1 /*
2 * Copyright (c) 2001, 2018, 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
47 //----------------------------GraphKit-----------------------------------------
48 // Main utility constructor.
GraphKit(JVMState * jvms)49 GraphKit::GraphKit(JVMState* jvms)
50 : Phase(Phase::Parser),
51 _env(C->env()),
52 _gvn(*C->initial_gvn()),
53 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
54 {
55 _exceptions = jvms->map()->next_exception();
56 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
57 set_jvms(jvms);
58 }
59
60 // Private constructor for parser.
GraphKit()61 GraphKit::GraphKit()
62 : Phase(Phase::Parser),
63 _env(C->env()),
64 _gvn(*C->initial_gvn()),
65 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
66 {
67 _exceptions = NULL;
68 set_map(NULL);
69 debug_only(_sp = -99);
70 debug_only(set_bci(-99));
71 }
72
73
74
75 //---------------------------clean_stack---------------------------------------
76 // Clear away rubbish from the stack area of the JVM state.
77 // This destroys any arguments that may be waiting on the stack.
clean_stack(int from_sp)78 void GraphKit::clean_stack(int from_sp) {
79 SafePointNode* map = this->map();
80 JVMState* jvms = this->jvms();
81 int stk_size = jvms->stk_size();
82 int stkoff = jvms->stkoff();
83 Node* top = this->top();
84 for (int i = from_sp; i < stk_size; i++) {
85 if (map->in(stkoff + i) != top) {
86 map->set_req(stkoff + i, top);
87 }
88 }
89 }
90
91
92 //--------------------------------sync_jvms-----------------------------------
93 // Make sure our current jvms agrees with our parse state.
sync_jvms() const94 JVMState* GraphKit::sync_jvms() const {
95 JVMState* jvms = this->jvms();
96 jvms->set_bci(bci()); // Record the new bci in the JVMState
97 jvms->set_sp(sp()); // Record the new sp in the JVMState
98 assert(jvms_in_sync(), "jvms is now in sync");
99 return jvms;
100 }
101
102 //--------------------------------sync_jvms_for_reexecute---------------------
103 // Make sure our current jvms agrees with our parse state. This version
104 // uses the reexecute_sp for reexecuting bytecodes.
sync_jvms_for_reexecute()105 JVMState* GraphKit::sync_jvms_for_reexecute() {
106 JVMState* jvms = this->jvms();
107 jvms->set_bci(bci()); // Record the new bci in the JVMState
108 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
109 return jvms;
110 }
111
112 #ifdef ASSERT
jvms_in_sync() const113 bool GraphKit::jvms_in_sync() const {
114 Parse* parse = is_Parse();
115 if (parse == NULL) {
116 if (bci() != jvms()->bci()) return false;
117 if (sp() != (int)jvms()->sp()) return false;
118 return true;
119 }
120 if (jvms()->method() != parse->method()) return false;
121 if (jvms()->bci() != parse->bci()) return false;
122 int jvms_sp = jvms()->sp();
123 if (jvms_sp != parse->sp()) return false;
124 int jvms_depth = jvms()->depth();
125 if (jvms_depth != parse->depth()) return false;
126 return true;
127 }
128
129 // Local helper checks for special internal merge points
130 // used to accumulate and merge exception states.
131 // They are marked by the region's in(0) edge being the map itself.
132 // Such merge points must never "escape" into the parser at large,
133 // until they have been handed to gvn.transform.
is_hidden_merge(Node * reg)134 static bool is_hidden_merge(Node* reg) {
135 if (reg == NULL) return false;
136 if (reg->is_Phi()) {
137 reg = reg->in(0);
138 if (reg == NULL) return false;
139 }
140 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
141 }
142
verify_map() const143 void GraphKit::verify_map() const {
144 if (map() == NULL) return; // null map is OK
145 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
146 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
147 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
148 }
149
verify_exception_state(SafePointNode * ex_map)150 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
151 assert(ex_map->next_exception() == NULL, "not already part of a chain");
152 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
153 }
154 #endif
155
156 //---------------------------stop_and_kill_map---------------------------------
157 // Set _map to NULL, signalling a stop to further bytecode execution.
