1 /*
2 * Copyright (c) 2001, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "interpreter/interpreter.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/graphKit.hpp"
36 #include "opto/idealKit.hpp"
37 #include "opto/intrinsicnode.hpp"
38 #include "opto/locknode.hpp"
39 #include "opto/machnode.hpp"
40 #include "opto/opaquenode.hpp"
41 #include "opto/parse.hpp"
42 #include "opto/rootnode.hpp"
43 #include "opto/runtime.hpp"
44 #include "runtime/deoptimization.hpp"
45 #include "runtime/sharedRuntime.hpp"
46 #include "utilities/bitMap.inline.hpp"
47
48 //----------------------------GraphKit-----------------------------------------
49 // Main utility constructor.
GraphKit(JVMState * jvms)50 GraphKit::GraphKit(JVMState* jvms)
51 : Phase(Phase::Parser),
52 _env(C->env()),
53 _gvn(*C->initial_gvn()),
54 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
55 {
56 _exceptions = jvms->map()->next_exception();
57 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
58 set_jvms(jvms);
59 }
60
61 // Private constructor for parser.
GraphKit()62 GraphKit::GraphKit()
63 : Phase(Phase::Parser),
64 _env(C->env()),
65 _gvn(*C->initial_gvn()),
66 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
67 {
68 _exceptions = NULL;
69 set_map(NULL);
70 debug_only(_sp = -99);
71 debug_only(set_bci(-99));
72 }
73
74
75
76 //---------------------------clean_stack---------------------------------------
77 // Clear away rubbish from the stack area of the JVM state.
78 // This destroys any arguments that may be waiting on the stack.
clean_stack(int from_sp)79 void GraphKit::clean_stack(int from_sp) {
80 SafePointNode* map = this->map();
81 JVMState* jvms = this->jvms();
82 int stk_size = jvms->stk_size();
83 int stkoff = jvms->stkoff();
84 Node* top = this->top();
85 for (int i = from_sp; i < stk_size; i++) {
86 if (map->in(stkoff + i) != top) {
87 map->set_req(stkoff + i, top);
88 }
89 }
90 }
91
92
93 //--------------------------------sync_jvms-----------------------------------
94 // Make sure our current jvms agrees with our parse state.
sync_jvms() const95 JVMState* GraphKit::sync_jvms() const {
96 JVMState* jvms = this->jvms();
97 jvms->set_bci(bci()); // Record the new bci in the JVMState
98 jvms->set_sp(sp()); // Record the new sp in the JVMState
99 assert(jvms_in_sync(), "jvms is now in sync");
100 return jvms;
101 }
102
103 //--------------------------------sync_jvms_for_reexecute---------------------
104 // Make sure our current jvms agrees with our parse state. This version
105 // uses the reexecute_sp for reexecuting bytecodes.
sync_jvms_for_reexecute()106 JVMState* GraphKit::sync_jvms_for_reexecute() {
107 JVMState* jvms = this->jvms();
108 jvms->set_bci(bci()); // Record the new bci in the JVMState
109 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
110 return jvms;
111 }
112
113 #ifdef ASSERT
jvms_in_sync() const114 bool GraphKit::jvms_in_sync() const {
115 Parse* parse = is_Parse();
116 if (parse == NULL) {
117 if (bci() != jvms()->bci()) return false;
118 if (sp() != (int)jvms()->sp()) return false;
119 return true;
120 }
121 if (jvms()->method() != parse->method()) return false;
122 if (jvms()->bci() != parse->bci()) return false;
123 int jvms_sp = jvms()->sp();
124 if (jvms_sp != parse->sp()) return false;
125 int jvms_depth = jvms()->depth();
126 if (jvms_depth != parse->depth()) return false;
127 return true;
128 }
129
130 // Local helper checks for special internal merge points
131 // used to accumulate and merge exception states.
132 // They are marked by the region's in(0) edge being the map itself.
133 // Such merge points must never "escape" into the parser at large,
134 // until they have been handed to gvn.transform.
is_hidden_merge(Node * reg)135 static bool is_hidden_merge(Node* reg) {
136 if (reg == NULL) return false;
137 if (reg->is_Phi()) {
138 reg = reg->in(0);
139 if (reg == NULL) return false;
140 }
141 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
142 }
143
verify_map() const144 void GraphKit::verify_map() const {
145 if (map() == NULL) return; // null map is OK
146 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
147 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
148 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
149 }
150
verify_exception_state(SafePointNode * ex_map)151 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
152 assert(ex_map->next_exception() == NULL, "not already part of a chain");
153 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
154 }
155 #endif
156
157 //---------------------------stop_and_kill_map---------------------------------
158 // Set _map to NULL, signalling a stop to further bytecode execution.
159 // First smash the current map's control to a constant, to mark it dead.
stop_and_kill_map()160 void GraphKit::stop_and_kill_map() {
161 SafePointNode* dead_map = stop();
162 if (dead_map != NULL) {
163 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
164 assert(dead_map->is_killed(), "must be so marked");
165 }
166 }
167
168
169 //--------------------------------stopped--------------------------------------
170 // Tell if _map is NULL, or control is top.
stopped()171 bool GraphKit::stopped() {
172 if (map() == NULL) return true;
173 else if (control() == top()) return true;
174 else return false;
175 }
176
177
178 //-----------------------------has_ex_handler----------------------------------
179 // Tell if this method or any caller method has exception handlers.
has_ex_handler()180 bool GraphKit::has_ex_handler() {
181 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
182 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
183 return true;
184 }
185 }
186 return false;
187 }
188
189 //------------------------------save_ex_oop------------------------------------
190 // Save an exception without blowing stack contents or other JVM state.
set_saved_ex_oop(SafePointNode * ex_map,Node * ex_oop)191 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
192 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
193 ex_map->add_req(ex_oop);
194 debug_only(verify_exception_state(ex_map));
195 }
196
common_saved_ex_oop(SafePointNode * ex_map,bool clear_it)197 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
198 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
199 Node* ex_oop = ex_map->in(ex_map->req()-1);
200 if (clear_it) ex_map->del_req(ex_map->req()-1);
201 return ex_oop;
202 }
203
204 //-----------------------------saved_ex_oop------------------------------------
205 // Recover a saved exception from its map.
saved_ex_oop(SafePointNode * ex_map)206 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
207 return common_saved_ex_oop(ex_map, false);
208 }
209
210 //--------------------------clear_saved_ex_oop---------------------------------
211 // Erase a previously saved exception from its map.
clear_saved_ex_oop(SafePointNode * ex_map)212 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
213 return common_saved_ex_oop(ex_map, true);
214 }
215
216 #ifdef ASSERT
217 //---------------------------has_saved_ex_oop----------------------------------
218 // Erase a previously saved exception from its map.
has_saved_ex_oop(SafePointNode * ex_map)219 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
220 return ex_map->req() == ex_map->jvms()->endoff()+1;
221 }
222 #endif
223
224 //-------------------------make_exception_state--------------------------------
225 // Turn the current JVM state into an exception state, appending the ex_oop.
make_exception_state(Node * ex_oop)226 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
227 sync_jvms();
228 SafePointNode* ex_map = stop(); // do not manipulate this map any more
229 set_saved_ex_oop(ex_map, ex_oop);
230 return ex_map;
231 }
232
233
234 //--------------------------add_exception_state--------------------------------
235 // Add an exception to my list of exceptions.
add_exception_state(SafePointNode * ex_map)236 void GraphKit::add_exception_state(SafePointNode* ex_map) {
237 if (ex_map == NULL || ex_map->control() == top()) {
238 return;
239 }
240 #ifdef ASSERT
241 verify_exception_state(ex_map);
242 if (has_exceptions()) {
243 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
244 }
245 #endif
246
247 // If there is already an exception of exactly this type, merge with it.
248 // In particular, null-checks and other low-level exceptions common up here.
249 Node* ex_oop = saved_ex_oop(ex_map);
250 const Type* ex_type = _gvn.type(ex_oop);
251 if (ex_oop == top()) {
252 // No action needed.
253 return;
254 }
255 assert(ex_type->isa_instptr(), "exception must be an instance");
256 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
257 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
258 // We check sp also because call bytecodes can generate exceptions
259 // both before and after arguments are popped!
260 if (ex_type2 == ex_type
261 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
262 combine_exception_states(ex_map, e2);
263 return;
264 }
265 }
266
267 // No pre-existing exception of the same type. Chain it on the list.
268 push_exception_state(ex_map);
269 }
270
271 //-----------------------add_exception_states_from-----------------------------
add_exception_states_from(JVMState * jvms)272 void GraphKit::add_exception_states_from(JVMState* jvms) {
273 SafePointNode* ex_map = jvms->map()->next_exception();
274 if (ex_map != NULL) {
275 jvms->map()->set_next_exception(NULL);
276 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
277 next_map = ex_map->next_exception();
278 ex_map->set_next_exception(NULL);
279 add_exception_state(ex_map);
280 }
281 }
282 }
283
284 //-----------------------transfer_exceptions_into_jvms-------------------------
transfer_exceptions_into_jvms()285 JVMState* GraphKit::transfer_exceptions_into_jvms() {
286 if (map() == NULL) {
287 // We need a JVMS to carry the exceptions, but the map has gone away.
288 // Create a scratch JVMS, cloned from any of the exception states...
289 if (has_exceptions()) {
290 _map = _exceptions;
291 _map = clone_map();
292 _map->set_next_exception(NULL);
293 clear_saved_ex_oop(_map);
294 debug_only(verify_map());
295 } else {
296 // ...or created from scratch
297 JVMState* jvms = new (C) JVMState(_method, NULL);
298 jvms->set_bci(_bci);
299 jvms->set_sp(_sp);
300 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
301 set_jvms(jvms);
302 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
303 set_all_memory(top());
304 while (map()->req() < jvms->endoff()) map()->add_req(top());
305 }
306 // (This is a kludge, in case you didn't notice.)
307 set_control(top());
308 }
309 JVMState* jvms = sync_jvms();
310 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
311 jvms->map()->set_next_exception(_exceptions);
312 _exceptions = NULL; // done with this set of exceptions
313 return jvms;
314 }
315
add_n_reqs(Node * dstphi,Node * srcphi)316 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
317 assert(is_hidden_merge(dstphi), "must be a special merge node");
318 assert(is_hidden_merge(srcphi), "must be a special merge node");
319 uint limit = srcphi->req();
320 for (uint i = PhiNode::Input; i < limit; i++) {
321 dstphi->add_req(srcphi->in(i));
322 }
323 }
add_one_req(Node * dstphi,Node * src)324 static inline void add_one_req(Node* dstphi, Node* src) {
325 assert(is_hidden_merge(dstphi), "must be a special merge node");
326 assert(!is_hidden_merge(src), "must not be a special merge node");
327 dstphi->add_req(src);
328 }
329
330 //-----------------------combine_exception_states------------------------------
331 // This helper function combines exception states by building phis on a
332 // specially marked state-merging region. These regions and phis are
333 // untransformed, and can build up gradually. The region is marked by
334 // having a control input of its exception map, rather than NULL. Such
335 // regions do not appear except in this function, and in use_exception_state.
combine_exception_states(SafePointNode * ex_map,SafePointNode * phi_map)336 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
337 if (failing()) return; // dying anyway...
338 JVMState* ex_jvms = ex_map->_jvms;
339 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
340 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
341 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
342 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
343 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
344 assert(ex_map->req() == phi_map->req(), "matching maps");
345 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
346 Node* hidden_merge_mark = root();
347 Node* region = phi_map->control();
348 MergeMemNode* phi_mem = phi_map->merged_memory();
349 MergeMemNode* ex_mem = ex_map->merged_memory();
350 if (region->in(0) != hidden_merge_mark) {
351 // The control input is not (yet) a specially-marked region in phi_map.
352 // Make it so, and build some phis.
353 region = new RegionNode(2);
354 _gvn.set_type(region, Type::CONTROL);
355 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
356 region->init_req(1, phi_map->control());
357 phi_map->set_control(region);
358 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
359 record_for_igvn(io_phi);
360 _gvn.set_type(io_phi, Type::ABIO);
361 phi_map->set_i_o(io_phi);
362 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
363 Node* m = mms.memory();
364 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
365 record_for_igvn(m_phi);
366 _gvn.set_type(m_phi, Type::MEMORY);
367 mms.set_memory(m_phi);
368 }
369 }
370
371 // Either or both of phi_map and ex_map might already be converted into phis.
372 Node* ex_control = ex_map->control();
373 // if there is special marking on ex_map also, we add multiple edges from src
374 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
375 // how wide was the destination phi_map, originally?
376 uint orig_width = region->req();
377
378 if (add_multiple) {
379 add_n_reqs(region, ex_control);
380 add_n_reqs(phi_map->i_o(), ex_map->i_o());
381 } else {
382 // ex_map has no merges, so we just add single edges everywhere
383 add_one_req(region, ex_control);
384 add_one_req(phi_map->i_o(), ex_map->i_o());
385 }
386 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
387 if (mms.is_empty()) {
388 // get a copy of the base memory, and patch some inputs into it
389 const TypePtr* adr_type = mms.adr_type(C);
390 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
391 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
392 mms.set_memory(phi);
393 // Prepare to append interesting stuff onto the newly sliced phi:
394 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
395 }
396 // Append stuff from ex_map:
397 if (add_multiple) {
398 add_n_reqs(mms.memory(), mms.memory2());
399 } else {
400 add_one_req(mms.memory(), mms.memory2());
401 }
402 }
403 uint limit = ex_map->req();
404 for (uint i = TypeFunc::Parms; i < limit; i++) {
405 // Skip everything in the JVMS after tos. (The ex_oop follows.)
406 if (i == tos) i = ex_jvms->monoff();
407 Node* src = ex_map->in(i);
408 Node* dst = phi_map->in(i);
409 if (src != dst) {
410 PhiNode* phi;
411 if (dst->in(0) != region) {
412 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
413 record_for_igvn(phi);
414 _gvn.set_type(phi, phi->type());
415 phi_map->set_req(i, dst);
416 // Prepare to append interesting stuff onto the new phi:
417 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
418 } else {
419 assert(dst->is_Phi(), "nobody else uses a hidden region");
420 phi = dst->as_Phi();
421 }
422 if (add_multiple && src->in(0) == ex_control) {
423 // Both are phis.
424 add_n_reqs(dst, src);
425 } else {
426 while (dst->req() < region->req()) add_one_req(dst, src);
427 }
428 const Type* srctype = _gvn.type(src);
429 if (phi->type() != srctype) {
430 const Type* dsttype = phi->type()->meet_speculative(srctype);
431 if (phi->type() != dsttype) {
432 phi->set_type(dsttype);
433 _gvn.set_type(phi, dsttype);
434 }
435 }
436 }
437 }
438 phi_map->merge_replaced_nodes_with(ex_map);
439 }
440
441 //--------------------------use_exception_state--------------------------------
use_exception_state(SafePointNode * phi_map)442 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
443 if (failing()) { stop(); return top(); }
444 Node* region = phi_map->control();
445 Node* hidden_merge_mark = root();
446 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
447 Node* ex_oop = clear_saved_ex_oop(phi_map);
448 if (region->in(0) == hidden_merge_mark) {
449 // Special marking for internal ex-states. Process the phis now.
450 region->set_req(0, region); // now it's an ordinary region
451 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
452 // Note: Setting the jvms also sets the bci and sp.
453 set_control(_gvn.transform(region));
454 uint tos = jvms()->stkoff() + sp();
455 for (uint i = 1; i < tos; i++) {
456 Node* x = phi_map->in(i);
457 if (x->in(0) == region) {
458 assert(x->is_Phi(), "expected a special phi");
459 phi_map->set_req(i, _gvn.transform(x));
460 }
461 }
462 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
463 Node* x = mms.memory();
464 if (x->in(0) == region) {
465 assert(x->is_Phi(), "nobody else uses a hidden region");
466 mms.set_memory(_gvn.transform(x));
467 }
468 }
469 if (ex_oop->in(0) == region) {
470 assert(ex_oop->is_Phi(), "expected a special phi");
471 ex_oop = _gvn.transform(ex_oop);
472 }
473 } else {
474 set_jvms(phi_map->jvms());
475 }
476
477 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
478 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
479 return ex_oop;
480 }
481
482 //---------------------------------java_bc-------------------------------------
java_bc() const483 Bytecodes::Code GraphKit::java_bc() const {
484 ciMethod* method = this->method();
485 int bci = this->bci();
486 if (method != NULL && bci != InvocationEntryBci)
487 return method->java_code_at_bci(bci);
488 else
489 return Bytecodes::_illegal;
490 }
491
uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,bool must_throw)492 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
493 bool must_throw) {
494 // if the exception capability is set, then we will generate code
495 // to check the JavaThread.should_post_on_exceptions flag to see
496 // if we actually need to report exception events (for this
497 // thread). If we don't need to report exception events, we will
498 // take the normal fast path provided by add_exception_events. If
499 // exception event reporting is enabled for this thread, we will
500 // take the uncommon_trap in the BuildCutout below.
501
502 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
503 Node* jthread = _gvn.transform(new ThreadLocalNode());
504 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
505 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
506
507 // Test the should_post_on_exceptions_flag vs. 0
508 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
509 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
510
511 // Branch to slow_path if should_post_on_exceptions_flag was true
512 { BuildCutout unless(this, tst, PROB_MAX);
513 // Do not try anything fancy if we're notifying the VM on every throw.
514 // Cf. case Bytecodes::_athrow in parse2.cpp.
515 uncommon_trap(reason, Deoptimization::Action_none,
516 (ciKlass*)NULL, (char*)NULL, must_throw);
517 }
518
519 }
520
521 //------------------------------builtin_throw----------------------------------
builtin_throw(Deoptimization::DeoptReason reason,Node * arg)522 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
523 bool must_throw = true;
524
525 if (env()->jvmti_can_post_on_exceptions()) {
526 // check if we must post exception events, take uncommon trap if so
527 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
528 // here if should_post_on_exceptions is false
529 // continue on with the normal codegen
530 }
531
532 // If this particular condition has not yet happened at this
533 // bytecode, then use the uncommon trap mechanism, and allow for
534 // a future recompilation if several traps occur here.
