improved type
2156 */
2157 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2158 const Type* current_type = _gvn.type(n);
2159 assert(UseTypeSpeculation, "type speculation must be on");
2160
2161 const TypePtr* speculative = current_type->speculative();
2162
2163 // Should the klass from the profile be recorded in the speculative type?
2164 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2165 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2166 const TypeOopPtr* xtype = tklass->as_instance_type();
2167 assert(xtype->klass_is_exact(), "Should be exact");
2168 // Any reason to believe n is not null (from this profiling or a previous one)?
2169 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2170 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2171 // record the new speculative type's depth
2172 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2173 speculative = speculative->with_inline_depth(jvms()->depth());
2174 } else if (current_type->would_improve_ptr(ptr_kind)) {
2175 // Profiling report that null was never seen so we can change the
2176 // speculative type to non null ptr.
2177 if (ptr_kind == ProfileAlwaysNull) {
2178 speculative = TypePtr::NULL_PTR;
2179 } else {
2180 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2181 const TypePtr* ptr = TypePtr::NOTNULL;
2182 if (speculative != NULL) {
2183 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2184 } else {
2185 speculative = ptr;
2186 }
2187 }
2188 }
2189
2190 if (speculative != current_type->speculative()) {
2191 // Build a type with a speculative type (what we think we know
2192 // about the type but will need a guard when we use it)
2193 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2194 // We're changing the type, we need a new CheckCast node to carry
2195 // the new type. The new type depends on the control: what
2196 // profiling tells us is only valid from here as far as we can
2197 // tell.
2198 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2199 cast = _gvn.transform(cast);
2200 replace_in_map(n, cast);
2201 n = cast;
2202 }
2203
2204 return n;
2205 }
2206
2207 /**
2208 * Record profiling data from receiver profiling at an invoke with the
2209 * type system so that it can propagate it (speculation)
2210 *
2211 * @param n receiver node
2212 *
2213 * @return node with improved type
2214 */
2215 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2216 if (!UseTypeSpeculation) {
2217 return n;
2218 }
2219 ciKlass* exact_kls = profile_has_unique_klass();
2220 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2221 if ((java_bc() == Bytecodes::_checkcast ||
2222 java_bc() == Bytecodes::_instanceof ||
2223 java_bc() == Bytecodes::_aastore) &&
2224 method()->method_data()->is_mature()) {
2225 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2226 if (data != NULL) {
2227 if (!data->as_BitData()->null_seen()) {
2228 ptr_kind = ProfileNeverNull;
2229 } else {
2230 assert(data->is_ReceiverTypeData(), "bad profile data type");
2231 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2232 uint i = 0;
2233 for (; i < call->row_limit(); i++) {
2234 ciKlass* receiver = call->receiver(i);
2235 if (receiver != NULL) {
2236 break;
2237 }
2238 }
2239 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2240 }
2241 }
2242 }
2243 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2244 }
2245
2246 /**
2247 * Record profiling data from argument profiling at an invoke with the
2248 * type system so that it can propagate it (speculation)
2249 *
2250 * @param dest_method target method for the call
2251 * @param bc what invoke bytecode is this?
2252 */
2253 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2254 if (!UseTypeSpeculation) {
2255 return;
2256 }
2257 const TypeFunc* tf = TypeFunc::make(dest_method);
2258 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2259 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2260 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2261 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2262 if (is_reference_type(targ->basic_type())) {
2263 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2264 ciKlass* better_type = NULL;
2265 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2266 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2267 }
2268 i++;
2269 }
2270 }
2271 }
2272
2273 /**
2274 * Record profiling data from parameter profiling at an invoke with
2275 * the type system so that it can propagate it (speculation)
2276 */
2277 void GraphKit::record_profiled_parameters_for_speculation() {
2278 if (!UseTypeSpeculation) {
2279 return;
2280 }
2281 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2282 if (_gvn.type(local(i))->isa_oopptr()) {
2283 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2284 ciKlass* better_type = NULL;
2285 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2286 record_profile_for_speculation(local(i), better_type, ptr_kind);
2287 }
2288 j++;
2289 }
2290 }
2291 }
2292
2293 /**
2294 * Record profiling data from return value profiling at an invoke with
2295 * the type system so that it can propagate it (speculation)
2296 */
2297 void GraphKit::record_profiled_return_for_speculation() {
2298 if (!UseTypeSpeculation) {
2299 return;
2300 }
2301 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2302 ciKlass* better_type = NULL;
2303 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2304 // If profiling reports a single type for the return value,
2305 // feed it to the type system so it can propagate it as a
2306 // speculative type
2307 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2308 }
2309 }
2310
2311 void GraphKit::round_double_result(ciMethod* dest_method) {
2312 if (Matcher::strict_fp_requires_explicit_rounding) {
2313 // If a strict caller invokes a non-strict callee, round a double result.
2314 // A non-strict method may return a double value which has an extended exponent,
2315 // but this must not be visible in a caller which is strict.
2316 BasicType result_type = dest_method->return_type()->basic_type();
2317 assert(method() != NULL, "must have caller context");
2318 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2319 // Destination method's return value is on top of stack
2320 // dstore_rounding() does gvn.transform
2321 Node *result = pop_pair();
2322 result = dstore_rounding(result);
2323 push_pair(result);
2324 }
2325 }
2326 }
2327
2328 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2329 if (Matcher::strict_fp_requires_explicit_rounding) {
2330 // (Note: TypeFunc::make has a cache that makes this fast.)
2331 const TypeFunc* tf = TypeFunc::make(dest_method);
2332 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2333 for (int j = 0; j < nargs; j++) {
2334 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2335 if (targ->basic_type() == T_DOUBLE) {
2336 // If any parameters are doubles, they must be rounded before
2337 // the call, dstore_rounding does gvn.transform
2338 Node *arg = argument(j);
2339 arg = dstore_rounding(arg);
2340 set_argument(j, arg);
2341 }
2342 }
2343 }
2344 }
2345
2346 // rounding for strict float precision conformance
2347 Node* GraphKit::precision_rounding(Node* n) {
2348 if (Matcher::strict_fp_requires_explicit_rounding) {
2349 #ifdef IA32
2350 if (_method->flags().is_strict() && UseSSE == 0) {
2351 return _gvn.transform(new RoundFloatNode(0, n));
2352 }
2353 #else
2354 Unimplemented();
2355 #endif // IA32
2356 }
2357 return n;
2358 }
2359
2360 // rounding for strict double precision conformance
2361 Node* GraphKit::dprecision_rounding(Node *n) {
2362 if (Matcher::strict_fp_requires_explicit_rounding) {
2363 #ifdef IA32
2364 if (_method->flags().is_strict() && UseSSE < 2) {
2365 return _gvn.transform(new RoundDoubleNode(0, n));
2366 }
2367 #else
2368 Unimplemented();
2369 #endif // IA32
2370 }
2371 return n;
2372 }
2373
2374 // rounding for non-strict double stores
2375 Node* GraphKit::dstore_rounding(Node* n) {
2376 if (Matcher::strict_fp_requires_explicit_rounding) {
2377 #ifdef IA32
2378 if (UseSSE < 2) {
2379 return _gvn.transform(new RoundDoubleNode(0, n));
2380 }
2381 #else
2382 Unimplemented();
2383 #endif // IA32
2384 }
2385 return n;
2386 }
2387
2388 //=============================================================================
2389 // Generate a fast path/slow path idiom. Graph looks like:
2390 // [foo] indicates that 'foo' is a parameter
2391 //
2392 // [in] NULL
2393 // \ /
2394 // CmpP
2395 // Bool ne
2396 // If
2397 // / \
2398 // True False-<2>
2399 // / |
2400 // / cast_not_null
2401 // Load | | ^
2402 // [fast_test] | |
2403 // gvn to opt_test | |
2404 // / \ | <1>
2405 // True False |
2406 // | \\ |
2407 // [slow_call] \[fast_result]
2408 // Ctl Val \ \
2409 // | \ \
2410 // Catch <1> \ \
2411 // / \ ^ \ \
2412 // Ex No_Ex | \ \
2413 // | \ \ | \ <2> \
2414 // ... \ [slow_res] | | \ [null_result]
2415 // \ \--+--+--- | |
2416 // \ | / \ | /
2417 // --------Region Phi
2418 //
2419 //=============================================================================
2420 // Code is structured as a series of driver functions all called 'do_XXX' that
2421 // call a set of helper functions. Helper functions first, then drivers.