158 // First smash the current map's control to a constant, to mark it dead.
stop_and_kill_map()159 void GraphKit::stop_and_kill_map() {
160 SafePointNode* dead_map = stop();
161 if (dead_map != NULL) {
162 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
163 assert(dead_map->is_killed(), "must be so marked");
164 }
165 }
166
167
168 //--------------------------------stopped--------------------------------------
169 // Tell if _map is NULL, or control is top.
stopped()170 bool GraphKit::stopped() {
171 if (map() == NULL) return true;
172 else if (control() == top()) return true;
173 else return false;
174 }
175
176
177 //-----------------------------has_ex_handler----------------------------------
178 // Tell if this method or any caller method has exception handlers.
has_ex_handler()179 bool GraphKit::has_ex_handler() {
180 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
181 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
182 return true;
183 }
184 }
185 return false;
186 }
187
188 //------------------------------save_ex_oop------------------------------------
189 // Save an exception without blowing stack contents or other JVM state.
set_saved_ex_oop(SafePointNode * ex_map,Node * ex_oop)190 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
191 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
192 ex_map->add_req(ex_oop);
193 debug_only(verify_exception_state(ex_map));
194 }
195
common_saved_ex_oop(SafePointNode * ex_map,bool clear_it)196 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
197 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
198 Node* ex_oop = ex_map->in(ex_map->req()-1);
199 if (clear_it) ex_map->del_req(ex_map->req()-1);
200 return ex_oop;
201 }
202
203 //-----------------------------saved_ex_oop------------------------------------
204 // Recover a saved exception from its map.
saved_ex_oop(SafePointNode * ex_map)205 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
206 return common_saved_ex_oop(ex_map, false);
207 }
208
209 //--------------------------clear_saved_ex_oop---------------------------------
210 // Erase a previously saved exception from its map.
clear_saved_ex_oop(SafePointNode * ex_map)211 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
212 return common_saved_ex_oop(ex_map, true);
213 }
214
215 #ifdef ASSERT
216 //---------------------------has_saved_ex_oop----------------------------------
217 // Erase a previously saved exception from its map.
has_saved_ex_oop(SafePointNode * ex_map)218 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
219 return ex_map->req() == ex_map->jvms()->endoff()+1;
220 }
221 #endif
222
223 //-------------------------make_exception_state--------------------------------
224 // Turn the current JVM state into an exception state, appending the ex_oop.
make_exception_state(Node * ex_oop)225 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
226 sync_jvms();
227 SafePointNode* ex_map = stop(); // do not manipulate this map any more
228 set_saved_ex_oop(ex_map, ex_oop);
229 return ex_map;
230 }
231
232
233 //--------------------------add_exception_state--------------------------------
234 // Add an exception to my list of exceptions.
add_exception_state(SafePointNode * ex_map)235 void GraphKit::add_exception_state(SafePointNode* ex_map) {
236 if (ex_map == NULL || ex_map->control() == top()) {
237 return;
238 }
239 #ifdef ASSERT
240 verify_exception_state(ex_map);
241 if (has_exceptions()) {
242 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
243 }
244 #endif
245
246 // If there is already an exception of exactly this type, merge with it.
247 // In particular, null-checks and other low-level exceptions common up here.
248 Node* ex_oop = saved_ex_oop(ex_map);
249 const Type* ex_type = _gvn.type(ex_oop);
250 if (ex_oop == top()) {
251 // No action needed.
252 return;
253 }
254 assert(ex_type->isa_instptr(), "exception must be an instance");
255 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
256 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
257 // We check sp also because call bytecodes can generate exceptions
258 // both before and after arguments are popped!
259 if (ex_type2 == ex_type
260 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
261 combine_exception_states(ex_map, e2);
262 return;
263 }
264 }
265
266 // No pre-existing exception of the same type. Chain it on the list.
267 push_exception_state(ex_map);
268 }
269
270 //-----------------------add_exception_states_from-----------------------------
add_exception_states_from(JVMState * jvms)271 void GraphKit::add_exception_states_from(JVMState* jvms) {
272 SafePointNode* ex_map = jvms->map()->next_exception();
273 if (ex_map != NULL) {
274 jvms->map()->set_next_exception(NULL);
275 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
276 next_map = ex_map->next_exception();
277 ex_map->set_next_exception(NULL);
278 add_exception_state(ex_map);
279 }
280 }
281 }
282
283 //-----------------------transfer_exceptions_into_jvms-------------------------
transfer_exceptions_into_jvms()284 JVMState* GraphKit::transfer_exceptions_into_jvms() {
285 if (map() == NULL) {
286 // We need a JVMS to carry the exceptions, but the map has gone away.
287 // Create a scratch JVMS, cloned from any of the exception states...
288 if (has_exceptions()) {
289 _map = _exceptions;
290 _map = clone_map();
291 _map->set_next_exception(NULL);
292 clear_saved_ex_oop(_map);
293 debug_only(verify_map());
294 } else {
295 // ...or created from scratch
296 JVMState* jvms = new (C) JVMState(_method, NULL);
297 jvms->set_bci(_bci);
298 jvms->set_sp(_sp);
299 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
300 set_jvms(jvms);
301 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
302 set_all_memory(top());
303 while (map()->req() < jvms->endoff()) map()->add_req(top());
304 }
305 // (This is a kludge, in case you didn't notice.)
306 set_control(top());
307 }
308 JVMState* jvms = sync_jvms();
309 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
310 jvms->map()->set_next_exception(_exceptions);
311 _exceptions = NULL; // done with this set of exceptions
312 return jvms;
313 }
314
add_n_reqs(Node * dstphi,Node * srcphi)315 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
316 assert(is_hidden_merge(dstphi), "must be a special merge node");
317 assert(is_hidden_merge(srcphi), "must be a special merge node");
318 uint limit = srcphi->req();
319 for (uint i = PhiNode::Input; i < limit; i++) {
320 dstphi->add_req(srcphi->in(i));
321 }
322 }
add_one_req(Node * dstphi,Node * src)323 static inline void add_one_req(Node* dstphi, Node* src) {
324 assert(is_hidden_merge(dstphi), "must be a special merge node");
325 assert(!is_hidden_merge(src), "must not be a special merge node");
326 dstphi->add_req(src);
327 }
328
329 //-----------------------combine_exception_states------------------------------
330 // This helper function combines exception states by building phis on a
331 // specially marked state-merging region. These regions and phis are
332 // untransformed, and can build up gradually. The region is marked by
333 // having a control input of its exception map, rather than NULL. Such
334 // regions do not appear except in this function, and in use_exception_state.
combine_exception_states(SafePointNode * ex_map,SafePointNode * phi_map)335 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
336 if (failing()) return; // dying anyway...
337 JVMState* ex_jvms = ex_map->_jvms;
338 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
339 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
340 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
341 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
342 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
343 assert(ex_map->req() == phi_map->req(), "matching maps");
344 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
345 Node* hidden_merge_mark = root();
346 Node* region = phi_map->control();
347 MergeMemNode* phi_mem = phi_map->merged_memory();
348 MergeMemNode* ex_mem = ex_map->merged_memory();
349 if (region->in(0) != hidden_merge_mark) {
350 // The control input is not (yet) a specially-marked region in phi_map.
351 // Make it so, and build some phis.
352 region = new RegionNode(2);
353 _gvn.set_type(region, Type::CONTROL);
354 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
355 region->init_req(1, phi_map->control());
356 phi_map->set_control(region);
357 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
358 record_for_igvn(io_phi);
359 _gvn.set_type(io_phi, Type::ABIO);
360 phi_map->set_i_o(io_phi);
361 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
362 Node* m = mms.memory();
363 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
364 record_for_igvn(m_phi);
365 _gvn.set_type(m_phi, Type::MEMORY);
366 mms.set_memory(m_phi);
367 }
368 }
369
370 // Either or both of phi_map and ex_map might already be converted into phis.
371 Node* ex_control = ex_map->control();
372 // if there is special marking on ex_map also, we add multiple edges from src
373 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
374 // how wide was the destination phi_map, originally?
375 uint orig_width = region->req();
376
377 if (add_multiple) {
378 add_n_reqs(region, ex_control);
379 add_n_reqs(phi_map->i_o(), ex_map->i_o());
380 } else {
381 // ex_map has no merges, so we just add single edges everywhere
382 add_one_req(region, ex_control);
383 add_one_req(phi_map->i_o(), ex_map->i_o());
384 }
385 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
386 if (mms.is_empty()) {
387 // get a copy of the base memory, and patch some inputs into it
388 const TypePtr* adr_type = mms.adr_type(C);
389 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
390 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
391 mms.set_memory(phi);
392 // Prepare to append interesting stuff onto the newly sliced phi:
393 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
394 }
395 // Append stuff from ex_map:
396 if (add_multiple) {
397 add_n_reqs(mms.memory(), mms.memory2());
398 } else {
399 add_one_req(mms.memory(), mms.memory2());
400 }
401 }
402 uint limit = ex_map->req();
403 for (uint i = TypeFunc::Parms; i < limit; i++) {
404 // Skip everything in the JVMS after tos. (The ex_oop follows.)
405 if (i == tos) i = ex_jvms->monoff();
406 Node* src = ex_map->in(i);
407 Node* dst = phi_map->in(i);
408 if (src != dst) {
409 PhiNode* phi;
410 if (dst->in(0) != region) {
411 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
412 record_for_igvn(phi);
413 _gvn.set_type(phi, phi->type());
414 phi_map->set_req(i, dst);
415 // Prepare to append interesting stuff onto the new phi:
416 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
417 } else {
418 assert(dst->is_Phi(), "nobody else uses a hidden region");
419 phi = dst->as_Phi();
420 }
421 if (add_multiple && src->in(0) == ex_control) {
422 // Both are phis.
423 add_n_reqs(dst, src);
424 } else {
425 while (dst->req() < region->req()) add_one_req(dst, src);
426 }
427 const Type* srctype = _gvn.type(src);
428 if (phi->type() != srctype) {
429 const Type* dsttype = phi->type()->meet_speculative(srctype);
430 if (phi->type() != dsttype) {
431 phi->set_type(dsttype);
432 _gvn.set_type(phi, dsttype);
433 }
434 }
435 }
436 }
437 phi_map->merge_replaced_nodes_with(ex_map);
438 }
439
440 //--------------------------use_exception_state--------------------------------
use_exception_state(SafePointNode * phi_map)441 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
442 if (failing()) { stop(); return top(); }
443 Node* region = phi_map->control();
444 Node* hidden_merge_mark = root();
445 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
446 Node* ex_oop = clear_saved_ex_oop(phi_map);
447 if (region->in(0) == hidden_merge_mark) {
448 // Special marking for internal ex-states. Process the phis now.
449 region->set_req(0, region); // now it's an ordinary region
450 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
451 // Note: Setting the jvms also sets the bci and sp.
452 set_control(_gvn.transform(region));
453 uint tos = jvms()->stkoff() + sp();
454 for (uint i = 1; i < tos; i++) {
455 Node* x = phi_map->in(i);
456 if (x->in(0) == region) {
457 assert(x->is_Phi(), "expected a special phi");
458 phi_map->set_req(i, _gvn.transform(x));
459 }
460 }
461 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
462 Node* x = mms.memory();
463 if (x->in(0) == region) {
464 assert(x->is_Phi(), "nobody else uses a hidden region");
465 mms.set_memory(_gvn.transform(x));
466 }
467 }
468 if (ex_oop->in(0) == region) {
469 assert(ex_oop->is_Phi(), "expected a special phi");
470 ex_oop = _gvn.transform(ex_oop);
471 }
472 } else {
473 set_jvms(phi_map->jvms());
474 }
475
476 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
477 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
478 return ex_oop;
479 }
480
481 //---------------------------------java_bc-------------------------------------
java_bc() const482 Bytecodes::Code GraphKit::java_bc() const {
483 ciMethod* method = this->method();
484 int bci = this->bci();
485 if (method != NULL && bci != InvocationEntryBci)
486 return method->java_code_at_bci(bci);
487 else
488 return Bytecodes::_illegal;
489 }
490
uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,bool must_throw)491 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
492 bool must_throw) {
493 // if the exception capability is set, then we will generate code
494 // to check the JavaThread.should_post_on_exceptions flag to see
495 // if we actually need to report exception events (for this
496 // thread). If we don't need to report exception events, we will
497 // take the normal fast path provided by add_exception_events. If
498 // exception event reporting is enabled for this thread, we will
499 // take the uncommon_trap in the BuildCutout below.
500
501 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
502 Node* jthread = _gvn.transform(new ThreadLocalNode());
503 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
504 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
505
506 // Test the should_post_on_exceptions_flag vs. 0
507 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
508 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
509
510 // Branch to slow_path if should_post_on_exceptions_flag was true
511 { BuildCutout unless(this, tst, PROB_MAX);
512 // Do not try anything fancy if we're notifying the VM on every throw.
513 // Cf. case Bytecodes::_athrow in parse2.cpp.
514 uncommon_trap(reason, Deoptimization::Action_none,
515 (ciKlass*)NULL, (char*)NULL, must_throw);
516 }
517
518 }
519
520 //------------------------------builtin_throw----------------------------------
builtin_throw(Deoptimization::DeoptReason reason,Node * arg)521 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
522 bool must_throw = true;
523
524 if (env()->jvmti_can_post_on_exceptions()) {
525 // check if we must post exception events, take uncommon trap if so
526 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
527 // here if should_post_on_exceptions is false
528 // continue on with the normal codegen
529 }
530
531 // If this particular condition has not yet happened at this
532 // bytecode, then use the uncommon trap mechanism, and allow for
533 // a future recompilation if several traps occur here.
534 // If the throw is hot, try to use a more complicated inline mechanism
535 // which keeps execution inside the compiled code.
536 bool treat_throw_as_hot = false;
537 ciMethodData* md = method()->method_data();
538
539 if (ProfileTraps) {
540 if (too_many_traps(reason)) {
541 treat_throw_as_hot = true;
542 }
543 // (If there is no MDO at all, assume it is early in
544 // execution, and that any deopts are part of the
545 // startup transient, and don't need to be remembered.)
546
547 // Also, if there is a local exception handler, treat all throws
548 // as hot if there has been at least one in this method.
549 if (C->trap_count(reason) != 0
550 && method()->method_data()->trap_count(reason) != 0
551 && has_ex_handler()) {
552 treat_throw_as_hot = true;
553 }
554 }
555
556 // If this throw happens frequently, an uncommon trap might cause
557 // a performance pothole. If there is a local exception handler,
558 // and if this particular bytecode appears to be deoptimizing often,
559 // let us handle the throw inline, with a preconstructed instance.
560 // Note: If the deopt count has blown up, the uncommon trap
561 // runtime is going to flush this nmethod, not matter what.
562 if (treat_throw_as_hot
563 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
564 // If the throw is local, we use a pre-existing instance and
565 // punt on the backtrace. This would lead to a missing backtrace
566 // (a repeat of 4292742) if the backtrace object is ever asked
567 // for its backtrace.
568 // