1 /*
2 * Copyright (c) 2001, 2021, 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/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciUtilities.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "ci/ciNativeEntryPoint.hpp"
31 #include "ci/ciObjArray.hpp"
32 #include "asm/register.hpp"
33 #include "compiler/compileLog.hpp"
34 #include "gc/shared/barrierSet.hpp"
35 #include "gc/shared/c2/barrierSetC2.hpp"
36 #include "interpreter/interpreter.hpp"
37 #include "memory/resourceArea.hpp"
38 #include "opto/addnode.hpp"
39 #include "opto/castnode.hpp"
40 #include "opto/convertnode.hpp"
41 #include "opto/graphKit.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/inlinetypenode.hpp"
44 #include "opto/intrinsicnode.hpp"
45 #include "opto/locknode.hpp"
46 #include "opto/machnode.hpp"
47 #include "opto/narrowptrnode.hpp"
48 #include "opto/opaquenode.hpp"
49 #include "opto/parse.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/runtime.hpp"
52 #include "opto/subtypenode.hpp"
53 #include "runtime/deoptimization.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #include "utilities/bitMap.inline.hpp"
56 #include "utilities/powerOfTwo.hpp"
57 #include "utilities/growableArray.hpp"
58
59 //----------------------------GraphKit-----------------------------------------
60 // Main utility constructor.
61 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
62 : Phase(Phase::Parser),
63 _env(C->env()),
64 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()),
65 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
66 {
67 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
68 _exceptions = jvms->map()->next_exception();
69 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
70 set_jvms(jvms);
71 #ifdef ASSERT
72 if (_gvn.is_IterGVN() != NULL) {
73 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
74 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
75 _worklist_size = _gvn.C->for_igvn()->size();
76 }
77 #endif
78 }
79
80 // Private constructor for parser.
81 GraphKit::GraphKit()
82 : Phase(Phase::Parser),
83 _env(C->env()),
84 _gvn(*C->initial_gvn()),
85 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
86 {
87 _exceptions = NULL;
88 set_map(NULL);
89 debug_only(_sp = -99);
90 debug_only(set_bci(-99));
91 }
92
93
94
95 //---------------------------clean_stack---------------------------------------
96 // Clear away rubbish from the stack area of the JVM state.
97 // This destroys any arguments that may be waiting on the stack.
98 void GraphKit::clean_stack(int from_sp) {
99 SafePointNode* map = this->map();
100 JVMState* jvms = this->jvms();
101 int stk_size = jvms->stk_size();
102 int stkoff = jvms->stkoff();
103 Node* top = this->top();
104 for (int i = from_sp; i < stk_size; i++) {
105 if (map->in(stkoff + i) != top) {
106 map->set_req(stkoff + i, top);
107 }
108 }
109 }
110
111
112 //--------------------------------sync_jvms-----------------------------------
113 // Make sure our current jvms agrees with our parse state.
114 JVMState* GraphKit::sync_jvms() const {
115 JVMState* jvms = this->jvms();
116 jvms->set_bci(bci()); // Record the new bci in the JVMState
117 jvms->set_sp(sp()); // Record the new sp in the JVMState
118 assert(jvms_in_sync(), "jvms is now in sync");
119 return jvms;
120 }
121
122 //--------------------------------sync_jvms_for_reexecute---------------------
123 // Make sure our current jvms agrees with our parse state. This version
124 // uses the reexecute_sp for reexecuting bytecodes.
125 JVMState* GraphKit::sync_jvms_for_reexecute() {
126 JVMState* jvms = this->jvms();
127 jvms->set_bci(bci()); // Record the new bci in the JVMState
128 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
129 return jvms;
130 }
131
132 #ifdef ASSERT
133 bool GraphKit::jvms_in_sync() const {
134 Parse* parse = is_Parse();
135 if (parse == NULL) {
136 if (bci() != jvms()->bci()) return false;
137 if (sp() != (int)jvms()->sp()) return false;
138 return true;
139 }
140 if (jvms()->method() != parse->method()) return false;
141 if (jvms()->bci() != parse->bci()) return false;
142 int jvms_sp = jvms()->sp();
143 if (jvms_sp != parse->sp()) return false;
144 int jvms_depth = jvms()->depth();
145 if (jvms_depth != parse->depth()) return false;
146 return true;
147 }
148
149 // Local helper checks for special internal merge points
150 // used to accumulate and merge exception states.
151 // They are marked by the region's in(0) edge being the map itself.
152 // Such merge points must never "escape" into the parser at large,
153 // until they have been handed to gvn.transform.
154 static bool is_hidden_merge(Node* reg) {
155 if (reg == NULL) return false;
156 if (reg->is_Phi()) {
157 reg = reg->in(0);
158 if (reg == NULL) return false;
159 }
160 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
161 }
162
163 void GraphKit::verify_map() const {
164 if (map() == NULL) return; // null map is OK
165 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
166 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
167 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
168 }
169
170 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
171 assert(ex_map->next_exception() == NULL, "not already part of a chain");
172 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
173 }
174 #endif
175
176 //---------------------------stop_and_kill_map---------------------------------
177 // Set _map to NULL, signalling a stop to further bytecode execution.
178 // First smash the current map's control to a constant, to mark it dead.
179 void GraphKit::stop_and_kill_map() {
180 SafePointNode* dead_map = stop();
181 if (dead_map != NULL) {
182 dead_map->disconnect_inputs(C); // Mark the map as killed.
183 assert(dead_map->is_killed(), "must be so marked");
184 }
185 }
186
187
188 //--------------------------------stopped--------------------------------------
189 // Tell if _map is NULL, or control is top.
190 bool GraphKit::stopped() {
191 if (map() == NULL) return true;
192 else if (control() == top()) return true;
193 else return false;
194 }
195
196
197 //-----------------------------has_ex_handler----------------------------------
198 // Tell if this method or any caller method has exception handlers.
199 bool GraphKit::has_ex_handler() {
200 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
201 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
202 return true;
203 }
204 }
205 return false;
206 }
207
208 //------------------------------save_ex_oop------------------------------------
209 // Save an exception without blowing stack contents or other JVM state.
210 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
211 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
212 ex_map->add_req(ex_oop);
213 debug_only(verify_exception_state(ex_map));
214 }
215
216 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
217 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
218 Node* ex_oop = ex_map->in(ex_map->req()-1);
219 if (clear_it) ex_map->del_req(ex_map->req()-1);
220 return ex_oop;
221 }
222
223 //-----------------------------saved_ex_oop------------------------------------
224 // Recover a saved exception from its map.
225 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
226 return common_saved_ex_oop(ex_map, false);
227 }
228
229 //--------------------------clear_saved_ex_oop---------------------------------
230 // Erase a previously saved exception from its map.
231 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
232 return common_saved_ex_oop(ex_map, true);
233 }
234
235 #ifdef ASSERT
236 //---------------------------has_saved_ex_oop----------------------------------
237 // Erase a previously saved exception from its map.
238 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
239 return ex_map->req() == ex_map->jvms()->endoff()+1;
240 }
241 #endif
242
243 //-------------------------make_exception_state--------------------------------
244 // Turn the current JVM state into an exception state, appending the ex_oop.
245 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
246 sync_jvms();
247 SafePointNode* ex_map = stop(); // do not manipulate this map any more
248 set_saved_ex_oop(ex_map, ex_oop);
249 return ex_map;
250 }
251
252
253 //--------------------------add_exception_state--------------------------------
254 // Add an exception to my list of exceptions.
255 void GraphKit::add_exception_state(SafePointNode* ex_map) {
256 if (ex_map == NULL || ex_map->control() == top()) {
257 return;
258 }
259 #ifdef ASSERT
260 verify_exception_state(ex_map);
261 if (has_exceptions()) {
262 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
263 }
264 #endif
265
266 // If there is already an exception of exactly this type, merge with it.
267 // In particular, null-checks and other low-level exceptions common up here.
268 Node* ex_oop = saved_ex_oop(ex_map);
269 const Type* ex_type = _gvn.type(ex_oop);
270 if (ex_oop == top()) {
271 // No action needed.
272 return;
273 }
274 assert(ex_type->isa_instptr(), "exception must be an instance");
275 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
276 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
277 // We check sp also because call bytecodes can generate exceptions
278 // both before and after arguments are popped!
279 if (ex_type2 == ex_type
280 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
281 combine_exception_states(ex_map, e2);
282 return;
283 }
284 }
285
286 // No pre-existing exception of the same type. Chain it on the list.
287 push_exception_state(ex_map);
288 }
289
290 //-----------------------add_exception_states_from-----------------------------
291 void GraphKit::add_exception_states_from(JVMState* jvms) {
292 SafePointNode* ex_map = jvms->map()->next_exception();
293 if (ex_map != NULL) {
294 jvms->map()->set_next_exception(NULL);
295 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
296 next_map = ex_map->next_exception();
297 ex_map->set_next_exception(NULL);
298 add_exception_state(ex_map);
299 }
300 }
301 }
302
303 //-----------------------transfer_exceptions_into_jvms-------------------------
304 JVMState* GraphKit::transfer_exceptions_into_jvms() {
305 if (map() == NULL) {
306 // We need a JVMS to carry the exceptions, but the map has gone away.
307 // Create a scratch JVMS, cloned from any of the exception states...
308 if (has_exceptions()) {
309 _map = _exceptions;
310 _map = clone_map();
311 _map->set_next_exception(NULL);
312 clear_saved_ex_oop(_map);
313 debug_only(verify_map());
314 } else {
315 // ...or created from scratch
316 JVMState* jvms = new (C) JVMState(_method, NULL);
317 jvms->set_bci(_bci);
318 jvms->set_sp(_sp);
319 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
320 set_jvms(jvms);
321 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
322 set_all_memory(top());
323 while (map()->req() < jvms->endoff()) map()->add_req(top());
324 }
325 // (This is a kludge, in case you didn't notice.)
326 set_control(top());
327 }
328 JVMState* jvms = sync_jvms();
329 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
330 jvms->map()->set_next_exception(_exceptions);
331 _exceptions = NULL; // done with this set of exceptions
332 return jvms;
333 }
334
335 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
336 assert(is_hidden_merge(dstphi), "must be a special merge node");
337 assert(is_hidden_merge(srcphi), "must be a special merge node");
338 uint limit = srcphi->req();
339 for (uint i = PhiNode::Input; i < limit; i++) {
340 dstphi->add_req(srcphi->in(i));
341 }
342 }
343 static inline void add_one_req(Node* dstphi, Node* src) {
344 assert(is_hidden_merge(dstphi), "must be a special merge node");
345 assert(!is_hidden_merge(src), "must not be a special merge node");
346 dstphi->add_req(src);
347 }
348
349 //-----------------------combine_exception_states------------------------------
350 // This helper function combines exception states by building phis on a
351 // specially marked state-merging region. These regions and phis are
352 // untransformed, and can build up gradually. The region is marked by
353 // having a control input of its exception map, rather than NULL. Such
354 // regions do not appear except in this function, and in use_exception_state.
355 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
356 if (failing()) return; // dying anyway...
357 JVMState* ex_jvms = ex_map->_jvms;
358 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
359 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
360 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
361 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
362 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
363 assert(ex_map->req() == phi_map->req(), "matching maps");
364 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
365 Node* hidden_merge_mark = root();
366 Node* region = phi_map->control();
367 MergeMemNode* phi_mem = phi_map->merged_memory();
368 MergeMemNode* ex_mem = ex_map->merged_memory();
369 if (region->in(0) != hidden_merge_mark) {
370 // The control input is not (yet) a specially-marked region in phi_map.
371 // Make it so, and build some phis.
372 region = new RegionNode(2);
373 _gvn.set_type(region, Type::CONTROL);
374 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
375 region->init_req(1, phi_map->control());
376 phi_map->set_control(region);
377 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
378 record_for_igvn(io_phi);
379 _gvn.set_type(io_phi, Type::ABIO);
380 phi_map->set_i_o(io_phi);
381 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
382 Node* m = mms.memory();
383 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
384 record_for_igvn(m_phi);
385 _gvn.set_type(m_phi, Type::MEMORY);
386 mms.set_memory(m_phi);
387 }
388 }
389
390 // Either or both of phi_map and ex_map might already be converted into phis.
391 Node* ex_control = ex_map->control();
392 // if there is special marking on ex_map also, we add multiple edges from src
393 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
394 // how wide was the destination phi_map, originally?
395 uint orig_width = region->req();
396
397 if (add_multiple) {
398 add_n_reqs(region, ex_control);
399 add_n_reqs(phi_map->i_o(), ex_map->i_o());
400 } else {
401 // ex_map has no merges, so we just add single edges everywhere
402 add_one_req(region, ex_control);
403 add_one_req(phi_map->i_o(), ex_map->i_o());
404 }
405 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
406 if (mms.is_empty()) {
407 // get a copy of the base memory, and patch some inputs into it
408 const TypePtr* adr_type = mms.adr_type(C);
409 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
410 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
411 mms.set_memory(phi);
412 // Prepare to append interesting stuff onto the newly sliced phi:
413 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
414 }
415 // Append stuff from ex_map:
416 if (add_multiple) {
417 add_n_reqs(mms.memory(), mms.memory2());
418 } else {
419 add_one_req(mms.memory(), mms.memory2());
420 }
421 }
422 uint limit = ex_map->req();
423 for (uint i = TypeFunc::Parms; i < limit; i++) {
424 // Skip everything in the JVMS after tos. (The ex_oop follows.)
425 if (i == tos) i = ex_jvms->monoff();
426 Node* src = ex_map->in(i);
427 Node* dst = phi_map->in(i);
428 if (src != dst) {
429 PhiNode* phi;
430 if (dst->in(0) != region) {
431 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
432 record_for_igvn(phi);
433 _gvn.set_type(phi, phi->type());
434 phi_map->set_req(i, dst);
435 // Prepare to append interesting stuff onto the new phi:
436 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
437 } else {
438 assert(dst->is_Phi(), "nobody else uses a hidden region");
439 phi = dst->as_Phi();
440 }
441 if (add_multiple && src->in(0) == ex_control) {
442 // Both are phis.
443 add_n_reqs(dst, src);
444 } else {
445 while (dst->req() < region->req()) add_one_req(dst, src);
446 }
447 const Type* srctype = _gvn.type(src);
448 if (phi->type() != srctype) {
449 const Type* dsttype = phi->type()->meet_speculative(srctype);
450 if (phi->type() != dsttype) {
451 phi->set_type(dsttype);
452 _gvn.set_type(phi, dsttype);
453 }
454 }
455 }
456 }
457 phi_map->merge_replaced_nodes_with(ex_map);
458 }
459
460 //--------------------------use_exception_state--------------------------------
461 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
462 if (failing()) { stop(); return top(); }
463 Node* region = phi_map->control();
464 Node* hidden_merge_mark = root();
465 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
466 Node* ex_oop = clear_saved_ex_oop(phi_map);
467 if (region->in(0) == hidden_merge_mark) {
468 // Special marking for internal ex-states. Process the phis now.
469 region->set_req(0, region); // now it's an ordinary region
470 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
471 // Note: Setting the jvms also sets the bci and sp.
472 set_control(_gvn.transform(region));
473 uint tos = jvms()->stkoff() + sp();
474 for (uint i = 1; i < tos; i++) {
475 Node* x = phi_map->in(i);
476 if (x->in(0) == region) {
477 assert(x->is_Phi(), "expected a special phi");
478 phi_map->set_req(i, _gvn.transform(x));
479 }
480 }
481 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
482 Node* x = mms.memory();
483 if (x->in(0) == region) {
484 assert(x->is_Phi(), "nobody else uses a hidden region");
485 mms.set_memory(_gvn.transform(x));
486 }
487 }
488 if (ex_oop->in(0) == region) {
489 assert(ex_oop->is_Phi(), "expected a special phi");
490 ex_oop = _gvn.transform(ex_oop);
491 }
492 } else {
493 set_jvms(phi_map->jvms());
494 }
495
496 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
497 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
498 return ex_oop;
499 }
500
501 //---------------------------------java_bc-------------------------------------
502 Bytecodes::Code GraphKit::java_bc() const {
503 ciMethod* method = this->method();
504 int bci = this->bci();
505 if (method != NULL && bci != InvocationEntryBci)
506 return method->java_code_at_bci(bci);
507 else
508 return Bytecodes::_illegal;
509 }
510
511 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
512 bool must_throw) {
513 // if the exception capability is set, then we will generate code
514 // to check the JavaThread.should_post_on_exceptions flag to see
515 // if we actually need to report exception events (for this
516 // thread). If we don't need to report exception events, we will
517 // take the normal fast path provided by add_exception_events. If
518 // exception event reporting is enabled for this thread, we will
519 // take the uncommon_trap in the BuildCutout below.
520
521 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
522 Node* jthread = _gvn.transform(new ThreadLocalNode());
523 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
524 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
525
526 // Test the should_post_on_exceptions_flag vs. 0
527 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
528 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
529
530 // Branch to slow_path if should_post_on_exceptions_flag was true
531 { BuildCutout unless(this, tst, PROB_MAX);
532 // Do not try anything fancy if we're notifying the VM on every throw.
533 // Cf. case Bytecodes::_athrow in parse2.cpp.
534 uncommon_trap(reason, Deoptimization::Action_none,
535 (ciKlass*)NULL, (char*)NULL, must_throw);
536 }
537
538 }
539
540 //------------------------------builtin_throw----------------------------------
541 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
542 bool must_throw = true;
543
544 // If this particular condition has not yet happened at this
545 // bytecode, then use the uncommon trap mechanism, and allow for
546 // a future recompilation if several traps occur here.
547 // If the throw is hot, try to use a more complicated inline mechanism
548 // which keeps execution inside the compiled code.
549 bool treat_throw_as_hot = false;
550 ciMethodData* md = method()->method_data();
551
552 if (ProfileTraps) {
553 if (too_many_traps(reason)) {
554 treat_throw_as_hot = true;
555 }
556 // (If there is no MDO at all, assume it is early in
557 // execution, and that any deopts are part of the
558 // startup transient, and don't need to be remembered.)
559
560 // Also, if there is a local exception handler, treat all throws
561 // as hot if there has been at least one in this method.
562 if (C->trap_count(reason) != 0
563 && method()->method_data()->trap_count(reason) != 0
564 && has_ex_handler()) {
565 treat_throw_as_hot = true;
566 }
567 }
568
569 // If this throw happens frequently, an uncommon trap might cause
570 // a performance pothole. If there is a local exception handler,
571 // and if this particular bytecode appears to be deoptimizing often,
572 // let us handle the throw inline, with a preconstructed instance.
573 // Note: If the deopt count has blown up, the uncommon trap
574 // runtime is going to flush this nmethod, not matter what.
575 if (treat_throw_as_hot
576 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
577 // If the throw is local, we use a pre-existing instance and
578 // punt on the backtrace. This would lead to a missing backtrace
579 // (a repeat of 4292742) if the backtrace object is ever asked
580 // for its backtrace.
581 // Fixing this remaining case of 4292742 requires some flavor of
582 // escape analysis. Leave that for the future.
583 ciInstance* ex_obj = NULL;
584 switch (reason) {
585 case Deoptimization::Reason_null_check:
586 ex_obj = env()->NullPointerException_instance();
587 break;
588 case Deoptimization::Reason_div0_check:
589 ex_obj = env()->ArithmeticException_instance();
590 break;
591 case Deoptimization::Reason_range_check:
592 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
593 break;
594 case Deoptimization::Reason_class_check:
595 if (java_bc() == Bytecodes::_aastore) {
596 ex_obj = env()->ArrayStoreException_instance();
597 } else {
598 ex_obj = env()->ClassCastException_instance();
599 }
600 break;
601 default:
602 break;
603 }
604 if (failing()) { stop(); return; } // exception allocation might fail
605 if (ex_obj != NULL) {
606 if (env()->jvmti_can_post_on_exceptions()) {
607 // check if we must post exception events, take uncommon trap if so
608 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
609 // here if should_post_on_exceptions is false
610 // continue on with the normal codegen
611 }
612
613 // Cheat with a preallocated exception object.
614 if (C->log() != NULL)
615 C->log()->elem("hot_throw preallocated='1' reason='%s'",
616 Deoptimization::trap_reason_name(reason));
617 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
618 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
619
620 // Clear the detail message of the preallocated exception object.