2422
2423 //------------------------------null_check_oop---------------------------------
2424 // Null check oop. Set null-path control into Region in slot 3.
2425 // Make a cast-not-nullness use the other not-null control. Return cast.
2426 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2427 bool never_see_null,
2428 bool safe_for_replace,
2429 bool speculative) {
2430 // Initial NULL check taken path
2431 (*null_control) = top();
2432 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2433
2434 // Generate uncommon_trap:
2435 if (never_see_null && (*null_control) != top()) {
2436 // If we see an unexpected null at a check-cast we record it and force a
2437 // recompile; the offending check-cast will be compiled to handle NULLs.
2438 // If we see more than one offending BCI, then all checkcasts in the
2439 // method will be compiled to handle NULLs.
2440 PreserveJVMState pjvms(this);
2441 set_control(*null_control);
2442 replace_in_map(value, null());
2443 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2444 uncommon_trap(reason,
2445 Deoptimization::Action_make_not_entrant);
2446 (*null_control) = top(); // NULL path is dead
2447 }
2448 if ((*null_control) == top() && safe_for_replace) {
2449 replace_in_map(value, cast);
2450 }
2451
2452 // Cast away null-ness on the result
2453 return cast;
2454 }
2455
2456 //------------------------------opt_iff----------------------------------------
2457 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2458 // Return slow-path control.
2459 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2460 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2461
2462 // Fast path taken; set region slot 2
2463 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2464 region->init_req(2,fast_taken); // Capture fast-control
2465
2466 // Fast path not-taken, i.e. slow path
2467 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2468 return slow_taken;
2469 }
2470
2471 //-----------------------------make_runtime_call-------------------------------
2472 Node* GraphKit::make_runtime_call(int flags,
2473 const TypeFunc* call_type, address call_addr,
2474 const char* call_name,
2475 const TypePtr* adr_type,
2476 // The following parms are all optional.
2477 // The first NULL ends the list.
2478 Node* parm0, Node* parm1,
2479 Node* parm2, Node* parm3,
2480 Node* parm4, Node* parm5,
2481 Node* parm6, Node* parm7) {
2482 assert(call_addr != NULL, "must not call NULL targets");
2483
2484 // Slow-path call
2485 bool is_leaf = !(flags & RC_NO_LEAF);
2486 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2487 if (call_name == NULL) {
2488 assert(!is_leaf, "must supply name for leaf");
2489 call_name = OptoRuntime::stub_name(call_addr);
2490 }
2491 CallNode* call;
2492 if (!is_leaf) {
2493 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2494 bci(), adr_type);
2495 } else if (flags & RC_NO_FP) {
2496 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2497 } else {
2498 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2499 }
2500
2501 // The following is similar to set_edges_for_java_call,
2502 // except that the memory effects of the call are restricted to AliasIdxRaw.
2503
2504 // Slow path call has no side-effects, uses few values
2505 bool wide_in = !(flags & RC_NARROW_MEM);
2506 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2507
2508 Node* prev_mem = NULL;
2509 if (wide_in) {
2510 prev_mem = set_predefined_input_for_runtime_call(call);
2511 } else {
2512 assert(!wide_out, "narrow in => narrow out");
2513 Node* narrow_mem = memory(adr_type);
2514 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2515 }
2516
2517 // Hook each parm in order. Stop looking at the first NULL.
2518 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2519 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2520 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2521 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2522 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2523 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2524 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2525 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2526 /* close each nested if ===> */ } } } } } } } }
2527 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2528
2529 if (!is_leaf) {
2530 // Non-leaves can block and take safepoints:
2531 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2532 }
2533 // Non-leaves can throw exceptions:
2534 if (has_io) {
2535 call->set_req(TypeFunc::I_O, i_o());
2536 }
2537
2538 if (flags & RC_UNCOMMON) {
2539 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2540 // (An "if" probability corresponds roughly to an unconditional count.
2541 // Sort of.)
2542 call->set_cnt(PROB_UNLIKELY_MAG(4));
2543 }
2544
2545 Node* c = _gvn.transform(call);
2546 assert(c == call, "cannot disappear");
2547
2548 if (wide_out) {
2549 // Slow path call has full side-effects.
2550 set_predefined_output_for_runtime_call(call);
2551 } else {
2552 // Slow path call has few side-effects, and/or sets few values.
2553 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2554 }
2555
2556 if (has_io) {
2557 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2558 }
2559 return call;
2560
2561 }
2562
2563 //------------------------------merge_memory-----------------------------------
2564 // Merge memory from one path into the current memory state.
2565 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2566 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2567 Node* old_slice = mms.force_memory();
2568 Node* new_slice = mms.memory2();
2569 if (old_slice != new_slice) {
2570 PhiNode* phi;
2571 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2572 if (mms.is_empty()) {
2573 // clone base memory Phi's inputs for this memory slice
2574 assert(old_slice == mms.base_memory(), "sanity");
2575 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2576 _gvn.set_type(phi, Type::MEMORY);
2577 for (uint i = 1; i < phi->req(); i++) {
2578 phi->init_req(i, old_slice->in(i));
2579 }
2580 } else {
2581 phi = old_slice->as_Phi(); // Phi was generated already
2582 }
2583 } else {
2584 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2585 _gvn.set_type(phi, Type::MEMORY);
2586 }
2587 phi->set_req(new_path, new_slice);
2588 mms.set_memory(phi);
2589 }
2590 }
2591 }
2592
2593 //------------------------------make_slow_call_ex------------------------------
2594 // Make the exception handler hookups for the slow call
2595 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2596 if (stopped()) return;
2597
2598 // Make a catch node with just two handlers: fall-through and catch-all
2599 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2600 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2601 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2602 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2603
2604 { PreserveJVMState pjvms(this);
2605 set_control(excp);
2606 set_i_o(i_o);
2607
2608 if (excp != top()) {
2609 if (deoptimize) {
2610 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2611 uncommon_trap(Deoptimization::Reason_unhandled,
2612 Deoptimization::Action_none);
2613 } else {
2614 // Create an exception state also.
2615 // Use an exact type if the caller has a specific exception.
2616 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2617 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2618 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2619 }
2620 }
2621 }
2622
2623 // Get the no-exception control from the CatchNode.
2624 set_control(norm);
2625 }
2626
2627 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2628 Node* cmp = NULL;
2629 switch(bt) {
2630 case T_INT: cmp = new CmpINode(in1, in2); break;
2631 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2632 default: fatal("unexpected comparison type %s", type2name(bt));
2633 }
2634 gvn.transform(cmp);
2635 Node* bol = gvn.transform(new BoolNode(cmp, test));
2636 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2637 gvn.transform(iff);
2638 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2639 return iff;
2640 }
2641
2642 //-------------------------------gen_subtype_check-----------------------------
2643 // Generate a subtyping check. Takes as input the subtype and supertype.
2644 // Returns 2 values: sets the default control() to the true path and returns
2645 // the false path. Only reads invariant memory; sets no (visible) memory.
2646 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2647 // but that's not exposed to the optimizer. This call also doesn't take in an
2648 // Object; if you wish to check an Object you need to load the Object's class
2649 // prior to coming here.
2650 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn) {
2651 Compile* C = gvn.C;
2652 if ((*ctrl)->is_top()) {
2653 return C->top();
2654 }
2655
2656 // Fast check for identical types, perhaps identical constants.
2657 // The types can even be identical non-constants, in cases
2658 // involving Array.newInstance, Object.clone, etc.
2659 if (subklass == superklass)
2660 return C->top(); // false path is dead; no test needed.
2661
2662 if (gvn.type(superklass)->singleton()) {
2663 ciKlass* superk = gvn.type(superklass)->is_klassptr()->klass();
2664 ciKlass* subk = gvn.type(subklass)->is_klassptr()->klass();
2665
2666 // In the common case of an exact superklass, try to fold up the
2667 // test before generating code. You may ask, why not just generate
2668 // the code and then let it fold up? The answer is that the generated
2669 // code will necessarily include null checks, which do not always
2670 // completely fold away. If they are also needless, then they turn
2671 // into a performance loss. Example:
2672 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2673 // Here, the type of 'fa' is often exact, so the store check
2674 // of fa[1]=x will fold up, without testing the nullness of x.
2675 switch (C->static_subtype_check(superk, subk)) {
2676 case Compile::SSC_always_false:
2677 {
2678 Node* always_fail = *ctrl;
2679 *ctrl = gvn.C->top();
2680 return always_fail;
2681 }
2682 case Compile::SSC_always_true:
2683 return C->top();
2684 case Compile::SSC_easy_test:
2685 {
2686 // Just do a direct pointer compare and be done.
2687 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2688 *ctrl = gvn.transform(new IfTrueNode(iff));
2689 return gvn.transform(new IfFalseNode(iff));
2690 }
2691 case Compile::SSC_full_test:
2692 break;
2693 default:
2694 ShouldNotReachHere();
2695 }
2696 }
2697
2698 // %%% Possible further optimization: Even if the superklass is not exact,
2699 // if the subklass is the unique subtype of the superklass, the check
2700 // will always succeed. We could leave a dependency behind to ensure this.
2701
2702 // First load the super-klass's check-offset
2703 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2704 Node* m = C->immutable_memory();
2705 Node *chk_off = gvn.transform(new LoadINode(NULL, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2706 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2707 bool might_be_cache = (gvn.find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2708
2709 // Load from the sub-klass's super-class display list, or a 1-word cache of
2710 // the secondary superclass list, or a failing value with a sentinel offset
2711 // if the super-klass is an interface or exceptionally deep in the Java
2712 // hierarchy and we have to scan the secondary superclass list the hard way.
2713 // Worst-case type is a little odd: NULL is allowed as a result (usually
2714 // klass loads can never produce a NULL).
2715 Node *chk_off_X = chk_off;
2716 #ifdef _LP64
2717 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2718 #endif
2719 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2720 // For some types like interfaces the following loadKlass is from a 1-word
2721 // cache which is mutable so can't use immutable memory. Other
2722 // types load from the super-class display table which is immutable.
2723 Node *kmem = C->immutable_memory();
2724 // secondary_super_cache is not immutable but can be treated as such because:
2725 // - no ideal node writes to it in a way that could cause an
2726 // incorrect/missed optimization of the following Load.
2727 // - it's a cache so, worse case, not reading the latest value
2728 // wouldn't cause incorrect execution
2729 if (might_be_cache && mem != NULL) {
2730 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
2731 }
2732 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, NULL, kmem, p2, gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2733
2734 // Compile speed common case: ARE a subtype and we canNOT fail
2735 if( superklass == nkls )
2736 return C->top(); // false path is dead; no test needed.
2737
2738 // See if we get an immediate positive hit. Happens roughly 83% of the
2739 // time. Test to see if the value loaded just previously from the subklass
2740 // is exactly the superklass.
2741 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2742 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
2743 *ctrl = gvn.transform(new IfFalseNode(iff1));
2744
2745 // Compile speed common case: Check for being deterministic right now. If
2746 // chk_off is a constant and not equal to cacheoff then we are NOT a
2747 // subklass. In this case we need exactly the 1 test above and we can
2748 // return those results immediately.
2749 if (!might_be_cache) {
2750 Node* not_subtype_ctrl = *ctrl;
2751 *ctrl = iftrue1; // We need exactly the 1 test above
2752 return not_subtype_ctrl;
2753 }
2754
2755 // Gather the various success & failures here
2756 RegionNode *r_ok_subtype = new RegionNode(4);
2757 gvn.record_for_igvn(r_ok_subtype);
2758 RegionNode *r_not_subtype = new RegionNode(3);
2759 gvn.record_for_igvn(r_not_subtype);
2760
2761 r_ok_subtype->init_req(1, iftrue1);
2762
2763 // Check for immediate negative hit. Happens roughly 11% of the time (which
2764 // is roughly 63% of the remaining cases). Test to see if the loaded
2765 // check-offset points into the subklass display list or the 1-element
2766 // cache. If it points to the display (and NOT the cache) and the display
2767 // missed then it's not a subtype.
2768 Node *cacheoff = gvn.intcon(cacheoff_con);
2769 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2770 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2771 *ctrl = gvn.transform(new IfFalseNode(iff2));
2772
2773 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2774 // No performance impact (too rare) but allows sharing of secondary arrays
2775 // which has some footprint reduction.
2776 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2777 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2778 *ctrl = gvn.transform(new IfFalseNode(iff3));
2779
2780 // -- Roads not taken here: --
2781 // We could also have chosen to perform the self-check at the beginning
2782 // of this code sequence, as the assembler does. This would not pay off
2783 // the same way, since the optimizer, unlike the assembler, can perform
2784 // static type analysis to fold away many successful self-checks.
2785 // Non-foldable self checks work better here in second position, because
2786 // the initial primary superclass check subsumes a self-check for most
2787 // types. An exception would be a secondary type like array-of-interface,
2788 // which does not appear in its own primary supertype display.
2789 // Finally, we could have chosen to move the self-check into the
2790 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2791 // dependent manner. But it is worthwhile to have the check here,
2792 // where it can be perhaps be optimized. The cost in code space is
2793 // small (register compare, branch).
2794
2795 // Now do a linear scan of the secondary super-klass array. Again, no real
2796 // performance impact (too rare) but it's gotta be done.
2797 // Since the code is rarely used, there is no penalty for moving it
2798 // out of line, and it can only improve I-cache density.
2799 // The decision to inline or out-of-line this final check is platform
2800 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2801 Node* psc = gvn.transform(
2802 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2803
2804 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2805 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2806 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2807
2808 // Return false path; set default control to true path.
2809 *ctrl = gvn.transform(r_ok_subtype);
2810 return gvn.transform(r_not_subtype);
2811 }
2812
2813 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2814 if (ExpandSubTypeCheckAtParseTime) {
2815 MergeMemNode* mem = merged_memory();
2816 Node* ctrl = control();
2817 Node* subklass = obj_or_subklass;
2818 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2819 subklass = load_object_klass(obj_or_subklass);
2820 }
2821
2822 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn);
2823 set_control(ctrl);
2824 return n;
2825 }
2826
2827 const TypePtr* adr_type = TypeKlassPtr::make(TypePtr::NotNull, C->env()->Object_klass(), Type::OffsetBot);
2828 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass));
2829 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2830 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2831 set_control(_gvn.transform(new IfTrueNode(iff)));
2832 return _gvn.transform(new IfFalseNode(iff));
2833 }
2834
2835 // Profile-driven exact type check:
2836 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2837 float prob,
2838 Node* *casted_receiver) {
2839 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2840 Node* recv_klass = load_object_klass(receiver);
2841 Node* want_klass = makecon(tklass);
2842 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2843 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2844 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2845 set_control( _gvn.transform( new IfTrueNode (iff) ));
2846 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2847
2848 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2849 assert(recv_xtype->klass_is_exact(), "");
2850
2851 // Subsume downstream occurrences of receiver with a cast to
2852 // recv_xtype, since now we know what the type will be.
2853 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2854 (*casted_receiver) = _gvn.transform(cast);
2855 // (User must make the replace_in_map call.)
2856
2857 return fail;
2858 }
2859
2860 //------------------------------subtype_check_receiver-------------------------
2861 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2862 Node** casted_receiver) {
2863 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2864 Node* want_klass = makecon(tklass);
2865
2866 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
2867
2868 // Cast receiver after successful check
2869 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2870 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2871 (*casted_receiver) = _gvn.transform(cast);
2872
2873 return slow_ctl;
2874 }
2875
2876 //------------------------------seems_never_null-------------------------------
2877 // Use null_seen information if it is available from the profile.
2878 // If we see an unexpected null at a type check we record it and force a
2879 // recompile; the offending check will be recompiled to handle NULLs.
2880 // If we see several offending BCIs, then all checks in the
2881 // method will be recompiled.
2882 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2883 speculating = !_gvn.type(obj)->speculative_maybe_null();
2884 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2885 if (UncommonNullCast // Cutout for this technique
2886 && obj != null() // And not the -Xcomp stupid case?
2887 && !too_many_traps(reason)
2888 ) {
2889 if (speculating) {
2890 return true;
2891 }
2892 if (data == NULL)
2893 // Edge case: no mature data. Be optimistic here.
2894 return true;
2895 // If the profile has not seen a null, assume it won't happen.