535 // If the throw is hot, try to use a more complicated inline mechanism
536 // which keeps execution inside the compiled code.
537 bool treat_throw_as_hot = false;
538 ciMethodData* md = method()->method_data();
539
540 if (ProfileTraps) {
541 if (too_many_traps(reason)) {
542 treat_throw_as_hot = true;
543 }
544 // (If there is no MDO at all, assume it is early in
545 // execution, and that any deopts are part of the
546 // startup transient, and don't need to be remembered.)
547
548 // Also, if there is a local exception handler, treat all throws
549 // as hot if there has been at least one in this method.
550 if (C->trap_count(reason) != 0
551 && method()->method_data()->trap_count(reason) != 0
552 && has_ex_handler()) {
553 treat_throw_as_hot = true;
554 }
555 }
556
557 // If this throw happens frequently, an uncommon trap might cause
558 // a performance pothole. If there is a local exception handler,
559 // and if this particular bytecode appears to be deoptimizing often,
560 // let us handle the throw inline, with a preconstructed instance.
561 // Note: If the deopt count has blown up, the uncommon trap
562 // runtime is going to flush this nmethod, not matter what.
563 if (treat_throw_as_hot
564 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
565 // If the throw is local, we use a pre-existing instance and
566 // punt on the backtrace. This would lead to a missing backtrace
567 // (a repeat of 4292742) if the backtrace object is ever asked
568 // for its backtrace.
569 // Fixing this remaining case of 4292742 requires some flavor of
570 // escape analysis. Leave that for the future.
571 ciInstance* ex_obj = NULL;
572 switch (reason) {
573 case Deoptimization::Reason_null_check:
574 ex_obj = env()->NullPointerException_instance();
575 break;
576 case Deoptimization::Reason_div0_check:
577 ex_obj = env()->ArithmeticException_instance();
578 break;
579 case Deoptimization::Reason_range_check:
580 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
581 break;
582 case Deoptimization::Reason_class_check:
583 if (java_bc() == Bytecodes::_aastore) {
584 ex_obj = env()->ArrayStoreException_instance();
585 } else {
586 ex_obj = env()->ClassCastException_instance();
587 }
588 break;
589 default:
590 break;
591 }
592 if (failing()) { stop(); return; } // exception allocation might fail
593 if (ex_obj != NULL) {
594 // Cheat with a preallocated exception object.
595 if (C->log() != NULL)
596 C->log()->elem("hot_throw preallocated='1' reason='%s'",
597 Deoptimization::trap_reason_name(reason));
598 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
599 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
600
601 // Clear the detail message of the preallocated exception object.
602 // Weblogic sometimes mutates the detail message of exceptions
603 // using reflection.
604 int offset = java_lang_Throwable::get_detailMessage_offset();
605 const TypePtr* adr_typ = ex_con->add_offset(offset);
606
607 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
608 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
609 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
610
611 add_exception_state(make_exception_state(ex_node));
612 return;
613 }
614 }
615
616 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
617 // It won't be much cheaper than bailing to the interp., since we'll
618 // have to pass up all the debug-info, and the runtime will have to
619 // create the stack trace.
620
621 // Usual case: Bail to interpreter.
622 // Reserve the right to recompile if we haven't seen anything yet.
623
624 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
625 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
626 if (treat_throw_as_hot
627 && (method()->method_data()->trap_recompiled_at(bci(), m)
628 || C->too_many_traps(reason))) {
629 // We cannot afford to take more traps here. Suffer in the interpreter.
630 if (C->log() != NULL)
631 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
632 Deoptimization::trap_reason_name(reason),
633 C->trap_count(reason));
634 action = Deoptimization::Action_none;
635 }
636
637 // "must_throw" prunes the JVM state to include only the stack, if there
638 // are no local exception handlers. This should cut down on register
639 // allocation time and code size, by drastically reducing the number
640 // of in-edges on the call to the uncommon trap.
641
642 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
643 }
644
645
646 //----------------------------PreserveJVMState---------------------------------
PreserveJVMState(GraphKit * kit,bool clone_map)647 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
648 debug_only(kit->verify_map());
649 _kit = kit;
650 _map = kit->map(); // preserve the map
651 _sp = kit->sp();
652 kit->set_map(clone_map ? kit->clone_map() : NULL);
653 #ifdef ASSERT
654 _bci = kit->bci();
655 Parse* parser = kit->is_Parse();
656 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
657 _block = block;
658 #endif
659 }
~PreserveJVMState()660 PreserveJVMState::~PreserveJVMState() {
661 GraphKit* kit = _kit;
662 #ifdef ASSERT
663 assert(kit->bci() == _bci, "bci must not shift");
664 Parse* parser = kit->is_Parse();
665 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
666 assert(block == _block, "block must not shift");
667 #endif
668 kit->set_map(_map);
669 kit->set_sp(_sp);
670 }
671
672
673 //-----------------------------BuildCutout-------------------------------------
BuildCutout(GraphKit * kit,Node * p,float prob,float cnt)674 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
675 : PreserveJVMState(kit)
676 {
677 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
678 SafePointNode* outer_map = _map; // preserved map is caller's
679 SafePointNode* inner_map = kit->map();
680 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
681 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
682 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
683 }
~BuildCutout()684 BuildCutout::~BuildCutout() {
685 GraphKit* kit = _kit;
686 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
687 }
688
689 //---------------------------PreserveReexecuteState----------------------------
PreserveReexecuteState(GraphKit * kit)690 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
691 assert(!kit->stopped(), "must call stopped() before");
692 _kit = kit;
693 _sp = kit->sp();
694 _reexecute = kit->jvms()->_reexecute;
695 }
~PreserveReexecuteState()696 PreserveReexecuteState::~PreserveReexecuteState() {
697 if (_kit->stopped()) return;
698 _kit->jvms()->_reexecute = _reexecute;
699 _kit->set_sp(_sp);
700 }
701
702 //------------------------------clone_map--------------------------------------
703 // Implementation of PreserveJVMState
704 //
705 // Only clone_map(...) here. If this function is only used in the
706 // PreserveJVMState class we may want to get rid of this extra
707 // function eventually and do it all there.
708
clone_map()709 SafePointNode* GraphKit::clone_map() {
710 if (map() == NULL) return NULL;
711
712 // Clone the memory edge first
713 Node* mem = MergeMemNode::make(map()->memory());
714 gvn().set_type_bottom(mem);
715
716 SafePointNode *clonemap = (SafePointNode*)map()->clone();
717 JVMState* jvms = this->jvms();
718 JVMState* clonejvms = jvms->clone_shallow(C);
719 clonemap->set_memory(mem);
720 clonemap->set_jvms(clonejvms);
721 clonejvms->set_map(clonemap);
722 record_for_igvn(clonemap);
723 gvn().set_type_bottom(clonemap);
724 return clonemap;
725 }
726
727
728 //-----------------------------set_map_clone-----------------------------------
set_map_clone(SafePointNode * m)729 void GraphKit::set_map_clone(SafePointNode* m) {
730 _map = m;
731 _map = clone_map();
732 _map->set_next_exception(NULL);
733 debug_only(verify_map());
734 }
735
736
737 //----------------------------kill_dead_locals---------------------------------
738 // Detect any locals which are known to be dead, and force them to top.
kill_dead_locals()739 void GraphKit::kill_dead_locals() {
740 // Consult the liveness information for the locals. If any
741 // of them are unused, then they can be replaced by top(). This
742 // should help register allocation time and cut down on the size
743 // of the deoptimization information.
744
745 // This call is made from many of the bytecode handling
746 // subroutines called from the Big Switch in do_one_bytecode.
747 // Every bytecode which might include a slow path is responsible
748 // for killing its dead locals. The more consistent we
749 // are about killing deads, the fewer useless phis will be
750 // constructed for them at various merge points.
751
752 // bci can be -1 (InvocationEntryBci). We return the entry
753 // liveness for the method.
754
755 if (method() == NULL || method()->code_size() == 0) {
756 // We are building a graph for a call to a native method.
757 // All locals are live.
758 return;
759 }
760
761 ResourceMark rm;
762
763 // Consult the liveness information for the locals. If any
764 // of them are unused, then they can be replaced by top(). This
765 // should help register allocation time and cut down on the size
766 // of the deoptimization information.
767 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
768
769 int len = (int)live_locals.size();
770 assert(len <= jvms()->loc_size(), "too many live locals");
771 for (int local = 0; local < len; local++) {
772 if (!live_locals.at(local)) {
773 set_local(local, top());
774 }
775 }
776 }
777
778 #ifdef ASSERT
779 //-------------------------dead_locals_are_killed------------------------------
780 // Return true if all dead locals are set to top in the map.
781 // Used to assert "clean" debug info at various points.
dead_locals_are_killed()782 bool GraphKit::dead_locals_are_killed() {
783 if (method() == NULL || method()->code_size() == 0) {
784 // No locals need to be dead, so all is as it should be.
785 return true;
786 }
787
788 // Make sure somebody called kill_dead_locals upstream.
789 ResourceMark rm;
790 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
791 if (jvms->loc_size() == 0) continue; // no locals to consult
792 SafePointNode* map = jvms->map();
793 ciMethod* method = jvms->method();
794 int bci = jvms->bci();
795 if (jvms == this->jvms()) {
796 bci = this->bci(); // it might not yet be synched
797 }
798 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
799 int len = (int)live_locals.size();
800 if (!live_locals.is_valid() || len == 0)
801 // This method is trivial, or is poisoned by a breakpoint.
802 return true;
803 assert(len == jvms->loc_size(), "live map consistent with locals map");
804 for (int local = 0; local < len; local++) {
805 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
806 if (PrintMiscellaneous && (Verbose || WizardMode)) {
807 tty->print_cr("Zombie local %d: ", local);
808 jvms->dump();
809 }
810 return false;
811 }
812 }
813 }
814 return true;
815 }
816
817 #endif //ASSERT
818
819 // Helper function for enforcing certain bytecodes to reexecute if
820 // deoptimization happens
should_reexecute_implied_by_bytecode(JVMState * jvms,bool is_anewarray)821 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
822 ciMethod* cur_method = jvms->method();
823 int cur_bci = jvms->bci();
824 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
825 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
826 return Interpreter::bytecode_should_reexecute(code) ||
827 (is_anewarray && code == Bytecodes::_multianewarray);
828 // Reexecute _multianewarray bytecode which was replaced with
829 // sequence of [a]newarray. See Parse::do_multianewarray().
830 //
831 // Note: interpreter should not have it set since this optimization
832 // is limited by dimensions and guarded by flag so in some cases
833 // multianewarray() runtime calls will be generated and
834 // the bytecode should not be reexecutes (stack will not be reset).
835 } else
836 return false;
837 }
838
839 // Helper function for adding JVMState and debug information to node
add_safepoint_edges(SafePointNode * call,bool must_throw)840 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
841 // Add the safepoint edges to the call (or other safepoint).
842
843 // Make sure dead locals are set to top. This
844 // should help register allocation time and cut down on the size
845 // of the deoptimization information.
846 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
847
848 // Walk the inline list to fill in the correct set of JVMState's
849 // Also fill in the associated edges for each JVMState.
850
851 // If the bytecode needs to be reexecuted we need to put
852 // the arguments back on the stack.
853 const bool should_reexecute = jvms()->should_reexecute();
854 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
855
856 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
857 // undefined if the bci is different. This is normal for Parse but it
858 // should not happen for LibraryCallKit because only one bci is processed.
859 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
860 "in LibraryCallKit the reexecute bit should not change");
861
862 // If we are guaranteed to throw, we can prune everything but the
863 // input to the current bytecode.
864 bool can_prune_locals = false;
865 uint stack_slots_not_pruned = 0;
866 int inputs = 0, depth = 0;
867 if (must_throw) {
868 assert(method() == youngest_jvms->method(), "sanity");
869 if (compute_stack_effects(inputs, depth)) {
870 can_prune_locals = true;
871 stack_slots_not_pruned = inputs;
872 }
873 }
874
875 if (env()->should_retain_local_variables()) {
876 // At any safepoint, this method can get breakpointed, which would
877 // then require an immediate deoptimization.
878 can_prune_locals = false; // do not prune locals
879 stack_slots_not_pruned = 0;
880 }
881
882 // do not scribble on the input jvms
883 JVMState* out_jvms = youngest_jvms->clone_deep(C);
884 call->set_jvms(out_jvms); // Start jvms list for call node
885
886 // For a known set of bytecodes, the interpreter should reexecute them if
887 // deoptimization happens. We set the reexecute state for them here
888 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
889 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
890 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
891 }
892
893 // Presize the call:
894 DEBUG_ONLY(uint non_debug_edges = call->req());
895 call->add_req_batch(top(), youngest_jvms->debug_depth());
896 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
897
898 // Set up edges so that the call looks like this:
899 // Call [state:] ctl io mem fptr retadr
900 // [parms:] parm0 ... parmN
901 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
902 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
903 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
904 // Note that caller debug info precedes callee debug info.
905
906 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
907 uint debug_ptr = call->req();
908
909 // Loop over the map input edges associated with jvms, add them
910 // to the call node, & reset all offsets to match call node array.
911 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
912 uint debug_end = debug_ptr;
913 uint debug_start = debug_ptr - in_jvms->debug_size();
914 debug_ptr = debug_start; // back up the ptr
915
916 uint p = debug_start; // walks forward in [debug_start, debug_end)
917 uint j, k, l;
918 SafePointNode* in_map = in_jvms->map();
919 out_jvms->set_map(call);
920
921 if (can_prune_locals) {
922 assert(in_jvms->method() == out_jvms->method(), "sanity");
923 // If the current throw can reach an exception handler in this JVMS,
924 // then we must keep everything live that can reach that handler.
925 // As a quick and dirty approximation, we look for any handlers at all.
926 if (in_jvms->method()->has_exception_handlers()) {
927 can_prune_locals = false;
928 }
929 }
930
931 // Add the Locals
932 k = in_jvms->locoff();
933 l = in_jvms->loc_size();
934 out_jvms->set_locoff(p);
935 if (!can_prune_locals) {
936 for (j = 0; j < l; j++)
937 call->set_req(p++, in_map->in(k+j));
938 } else {
939 p += l; // already set to top above by add_req_batch
940 }
941
942 // Add the Expression Stack
943 k = in_jvms->stkoff();
944 l = in_jvms->sp();
945 out_jvms->set_stkoff(p);
946 if (!can_prune_locals) {
947 for (j = 0; j < l; j++)
948 call->set_req(p++, in_map->in(k+j));
949 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
950 // Divide stack into {S0,...,S1}, where S0 is set to top.
951 uint s1 = stack_slots_not_pruned;
952 stack_slots_not_pruned = 0; // for next iteration
953 if (s1 > l) s1 = l;
954 uint s0 = l - s1;
955 p += s0; // skip the tops preinstalled by add_req_batch
956 for (j = s0; j < l; j++)
957 call->set_req(p++, in_map->in(k+j));
958 } else {
959 p += l; // already set to top above by add_req_batch
960 }
961
962 // Add the Monitors
963 k = in_jvms->monoff();
964 l = in_jvms->mon_size();
965 out_jvms->set_monoff(p);
966 for (j = 0; j < l; j++)
967 call->set_req(p++, in_map->in(k+j));
968
969 // Copy any scalar object fields.
970 k = in_jvms->scloff();
971 l = in_jvms->scl_size();
972 out_jvms->set_scloff(p);
973 for (j = 0; j < l; j++)
974 call->set_req(p++, in_map->in(k+j));
975
976 // Finish the new jvms.
977 out_jvms->set_endoff(p);
978
979 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
980 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
981 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
982 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
983 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
984 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
985
986 // Update the two tail pointers in parallel.
987 out_jvms = out_jvms->caller();
988 in_jvms = in_jvms->caller();
989 }
990
991 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
992
993 // Test the correctness of JVMState::debug_xxx accessors:
994 assert(call->jvms()->debug_start() == non_debug_edges, "");
995 assert(call->jvms()->debug_end() == call->req(), "");
996 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
997 }
998
compute_stack_effects(int & inputs,int & depth)999 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1000 Bytecodes::Code code = java_bc();
1001 if (code == Bytecodes::_wide) {
1002 code = method()->java_code_at_bci(bci() + 1);
1003 }
1004
1005 BasicType rtype = T_ILLEGAL;
1006 int rsize = 0;
1007
1008 if (code != Bytecodes::_illegal) {
1009 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1010 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1011 if (rtype < T_CONFLICT)
1012 rsize = type2size[rtype];
1013 }
1014
1015 switch (code) {
1016 case Bytecodes::_illegal:
1017 return false;
1018
1019 case Bytecodes::_ldc:
1020 case Bytecodes::_ldc_w:
1021 case Bytecodes::_ldc2_w:
1022 inputs = 0;
1023 break;
1024
1025 case Bytecodes::_dup: inputs = 1; break;
1026 case Bytecodes::_dup_x1: inputs = 2; break;
1027 case Bytecodes::_dup_x2: inputs = 3; break;
1028 case Bytecodes::_dup2: inputs = 2; break;
1029 case Bytecodes::_dup2_x1: inputs = 3; break;
1030 case Bytecodes::_dup2_x2: inputs = 4; break;
1031 case Bytecodes::_swap: inputs = 2; break;
1032 case Bytecodes::_arraylength: inputs = 1; break;
1033
1034 case Bytecodes::_getstatic:
1035 case Bytecodes::_putstatic:
1036 case Bytecodes::_getfield:
1037 case Bytecodes::_putfield:
1038 {
1039 bool ignored_will_link;
1040 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1041 int size = field->type()->size();
1042 bool is_get = (depth >= 0), is_static = (depth & 1);
1043 inputs = (is_static ? 0 : 1);
1044 if (is_get) {
1045 depth = size - inputs;
1046 } else {
1047 inputs += size; // putxxx pops the value from the stack
1048 depth = - inputs;
1049 }
1050 }
1051 break;
1052
1053 case Bytecodes::_invokevirtual:
1054 case Bytecodes::_invokespecial:
1055 case Bytecodes::_invokestatic:
1056 case Bytecodes::_invokedynamic:
1057 case Bytecodes::_invokeinterface:
1058 {
1059 bool ignored_will_link;
1060 ciSignature* declared_signature = NULL;
1061 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1062 assert(declared_signature != NULL, "cannot be null");
1063 inputs = declared_signature->arg_size_for_bc(code);
1064 int size = declared_signature->return_type()->size();
1065 depth = size - inputs;
1066 }
1067 break;
1068
1069 case Bytecodes::_multianewarray:
1070 {
1071 ciBytecodeStream iter(method());
1072 iter.reset_to_bci(bci());
1073 iter.next();
1074 inputs = iter.get_dimensions();
1075 assert(rsize == 1, "");
1076 depth = rsize - inputs;
1077 }
1078 break;
1079
1080 case Bytecodes::_ireturn:
1081 case Bytecodes::_lreturn:
1082 case Bytecodes::_freturn:
1083 case Bytecodes::_dreturn:
1084 case Bytecodes::_areturn:
1085 assert(rsize == -depth, "");
1086 inputs = rsize;
1087 break;
1088
1089 case Bytecodes::_jsr:
1090 case Bytecodes::_jsr_w:
1091 inputs = 0;
1092 depth = 1; // S.B. depth=1, not zero
1093 break;
1094
1095 default:
1096 // bytecode produces a typed result
1097 inputs = rsize - depth;
1098 assert(inputs >= 0, "");
1099 break;
1100 }
1101
1102 #ifdef ASSERT
1103 // spot check
1104 int outputs = depth + inputs;
1105 assert(outputs >= 0, "sanity");
1106 switch (code) {
1107 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1108 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1109 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1110 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1111 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1112 default: break;
1113 }
1114 #endif //ASSERT
1115
1116 return true;
1117 }
1118
1119
1120
1121 //------------------------------basic_plus_adr---------------------------------
basic_plus_adr(Node * base,Node * ptr,Node * offset)1122 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1123 // short-circuit a common case
1124 if (offset == intcon(0)) return ptr;
1125 return _gvn.transform( new AddPNode(base, ptr, offset) );
1126 }
1127
ConvI2L(Node * offset)1128 Node* GraphKit::ConvI2L(Node* offset) {
1129 // short-circuit a common case
1130 jint offset_con = find_int_con(offset, Type::OffsetBot);
1131 if (offset_con != Type::OffsetBot) {
1132 return longcon((jlong) offset_con);
1133 }
1134 return _gvn.transform( new ConvI2LNode(offset));
1135 }
1136
ConvI2UL(Node * offset)1137 Node* GraphKit::ConvI2UL(Node* offset) {
1138 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1139 if (offset_con != (juint) Type::OffsetBot) {
1140 return longcon((julong) offset_con);
1141 }
1142 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1143 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1144 return _gvn.transform( new AndLNode(conv, mask) );
1145 }
1146
ConvL2I(Node * offset)1147 Node* GraphKit::ConvL2I(Node* offset) {
1148 // short-circuit a common case
1149 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1150 if (offset_con != (jlong)Type::OffsetBot) {
1151 return intcon((int) offset_con);
1152 }
1153 return _gvn.transform( new ConvL2INode(offset));
1154 }
1155
1156 //-------------------------load_object_klass-----------------------------------
load_object_klass(Node * obj)1157 Node* GraphKit::load_object_klass(Node* obj) {
1158 // Special-case a fresh allocation to avoid building nodes:
1159 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1160 if (akls != NULL) return akls;
1161 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1162 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1163 }
1164
1165 //-------------------------load_array_length-----------------------------------
load_array_length(Node * array)1166 Node* GraphKit::load_array_length(Node* array) {
1167 // Special-case a fresh allocation to avoid building nodes:
1168 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1169 Node *alen;
1170 if (alloc == NULL) {
1171 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1172 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1173 } else {
1174 alen = alloc->Ideal_length();
1175 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1176 if (ccast != alen) {
1177 alen = _gvn.transform(ccast);
1178 }
1179 }
1180 return alen;
1181 }
1182
1183 //------------------------------do_null_check----------------------------------
1184 // Helper function to do a NULL pointer check. Returned value is
1185 // the incoming address with NULL casted away. You are allowed to use the
1186 // not-null value only if you are control dependent on the test.
1187 #ifndef PRODUCT
1188 extern int explicit_null_checks_inserted,
1189 explicit_null_checks_elided;
1190 #endif
null_check_common(Node * value,BasicType type,bool assert_null,Node ** null_control,bool speculative)1191 Node* GraphKit::null_check_common(Node* value, BasicType type,
1192 // optional arguments for variations:
1193 bool assert_null,
1194 Node* *null_control,
1195 bool speculative) {
1196 assert(!assert_null || null_control == NULL, "not both at once");
1197 if (stopped()) return top();
1198 NOT_PRODUCT(explicit_null_checks_inserted++);
1199
1200 // Construct NULL check
1201 Node *chk = NULL;
1202 switch(type) {
1203 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1204 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1205 case T_ARRAY : // fall through
1206 type = T_OBJECT; // simplify further tests
1207 case T_OBJECT : {
1208 const Type *t = _gvn.type( value );
1209
1210 const TypeOopPtr* tp = t->isa_oopptr();
1211 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1212 // Only for do_null_check, not any of its siblings:
1213 && !assert_null && null_control == NULL) {
1214 // Usually, any field access or invocation on an unloaded oop type
1215 // will simply fail to link, since the statically linked class is
1216 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1217 // the static class is loaded but the sharper oop type is not.
1218 // Rather than checking for this obscure case in lots of places,
1219 // we simply observe that a null check on an unloaded class
1220 // will always be followed by a nonsense operation, so we
1221 // can just issue the uncommon trap here.
1222 // Our access to the unloaded class will only be correct
1223 // after it has been loaded and initialized, which requires
1224 // a trip through the interpreter.
1225 #ifndef PRODUCT
1226 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1227 #endif
1228 uncommon_trap(Deoptimization::Reason_unloaded,
1229 Deoptimization::Action_reinterpret,
1230 tp->klass(), "!loaded");
1231 return top();
1232 }
1233
1234 if (assert_null) {
1235 // See if the type is contained in NULL_PTR.
1236 // If so, then the value is already null.
1237 if (t->higher_equal(TypePtr::NULL_PTR)) {
1238 NOT_PRODUCT(explicit_null_checks_elided++);
1239 return value; // Elided null assert quickly!
1240 }
1241 } else {
1242 // See if mixing in the NULL pointer changes type.
1243 // If so, then the NULL pointer was not allowed in the original
1244 // type. In other words, "value" was not-null.
1245 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1246 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1247 NOT_PRODUCT(explicit_null_checks_elided++);
1248 return value; // Elided null check quickly!
1249 }
1250 }
1251 chk = new CmpPNode( value, null() );
1252 break;
1253 }
1254
1255 default:
1256 fatal("unexpected type: %s", type2name(type));
1257 }
1258 assert(chk != NULL, "sanity check");
1259 chk = _gvn.transform(chk);
1260
1261 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1262 BoolNode *btst = new BoolNode( chk, btest);
1263 Node *tst = _gvn.transform( btst );
1264
1265 //-----------
1266 // if peephole optimizations occurred, a prior test existed.
1267 // If a prior test existed, maybe it dominates as we can avoid this test.
1268 if (tst != btst && type == T_OBJECT) {
1269 // At this point we want to scan up the CFG to see if we can
1270 // find an identical test (and so avoid this test altogether).
1271 Node *cfg = control();
1272 int depth = 0;
1273 while( depth < 16 ) { // Limit search depth for speed
1274 if( cfg->Opcode() == Op_IfTrue &&
1275 cfg->in(0)->in(1) == tst ) {
1276 // Found prior test. Use "cast_not_null" to construct an identical
1277 // CastPP (and hence hash to) as already exists for the prior test.
1278 // Return that casted value.
1279 if (assert_null) {
1280 replace_in_map(value, null());
1281 return null(); // do not issue the redundant test
1282 }
1283 Node *oldcontrol = control();
1284 set_control(cfg);
1285 Node *res = cast_not_null(value);
1286 set_control(oldcontrol);
1287 NOT_PRODUCT(explicit_null_checks_elided++);
1288 return res;
1289 }
1290 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1291 if (cfg == NULL) break; // Quit at region nodes
1292 depth++;
1293 }
1294 }
1295
1296 //-----------
1297 // Branch to failure if null
1298 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1299 Deoptimization::DeoptReason reason;
1300 if (assert_null) {
1301 reason = Deoptimization::reason_null_assert(speculative);
1302 } else if (type == T_OBJECT) {
1303 reason = Deoptimization::reason_null_check(speculative);
1304 } else {
1305 reason = Deoptimization::Reason_div0_check;
1306 }
1307 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1308 // ciMethodData::has_trap_at will return a conservative -1 if any
1309 // must-be-null assertion has failed. This could cause performance
1310 // problems for a method after its first do_null_assert failure.
1311 // Consider using 'Reason_class_check' instead?
1312
1313 // To cause an implicit null check, we set the not-null probability
1314 // to the maximum (PROB_MAX). For an explicit check the probability
1315 // is set to a smaller value.
1316 if (null_control != NULL || too_many_traps(reason)) {
1317 // probability is less likely
1318 ok_prob = PROB_LIKELY_MAG(3);
1319 } else if (!assert_null &&
1320 (ImplicitNullCheckThreshold > 0) &&
1321 method() != NULL &&
1322 (method()->method_data()->trap_count(reason)
1323 >= (uint)ImplicitNullCheckThreshold)) {
1324 ok_prob = PROB_LIKELY_MAG(3);
1325 }
1326
1327 if (null_control != NULL) {
1328 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1329 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1330 set_control( _gvn.transform( new IfTrueNode(iff)));
1331 #ifndef PRODUCT
1332 if (null_true == top()) {
1333 explicit_null_checks_elided++;
1334 }
1335 #endif
1336 (*null_control) = null_true;
1337 } else {
1338 BuildCutout unless(this, tst, ok_prob);
1339 // Check for optimizer eliding test at parse time
1340 if (stopped()) {
1341 // Failure not possible; do not bother making uncommon trap.
1342 NOT_PRODUCT(explicit_null_checks_elided++);
1343 } else if (assert_null) {
1344 uncommon_trap(reason,
1345 Deoptimization::Action_make_not_entrant,
1346 NULL, "assert_null");
1347 } else {
1348 replace_in_map(value, zerocon(type));
1349 builtin_throw(reason);
1350 }
1351 }
1352
1353 // Must throw exception, fall-thru not possible?
1354 if (stopped()) {
1355 return top(); // No result
1356 }
1357
1358 if (assert_null) {
1359 // Cast obj to null on this path.
1360 replace_in_map(value, zerocon(type));
1361 return zerocon(type);
1362 }
1363
1364 // Cast obj to not-null on this path, if there is no null_control.
1365 // (If there is a null_control, a non-null value may come back to haunt us.)
1366 return cast_not_null(value, (null_control == NULL || (*null_control) == top()));
1367 }
1368
1369
1370 //------------------------------cast_not_null----------------------------------
1371 // Cast obj to not-null on this path
cast_not_null(Node * obj,bool do_replace_in_map)1372 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1373 Node* cast = NULL;
1374 const Type* t = _gvn.type(obj);
1375 if (t->make_ptr() != NULL) {
1376 const Type* t_not_null = t->join_speculative(TypePtr::NOTNULL);
1377 // Object is already not-null?
1378 if (t == t_not_null) {
1379 return obj;
1380 }
1381 cast = ConstraintCastNode::make_cast(Op_CastPP, control(), obj, t_not_null, false);
1382 } else if (t->isa_int() != NULL) {
1383 cast = ConstraintCastNode::make_cast(Op_CastII, control(), obj, TypeInt::INT, true);
1384 } else if (t->isa_long() != NULL) {
1385 cast = ConstraintCastNode::make_cast(Op_CastLL, control(), obj, TypeLong::LONG, true);
1386 } else {
1387 fatal("unexpected type: %s", type2name(t->basic_type()));
1388 }
1389 cast = _gvn.transform(cast);
1390
1391 // Scan for instances of 'obj' in the current JVM mapping.
1392 // These instances are known to be not-null after the test.
1393 if (do_replace_in_map) {
1394 replace_in_map(obj, cast);
1395 }
1396 return cast;
1397 }
1398
1399 // Sometimes in intrinsics, we implicitly know an object is not null
1400 // (there's no actual null check) so we can cast it to not null. In
1401 // the course of optimizations, the input to the cast can become null.
1402 // In that case that data path will die and we need the control path
1403 // to become dead as well to keep the graph consistent. So we have to
1404 // add a check for null for which one branch can't be taken. It uses
1405 // an Opaque4 node that will cause the check to be removed after loop
1406 // opts so the test goes away and the compiled code doesn't execute a
1407 // useless check.
must_be_not_null(Node * value,bool do_replace_in_map)1408 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1409 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1410 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1411 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1412 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1413 _gvn.set_type(iff, iff->Value(&_gvn));
1414 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1415 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1416 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1417 C->root()->add_req(halt);
1418 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1419 set_control(if_t);
1420 return cast_not_null(value, do_replace_in_map);
1421 }
1422
1423
1424 //--------------------------replace_in_map-------------------------------------
replace_in_map(Node * old,Node * neww)1425 void GraphKit::replace_in_map(Node* old, Node* neww) {
1426 if (old == neww) {
1427 return;
1428 }
1429
1430 map()->replace_edge(old, neww);
1431
1432 // Note: This operation potentially replaces any edge
1433 // on the map. This includes locals, stack, and monitors
1434 // of the current (innermost) JVM state.
1435
1436 // don't let inconsistent types from profiling escape this
1437 // method
1438
1439 const Type* told = _gvn.type(old);
1440 const Type* tnew = _gvn.type(neww);
1441
1442 if (!tnew->higher_equal(told)) {
1443 return;
1444 }
1445
1446 map()->record_replaced_node(old, neww);
1447 }
1448
1449
1450 //=============================================================================
1451 //--------------------------------memory---------------------------------------
memory(uint alias_idx)1452 Node* GraphKit::memory(uint alias_idx) {
1453 MergeMemNode* mem = merged_memory();
1454 Node* p = mem->memory_at(alias_idx);
1455 _gvn.set_type(p, Type::MEMORY); // must be mapped
1456 return p;
1457 }
1458
1459 //-----------------------------reset_memory------------------------------------
reset_memory()1460 Node* GraphKit::reset_memory() {
1461 Node* mem = map()->memory();
1462 // do not use this node for any more parsing!
1463 debug_only( map()->set_memory((Node*)NULL) );
1464 return _gvn.transform( mem );
1465 }
1466
1467 //------------------------------set_all_memory---------------------------------
set_all_memory(Node * newmem)1468 void GraphKit::set_all_memory(Node* newmem) {
1469 Node* mergemem = MergeMemNode::make(newmem);
1470 gvn().set_type_bottom(mergemem);
1471 map()->set_memory(mergemem);
1472 }
1473
1474 //------------------------------set_all_memory_call----------------------------
set_all_memory_call(Node * call,bool separate_io_proj)1475 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1476 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1477 set_all_memory(newmem);
1478 }
1479
1480 //=============================================================================
1481 //
1482 // parser factory methods for MemNodes
1483 //
1484 // These are layered on top of the factory methods in LoadNode and StoreNode,
1485 // and integrate with the parser's memory state and _gvn engine.
1486 //
1487
1488 // factory methods in "int adr_idx"
make_load(Node * ctl,Node * adr,const Type * t,BasicType bt,int adr_idx,MemNode::MemOrd mo,LoadNode::ControlDependency control_dependency,bool require_atomic_access,bool unaligned,bool mismatched,bool unsafe,uint8_t barrier_data)1489 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1490 int adr_idx,
1491 MemNode::MemOrd mo,
1492 LoadNode::ControlDependency control_dependency,
1493 bool require_atomic_access,
1494 bool unaligned,
1495 bool mismatched,
1496 bool unsafe,
1497 uint8_t barrier_data) {
1498 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1499 const TypePtr* adr_type = NULL; // debug-mode-only argument
1500 debug_only(adr_type = C->get_adr_type(adr_idx));
1501 Node* mem = memory(adr_idx);
1502 Node* ld;
1503 if (require_atomic_access && bt == T_LONG) {
1504 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1505 } else if (require_atomic_access && bt == T_DOUBLE) {
1506 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1507 } else {
1508 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1509 }
1510 ld = _gvn.transform(ld);
1511 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1512 // Improve graph before escape analysis and boxing elimination.
1513 record_for_igvn(ld);
1514 }
1515 return ld;
1516 }
1517
store_to_memory(Node * ctl,Node * adr,Node * val,BasicType bt,int adr_idx,MemNode::MemOrd mo,bool require_atomic_access,bool unaligned,bool mismatched,bool unsafe)1518 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1519 int adr_idx,
1520 MemNode::MemOrd mo,
1521 bool require_atomic_access,
1522 bool unaligned,
1523 bool mismatched,
1524 bool unsafe) {
1525 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1526 const TypePtr* adr_type = NULL;
1527 debug_only(adr_type = C->get_adr_type(adr_idx));
1528 Node *mem = memory(adr_idx);
1529 Node* st;
1530 if (require_atomic_access && bt == T_LONG) {
1531 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1532 } else if (require_atomic_access && bt == T_DOUBLE) {
1533 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1534 } else {
1535 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1536 }
1537 if (unaligned) {
1538 st->as_Store()->set_unaligned_access();
1539 }
1540 if (mismatched) {
1541 st->as_Store()->set_mismatched_access();
1542 }
1543 if (unsafe) {
1544 st->as_Store()->set_unsafe_access();
1545 }
1546 st = _gvn.transform(st);
1547 set_memory(st, adr_idx);
1548 // Back-to-back stores can only remove intermediate store with DU info
1549 // so push on worklist for optimizer.