621 // Weblogic sometimes mutates the detail message of exceptions
622 // using reflection.
623 int offset = java_lang_Throwable::get_detailMessage_offset();
624 const TypePtr* adr_typ = ex_con->add_offset(offset);
625
626 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
627 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
628 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
629
630 add_exception_state(make_exception_state(ex_node));
631 return;
632 }
633 }
634
635 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
636 // It won't be much cheaper than bailing to the interp., since we'll
637 // have to pass up all the debug-info, and the runtime will have to
638 // create the stack trace.
639
640 // Usual case: Bail to interpreter.
641 // Reserve the right to recompile if we haven't seen anything yet.
642
643 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
644 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
645 if (treat_throw_as_hot
646 && (method()->method_data()->trap_recompiled_at(bci(), m)
647 || C->too_many_traps(reason))) {
648 // We cannot afford to take more traps here. Suffer in the interpreter.
649 if (C->log() != NULL)
650 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
651 Deoptimization::trap_reason_name(reason),
652 C->trap_count(reason));
653 action = Deoptimization::Action_none;
654 }
655
656 // "must_throw" prunes the JVM state to include only the stack, if there
657 // are no local exception handlers. This should cut down on register
658 // allocation time and code size, by drastically reducing the number
659 // of in-edges on the call to the uncommon trap.
660
661 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
662 }
663
664
665 //----------------------------PreserveJVMState---------------------------------
666 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
667 debug_only(kit->verify_map());
668 _kit = kit;
669 _map = kit->map(); // preserve the map
670 _sp = kit->sp();
671 kit->set_map(clone_map ? kit->clone_map() : NULL);
672 #ifdef ASSERT
673 _bci = kit->bci();
674 Parse* parser = kit->is_Parse();
675 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
676 _block = block;
677 #endif
678 }
679 PreserveJVMState::~PreserveJVMState() {
680 GraphKit* kit = _kit;
681 #ifdef ASSERT
682 assert(kit->bci() == _bci, "bci must not shift");
683 Parse* parser = kit->is_Parse();
684 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
685 assert(block == _block, "block must not shift");
686 #endif
687 kit->set_map(_map);
688 kit->set_sp(_sp);
689 }
690
691
692 //-----------------------------BuildCutout-------------------------------------
693 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
694 : PreserveJVMState(kit)
695 {
696 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
697 SafePointNode* outer_map = _map; // preserved map is caller's
698 SafePointNode* inner_map = kit->map();
699 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
700 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
701 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
702 }
703 BuildCutout::~BuildCutout() {
704 GraphKit* kit = _kit;
705 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
706 }
707
708 //---------------------------PreserveReexecuteState----------------------------
709 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
710 assert(!kit->stopped(), "must call stopped() before");
711 _kit = kit;
712 _sp = kit->sp();
713 _reexecute = kit->jvms()->_reexecute;
714 }
715 PreserveReexecuteState::~PreserveReexecuteState() {
716 if (_kit->stopped()) return;
717 _kit->jvms()->_reexecute = _reexecute;
718 _kit->set_sp(_sp);
719 }
720
721 //------------------------------clone_map--------------------------------------
722 // Implementation of PreserveJVMState
723 //
724 // Only clone_map(...) here. If this function is only used in the
725 // PreserveJVMState class we may want to get rid of this extra
726 // function eventually and do it all there.
727
728 SafePointNode* GraphKit::clone_map() {
729 if (map() == NULL) return NULL;
730
731 // Clone the memory edge first
732 Node* mem = MergeMemNode::make(map()->memory());
733 gvn().set_type_bottom(mem);
734
735 SafePointNode *clonemap = (SafePointNode*)map()->clone();
736 JVMState* jvms = this->jvms();
737 JVMState* clonejvms = jvms->clone_shallow(C);
738 clonemap->set_memory(mem);
739 clonemap->set_jvms(clonejvms);
740 clonejvms->set_map(clonemap);
741 record_for_igvn(clonemap);
742 gvn().set_type_bottom(clonemap);
743 return clonemap;
744 }
745
746
747 //-----------------------------set_map_clone-----------------------------------
748 void GraphKit::set_map_clone(SafePointNode* m) {
749 _map = m;
750 _map = clone_map();
751 _map->set_next_exception(NULL);
752 debug_only(verify_map());
753 }
754
755
756 //----------------------------kill_dead_locals---------------------------------
757 // Detect any locals which are known to be dead, and force them to top.
758 void GraphKit::kill_dead_locals() {
759 // Consult the liveness information for the locals. If any
760 // of them are unused, then they can be replaced by top(). This
761 // should help register allocation time and cut down on the size
762 // of the deoptimization information.
763
764 // This call is made from many of the bytecode handling
765 // subroutines called from the Big Switch in do_one_bytecode.
766 // Every bytecode which might include a slow path is responsible
767 // for killing its dead locals. The more consistent we
768 // are about killing deads, the fewer useless phis will be
769 // constructed for them at various merge points.
770
771 // bci can be -1 (InvocationEntryBci). We return the entry
772 // liveness for the method.
773
774 if (method() == NULL || method()->code_size() == 0) {
775 // We are building a graph for a call to a native method.
776 // All locals are live.
777 return;
778 }
779
780 ResourceMark rm;
781
782 // Consult the liveness information for the locals. If any
783 // of them are unused, then they can be replaced by top(). This
784 // should help register allocation time and cut down on the size
785 // of the deoptimization information.
786 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
787
788 int len = (int)live_locals.size();
789 assert(len <= jvms()->loc_size(), "too many live locals");
790 for (int local = 0; local < len; local++) {
791 if (!live_locals.at(local)) {
792 set_local(local, top());
793 }
794 }
795 }
796
797 #ifdef ASSERT
798 //-------------------------dead_locals_are_killed------------------------------
799 // Return true if all dead locals are set to top in the map.
800 // Used to assert "clean" debug info at various points.
801 bool GraphKit::dead_locals_are_killed() {
802 if (method() == NULL || method()->code_size() == 0) {
803 // No locals need to be dead, so all is as it should be.
804 return true;
805 }
806
807 // Make sure somebody called kill_dead_locals upstream.
808 ResourceMark rm;
809 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
810 if (jvms->loc_size() == 0) continue; // no locals to consult
811 SafePointNode* map = jvms->map();
812 ciMethod* method = jvms->method();
813 int bci = jvms->bci();
814 if (jvms == this->jvms()) {
815 bci = this->bci(); // it might not yet be synched
816 }
817 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
818 int len = (int)live_locals.size();
819 if (!live_locals.is_valid() || len == 0)
820 // This method is trivial, or is poisoned by a breakpoint.
821 return true;
822 assert(len == jvms->loc_size(), "live map consistent with locals map");
823 for (int local = 0; local < len; local++) {
824 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
825 if (PrintMiscellaneous && (Verbose || WizardMode)) {
826 tty->print_cr("Zombie local %d: ", local);
827 jvms->dump();
828 }
829 return false;
830 }
831 }
832 }
833 return true;
834 }
835
836 #endif //ASSERT
837
838 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
839 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
840 ciMethod* cur_method = jvms->method();
841 int cur_bci = jvms->bci();
842 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
843 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
844 return Interpreter::bytecode_should_reexecute(code) ||
845 (is_anewarray && (code == Bytecodes::_multianewarray));
846 // Reexecute _multianewarray bytecode which was replaced with
847 // sequence of [a]newarray. See Parse::do_multianewarray().
848 //
849 // Note: interpreter should not have it set since this optimization
850 // is limited by dimensions and guarded by flag so in some cases
851 // multianewarray() runtime calls will be generated and
852 // the bytecode should not be reexecutes (stack will not be reset).
853 } else {
854 return false;
855 }
856 }
857
858 // Helper function for adding JVMState and debug information to node
859 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
860 // Add the safepoint edges to the call (or other safepoint).
861
862 // Make sure dead locals are set to top. This
863 // should help register allocation time and cut down on the size
864 // of the deoptimization information.
865 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
866
867 // Walk the inline list to fill in the correct set of JVMState's
868 // Also fill in the associated edges for each JVMState.
869
870 // If the bytecode needs to be reexecuted we need to put
871 // the arguments back on the stack.
872 const bool should_reexecute = jvms()->should_reexecute();
873 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
874
875 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
876 // undefined if the bci is different. This is normal for Parse but it
877 // should not happen for LibraryCallKit because only one bci is processed.
878 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
879 "in LibraryCallKit the reexecute bit should not change");
880
881 // If we are guaranteed to throw, we can prune everything but the
882 // input to the current bytecode.
883 bool can_prune_locals = false;
884 uint stack_slots_not_pruned = 0;
885 int inputs = 0, depth = 0;
886 if (must_throw) {
887 assert(method() == youngest_jvms->method(), "sanity");
888 if (compute_stack_effects(inputs, depth)) {
889 can_prune_locals = true;
890 stack_slots_not_pruned = inputs;
891 }
892 }
893
894 if (env()->should_retain_local_variables()) {
895 // At any safepoint, this method can get breakpointed, which would
896 // then require an immediate deoptimization.
897 can_prune_locals = false; // do not prune locals
898 stack_slots_not_pruned = 0;
899 }
900
901 // do not scribble on the input jvms
902 JVMState* out_jvms = youngest_jvms->clone_deep(C);
903 call->set_jvms(out_jvms); // Start jvms list for call node
904
905 // For a known set of bytecodes, the interpreter should reexecute them if
906 // deoptimization happens. We set the reexecute state for them here
907 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
908 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
909 #ifdef ASSERT
910 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
911 assert(method() == youngest_jvms->method(), "sanity");
912 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
913 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
914 #endif // ASSERT
915 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
916 }
917
918 // Presize the call:
919 DEBUG_ONLY(uint non_debug_edges = call->req());
920 call->add_req_batch(top(), youngest_jvms->debug_depth());
921 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
922
923 // Set up edges so that the call looks like this:
924 // Call [state:] ctl io mem fptr retadr
925 // [parms:] parm0 ... parmN
926 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
927 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
928 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
929 // Note that caller debug info precedes callee debug info.
930
931 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
932 uint debug_ptr = call->req();
933
934 // Loop over the map input edges associated with jvms, add them
935 // to the call node, & reset all offsets to match call node array.
936 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
937 uint debug_end = debug_ptr;
938 uint debug_start = debug_ptr - in_jvms->debug_size();
939 debug_ptr = debug_start; // back up the ptr
940
941 uint p = debug_start; // walks forward in [debug_start, debug_end)
942 uint j, k, l;
943 SafePointNode* in_map = in_jvms->map();
944 out_jvms->set_map(call);
945
946 if (can_prune_locals) {
947 assert(in_jvms->method() == out_jvms->method(), "sanity");
948 // If the current throw can reach an exception handler in this JVMS,
949 // then we must keep everything live that can reach that handler.
950 // As a quick and dirty approximation, we look for any handlers at all.
951 if (in_jvms->method()->has_exception_handlers()) {
952 can_prune_locals = false;
953 }
954 }
955
956 // Add the Locals
957 k = in_jvms->locoff();
958 l = in_jvms->loc_size();
959 out_jvms->set_locoff(p);
960 if (!can_prune_locals) {
961 for (j = 0; j < l; j++)
962 call->set_req(p++, in_map->in(k+j));
963 } else {
964 p += l; // already set to top above by add_req_batch
965 }
966
967 // Add the Expression Stack
968 k = in_jvms->stkoff();
969 l = in_jvms->sp();
970 out_jvms->set_stkoff(p);
971 if (!can_prune_locals) {
972 for (j = 0; j < l; j++)
973 call->set_req(p++, in_map->in(k+j));
974 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
975 // Divide stack into {S0,...,S1}, where S0 is set to top.
976 uint s1 = stack_slots_not_pruned;
977 stack_slots_not_pruned = 0; // for next iteration
978 if (s1 > l) s1 = l;
979 uint s0 = l - s1;
980 p += s0; // skip the tops preinstalled by add_req_batch
981 for (j = s0; j < l; j++)
982 call->set_req(p++, in_map->in(k+j));
983 } else {
984 p += l; // already set to top above by add_req_batch
985 }
986
987 // Add the Monitors
988 k = in_jvms->monoff();
989 l = in_jvms->mon_size();
990 out_jvms->set_monoff(p);
991 for (j = 0; j < l; j++)
992 call->set_req(p++, in_map->in(k+j));
993
994 // Copy any scalar object fields.
995 k = in_jvms->scloff();
996 l = in_jvms->scl_size();
997 out_jvms->set_scloff(p);
998 for (j = 0; j < l; j++)
999 call->set_req(p++, in_map->in(k+j));
1000
1001 // Finish the new jvms.
1002 out_jvms->set_endoff(p);
1003
1004 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1005 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1006 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1007 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1008 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1009 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1010
1011 // Update the two tail pointers in parallel.
1012 out_jvms = out_jvms->caller();
1013 in_jvms = in_jvms->caller();
1014 }
1015
1016 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1017
1018 // Test the correctness of JVMState::debug_xxx accessors:
1019 assert(call->jvms()->debug_start() == non_debug_edges, "");
1020 assert(call->jvms()->debug_end() == call->req(), "");
1021 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1022 }
1023
1024 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1025 Bytecodes::Code code = java_bc();
1026 if (code == Bytecodes::_wide) {
1027 code = method()->java_code_at_bci(bci() + 1);
1028 }
1029
1030 BasicType rtype = T_ILLEGAL;
1031 int rsize = 0;
1032
1033 if (code != Bytecodes::_illegal) {
1034 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1035 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1036 if (rtype < T_CONFLICT)
1037 rsize = type2size[rtype];
1038 }
1039
1040 switch (code) {
1041 case Bytecodes::_illegal:
1042 return false;
1043
1044 case Bytecodes::_ldc:
1045 case Bytecodes::_ldc_w:
1046 case Bytecodes::_ldc2_w:
1047 inputs = 0;
1048 break;
1049
1050 case Bytecodes::_dup: inputs = 1; break;
1051 case Bytecodes::_dup_x1: inputs = 2; break;
1052 case Bytecodes::_dup_x2: inputs = 3; break;
1053 case Bytecodes::_dup2: inputs = 2; break;
1054 case Bytecodes::_dup2_x1: inputs = 3; break;
1055 case Bytecodes::_dup2_x2: inputs = 4; break;
1056 case Bytecodes::_swap: inputs = 2; break;
1057 case Bytecodes::_arraylength: inputs = 1; break;
1058
1059 case Bytecodes::_getstatic:
1060 case Bytecodes::_putstatic:
1061 case Bytecodes::_getfield:
1062 case Bytecodes::_putfield:
1063 {
1064 bool ignored_will_link;
1065 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1066 int size = field->type()->size();
1067 bool is_get = (depth >= 0), is_static = (depth & 1);
1068 inputs = (is_static ? 0 : 1);
1069 if (is_get) {
1070 depth = size - inputs;
1071 } else {
1072 inputs += size; // putxxx pops the value from the stack
1073 depth = - inputs;
1074 }
1075 }
1076 break;
1077
1078 case Bytecodes::_invokevirtual:
1079 case Bytecodes::_invokespecial:
1080 case Bytecodes::_invokestatic:
1081 case Bytecodes::_invokedynamic:
1082 case Bytecodes::_invokeinterface:
1083 {
1084 bool ignored_will_link;
1085 ciSignature* declared_signature = NULL;
1086 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1087 assert(declared_signature != NULL, "cannot be null");
1088 inputs = declared_signature->arg_size_for_bc(code);
1089 int size = declared_signature->return_type()->size();
1090 depth = size - inputs;
1091 }
1092 break;
1093
1094 case Bytecodes::_multianewarray:
1095 {
1096 ciBytecodeStream iter(method());
1097 iter.reset_to_bci(bci());
1098 iter.next();
1099 inputs = iter.get_dimensions();
1100 assert(rsize == 1, "");
1101 depth = rsize - inputs;
1102 }
1103 break;
1104
1105 case Bytecodes::_withfield: {
1106 bool ignored_will_link;
1107 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1108 int size = field->type()->size();
1109 inputs = size+1;
1110 depth = rsize - inputs;
1111 break;
1112 }
1113
1114 case Bytecodes::_ireturn:
1115 case Bytecodes::_lreturn:
1116 case Bytecodes::_freturn:
1117 case Bytecodes::_dreturn:
1118 case Bytecodes::_areturn:
1119 assert(rsize == -depth, "");
1120 inputs = rsize;
1121 break;
1122
1123 case Bytecodes::_jsr:
1124 case Bytecodes::_jsr_w:
1125 inputs = 0;
1126 depth = 1; // S.B. depth=1, not zero
1127 break;
1128
1129 default:
1130 // bytecode produces a typed result
1131 inputs = rsize - depth;
1132 assert(inputs >= 0, "");
1133 break;
1134 }
1135
1136 #ifdef ASSERT
1137 // spot check
1138 int outputs = depth + inputs;
1139 assert(outputs >= 0, "sanity");
1140 switch (code) {
1141 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1142 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1143 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1144 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1145 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1146 default: break;
1147 }
1148 #endif //ASSERT
1149
1150 return true;
1151 }
1152
1153
1154
1155 //------------------------------basic_plus_adr---------------------------------
1156 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1157 // short-circuit a common case
1158 if (offset == intcon(0)) return ptr;
1159 return _gvn.transform( new AddPNode(base, ptr, offset) );
1160 }
1161
1162 Node* GraphKit::ConvI2L(Node* offset) {
1163 // short-circuit a common case
1164 jint offset_con = find_int_con(offset, Type::OffsetBot);
1165 if (offset_con != Type::OffsetBot) {
1166 return longcon((jlong) offset_con);
1167 }
1168 return _gvn.transform( new ConvI2LNode(offset));
1169 }
1170
1171 Node* GraphKit::ConvI2UL(Node* offset) {
1172 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1173 if (offset_con != (juint) Type::OffsetBot) {
1174 return longcon((julong) offset_con);
1175 }
1176 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1177 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1178 return _gvn.transform( new AndLNode(conv, mask) );
1179 }
1180
1181 Node* GraphKit::ConvL2I(Node* offset) {
1182 // short-circuit a common case
1183 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1184 if (offset_con != (jlong)Type::OffsetBot) {
1185 return intcon((int) offset_con);
1186 }
1187 return _gvn.transform( new ConvL2INode(offset));
1188 }
1189
1190 //-------------------------load_object_klass-----------------------------------
1191 Node* GraphKit::load_object_klass(Node* obj) {
1192 // Special-case a fresh allocation to avoid building nodes:
1193 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1194 if (akls != NULL) return akls;
1195 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1196 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1197 }
1198
1199 //-------------------------load_array_length-----------------------------------
1200 Node* GraphKit::load_array_length(Node* array) {
1201 // Special-case a fresh allocation to avoid building nodes:
1202 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1203 Node *alen;
1204 if (alloc == NULL) {
1205 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1206 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1207 } else {
1208 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1209 }
1210 return alen;
1211 }
1212
1213 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1214 const TypeOopPtr* oop_type,
1215 bool replace_length_in_map) {
1216 Node* length = alloc->Ideal_length();
1217 if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1218 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1219 if (ccast != length) {
1220 // do not transfrom ccast here, it might convert to top node for
1221 // negative array length and break assumptions in parsing stage.
1222 _gvn.set_type_bottom(ccast);
1223 record_for_igvn(ccast);
1224 if (replace_length_in_map) {
1225 replace_in_map(length, ccast);
1226 }
1227 return ccast;
1228 }
1229 }
1230 return length;
1231 }
1232
1233 //------------------------------do_null_check----------------------------------
1234 // Helper function to do a NULL pointer check. Returned value is
1235 // the incoming address with NULL casted away. You are allowed to use the
1236 // not-null value only if you are control dependent on the test.
1237 #ifndef PRODUCT
1238 extern int explicit_null_checks_inserted,
1239 explicit_null_checks_elided;
1240 #endif
1241 Node* GraphKit::null_check_common(Node* value, BasicType type,
1242 // optional arguments for variations:
1243 bool assert_null,
1244 Node* *null_control,
1245 bool speculative,
1246 bool is_init_check) {
1247 assert(!assert_null || null_control == NULL, "not both at once");
1248 if (stopped()) return top();
1249 NOT_PRODUCT(explicit_null_checks_inserted++);
1250
1251 if (value->is_InlineTypePtr()) {
1252 // Null checking a scalarized but nullable inline type. Check the is_init
1253 // input instead of the oop input to avoid keeping buffer allocations alive.