2896 assert(java_bc() == Bytecodes::_checkcast ||
2897 java_bc() == Bytecodes::_instanceof ||
2898 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2899 return !data->as_BitData()->null_seen();
2900 }
2901 speculating = false;
2902 return false;
2903 }
2904
2905 void GraphKit::guard_klass_being_initialized(Node* klass) {
2906 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
2907 Node* adr = basic_plus_adr(top(), klass, init_state_off);
2908 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
2909 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
2910 T_BYTE, MemNode::unordered);
2911 init_state = _gvn.transform(init_state);
2912
2913 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
2914
2915 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
2916 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
2917
2918 { BuildCutout unless(this, tst, PROB_MAX);
2919 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
2920 }
2921 }
2922
2923 void GraphKit::guard_init_thread(Node* klass) {
2924 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
2925 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
2926
2927 Node* init_thread = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
2928 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
2929 T_ADDRESS, MemNode::unordered);
2930 init_thread = _gvn.transform(init_thread);
2931
2932 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
2933
2934 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
2935 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
2936
2937 { BuildCutout unless(this, tst, PROB_MAX);
2938 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
2939 }
2940 }
2941
2942 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
2943 if (ik->is_being_initialized()) {
2944 if (C->needs_clinit_barrier(ik, context)) {
2945 Node* klass = makecon(TypeKlassPtr::make(ik));
2946 guard_klass_being_initialized(klass);
2947 guard_init_thread(klass);
2948 insert_mem_bar(Op_MemBarCPUOrder);
2949 }
2950 } else if (ik->is_initialized()) {
2951 return; // no barrier needed
2952 } else {
2953 uncommon_trap(Deoptimization::Reason_uninitialized,
2954 Deoptimization::Action_reinterpret,
2955 NULL);
2956 }
2957 }
2958
2959 //------------------------maybe_cast_profiled_receiver-------------------------
2960 // If the profile has seen exactly one type, narrow to exactly that type.
2961 // Subsequent type checks will always fold up.
2962 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2963 ciKlass* require_klass,
2964 ciKlass* spec_klass,
2965 bool safe_for_replace) {
2966 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2967
2968 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2969
2970 // Make sure we haven't already deoptimized from this tactic.
2971 if (too_many_traps_or_recompiles(reason))
2972 return NULL;
2973
2974 // (No, this isn't a call, but it's enough like a virtual call
2975 // to use the same ciMethod accessor to get the profile info...)
2976 // If we have a speculative type use it instead of profiling (which
2977 // may not help us)
2978 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2979 if (exact_kls != NULL) {// no cast failures here
2980 if (require_klass == NULL ||
2981 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2982 // If we narrow the type to match what the type profile sees or
2983 // the speculative type, we can then remove the rest of the
2984 // cast.
2985 // This is a win, even if the exact_kls is very specific,
2986 // because downstream operations, such as method calls,
2987 // will often benefit from the sharper type.
2988 Node* exact_obj = not_null_obj; // will get updated in place...
2989 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2990 &exact_obj);
2991 { PreserveJVMState pjvms(this);
2992 set_control(slow_ctl);
2993 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2994 }
2995 if (safe_for_replace) {
2996 replace_in_map(not_null_obj, exact_obj);
2997 }
2998 return exact_obj;
2999 }
3000 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3001 }
3002
3003 return NULL;
3004 }
3005
3006 /**
3007 * Cast obj to type and emit guard unless we had too many traps here
3008 * already
3009 *
3010 * @param obj node being casted
3011 * @param type type to cast the node to
3012 * @param not_null true if we know node cannot be null
3013 */
3014 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3015 ciKlass* type,
3016 bool not_null) {
3017 if (stopped()) {
3018 return obj;
3019 }
3020
3021 // type == NULL if profiling tells us this object is always null
3022 if (type != NULL) {
3023 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3024 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3025
3026 if (!too_many_traps_or_recompiles(null_reason) &&
3027 !too_many_traps_or_recompiles(class_reason)) {
3028 Node* not_null_obj = NULL;
3029 // not_null is true if we know the object is not null and
3030 // there's no need for a null check
3031 if (!not_null) {
3032 Node* null_ctl = top();
3033 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3034 assert(null_ctl->is_top(), "no null control here");
3035 } else {
3036 not_null_obj = obj;
3037 }
3038
3039 Node* exact_obj = not_null_obj;
3040 ciKlass* exact_kls = type;
3041 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3042 &exact_obj);
3043 {
3044 PreserveJVMState pjvms(this);
3045 set_control(slow_ctl);
3046 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3047 }
3048 replace_in_map(not_null_obj, exact_obj);
3049 obj = exact_obj;
3050 }
3051 } else {
3052 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3053 Node* exact_obj = null_assert(obj);
3054 replace_in_map(obj, exact_obj);
3055 obj = exact_obj;
3056 }
3057 }
3058 return obj;
3059 }
3060
3061 //-------------------------------gen_instanceof--------------------------------
3062 // Generate an instance-of idiom. Used by both the instance-of bytecode
3063 // and the reflective instance-of call.
3064 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3065 kill_dead_locals(); // Benefit all the uncommon traps
3066 assert( !stopped(), "dead parse path should be checked in callers" );
3067 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3068 "must check for not-null not-dead klass in callers");
3069
3070 // Make the merge point
3071 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3072 RegionNode* region = new RegionNode(PATH_LIMIT);
3073 Node* phi = new PhiNode(region, TypeInt::BOOL);
3074 C->set_has_split_ifs(true); // Has chance for split-if optimization
3075
3076 ciProfileData* data = NULL;
3077 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3078 data = method()->method_data()->bci_to_data(bci());
3079 }
3080 bool speculative_not_null = false;
3081 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3082 && seems_never_null(obj, data, speculative_not_null));
3083
3084 // Null check; get casted pointer; set region slot 3
3085 Node* null_ctl = top();
3086 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3087
3088 // If not_null_obj is dead, only null-path is taken
3089 if (stopped()) { // Doing instance-of on a NULL?
3090 set_control(null_ctl);
3091 return intcon(0);
3092 }
3093 region->init_req(_null_path, null_ctl);
3094 phi ->init_req(_null_path, intcon(0)); // Set null path value
3095 if (null_ctl == top()) {
3096 // Do this eagerly, so that pattern matches like is_diamond_phi
3097 // will work even during parsing.
3098 assert(_null_path == PATH_LIMIT-1, "delete last");
3099 region->del_req(_null_path);
3100 phi ->del_req(_null_path);
3101 }
3102
3103 // Do we know the type check always succeed?
3104 bool known_statically = false;
3105 if (_gvn.type(superklass)->singleton()) {
3106 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3107 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3108 if (subk != NULL && subk->is_loaded()) {
3109 int static_res = C->static_subtype_check(superk, subk);
3110 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3111 }
3112 }
3113
3114 if (!known_statically) {
3115 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3116 // We may not have profiling here or it may not help us. If we
3117 // have a speculative type use it to perform an exact cast.
3118 ciKlass* spec_obj_type = obj_type->speculative_type();
3119 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3120 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3121 if (stopped()) { // Profile disagrees with this path.
3122 set_control(null_ctl); // Null is the only remaining possibility.
3123 return intcon(0);
3124 }
3125 if (cast_obj != NULL) {
3126 not_null_obj = cast_obj;
3127 }
3128 }
3129 }
3130
3131 // Generate the subtype check
3132 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3133
3134 // Plug in the success path to the general merge in slot 1.
3135 region->init_req(_obj_path, control());
3136 phi ->init_req(_obj_path, intcon(1));
3137
3138 // Plug in the failing path to the general merge in slot 2.
3139 region->init_req(_fail_path, not_subtype_ctrl);
3140 phi ->init_req(_fail_path, intcon(0));
3141
3142 // Return final merged results
3143 set_control( _gvn.transform(region) );
3144 record_for_igvn(region);
3145
3146 // If we know the type check always succeeds then we don't use the
3147 // profiling data at this bytecode. Don't lose it, feed it to the
3148 // type system as a speculative type.
3149 if (safe_for_replace) {
3150 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3151 replace_in_map(obj, casted_obj);
3152 }
3153
3154 return _gvn.transform(phi);
3155 }
3156
3157 //-------------------------------gen_checkcast---------------------------------
3158 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3159 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3160 // uncommon-trap paths work. Adjust stack after this call.
3161 // If failure_control is supplied and not null, it is filled in with
3162 // the control edge for the cast failure. Otherwise, an appropriate
3163 // uncommon trap or exception is thrown.
3164 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3165 Node* *failure_control) {
3166 kill_dead_locals(); // Benefit all the uncommon traps
3167 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3168 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3169
3170 // Fast cutout: Check the case that the cast is vacuously true.
3171 // This detects the common cases where the test will short-circuit
3172 // away completely. We do this before we perform the null check,
3173 // because if the test is going to turn into zero code, we don't
3174 // want a residual null check left around. (Causes a slowdown,
3175 // for example, in some objArray manipulations, such as a[i]=a[j].)
3176 if (tk->singleton()) {
3177 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3178 if (objtp != NULL && objtp->klass() != NULL) {
3179 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3180 case Compile::SSC_always_true:
3181 // If we know the type check always succeed then we don't use
3182 // the profiling data at this bytecode. Don't lose it, feed it
3183 // to the type system as a speculative type.
3184 return record_profiled_receiver_for_speculation(obj);
3185 case Compile::SSC_always_false:
3186 // It needs a null check because a null will *pass* the cast check.
3187 // A non-null value will always produce an exception.
3188 return null_assert(obj);
3189 }
3190 }
3191 }
3192
3193 ciProfileData* data = NULL;
3194 bool safe_for_replace = false;
3195 if (failure_control == NULL) { // use MDO in regular case only
3196 assert(java_bc() == Bytecodes::_aastore ||
3197 java_bc() == Bytecodes::_checkcast,
3198 "interpreter profiles type checks only for these BCs");
3199 data = method()->method_data()->bci_to_data(bci());
3200 safe_for_replace = true;
3201 }
3202
3203 // Make the merge point
3204 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3205 RegionNode* region = new RegionNode(PATH_LIMIT);
3206 Node* phi = new PhiNode(region, toop);
3207 C->set_has_split_ifs(true); // Has chance for split-if optimization
3208
3209 // Use null-cast information if it is available
3210 bool speculative_not_null = false;
3211 bool never_see_null = ((failure_control == NULL) // regular case only
3212 && seems_never_null(obj, data, speculative_not_null));
3213
3214 // Null check; get casted pointer; set region slot 3
3215 Node* null_ctl = top();
3216 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3217
3218 // If not_null_obj is dead, only null-path is taken
3219 if (stopped()) { // Doing instance-of on a NULL?
3220 set_control(null_ctl);
3221 return null();
3222 }
3223 region->init_req(_null_path, null_ctl);
3224 phi ->init_req(_null_path, null()); // Set null path value
3225 if (null_ctl == top()) {
3226 // Do this eagerly, so that pattern matches like is_diamond_phi
3227 // will work even during parsing.
3228 assert(_null_path == PATH_LIMIT-1, "delete last");
3229 region->del_req(_null_path);
3230 phi ->del_req(_null_path);
3231 }
3232
3233 Node* cast_obj = NULL;
3234 if (tk->klass_is_exact()) {
3235 // The following optimization tries to statically cast the speculative type of the object
3236 // (for example obtained during profiling) to the type of the superklass and then do a
3237 // dynamic check that the type of the object is what we expect. To work correctly
3238 // for checkcast and aastore the type of superklass should be exact.
3239 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3240 // We may not have profiling here or it may not help us. If we have
3241 // a speculative type use it to perform an exact cast.
3242 ciKlass* spec_obj_type = obj_type->speculative_type();
3243 if (spec_obj_type != NULL || data != NULL) {
3244 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3245 if (cast_obj != NULL) {
3246 if (failure_control != NULL) // failure is now impossible
3247 (*failure_control) = top();
3248 // adjust the type of the phi to the exact klass:
3249 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3250 }
3251 }
3252 }
3253
3254 if (cast_obj == NULL) {
3255 // Generate the subtype check
3256 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass );
3257
3258 // Plug in success path into the merge
3259 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3260 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3261 if (failure_control == NULL) {
3262 if (not_subtype_ctrl != top()) { // If failure is possible
3263 PreserveJVMState pjvms(this);
3264 set_control(not_subtype_ctrl);
3265 builtin_throw(Deoptimization::Reason_class_check, load_object_klass(not_null_obj));
3266 }
3267 } else {
3268 (*failure_control) = not_subtype_ctrl;
3269 }
3270 }
3271
3272 region->init_req(_obj_path, control());
3273 phi ->init_req(_obj_path, cast_obj);
3274
3275 // A merge of NULL or Casted-NotNull obj
3276 Node* res = _gvn.transform(phi);
3277
3278 // Note I do NOT always 'replace_in_map(obj,result)' here.
3279 // if( tk->klass()->can_be_primary_super() )
3280 // This means that if I successfully store an Object into an array-of-String
3281 // I 'forget' that the Object is really now known to be a String. I have to
3282 // do this because we don't have true union types for interfaces - if I store
3283 // a Baz into an array-of-Interface and then tell the optimizer it's an
3284 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3285 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3286 // replace_in_map( obj, res );
3287
3288 // Return final merged results
3289 set_control( _gvn.transform(region) );
3290 record_for_igvn(region);
3291
3292 return record_profiled_receiver_for_speculation(res);
3293 }
3294
3295 //------------------------------next_monitor-----------------------------------
3296 // What number should be given to the next monitor?
3297 int GraphKit::next_monitor() {
3298 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3299 int next = current + C->sync_stack_slots();
3300 // Keep the toplevel high water mark current:
3301 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3302 return current;
3303 }
3304
3305 //------------------------------insert_mem_bar---------------------------------
3306 // Memory barrier to avoid floating things around
3307 // The membar serves as a pinch point between both control and all memory slices.
3308 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3309 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3310 mb->init_req(TypeFunc::Control, control());
3311 mb->init_req(TypeFunc::Memory, reset_memory());
3312 Node* membar = _gvn.transform(mb);
3313 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3314 set_all_memory_call(membar);
3315 return membar;
3316 }
3317
3318 //-------------------------insert_mem_bar_volatile----------------------------
3319 // Memory barrier to avoid floating things around
3320 // The membar serves as a pinch point between both control and memory(alias_idx).
3321 // If you want to make a pinch point on all memory slices, do not use this
3322 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3323 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3324 // When Parse::do_put_xxx updates a volatile field, it appends a series
3325 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3326 // The first membar is on the same memory slice as the field store opcode.
3327 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3328 // All the other membars (for other volatile slices, including AliasIdxBot,
3329 // which stands for all unknown volatile slices) are control-dependent
3330 // on the first membar. This prevents later volatile loads or stores
3331 // from sliding up past the just-emitted store.
3332
3333 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3334 mb->set_req(TypeFunc::Control,control());
3335 if (alias_idx == Compile::AliasIdxBot) {
3336 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3337 } else {
3338 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3339 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3340 }
3341 Node* membar = _gvn.transform(mb);
3342 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3343 if (alias_idx == Compile::AliasIdxBot) {
3344 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3345 } else {
3346 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3347 }
3348 return membar;
3349 }
3350
3351 //------------------------------shared_lock------------------------------------
3352 // Emit locking code.
3353 FastLockNode* GraphKit::shared_lock(Node* obj) {
3354 // bci is either a monitorenter bc or InvocationEntryBci
3355 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3356 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3357
3358 if( !GenerateSynchronizationCode )
3359 return NULL; // Not locking things?
3360 if (stopped()) // Dead monitor?
3361 return NULL;
3362
3363 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3364
3365 // Box the stack location
3366 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3367 Node* mem = reset_memory();
3368
3369 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3370 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3371 // Create the counters for this fast lock.
3372 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3373 }
3374
3375 // Create the rtm counters for this fast lock if needed.
3376 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3377
3378 // Add monitor to debug info for the slow path. If we block inside the
3379 // slow path and de-opt, we need the monitor hanging around
3380 map()->push_monitor( flock );
3381
3382 const TypeFunc *tf = LockNode::lock_type();
3383 LockNode *lock = new LockNode(C, tf);
3384
3385 lock->init_req( TypeFunc::Control, control() );
3386 lock->init_req( TypeFunc::Memory , mem );
3387 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3388 lock->init_req( TypeFunc::FramePtr, frameptr() );
3389 lock->init_req( TypeFunc::ReturnAdr, top() );
3390
3391 lock->init_req(TypeFunc::Parms + 0, obj);
3392 lock->init_req(TypeFunc::Parms + 1, box);
3393 lock->init_req(TypeFunc::Parms + 2, flock);
3394 add_safepoint_edges(lock);
3395
3396 lock = _gvn.transform( lock )->as_Lock();
3397
3398 // lock has no side-effects, sets few values
3399 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3400
3401 insert_mem_bar(Op_MemBarAcquireLock);
3402
3403 // Add this to the worklist so that the lock can be eliminated
3404 record_for_igvn(lock);
3405
3406 #ifndef PRODUCT
3407 if (PrintLockStatistics) {
3408 // Update the counter for this lock. Don't bother using an atomic
3409 // operation since we don't require absolute accuracy.
3410 lock->create_lock_counter(map()->jvms());
3411 increment_counter(lock->counter()->addr());
3412 }
3413 #endif
3414
3415 return flock;
3416 }
3417
3418
3419 //------------------------------shared_unlock----------------------------------
3420 // Emit unlocking code.
3421 void GraphKit::shared_unlock(Node* box, Node* obj) {
3422 // bci is either a monitorenter bc or InvocationEntryBci
3423 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3424 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3425
3426 if( !GenerateSynchronizationCode )
3427 return;
3428 if (stopped()) { // Dead monitor?
3429 map()->pop_monitor(); // Kill monitor from debug info
3430 return;
3431 }
3432
3433 // Memory barrier to avoid floating things down past the locked region
3434 insert_mem_bar(Op_MemBarReleaseLock);
3435
3436 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3437 UnlockNode *unlock = new UnlockNode(C, tf);
3438 #ifdef ASSERT
3439 unlock->set_dbg_jvms(sync_jvms());
3440 #endif
3441 uint raw_idx = Compile::AliasIdxRaw;
3442 unlock->init_req( TypeFunc::Control, control() );
3443 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3444 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3445 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3446 unlock->init_req( TypeFunc::ReturnAdr, top() );
3447
3448 unlock->init_req(TypeFunc::Parms + 0, obj);
3449 unlock->init_req(TypeFunc::Parms + 1, box);
3450 unlock = _gvn.transform(unlock)->as_Unlock();
3451
3452 Node* mem = reset_memory();
3453
3454 // unlock has no side-effects, sets few values
3455 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3456
3457 // Kill monitor from debug info
3458 map()->pop_monitor( );
3459 }
3460
3461 //-------------------------------get_layout_helper-----------------------------
3462 // If the given klass is a constant or known to be an array,
3463 // fetch the constant layout helper value into constant_value
3464 // and return (Node*)NULL. Otherwise, load the non-constant
3465 // layout helper value, and return the node which represents it.
3466 // This two-faced routine is useful because allocation sites
3467 // almost always feature constant types.
3468 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3469 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3470 if (!StressReflectiveCode && inst_klass != NULL) {
3471 ciKlass* klass = inst_klass->klass();
3472 bool xklass = inst_klass->klass_is_exact();
3473 if (xklass || klass->is_array_klass()) {
3474 jint lhelper = klass->layout_helper();
3475 if (lhelper != Klass::_lh_neutral_value) {
3476 constant_value = lhelper;
3477 return (Node*) NULL;
3478 }
3479 }
3480 }
3481 constant_value = Klass::_lh_neutral_value; // put in a known value
3482 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3483 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3484 }
3485
3486 // We just put in an allocate/initialize with a big raw-memory effect.
3487 // Hook selected additional alias categories on the initialization.
3488 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3489 MergeMemNode* init_in_merge,
3490 Node* init_out_raw) {
3491 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3492 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3493
3494 Node* prevmem = kit.memory(alias_idx);
3495 init_in_merge->set_memory_at(alias_idx, prevmem);
3496 kit.set_memory(init_out_raw, alias_idx);
3497 }
3498
3499 //---------------------------set_output_for_allocation-------------------------
3500 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3501 const TypeOopPtr* oop_type,
3502 bool deoptimize_on_exception) {
3503 int rawidx = Compile::AliasIdxRaw;
3504 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3505 add_safepoint_edges(alloc);
3506 Node* allocx = _gvn.transform(alloc);
3507 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3508 // create memory projection for i_o
3509 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3510 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3511
3512 // create a memory projection as for the normal control path
3513 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3514 set_memory(malloc, rawidx);
3515
3516 // a normal slow-call doesn't change i_o, but an allocation does
3517 // we create a separate i_o projection for the normal control path
3518 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3519 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3520
3521 // put in an initialization barrier
3522 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3523 rawoop)->as_Initialize();
3524 assert(alloc->initialization() == init, "2-way macro link must work");
3525 assert(init ->allocation() == alloc, "2-way macro link must work");
3526 {
3527 // Extract memory strands which may participate in the new object's
3528 // initialization, and source them from the new InitializeNode.
3529 // This will allow us to observe initializations when they occur,
3530 // and link them properly (as a group) to the InitializeNode.
3531 assert(init->in(InitializeNode::Memory) == malloc, "");
3532 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3533 init->set_req(InitializeNode::Memory, minit_in);
3534 record_for_igvn(minit_in); // fold it up later, if possible
3535 Node* minit_out = memory(rawidx);
3536 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3537 // Add an edge in the MergeMem for the header fields so an access
3538 // to one of those has correct memory state
3539 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3540 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3541 if (oop_type->isa_aryptr()) {
3542 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3543 int elemidx = C->get_alias_index(telemref);
3544 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3545 } else if (oop_type->isa_instptr()) {
3546 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3547 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3548 ciField* field = ik->nonstatic_field_at(i);
3549 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3550 continue; // do not bother to track really large numbers of fields
3551 // Find (or create) the alias category for this field:
3552 int fieldidx = C->alias_type(field)->index();
3553 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3554 }
3555 }
3556 }
3557
3558 // Cast raw oop to the real thing...
3559 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3560 javaoop = _gvn.transform(javaoop);
3561 C->set_recent_alloc(control(), javaoop);
3562 assert(just_allocated_object(control()) == javaoop, "just allocated");
3563
3564 #ifdef ASSERT
3565 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3566 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3567 "Ideal_allocation works");
3568 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3569 "Ideal_allocation works");
3570 if (alloc->is_AllocateArray()) {
3571 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3572 "Ideal_allocation works");
3573 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3574 "Ideal_allocation works");
3575 } else {
3576 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3577 }
3578 }
3579 #endif //ASSERT
3580
3581 return javaoop;
3582 }
3583
3584 //---------------------------new_instance--------------------------------------
3585 // This routine takes a klass_node which may be constant (for a static type)
3586 // or may be non-constant (for reflective code). It will work equally well
3587 // for either, and the graph will fold nicely if the optimizer later reduces
3588 // the type to a constant.
3589 // The optional arguments are for specialized use by intrinsics:
3590 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3591 // - If 'return_size_val', report the the total object size to the caller.
3592 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3593 Node* GraphKit::new_instance(Node* klass_node,
3594 Node* extra_slow_test,
3595 Node* *return_size_val,
3596 bool deoptimize_on_exception) {
3597 // Compute size in doublewords
3598 // The size is always an integral number of doublewords, represented
3599 // as a positive bytewise size stored in the klass's layout_helper.
3600 // The layout_helper also encodes (in a low bit) the need for a slow path.
3601 jint layout_con = Klass::_lh_neutral_value;
3602 Node* layout_val = get_layout_helper(klass_node, layout_con);
3603 int layout_is_con = (layout_val == NULL);
3604
3605 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3606 // Generate the initial go-slow test. It's either ALWAYS (return a
3607 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3608 // case) a computed value derived from the layout_helper.
3609 Node* initial_slow_test = NULL;
3610 if (layout_is_con) {
3611 assert(!StressReflectiveCode, "stress mode does not use these paths");
3612 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3613 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3614 } else { // reflective case
3615 // This reflective path is used by Unsafe.allocateInstance.
3616 // (It may be stress-tested by specifying StressReflectiveCode.)
3617 // Basically, we want to get into the VM is there's an illegal argument.
3618 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3619 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3620 if (extra_slow_test != intcon(0)) {
3621 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3622 }
3623 // (Macro-expander will further convert this to a Bool, if necessary.)
3624 }
3625
3626 // Find the size in bytes. This is easy; it's the layout_helper.
3627 // The size value must be valid even if the slow path is taken.
3628 Node* size = NULL;
3629 if (layout_is_con) {
3630 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3631 } else { // reflective case
3632 // This reflective path is used by clone and Unsafe.allocateInstance.
3633 size = ConvI2X(layout_val);
3634
3635 // Clear the low bits to extract layout_helper_size_in_bytes:
3636 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3637 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3638 size = _gvn.transform( new AndXNode(size, mask) );
3639 }
3640 if (return_size_val != NULL) {
3641 (*return_size_val) = size;
3642 }
3643
3644 // This is a precise notnull oop of the klass.
3645 // (Actually, it need not be precise if this is a reflective allocation.)