1550 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1551 record_for_igvn(st);
1552
1553 return st;
1554 }
1555
access_store_at(Node * obj,Node * adr,const TypePtr * adr_type,Node * val,const Type * val_type,BasicType bt,DecoratorSet decorators)1556 Node* GraphKit::access_store_at(Node* obj,
1557 Node* adr,
1558 const TypePtr* adr_type,
1559 Node* val,
1560 const Type* val_type,
1561 BasicType bt,
1562 DecoratorSet decorators) {
1563 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1564 // could be delayed during Parse (for example, in adjust_map_after_if()).
1565 // Execute transformation here to avoid barrier generation in such case.
1566 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1567 val = _gvn.makecon(TypePtr::NULL_PTR);
1568 }
1569
1570 if (stopped()) {
1571 return top(); // Dead path ?
1572 }
1573
1574 assert(val != NULL, "not dead path");
1575
1576 C2AccessValuePtr addr(adr, adr_type);
1577 C2AccessValue value(val, val_type);
1578 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1579 if (access.is_raw()) {
1580 return _barrier_set->BarrierSetC2::store_at(access, value);
1581 } else {
1582 return _barrier_set->store_at(access, value);
1583 }
1584 }
1585
access_load_at(Node * obj,Node * adr,const TypePtr * adr_type,const Type * val_type,BasicType bt,DecoratorSet decorators)1586 Node* GraphKit::access_load_at(Node* obj, // containing obj
1587 Node* adr, // actual adress to store val at
1588 const TypePtr* adr_type,
1589 const Type* val_type,
1590 BasicType bt,
1591 DecoratorSet decorators) {
1592 if (stopped()) {
1593 return top(); // Dead path ?
1594 }
1595
1596 C2AccessValuePtr addr(adr, adr_type);
1597 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1598 if (access.is_raw()) {
1599 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1600 } else {
1601 return _barrier_set->load_at(access, val_type);
1602 }
1603 }
1604
access_load(Node * adr,const Type * val_type,BasicType bt,DecoratorSet decorators)1605 Node* GraphKit::access_load(Node* adr, // actual adress to load val at
1606 const Type* val_type,
1607 BasicType bt,
1608 DecoratorSet decorators) {
1609 if (stopped()) {
1610 return top(); // Dead path ?
1611 }
1612
1613 C2AccessValuePtr addr(adr, NULL);
1614 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1615 if (access.is_raw()) {
1616 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1617 } else {
1618 return _barrier_set->load_at(access, val_type);
1619 }
1620 }
1621
access_atomic_cmpxchg_val_at(Node * obj,Node * adr,const TypePtr * adr_type,int alias_idx,Node * expected_val,Node * new_val,const Type * value_type,BasicType bt,DecoratorSet decorators)1622 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1623 Node* adr,
1624 const TypePtr* adr_type,
1625 int alias_idx,
1626 Node* expected_val,
1627 Node* new_val,
1628 const Type* value_type,
1629 BasicType bt,
1630 DecoratorSet decorators) {
1631 C2AccessValuePtr addr(adr, adr_type);
1632 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1633 bt, obj, addr, alias_idx);
1634 if (access.is_raw()) {
1635 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1636 } else {
1637 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1638 }
1639 }
1640
access_atomic_cmpxchg_bool_at(Node * obj,Node * adr,const TypePtr * adr_type,int alias_idx,Node * expected_val,Node * new_val,const Type * value_type,BasicType bt,DecoratorSet decorators)1641 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1642 Node* adr,
1643 const TypePtr* adr_type,
1644 int alias_idx,
1645 Node* expected_val,
1646 Node* new_val,
1647 const Type* value_type,
1648 BasicType bt,
1649 DecoratorSet decorators) {
1650 C2AccessValuePtr addr(adr, adr_type);
1651 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1652 bt, obj, addr, alias_idx);
1653 if (access.is_raw()) {
1654 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1655 } else {
1656 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1657 }
1658 }
1659
access_atomic_xchg_at(Node * obj,Node * adr,const TypePtr * adr_type,int alias_idx,Node * new_val,const Type * value_type,BasicType bt,DecoratorSet decorators)1660 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1661 Node* adr,
1662 const TypePtr* adr_type,
1663 int alias_idx,
1664 Node* new_val,
1665 const Type* value_type,
1666 BasicType bt,
1667 DecoratorSet decorators) {
1668 C2AccessValuePtr addr(adr, adr_type);
1669 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1670 bt, obj, addr, alias_idx);
1671 if (access.is_raw()) {
1672 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1673 } else {
1674 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1675 }
1676 }
1677
access_atomic_add_at(Node * obj,Node * adr,const TypePtr * adr_type,int alias_idx,Node * new_val,const Type * value_type,BasicType bt,DecoratorSet decorators)1678 Node* GraphKit::access_atomic_add_at(Node* obj,
1679 Node* adr,
1680 const TypePtr* adr_type,
1681 int alias_idx,
1682 Node* new_val,
1683 const Type* value_type,
1684 BasicType bt,
1685 DecoratorSet decorators) {
1686 C2AccessValuePtr addr(adr, adr_type);
1687 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1688 if (access.is_raw()) {
1689 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1690 } else {
1691 return _barrier_set->atomic_add_at(access, new_val, value_type);
1692 }
1693 }
1694
access_clone(Node * src,Node * dst,Node * size,bool is_array)1695 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1696 return _barrier_set->clone(this, src, dst, size, is_array);
1697 }
1698
1699 //-------------------------array_element_address-------------------------
array_element_address(Node * ary,Node * idx,BasicType elembt,const TypeInt * sizetype,Node * ctrl)1700 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1701 const TypeInt* sizetype, Node* ctrl) {
1702 uint shift = exact_log2(type2aelembytes(elembt));
1703 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1704
1705 // short-circuit a common case (saves lots of confusing waste motion)
1706 jint idx_con = find_int_con(idx, -1);
1707 if (idx_con >= 0) {
1708 intptr_t offset = header + ((intptr_t)idx_con << shift);
1709 return basic_plus_adr(ary, offset);
1710 }
1711
1712 // must be correct type for alignment purposes
1713 Node* base = basic_plus_adr(ary, header);
1714 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1715 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1716 return basic_plus_adr(ary, base, scale);
1717 }
1718
1719 //-------------------------load_array_element-------------------------
load_array_element(Node * ctl,Node * ary,Node * idx,const TypeAryPtr * arytype)1720 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1721 const Type* elemtype = arytype->elem();
1722 BasicType elembt = elemtype->array_element_basic_type();
1723 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1724 if (elembt == T_NARROWOOP) {
1725 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1726 }
1727 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1728 return ld;
1729 }
1730
1731 //-------------------------set_arguments_for_java_call-------------------------
1732 // Arguments (pre-popped from the stack) are taken from the JVMS.
set_arguments_for_java_call(CallJavaNode * call)1733 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1734 // Add the call arguments:
1735 uint nargs = call->method()->arg_size();
1736 for (uint i = 0; i < nargs; i++) {
1737 Node* arg = argument(i);
1738 call->init_req(i + TypeFunc::Parms, arg);
1739 }
1740 }
1741
1742 //---------------------------set_edges_for_java_call---------------------------
1743 // Connect a newly created call into the current JVMS.
1744 // A return value node (if any) is returned from set_edges_for_java_call.
set_edges_for_java_call(CallJavaNode * call,bool must_throw,bool separate_io_proj)1745 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1746
1747 // Add the predefined inputs:
1748 call->init_req( TypeFunc::Control, control() );
1749 call->init_req( TypeFunc::I_O , i_o() );
1750 call->init_req( TypeFunc::Memory , reset_memory() );
1751 call->init_req( TypeFunc::FramePtr, frameptr() );
1752 call->init_req( TypeFunc::ReturnAdr, top() );
1753
1754 add_safepoint_edges(call, must_throw);
1755
1756 Node* xcall = _gvn.transform(call);
1757
1758 if (xcall == top()) {
1759 set_control(top());
1760 return;
1761 }
1762 assert(xcall == call, "call identity is stable");
1763
1764 // Re-use the current map to produce the result.
1765
1766 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1767 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1768 set_all_memory_call(xcall, separate_io_proj);
1769
1770 //return xcall; // no need, caller already has it
1771 }
1772
set_results_for_java_call(CallJavaNode * call,bool separate_io_proj,bool deoptimize)1773 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1774 if (stopped()) return top(); // maybe the call folded up?
1775
1776 // Capture the return value, if any.
1777 Node* ret;
1778 if (call->method() == NULL ||
1779 call->method()->return_type()->basic_type() == T_VOID)
1780 ret = top();
1781 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1782
1783 // Note: Since any out-of-line call can produce an exception,
1784 // we always insert an I_O projection from the call into the result.
1785
1786 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1787
1788 if (separate_io_proj) {
1789 // The caller requested separate projections be used by the fall
1790 // through and exceptional paths, so replace the projections for
1791 // the fall through path.
1792 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1793 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1794 }
1795 return ret;
1796 }
1797
1798 //--------------------set_predefined_input_for_runtime_call--------------------
1799 // Reading and setting the memory state is way conservative here.
1800 // The real problem is that I am not doing real Type analysis on memory,
1801 // so I cannot distinguish card mark stores from other stores. Across a GC
1802 // point the Store Barrier and the card mark memory has to agree. I cannot
1803 // have a card mark store and its barrier split across the GC point from
1804 // either above or below. Here I get that to happen by reading ALL of memory.
1805 // A better answer would be to separate out card marks from other memory.
1806 // For now, return the input memory state, so that it can be reused
1807 // after the call, if this call has restricted memory effects.
set_predefined_input_for_runtime_call(SafePointNode * call,Node * narrow_mem)1808 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1809 // Set fixed predefined input arguments
1810 Node* memory = reset_memory();
1811 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1812 call->init_req( TypeFunc::Control, control() );
1813 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1814 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1815 call->init_req( TypeFunc::FramePtr, frameptr() );
1816 call->init_req( TypeFunc::ReturnAdr, top() );
1817 return memory;
1818 }
1819
1820 //-------------------set_predefined_output_for_runtime_call--------------------
1821 // Set control and memory (not i_o) from the call.
1822 // If keep_mem is not NULL, use it for the output state,
1823 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1824 // If hook_mem is NULL, this call produces no memory effects at all.
1825 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1826 // then only that memory slice is taken from the call.
1827 // In the last case, we must put an appropriate memory barrier before
1828 // the call, so as to create the correct anti-dependencies on loads
1829 // preceding the call.
set_predefined_output_for_runtime_call(Node * call,Node * keep_mem,const TypePtr * hook_mem)1830 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1831 Node* keep_mem,
1832 const TypePtr* hook_mem) {
1833 // no i/o
1834 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1835 if (keep_mem) {
1836 // First clone the existing memory state
1837 set_all_memory(keep_mem);
1838 if (hook_mem != NULL) {
1839 // Make memory for the call
1840 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1841 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1842 // We also use hook_mem to extract specific effects from arraycopy stubs.
1843 set_memory(mem, hook_mem);
1844 }
1845 // ...else the call has NO memory effects.
1846
1847 // Make sure the call advertises its memory effects precisely.
1848 // This lets us build accurate anti-dependences in gcm.cpp.
1849 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1850 "call node must be constructed correctly");
1851 } else {
1852 assert(hook_mem == NULL, "");
1853 // This is not a "slow path" call; all memory comes from the call.
1854 set_all_memory_call(call);
1855 }
1856 }
1857
1858 // Keep track of MergeMems feeding into other MergeMems
add_mergemem_users_to_worklist(Unique_Node_List & wl,Node * mem)1859 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
1860 if (!mem->is_MergeMem()) {
1861 return;
1862 }
1863 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
1864 Node* use = i.get();
1865 if (use->is_MergeMem()) {
1866 wl.push(use);
1867 }
1868 }
1869 }
1870
1871 // Replace the call with the current state of the kit.
replace_call(CallNode * call,Node * result,bool do_replaced_nodes)1872 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1873 JVMState* ejvms = NULL;
1874 if (has_exceptions()) {
1875 ejvms = transfer_exceptions_into_jvms();
1876 }
1877
1878 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1879 ReplacedNodes replaced_nodes_exception;
1880 Node* ex_ctl = top();
1881
1882 SafePointNode* final_state = stop();
1883
1884 // Find all the needed outputs of this call
1885 CallProjections callprojs;
1886 call->extract_projections(&callprojs, true);
1887
1888 Unique_Node_List wl;
1889 Node* init_mem = call->in(TypeFunc::Memory);
1890 Node* final_mem = final_state->in(TypeFunc::Memory);
1891 Node* final_ctl = final_state->in(TypeFunc::Control);
1892 Node* final_io = final_state->in(TypeFunc::I_O);
1893
1894 // Replace all the old call edges with the edges from the inlining result
1895 if (callprojs.fallthrough_catchproj != NULL) {
1896 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1897 }
1898 if (callprojs.fallthrough_memproj != NULL) {
1899 if (final_mem->is_MergeMem()) {
1900 // Parser's exits MergeMem was not transformed but may be optimized
1901 final_mem = _gvn.transform(final_mem);
1902 }
1903 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1904 add_mergemem_users_to_worklist(wl, final_mem);
1905 }
1906 if (callprojs.fallthrough_ioproj != NULL) {
1907 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1908 }
1909
1910 // Replace the result with the new result if it exists and is used
1911 if (callprojs.resproj != NULL && result != NULL) {
1912 C->gvn_replace_by(callprojs.resproj, result);
1913 }
1914
1915 if (ejvms == NULL) {
1916 // No exception edges to simply kill off those paths
1917 if (callprojs.catchall_catchproj != NULL) {
1918 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1919 }
1920 if (callprojs.catchall_memproj != NULL) {
1921 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1922 }
1923 if (callprojs.catchall_ioproj != NULL) {
1924 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1925 }
1926 // Replace the old exception object with top
1927 if (callprojs.exobj != NULL) {
1928 C->gvn_replace_by(callprojs.exobj, C->top());
1929 }
1930 } else {
1931 GraphKit ekit(ejvms);
1932
1933 // Load my combined exception state into the kit, with all phis transformed:
1934 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1935 replaced_nodes_exception = ex_map->replaced_nodes();
1936
1937 Node* ex_oop = ekit.use_exception_state(ex_map);
1938
1939 if (callprojs.catchall_catchproj != NULL) {
1940 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1941 ex_ctl = ekit.control();
1942 }
1943 if (callprojs.catchall_memproj != NULL) {
1944 Node* ex_mem = ekit.reset_memory();
1945 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
1946 add_mergemem_users_to_worklist(wl, ex_mem);
1947 }
1948 if (callprojs.catchall_ioproj != NULL) {
1949 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1950 }
1951
1952 // Replace the old exception object with the newly created one
1953 if (callprojs.exobj != NULL) {
1954 C->gvn_replace_by(callprojs.exobj, ex_oop);
1955 }
1956 }
1957
1958 // Disconnect the call from the graph
1959 call->disconnect_inputs(NULL, C);
1960 C->gvn_replace_by(call, C->top());
1961
1962 // Clean up any MergeMems that feed other MergeMems since the
1963 // optimizer doesn't like that.
1964 while (wl.size() > 0) {
1965 _gvn.transform(wl.pop());
1966 }
1967
1968 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1969 replaced_nodes.apply(C, final_ctl);
1970 }
1971 if (!ex_ctl->is_top() && do_replaced_nodes) {
1972 replaced_nodes_exception.apply(C, ex_ctl);
1973 }
1974 }
1975
1976
1977 //------------------------------increment_counter------------------------------
1978 // for statistics: increment a VM counter by 1
1979
increment_counter(address counter_addr)1980 void GraphKit::increment_counter(address counter_addr) {
1981 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1982 increment_counter(adr1);
1983 }
1984
increment_counter(Node * counter_addr)1985 void GraphKit::increment_counter(Node* counter_addr) {
1986 int adr_type = Compile::AliasIdxRaw;
1987 Node* ctrl = control();
1988 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1989 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1990 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1991 }
1992
1993
1994 //------------------------------uncommon_trap----------------------------------
1995 // Bail out to the interpreter in mid-method. Implemented by calling the
1996 // uncommon_trap blob. This helper function inserts a runtime call with the
1997 // right debug info.
uncommon_trap(int trap_request,ciKlass * klass,const char * comment,bool must_throw,bool keep_exact_action)1998 void GraphKit::uncommon_trap(int trap_request,
1999 ciKlass* klass, const char* comment,
2000 bool must_throw,
2001 bool keep_exact_action) {
2002 if (failing()) stop();
2003 if (stopped()) return; // trap reachable?
2004
2005 // Note: If ProfileTraps is true, and if a deopt. actually
2006 // occurs here, the runtime will make sure an MDO exists. There is
2007 // no need to call method()->ensure_method_data() at this point.
2008
2009 // Set the stack pointer to the right value for reexecution:
2010 set_sp(reexecute_sp());
2011
2012 #ifdef ASSERT
2013 if (!must_throw) {
2014 // Make sure the stack has at least enough depth to execute
2015 // the current bytecode.
2016 int inputs, ignored_depth;
2017 if (compute_stack_effects(inputs, ignored_depth)) {
2018 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2019 Bytecodes::name(java_bc()), sp(), inputs);
2020 }
2021 }
2022 #endif
2023
2024 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2025 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2026
2027 switch (action) {
2028 case Deoptimization::Action_maybe_recompile:
2029 case Deoptimization::Action_reinterpret:
2030 // Temporary fix for 6529811 to allow virtual calls to be sure they
2031 // get the chance to go from mono->bi->mega
2032 if (!keep_exact_action &&
2033 Deoptimization::trap_request_index(trap_request) < 0 &&
2034 too_many_recompiles(reason)) {
2035 // This BCI is causing too many recompilations.