1254 InlineTypePtrNode* vtptr = value->as_InlineTypePtr();
1255 while (vtptr->get_oop()->is_InlineTypePtr()) {
1256 vtptr = vtptr->get_oop()->as_InlineTypePtr();
1257 }
1258 null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true);
1259 if (stopped()) {
1260 return top();
1261 }
1262 bool do_replace_in_map = (null_control == NULL || (*null_control) == top());
1263 return cast_not_null(value, do_replace_in_map);
1264 }
1265
1266 // Construct NULL check
1267 Node *chk = NULL;
1268 switch(type) {
1269 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1270 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1271 case T_INLINE_TYPE : // fall through
1272 case T_ARRAY : // fall through
1273 type = T_OBJECT; // simplify further tests
1274 case T_OBJECT : {
1275 const Type *t = _gvn.type( value );
1276
1277 const TypeOopPtr* tp = t->isa_oopptr();
1278 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1279 // Only for do_null_check, not any of its siblings:
1280 && !assert_null && null_control == NULL) {
1281 // Usually, any field access or invocation on an unloaded oop type
1282 // will simply fail to link, since the statically linked class is
1283 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1284 // the static class is loaded but the sharper oop type is not.
1285 // Rather than checking for this obscure case in lots of places,
1286 // we simply observe that a null check on an unloaded class
1287 // will always be followed by a nonsense operation, so we
1288 // can just issue the uncommon trap here.
1289 // Our access to the unloaded class will only be correct
1290 // after it has been loaded and initialized, which requires
1291 // a trip through the interpreter.
1292 #ifndef PRODUCT
1293 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1294 #endif
1295 uncommon_trap(Deoptimization::Reason_unloaded,
1296 Deoptimization::Action_reinterpret,
1297 tp->klass(), "!loaded");
1298 return top();
1299 }
1300
1301 if (assert_null) {
1302 // See if the type is contained in NULL_PTR.
1303 // If so, then the value is already null.
1304 if (t->higher_equal(TypePtr::NULL_PTR)) {
1305 NOT_PRODUCT(explicit_null_checks_elided++);
1306 return value; // Elided null assert quickly!
1307 }
1308 } else {
1309 // See if mixing in the NULL pointer changes type.
1310 // If so, then the NULL pointer was not allowed in the original
1311 // type. In other words, "value" was not-null.
1312 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1313 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1314 NOT_PRODUCT(explicit_null_checks_elided++);
1315 return value; // Elided null check quickly!
1316 }
1317 }
1318 chk = new CmpPNode( value, null() );
1319 break;
1320 }
1321
1322 default:
1323 fatal("unexpected type: %s", type2name(type));
1324 }
1325 assert(chk != NULL, "sanity check");
1326 chk = _gvn.transform(chk);
1327
1328 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1329 BoolNode *btst = new BoolNode( chk, btest);
1330 Node *tst = _gvn.transform( btst );
1331
1332 //-----------
1333 // if peephole optimizations occurred, a prior test existed.
1334 // If a prior test existed, maybe it dominates as we can avoid this test.
1335 if (tst != btst && type == T_OBJECT) {
1336 // At this point we want to scan up the CFG to see if we can
1337 // find an identical test (and so avoid this test altogether).
1338 Node *cfg = control();
1339 int depth = 0;
1340 while( depth < 16 ) { // Limit search depth for speed
1341 if( cfg->Opcode() == Op_IfTrue &&
1342 cfg->in(0)->in(1) == tst ) {
1343 // Found prior test. Use "cast_not_null" to construct an identical
1344 // CastPP (and hence hash to) as already exists for the prior test.
1345 // Return that casted value.
1346 if (assert_null) {
1347 replace_in_map(value, null());
1348 return null(); // do not issue the redundant test
1349 }
1350 Node *oldcontrol = control();
1351 set_control(cfg);
1352 Node *res = cast_not_null(value);
1353 set_control(oldcontrol);
1354 NOT_PRODUCT(explicit_null_checks_elided++);
1355 return res;
1356 }
1357 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1358 if (cfg == NULL) break; // Quit at region nodes
1359 depth++;
1360 }
1361 }
1362
1363 //-----------
1364 // Branch to failure if null
1365 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1366 Deoptimization::DeoptReason reason;
1367 if (assert_null) {
1368 reason = Deoptimization::reason_null_assert(speculative);
1369 } else if (type == T_OBJECT || is_init_check) {
1370 reason = Deoptimization::reason_null_check(speculative);
1371 } else {
1372 reason = Deoptimization::Reason_div0_check;
1373 }
1374 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1375 // ciMethodData::has_trap_at will return a conservative -1 if any
1376 // must-be-null assertion has failed. This could cause performance
1377 // problems for a method after its first do_null_assert failure.
1378 // Consider using 'Reason_class_check' instead?
1379
1380 // To cause an implicit null check, we set the not-null probability
1381 // to the maximum (PROB_MAX). For an explicit check the probability
1382 // is set to a smaller value.
1383 if (null_control != NULL || too_many_traps(reason)) {
1384 // probability is less likely
1385 ok_prob = PROB_LIKELY_MAG(3);
1386 } else if (!assert_null &&
1387 (ImplicitNullCheckThreshold > 0) &&
1388 method() != NULL &&
1389 (method()->method_data()->trap_count(reason)
1390 >= (uint)ImplicitNullCheckThreshold)) {
1391 ok_prob = PROB_LIKELY_MAG(3);
1392 }
1393
1394 if (null_control != NULL) {
1395 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1396 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1397 set_control( _gvn.transform( new IfTrueNode(iff)));
1398 #ifndef PRODUCT
1399 if (null_true == top()) {
1400 explicit_null_checks_elided++;
1401 }
1402 #endif
1403 (*null_control) = null_true;
1404 } else {
1405 BuildCutout unless(this, tst, ok_prob);
1406 // Check for optimizer eliding test at parse time
1407 if (stopped()) {
1408 // Failure not possible; do not bother making uncommon trap.
1409 NOT_PRODUCT(explicit_null_checks_elided++);
1410 } else if (assert_null) {
1411 uncommon_trap(reason,
1412 Deoptimization::Action_make_not_entrant,
1413 NULL, "assert_null");
1414 } else {
1415 replace_in_map(value, zerocon(type));
1416 builtin_throw(reason);
1417 }
1418 }
1419
1420 // Must throw exception, fall-thru not possible?
1421 if (stopped()) {
1422 return top(); // No result
1423 }
1424
1425 if (assert_null) {
1426 // Cast obj to null on this path.
1427 replace_in_map(value, zerocon(type));
1428 return zerocon(type);
1429 }
1430
1431 // Cast obj to not-null on this path, if there is no null_control.
1432 // (If there is a null_control, a non-null value may come back to haunt us.)
1433 if (type == T_OBJECT) {
1434 Node* cast = cast_not_null(value, false);
1435 if (null_control == NULL || (*null_control) == top())
1436 replace_in_map(value, cast);
1437 value = cast;
1438 }
1439
1440 return value;
1441 }
1442
1443 //------------------------------cast_not_null----------------------------------
1444 // Cast obj to not-null on this path
1445 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1446 if (obj->is_InlineType()) {
1447 return obj;
1448 } else if (obj->is_InlineTypePtr()) {
1449 // Cast oop input instead
1450 Node* cast = cast_not_null(obj->as_InlineTypePtr()->get_oop(), do_replace_in_map);
1451 if (cast->is_top()) {
1452 // Always null
1453 return top();
1454 }
1455 // Create a new node with the casted oop input and is_init set
1456 InlineTypeBaseNode* vt = obj->clone()->as_InlineTypePtr();
1457 vt->set_oop(cast);
1458 vt->set_is_init(_gvn);
1459 vt = _gvn.transform(vt)->as_InlineTypePtr();
1460 if (do_replace_in_map) {
1461 replace_in_map(obj, vt);
1462 }
1463 return vt;
1464 }
1465 const Type *t = _gvn.type(obj);
1466 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1467 // Object is already not-null?
1468 if( t == t_not_null ) return obj;
1469
1470 Node *cast = new CastPPNode(obj,t_not_null);
1471 cast->init_req(0, control());
1472 cast = _gvn.transform( cast );
1473
1474 // Scan for instances of 'obj' in the current JVM mapping.
1475 // These instances are known to be not-null after the test.
1476 if (do_replace_in_map)
1477 replace_in_map(obj, cast);
1478
1479 return cast; // Return casted value
1480 }
1481
1482 // Sometimes in intrinsics, we implicitly know an object is not null
1483 // (there's no actual null check) so we can cast it to not null. In
1484 // the course of optimizations, the input to the cast can become null.
1485 // In that case that data path will die and we need the control path
1486 // to become dead as well to keep the graph consistent. So we have to
1487 // add a check for null for which one branch can't be taken. It uses
1488 // an Opaque4 node that will cause the check to be removed after loop
1489 // opts so the test goes away and the compiled code doesn't execute a
1490 // useless check.
1491 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1492 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1493 return value;
1494 }
1495 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1496 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1497 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1498 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1499 _gvn.set_type(iff, iff->Value(&_gvn));
1500 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1501 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1502 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1503 C->root()->add_req(halt);
1504 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1505 set_control(if_t);
1506 return cast_not_null(value, do_replace_in_map);
1507 }
1508
1509
1510 //--------------------------replace_in_map-------------------------------------
1511 void GraphKit::replace_in_map(Node* old, Node* neww) {
1512 if (old == neww) {
1513 return;
1514 }
1515
1516 map()->replace_edge(old, neww);
1517
1518 // Note: This operation potentially replaces any edge
1519 // on the map. This includes locals, stack, and monitors
1520 // of the current (innermost) JVM state.
1521
1522 // don't let inconsistent types from profiling escape this
1523 // method
1524
1525 const Type* told = _gvn.type(old);
1526 const Type* tnew = _gvn.type(neww);
1527
1528 if (!tnew->higher_equal(told)) {
1529 return;
1530 }
1531
1532 map()->record_replaced_node(old, neww);
1533 }
1534
1535
1536 //=============================================================================
1537 //--------------------------------memory---------------------------------------
1538 Node* GraphKit::memory(uint alias_idx) {
1539 MergeMemNode* mem = merged_memory();
1540 Node* p = mem->memory_at(alias_idx);
1541 assert(p != mem->empty_memory(), "empty");
1542 _gvn.set_type(p, Type::MEMORY); // must be mapped
1543 return p;
1544 }
1545
1546 //-----------------------------reset_memory------------------------------------
1547 Node* GraphKit::reset_memory() {
1548 Node* mem = map()->memory();
1549 // do not use this node for any more parsing!
1550 debug_only( map()->set_memory((Node*)NULL) );
1551 return _gvn.transform( mem );
1552 }
1553
1554 //------------------------------set_all_memory---------------------------------
1555 void GraphKit::set_all_memory(Node* newmem) {
1556 Node* mergemem = MergeMemNode::make(newmem);
1557 gvn().set_type_bottom(mergemem);
1558 map()->set_memory(mergemem);
1559 }
1560
1561 //------------------------------set_all_memory_call----------------------------
1562 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1563 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1564 set_all_memory(newmem);
1565 }
1566
1567 //=============================================================================
1568 //
1569 // parser factory methods for MemNodes
1570 //
1571 // These are layered on top of the factory methods in LoadNode and StoreNode,
1572 // and integrate with the parser's memory state and _gvn engine.
1573 //
1574
1575 // factory methods in "int adr_idx"
1576 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1577 int adr_idx,
1578 MemNode::MemOrd mo,
1579 LoadNode::ControlDependency control_dependency,
1580 bool require_atomic_access,
1581 bool unaligned,
1582 bool mismatched,
1583 bool unsafe,
1584 uint8_t barrier_data) {
1585 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1586 const TypePtr* adr_type = NULL; // debug-mode-only argument
1587 debug_only(adr_type = C->get_adr_type(adr_idx));
1588 Node* mem = memory(adr_idx);
1589 Node* ld;
1590 if (require_atomic_access && bt == T_LONG) {
1591 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1592 } else if (require_atomic_access && bt == T_DOUBLE) {
1593 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1594 } else {
1595 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1596 }
1597 ld = _gvn.transform(ld);
1598
1599 if (((bt == T_OBJECT || bt == T_INLINE_TYPE) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1600 // Improve graph before escape analysis and boxing elimination.
1601 record_for_igvn(ld);
1602 }
1603 return ld;
1604 }
1605
1606 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1607 int adr_idx,
1608 MemNode::MemOrd mo,
1609 bool require_atomic_access,
1610 bool unaligned,
1611 bool mismatched,
1612 bool unsafe) {
1613 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1614 const TypePtr* adr_type = NULL;
1615 debug_only(adr_type = C->get_adr_type(adr_idx));
1616 Node *mem = memory(adr_idx);
1617 Node* st;
1618 if (require_atomic_access && bt == T_LONG) {
1619 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1620 } else if (require_atomic_access && bt == T_DOUBLE) {
1621 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1622 } else {
1623 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1624 }
1625 if (unaligned) {
1626 st->as_Store()->set_unaligned_access();
1627 }
1628 if (mismatched) {
1629 st->as_Store()->set_mismatched_access();
1630 }
1631 if (unsafe) {
1632 st->as_Store()->set_unsafe_access();
1633 }
1634 st = _gvn.transform(st);
1635 set_memory(st, adr_idx);
1636 // Back-to-back stores can only remove intermediate store with DU info
1637 // so push on worklist for optimizer.
1638 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1639 record_for_igvn(st);
1640
1641 return st;
1642 }
1643
1644 Node* GraphKit::access_store_at(Node* obj,
1645 Node* adr,
1646 const TypePtr* adr_type,
1647 Node* val,
1648 const Type* val_type,
1649 BasicType bt,
1650 DecoratorSet decorators,
1651 bool safe_for_replace) {
1652 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1653 // could be delayed during Parse (for example, in adjust_map_after_if()).
1654 // Execute transformation here to avoid barrier generation in such case.
1655 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1656 val = _gvn.makecon(TypePtr::NULL_PTR);
1657 }
1658
1659 if (stopped()) {
1660 return top(); // Dead path ?
1661 }
1662
1663 assert(val != NULL, "not dead path");
1664 if (val->is_InlineType()) {
1665 // Store to non-flattened field. Buffer the inline type and make sure
1666 // the store is re-executed if the allocation triggers deoptimization.
1667 PreserveReexecuteState preexecs(this);
1668 jvms()->set_should_reexecute(true);
1669 val = val->as_InlineType()->buffer(this, safe_for_replace);
1670 }
1671
1672 C2AccessValuePtr addr(adr, adr_type);
1673 C2AccessValue value(val, val_type);
1674 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1675 if (access.is_raw()) {
1676 return _barrier_set->BarrierSetC2::store_at(access, value);
1677 } else {
1678 return _barrier_set->store_at(access, value);
1679 }
1680 }
1681
1682 Node* GraphKit::access_load_at(Node* obj, // containing obj
1683 Node* adr, // actual adress to store val at
1684 const TypePtr* adr_type,
1685 const Type* val_type,
1686 BasicType bt,
1687 DecoratorSet decorators,
1688 Node* ctl) {
1689 if (stopped()) {
1690 return top(); // Dead path ?
1691 }
1692
1693 C2AccessValuePtr addr(adr, adr_type);
1694 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1695 if (access.is_raw()) {
1696 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1697 } else {
1698 return _barrier_set->load_at(access, val_type);
1699 }
1700 }
1701
1702 Node* GraphKit::access_load(Node* adr, // actual adress to load val at
1703 const Type* val_type,
1704 BasicType bt,
1705 DecoratorSet decorators) {
1706 if (stopped()) {
1707 return top(); // Dead path ?
1708 }
1709
1710 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1711 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1712 if (access.is_raw()) {
1713 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1714 } else {
1715 return _barrier_set->load_at(access, val_type);
1716 }
1717 }
1718
1719 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1720 Node* adr,
1721 const TypePtr* adr_type,
1722 int alias_idx,
1723 Node* expected_val,
1724 Node* new_val,
1725 const Type* value_type,
1726 BasicType bt,
1727 DecoratorSet decorators) {
1728 C2AccessValuePtr addr(adr, adr_type);
1729 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1730 bt, obj, addr, alias_idx);
1731 if (access.is_raw()) {
1732 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1733 } else {
1734 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1735 }
1736 }
1737
1738 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1739 Node* adr,
1740 const TypePtr* adr_type,
1741 int alias_idx,
1742 Node* expected_val,
1743 Node* new_val,
1744 const Type* value_type,
1745 BasicType bt,
1746 DecoratorSet decorators) {
1747 C2AccessValuePtr addr(adr, adr_type);
1748 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1749 bt, obj, addr, alias_idx);
1750 if (access.is_raw()) {
1751 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1752 } else {
1753 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1754 }
1755 }
1756
1757 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1758 Node* adr,
1759 const TypePtr* adr_type,
1760 int alias_idx,
1761 Node* new_val,
1762 const Type* value_type,
1763 BasicType bt,
1764 DecoratorSet decorators) {
1765 C2AccessValuePtr addr(adr, adr_type);
1766 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1767 bt, obj, addr, alias_idx);
1768 if (access.is_raw()) {
1769 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1770 } else {
1771 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1772 }
1773 }
1774
1775 Node* GraphKit::access_atomic_add_at(Node* obj,
1776 Node* adr,
1777 const TypePtr* adr_type,
1778 int alias_idx,
1779 Node* new_val,
1780 const Type* value_type,
1781 BasicType bt,
1782 DecoratorSet decorators) {
1783 C2AccessValuePtr addr(adr, adr_type);
1784 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1785 if (access.is_raw()) {
1786 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1787 } else {
1788 return _barrier_set->atomic_add_at(access, new_val, value_type);
1789 }
1790 }
1791
1792 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1793 return _barrier_set->clone(this, src, dst, size, is_array);
1794 }
1795
1796 //-------------------------array_element_address-------------------------
1797 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1798 const TypeInt* sizetype, Node* ctrl) {
1799 uint shift = exact_log2(type2aelembytes(elembt));
1800 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1801 if (arytype_klass != NULL && arytype_klass->is_flat_array_klass()) {
1802 ciFlatArrayKlass* vak = arytype_klass->as_flat_array_klass();
1803 shift = vak->log2_element_size();
1804 }
1805 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1806
1807 // short-circuit a common case (saves lots of confusing waste motion)
1808 jint idx_con = find_int_con(idx, -1);
1809 if (idx_con >= 0) {
1810 intptr_t offset = header + ((intptr_t)idx_con << shift);
1811 return basic_plus_adr(ary, offset);
1812 }
1813
1814 // must be correct type for alignment purposes
1815 Node* base = basic_plus_adr(ary, header);
1816 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1817 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1818 return basic_plus_adr(ary, base, scale);
1819 }
1820
1821 //-------------------------load_array_element-------------------------
1822 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1823 const Type* elemtype = arytype->elem();
1824 BasicType elembt = elemtype->array_element_basic_type();
1825 assert(elembt != T_INLINE_TYPE, "inline types are not supported by this method");
1826 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1827 if (elembt == T_NARROWOOP) {
1828 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1829 }
1830 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1831 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1832 return ld;
1833 }
1834
1835 //-------------------------set_arguments_for_java_call-------------------------
1836 // Arguments (pre-popped from the stack) are taken from the JVMS.
1837 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1838 PreserveReexecuteState preexecs(this);
1839 if (EnableValhalla) {
1840 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1841 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1842 jvms()->set_should_reexecute(true);
1843 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1844 inc_sp(arg_size);
1845 }
1846 // Add the call arguments
1847 const TypeTuple* domain = call->tf()->domain_sig();
1848 uint nargs = domain->cnt();
1849 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1850 Node* arg = argument(i-TypeFunc::Parms);
1851 const Type* t = domain->field_at(i);
1852 if (call->method()->has_scalarized_args() && t->is_inlinetypeptr() && !t->maybe_null() && t->inline_klass()->can_be_passed_as_fields()) {
1853 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1854 InlineTypeBaseNode* vt = arg->as_InlineTypeBase();
1855 vt->pass_fields(this, call, idx);
1856 // If an inline type argument is passed as fields, attach the Method* to the call site
1857 // to be able to access the extended signature later via attached_method_before_pc().
1858 // For example, see CompiledMethod::preserve_callee_argument_oops().
1859 call->set_override_symbolic_info(true);
1860 continue;
1861 } else if (arg->is_InlineType()) {
1862 // Pass inline type argument via oop to callee
1863 arg = arg->as_InlineType()->buffer(this);
1864 if (!is_late_inline) {
1865 arg = arg->as_InlineTypePtr()->get_oop();
1866 }
1867 }
1868 call->init_req(idx++, arg);
1869 }
1870 }
1871
1872 //---------------------------set_edges_for_java_call---------------------------
1873 // Connect a newly created call into the current JVMS.
1874 // A return value node (if any) is returned from set_edges_for_java_call.
1875 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1876
1877 // Add the predefined inputs:
1878 call->init_req( TypeFunc::Control, control() );
1879 call->init_req( TypeFunc::I_O , i_o() );
1880 call->init_req( TypeFunc::Memory , reset_memory() );
1881 call->init_req( TypeFunc::FramePtr, frameptr() );
1882 call->init_req( TypeFunc::ReturnAdr, top() );
1883
1884 add_safepoint_edges(call, must_throw);
1885
1886 Node* xcall = _gvn.transform(call);
1887
1888 if (xcall == top()) {
1889 set_control(top());
1890 return;
1891 }
1892 assert(xcall == call, "call identity is stable");
1893
1894 // Re-use the current map to produce the result.
1895
1896 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1897 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1898 set_all_memory_call(xcall, separate_io_proj);
1899
1900 //return xcall; // no need, caller already has it
1901 }
1902
1903 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1904 if (stopped()) return top(); // maybe the call folded up?
1905
1906 // Note: Since any out-of-line call can produce an exception,
1907 // we always insert an I_O projection from the call into the result.
1908
1909 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1910
1911 if (separate_io_proj) {
1912 // The caller requested separate projections be used by the fall
1913 // through and exceptional paths, so replace the projections for
1914 // the fall through path.
1915 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1916 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1917 }
1918
1919 // Capture the return value, if any.
1920 Node* ret;
1921 if (call->method() == NULL || call->method()->return_type()->basic_type() == T_VOID) {
1922 ret = top();
1923 } else if (call->method()->return_type()->is_inlinetype()) {
1924 const Type* ret_type = call->tf()->range_sig()->field_at(TypeFunc::Parms);
1925 if (call->tf()->returns_inline_type_as_fields()) {
1926 // Return of multiple values (inline type fields): we create a
1927 // InlineType node, each field is a projection from the call.
1928 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
1929 uint base_input = TypeFunc::Parms;
1930 ret = InlineTypeNode::make_from_multi(this, call, ret_type->inline_klass(), base_input, false);
1931 } else {
1932 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1933 ret = _gvn.transform(InlineTypeNode::make_from_oop(this, ret, ret_type->inline_klass(), !ret_type->maybe_null()));
1934 }
1935 } else {
1936 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1937 }
1938
1939 return ret;
1940 }
1941
1942 //--------------------set_predefined_input_for_runtime_call--------------------
1943 // Reading and setting the memory state is way conservative here.
1944 // The real problem is that I am not doing real Type analysis on memory,
1945 // so I cannot distinguish card mark stores from other stores. Across a GC
1946 // point the Store Barrier and the card mark memory has to agree. I cannot
1947 // have a card mark store and its barrier split across the GC point from
1948 // either above or below. Here I get that to happen by reading ALL of memory.
1949 // A better answer would be to separate out card marks from other memory.
1950 // For now, return the input memory state, so that it can be reused
1951 // after the call, if this call has restricted memory effects.
1952 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1953 // Set fixed predefined input arguments
1954 Node* memory = reset_memory();
1955 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1956 call->init_req( TypeFunc::Control, control() );
1957 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1958 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1959 call->init_req( TypeFunc::FramePtr, frameptr() );
1960 call->init_req( TypeFunc::ReturnAdr, top() );
1961 return memory;
1962 }
1963
1964 //-------------------set_predefined_output_for_runtime_call--------------------
1965 // Set control and memory (not i_o) from the call.
1966 // If keep_mem is not NULL, use it for the output state,
1967 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1968 // If hook_mem is NULL, this call produces no memory effects at all.
1969 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1970 // then only that memory slice is taken from the call.
1971 // In the last case, we must put an appropriate memory barrier before
1972 // the call, so as to create the correct anti-dependencies on loads
1973 // preceding the call.
1974 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1975 Node* keep_mem,
1976 const TypePtr* hook_mem) {
1977 // no i/o
1978 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1979 if (keep_mem) {
1980 // First clone the existing memory state
1981 set_all_memory(keep_mem);
1982 if (hook_mem != NULL) {
1983 // Make memory for the call
1984 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1985 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1986 // We also use hook_mem to extract specific effects from arraycopy stubs.
1987 set_memory(mem, hook_mem);
1988 }
1989 // ...else the call has NO memory effects.
1990
1991 // Make sure the call advertises its memory effects precisely.
1992 // This lets us build accurate anti-dependences in gcm.cpp.
1993 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1994 "call node must be constructed correctly");
1995 } else {
1996 assert(hook_mem == NULL, "");
1997 // This is not a "slow path" call; all memory comes from the call.
1998 set_all_memory_call(call);
1999 }
2000 }
2001
2002 // Keep track of MergeMems feeding into other MergeMems
2003 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
2004 if (!mem->is_MergeMem()) {
2005 return;
2006 }
2007 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
2008 Node* use = i.get();
2009 if (use->is_MergeMem()) {
2010 wl.push(use);
2011 }
2012 }
2013 }
2014
2015 // Replace the call with the current state of the kit.
2016 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
2017 JVMState* ejvms = NULL;
2018 if (has_exceptions()) {
2019 ejvms = transfer_exceptions_into_jvms();
2020 }
2021
2022 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2023 ReplacedNodes replaced_nodes_exception;
2024 Node* ex_ctl = top();
2025
2026 SafePointNode* final_state = stop();
2027
2028 // Find all the needed outputs of this call
2029 CallProjections* callprojs = call->extract_projections(true);
2030
2031 Unique_Node_List wl;
2032 Node* init_mem = call->in(TypeFunc::Memory);
2033 Node* final_mem = final_state->in(TypeFunc::Memory);
2034 Node* final_ctl = final_state->in(TypeFunc::Control);
2035 Node* final_io = final_state->in(TypeFunc::I_O);
2036
2037 // Replace all the old call edges with the edges from the inlining result
2038 if (callprojs->fallthrough_catchproj != NULL) {
2039 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2040 }
2041 if (callprojs->fallthrough_memproj != NULL) {
2042 if (final_mem->is_MergeMem()) {
2043 // Parser's exits MergeMem was not transformed but may be optimized
2044 final_mem = _gvn.transform(final_mem);
2045 }
2046 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2047 add_mergemem_users_to_worklist(wl, final_mem);
2048 }
2049 if (callprojs->fallthrough_ioproj != NULL) {
2050 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2051 }
2052
2053 // Replace the result with the new result if it exists and is used
2054 if (callprojs->resproj[0] != NULL && result != NULL) {
2055 // If the inlined code is dead, the result projections for an inline type returned as
2056 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2057 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2058 "unexpected number of results");
2059 C->gvn_replace_by(callprojs->resproj[0], result);
2060 }
2061
2062 if (ejvms == NULL) {
2063 // No exception edges to simply kill off those paths
2064 if (callprojs->catchall_catchproj != NULL) {
2065 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2066 }
2067 if (callprojs->catchall_memproj != NULL) {
2068 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2069 }
2070 if (callprojs->catchall_ioproj != NULL) {
2071 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2072 }
2073 // Replace the old exception object with top
2074 if (callprojs->exobj != NULL) {
2075 C->gvn_replace_by(callprojs->exobj, C->top());
2076 }
2077 } else {
2078 GraphKit ekit(ejvms);
2079
2080 // Load my combined exception state into the kit, with all phis transformed:
2081 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2082 replaced_nodes_exception = ex_map->replaced_nodes();
2083
2084 Node* ex_oop = ekit.use_exception_state(ex_map);
2085
2086 if (callprojs->catchall_catchproj != NULL) {
2087 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2088 ex_ctl = ekit.control();
2089 }
2090 if (callprojs->catchall_memproj != NULL) {
2091 Node* ex_mem = ekit.reset_memory();
2092 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2093 add_mergemem_users_to_worklist(wl, ex_mem);
2094 }
2095 if (callprojs->catchall_ioproj != NULL) {
2096 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2097 }
2098
2099 // Replace the old exception object with the newly created one
2100 if (callprojs->exobj != NULL) {
2101 C->gvn_replace_by(callprojs->exobj, ex_oop);
2102 }
2103 }
2104
2105 // Disconnect the call from the graph
2106 call->disconnect_inputs(C);
2107 C->gvn_replace_by(call, C->top());
2108
2109 // Clean up any MergeMems that feed other MergeMems since the
2110 // optimizer doesn't like that.
2111 while (wl.size() > 0) {
2112 _gvn.transform(wl.pop());
2113 }
2114
2115 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2116 replaced_nodes.apply(C, final_ctl);
2117 }
2118 if (!ex_ctl->is_top() && do_replaced_nodes) {
2119 replaced_nodes_exception.apply(C, ex_ctl);
2120 }
2121 }
2122
2123
2124 //------------------------------increment_counter------------------------------
2125 // for statistics: increment a VM counter by 1
2126
2127 void GraphKit::increment_counter(address counter_addr) {
2128 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2129 increment_counter(adr1);
2130 }
2131
2132 void GraphKit::increment_counter(Node* counter_addr) {
2133 int adr_type = Compile::AliasIdxRaw;
2134 Node* ctrl = control();
2135 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered);
2136 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2137 store_to_memory(ctrl, counter_addr, incr, T_LONG, adr_type, MemNode::unordered);
2138 }
2139
2140
2141 //------------------------------uncommon_trap----------------------------------
2142 // Bail out to the interpreter in mid-method. Implemented by calling the
2143 // uncommon_trap blob. This helper function inserts a runtime call with the
2144 // right debug info.
2145 void GraphKit::uncommon_trap(int trap_request,
2146 ciKlass* klass, const char* comment,
2147 bool must_throw,
2148 bool keep_exact_action) {
2149 if (failing()) stop();
2150 if (stopped()) return; // trap reachable?
2151
2152 // Note: If ProfileTraps is true, and if a deopt. actually
2153 // occurs here, the runtime will make sure an MDO exists. There is
2154 // no need to call method()->ensure_method_data() at this point.
2155
2156 // Set the stack pointer to the right value for reexecution:
2157 set_sp(reexecute_sp());
2158
2159 #ifdef ASSERT
2160 if (!must_throw) {
2161 // Make sure the stack has at least enough depth to execute
2162 // the current bytecode.
2163 int inputs, ignored_depth;
2164 if (compute_stack_effects(inputs, ignored_depth)) {
2165 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2166 Bytecodes::name(java_bc()), sp(), inputs);
2167 }
2168 }
2169 #endif
2170
2171 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2172 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2173
2174 switch (action) {
2175 case Deoptimization::Action_maybe_recompile:
2176 case Deoptimization::Action_reinterpret:
2177 // Temporary fix for 6529811 to allow virtual calls to be sure they
2178 // get the chance to go from mono->bi->mega
2179 if (!keep_exact_action &&
2180 Deoptimization::trap_request_index(trap_request) < 0 &&
2181 too_many_recompiles(reason)) {
2182 // This BCI is causing too many recompilations.
2183 if (C->log() != NULL) {
2184 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2185 Deoptimization::trap_reason_name(reason),
2186 Deoptimization::trap_action_name(action));
2187 }
2188 action = Deoptimization::Action_none;
2189 trap_request = Deoptimization::make_trap_request(reason, action);
2190 } else {
2191 C->set_trap_can_recompile(true);
2192 }
2193 break;
2194 case Deoptimization::Action_make_not_entrant:
2195 C->set_trap_can_recompile(true);
2196 break;
2197 case Deoptimization::Action_none:
2198 case Deoptimization::Action_make_not_compilable:
2199 break;
2200 default:
2201 #ifdef ASSERT
2202 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2203 #endif
2204 break;
2205 }
2206
2207 if (TraceOptoParse) {
2208 char buf[100];
2209 tty->print_cr("Uncommon trap %s at bci:%d",
2210 Deoptimization::format_trap_request(buf, sizeof(buf),
2211 trap_request), bci());
2212 }
2213
2214 CompileLog* log = C->log();
2215 if (log != NULL) {
2216 int kid = (klass == NULL)? -1: log->identify(klass);
2217 log->begin_elem("uncommon_trap bci='%d'", bci());
2218 char buf[100];
2219 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2220 trap_request));
2221 if (kid >= 0) log->print(" klass='%d'", kid);
2222 if (comment != NULL) log->print(" comment='%s'", comment);
2223 log->end_elem();
2224 }
2225
2226 // Make sure any guarding test views this path as very unlikely
2227 Node *i0 = control()->in(0);
2228 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2229 IfNode *iff = i0->as_If();
2230 float f = iff->_prob; // Get prob
2231 if (control()->Opcode() == Op_IfTrue) {
2232 if (f > PROB_UNLIKELY_MAG(4))
2233 iff->_prob = PROB_MIN;
2234 } else {
2235 if (f < PROB_LIKELY_MAG(4))
2236 iff->_prob = PROB_MAX;
2237 }
2238 }
2239
2240 // Clear out dead values from the debug info.
2241 kill_dead_locals();
2242
2243 // Now insert the uncommon trap subroutine call
2244 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2245 const TypePtr* no_memory_effects = NULL;
2246 // Pass the index of the class to be loaded
2247 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2248 (must_throw ? RC_MUST_THROW : 0),
2249 OptoRuntime::uncommon_trap_Type(),
2250 call_addr, "uncommon_trap", no_memory_effects,
2251 intcon(trap_request));
2252 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2253 "must extract request correctly from the graph");
2254 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2255
2256 call->set_req(TypeFunc::ReturnAdr, returnadr());
2257 // The debug info is the only real input to this call.
2258
2259 // Halt-and-catch fire here. The above call should never return!
2260 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2261 PRODUCT_ONLY(COMMA /*reachable*/false));
2262 _gvn.set_type_bottom(halt);
2263 root()->add_req(halt);
2264
2265 stop_and_kill_map();
2266 }
2267
2268
2269 //--------------------------just_allocated_object------------------------------
2270 // Report the object that was just allocated.
2271 // It must be the case that there are no intervening safepoints.
2272 // We use this to determine if an object is so "fresh" that
2273 // it does not require card marks.
2274 Node* GraphKit::just_allocated_object(Node* current_control) {
2275 Node* ctrl = current_control;
2276 // Object::<init> is invoked after allocation, most of invoke nodes
2277 // will be reduced, but a region node is kept in parse time, we check
2278 // the pattern and skip the region node if it degraded to a copy.
2279 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 &&
2280 ctrl->as_Region()->is_copy()) {
2281 ctrl = ctrl->as_Region()->is_copy();
2282 }
2283 if (C->recent_alloc_ctl() == ctrl) {
2284 return C->recent_alloc_obj();
2285 }
2286 return NULL;
2287 }
2288
2289
2290 /**
2291 * Record profiling data exact_kls for Node n with the type system so
2292 * that it can propagate it (speculation)
2293 *
2294 * @param n node that the type applies to
2295 * @param exact_kls type from profiling
2296 * @param maybe_null did profiling see null?
2297 *
2298 * @return node with improved type
2299 */
2300 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2301 const Type* current_type = _gvn.type(n);
2302 assert(UseTypeSpeculation, "type speculation must be on");
2303
2304 const TypePtr* speculative = current_type->speculative();
2305
2306 // Should the klass from the profile be recorded in the speculative type?
2307 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2308 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2309 const TypeOopPtr* xtype = tklass->as_instance_type();
2310 assert(xtype->klass_is_exact(), "Should be exact");
2311 // Any reason to believe n is not null (from this profiling or a previous one)?
2312 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2313 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2314 // record the new speculative type's depth
2315 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2316 speculative = speculative->with_inline_depth(jvms()->depth());
2317 } else if (current_type->would_improve_ptr(ptr_kind)) {
2318 // Profiling report that null was never seen so we can change the
2319 // speculative type to non null ptr.
2320 if (ptr_kind == ProfileAlwaysNull) {
2321 speculative = TypePtr::NULL_PTR;
2322 } else {
2323 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2324 const TypePtr* ptr = TypePtr::NOTNULL;
2325 if (speculative != NULL) {
2326 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2327 } else {
2328 speculative = ptr;
2329 }
2330 }
2331 }
2332
2333 if (speculative != current_type->speculative()) {
2334 // Build a type with a speculative type (what we think we know
2335 // about the type but will need a guard when we use it)
2336 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2337 // We're changing the type, we need a new CheckCast node to carry
2338 // the new type. The new type depends on the control: what
2339 // profiling tells us is only valid from here as far as we can
2340 // tell.
2341 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2342 cast = _gvn.transform(cast);
2343 replace_in_map(n, cast);
2344 n = cast;
2345 }
2346
2347 return n;
2348 }
2349
2350 /**
2351 * Record profiling data from receiver profiling at an invoke with the
2352 * type system so that it can propagate it (speculation)
2353 *
2354 * @param n receiver node
2355 *
2356 * @return node with improved type
2357 */
2358 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2359 if (!UseTypeSpeculation) {
2360 return n;
2361 }
2362 ciKlass* exact_kls = profile_has_unique_klass();
2363 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2364 if ((java_bc() == Bytecodes::_checkcast ||
2365 java_bc() == Bytecodes::_instanceof ||
2366 java_bc() == Bytecodes::_aastore) &&
2367 method()->method_data()->is_mature()) {
2368 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2369 if (data != NULL) {
2370 if (java_bc() == Bytecodes::_aastore) {
2371 ciKlass* array_type = NULL;
2372 ciKlass* element_type = NULL;
2373 ProfilePtrKind element_ptr = ProfileMaybeNull;
2374 bool flat_array = true;
2375 bool null_free_array = true;
2376 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2377 exact_kls = element_type;
2378 ptr_kind = element_ptr;
2379 } else {
2380 if (!data->as_BitData()->null_seen()) {
2381 ptr_kind = ProfileNeverNull;
2382 } else {
2383 assert(data->is_ReceiverTypeData(), "bad profile data type");
2384 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2385 uint i = 0;
2386 for (; i < call->row_limit(); i++) {
2387 ciKlass* receiver = call->receiver(i);
2388 if (receiver != NULL) {
2389 break;
2390 }
2391 }
2392 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2393 }
2394 }
2395 }
2396 }
2397 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2398 }
2399
2400 /**
2401 * Record profiling data from argument profiling at an invoke with the
2402 * type system so that it can propagate it (speculation)
2403 *
2404 * @param dest_method target method for the call
2405 * @param bc what invoke bytecode is this?
2406 */
2407 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2408 if (!UseTypeSpeculation) {
2409 return;
2410 }
2411 const TypeFunc* tf = TypeFunc::make(dest_method);
2412 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2413 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2414 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2415 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2416 if (is_reference_type(targ->basic_type())) {
2417 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2418 ciKlass* better_type = NULL;
2419 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2420 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2421 }
2422 i++;
2423 }
2424 }
2425 }
2426
2427 /**
2428 * Record profiling data from parameter profiling at an invoke with
2429 * the type system so that it can propagate it (speculation)
2430 */
2431 void GraphKit::record_profiled_parameters_for_speculation() {
2432 if (!UseTypeSpeculation) {
2433 return;
2434 }
2435 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2436 if (_gvn.type(local(i))->isa_oopptr()) {
2437 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2438 ciKlass* better_type = NULL;
2439 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2440 record_profile_for_speculation(local(i), better_type, ptr_kind);
2441 }
2442 j++;
2443 }
2444 }
2445 }
2446
2447 /**
2448 * Record profiling data from return value profiling at an invoke with
2449 * the type system so that it can propagate it (speculation)
2450 */
2451 void GraphKit::record_profiled_return_for_speculation() {
2452 if (!UseTypeSpeculation) {
2453 return;
2454 }
2455 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2456 ciKlass* better_type = NULL;
2457 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2458 // If profiling reports a single type for the return value,
2459 // feed it to the type system so it can propagate it as a
2460 // speculative type
2461 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2462 }
2463 }
2464
2465 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2466 if (Matcher::strict_fp_requires_explicit_rounding) {
2467 // (Note: TypeFunc::make has a cache that makes this fast.)
2468 const TypeFunc* tf = TypeFunc::make(dest_method);
2469 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2470 for (int j = 0; j < nargs; j++) {
2471 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2472 if (targ->basic_type() == T_DOUBLE) {
2473 // If any parameters are doubles, they must be rounded before
2474 // the call, dstore_rounding does gvn.transform
2475 Node *arg = argument(j);
2476 arg = dstore_rounding(arg);
2477 set_argument(j, arg);
2478 }
2479 }
2480 }
2481 }
2482
2483 // rounding for strict float precision conformance
2484 Node* GraphKit::precision_rounding(Node* n) {
2485 if (Matcher::strict_fp_requires_explicit_rounding) {
2486 #ifdef IA32
2487 if (UseSSE == 0) {
2488 return _gvn.transform(new RoundFloatNode(0, n));
2489 }
2490 #else
2491 Unimplemented();
2492 #endif // IA32
2493 }
2494 return n;
2495 }
2496
2497 // rounding for strict double precision conformance
2498 Node* GraphKit::dprecision_rounding(Node *n) {
2499 if (Matcher::strict_fp_requires_explicit_rounding) {
2500 #ifdef IA32
2501 if (UseSSE < 2) {
2502 return _gvn.transform(new RoundDoubleNode(0, n));
2503 }
2504 #else
2505 Unimplemented();
2506 #endif // IA32
2507 }
2508 return n;
2509 }
2510
2511 // rounding for non-strict double stores
2512 Node* GraphKit::dstore_rounding(Node* n) {
2513 if (Matcher::strict_fp_requires_explicit_rounding) {
2514 #ifdef IA32
2515 if (UseSSE < 2) {
2516 return _gvn.transform(new RoundDoubleNode(0, n));
2517 }
2518 #else
2519 Unimplemented();
2520 #endif // IA32
2521 }
2522 return n;
2523 }
2524
2525 //=============================================================================
2526 // Generate a fast path/slow path idiom. Graph looks like:
2527 // [foo] indicates that 'foo' is a parameter
2528 //
2529 // [in] NULL
2530 // \ /
2531 // CmpP
2532 // Bool ne
2533 // If
2534 // / \
2535 // True False-<2>
2536 // / |
2537 // / cast_not_null
2538 // Load | | ^
2539 // [fast_test] | |
2540 // gvn to opt_test | |
2541 // / \ | <1>
2542 // True False |
2543 // | \\ |
2544 // [slow_call] \[fast_result]
2545 // Ctl Val \ \
2546 // | \ \
2547 // Catch <1> \ \
2548 // / \ ^ \ \
2549 // Ex No_Ex | \ \
2550 // | \ \ | \ <2> \
2551 // ... \ [slow_res] | | \ [null_result]
2552 // \ \--+--+--- | |
2553 // \ | / \ | /
2554 // --------Region Phi
2555 //
2556 //=============================================================================
2557 // Code is structured as a series of driver functions all called 'do_XXX' that
2558 // call a set of helper functions. Helper functions first, then drivers.
2559
2560 //------------------------------null_check_oop---------------------------------
2561 // Null check oop. Set null-path control into Region in slot 3.
2562 // Make a cast-not-nullness use the other not-null control. Return cast.
2563 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2564 bool never_see_null,
2565 bool safe_for_replace,
2566 bool speculative) {
2567 // Initial NULL check taken path
2568 (*null_control) = top();
2569 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2570
2571 // Generate uncommon_trap:
2572 if (never_see_null && (*null_control) != top()) {
2573 // If we see an unexpected null at a check-cast we record it and force a
2574 // recompile; the offending check-cast will be compiled to handle NULLs.
2575 // If we see more than one offending BCI, then all checkcasts in the
2576 // method will be compiled to handle NULLs.
2577 PreserveJVMState pjvms(this);
2578 set_control(*null_control);
2579 replace_in_map(value, null());
2580 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2581 uncommon_trap(reason,
2582 Deoptimization::Action_make_not_entrant);
2583 (*null_control) = top(); // NULL path is dead
2584 }
2585 if ((*null_control) == top() && safe_for_replace) {
2586 replace_in_map(value, cast);
2587 }
2588
2589 // Cast away null-ness on the result
2590 return cast;
2591 }
2592
2593 //------------------------------opt_iff----------------------------------------
2594 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2595 // Return slow-path control.
2596 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2597 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2598
2599 // Fast path taken; set region slot 2
2600 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2601 region->init_req(2,fast_taken); // Capture fast-control
2602
2603 // Fast path not-taken, i.e. slow path
2604 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2605 return slow_taken;
2606 }
2607
2608 //-----------------------------make_runtime_call-------------------------------
2609 Node* GraphKit::make_runtime_call(int flags,
2610 const TypeFunc* call_type, address call_addr,
2611 const char* call_name,
2612 const TypePtr* adr_type,
2613 // The following parms are all optional.
2614 // The first NULL ends the list.
2615 Node* parm0, Node* parm1,
2616 Node* parm2, Node* parm3,
2617 Node* parm4, Node* parm5,
2618 Node* parm6, Node* parm7,
2619 Node* parm8) {
2620 assert(call_addr != NULL, "must not call NULL targets");
2621
2622 // Slow-path call
2623 bool is_leaf = !(flags & RC_NO_LEAF);
2624 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2625 if (call_name == NULL) {
2626 assert(!is_leaf, "must supply name for leaf");
2627 call_name = OptoRuntime::stub_name(call_addr);
2628 }
2629 CallNode* call;
2630 if (!is_leaf) {
2631 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2632 } else if (flags & RC_NO_FP) {
2633 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2634 } else if (flags & RC_VECTOR){
2635 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2636 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2637 } else {
2638 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2639 }
2640
2641 // The following is similar to set_edges_for_java_call,
2642 // except that the memory effects of the call are restricted to AliasIdxRaw.
2643
2644 // Slow path call has no side-effects, uses few values
2645 bool wide_in = !(flags & RC_NARROW_MEM);
2646 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2647
2648 Node* prev_mem = NULL;
2649 if (wide_in) {
2650 prev_mem = set_predefined_input_for_runtime_call(call);
2651 } else {
2652 assert(!wide_out, "narrow in => narrow out");
2653 Node* narrow_mem = memory(adr_type);
2654 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2655 }
2656
2657 // Hook each parm in order. Stop looking at the first NULL.
2658 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2659 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2660 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2661 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2662 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2663 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2664 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2665 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2666 if (parm8 != NULL) { call->init_req(TypeFunc::Parms+8, parm8);
2667 /* close each nested if ===> */ } } } } } } } } }
2668 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2669
2670 if (!is_leaf) {
2671 // Non-leaves can block and take safepoints:
2672 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2673 }
2674 // Non-leaves can throw exceptions:
2675 if (has_io) {
2676 call->set_req(TypeFunc::I_O, i_o());
2677 }
2678
2679 if (flags & RC_UNCOMMON) {
2680 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2681 // (An "if" probability corresponds roughly to an unconditional count.
2682 // Sort of.)
2683 call->set_cnt(PROB_UNLIKELY_MAG(4));
2684 }
2685
2686 Node* c = _gvn.transform(call);
2687 assert(c == call, "cannot disappear");
2688
2689 if (wide_out) {
2690 // Slow path call has full side-effects.
2691 set_predefined_output_for_runtime_call(call);
2692 } else {
2693 // Slow path call has few side-effects, and/or sets few values.
2694 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2695 }
2696
2697 if (has_io) {
2698 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2699 }
2700 return call;
2701
2702 }
2703
2704 // i2b
2705 Node* GraphKit::sign_extend_byte(Node* in) {
2706 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2707 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2708 }
2709
2710 // i2s
2711 Node* GraphKit::sign_extend_short(Node* in) {
2712 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2713 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2714 }
2715
2716 //-----------------------------make_native_call-------------------------------
2717 Node* GraphKit::make_native_call(address call_addr, const TypeFunc* call_type, uint nargs, ciNativeEntryPoint* nep) {
2718 // Select just the actual call args to pass on
2719 // [MethodHandle fallback, long addr, HALF addr, ... args , NativeEntryPoint nep]
2720 // | |
2721 // V V
2722 // [ ... args ]
2723 uint n_filtered_args = nargs - 4; // -fallback, -addr (2), -nep;
2724 ResourceMark rm;
2725 Node** argument_nodes = NEW_RESOURCE_ARRAY(Node*, n_filtered_args);
2726 const Type** arg_types = TypeTuple::fields(n_filtered_args);
2727 GrowableArray<VMReg> arg_regs(C->comp_arena(), n_filtered_args, n_filtered_args, VMRegImpl::Bad());
2728
2729 VMReg* argRegs = nep->argMoves();
2730 {
2731 for (uint vm_arg_pos = 0, java_arg_read_pos = 0;
2732 vm_arg_pos < n_filtered_args; vm_arg_pos++) {
2733 uint vm_unfiltered_arg_pos = vm_arg_pos + 3; // +3 to skip fallback handle argument and addr (2 since long)
2734 Node* node = argument(vm_unfiltered_arg_pos);
2735 const Type* type = call_type->domain_sig()->field_at(TypeFunc::Parms + vm_unfiltered_arg_pos);
2736 VMReg reg = type == Type::HALF
2737 ? VMRegImpl::Bad()
2738 : argRegs[java_arg_read_pos++];
2739
2740 argument_nodes[vm_arg_pos] = node;
2741 arg_types[TypeFunc::Parms + vm_arg_pos] = type;
2742 arg_regs.at_put(vm_arg_pos, reg);
2743 }
2744 }
2745
2746 uint n_returns = call_type->range_sig()->cnt() - TypeFunc::Parms;
2747 GrowableArray<VMReg> ret_regs(C->comp_arena(), n_returns, n_returns, VMRegImpl::Bad());
2748 const Type** ret_types = TypeTuple::fields(n_returns);
2749
2750 VMReg* retRegs = nep->returnMoves();
2751 {
2752 for (uint vm_ret_pos = 0, java_ret_read_pos = 0;
2753 vm_ret_pos < n_returns; vm_ret_pos++) { // 0 or 1
2754 const Type* type = call_type->range_sig()->field_at(TypeFunc::Parms + vm_ret_pos);
2755 VMReg reg = type == Type::HALF
2756 ? VMRegImpl::Bad()
2757 : retRegs[java_ret_read_pos++];
2758
2759 ret_regs.at_put(vm_ret_pos, reg);
2760 ret_types[TypeFunc::Parms + vm_ret_pos] = type;
2761 }
2762 }
2763
2764 const TypeFunc* new_call_type = TypeFunc::make(
2765 TypeTuple::make(TypeFunc::Parms + n_filtered_args, arg_types),
2766 TypeTuple::make(TypeFunc::Parms + n_returns, ret_types)
2767 );
2768
2769 if (nep->need_transition()) {
2770 RuntimeStub* invoker = SharedRuntime::make_native_invoker(call_addr,
2771 nep->shadow_space(),
2772 arg_regs, ret_regs);
2773 if (invoker == NULL) {
2774 C->record_failure("native invoker not implemented on this platform");
2775 return NULL;
2776 }
2777 C->add_native_invoker(invoker);
2778 call_addr = invoker->code_begin();
2779 }
2780 assert(call_addr != NULL, "sanity");
2781
2782 CallNativeNode* call = new CallNativeNode(new_call_type, call_addr, nep->name(), TypePtr::BOTTOM,
2783 arg_regs,
2784 ret_regs,
2785 nep->shadow_space(),
2786 nep->need_transition());
2787
2788 if (call->_need_transition) {
2789 add_safepoint_edges(call);
2790 }
2791
2792 set_predefined_input_for_runtime_call(call);
2793
2794 for (uint i = 0; i < n_filtered_args; i++) {
2795 call->init_req(i + TypeFunc::Parms, argument_nodes[i]);
2796 }
2797
2798 Node* c = gvn().transform(call);
2799 assert(c == call, "cannot disappear");
2800
2801 set_predefined_output_for_runtime_call(call);
2802
2803 Node* ret;
2804 if (method() == NULL || method()->return_type()->basic_type() == T_VOID) {
2805 ret = top();
2806 } else {
2807 ret = gvn().transform(new ProjNode(call, TypeFunc::Parms));
2808 // Unpack native results if needed
2809 // Need this method type since it's unerased
2810 switch (nep->method_type()->rtype()->basic_type()) {
2811 case T_CHAR:
2812 ret = _gvn.transform(new AndINode(ret, _gvn.intcon(0xFFFF)));
2813 break;
2814 case T_BYTE:
2815 ret = sign_extend_byte(ret);
2816 break;
2817 case T_SHORT:
2818 ret = sign_extend_short(ret);
2819 break;
2820 default: // do nothing
2821 break;
2822 }
2823 }
2824
2825 push_node(method()->return_type()->basic_type(), ret);
2826
2827 return call;
2828 }
2829
2830 //------------------------------merge_memory-----------------------------------
2831 // Merge memory from one path into the current memory state.
2832 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2833 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2834 Node* old_slice = mms.force_memory();
2835 Node* new_slice = mms.memory2();
2836 if (old_slice != new_slice) {
2837 PhiNode* phi;
2838 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2839 if (mms.is_empty()) {
2840 // clone base memory Phi's inputs for this memory slice
2841 assert(old_slice == mms.base_memory(), "sanity");
2842 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2843 _gvn.set_type(phi, Type::MEMORY);
2844 for (uint i = 1; i < phi->req(); i++) {
2845 phi->init_req(i, old_slice->in(i));
2846 }
2847 } else {
2848 phi = old_slice->as_Phi(); // Phi was generated already
2849 }
2850 } else {
2851 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2852 _gvn.set_type(phi, Type::MEMORY);
2853 }
2854 phi->set_req(new_path, new_slice);
2855 mms.set_memory(phi);
2856 }
2857 }
2858 }
2859
2860 //------------------------------make_slow_call_ex------------------------------
2861 // Make the exception handler hookups for the slow call
2862 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2863 if (stopped()) return;
2864
2865 // Make a catch node with just two handlers: fall-through and catch-all
2866 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2867 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2868 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2869 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2870
2871 { PreserveJVMState pjvms(this);
2872 set_control(excp);
2873 set_i_o(i_o);
2874
2875 if (excp != top()) {
2876 if (deoptimize) {
2877 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2878 uncommon_trap(Deoptimization::Reason_unhandled,
2879 Deoptimization::Action_none);
2880 } else {
2881 // Create an exception state also.
2882 // Use an exact type if the caller has a specific exception.
2883 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2884 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2885 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2886 }
2887 }
2888 }
2889
2890 // Get the no-exception control from the CatchNode.
2891 set_control(norm);
2892 }
2893
2894 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2895 Node* cmp = NULL;
2896 switch(bt) {
2897 case T_INT: cmp = new CmpINode(in1, in2); break;
2898 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2899 default: fatal("unexpected comparison type %s", type2name(bt));
2900 }
2901 gvn.transform(cmp);
2902 Node* bol = gvn.transform(new BoolNode(cmp, test));
2903 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2904 gvn.transform(iff);
2905 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2906 return iff;
2907 }
2908
2909 //-------------------------------gen_subtype_check-----------------------------
2910 // Generate a subtyping check. Takes as input the subtype and supertype.
2911 // Returns 2 values: sets the default control() to the true path and returns
2912 // the false path. Only reads invariant memory; sets no (visible) memory.
2913 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2914 // but that's not exposed to the optimizer. This call also doesn't take in an
2915 // Object; if you wish to check an Object you need to load the Object's class
2916 // prior to coming here.
2917 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn) {
2918 Compile* C = gvn.C;
2919 if ((*ctrl)->is_top()) {
2920 return C->top();
2921 }
2922
2923 // Fast check for identical types, perhaps identical constants.
2924 // The types can even be identical non-constants, in cases
2925 // involving Array.newInstance, Object.clone, etc.
2926 if (subklass == superklass)
2927 return C->top(); // false path is dead; no test needed.
2928
2929 if (gvn.type(superklass)->singleton()) {
2930 ciKlass* superk = gvn.type(superklass)->is_klassptr()->klass();
2931 ciKlass* subk = gvn.type(subklass)->is_klassptr()->klass();
2932
2933 // In the common case of an exact superklass, try to fold up the
2934 // test before generating code. You may ask, why not just generate
2935 // the code and then let it fold up? The answer is that the generated
2936 // code will necessarily include null checks, which do not always
2937 // completely fold away. If they are also needless, then they turn
2938 // into a performance loss. Example:
2939 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2940 // Here, the type of 'fa' is often exact, so the store check
2941 // of fa[1]=x will fold up, without testing the nullness of x.
2942 switch (C->static_subtype_check(superk, subk)) {
2943 case Compile::SSC_always_false:
2944 {
2945 Node* always_fail = *ctrl;
2946 *ctrl = gvn.C->top();
2947 return always_fail;
2948 }
2949 case Compile::SSC_always_true:
2950 return C->top();
2951 case Compile::SSC_easy_test:
2952 {
2953 // Just do a direct pointer compare and be done.
2954 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2955 *ctrl = gvn.transform(new IfTrueNode(iff));
2956 return gvn.transform(new IfFalseNode(iff));
2957 }
2958 case Compile::SSC_full_test:
2959 break;
2960 default:
2961 ShouldNotReachHere();
2962 }
2963 }
2964
2965 // %%% Possible further optimization: Even if the superklass is not exact,
2966 // if the subklass is the unique subtype of the superklass, the check
2967 // will always succeed. We could leave a dependency behind to ensure this.
2968
2969 // First load the super-klass's check-offset
2970 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2971 Node* m = C->immutable_memory();
2972 Node *chk_off = gvn.transform(new LoadINode(NULL, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2973 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2974 bool might_be_cache = (gvn.find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2975
2976 // Load from the sub-klass's super-class display list, or a 1-word cache of
2977 // the secondary superclass list, or a failing value with a sentinel offset
2978 // if the super-klass is an interface or exceptionally deep in the Java
2979 // hierarchy and we have to scan the secondary superclass list the hard way.
2980 // Worst-case type is a little odd: NULL is allowed as a result (usually
2981 // klass loads can never produce a NULL).
2982 Node *chk_off_X = chk_off;
2983 #ifdef _LP64
2984 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2985 #endif
2986 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2987 // For some types like interfaces the following loadKlass is from a 1-word
2988 // cache which is mutable so can't use immutable memory. Other
2989 // types load from the super-class display table which is immutable.
2990 Node *kmem = C->immutable_memory();
2991 // secondary_super_cache is not immutable but can be treated as such because:
2992 // - no ideal node writes to it in a way that could cause an
2993 // incorrect/missed optimization of the following Load.
2994 // - it's a cache so, worse case, not reading the latest value
2995 // wouldn't cause incorrect execution
2996 if (might_be_cache && mem != NULL) {
2997 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
2998 }
2999 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, NULL, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
3000
3001 // Compile speed common case: ARE a subtype and we canNOT fail
3002 if( superklass == nkls )
3003 return C->top(); // false path is dead; no test needed.
3004
3005 // See if we get an immediate positive hit. Happens roughly 83% of the
3006 // time. Test to see if the value loaded just previously from the subklass
3007 // is exactly the superklass.
3008 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3009 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
3010 *ctrl = gvn.transform(new IfFalseNode(iff1));
3011
3012 // Compile speed common case: Check for being deterministic right now. If
3013 // chk_off is a constant and not equal to cacheoff then we are NOT a
3014 // subklass. In this case we need exactly the 1 test above and we can
3015 // return those results immediately.
3016 if (!might_be_cache) {
3017 Node* not_subtype_ctrl = *ctrl;
3018 *ctrl = iftrue1; // We need exactly the 1 test above
3019 return not_subtype_ctrl;
3020 }
3021
3022 // Gather the various success & failures here
3023 RegionNode *r_ok_subtype = new RegionNode(4);
3024 gvn.record_for_igvn(r_ok_subtype);
3025 RegionNode *r_not_subtype = new RegionNode(3);
3026 gvn.record_for_igvn(r_not_subtype);
3027
3028 r_ok_subtype->init_req(1, iftrue1);
3029
3030 // Check for immediate negative hit. Happens roughly 11% of the time (which
3031 // is roughly 63% of the remaining cases). Test to see if the loaded
3032 // check-offset points into the subklass display list or the 1-element
3033 // cache. If it points to the display (and NOT the cache) and the display
3034 // missed then it's not a subtype.
3035 Node *cacheoff = gvn.intcon(cacheoff_con);
3036 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3037 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3038 *ctrl = gvn.transform(new IfFalseNode(iff2));
3039
3040 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3041 // No performance impact (too rare) but allows sharing of secondary arrays
3042 // which has some footprint reduction.
3043 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3044 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3045 *ctrl = gvn.transform(new IfFalseNode(iff3));
3046
3047 // -- Roads not taken here: --
3048 // We could also have chosen to perform the self-check at the beginning
3049 // of this code sequence, as the assembler does. This would not pay off
3050 // the same way, since the optimizer, unlike the assembler, can perform
3051 // static type analysis to fold away many successful self-checks.
3052 // Non-foldable self checks work better here in second position, because
3053 // the initial primary superclass check subsumes a self-check for most
3054 // types. An exception would be a secondary type like array-of-interface,
3055 // which does not appear in its own primary supertype display.
3056 // Finally, we could have chosen to move the self-check into the
3057 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3058 // dependent manner. But it is worthwhile to have the check here,
3059 // where it can be perhaps be optimized. The cost in code space is
3060 // small (register compare, branch).
3061
3062 // Now do a linear scan of the secondary super-klass array. Again, no real
3063 // performance impact (too rare) but it's gotta be done.
3064 // Since the code is rarely used, there is no penalty for moving it
3065 // out of line, and it can only improve I-cache density.
3066 // The decision to inline or out-of-line this final check is platform
3067 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3068 Node* psc = gvn.transform(
3069 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3070
3071 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3072 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3073 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3074
3075 // Return false path; set default control to true path.
3076 *ctrl = gvn.transform(r_ok_subtype);
3077 return gvn.transform(r_not_subtype);
3078 }
3079
3080 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3081 const Type* sub_t = _gvn.type(obj_or_subklass);
3082 if (sub_t->isa_inlinetype()) {
3083 obj_or_subklass = makecon(TypeKlassPtr::make(sub_t->inline_klass()));
3084 }
3085 bool expand_subtype_check = C->post_loop_opts_phase() || // macro node expansion is over
3086 ExpandSubTypeCheckAtParseTime; // forced expansion
3087 if (expand_subtype_check) {
3088 MergeMemNode* mem = merged_memory();
3089 Node* ctrl = control();
3090 Node* subklass = obj_or_subklass;
3091 if (!sub_t->isa_klassptr() && !sub_t->isa_inlinetype()) {
3092 subklass = load_object_klass(obj_or_subklass);
3093 }
3094 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn);
3095 set_control(ctrl);
3096 return n;
3097 }
3098
3099 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass));
3100 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3101 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3102 set_control(_gvn.transform(new IfTrueNode(iff)));
3103 return _gvn.transform(new IfFalseNode(iff));
3104 }
3105
3106 // Profile-driven exact type check:
3107 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3108 float prob, Node* *casted_receiver) {
3109 assert(!klass->is_interface(), "no exact type check on interfaces");
3110 Node* fail = top();
3111 const Type* rec_t = _gvn.type(receiver);
3112 if (rec_t->isa_inlinetype()) {
3113 if (klass->equals(rec_t->inline_klass())) {
3114 (*casted_receiver) = receiver; // Always passes
3115 } else {
3116 (*casted_receiver) = top(); // Always fails
3117 fail = control();
3118 set_control(top());
3119 }
3120 return fail;
3121 }
3122 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
3123 Node* recv_klass = load_object_klass(receiver);
3124 fail = type_check(recv_klass, tklass, prob);
3125
3126 if (!stopped()) {
3127 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3128 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3129 assert(recv_xtype->klass_is_exact(), "");
3130
3131 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3132 // Subsume downstream occurrences of receiver with a cast to
3133 // recv_xtype, since now we know what the type will be.
3134 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3135 Node* res = _gvn.transform(cast);
3136 if (recv_xtype->is_inlinetypeptr()) {
3137 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3138 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass())->as_InlineTypeBase()->as_ptr(&gvn());
3139 }
3140 (*casted_receiver) = res;
3141 // (User must make the replace_in_map call.)
3142 }
3143 }
3144
3145 return fail;
3146 }
3147
3148 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3149 float prob) {
3150 Node* want_klass = makecon(tklass);
3151 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3152 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3153 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3154 set_control(_gvn.transform(new IfTrueNode (iff)));
3155 Node* fail = _gvn.transform(new IfFalseNode(iff));
3156 return fail;
3157 }
3158
3159 //------------------------------subtype_check_receiver-------------------------
3160 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3161 Node** casted_receiver) {
3162 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
3163 Node* want_klass = makecon(tklass);
3164
3165 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3166
3167 // Ignore interface type information until interface types are properly tracked.
3168 if (!stopped() && !klass->is_interface()) {
3169 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3170 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3171 if (receiver_type != NULL && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3172 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
3173 (*casted_receiver) = _gvn.transform(cast);
3174 }
3175 }
3176
3177 return slow_ctl;
3178 }
3179
3180 //------------------------------seems_never_null-------------------------------
3181 // Use null_seen information if it is available from the profile.
3182 // If we see an unexpected null at a type check we record it and force a
3183 // recompile; the offending check will be recompiled to handle NULLs.
3184 // If we see several offending BCIs, then all checks in the
3185 // method will be recompiled.
3186 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3187 speculating = !_gvn.type(obj)->speculative_maybe_null();
3188 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3189 if (UncommonNullCast // Cutout for this technique
3190 && obj != null() // And not the -Xcomp stupid case?
3191 && !too_many_traps(reason)
3192 ) {
3193 if (speculating) {
3194 return true;
3195 }
3196 if (data == NULL)
3197 // Edge case: no mature data. Be optimistic here.
3198 return true;
3199 // If the profile has not seen a null, assume it won't happen.
3200 assert(java_bc() == Bytecodes::_checkcast ||
3201 java_bc() == Bytecodes::_instanceof ||
3202 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3203 if (java_bc() == Bytecodes::_aastore) {
3204 return ((ciArrayLoadStoreData*)data->as_ArrayLoadStoreData())->element()->ptr_kind() == ProfileNeverNull;
3205 }
3206 return !data->as_BitData()->null_seen();
3207 }
3208 speculating = false;
3209 return false;
3210 }
3211
3212 void GraphKit::guard_klass_being_initialized(Node* klass) {
3213 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3214 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3215 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
3216 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3217 T_BYTE, MemNode::unordered);
3218 init_state = _gvn.transform(init_state);
3219
3220 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3221
3222 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3223 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3224
3225 { BuildCutout unless(this, tst, PROB_MAX);
3226 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3227 }
3228 }
3229
3230 void GraphKit::guard_init_thread(Node* klass) {
3231 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3232 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3233
3234 Node* init_thread = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
3235 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3236 T_ADDRESS, MemNode::unordered);
3237 init_thread = _gvn.transform(init_thread);
3238
3239 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3240
3241 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3242 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3243
3244 { BuildCutout unless(this, tst, PROB_MAX);
3245 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3246 }
3247 }
3248
3249 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3250 if (ik->is_being_initialized()) {
3251 if (C->needs_clinit_barrier(ik, context)) {
3252 Node* klass = makecon(TypeKlassPtr::make(ik));
3253 guard_klass_being_initialized(klass);
3254 guard_init_thread(klass);
3255 insert_mem_bar(Op_MemBarCPUOrder);
3256 }
3257 } else if (ik->is_initialized()) {
3258 return; // no barrier needed
3259 } else {
3260 uncommon_trap(Deoptimization::Reason_uninitialized,
3261 Deoptimization::Action_reinterpret,
3262 NULL);
3263 }
3264 }
3265
3266 //------------------------maybe_cast_profiled_receiver-------------------------
3267 // If the profile has seen exactly one type, narrow to exactly that type.
3268 // Subsequent type checks will always fold up.
3269 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3270 ciKlass* require_klass,
3271 ciKlass* spec_klass,
3272 bool safe_for_replace) {
3273 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
3274
3275 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
3276
3277 // Make sure we haven't already deoptimized from this tactic.
3278 if (too_many_traps_or_recompiles(reason))
3279 return NULL;
3280
3281 // (No, this isn't a call, but it's enough like a virtual call
3282 // to use the same ciMethod accessor to get the profile info...)
3283 // If we have a speculative type use it instead of profiling (which
3284 // may not help us)
3285 ciKlass* exact_kls = spec_klass;
3286 if (exact_kls == NULL) {
3287 if (java_bc() == Bytecodes::_aastore) {
3288 ciKlass* array_type = NULL;
3289 ciKlass* element_type = NULL;
3290 ProfilePtrKind element_ptr = ProfileMaybeNull;
3291 bool flat_array = true;
3292 bool null_free_array = true;
3293 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3294 exact_kls = element_type;
3295 } else {
3296 exact_kls = profile_has_unique_klass();
3297 }
3298 }
3299 if (exact_kls != NULL) {// no cast failures here
3300 if (require_klass == NULL ||
3301 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
3302 // If we narrow the type to match what the type profile sees or
3303 // the speculative type, we can then remove the rest of the
3304 // cast.
3305 // This is a win, even if the exact_kls is very specific,
3306 // because downstream operations, such as method calls,
3307 // will often benefit from the sharper type.
3308 Node* exact_obj = not_null_obj; // will get updated in place...
3309 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3310 &exact_obj);
3311 { PreserveJVMState pjvms(this);
3312 set_control(slow_ctl);
3313 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3314 }
3315 if (safe_for_replace) {
3316 replace_in_map(not_null_obj, exact_obj);
3317 }
3318 return exact_obj;
3319 }
3320 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3321 }
3322
3323 return NULL;
3324 }
3325
3326 /**
3327 * Cast obj to type and emit guard unless we had too many traps here
3328 * already
3329 *
3330 * @param obj node being casted
3331 * @param type type to cast the node to
3332 * @param not_null true if we know node cannot be null
3333 */
3334 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3335 ciKlass* type,
3336 bool not_null) {
3337 if (stopped()) {
3338 return obj;
3339 }
3340
3341 // type == NULL if profiling tells us this object is always null
3342 if (type != NULL) {
3343 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3344 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3345
3346 if (!too_many_traps_or_recompiles(null_reason) &&
3347 !too_many_traps_or_recompiles(class_reason)) {
3348 Node* not_null_obj = NULL;
3349 // not_null is true if we know the object is not null and
3350 // there's no need for a null check
3351 if (!not_null) {
3352 Node* null_ctl = top();
3353 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3354 assert(null_ctl->is_top(), "no null control here");
3355 } else {
3356 not_null_obj = obj;
3357 }
3358
3359 Node* exact_obj = not_null_obj;
3360 ciKlass* exact_kls = type;
3361 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3362 &exact_obj);
3363 {
3364 PreserveJVMState pjvms(this);
3365 set_control(slow_ctl);
3366 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3367 }
3368 replace_in_map(not_null_obj, exact_obj);
3369 obj = exact_obj;
3370 }
3371 } else {
3372 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3373 Node* exact_obj = null_assert(obj);
3374 replace_in_map(obj, exact_obj);
3375 obj = exact_obj;
3376 }
3377 }
3378 return obj;
3379 }
3380
3381 //-------------------------------gen_instanceof--------------------------------
3382 // Generate an instance-of idiom. Used by both the instance-of bytecode
3383 // and the reflective instance-of call.
3384 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3385 kill_dead_locals(); // Benefit all the uncommon traps
3386 assert( !stopped(), "dead parse path should be checked in callers" );
3387 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3388 "must check for not-null not-dead klass in callers");
3389
3390 // Make the merge point
3391 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3392 RegionNode* region = new RegionNode(PATH_LIMIT);
3393 Node* phi = new PhiNode(region, TypeInt::BOOL);
3394 C->set_has_split_ifs(true); // Has chance for split-if optimization
3395
3396 ciProfileData* data = NULL;
3397 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3398 data = method()->method_data()->bci_to_data(bci());
3399 }
3400 bool speculative_not_null = false;
3401 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3402 && seems_never_null(obj, data, speculative_not_null));
3403 bool is_value = obj->is_InlineType();
3404
3405 // Null check; get casted pointer; set region slot 3
3406 Node* null_ctl = top();
3407 Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3408
3409 // If not_null_obj is dead, only null-path is taken
3410 if (stopped()) { // Doing instance-of on a NULL?
3411 set_control(null_ctl);
3412 return intcon(0);
3413 }
3414 region->init_req(_null_path, null_ctl);
3415 phi ->init_req(_null_path, intcon(0)); // Set null path value
3416 if (null_ctl == top()) {
3417 // Do this eagerly, so that pattern matches like is_diamond_phi
3418 // will work even during parsing.
3419 assert(_null_path == PATH_LIMIT-1, "delete last");
3420 region->del_req(_null_path);
3421 phi ->del_req(_null_path);
3422 }
3423
3424 // Do we know the type check always succeed?
3425 if (!is_value) {
3426 bool known_statically = false;
3427 if (_gvn.type(superklass)->singleton()) {
3428 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3429 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3430 if (subk != NULL && subk->is_loaded()) {
3431 int static_res = C->static_subtype_check(superk, subk);
3432 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3433 }
3434 }
3435
3436 if (!known_statically) {
3437 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3438 // We may not have profiling here or it may not help us. If we
3439 // have a speculative type use it to perform an exact cast.
3440 ciKlass* spec_obj_type = obj_type->speculative_type();
3441 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3442 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3443 if (stopped()) { // Profile disagrees with this path.
3444 set_control(null_ctl); // Null is the only remaining possibility.
3445 return intcon(0);
3446 }
3447 if (cast_obj != NULL) {
3448 not_null_obj = cast_obj;
3449 is_value = not_null_obj->is_InlineType();
3450 }
3451 }
3452 }
3453 }
3454
3455 // Generate the subtype check
3456 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3457
3458 // Plug in the success path to the general merge in slot 1.
3459 region->init_req(_obj_path, control());
3460 phi ->init_req(_obj_path, intcon(1));
3461
3462 // Plug in the failing path to the general merge in slot 2.
3463 region->init_req(_fail_path, not_subtype_ctrl);
3464 phi ->init_req(_fail_path, intcon(0));
3465
3466 // Return final merged results
3467 set_control( _gvn.transform(region) );
3468 record_for_igvn(region);
3469
3470 // If we know the type check always succeeds then we don't use the
3471 // profiling data at this bytecode. Don't lose it, feed it to the
3472 // type system as a speculative type.
3473 if (safe_for_replace && !is_value) {
3474 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3475 replace_in_map(obj, casted_obj);
3476 }
3477
3478 return _gvn.transform(phi);
3479 }
3480
3481 //-------------------------------gen_checkcast---------------------------------
3482 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3483 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3484 // uncommon-trap paths work. Adjust stack after this call.
3485 // If failure_control is supplied and not null, it is filled in with
3486 // the control edge for the cast failure. Otherwise, an appropriate
3487 // uncommon trap or exception is thrown.
3488 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool null_free) {
3489 kill_dead_locals(); // Benefit all the uncommon traps
3490 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr();
3491 const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass());
3492 bool safe_for_replace = (failure_control == NULL);
3493 bool from_inline = obj->is_InlineType();
3494 assert(!null_free || toop->is_inlinetypeptr(), "must be an inline type pointer");
3495
3496 // Fast cutout: Check the case that the cast is vacuously true.
3497 // This detects the common cases where the test will short-circuit
3498 // away completely. We do this before we perform the null check,
3499 // because if the test is going to turn into zero code, we don't
3500 // want a residual null check left around. (Causes a slowdown,
3501 // for example, in some objArray manipulations, such as a[i]=a[j].)
3502 if (tk->singleton()) {
3503 ciKlass* klass = NULL;
3504 if (obj->is_InlineTypeBase()) {
3505 klass = _gvn.type(obj)->inline_klass();
3506 } else {
3507 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3508 if (objtp != NULL) {
3509 klass = objtp->klass();
3510 }
3511 }
3512 if (klass != NULL) {
3513 switch (C->static_subtype_check(tk->klass(), klass)) {
3514 case Compile::SSC_always_true:
3515 // If we know the type check always succeed then we don't use
3516 // the profiling data at this bytecode. Don't lose it, feed it
3517 // to the type system as a speculative type.
3518 if (!from_inline) {
3519 obj = record_profiled_receiver_for_speculation(obj);
3520 if (null_free) {
3521 assert(safe_for_replace, "must be");
3522 obj = null_check(obj);
3523 }
3524 assert(stopped() || !toop->is_inlinetypeptr() ||
3525 obj->is_InlineTypeBase(), "should have been scalarized");
3526 }
3527 return obj;
3528 case Compile::SSC_always_false:
3529 if (from_inline || null_free) {
3530 if (!from_inline) {
3531 assert(safe_for_replace, "must be");
3532 null_check(obj);
3533 }
3534 // Inline type is null-free. Always throw an exception.
3535 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(klass)));
3536 return top();
3537 } else {
3538 // It needs a null check because a null will *pass* the cast check.
3539 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3540 if (!objtp->maybe_null()) {
3541 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(objtp->klass())));
3542 return top();
3543 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3544 return null_assert(obj);
3545 }
3546 break; // Fall through to full check
3547 }
3548 }
3549 }
3550 }
3551
3552 ciProfileData* data = NULL;
3553 if (failure_control == NULL) { // use MDO in regular case only
3554 assert(java_bc() == Bytecodes::_aastore ||
3555 java_bc() == Bytecodes::_checkcast,
3556 "interpreter profiles type checks only for these BCs");
3557 if (method()->method_data()->is_mature()) {
3558 data = method()->method_data()->bci_to_data(bci());
3559 }
3560 }
3561
3562 // Make the merge point
3563 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3564 RegionNode* region = new RegionNode(PATH_LIMIT);
3565 Node* phi = new PhiNode(region, toop);
3566 _gvn.set_type(region, Type::CONTROL);
3567 _gvn.set_type(phi, toop);
3568
3569 C->set_has_split_ifs(true); // Has chance for split-if optimization
3570
3571 // Use null-cast information if it is available
3572 bool speculative_not_null = false;
3573 bool never_see_null = ((failure_control == NULL) // regular case only
3574 && seems_never_null(obj, data, speculative_not_null));
3575
3576 // Null check; get casted pointer; set region slot 3
3577 Node* null_ctl = top();
3578 Node* not_null_obj = NULL;
3579 if (from_inline) {
3580 not_null_obj = obj;
3581 } else if (null_free) {
3582 assert(safe_for_replace, "must be");
3583 not_null_obj = null_check(obj);
3584 } else {
3585 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3586 }
3587
3588 // If not_null_obj is dead, only null-path is taken
3589 if (stopped()) { // Doing instance-of on a NULL?
3590 set_control(null_ctl);
3591 return null();
3592 }
3593 region->init_req(_null_path, null_ctl);
3594 phi ->init_req(_null_path, null()); // Set null path value
3595 if (null_ctl == top()) {
3596 // Do this eagerly, so that pattern matches like is_diamond_phi
3597 // will work even during parsing.
3598 assert(_null_path == PATH_LIMIT-1, "delete last");
3599 region->del_req(_null_path);
3600 phi ->del_req(_null_path);
3601 }
3602
3603 Node* cast_obj = NULL;
3604 if (!from_inline && tk->klass_is_exact()) {
3605 // The following optimization tries to statically cast the speculative type of the object
3606 // (for example obtained during profiling) to the type of the superklass and then do a
3607 // dynamic check that the type of the object is what we expect. To work correctly
3608 // for checkcast and aastore the type of superklass should be exact.
3609 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3610 // We may not have profiling here or it may not help us. If we have
3611 // a speculative type use it to perform an exact cast.
3612 ciKlass* spec_obj_type = obj_type->speculative_type();
3613 if (spec_obj_type != NULL || data != NULL) {
3614 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3615 if (cast_obj != NULL) {
3616 if (failure_control != NULL) // failure is now impossible
3617 (*failure_control) = top();
3618 // adjust the type of the phi to the exact klass:
3619 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3620 }
3621 }
3622 }
3623
3624 if (cast_obj == NULL) {
3625 // Generate the subtype check
3626 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3627
3628 // Plug in success path into the merge
3629 cast_obj = from_inline ? not_null_obj : _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3630 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3631 if (failure_control == NULL) {
3632 if (not_subtype_ctrl != top()) { // If failure is possible
3633 PreserveJVMState pjvms(this);
3634 set_control(not_subtype_ctrl);
3635 Node* obj_klass = NULL;
3636 if (not_null_obj->is_InlineTypeBase()) {
3637 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
3638 } else {
3639 obj_klass = load_object_klass(not_null_obj);
3640 }
3641 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3642 }
3643 } else {
3644 (*failure_control) = not_subtype_ctrl;
3645 }
3646 }
3647
3648 region->init_req(_obj_path, control());
3649 phi ->init_req(_obj_path, cast_obj);
3650
3651 // A merge of NULL or Casted-NotNull obj
3652 Node* res = _gvn.transform(phi);
3653
3654 // Note I do NOT always 'replace_in_map(obj,result)' here.
3655 // if( tk->klass()->can_be_primary_super() )
3656 // This means that if I successfully store an Object into an array-of-String
3657 // I 'forget' that the Object is really now known to be a String. I have to
3658 // do this because we don't have true union types for interfaces - if I store
3659 // a Baz into an array-of-Interface and then tell the optimizer it's an
3660 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3661 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3662 // replace_in_map( obj, res );
3663
3664 // Return final merged results
3665 set_control( _gvn.transform(region) );
3666 record_for_igvn(region);
3667
3668 bool not_inline = !toop->can_be_inline_type();
3669 bool not_flattened = !UseFlatArray || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flatten_array());
3670 if (EnableValhalla && not_flattened) {
3671 // Check if obj has been loaded from an array
3672 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3673 Node* array = NULL;
3674 if (obj->isa_Load()) {
3675 Node* address = obj->in(MemNode::Address);
3676 if (address->isa_AddP()) {
3677 array = address->as_AddP()->in(AddPNode::Base);
3678 }
3679 } else if (obj->is_Phi()) {
3680 Node* region = obj->in(0);
3681 // TODO make this more robust (see JDK-8231346)
3682 if (region->req() == 3 && region->in(2) != NULL && region->in(2)->in(0) != NULL) {
3683 IfNode* iff = region->in(2)->in(0)->isa_If();
3684 if (iff != NULL) {
3685 iff->is_flat_array_check(&_gvn, &array);
3686 }
3687 }
3688 }
3689 if (array != NULL) {
3690 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3691 if (ary_t != NULL) {
3692 if (!ary_t->is_not_null_free() && not_inline) {
3693 // Casting array element to a non-inline-type, mark array as not null-free.
3694 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3695 replace_in_map(array, cast);
3696 } else if (!ary_t->is_not_flat()) {
3697 // Casting array element to a non-flattened type, mark array as not flat.
3698 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3699 replace_in_map(array, cast);
3700 }
3701 }
3702 }
3703 }
3704
3705 if (!stopped() && !res->is_InlineTypeBase()) {
3706 res = record_profiled_receiver_for_speculation(res);
3707 if (toop->is_inlinetypeptr()) {
3708 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass(), !gvn().type(res)->maybe_null());
3709 res = vt;
3710 if (safe_for_replace) {
3711 if (vt->is_InlineType() && C->inlining_incrementally()) {
3712 vt = vt->as_InlineType()->as_ptr(&_gvn);
3713 }
3714 replace_in_map(obj, vt);
3715 replace_in_map(not_null_obj, vt);
3716 replace_in_map(res, vt);
3717 }
3718 }
3719 }
3720 return res;
3721 }
3722
3723 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3724 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3725 Node* mark = make_load(NULL, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3726 Node* mask = MakeConX(markWord::inline_type_pattern);
3727 Node* masked = _gvn.transform(new AndXNode(mark, mask));
3728 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3729 return _gvn.transform(new BoolNode(cmp, is_inline ? BoolTest::eq : BoolTest::ne));
3730 }
3731
3732 Node* GraphKit::is_val_mirror(Node* mirror) {
3733 Node* p = basic_plus_adr(mirror, java_lang_Class::secondary_mirror_offset());
3734 Node* secondary_mirror = access_load_at(mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR), T_OBJECT, IN_HEAP);
3735 Node* cmp = _gvn.transform(new CmpPNode(mirror, secondary_mirror));
3736 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3737 }
3738
3739 Node* GraphKit::array_lh_test(Node* klass, jint mask, jint val, bool eq) {
3740 Node* lh_adr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
3741 // Make sure to use immutable memory here to enable hoisting the check out of loops
3742 Node* lh_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lh_adr, lh_adr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3743 Node* masked = _gvn.transform(new AndINode(lh_val, intcon(mask)));
3744 Node* cmp = _gvn.transform(new CmpINode(masked, intcon(val)));
3745 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3746 }
3747
3748 Node* GraphKit::flat_array_test(Node* ary, bool flat) {
3749 // We can't use immutable memory here because the mark word is mutable.
3750 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3751 // check is moved out of loops (mainly to enable loop unswitching).
3752 Node* mem = UseArrayMarkWordCheck ? memory(Compile::AliasIdxRaw) : immutable_memory();
3753 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, mem, ary));
3754 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3755 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3756 }
3757
3758 Node* GraphKit::null_free_array_test(Node* klass, bool null_free) {
3759 return array_lh_test(klass, Klass::_lh_null_free_array_bit_inplace, 0, !null_free);
3760 }
3761
3762 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3763 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3764 const Type* val_t = _gvn.type(val);
3765 if (val->is_InlineType() || !TypePtr::NULL_PTR->higher_equal(val_t)) {
3766 return ary; // Never null
3767 }
3768 RegionNode* region = new RegionNode(3);
3769 Node* null_ctl = top();
3770 null_check_oop(val, &null_ctl);
3771 if (null_ctl != top()) {
3772 PreserveJVMState pjvms(this);
3773 set_control(null_ctl);
3774 {
3775 // Deoptimize if null-free array
3776 BuildCutout unless(this, null_free_array_test(load_object_klass(ary), /* null_free = */ false), PROB_MAX);
3777 inc_sp(nargs);
3778 uncommon_trap(Deoptimization::Reason_null_check,
3779 Deoptimization::Action_none);
3780 }
3781 region->init_req(1, control());
3782 }
3783 region->init_req(2, control());
3784 set_control(_gvn.transform(region));
3785 record_for_igvn(region);
3786 if (val_t == TypePtr::NULL_PTR) {
3787 // Since we were just successfully storing null, the array can't be null free.
3788 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3789 ary_t = ary_t->cast_to_not_null_free();
3790 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3791 if (safe_for_replace) {
3792 replace_in_map(ary, cast);
3793 }
3794 ary = cast;
3795 }
3796 return ary;
3797 }
3798
3799 //------------------------------next_monitor-----------------------------------
3800 // What number should be given to the next monitor?
3801 int GraphKit::next_monitor() {
3802 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3803 int next = current + C->sync_stack_slots();
3804 // Keep the toplevel high water mark current:
3805 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3806 return current;
3807 }
3808
3809 //------------------------------insert_mem_bar---------------------------------
3810 // Memory barrier to avoid floating things around
3811 // The membar serves as a pinch point between both control and all memory slices.
3812 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3813 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3814 mb->init_req(TypeFunc::Control, control());
3815 mb->init_req(TypeFunc::Memory, reset_memory());
3816 Node* membar = _gvn.transform(mb);
3817 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3818 set_all_memory_call(membar);
3819 return membar;
3820 }
3821
3822 //-------------------------insert_mem_bar_volatile----------------------------
3823 // Memory barrier to avoid floating things around
3824 // The membar serves as a pinch point between both control and memory(alias_idx).
3825 // If you want to make a pinch point on all memory slices, do not use this
3826 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3827 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3828 // When Parse::do_put_xxx updates a volatile field, it appends a series
3829 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3830 // The first membar is on the same memory slice as the field store opcode.
3831 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3832 // All the other membars (for other volatile slices, including AliasIdxBot,
3833 // which stands for all unknown volatile slices) are control-dependent
3834 // on the first membar. This prevents later volatile loads or stores
3835 // from sliding up past the just-emitted store.
3836
3837 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3838 mb->set_req(TypeFunc::Control,control());
3839 if (alias_idx == Compile::AliasIdxBot) {
3840 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3841 } else {
3842 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3843 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3844 }
3845 Node* membar = _gvn.transform(mb);
3846 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3847 if (alias_idx == Compile::AliasIdxBot) {
3848 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3849 } else {
3850 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3851 }
3852 return membar;
3853 }
3854
3855 //------------------------------shared_lock------------------------------------
3856 // Emit locking code.
3857 FastLockNode* GraphKit::shared_lock(Node* obj) {
3858 // bci is either a monitorenter bc or InvocationEntryBci
3859 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3860 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3861
3862 if( !GenerateSynchronizationCode )
3863 return NULL; // Not locking things?
3864
3865 if (stopped()) // Dead monitor?
3866 return NULL;
3867
3868 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3869
3870 // Box the stack location
3871 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3872 Node* mem = reset_memory();
3873
3874 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3875
3876 // Create the rtm counters for this fast lock if needed.
3877 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3878
3879 // Add monitor to debug info for the slow path. If we block inside the
3880 // slow path and de-opt, we need the monitor hanging around
3881 map()->push_monitor( flock );
3882
3883 const TypeFunc *tf = LockNode::lock_type();
3884 LockNode *lock = new LockNode(C, tf);
3885
3886 lock->init_req( TypeFunc::Control, control() );
3887 lock->init_req( TypeFunc::Memory , mem );
3888 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3889 lock->init_req( TypeFunc::FramePtr, frameptr() );
3890 lock->init_req( TypeFunc::ReturnAdr, top() );
3891
3892 lock->init_req(TypeFunc::Parms + 0, obj);
3893 lock->init_req(TypeFunc::Parms + 1, box);
3894 lock->init_req(TypeFunc::Parms + 2, flock);
3895 add_safepoint_edges(lock);
3896
3897 lock = _gvn.transform( lock )->as_Lock();
3898
3899 // lock has no side-effects, sets few values
3900 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3901
3902 insert_mem_bar(Op_MemBarAcquireLock);
3903
3904 // Add this to the worklist so that the lock can be eliminated
3905 record_for_igvn(lock);
3906
3907 #ifndef PRODUCT
3908 if (PrintLockStatistics) {
3909 // Update the counter for this lock. Don't bother using an atomic
3910 // operation since we don't require absolute accuracy.
3911 lock->create_lock_counter(map()->jvms());
3912 increment_counter(lock->counter()->addr());
3913 }
3914 #endif
3915
3916 return flock;
3917 }
3918
3919
3920 //------------------------------shared_unlock----------------------------------
3921 // Emit unlocking code.
3922 void GraphKit::shared_unlock(Node* box, Node* obj) {
3923 // bci is either a monitorenter bc or InvocationEntryBci
3924 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3925 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3926
3927 if( !GenerateSynchronizationCode )
3928 return;
3929 if (stopped()) { // Dead monitor?
3930 map()->pop_monitor(); // Kill monitor from debug info
3931 return;
3932 }
3933 assert(!obj->is_InlineTypeBase(), "should not unlock on inline type");
3934
3935 // Memory barrier to avoid floating things down past the locked region
3936 insert_mem_bar(Op_MemBarReleaseLock);
3937
3938 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3939 UnlockNode *unlock = new UnlockNode(C, tf);
3940 #ifdef ASSERT
3941 unlock->set_dbg_jvms(sync_jvms());
3942 #endif
3943 uint raw_idx = Compile::AliasIdxRaw;
3944 unlock->init_req( TypeFunc::Control, control() );
3945 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3946 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3947 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3948 unlock->init_req( TypeFunc::ReturnAdr, top() );
3949
3950 unlock->init_req(TypeFunc::Parms + 0, obj);
3951 unlock->init_req(TypeFunc::Parms + 1, box);
3952 unlock = _gvn.transform(unlock)->as_Unlock();
3953
3954 Node* mem = reset_memory();
3955
3956 // unlock has no side-effects, sets few values
3957 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3958
3959 // Kill monitor from debug info
3960 map()->pop_monitor( );
3961 }
3962
3963 //-------------------------------get_layout_helper-----------------------------
3964 // If the given klass is a constant or known to be an array,
3965 // fetch the constant layout helper value into constant_value
3966 // and return (Node*)NULL. Otherwise, load the non-constant
3967 // layout helper value, and return the node which represents it.
3968 // This two-faced routine is useful because allocation sites
3969 // almost always feature constant types.
3970 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3971 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3972 if (!StressReflectiveCode && inst_klass != NULL) {
3973 ciKlass* klass = inst_klass->klass();
3974 assert(klass != NULL, "klass should not be NULL");
3975 bool xklass = inst_klass->klass_is_exact();
3976 bool can_be_flattened = false;
3977 if (UseFlatArray && klass->is_obj_array_klass() && !klass->as_obj_array_klass()->is_elem_null_free()) {
3978 // The runtime type of [LMyValue might be [QMyValue due to [QMyValue <: [LMyValue.
3979 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
3980 can_be_flattened = elem->can_be_inline_klass() && (!elem->is_inlinetype() || elem->flatten_array());
3981 }
3982 if (!can_be_flattened && (xklass || klass->is_array_klass())) {
3983 jint lhelper = klass->layout_helper();
3984 if (lhelper != Klass::_lh_neutral_value) {
3985 constant_value = lhelper;
3986 return (Node*) NULL;
3987 }
3988 }
3989 }
3990 constant_value = Klass::_lh_neutral_value; // put in a known value
3991 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3992 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3993 }
3994
3995 // We just put in an allocate/initialize with a big raw-memory effect.
3996 // Hook selected additional alias categories on the initialization.
3997 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3998 MergeMemNode* init_in_merge,
3999 Node* init_out_raw) {
4000 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4001 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4002
4003 Node* prevmem = kit.memory(alias_idx);
4004 init_in_merge->set_memory_at(alias_idx, prevmem);
4005 if (init_out_raw != NULL) {
4006 kit.set_memory(init_out_raw, alias_idx);
4007 }
4008 }
4009
4010 //---------------------------set_output_for_allocation-------------------------
4011 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4012 const TypeOopPtr* oop_type,
4013 bool deoptimize_on_exception) {
4014 int rawidx = Compile::AliasIdxRaw;
4015 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4016 add_safepoint_edges(alloc);
4017 Node* allocx = _gvn.transform(alloc);
4018 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4019 // create memory projection for i_o
4020 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4021 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4022
4023 // create a memory projection as for the normal control path
4024 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4025 set_memory(malloc, rawidx);
4026
4027 // a normal slow-call doesn't change i_o, but an allocation does
4028 // we create a separate i_o projection for the normal control path
4029 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4030 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4031
4032 // put in an initialization barrier
4033 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4034 rawoop)->as_Initialize();
4035 assert(alloc->initialization() == init, "2-way macro link must work");
4036 assert(init ->allocation() == alloc, "2-way macro link must work");
4037 {
4038 // Extract memory strands which may participate in the new object's
4039 // initialization, and source them from the new InitializeNode.
4040 // This will allow us to observe initializations when they occur,
4041 // and link them properly (as a group) to the InitializeNode.
4042 assert(init->in(InitializeNode::Memory) == malloc, "");
4043 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4044 init->set_req(InitializeNode::Memory, minit_in);
4045 record_for_igvn(minit_in); // fold it up later, if possible
4046 _gvn.set_type(minit_in, Type::MEMORY);
4047 Node* minit_out = memory(rawidx);
4048 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4049 // Add an edge in the MergeMem for the header fields so an access
4050 // to one of those has correct memory state
4051 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
4052 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
4053 if (oop_type->isa_aryptr()) {
4054 const TypeAryPtr* arytype = oop_type->is_aryptr();
4055 if (arytype->klass()->is_flat_array_klass()) {
4056 // Initially all flattened array accesses share a single slice
4057 // but that changes after parsing. Prepare the memory graph so
4058 // it can optimize flattened array accesses properly once they
4059 // don't share a single slice.
4060 assert(C->flattened_accesses_share_alias(), "should be set at parse time");
4061 C->set_flattened_accesses_share_alias(false);
4062 ciFlatArrayKlass* vak = arytype->klass()->as_flat_array_klass();
4063 ciInlineKlass* vk = vak->element_klass()->as_inline_klass();
4064 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
4065 ciField* field = vk->nonstatic_field_at(i);
4066 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
4067 continue; // do not bother to track really large numbers of fields
4068 int off_in_vt = field->offset() - vk->first_field_offset();
4069 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
4070 int fieldidx = C->get_alias_index(adr_type, true);
4071 // Pass NULL for init_out. Having per flat array element field memory edges as uses of the Initialize node
4072 // can result in per flat array field Phis to be created which confuses the logic of
4073 // Compile::adjust_flattened_array_access_aliases().
4074 hook_memory_on_init(*this, fieldidx, minit_in, NULL);
4075 }
4076 C->set_flattened_accesses_share_alias(true);
4077 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
4078 } else {
4079 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4080 int elemidx = C->get_alias_index(telemref);
4081 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
4082 }
4083 } else if (oop_type->isa_instptr()) {
4084 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
4085 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
4086 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4087 ciField* field = ik->nonstatic_field_at(i);
4088 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
4089 continue; // do not bother to track really large numbers of fields
4090 // Find (or create) the alias category for this field:
4091 int fieldidx = C->alias_type(field)->index();
4092 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
4093 }
4094 }
4095 }
4096
4097 // Cast raw oop to the real thing...
4098 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4099 javaoop = _gvn.transform(javaoop);
4100 C->set_recent_alloc(control(), javaoop);
4101 assert(just_allocated_object(control()) == javaoop, "just allocated");
4102
4103 #ifdef ASSERT
4104 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4105 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
4106 "Ideal_allocation works");
4107 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
4108 "Ideal_allocation works");
4109 if (alloc->is_AllocateArray()) {
4110 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
4111 "Ideal_allocation works");
4112 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
4113 "Ideal_allocation works");
4114 } else {
4115 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4116 }
4117 }
4118 #endif //ASSERT
4119
4120 return javaoop;
4121 }
4122
4123 //---------------------------new_instance--------------------------------------
4124 // This routine takes a klass_node which may be constant (for a static type)
4125 // or may be non-constant (for reflective code). It will work equally well
4126 // for either, and the graph will fold nicely if the optimizer later reduces
4127 // the type to a constant.
4128 // The optional arguments are for specialized use by intrinsics:
4129 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4130 // - If 'return_size_val', report the the total object size to the caller.
4131 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4132 Node* GraphKit::new_instance(Node* klass_node,
4133 Node* extra_slow_test,
4134 Node* *return_size_val,
4135 bool deoptimize_on_exception,
4136 InlineTypeBaseNode* inline_type_node) {
4137 // Compute size in doublewords
4138 // The size is always an integral number of doublewords, represented
4139 // as a positive bytewise size stored in the klass's layout_helper.
4140 // The layout_helper also encodes (in a low bit) the need for a slow path.
4141 jint layout_con = Klass::_lh_neutral_value;
4142 Node* layout_val = get_layout_helper(klass_node, layout_con);
4143 bool layout_is_con = (layout_val == NULL);
4144
4145 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
4146 // Generate the initial go-slow test. It's either ALWAYS (return a
4147 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
4148 // case) a computed value derived from the layout_helper.
4149 Node* initial_slow_test = NULL;
4150 if (layout_is_con) {
4151 assert(!StressReflectiveCode, "stress mode does not use these paths");
4152 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4153 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4154 } else { // reflective case
4155 // This reflective path is used by Unsafe.allocateInstance.
4156 // (It may be stress-tested by specifying StressReflectiveCode.)
4157 // Basically, we want to get into the VM is there's an illegal argument.
4158 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4159 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4160 if (extra_slow_test != intcon(0)) {
4161 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4162 }
4163 // (Macro-expander will further convert this to a Bool, if necessary.)
4164 }
4165
4166 // Find the size in bytes. This is easy; it's the layout_helper.
4167 // The size value must be valid even if the slow path is taken.
4168 Node* size = NULL;
4169 if (layout_is_con) {
4170 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4171 } else { // reflective case
4172 // This reflective path is used by clone and Unsafe.allocateInstance.
4173 size = ConvI2X(layout_val);
4174
4175 // Clear the low bits to extract layout_helper_size_in_bytes:
4176 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4177 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4178 size = _gvn.transform( new AndXNode(size, mask) );
4179 }
4180 if (return_size_val != NULL) {
4181 (*return_size_val) = size;
4182 }
4183
4184 // This is a precise notnull oop of the klass.
4185 // (Actually, it need not be precise if this is a reflective allocation.)
4186 // It's what we cast the result to.
4187 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4188 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4189 const TypeOopPtr* oop_type = tklass->as_instance_type();
4190
4191 // Now generate allocation code
4192
4193 // The entire memory state is needed for slow path of the allocation
4194 // since GC and deoptimization can happen.
4195 Node *mem = reset_memory();
4196 set_all_memory(mem); // Create new memory state
4197
4198 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4199 control(), mem, i_o(),
4200 size, klass_node,
4201 initial_slow_test, inline_type_node);
4202
4203 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4204 }
4205
4206 //-------------------------------new_array-------------------------------------
4207 // helper for newarray and anewarray
4208 // The 'length' parameter is (obviously) the length of the array.
4209 // See comments on new_instance for the meaning of the other arguments.
4210 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4211 Node* length, // number of array elements
4212 int nargs, // number of arguments to push back for uncommon trap
4213 Node* *return_size_val,
4214 bool deoptimize_on_exception) {
4215 jint layout_con = Klass::_lh_neutral_value;
4216 Node* layout_val = get_layout_helper(klass_node, layout_con);
4217 bool layout_is_con = (layout_val == NULL);
4218
4219 if (!layout_is_con && !StressReflectiveCode &&
4220 !too_many_traps(Deoptimization::Reason_class_check)) {
4221 // This is a reflective array creation site.
4222 // Optimistically assume that it is a subtype of Object[],
4223 // so that we can fold up all the address arithmetic.
4224 layout_con = Klass::array_layout_helper(T_OBJECT);
4225 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4226 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4227 { BuildCutout unless(this, bol_lh, PROB_MAX);
4228 inc_sp(nargs);
4229 uncommon_trap(Deoptimization::Reason_class_check,
4230 Deoptimization::Action_maybe_recompile);
4231 }
4232 layout_val = NULL;
4233 layout_is_con = true;
4234 }
4235
4236 // Generate the initial go-slow test. Make sure we do not overflow
4237 // if length is huge (near 2Gig) or negative! We do not need
4238 // exact double-words here, just a close approximation of needed
4239 // double-words. We can't add any offset or rounding bits, lest we
4240 // take a size -1 of bytes and make it positive. Use an unsigned
4241 // compare, so negative sizes look hugely positive.
4242 int fast_size_limit = FastAllocateSizeLimit;
4243 if (layout_is_con) {
4244 assert(!StressReflectiveCode, "stress mode does not use these paths");
4245 // Increase the size limit if we have exact knowledge of array type.
4246 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4247 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4248 }
4249
4250 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4251 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4252
4253 // --- Size Computation ---
4254 // array_size = round_to_heap(array_header + (length << elem_shift));
4255 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4256 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4257 // The rounding mask is strength-reduced, if possible.
4258 int round_mask = MinObjAlignmentInBytes - 1;
4259 Node* header_size = NULL;
4260 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4261 // (T_BYTE has the weakest alignment and size restrictions...)
4262 if (layout_is_con) {
4263 int hsize = Klass::layout_helper_header_size(layout_con);
4264 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4265 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4266 if ((round_mask & ~right_n_bits(eshift)) == 0)
4267 round_mask = 0; // strength-reduce it if it goes away completely
4268 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4269 assert(header_size_min <= hsize, "generic minimum is smallest");
4270 header_size_min = hsize;
4271 header_size = intcon(hsize + round_mask);
4272 } else {
4273 Node* hss = intcon(Klass::_lh_header_size_shift);
4274 Node* hsm = intcon(Klass::_lh_header_size_mask);
4275 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
4276 hsize = _gvn.transform( new AndINode(hsize, hsm) );
4277 Node* mask = intcon(round_mask);
4278 header_size = _gvn.transform( new AddINode(hsize, mask) );
4279 }
4280
4281 Node* elem_shift = NULL;
4282 if (layout_is_con) {
4283 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4284 if (eshift != 0)
4285 elem_shift = intcon(eshift);
4286 } else {
4287 // There is no need to mask or shift this value.
4288 // The semantics of LShiftINode include an implicit mask to 0x1F.
4289 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4290 elem_shift = layout_val;
4291 }
4292
4293 // Transition to native address size for all offset calculations:
4294 Node* lengthx = ConvI2X(length);
4295 Node* headerx = ConvI2X(header_size);
4296 #ifdef _LP64
4297 { const TypeInt* tilen = _gvn.find_int_type(length);
4298 if (tilen != NULL && tilen->_lo < 0) {
4299 // Add a manual constraint to a positive range. Cf. array_element_address.
4300 jint size_max = fast_size_limit;
4301 if (size_max > tilen->_hi) size_max = tilen->_hi;
4302 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4303
4304 // Only do a narrow I2L conversion if the range check passed.
4305 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4306 _gvn.transform(iff);
4307 RegionNode* region = new RegionNode(3);
4308 _gvn.set_type(region, Type::CONTROL);
4309 lengthx = new PhiNode(region, TypeLong::LONG);
4310 _gvn.set_type(lengthx, TypeLong::LONG);
4311
4312 // Range check passed. Use ConvI2L node with narrow type.
4313 Node* passed = IfFalse(iff);
4314 region->init_req(1, passed);
4315 // Make I2L conversion control dependent to prevent it from
4316 // floating above the range check during loop optimizations.
4317 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4318
4319 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4320 region->init_req(2, IfTrue(iff));
4321 lengthx->init_req(2, ConvI2X(length));
4322
4323 set_control(region);
4324 record_for_igvn(region);
4325 record_for_igvn(lengthx);
4326 }
4327 }
4328 #endif
4329
4330 // Combine header size (plus rounding) and body size. Then round down.
4331 // This computation cannot overflow, because it is used only in two
4332 // places, one where the length is sharply limited, and the other
4333 // after a successful allocation.
4334 Node* abody = lengthx;
4335 if (elem_shift != NULL)
4336 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
4337 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
4338 if (round_mask != 0) {
4339 Node* mask = MakeConX(~round_mask);
4340 size = _gvn.transform( new AndXNode(size, mask) );
4341 }
4342 // else if round_mask == 0, the size computation is self-rounding
4343
4344 if (return_size_val != NULL) {
4345 // This is the size
4346 (*return_size_val) = size;
4347 }
4348
4349 // Now generate allocation code
4350
4351 // The entire memory state is needed for slow path of the allocation
4352 // since GC and deoptimization can happen.
4353 Node *mem = reset_memory();
4354 set_all_memory(mem); // Create new memory state
4355
4356 if (initial_slow_test->is_Bool()) {
4357 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4358 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4359 }
4360
4361 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4362 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4363 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
4364
4365 // Inline type array variants:
4366 // - null-ok: MyValue.ref[] (ciObjArrayKlass "[LMyValue")
4367 // - null-free: MyValue.val[] (ciObjArrayKlass "[QMyValue")
4368 // - null-free, flattened: MyValue.val[] (ciFlatArrayKlass "[QMyValue")
4369 // Check if array is a null-free, non-flattened inline type array
4370 // that needs to be initialized with the default inline type.
4371 Node* default_value = NULL;
4372 Node* raw_default_value = NULL;
4373 if (ary_ptr != NULL && ary_ptr->klass_is_exact()) {
4374 // Array type is known
4375 if (ary_ptr->klass()->as_array_klass()->is_elem_null_free()) {
4376 ciInlineKlass* vk = ary_ptr->klass()->as_array_klass()->element_klass()->as_inline_klass();
4377 if (!vk->flatten_array()) {
4378 default_value = InlineTypeNode::default_oop(gvn(), vk);
4379 }
4380 }
4381 } else if (ary_klass->klass()->can_be_inline_array_klass()) {
4382 // Array type is not known, add runtime checks
4383 assert(!ary_klass->klass_is_exact(), "unexpected exact type");
4384 Node* r = new RegionNode(3);
4385 default_value = new PhiNode(r, TypeInstPtr::BOTTOM);
4386
4387 Node* bol = array_lh_test(klass_node, Klass::_lh_array_tag_flat_value_bit_inplace | Klass::_lh_null_free_array_bit_inplace, Klass::_lh_null_free_array_bit_inplace);
4388 IfNode* iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4389
4390 // Null-free, non-flattened inline type array, initialize with the default value
4391 set_control(_gvn.transform(new IfTrueNode(iff)));
4392 Node* p = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset()));
4393 Node* eklass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), p, TypeInstPtr::KLASS));
4394 Node* adr_fixed_block_addr = basic_plus_adr(eklass, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()));
4395 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4396 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(InlineKlass::default_value_offset_offset()));
4397 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered);
4398 Node* elem_mirror = load_mirror_from_klass(eklass);
4399 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset));
4400 Node* val = access_load_at(elem_mirror, default_value_addr, _gvn.type(default_value_addr)->is_ptr(), TypeInstPtr::BOTTOM, T_OBJECT, IN_HEAP);
4401 r->init_req(1, control());
4402 default_value->init_req(1, val);
4403
4404 // Otherwise initialize with all zero
4405 r->init_req(2, _gvn.transform(new IfFalseNode(iff)));
4406 default_value->init_req(2, null());
4407
4408 set_control(_gvn.transform(r));
4409 default_value = _gvn.transform(default_value);
4410 }
4411 if (default_value != NULL) {
4412 if (UseCompressedOops) {
4413 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4414 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4415 Node* lower = _gvn.transform(new CastP2XNode(control(), default_value));
4416 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4417 raw_default_value = _gvn.transform(new OrLNode(lower, upper));
4418 } else {
4419 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4420 }
4421 }
4422
4423 // Create the AllocateArrayNode and its result projections
4424 AllocateArrayNode* alloc = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4425 control(), mem, i_o(),
4426 size, klass_node,
4427 initial_slow_test,
4428 length, default_value,
4429 raw_default_value);
4430
4431 // Cast to correct type. Note that the klass_node may be constant or not,
4432 // and in the latter case the actual array type will be inexact also.
4433 // (This happens via a non-constant argument to inline_native_newArray.)
4434 // In any case, the value of klass_node provides the desired array type.
4435 const TypeInt* length_type = _gvn.find_int_type(length);
4436 if (ary_type->isa_aryptr() && length_type != NULL) {
4437 // Try to get a better type than POS for the size
4438 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4439 }
4440
4441 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4442
4443 array_ideal_length(alloc, ary_type, true);
4444 return javaoop;
4445 }
4446
4447 // The following "Ideal_foo" functions are placed here because they recognize
4448 // the graph shapes created by the functions immediately above.
4449
4450 //---------------------------Ideal_allocation----------------------------------
4451 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4452 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
4453 if (ptr == NULL) { // reduce dumb test in callers
4454 return NULL;
4455 }
4456
4457 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4458 ptr = bs->step_over_gc_barrier(ptr);
4459
4460 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4461 ptr = ptr->in(1);
4462 if (ptr == NULL) return NULL;
4463 }
4464 // Return NULL for allocations with several casts:
4465 // j.l.reflect.Array.newInstance(jobject, jint)
4466 // Object.clone()
4467 // to keep more precise type from last cast.
4468 if (ptr->is_Proj()) {
4469 Node* allo = ptr->in(0);
4470 if (allo != NULL && allo->is_Allocate()) {
4471 return allo->as_Allocate();
4472 }
4473 }
4474 // Report failure to match.
4475 return NULL;
4476 }
4477
4478 // Fancy version which also strips off an offset (and reports it to caller).
4479 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
4480 intptr_t& offset) {
4481 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4482 if (base == NULL) return NULL;
4483 return Ideal_allocation(base, phase);
4484 }
4485
4486 // Trace Initialize <- Proj[Parm] <- Allocate
4487 AllocateNode* InitializeNode::allocation() {
4488 Node* rawoop = in(InitializeNode::RawAddress);
4489 if (rawoop->is_Proj()) {
4490 Node* alloc = rawoop->in(0);
4491 if (alloc->is_Allocate()) {
4492 return alloc->as_Allocate();
4493 }
4494 }
4495 return NULL;
4496 }
4497
4498 // Trace Allocate -> Proj[Parm] -> Initialize
4499 InitializeNode* AllocateNode::initialization() {
4500 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4501 if (rawoop == NULL) return NULL;
4502 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4503 Node* init = rawoop->fast_out(i);
4504 if (init->is_Initialize()) {
4505 assert(init->as_Initialize()->allocation() == this, "2-way link");
4506 return init->as_Initialize();
4507 }
4508 }
4509 return NULL;
4510 }
4511
4512 //----------------------------- loop predicates ---------------------------
4513
4514 //------------------------------add_predicate_impl----------------------------
4515 void GraphKit::add_empty_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
4516 // Too many traps seen?
4517 if (too_many_traps(reason)) {
4518 #ifdef ASSERT
4519 if (TraceLoopPredicate) {
4520 int tc = C->trap_count(reason);
4521 tty->print("too many traps=%s tcount=%d in ",
4522 Deoptimization::trap_reason_name(reason), tc);
4523 method()->print(); // which method has too many predicate traps
4524 tty->cr();
4525 }
4526 #endif
4527 // We cannot afford to take more traps here,
4528 // do not generate predicate.
4529 return;
4530 }
4531
4532 Node *cont = _gvn.intcon(1);
4533 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
4534 Node *bol = _gvn.transform(new Conv2BNode(opq));
4535 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4536 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
4537 C->add_predicate_opaq(opq);
4538 {
4539 PreserveJVMState pjvms(this);
4540 set_control(iffalse);
4541 inc_sp(nargs);
4542 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4543 }
4544 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
4545 set_control(iftrue);
4546 }
4547
4548 //------------------------------add_predicate---------------------------------
4549 void GraphKit::add_empty_predicates(int nargs) {
4550 // These loop predicates remain empty. All concrete loop predicates are inserted above the corresponding
4551 // empty loop predicate later by 'PhaseIdealLoop::create_new_if_for_predicate'. All concrete loop predicates of
4552 // a specific kind (normal, profile or limit check) share the same uncommon trap as the empty loop predicate.
4553 if (UseLoopPredicate) {
4554 add_empty_predicate_impl(Deoptimization::Reason_predicate, nargs);
4555 }
4556 if (UseProfiledLoopPredicate) {
4557 add_empty_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
4558 }
4559 // loop's limit check predicate should be near the loop.
4560 add_empty_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
4561 }
4562
4563 void GraphKit::sync_kit(IdealKit& ideal) {
4564 set_all_memory(ideal.merged_memory());
4565 set_i_o(ideal.i_o());
4566 set_control(ideal.ctrl());
4567 }
4568
4569 void GraphKit::final_sync(IdealKit& ideal) {
4570 // Final sync IdealKit and graphKit.
4571 sync_kit(ideal);
4572 }
4573
4574 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4575 Node* len = load_array_length(load_String_value(str, set_ctrl));
4576 Node* coder = load_String_coder(str, set_ctrl);
4577 // Divide length by 2 if coder is UTF16
4578 return _gvn.transform(new RShiftINode(len, coder));
4579 }
4580
4581 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4582 int value_offset = java_lang_String::value_offset();
4583 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4584 false, NULL, Type::Offset(0));
4585 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4586 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4587 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, true, true),
4588 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4589 Node* p = basic_plus_adr(str, str, value_offset);
4590 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4591 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4592 return load;
4593 }
4594
4595 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4596 if (!CompactStrings) {
4597 return intcon(java_lang_String::CODER_UTF16);
4598 }
4599 int coder_offset = java_lang_String::coder_offset();
4600 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4601 false, NULL, Type::Offset(0));
4602 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4603
4604 Node* p = basic_plus_adr(str, str, coder_offset);
4605 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4606 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4607 return load;
4608 }
4609
4610 void GraphKit::store_String_value(Node* str, Node* value) {
4611 int value_offset = java_lang_String::value_offset();
4612 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4613 false, NULL, Type::Offset(0));
4614 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4615
4616 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4617 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4618 }
4619
4620 void GraphKit::store_String_coder(Node* str, Node* value) {
4621 int coder_offset = java_lang_String::coder_offset();
4622 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4623 false, NULL, Type::Offset(0));
4624 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4625
4626 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4627 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4628 }
4629
4630 // Capture src and dst memory state with a MergeMemNode
4631 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4632 if (src_type == dst_type) {
4633 // Types are equal, we don't need a MergeMemNode
4634 return memory(src_type);
4635 }
4636 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4637 record_for_igvn(merge); // fold it up later, if possible
4638 int src_idx = C->get_alias_index(src_type);
4639 int dst_idx = C->get_alias_index(dst_type);
4640 merge->set_memory_at(src_idx, memory(src_idx));
4641 merge->set_memory_at(dst_idx, memory(dst_idx));
4642 return merge;
4643 }
4644
4645 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4646 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4647 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4648 // If input and output memory types differ, capture both states to preserve
4649 // the dependency between preceding and subsequent loads/stores.
4650 // For example, the following program:
4651 // StoreB
4652 // compress_string
4653 // LoadB
4654 // has this memory graph (use->def):
4655 // LoadB -> compress_string -> CharMem
4656 // ... -> StoreB -> ByteMem
4657 // The intrinsic hides the dependency between LoadB and StoreB, causing
4658 // the load to read from memory not containing the result of the StoreB.
4659 // The correct memory graph should look like this:
4660 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4661 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4662 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4663 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4664 set_memory(res_mem, TypeAryPtr::BYTES);
4665 return str;
4666 }
4667
4668 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4669 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4670 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4671 // Capture src and dst memory (see comment in 'compress_string').
4672 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4673 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4674 set_memory(_gvn.transform(str), dst_type);
4675 }
4676
4677 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4678 /**
4679 * int i_char = start;
4680 * for (int i_byte = 0; i_byte < count; i_byte++) {
4681 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4682 * }
4683 */
4684 add_empty_predicates();
4685 C->set_has_loops(true);
4686
4687 RegionNode* head = new RegionNode(3);
4688 head->init_req(1, control());
4689 gvn().set_type(head, Type::CONTROL);
4690 record_for_igvn(head);
4691
4692 Node* i_byte = new PhiNode(head, TypeInt::INT);
4693 i_byte->init_req(1, intcon(0));
4694 gvn().set_type(i_byte, TypeInt::INT);
4695 record_for_igvn(i_byte);
4696
4697 Node* i_char = new PhiNode(head, TypeInt::INT);
4698 i_char->init_req(1, start);
4699 gvn().set_type(i_char, TypeInt::INT);
4700 record_for_igvn(i_char);
4701
4702 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4703 gvn().set_type(mem, Type::MEMORY);
4704 record_for_igvn(mem);
4705 set_control(head);
4706 set_memory(mem, TypeAryPtr::BYTES);
4707 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4708 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4709 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4710 false, false, true /* mismatched */);
4711
4712 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4713 head->init_req(2, IfTrue(iff));
4714 mem->init_req(2, st);
4715 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4716 i_char->init_req(2, AddI(i_char, intcon(2)));
4717
4718 set_control(IfFalse(iff));
4719 set_memory(st, TypeAryPtr::BYTES);
4720 }
4721
4722 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4723 if (!field->is_constant()) {
4724 return NULL; // Field not marked as constant.
4725 }
4726 ciInstance* holder = NULL;
4727 if (!field->is_static()) {
4728 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4729 if (const_oop != NULL && const_oop->is_instance()) {
4730 holder = const_oop->as_instance();
4731 }
4732 }
4733 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4734 /*is_unsigned_load=*/false);
4735 if (con_type != NULL) {
4736 Node* con = makecon(con_type);
4737 if (field->type()->is_inlinetype()) {
4738 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass(), field->is_null_free());
4739 } else if (con_type->is_inlinetypeptr()) {
4740 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass(), field->is_null_free());
4741 }
4742 return con;
4743 }
4744 return NULL;
4745 }
4746
4747 //---------------------------load_mirror_from_klass----------------------------
4748 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4749 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4750 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4751 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4752 // mirror = ((OopHandle)mirror)->resolve();
4753 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4754 }