3646 // It's what we cast the result to.
3647 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3648 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3649 const TypeOopPtr* oop_type = tklass->as_instance_type();
3650
3651 // Now generate allocation code
3652
3653 // The entire memory state is needed for slow path of the allocation
3654 // since GC and deoptimization can happened.
3655 Node *mem = reset_memory();
3656 set_all_memory(mem); // Create new memory state
3657
3658 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3659 control(), mem, i_o(),
3660 size, klass_node,
3661 initial_slow_test);
3662
3663 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3664 }
3665
3666 //-------------------------------new_array-------------------------------------
3667 // helper for both newarray and anewarray
3668 // The 'length' parameter is (obviously) the length of the array.
3669 // See comments on new_instance for the meaning of the other arguments.
3670 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3671 Node* length, // number of array elements
3672 int nargs, // number of arguments to push back for uncommon trap
3673 Node* *return_size_val,
3674 bool deoptimize_on_exception) {
3675 jint layout_con = Klass::_lh_neutral_value;
3676 Node* layout_val = get_layout_helper(klass_node, layout_con);
3677 int layout_is_con = (layout_val == NULL);
3678
3679 if (!layout_is_con && !StressReflectiveCode &&
3680 !too_many_traps(Deoptimization::Reason_class_check)) {
3681 // This is a reflective array creation site.
3682 // Optimistically assume that it is a subtype of Object[],
3683 // so that we can fold up all the address arithmetic.
3684 layout_con = Klass::array_layout_helper(T_OBJECT);
3685 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3686 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3687 { BuildCutout unless(this, bol_lh, PROB_MAX);
3688 inc_sp(nargs);
3689 uncommon_trap(Deoptimization::Reason_class_check,
3690 Deoptimization::Action_maybe_recompile);
3691 }
3692 layout_val = NULL;
3693 layout_is_con = true;
3694 }
3695
3696 // Generate the initial go-slow test. Make sure we do not overflow
3697 // if length is huge (near 2Gig) or negative! We do not need
3698 // exact double-words here, just a close approximation of needed
3699 // double-words. We can't add any offset or rounding bits, lest we
3700 // take a size -1 of bytes and make it positive. Use an unsigned
3701 // compare, so negative sizes look hugely positive.
3702 int fast_size_limit = FastAllocateSizeLimit;
3703 if (layout_is_con) {
3704 assert(!StressReflectiveCode, "stress mode does not use these paths");
3705 // Increase the size limit if we have exact knowledge of array type.
3706 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3707 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3708 }
3709
3710 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3711 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3712
3713 // --- Size Computation ---
3714 // array_size = round_to_heap(array_header + (length << elem_shift));
3715 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3716 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3717 // The rounding mask is strength-reduced, if possible.
3718 int round_mask = MinObjAlignmentInBytes - 1;
3719 Node* header_size = NULL;
3720 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3721 // (T_BYTE has the weakest alignment and size restrictions...)
3722 if (layout_is_con) {
3723 int hsize = Klass::layout_helper_header_size(layout_con);
3724 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3725 BasicType etype = Klass::layout_helper_element_type(layout_con);
3726 if ((round_mask & ~right_n_bits(eshift)) == 0)
3727 round_mask = 0; // strength-reduce it if it goes away completely
3728 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3729 assert(header_size_min <= hsize, "generic minimum is smallest");
3730 header_size_min = hsize;
3731 header_size = intcon(hsize + round_mask);
3732 } else {
3733 Node* hss = intcon(Klass::_lh_header_size_shift);
3734 Node* hsm = intcon(Klass::_lh_header_size_mask);
3735 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3736 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3737 Node* mask = intcon(round_mask);
3738 header_size = _gvn.transform( new AddINode(hsize, mask) );
3739 }
3740
3741 Node* elem_shift = NULL;
3742 if (layout_is_con) {
3743 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3744 if (eshift != 0)
3745 elem_shift = intcon(eshift);
3746 } else {
3747 // There is no need to mask or shift this value.
3748 // The semantics of LShiftINode include an implicit mask to 0x1F.
3749 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3750 elem_shift = layout_val;
3751 }
3752
3753 // Transition to native address size for all offset calculations:
3754 Node* lengthx = ConvI2X(length);
3755 Node* headerx = ConvI2X(header_size);
3756 #ifdef _LP64
3757 { const TypeInt* tilen = _gvn.find_int_type(length);
3758 if (tilen != NULL && tilen->_lo < 0) {
3759 // Add a manual constraint to a positive range. Cf. array_element_address.
3760 jint size_max = fast_size_limit;
3761 if (size_max > tilen->_hi) size_max = tilen->_hi;
3762 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3763
3764 // Only do a narrow I2L conversion if the range check passed.
3765 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3766 _gvn.transform(iff);
3767 RegionNode* region = new RegionNode(3);
3768 _gvn.set_type(region, Type::CONTROL);
3769 lengthx = new PhiNode(region, TypeLong::LONG);
3770 _gvn.set_type(lengthx, TypeLong::LONG);
3771
3772 // Range check passed. Use ConvI2L node with narrow type.
3773 Node* passed = IfFalse(iff);
3774 region->init_req(1, passed);
3775 // Make I2L conversion control dependent to prevent it from
3776 // floating above the range check during loop optimizations.
3777 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3778
3779 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3780 region->init_req(2, IfTrue(iff));
3781 lengthx->init_req(2, ConvI2X(length));
3782
3783 set_control(region);
3784 record_for_igvn(region);
3785 record_for_igvn(lengthx);
3786 }
3787 }
3788 #endif
3789
3790 // Combine header size (plus rounding) and body size. Then round down.
3791 // This computation cannot overflow, because it is used only in two
3792 // places, one where the length is sharply limited, and the other
3793 // after a successful allocation.
3794 Node* abody = lengthx;
3795 if (elem_shift != NULL)
3796 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3797 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3798 if (round_mask != 0) {
3799 Node* mask = MakeConX(~round_mask);
3800 size = _gvn.transform( new AndXNode(size, mask) );
3801 }
3802 // else if round_mask == 0, the size computation is self-rounding
3803
3804 if (return_size_val != NULL) {
3805 // This is the size
3806 (*return_size_val) = size;
3807 }
3808
3809 // Now generate allocation code
3810
3811 // The entire memory state is needed for slow path of the allocation
3812 // since GC and deoptimization can happened.
3813 Node *mem = reset_memory();
3814 set_all_memory(mem); // Create new memory state
3815
3816 if (initial_slow_test->is_Bool()) {
3817 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3818 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3819 }
3820
3821 // Create the AllocateArrayNode and its result projections
3822 AllocateArrayNode* alloc
3823 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3824 control(), mem, i_o(),
3825 size, klass_node,
3826 initial_slow_test,
3827 length);
3828
3829 // Cast to correct type. Note that the klass_node may be constant or not,
3830 // and in the latter case the actual array type will be inexact also.
3831 // (This happens via a non-constant argument to inline_native_newArray.)
3832 // In any case, the value of klass_node provides the desired array type.
3833 const TypeInt* length_type = _gvn.find_int_type(length);
3834 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3835 if (ary_type->isa_aryptr() && length_type != NULL) {
3836 // Try to get a better type than POS for the size
3837 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3838 }
3839
3840 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3841
3842 // Cast length on remaining path to be as narrow as possible
3843 if (map()->find_edge(length) >= 0) {
3844 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3845 if (ccast != length) {
3846 _gvn.set_type_bottom(ccast);
3847 record_for_igvn(ccast);
3848 replace_in_map(length, ccast);
3849 }
3850 }
3851
3852 return javaoop;
3853 }
3854
3855 // The following "Ideal_foo" functions are placed here because they recognize
3856 // the graph shapes created by the functions immediately above.
3857
3858 //---------------------------Ideal_allocation----------------------------------
3859 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3860 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3861 if (ptr == NULL) { // reduce dumb test in callers
3862 return NULL;
3863 }
3864
3865 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
3866 ptr = bs->step_over_gc_barrier(ptr);
3867
3868 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3869 ptr = ptr->in(1);
3870 if (ptr == NULL) return NULL;
3871 }
3872 // Return NULL for allocations with several casts:
3873 // j.l.reflect.Array.newInstance(jobject, jint)
3874 // Object.clone()
3875 // to keep more precise type from last cast.
3876 if (ptr->is_Proj()) {
3877 Node* allo = ptr->in(0);
3878 if (allo != NULL && allo->is_Allocate()) {
3879 return allo->as_Allocate();
3880 }
3881 }
3882 // Report failure to match.