2036 if (C->log() != NULL) {
2037 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2038 Deoptimization::trap_reason_name(reason),
2039 Deoptimization::trap_action_name(action));
2040 }
2041 action = Deoptimization::Action_none;
2042 trap_request = Deoptimization::make_trap_request(reason, action);
2043 } else {
2044 C->set_trap_can_recompile(true);
2045 }
2046 break;
2047 case Deoptimization::Action_make_not_entrant:
2048 C->set_trap_can_recompile(true);
2049 break;
2050 case Deoptimization::Action_none:
2051 case Deoptimization::Action_make_not_compilable:
2052 break;
2053 default:
2054 #ifdef ASSERT
2055 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2056 #endif
2057 break;
2058 }
2059
2060 if (TraceOptoParse) {
2061 char buf[100];
2062 tty->print_cr("Uncommon trap %s at bci:%d",
2063 Deoptimization::format_trap_request(buf, sizeof(buf),
2064 trap_request), bci());
2065 }
2066
2067 CompileLog* log = C->log();
2068 if (log != NULL) {
2069 int kid = (klass == NULL)? -1: log->identify(klass);
2070 log->begin_elem("uncommon_trap bci='%d'", bci());
2071 char buf[100];
2072 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2073 trap_request));
2074 if (kid >= 0) log->print(" klass='%d'", kid);
2075 if (comment != NULL) log->print(" comment='%s'", comment);
2076 log->end_elem();
2077 }
2078
2079 // Make sure any guarding test views this path as very unlikely
2080 Node *i0 = control()->in(0);
2081 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2082 IfNode *iff = i0->as_If();
2083 float f = iff->_prob; // Get prob
2084 if (control()->Opcode() == Op_IfTrue) {
2085 if (f > PROB_UNLIKELY_MAG(4))
2086 iff->_prob = PROB_MIN;
2087 } else {
2088 if (f < PROB_LIKELY_MAG(4))
2089 iff->_prob = PROB_MAX;
2090 }
2091 }
2092
2093 // Clear out dead values from the debug info.
2094 kill_dead_locals();
2095
2096 // Now insert the uncommon trap subroutine call
2097 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2098 const TypePtr* no_memory_effects = NULL;
2099 // Pass the index of the class to be loaded
2100 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2101 (must_throw ? RC_MUST_THROW : 0),
2102 OptoRuntime::uncommon_trap_Type(),
2103 call_addr, "uncommon_trap", no_memory_effects,
2104 intcon(trap_request));
2105 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2106 "must extract request correctly from the graph");
2107 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2108
2109 call->set_req(TypeFunc::ReturnAdr, returnadr());
2110 // The debug info is the only real input to this call.
2111
2112 // Halt-and-catch fire here. The above call should never return!
2113 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen");
2114 _gvn.set_type_bottom(halt);
2115 root()->add_req(halt);
2116
2117 stop_and_kill_map();
2118 }
2119
2120
2121 //--------------------------just_allocated_object------------------------------
2122 // Report the object that was just allocated.
2123 // It must be the case that there are no intervening safepoints.
2124 // We use this to determine if an object is so "fresh" that
2125 // it does not require card marks.
just_allocated_object(Node * current_control)2126 Node* GraphKit::just_allocated_object(Node* current_control) {
2127 Node* ctrl = current_control;
2128 // Object::<init> is invoked after allocation, most of invoke nodes
2129 // will be reduced, but a region node is kept in parse time, we check
2130 // the pattern and skip the region node if it degraded to a copy.
2131 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 &&
2132 ctrl->as_Region()->is_copy()) {
2133 ctrl = ctrl->as_Region()->is_copy();
2134 }
2135 if (C->recent_alloc_ctl() == ctrl) {
2136 return C->recent_alloc_obj();
2137 }
2138 return NULL;
2139 }
2140
2141
round_double_arguments(ciMethod * dest_method)2142 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2143 // (Note: TypeFunc::make has a cache that makes this fast.)
2144 const TypeFunc* tf = TypeFunc::make(dest_method);
2145 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2146 for (int j = 0; j < nargs; j++) {
2147 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2148 if( targ->basic_type() == T_DOUBLE ) {
2149 // If any parameters are doubles, they must be rounded before
2150 // the call, dstore_rounding does gvn.transform
2151 Node *arg = argument(j);
2152 arg = dstore_rounding(arg);
2153 set_argument(j, arg);
2154 }
2155 }
2156 }
2157
2158 /**
2159 * Record profiling data exact_kls for Node n with the type system so
2160 * that it can propagate it (speculation)
2161 *
2162 * @param n node that the type applies to
2163 * @param exact_kls type from profiling
2164 * @param maybe_null did profiling see null?
2165 *
2166 * @return node with improved type
2167 */
record_profile_for_speculation(Node * n,ciKlass * exact_kls,ProfilePtrKind ptr_kind)2168 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2169 const Type* current_type = _gvn.type(n);
2170 assert(UseTypeSpeculation, "type speculation must be on");
2171
2172 const TypePtr* speculative = current_type->speculative();
2173
2174 // Should the klass from the profile be recorded in the speculative type?
2175 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2176 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2177 const TypeOopPtr* xtype = tklass->as_instance_type();
2178 assert(xtype->klass_is_exact(), "Should be exact");
2179 // Any reason to believe n is not null (from this profiling or a previous one)?
2180 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2181 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2182 // record the new speculative type's depth
2183 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2184 speculative = speculative->with_inline_depth(jvms()->depth());
2185 } else if (current_type->would_improve_ptr(ptr_kind)) {
2186 // Profiling report that null was never seen so we can change the
2187 // speculative type to non null ptr.
2188 if (ptr_kind == ProfileAlwaysNull) {
2189 speculative = TypePtr::NULL_PTR;
2190 } else {
2191 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2192 const TypePtr* ptr = TypePtr::NOTNULL;
2193 if (speculative != NULL) {
2194 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2195 } else {
2196 speculative = ptr;
2197 }
2198 }
2199 }
2200
2201 if (speculative != current_type->speculative()) {
2202 // Build a type with a speculative type (what we think we know
2203 // about the type but will need a guard when we use it)
2204 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2205 // We're changing the type, we need a new CheckCast node to carry
2206 // the new type. The new type depends on the control: what
2207 // profiling tells us is only valid from here as far as we can
2208 // tell.
2209 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2210 cast = _gvn.transform(cast);
2211 replace_in_map(n, cast);
2212 n = cast;
2213 }
2214
2215 return n;
2216 }
2217
2218 /**
2219 * Record profiling data from receiver profiling at an invoke with the
2220 * type system so that it can propagate it (speculation)
2221 *
2222 * @param n receiver node
2223 *
2224 * @return node with improved type
2225 */
record_profiled_receiver_for_speculation(Node * n)2226 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2227 if (!UseTypeSpeculation) {
2228 return n;
2229 }
2230 ciKlass* exact_kls = profile_has_unique_klass();
2231 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2232 if ((java_bc() == Bytecodes::_checkcast ||
2233 java_bc() == Bytecodes::_instanceof ||
2234 java_bc() == Bytecodes::_aastore) &&
2235 method()->method_data()->is_mature()) {
2236 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2237 if (data != NULL) {
2238 if (!data->as_BitData()->null_seen()) {
2239 ptr_kind = ProfileNeverNull;
2240 } else {
2241 assert(data->is_ReceiverTypeData(), "bad profile data type");
2242 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2243 uint i = 0;
2244 for (; i < call->row_limit(); i++) {
2245 ciKlass* receiver = call->receiver(i);
2246 if (receiver != NULL) {
2247 break;
2248 }
2249 }
2250 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2251 }
2252 }
2253 }
2254 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2255 }
2256
2257 /**
2258 * Record profiling data from argument profiling at an invoke with the
2259 * type system so that it can propagate it (speculation)
2260 *
2261 * @param dest_method target method for the call
2262 * @param bc what invoke bytecode is this?
2263 */
record_profiled_arguments_for_speculation(ciMethod * dest_method,Bytecodes::Code bc)2264 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2265 if (!UseTypeSpeculation) {
2266 return;
2267 }
2268 const TypeFunc* tf = TypeFunc::make(dest_method);
2269 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2270 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2271 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2272 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2273 if (is_reference_type(targ->basic_type())) {
2274 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2275 ciKlass* better_type = NULL;
2276 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2277 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2278 }
2279 i++;
2280 }
2281 }
2282 }
2283
2284 /**
2285 * Record profiling data from parameter profiling at an invoke with
2286 * the type system so that it can propagate it (speculation)
2287 */
record_profiled_parameters_for_speculation()2288 void GraphKit::record_profiled_parameters_for_speculation() {
2289 if (!UseTypeSpeculation) {
2290 return;
2291 }
2292 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2293 if (_gvn.type(local(i))->isa_oopptr()) {
2294 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2295 ciKlass* better_type = NULL;
2296 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2297 record_profile_for_speculation(local(i), better_type, ptr_kind);
2298 }
2299 j++;
2300 }
2301 }
2302 }
2303
2304 /**
2305 * Record profiling data from return value profiling at an invoke with
2306 * the type system so that it can propagate it (speculation)
2307 */
record_profiled_return_for_speculation()2308 void GraphKit::record_profiled_return_for_speculation() {
2309 if (!UseTypeSpeculation) {
2310 return;
2311 }
2312 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2313 ciKlass* better_type = NULL;
2314 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2315 // If profiling reports a single type for the return value,
2316 // feed it to the type system so it can propagate it as a
2317 // speculative type
2318 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2319 }
2320 }
2321
round_double_result(ciMethod * dest_method)2322 void GraphKit::round_double_result(ciMethod* dest_method) {
2323 // A non-strict method may return a double value which has an extended
2324 // exponent, but this must not be visible in a caller which is 'strict'
2325 // If a strict caller invokes a non-strict callee, round a double result
2326
2327 BasicType result_type = dest_method->return_type()->basic_type();
2328 assert( method() != NULL, "must have caller context");
2329 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2330 // Destination method's return value is on top of stack
2331 // dstore_rounding() does gvn.transform
2332 Node *result = pop_pair();
2333 result = dstore_rounding(result);
2334 push_pair(result);
2335 }
2336 }
2337
2338 // rounding for strict float precision conformance
precision_rounding(Node * n)2339 Node* GraphKit::precision_rounding(Node* n) {
2340 return UseStrictFP && _method->flags().is_strict()
2341 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2342 ? _gvn.transform( new RoundFloatNode(0, n) )
2343 : n;
2344 }
2345
2346 // rounding for strict double precision conformance
dprecision_rounding(Node * n)2347 Node* GraphKit::dprecision_rounding(Node *n) {
2348 return UseStrictFP && _method->flags().is_strict()
2349 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2350 ? _gvn.transform( new RoundDoubleNode(0, n) )
2351 : n;
2352 }
2353
2354 // rounding for non-strict double stores
dstore_rounding(Node * n)2355 Node* GraphKit::dstore_rounding(Node* n) {
2356 return Matcher::strict_fp_requires_explicit_rounding
2357 && UseSSE <= 1
2358 ? _gvn.transform( new RoundDoubleNode(0, n) )
2359 : n;
2360 }
2361
2362 //=============================================================================
2363 // Generate a fast path/slow path idiom. Graph looks like:
2364 // [foo] indicates that 'foo' is a parameter
2365 //
2366 // [in] NULL
2367 // \ /
2368 // CmpP
2369 // Bool ne
2370 // If
2371 // / \
2372 // True False-<2>
2373 // / |
2374 // / cast_not_null
2375 // Load | | ^
2376 // [fast_test] | |
2377 // gvn to opt_test | |
2378 // / \ | <1>
2379 // True False |
2380 // | \\ |
2381 // [slow_call] \[fast_result]
2382 // Ctl Val \ \
2383 // | \ \
2384 // Catch <1> \ \
2385 // / \ ^ \ \
2386 // Ex No_Ex | \ \
2387 // | \ \ | \ <2> \
2388 // ... \ [slow_res] | | \ [null_result]
2389 // \ \--+--+--- | |
2390 // \ | / \ | /
2391 // --------Region Phi
2392 //
2393 //=============================================================================
2394 // Code is structured as a series of driver functions all called 'do_XXX' that
2395 // call a set of helper functions. Helper functions first, then drivers.
2396
2397 //------------------------------null_check_oop---------------------------------
2398 // Null check oop. Set null-path control into Region in slot 3.
2399 // Make a cast-not-nullness use the other not-null control. Return cast.
null_check_oop(Node * value,Node ** null_control,bool never_see_null,bool safe_for_replace,bool speculative)2400 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2401 bool never_see_null,
2402 bool safe_for_replace,
2403 bool speculative) {
2404 // Initial NULL check taken path
2405 (*null_control) = top();
2406 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2407
2408 // Generate uncommon_trap:
2409 if (never_see_null && (*null_control) != top()) {
2410 // If we see an unexpected null at a check-cast we record it and force a
2411 // recompile; the offending check-cast will be compiled to handle NULLs.
2412 // If we see more than one offending BCI, then all checkcasts in the
2413 // method will be compiled to handle NULLs.
2414 PreserveJVMState pjvms(this);
2415 set_control(*null_control);
2416 replace_in_map(value, null());
2417 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2418 uncommon_trap(reason,
2419 Deoptimization::Action_make_not_entrant);
2420 (*null_control) = top(); // NULL path is dead
2421 }
2422 if ((*null_control) == top() && safe_for_replace) {
2423 replace_in_map(value, cast);
2424 }
2425
2426 // Cast away null-ness on the result
2427 return cast;
2428 }
2429
2430 //------------------------------opt_iff----------------------------------------
2431 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2432 // Return slow-path control.
opt_iff(Node * region,Node * iff)2433 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2434 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2435
2436 // Fast path taken; set region slot 2
2437 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2438 region->init_req(2,fast_taken); // Capture fast-control
2439
2440 // Fast path not-taken, i.e. slow path
2441 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2442 return slow_taken;
2443 }
2444
2445 //-----------------------------make_runtime_call-------------------------------
make_runtime_call(int flags,const TypeFunc * call_type,address call_addr,const char * call_name,const TypePtr * adr_type,Node * parm0,Node * parm1,Node * parm2,Node * parm3,Node * parm4,Node * parm5,Node * parm6,Node * parm7)2446 Node* GraphKit::make_runtime_call(int flags,
2447 const TypeFunc* call_type, address call_addr,
2448 const char* call_name,
2449 const TypePtr* adr_type,
2450 // The following parms are all optional.
2451 // The first NULL ends the list.
2452 Node* parm0, Node* parm1,
2453 Node* parm2, Node* parm3,
2454 Node* parm4, Node* parm5,
2455 Node* parm6, Node* parm7) {
2456 assert(call_addr != NULL, "must not call NULL targets");
2457
2458 // Slow-path call
2459 bool is_leaf = !(flags & RC_NO_LEAF);
2460 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2461 if (call_name == NULL) {
2462 assert(!is_leaf, "must supply name for leaf");
2463 call_name = OptoRuntime::stub_name(call_addr);
2464 }
2465 CallNode* call;
2466 if (!is_leaf) {
2467 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2468 bci(), adr_type);
2469 } else if (flags & RC_NO_FP) {
2470 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2471 } else {
2472 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2473 }
2474
2475 // The following is similar to set_edges_for_java_call,
2476 // except that the memory effects of the call are restricted to AliasIdxRaw.
2477
2478 // Slow path call has no side-effects, uses few values
2479 bool wide_in = !(flags & RC_NARROW_MEM);
2480 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2481
2482 Node* prev_mem = NULL;
2483 if (wide_in) {
2484 prev_mem = set_predefined_input_for_runtime_call(call);
2485 } else {
2486 assert(!wide_out, "narrow in => narrow out");
2487 Node* narrow_mem = memory(adr_type);
2488 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2489 }
2490
2491 // Hook each parm in order. Stop looking at the first NULL.
2492 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2493 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2494 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2495 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2496 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2497 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2498 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2499 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2500 /* close each nested if ===> */ } } } } } } } }
2501 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2502
2503 if (!is_leaf) {
2504 // Non-leaves can block and take safepoints:
2505 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2506 }
2507 // Non-leaves can throw exceptions:
2508 if (has_io) {
2509 call->set_req(TypeFunc::I_O, i_o());
2510 }
2511
2512 if (flags & RC_UNCOMMON) {
2513 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2514 // (An "if" probability corresponds roughly to an unconditional count.
2515 // Sort of.)
2516 call->set_cnt(PROB_UNLIKELY_MAG(4));
2517 }
2518
2519 Node* c = _gvn.transform(call);
2520 assert(c == call, "cannot disappear");
2521
2522 if (wide_out) {
2523 // Slow path call has full side-effects.
2524 set_predefined_output_for_runtime_call(call);
2525 } else {
2526 // Slow path call has few side-effects, and/or sets few values.
2527 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2528 }
2529
2530 if (has_io) {
2531 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2532 }
2533 return call;
2534
2535 }
2536
2537 //------------------------------merge_memory-----------------------------------
2538 // Merge memory from one path into the current memory state.
merge_memory(Node * new_mem,Node * region,int new_path)2539 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2540 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2541 Node* old_slice = mms.force_memory();
2542 Node* new_slice = mms.memory2();
2543 if (old_slice != new_slice) {
2544 PhiNode* phi;
2545 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2546 if (mms.is_empty()) {
2547 // clone base memory Phi's inputs for this memory slice
2548 assert(old_slice == mms.base_memory(), "sanity");
2549 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2550 _gvn.set_type(phi, Type::MEMORY);
2551 for (uint i = 1; i < phi->req(); i++) {
2552 phi->init_req(i, old_slice->in(i));
2553 }
2554 } else {
2555 phi = old_slice->as_Phi(); // Phi was generated already
2556 }
2557 } else {
2558 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2559 _gvn.set_type(phi, Type::MEMORY);
2560 }
2561 phi->set_req(new_path, new_slice);
2562 mms.set_memory(phi);
2563 }
2564 }
2565 }
2566
2567 //------------------------------make_slow_call_ex------------------------------
2568 // Make the exception handler hookups for the slow call
make_slow_call_ex(Node * call,ciInstanceKlass * ex_klass,bool separate_io_proj,bool deoptimize)2569 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2570 if (stopped()) return;
2571
2572 // Make a catch node with just two handlers: fall-through and catch-all
2573 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2574 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2575 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2576 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2577
2578 { PreserveJVMState pjvms(this);
2579 set_control(excp);
2580 set_i_o(i_o);
2581
2582 if (excp != top()) {
2583 if (deoptimize) {
2584 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2585 uncommon_trap(Deoptimization::Reason_unhandled,
2586 Deoptimization::Action_none);
2587 } else {
2588 // Create an exception state also.
2589 // Use an exact type if the caller has a specific exception.
2590 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2591 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2592 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2593 }
2594 }
2595 }
2596
2597 // Get the no-exception control from the CatchNode.
2598 set_control(norm);
2599 }
2600
gen_subtype_check_compare(Node * ctrl,Node * in1,Node * in2,BoolTest::mask test,float p,PhaseGVN * gvn,BasicType bt)2601 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2602 Node* cmp = NULL;
2603 switch(bt) {
2604 case T_INT: cmp = new CmpINode(in1, in2); break;
2605 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2606 default: fatal("unexpected comparison type %s", type2name(bt));
2607 }
2608 gvn->transform(cmp);
2609 Node* bol = gvn->transform(new BoolNode(cmp, test));
2610 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2611 gvn->transform(iff);
2612 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2613 return iff;
2614 }
2615
2616
2617 //-------------------------------gen_subtype_check-----------------------------
2618 // Generate a subtyping check. Takes as input the subtype and supertype.
2619 // Returns 2 values: sets the default control() to the true path and returns
2620 // the false path. Only reads invariant memory; sets no (visible) memory.
2621 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2622 // but that's not exposed to the optimizer. This call also doesn't take in an
2623 // Object; if you wish to check an Object you need to load the Object's class
2624 // prior to coming here.
gen_subtype_check(Node * subklass,Node * superklass,Node ** ctrl,MergeMemNode * mem,PhaseGVN * gvn)2625 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2626 Compile* C = gvn->C;
2627
2628 if ((*ctrl)->is_top()) {
2629 return C->top();
2630 }
2631
2632 // Fast check for identical types, perhaps identical constants.
2633 // The types can even be identical non-constants, in cases
2634 // involving Array.newInstance, Object.clone, etc.
2635 if (subklass == superklass)
2636 return C->top(); // false path is dead; no test needed.
2637
2638 if (gvn->type(superklass)->singleton()) {
2639 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2640 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2641
2642 // In the common case of an exact superklass, try to fold up the
2643 // test before generating code. You may ask, why not just generate
2644 // the code and then let it fold up? The answer is that the generated
2645 // code will necessarily include null checks, which do not always
2646 // completely fold away. If they are also needless, then they turn
2647 // into a performance loss. Example:
2648 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2649 // Here, the type of 'fa' is often exact, so the store check
2650 // of fa[1]=x will fold up, without testing the nullness of x.
2651 switch (C->static_subtype_check(superk, subk)) {
2652 case Compile::SSC_always_false:
2653 {
2654 Node* always_fail = *ctrl;
2655 *ctrl = gvn->C->top();
2656 return always_fail;
2657 }
2658 case Compile::SSC_always_true:
2659 return C->top();
2660 case Compile::SSC_easy_test:
2661 {
2662 // Just do a direct pointer compare and be done.
2663 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2664 *ctrl = gvn->transform(new IfTrueNode(iff));
2665 return gvn->transform(new IfFalseNode(iff));
2666 }
2667 case Compile::SSC_full_test:
2668 break;
2669 default:
2670 ShouldNotReachHere();
2671 }
2672 }
2673
2674 // %%% Possible further optimization: Even if the superklass is not exact,
2675 // if the subklass is the unique subtype of the superklass, the check
2676 // will always succeed. We could leave a dependency behind to ensure this.
2677
2678 // First load the super-klass's check-offset
2679 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2680 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2681 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2682 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2683 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2684
2685 // Load from the sub-klass's super-class display list, or a 1-word cache of
2686 // the secondary superclass list, or a failing value with a sentinel offset
2687 // if the super-klass is an interface or exceptionally deep in the Java
2688 // hierarchy and we have to scan the secondary superclass list the hard way.
2689 // Worst-case type is a little odd: NULL is allowed as a result (usually
2690 // klass loads can never produce a NULL).
2691 Node *chk_off_X = chk_off;
2692 #ifdef _LP64
2693 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2694 #endif
2695 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2696 // For some types like interfaces the following loadKlass is from a 1-word
2697 // cache which is mutable so can't use immutable memory. Other
2698 // types load from the super-class display table which is immutable.
2699 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2700 Node *kmem = might_be_cache ? m : C->immutable_memory();
2701 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2702
2703 // Compile speed common case: ARE a subtype and we canNOT fail
2704 if( superklass == nkls )
2705 return C->top(); // false path is dead; no test needed.
2706
2707 // See if we get an immediate positive hit. Happens roughly 83% of the
2708 // time. Test to see if the value loaded just previously from the subklass
2709 // is exactly the superklass.
2710 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2711 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2712 *ctrl = gvn->transform(new IfFalseNode(iff1));
2713
2714 // Compile speed common case: Check for being deterministic right now. If
2715 // chk_off is a constant and not equal to cacheoff then we are NOT a
2716 // subklass. In this case we need exactly the 1 test above and we can
2717 // return those results immediately.
2718 if (!might_be_cache) {
2719 Node* not_subtype_ctrl = *ctrl;
2720 *ctrl = iftrue1; // We need exactly the 1 test above
2721 return not_subtype_ctrl;
2722 }
2723
2724 // Gather the various success & failures here
2725 RegionNode *r_ok_subtype = new RegionNode(4);
2726 gvn->record_for_igvn(r_ok_subtype);
2727 RegionNode *r_not_subtype = new RegionNode(3);
2728 gvn->record_for_igvn(r_not_subtype);
2729
2730 r_ok_subtype->init_req(1, iftrue1);
2731
2732 // Check for immediate negative hit. Happens roughly 11% of the time (which
2733 // is roughly 63% of the remaining cases). Test to see if the loaded
2734 // check-offset points into the subklass display list or the 1-element
2735 // cache. If it points to the display (and NOT the cache) and the display
2736 // missed then it's not a subtype.
2737 Node *cacheoff = gvn->intcon(cacheoff_con);
2738 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2739 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2740 *ctrl = gvn->transform(new IfFalseNode(iff2));
2741
2742 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2743 // No performance impact (too rare) but allows sharing of secondary arrays
2744 // which has some footprint reduction.
2745 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2746 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2747 *ctrl = gvn->transform(new IfFalseNode(iff3));
2748
2749 // -- Roads not taken here: --
2750 // We could also have chosen to perform the self-check at the beginning
2751 // of this code sequence, as the assembler does. This would not pay off
2752 // the same way, since the optimizer, unlike the assembler, can perform
2753 // static type analysis to fold away many successful self-checks.
2754 // Non-foldable self checks work better here in second position, because
2755 // the initial primary superclass check subsumes a self-check for most
2756 // types. An exception would be a secondary type like array-of-interface,
2757 // which does not appear in its own primary supertype display.
2758 // Finally, we could have chosen to move the self-check into the
2759 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2760 // dependent manner. But it is worthwhile to have the check here,
2761 // where it can be perhaps be optimized. The cost in code space is
2762 // small (register compare, branch).
2763
2764 // Now do a linear scan of the secondary super-klass array. Again, no real
2765 // performance impact (too rare) but it's gotta be done.
2766 // Since the code is rarely used, there is no penalty for moving it
2767 // out of line, and it can only improve I-cache density.
2768 // The decision to inline or out-of-line this final check is platform
2769 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2770 Node* psc = gvn->transform(
2771 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2772
2773 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2774 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2775 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2776
2777 // Return false path; set default control to true path.
2778 *ctrl = gvn->transform(r_ok_subtype);
2779 return gvn->transform(r_not_subtype);
2780 }
2781
2782 // Profile-driven exact type check:
type_check_receiver(Node * receiver,ciKlass * klass,float prob,Node ** casted_receiver)2783 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2784 float prob,
2785 Node* *casted_receiver) {
2786 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2787 Node* recv_klass = load_object_klass(receiver);
2788 Node* want_klass = makecon(tklass);
2789 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2790 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2791 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2792 set_control( _gvn.transform( new IfTrueNode (iff) ));
2793 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2794
2795 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2796 assert(recv_xtype->klass_is_exact(), "");
2797
2798 // Subsume downstream occurrences of receiver with a cast to
2799 // recv_xtype, since now we know what the type will be.
2800 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2801 (*casted_receiver) = _gvn.transform(cast);
2802 // (User must make the replace_in_map call.)
2803
2804 return fail;
2805 }
2806
2807 //------------------------------subtype_check_receiver-------------------------
subtype_check_receiver(Node * receiver,ciKlass * klass,Node ** casted_receiver)2808 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2809 Node** casted_receiver) {
2810 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2811 Node* recv_klass = load_object_klass(receiver);
2812 Node* want_klass = makecon(tklass);
2813
2814 Node* slow_ctl = gen_subtype_check(recv_klass, want_klass);
2815
2816 // Cast receiver after successful check
2817 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2818 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2819 (*casted_receiver) = _gvn.transform(cast);
2820
2821 return slow_ctl;
2822 }
2823
2824 //------------------------------seems_never_null-------------------------------
2825 // Use null_seen information if it is available from the profile.
2826 // If we see an unexpected null at a type check we record it and force a
2827 // recompile; the offending check will be recompiled to handle NULLs.
2828 // If we see several offending BCIs, then all checks in the
2829 // method will be recompiled.
seems_never_null(Node * obj,ciProfileData * data,bool & speculating)2830 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2831 speculating = !_gvn.type(obj)->speculative_maybe_null();
2832 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2833 if (UncommonNullCast // Cutout for this technique
2834 && obj != null() // And not the -Xcomp stupid case?
2835 && !too_many_traps(reason)
2836 ) {
2837 if (speculating) {
2838 return true;
2839 }
2840 if (data == NULL)
2841 // Edge case: no mature data. Be optimistic here.
2842 return true;
2843 // If the profile has not seen a null, assume it won't happen.
2844 assert(java_bc() == Bytecodes::_checkcast ||
2845 java_bc() == Bytecodes::_instanceof ||
2846 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2847 return !data->as_BitData()->null_seen();
2848 }
2849 speculating = false;
2850 return false;
2851 }
2852
guard_klass_being_initialized(Node * klass)2853 void GraphKit::guard_klass_being_initialized(Node* klass) {
2854 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
2855 Node* adr = basic_plus_adr(top(), klass, init_state_off);
2856 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
2857 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
2858 T_BYTE, MemNode::unordered);
2859 init_state = _gvn.transform(init_state);
2860
2861 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
2862
2863 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
2864 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
2865
2866 { BuildCutout unless(this, tst, PROB_MAX);
2867 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
2868 }
2869 }
2870
guard_init_thread(Node * klass)2871 void GraphKit::guard_init_thread(Node* klass) {
2872 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
2873 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
2874
2875 Node* init_thread = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
2876 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
2877 T_ADDRESS, MemNode::unordered);
2878 init_thread = _gvn.transform(init_thread);
2879
2880 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
2881
2882 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
2883 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
2884
2885 { BuildCutout unless(this, tst, PROB_MAX);
2886 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
2887 }
2888 }
2889
clinit_barrier(ciInstanceKlass * ik,ciMethod * context)2890 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
2891 if (ik->is_being_initialized()) {
2892 if (C->needs_clinit_barrier(ik, context)) {
2893 Node* klass = makecon(TypeKlassPtr::make(ik));
2894 guard_klass_being_initialized(klass);
2895 guard_init_thread(klass);
2896 insert_mem_bar(Op_MemBarCPUOrder);
2897 }
2898 } else if (ik->is_initialized()) {
2899 return; // no barrier needed
2900 } else {
2901 uncommon_trap(Deoptimization::Reason_uninitialized,
2902 Deoptimization::Action_reinterpret,
2903 NULL);
2904 }
2905 }
2906
2907 //------------------------maybe_cast_profiled_receiver-------------------------
2908 // If the profile has seen exactly one type, narrow to exactly that type.
2909 // Subsequent type checks will always fold up.
maybe_cast_profiled_receiver(Node * not_null_obj,ciKlass * require_klass,ciKlass * spec_klass,bool safe_for_replace)2910 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2911 ciKlass* require_klass,
2912 ciKlass* spec_klass,
2913 bool safe_for_replace) {
2914 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2915
2916 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2917
2918 // Make sure we haven't already deoptimized from this tactic.
2919 if (too_many_traps_or_recompiles(reason))
2920 return NULL;
2921
2922 // (No, this isn't a call, but it's enough like a virtual call
2923 // to use the same ciMethod accessor to get the profile info...)
2924 // If we have a speculative type use it instead of profiling (which
2925 // may not help us)
2926 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2927 if (exact_kls != NULL) {// no cast failures here
2928 if (require_klass == NULL ||
2929 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2930 // If we narrow the type to match what the type profile sees or
2931 // the speculative type, we can then remove the rest of the
2932 // cast.
2933 // This is a win, even if the exact_kls is very specific,
2934 // because downstream operations, such as method calls,
2935 // will often benefit from the sharper type.
2936 Node* exact_obj = not_null_obj; // will get updated in place...
2937 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2938 &exact_obj);
2939 { PreserveJVMState pjvms(this);
2940 set_control(slow_ctl);
2941 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2942 }
2943 if (safe_for_replace) {
2944 replace_in_map(not_null_obj, exact_obj);
2945 }
2946 return exact_obj;
2947 }
2948 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2949 }
2950
2951 return NULL;
2952 }
2953
2954 /**
2955 * Cast obj to type and emit guard unless we had too many traps here
2956 * already
2957 *
2958 * @param obj node being casted
2959 * @param type type to cast the node to
2960 * @param not_null true if we know node cannot be null
2961 */
maybe_cast_profiled_obj(Node * obj,ciKlass * type,bool not_null)2962 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2963 ciKlass* type,
2964 bool not_null) {
2965 if (stopped()) {
2966 return obj;
2967 }
2968
2969 // type == NULL if profiling tells us this object is always null
2970 if (type != NULL) {
2971 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2972 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2973
2974 if (!too_many_traps_or_recompiles(null_reason) &&
2975 !too_many_traps_or_recompiles(class_reason)) {
2976 Node* not_null_obj = NULL;
2977 // not_null is true if we know the object is not null and
2978 // there's no need for a null check
2979 if (!not_null) {
2980 Node* null_ctl = top();
2981 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2982 assert(null_ctl->is_top(), "no null control here");
2983 } else {
2984 not_null_obj = obj;
2985 }
2986
2987 Node* exact_obj = not_null_obj;
2988 ciKlass* exact_kls = type;
2989 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2990 &exact_obj);
2991 {
2992 PreserveJVMState pjvms(this);
2993 set_control(slow_ctl);
2994 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2995 }
2996 replace_in_map(not_null_obj, exact_obj);
2997 obj = exact_obj;
2998 }
2999 } else {
3000 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3001 Node* exact_obj = null_assert(obj);
3002 replace_in_map(obj, exact_obj);
3003 obj = exact_obj;
3004 }
3005 }
3006 return obj;
3007 }
3008
3009 //-------------------------------gen_instanceof--------------------------------
3010 // Generate an instance-of idiom. Used by both the instance-of bytecode
3011 // and the reflective instance-of call.
gen_instanceof(Node * obj,Node * superklass,bool safe_for_replace)3012 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3013 kill_dead_locals(); // Benefit all the uncommon traps
3014 assert( !stopped(), "dead parse path should be checked in callers" );
3015 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3016 "must check for not-null not-dead klass in callers");
3017
3018 // Make the merge point
3019 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3020 RegionNode* region = new RegionNode(PATH_LIMIT);
3021 Node* phi = new PhiNode(region, TypeInt::BOOL);
3022 C->set_has_split_ifs(true); // Has chance for split-if optimization
3023
3024 ciProfileData* data = NULL;
3025 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3026 data = method()->method_data()->bci_to_data(bci());
3027 }
3028 bool speculative_not_null = false;
3029 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3030 && seems_never_null(obj, data, speculative_not_null));
3031
3032 // Null check; get casted pointer; set region slot 3
3033 Node* null_ctl = top();
3034 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3035
3036 // If not_null_obj is dead, only null-path is taken
3037 if (stopped()) { // Doing instance-of on a NULL?
3038 set_control(null_ctl);
3039 return intcon(0);
3040 }
3041 region->init_req(_null_path, null_ctl);
3042 phi ->init_req(_null_path, intcon(0)); // Set null path value
3043 if (null_ctl == top()) {
3044 // Do this eagerly, so that pattern matches like is_diamond_phi
3045 // will work even during parsing.
3046 assert(_null_path == PATH_LIMIT-1, "delete last");
3047 region->del_req(_null_path);
3048 phi ->del_req(_null_path);
3049 }
3050
3051 // Do we know the type check always succeed?
3052 bool known_statically = false;
3053 if (_gvn.type(superklass)->singleton()) {
3054 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3055 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3056 if (subk != NULL && subk->is_loaded()) {
3057 int static_res = C->static_subtype_check(superk, subk);
3058 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3059 }
3060 }
3061
3062 if (!known_statically) {
3063 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3064 // We may not have profiling here or it may not help us. If we
3065 // have a speculative type use it to perform an exact cast.
3066 ciKlass* spec_obj_type = obj_type->speculative_type();
3067 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3068 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3069 if (stopped()) { // Profile disagrees with this path.
3070 set_control(null_ctl); // Null is the only remaining possibility.
3071 return intcon(0);
3072 }
3073 if (cast_obj != NULL) {
3074 not_null_obj = cast_obj;
3075 }
3076 }
3077 }
3078
3079 // Load the object's klass
3080 Node* obj_klass = load_object_klass(not_null_obj);
3081
3082 // Generate the subtype check
3083 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3084
3085 // Plug in the success path to the general merge in slot 1.
3086 region->init_req(_obj_path, control());
3087 phi ->init_req(_obj_path, intcon(1));
3088
3089 // Plug in the failing path to the general merge in slot 2.
3090 region->init_req(_fail_path, not_subtype_ctrl);
3091 phi ->init_req(_fail_path, intcon(0));
3092
3093 // Return final merged results
3094 set_control( _gvn.transform(region) );
3095 record_for_igvn(region);
3096
3097 // If we know the type check always succeeds then we don't use the
3098 // profiling data at this bytecode. Don't lose it, feed it to the
3099 // type system as a speculative type.
3100 if (safe_for_replace) {
3101 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3102 replace_in_map(obj, casted_obj);
3103 }
3104
3105 return _gvn.transform(phi);
3106 }
3107
3108 //-------------------------------gen_checkcast---------------------------------
3109 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3110 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3111 // uncommon-trap paths work. Adjust stack after this call.
3112 // If failure_control is supplied and not null, it is filled in with
3113 // the control edge for the cast failure. Otherwise, an appropriate
3114 // uncommon trap or exception is thrown.
gen_checkcast(Node * obj,Node * superklass,Node ** failure_control)3115 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3116 Node* *failure_control) {
3117 kill_dead_locals(); // Benefit all the uncommon traps
3118 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3119 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3120
3121 // Fast cutout: Check the case that the cast is vacuously true.
3122 // This detects the common cases where the test will short-circuit
3123 // away completely. We do this before we perform the null check,
3124 // because if the test is going to turn into zero code, we don't
3125 // want a residual null check left around. (Causes a slowdown,
3126 // for example, in some objArray manipulations, such as a[i]=a[j].)
3127 if (tk->singleton()) {
3128 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3129 if (objtp != NULL && objtp->klass() != NULL) {
3130 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3131 case Compile::SSC_always_true:
3132 // If we know the type check always succeed then we don't use
3133 // the profiling data at this bytecode. Don't lose it, feed it
3134 // to the type system as a speculative type.
3135 return record_profiled_receiver_for_speculation(obj);
3136 case Compile::SSC_always_false:
3137 // It needs a null check because a null will *pass* the cast check.
3138 // A non-null value will always produce an exception.
3139 return null_assert(obj);
3140 }
3141 }
3142 }
3143
3144 ciProfileData* data = NULL;
3145 bool safe_for_replace = false;
3146 if (failure_control == NULL) { // use MDO in regular case only
3147 assert(java_bc() == Bytecodes::_aastore ||
3148 java_bc() == Bytecodes::_checkcast,
3149 "interpreter profiles type checks only for these BCs");
3150 data = method()->method_data()->bci_to_data(bci());
3151 safe_for_replace = true;
3152 }
3153
3154 // Make the merge point
3155 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3156 RegionNode* region = new RegionNode(PATH_LIMIT);
3157 Node* phi = new PhiNode(region, toop);
3158 C->set_has_split_ifs(true); // Has chance for split-if optimization
3159
3160 // Use null-cast information if it is available
3161 bool speculative_not_null = false;
3162 bool never_see_null = ((failure_control == NULL) // regular case only
3163 && seems_never_null(obj, data, speculative_not_null));
3164
3165 // Null check; get casted pointer; set region slot 3
3166 Node* null_ctl = top();
3167 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3168
3169 // If not_null_obj is dead, only null-path is taken
3170 if (stopped()) { // Doing instance-of on a NULL?
3171 set_control(null_ctl);
3172 return null();
3173 }
3174 region->init_req(_null_path, null_ctl);
3175 phi ->init_req(_null_path, null()); // Set null path value
3176 if (null_ctl == top()) {
3177 // Do this eagerly, so that pattern matches like is_diamond_phi
3178 // will work even during parsing.
3179 assert(_null_path == PATH_LIMIT-1, "delete last");
3180 region->del_req(_null_path);
3181 phi ->del_req(_null_path);
3182 }
3183
3184 Node* cast_obj = NULL;
3185 if (tk->klass_is_exact()) {
3186 // The following optimization tries to statically cast the speculative type of the object
3187 // (for example obtained during profiling) to the type of the superklass and then do a
3188 // dynamic check that the type of the object is what we expect. To work correctly
3189 // for checkcast and aastore the type of superklass should be exact.
3190 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3191 // We may not have profiling here or it may not help us. If we have
3192 // a speculative type use it to perform an exact cast.
3193 ciKlass* spec_obj_type = obj_type->speculative_type();
3194 if (spec_obj_type != NULL || data != NULL) {
3195 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3196 if (cast_obj != NULL) {
3197 if (failure_control != NULL) // failure is now impossible
3198 (*failure_control) = top();
3199 // adjust the type of the phi to the exact klass:
3200 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3201 }
3202 }
3203 }
3204
3205 if (cast_obj == NULL) {
3206 // Load the object's klass
3207 Node* obj_klass = load_object_klass(not_null_obj);
3208
3209 // Generate the subtype check
3210 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3211
3212 // Plug in success path into the merge
3213 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3214 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3215 if (failure_control == NULL) {
3216 if (not_subtype_ctrl != top()) { // If failure is possible
3217 PreserveJVMState pjvms(this);
3218 set_control(not_subtype_ctrl);
3219 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3220 }
3221 } else {
3222 (*failure_control) = not_subtype_ctrl;
3223 }
3224 }
3225
3226 region->init_req(_obj_path, control());
3227 phi ->init_req(_obj_path, cast_obj);
3228
3229 // A merge of NULL or Casted-NotNull obj
3230 Node* res = _gvn.transform(phi);
3231
3232 // Note I do NOT always 'replace_in_map(obj,result)' here.
3233 // if( tk->klass()->can_be_primary_super() )
3234 // This means that if I successfully store an Object into an array-of-String
3235 // I 'forget' that the Object is really now known to be a String. I have to
3236 // do this because we don't have true union types for interfaces - if I store
3237 // a Baz into an array-of-Interface and then tell the optimizer it's an
3238 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3239 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3240 // replace_in_map( obj, res );
3241
3242 // Return final merged results
3243 set_control( _gvn.transform(region) );
3244 record_for_igvn(region);
3245
3246 return record_profiled_receiver_for_speculation(res);
3247 }
3248
3249 //------------------------------next_monitor-----------------------------------
3250 // What number should be given to the next monitor?
next_monitor()3251 int GraphKit::next_monitor() {
3252 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3253 int next = current + C->sync_stack_slots();
3254 // Keep the toplevel high water mark current:
3255 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3256 return current;
3257 }
3258
3259 //------------------------------insert_mem_bar---------------------------------
3260 // Memory barrier to avoid floating things around
3261 // The membar serves as a pinch point between both control and all memory slices.
insert_mem_bar(int opcode,Node * precedent)3262 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3263 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3264 mb->init_req(TypeFunc::Control, control());
3265 mb->init_req(TypeFunc::Memory, reset_memory());
3266 Node* membar = _gvn.transform(mb);
3267 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3268 set_all_memory_call(membar);
3269 return membar;
3270 }
3271
3272 //-------------------------insert_mem_bar_volatile----------------------------
3273 // Memory barrier to avoid floating things around
3274 // The membar serves as a pinch point between both control and memory(alias_idx).
3275 // If you want to make a pinch point on all memory slices, do not use this
3276 // function (even with AliasIdxBot); use insert_mem_bar() instead.
insert_mem_bar_volatile(int opcode,int alias_idx,Node * precedent)3277 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3278 // When Parse::do_put_xxx updates a volatile field, it appends a series
3279 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3280 // The first membar is on the same memory slice as the field store opcode.
3281 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3282 // All the other membars (for other volatile slices, including AliasIdxBot,
3283 // which stands for all unknown volatile slices) are control-dependent
3284 // on the first membar. This prevents later volatile loads or stores
3285 // from sliding up past the just-emitted store.
3286
3287 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3288 mb->set_req(TypeFunc::Control,control());
3289 if (alias_idx == Compile::AliasIdxBot) {
3290 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3291 } else {
3292 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3293 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3294 }
3295 Node* membar = _gvn.transform(mb);
3296 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3297 if (alias_idx == Compile::AliasIdxBot) {
3298 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3299 } else {
3300 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3301 }
3302 return membar;
3303 }
3304
3305 //------------------------------shared_lock------------------------------------
3306 // Emit locking code.
shared_lock(Node * obj)3307 FastLockNode* GraphKit::shared_lock(Node* obj) {
3308 // bci is either a monitorenter bc or InvocationEntryBci
3309 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3310 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3311
3312 if( !GenerateSynchronizationCode )
3313 return NULL; // Not locking things?
3314 if (stopped()) // Dead monitor?
3315 return NULL;
3316
3317 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3318
3319 // Box the stack location
3320 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3321 Node* mem = reset_memory();
3322
3323 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3324 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3325 // Create the counters for this fast lock.
3326 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3327 }
3328
3329 // Create the rtm counters for this fast lock if needed.
3330 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3331
3332 // Add monitor to debug info for the slow path. If we block inside the
3333 // slow path and de-opt, we need the monitor hanging around
3334 map()->push_monitor( flock );
3335
3336 const TypeFunc *tf = LockNode::lock_type();
3337 LockNode *lock = new LockNode(C, tf);
3338
3339 lock->init_req( TypeFunc::Control, control() );
3340 lock->init_req( TypeFunc::Memory , mem );
3341 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3342 lock->init_req( TypeFunc::FramePtr, frameptr() );
3343 lock->init_req( TypeFunc::ReturnAdr, top() );
3344
3345 lock->init_req(TypeFunc::Parms + 0, obj);
3346 lock->init_req(TypeFunc::Parms + 1, box);
3347 lock->init_req(TypeFunc::Parms + 2, flock);
3348 add_safepoint_edges(lock);
3349
3350 lock = _gvn.transform( lock )->as_Lock();
3351
3352 // lock has no side-effects, sets few values
3353 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3354
3355 insert_mem_bar(Op_MemBarAcquireLock);
3356
3357 // Add this to the worklist so that the lock can be eliminated
3358 record_for_igvn(lock);
3359
3360 #ifndef PRODUCT
3361 if (PrintLockStatistics) {
3362 // Update the counter for this lock. Don't bother using an atomic
3363 // operation since we don't require absolute accuracy.
3364 lock->create_lock_counter(map()->jvms());
3365 increment_counter(lock->counter()->addr());
3366 }
3367 #endif
3368
3369 return flock;
3370 }
3371
3372
3373 //------------------------------shared_unlock----------------------------------
3374 // Emit unlocking code.
shared_unlock(Node * box,Node * obj)3375 void GraphKit::shared_unlock(Node* box, Node* obj) {
3376 // bci is either a monitorenter bc or InvocationEntryBci
3377 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3378 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3379
3380 if( !GenerateSynchronizationCode )
3381 return;
3382 if (stopped()) { // Dead monitor?
3383 map()->pop_monitor(); // Kill monitor from debug info
3384 return;
3385 }
3386
3387 // Memory barrier to avoid floating things down past the locked region
3388 insert_mem_bar(Op_MemBarReleaseLock);
3389
3390 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3391 UnlockNode *unlock = new UnlockNode(C, tf);
3392 #ifdef ASSERT
3393 unlock->set_dbg_jvms(sync_jvms());
3394 #endif
3395 uint raw_idx = Compile::AliasIdxRaw;
3396 unlock->init_req( TypeFunc::Control, control() );
3397 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3398 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3399 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3400 unlock->init_req( TypeFunc::ReturnAdr, top() );
3401
3402 unlock->init_req(TypeFunc::Parms + 0, obj);
3403 unlock->init_req(TypeFunc::Parms + 1, box);
3404 unlock = _gvn.transform(unlock)->as_Unlock();
3405
3406 Node* mem = reset_memory();
3407
3408 // unlock has no side-effects, sets few values
3409 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3410
3411 // Kill monitor from debug info
3412 map()->pop_monitor( );
3413 }
3414
3415 //-------------------------------get_layout_helper-----------------------------
3416 // If the given klass is a constant or known to be an array,
3417 // fetch the constant layout helper value into constant_value
3418 // and return (Node*)NULL. Otherwise, load the non-constant
3419 // layout helper value, and return the node which represents it.
3420 // This two-faced routine is useful because allocation sites
3421 // almost always feature constant types.
get_layout_helper(Node * klass_node,jint & constant_value)3422 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3423 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3424 if (!StressReflectiveCode && inst_klass != NULL) {
3425 ciKlass* klass = inst_klass->klass();
3426 bool xklass = inst_klass->klass_is_exact();
3427 if (xklass || klass->is_array_klass()) {
3428 jint lhelper = klass->layout_helper();
3429 if (lhelper != Klass::_lh_neutral_value) {
3430 constant_value = lhelper;
3431 return (Node*) NULL;
3432 }
3433 }
3434 }
3435 constant_value = Klass::_lh_neutral_value; // put in a known value
3436 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3437 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3438 }
3439
3440 // We just put in an allocate/initialize with a big raw-memory effect.
3441 // Hook selected additional alias categories on the initialization.
hook_memory_on_init(GraphKit & kit,int alias_idx,MergeMemNode * init_in_merge,Node * init_out_raw)3442 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3443 MergeMemNode* init_in_merge,
3444 Node* init_out_raw) {
3445 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3446 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3447
3448 Node* prevmem = kit.memory(alias_idx);
3449 init_in_merge->set_memory_at(alias_idx, prevmem);
3450 kit.set_memory(init_out_raw, alias_idx);
3451 }
3452
3453 //---------------------------set_output_for_allocation-------------------------
set_output_for_allocation(AllocateNode * alloc,const TypeOopPtr * oop_type,bool deoptimize_on_exception)3454 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3455 const TypeOopPtr* oop_type,
3456 bool deoptimize_on_exception) {
3457 int rawidx = Compile::AliasIdxRaw;
3458 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3459 add_safepoint_edges(alloc);
3460 Node* allocx = _gvn.transform(alloc);
3461 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3462 // create memory projection for i_o
3463 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3464 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3465
3466 // create a memory projection as for the normal control path
3467 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3468 set_memory(malloc, rawidx);
3469
3470 // a normal slow-call doesn't change i_o, but an allocation does
3471 // we create a separate i_o projection for the normal control path
3472 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3473 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3474
3475 // put in an initialization barrier
3476 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3477 rawoop)->as_Initialize();
3478 assert(alloc->initialization() == init, "2-way macro link must work");
3479 assert(init ->allocation() == alloc, "2-way macro link must work");
3480 {
3481 // Extract memory strands which may participate in the new object's
3482 // initialization, and source them from the new InitializeNode.
3483 // This will allow us to observe initializations when they occur,
3484 // and link them properly (as a group) to the InitializeNode.
3485 assert(init->in(InitializeNode::Memory) == malloc, "");
3486 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3487 init->set_req(InitializeNode::Memory, minit_in);
3488 record_for_igvn(minit_in); // fold it up later, if possible
3489 Node* minit_out = memory(rawidx);
3490 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3491 // Add an edge in the MergeMem for the header fields so an access
3492 // to one of those has correct memory state
3493 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3494 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3495 if (oop_type->isa_aryptr()) {
3496 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3497 int elemidx = C->get_alias_index(telemref);
3498 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3499 } else if (oop_type->isa_instptr()) {
3500 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3501 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3502 ciField* field = ik->nonstatic_field_at(i);
3503 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3504 continue; // do not bother to track really large numbers of fields
3505 // Find (or create) the alias category for this field:
3506 int fieldidx = C->alias_type(field)->index();
3507 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3508 }
3509 }
3510 }
3511
3512 // Cast raw oop to the real thing...
3513 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3514 javaoop = _gvn.transform(javaoop);
3515 C->set_recent_alloc(control(), javaoop);
3516 assert(just_allocated_object(control()) == javaoop, "just allocated");
3517
3518 #ifdef ASSERT
3519 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3520 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3521 "Ideal_allocation works");
3522 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3523 "Ideal_allocation works");
3524 if (alloc->is_AllocateArray()) {
3525 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3526 "Ideal_allocation works");
3527 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3528 "Ideal_allocation works");
3529 } else {
3530 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3531 }
3532 }
3533 #endif //ASSERT
3534
3535 return javaoop;
3536 }
3537
3538 //---------------------------new_instance--------------------------------------
3539 // This routine takes a klass_node which may be constant (for a static type)
3540 // or may be non-constant (for reflective code). It will work equally well
3541 // for either, and the graph will fold nicely if the optimizer later reduces
3542 // the type to a constant.
3543 // The optional arguments are for specialized use by intrinsics:
3544 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3545 // - If 'return_size_val', report the the total object size to the caller.
3546 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
new_instance(Node * klass_node,Node * extra_slow_test,Node ** return_size_val,bool deoptimize_on_exception)3547 Node* GraphKit::new_instance(Node* klass_node,
3548 Node* extra_slow_test,
3549 Node* *return_size_val,
3550 bool deoptimize_on_exception) {
3551 // Compute size in doublewords
3552 // The size is always an integral number of doublewords, represented
3553 // as a positive bytewise size stored in the klass's layout_helper.
3554 // The layout_helper also encodes (in a low bit) the need for a slow path.
3555 jint layout_con = Klass::_lh_neutral_value;
3556 Node* layout_val = get_layout_helper(klass_node, layout_con);
3557 int layout_is_con = (layout_val == NULL);
3558
3559 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3560 // Generate the initial go-slow test. It's either ALWAYS (return a
3561 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3562 // case) a computed value derived from the layout_helper.
3563 Node* initial_slow_test = NULL;
3564 if (layout_is_con) {
3565 assert(!StressReflectiveCode, "stress mode does not use these paths");
3566 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3567 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3568 } else { // reflective case
3569 // This reflective path is used by Unsafe.allocateInstance.
3570 // (It may be stress-tested by specifying StressReflectiveCode.)
3571 // Basically, we want to get into the VM is there's an illegal argument.
3572 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3573 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3574 if (extra_slow_test != intcon(0)) {
3575 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3576 }
3577 // (Macro-expander will further convert this to a Bool, if necessary.)
3578 }
3579
3580 // Find the size in bytes. This is easy; it's the layout_helper.
3581 // The size value must be valid even if the slow path is taken.
3582 Node* size = NULL;
3583 if (layout_is_con) {
3584 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3585 } else { // reflective case
3586 // This reflective path is used by clone and Unsafe.allocateInstance.
3587 size = ConvI2X(layout_val);
3588
3589 // Clear the low bits to extract layout_helper_size_in_bytes:
3590 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3591 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3592 size = _gvn.transform( new AndXNode(size, mask) );
3593 }
3594 if (return_size_val != NULL) {
3595 (*return_size_val) = size;
3596 }
3597
3598 // This is a precise notnull oop of the klass.
3599 // (Actually, it need not be precise if this is a reflective allocation.)
3600 // It's what we cast the result to.
3601 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3602 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3603 const TypeOopPtr* oop_type = tklass->as_instance_type();
3604
3605 // Now generate allocation code
3606
3607 // The entire memory state is needed for slow path of the allocation
3608 // since GC and deoptimization can happened.
3609 Node *mem = reset_memory();
3610 set_all_memory(mem); // Create new memory state
3611
3612 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3613 control(), mem, i_o(),
3614 size, klass_node,
3615 initial_slow_test);
3616
3617 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3618 }
3619
3620 //-------------------------------new_array-------------------------------------
3621 // helper for both newarray and anewarray
3622 // The 'length' parameter is (obviously) the length of the array.
3623 // See comments on new_instance for the meaning of the other arguments.
new_array(Node * klass_node,Node * length,int nargs,Node ** return_size_val,bool deoptimize_on_exception)3624 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3625 Node* length, // number of array elements
3626 int nargs, // number of arguments to push back for uncommon trap
3627 Node* *return_size_val,
3628 bool deoptimize_on_exception) {
3629 jint layout_con = Klass::_lh_neutral_value;
3630 Node* layout_val = get_layout_helper(klass_node, layout_con);
3631 int layout_is_con = (layout_val == NULL);
3632
3633 if (!layout_is_con && !StressReflectiveCode &&
3634 !too_many_traps(Deoptimization::Reason_class_check)) {
3635 // This is a reflective array creation site.
3636 // Optimistically assume that it is a subtype of Object[],
3637 // so that we can fold up all the address arithmetic.
3638 layout_con = Klass::array_layout_helper(T_OBJECT);
3639 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3640 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3641 { BuildCutout unless(this, bol_lh, PROB_MAX);
3642 inc_sp(nargs);
3643 uncommon_trap(Deoptimization::Reason_class_check,
3644 Deoptimization::Action_maybe_recompile);
3645 }
3646 layout_val = NULL;
3647 layout_is_con = true;
3648 }
3649
3650 // Generate the initial go-slow test. Make sure we do not overflow
3651 // if length is huge (near 2Gig) or negative! We do not need
3652 // exact double-words here, just a close approximation of needed
3653 // double-words. We can't add any offset or rounding bits, lest we
3654 // take a size -1 of bytes and make it positive. Use an unsigned
3655 // compare, so negative sizes look hugely positive.
3656 int fast_size_limit = FastAllocateSizeLimit;
3657 if (layout_is_con) {
3658 assert(!StressReflectiveCode, "stress mode does not use these paths");
3659 // Increase the size limit if we have exact knowledge of array type.
3660 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3661 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3662 }
3663
3664 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3665 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3666
3667 // --- Size Computation ---
3668 // array_size = round_to_heap(array_header + (length << elem_shift));
3669 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3670 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3671 // The rounding mask is strength-reduced, if possible.
3672 int round_mask = MinObjAlignmentInBytes - 1;
3673 Node* header_size = NULL;
3674 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3675 // (T_BYTE has the weakest alignment and size restrictions...)
3676 if (layout_is_con) {
3677 int hsize = Klass::layout_helper_header_size(layout_con);
3678 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3679 BasicType etype = Klass::layout_helper_element_type(layout_con);
3680 if ((round_mask & ~right_n_bits(eshift)) == 0)
3681 round_mask = 0; // strength-reduce it if it goes away completely
3682 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3683 assert(header_size_min <= hsize, "generic minimum is smallest");
3684 header_size_min = hsize;
3685 header_size = intcon(hsize + round_mask);
3686 } else {
3687 Node* hss = intcon(Klass::_lh_header_size_shift);
3688 Node* hsm = intcon(Klass::_lh_header_size_mask);
3689 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3690 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3691 Node* mask = intcon(round_mask);
3692 header_size = _gvn.transform( new AddINode(hsize, mask) );
3693 }
3694
3695 Node* elem_shift = NULL;
3696 if (layout_is_con) {
3697 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3698 if (eshift != 0)
3699 elem_shift = intcon(eshift);
3700 } else {
3701 // There is no need to mask or shift this value.
3702 // The semantics of LShiftINode include an implicit mask to 0x1F.
3703 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3704 elem_shift = layout_val;
3705 }
3706
3707 // Transition to native address size for all offset calculations:
3708 Node* lengthx = ConvI2X(length);
3709 Node* headerx = ConvI2X(header_size);
3710 #ifdef _LP64
3711 { const TypeInt* tilen = _gvn.find_int_type(length);
3712 if (tilen != NULL && tilen->_lo < 0) {
3713 // Add a manual constraint to a positive range. Cf. array_element_address.
3714 jint size_max = fast_size_limit;
3715 if (size_max > tilen->_hi) size_max = tilen->_hi;
3716 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3717
3718 // Only do a narrow I2L conversion if the range check passed.
3719 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3720 _gvn.transform(iff);
3721 RegionNode* region = new RegionNode(3);
3722 _gvn.set_type(region, Type::CONTROL);
3723 lengthx = new PhiNode(region, TypeLong::LONG);
3724 _gvn.set_type(lengthx, TypeLong::LONG);
3725
3726 // Range check passed. Use ConvI2L node with narrow type.
3727 Node* passed = IfFalse(iff);
3728 region->init_req(1, passed);
3729 // Make I2L conversion control dependent to prevent it from
3730 // floating above the range check during loop optimizations.
3731 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3732
3733 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3734 region->init_req(2, IfTrue(iff));
3735 lengthx->init_req(2, ConvI2X(length));
3736
3737 set_control(region);
3738 record_for_igvn(region);
3739 record_for_igvn(lengthx);
3740 }
3741 }
3742 #endif
3743
3744 // Combine header size (plus rounding) and body size. Then round down.
3745 // This computation cannot overflow, because it is used only in two
3746 // places, one where the length is sharply limited, and the other
3747 // after a successful allocation.
3748 Node* abody = lengthx;
3749 if (elem_shift != NULL)
3750 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3751 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3752 if (round_mask != 0) {
3753 Node* mask = MakeConX(~round_mask);
3754 size = _gvn.transform( new AndXNode(size, mask) );
3755 }
3756 // else if round_mask == 0, the size computation is self-rounding
3757
3758 if (return_size_val != NULL) {
3759 // This is the size
3760 (*return_size_val) = size;
3761 }
3762
3763 // Now generate allocation code
3764
3765 // The entire memory state is needed for slow path of the allocation
3766 // since GC and deoptimization can happened.
3767 Node *mem = reset_memory();
3768 set_all_memory(mem); // Create new memory state
3769
3770 if (initial_slow_test->is_Bool()) {
3771 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3772 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3773 }
3774
3775 // Create the AllocateArrayNode and its result projections
3776 AllocateArrayNode* alloc
3777 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3778 control(), mem, i_o(),
3779 size, klass_node,
3780 initial_slow_test,
3781 length);
3782
3783 // Cast to correct type. Note that the klass_node may be constant or not,
3784 // and in the latter case the actual array type will be inexact also.
3785 // (This happens via a non-constant argument to inline_native_newArray.)
3786 // In any case, the value of klass_node provides the desired array type.
3787 const TypeInt* length_type = _gvn.find_int_type(length);
3788 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3789 if (ary_type->isa_aryptr() && length_type != NULL) {
3790 // Try to get a better type than POS for the size
3791 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3792 }
3793
3794 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3795
3796 // Cast length on remaining path to be as narrow as possible
3797 if (map()->find_edge(length) >= 0) {
3798 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3799 if (ccast != length) {
3800 _gvn.set_type_bottom(ccast);
3801 record_for_igvn(ccast);
3802 replace_in_map(length, ccast);
3803 }
3804 }
3805
3806 return javaoop;
3807 }
3808
3809 // The following "Ideal_foo" functions are placed here because they recognize
3810 // the graph shapes created by the functions immediately above.
3811
3812 //---------------------------Ideal_allocation----------------------------------
3813 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
Ideal_allocation(Node * ptr,PhaseTransform * phase)3814 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3815 if (ptr == NULL) { // reduce dumb test in callers
3816 return NULL;
3817 }
3818
3819 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
3820 ptr = bs->step_over_gc_barrier(ptr);
3821
3822 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3823 ptr = ptr->in(1);
3824 if (ptr == NULL) return NULL;
3825 }
3826 // Return NULL for allocations with several casts:
3827 // j.l.reflect.Array.newInstance(jobject, jint)
3828 // Object.clone()
3829 // to keep more precise type from last cast.
3830 if (ptr->is_Proj()) {
3831 Node* allo = ptr->in(0);
3832 if (allo != NULL && allo->is_Allocate()) {
3833 return allo->as_Allocate();
3834 }
3835 }
3836 // Report failure to match.
3837 return NULL;
3838 }
3839
3840 // Fancy version which also strips off an offset (and reports it to caller).
Ideal_allocation(Node * ptr,PhaseTransform * phase,intptr_t & offset)3841 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3842 intptr_t& offset) {
3843 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3844 if (base == NULL) return NULL;
3845 return Ideal_allocation(base, phase);
3846 }
3847
3848 // Trace Initialize <- Proj[Parm] <- Allocate
allocation()3849 AllocateNode* InitializeNode::allocation() {
3850 Node* rawoop = in(InitializeNode::RawAddress);
3851 if (rawoop->is_Proj()) {
3852 Node* alloc = rawoop->in(0);
3853 if (alloc->is_Allocate()) {
3854 return alloc->as_Allocate();
3855 }
3856 }
3857 return NULL;
3858 }
3859
3860 // Trace Allocate -> Proj[Parm] -> Initialize
initialization()3861 InitializeNode* AllocateNode::initialization() {
3862 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
3863 if (rawoop == NULL) return NULL;
3864 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3865 Node* init = rawoop->fast_out(i);
3866 if (init->is_Initialize()) {
3867 assert(init->as_Initialize()->allocation() == this, "2-way link");
3868 return init->as_Initialize();
3869 }
3870 }
3871 return NULL;
3872 }
3873
3874 //----------------------------- loop predicates ---------------------------
3875
3876 //------------------------------add_predicate_impl----------------------------
add_empty_predicate_impl(Deoptimization::DeoptReason reason,int nargs)3877 void GraphKit::add_empty_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3878 // Too many traps seen?
3879 if (too_many_traps(reason)) {
3880 #ifdef ASSERT
3881 if (TraceLoopPredicate) {
3882 int tc = C->trap_count(reason);
3883 tty->print("too many traps=%s tcount=%d in ",
3884 Deoptimization::trap_reason_name(reason), tc);
3885 method()->print(); // which method has too many predicate traps
3886 tty->cr();
3887 }
3888 #endif
3889 // We cannot afford to take more traps here,
3890 // do not generate predicate.
3891 return;
3892 }
3893
3894 Node *cont = _gvn.intcon(1);
3895 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
3896 Node *bol = _gvn.transform(new Conv2BNode(opq));
3897 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3898 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3899 C->add_predicate_opaq(opq);
3900 {
3901 PreserveJVMState pjvms(this);
3902 set_control(iffalse);
3903 inc_sp(nargs);
3904 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3905 }
3906 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3907 set_control(iftrue);
3908 }
3909
3910 //------------------------------add_predicate---------------------------------
add_empty_predicates(int nargs)3911 void GraphKit::add_empty_predicates(int nargs) {
3912 // These loop predicates remain empty. All concrete loop predicates are inserted above the corresponding
3913 // empty loop predicate later by 'PhaseIdealLoop::create_new_if_for_predicate'. All concrete loop predicates of
3914 // a specific kind (normal, profile or limit check) share the same uncommon trap as the empty loop predicate.
3915 if (UseLoopPredicate) {
3916 add_empty_predicate_impl(Deoptimization::Reason_predicate, nargs);
3917 }
3918 if (UseProfiledLoopPredicate) {
3919 add_empty_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
3920 }
3921 // loop's limit check predicate should be near the loop.
3922 add_empty_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3923 }
3924
sync_kit(IdealKit & ideal)3925 void GraphKit::sync_kit(IdealKit& ideal) {
3926 set_all_memory(ideal.merged_memory());
3927 set_i_o(ideal.i_o());
3928 set_control(ideal.ctrl());
3929 }
3930
final_sync(IdealKit & ideal)3931 void GraphKit::final_sync(IdealKit& ideal) {
3932 // Final sync IdealKit and graphKit.
3933 sync_kit(ideal);
3934 }
3935
load_String_length(Node * str,bool set_ctrl)3936 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
3937 Node* len = load_array_length(load_String_value(str, set_ctrl));
3938 Node* coder = load_String_coder(str, set_ctrl);
3939 // Divide length by 2 if coder is UTF16
3940 return _gvn.transform(new RShiftINode(len, coder));
3941 }
3942
load_String_value(Node * str,bool set_ctrl)3943 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
3944 int value_offset = java_lang_String::value_offset_in_bytes();
3945 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3946 false, NULL, 0);
3947 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3948 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3949 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
3950 ciTypeArrayKlass::make(T_BYTE), true, 0);
3951 Node* p = basic_plus_adr(str, str, value_offset);
3952 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
3953 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
3954 return load;
3955 }
3956
load_String_coder(Node * str,bool set_ctrl)3957 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
3958 if (!CompactStrings) {
3959 return intcon(java_lang_String::CODER_UTF16);
3960 }
3961 int coder_offset = java_lang_String::coder_offset_in_bytes();
3962 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3963 false, NULL, 0);
3964 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
3965
3966 Node* p = basic_plus_adr(str, str, coder_offset);
3967 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
3968 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
3969 return load;
3970 }
3971
store_String_value(Node * str,Node * value)3972 void GraphKit::store_String_value(Node* str, Node* value) {
3973 int value_offset = java_lang_String::value_offset_in_bytes();
3974 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3975 false, NULL, 0);
3976 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3977
3978 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
3979 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
3980 }
3981
store_String_coder(Node * str,Node * value)3982 void GraphKit::store_String_coder(Node* str, Node* value) {
3983 int coder_offset = java_lang_String::coder_offset_in_bytes();
3984 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3985 false, NULL, 0);
3986 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
3987
3988 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
3989 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
3990 }
3991
3992 // Capture src and dst memory state with a MergeMemNode
capture_memory(const TypePtr * src_type,const TypePtr * dst_type)3993 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
3994 if (src_type == dst_type) {
3995 // Types are equal, we don't need a MergeMemNode
3996 return memory(src_type);
3997 }
3998 MergeMemNode* merge = MergeMemNode::make(map()->memory());
3999 record_for_igvn(merge); // fold it up later, if possible
4000 int src_idx = C->get_alias_index(src_type);
4001 int dst_idx = C->get_alias_index(dst_type);
4002 merge->set_memory_at(src_idx, memory(src_idx));
4003 merge->set_memory_at(dst_idx, memory(dst_idx));
4004 return merge;
4005 }
4006
compress_string(Node * src,const TypeAryPtr * src_type,Node * dst,Node * count)4007 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4008 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4009 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4010 // If input and output memory types differ, capture both states to preserve
4011 // the dependency between preceding and subsequent loads/stores.
4012 // For example, the following program:
4013 // StoreB
4014 // compress_string
4015 // LoadB
4016 // has this memory graph (use->def):
4017 // LoadB -> compress_string -> CharMem
4018 // ... -> StoreB -> ByteMem
4019 // The intrinsic hides the dependency between LoadB and StoreB, causing
4020 // the load to read from memory not containing the result of the StoreB.
4021 // The correct memory graph should look like this:
4022 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4023 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4024 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4025 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4026 set_memory(res_mem, TypeAryPtr::BYTES);
4027 return str;
4028 }
4029
inflate_string(Node * src,Node * dst,const TypeAryPtr * dst_type,Node * count)4030 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4031 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4032 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4033 // Capture src and dst memory (see comment in 'compress_string').
4034 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4035 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4036 set_memory(_gvn.transform(str), dst_type);
4037 }
4038
inflate_string_slow(Node * src,Node * dst,Node * start,Node * count)4039 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4040 /**
4041 * int i_char = start;
4042 * for (int i_byte = 0; i_byte < count; i_byte++) {
4043 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4044 * }
4045 */
4046 add_empty_predicates();
4047 RegionNode* head = new RegionNode(3);
4048 head->init_req(1, control());
4049 gvn().set_type(head, Type::CONTROL);
4050 record_for_igvn(head);
4051
4052 Node* i_byte = new PhiNode(head, TypeInt::INT);
4053 i_byte->init_req(1, intcon(0));
4054 gvn().set_type(i_byte, TypeInt::INT);
4055 record_for_igvn(i_byte);
4056
4057 Node* i_char = new PhiNode(head, TypeInt::INT);
4058 i_char->init_req(1, start);
4059 gvn().set_type(i_char, TypeInt::INT);
4060 record_for_igvn(i_char);
4061
4062 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4063 gvn().set_type(mem, Type::MEMORY);
4064 record_for_igvn(mem);
4065 set_control(head);
4066 set_memory(mem, TypeAryPtr::BYTES);
4067 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4068 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4069 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4070 false, false, true /* mismatched */);
4071
4072 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4073 head->init_req(2, IfTrue(iff));
4074 mem->init_req(2, st);
4075 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4076 i_char->init_req(2, AddI(i_char, intcon(2)));
4077
4078 set_control(IfFalse(iff));
4079 set_memory(st, TypeAryPtr::BYTES);
4080 }
4081
make_constant_from_field(ciField * field,Node * obj)4082 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4083 if (!field->is_constant()) {
4084 return NULL; // Field not marked as constant.
4085 }
4086 ciInstance* holder = NULL;
4087 if (!field->is_static()) {
4088 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4089 if (const_oop != NULL && const_oop->is_instance()) {
4090 holder = const_oop->as_instance();
4091 }
4092 }
4093 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4094 /*is_unsigned_load=*/false);
4095 if (con_type != NULL) {
4096 return makecon(con_type);
4097 }
4098 return NULL;
4099 }
4100