3883 return NULL;
3884 }
3885
3886 // Fancy version which also strips off an offset (and reports it to caller).
3887 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3888 intptr_t& offset) {
3889 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3890 if (base == NULL) return NULL;
3891 return Ideal_allocation(base, phase);
3892 }
3893
3894 // Trace Initialize <- Proj[Parm] <- Allocate
3895 AllocateNode* InitializeNode::allocation() {
3896 Node* rawoop = in(InitializeNode::RawAddress);
3897 if (rawoop->is_Proj()) {
3898 Node* alloc = rawoop->in(0);
3899 if (alloc->is_Allocate()) {
3900 return alloc->as_Allocate();
3901 }
3902 }
3903 return NULL;
3904 }
3905
3906 // Trace Allocate -> Proj[Parm] -> Initialize
3907 InitializeNode* AllocateNode::initialization() {
3908 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
3909 if (rawoop == NULL) return NULL;
3910 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3911 Node* init = rawoop->fast_out(i);
3912 if (init->is_Initialize()) {
3913 assert(init->as_Initialize()->allocation() == this, "2-way link");
3914 return init->as_Initialize();
3915 }
3916 }
3917 return NULL;
3918 }
3919
3920 //----------------------------- loop predicates ---------------------------
3921
3922 //------------------------------add_predicate_impl----------------------------
3923 void GraphKit::add_empty_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3924 // Too many traps seen?
3925 if (too_many_traps(reason)) {
3926 #ifdef ASSERT
3927 if (TraceLoopPredicate) {
3928 int tc = C->trap_count(reason);
3929 tty->print("too many traps=%s tcount=%d in ",
3930 Deoptimization::trap_reason_name(reason), tc);
3931 method()->print(); // which method has too many predicate traps
3932 tty->cr();
3933 }
3934 #endif
3935 // We cannot afford to take more traps here,
3936 // do not generate predicate.
3937 return;
3938 }
3939
3940 Node *cont = _gvn.intcon(1);
3941 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
3942 Node *bol = _gvn.transform(new Conv2BNode(opq));
3943 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3944 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3945 C->add_predicate_opaq(opq);
3946 {
3947 PreserveJVMState pjvms(this);
3948 set_control(iffalse);
3949 inc_sp(nargs);
3950 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3951 }
3952 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3953 set_control(iftrue);
3954 }
3955
3956 //------------------------------add_predicate---------------------------------
3957 void GraphKit::add_empty_predicates(int nargs) {
3958 // These loop predicates remain empty. All concrete loop predicates are inserted above the corresponding
3959 // empty loop predicate later by 'PhaseIdealLoop::create_new_if_for_predicate'. All concrete loop predicates of
3960 // a specific kind (normal, profile or limit check) share the same uncommon trap as the empty loop predicate.
3961 if (UseLoopPredicate) {
3962 add_empty_predicate_impl(Deoptimization::Reason_predicate, nargs);
3963 }
3964 if (UseProfiledLoopPredicate) {
3965 add_empty_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
3966 }
3967 // loop's limit check predicate should be near the loop.
3968 add_empty_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3969 }
3970
3971 void GraphKit::sync_kit(IdealKit& ideal) {
3972 set_all_memory(ideal.merged_memory());
3973 set_i_o(ideal.i_o());
3974 set_control(ideal.ctrl());
3975 }
3976
3977 void GraphKit::final_sync(IdealKit& ideal) {
3978 // Final sync IdealKit and graphKit.
3979 sync_kit(ideal);
3980 }
3981
3982 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
3983 Node* len = load_array_length(load_String_value(str, set_ctrl));
3984 Node* coder = load_String_coder(str, set_ctrl);
3985 // Divide length by 2 if coder is UTF16
3986 return _gvn.transform(new RShiftINode(len, coder));
3987 }
3988
3989 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
3990 int value_offset = java_lang_String::value_offset();
3991 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3992 false, NULL, 0);
3993 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3994 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3995 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
3996 ciTypeArrayKlass::make(T_BYTE), true, 0);
3997 Node* p = basic_plus_adr(str, str, value_offset);
3998 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
3999 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4000 return load;
4001 }
4002
4003 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4004 if (!CompactStrings) {
4005 return intcon(java_lang_String::CODER_UTF16);
4006 }
4007 int coder_offset = java_lang_String::coder_offset();
4008 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4009 false, NULL, 0);
4010 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4011
4012 Node* p = basic_plus_adr(str, str, coder_offset);
4013 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4014 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4015 return load;
4016 }
4017
4018 void GraphKit::store_String_value(Node* str, Node* value) {
4019 int value_offset = java_lang_String::value_offset();
4020 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4021 false, NULL, 0);
4022 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4023
4024 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4025 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4026 }
4027
4028 void GraphKit::store_String_coder(Node* str, Node* value) {
4029 int coder_offset = java_lang_String::coder_offset();
4030 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4031 false, NULL, 0);
4032 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4033
4034 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4035 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4036 }
4037
4038 // Capture src and dst memory state with a MergeMemNode
4039 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4040 if (src_type == dst_type) {
4041 // Types are equal, we don't need a MergeMemNode
4042 return memory(src_type);
4043 }
4044 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4045 record_for_igvn(merge); // fold it up later, if possible
4046 int src_idx = C->get_alias_index(src_type);
4047 int dst_idx = C->get_alias_index(dst_type);
4048 merge->set_memory_at(src_idx, memory(src_idx));
4049 merge->set_memory_at(dst_idx, memory(dst_idx));
4050 return merge;
4051 }
4052
4053 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4054 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4055 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4056 // If input and output memory types differ, capture both states to preserve
4057 // the dependency between preceding and subsequent loads/stores.
4058 // For example, the following program:
4059 // StoreB
4060 // compress_string
4061 // LoadB
4062 // has this memory graph (use->def):
4063 // LoadB -> compress_string -> CharMem
4064 // ... -> StoreB -> ByteMem
4065 // The intrinsic hides the dependency between LoadB and StoreB, causing
4066 // the load to read from memory not containing the result of the StoreB.
4067 // The correct memory graph should look like this:
4068 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4069 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4070 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4071 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4072 set_memory(res_mem, TypeAryPtr::BYTES);
4073 return str;
4074 }
4075
4076 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4077 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4078 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4079 // Capture src and dst memory (see comment in 'compress_string').
4080 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4081 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4082 set_memory(_gvn.transform(str), dst_type);
4083 }
4084
4085 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4086 /**
4087 * int i_char = start;
4088 * for (int i_byte = 0; i_byte < count; i_byte++) {
4089 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4090 * }
4091 */
4092 add_empty_predicates();
4093 RegionNode* head = new RegionNode(3);
4094 head->init_req(1, control());
4095 gvn().set_type(head, Type::CONTROL);
4096 record_for_igvn(head);
4097
4098 Node* i_byte = new PhiNode(head, TypeInt::INT);
4099 i_byte->init_req(1, intcon(0));
4100 gvn().set_type(i_byte, TypeInt::INT);
4101 record_for_igvn(i_byte);
4102
4103 Node* i_char = new PhiNode(head, TypeInt::INT);
4104 i_char->init_req(1, start);
4105 gvn().set_type(i_char, TypeInt::INT);
4106 record_for_igvn(i_char);
4107
4108 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4109 gvn().set_type(mem, Type::MEMORY);
4110 record_for_igvn(mem);
4111 set_control(head);
4112 set_memory(mem, TypeAryPtr::BYTES);
4113 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4114 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4115 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4116 false, false, true /* mismatched */);
4117
4118 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4119 head->init_req(2, IfTrue(iff));
4120 mem->init_req(2, st);
4121 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4122 i_char->init_req(2, AddI(i_char, intcon(2)));
4123
4124 set_control(IfFalse(iff));
4125 set_memory(st, TypeAryPtr::BYTES);
4126 }
4127
4128 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4129 if (!field->is_constant()) {
4130 return NULL; // Field not marked as constant.
4131 }
4132 ciInstance* holder = NULL;
4133 if (!field->is_static()) {
4134 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4135 if (const_oop != NULL && const_oop->is_instance()) {
4136 holder = const_oop->as_instance();
4137 }
4138 }
4139 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4140 /*is_unsigned_load=*/false);
4141 if (con_type != NULL) {
4142 return makecon(con_type);
4143 }
4144 return NULL;
4145 }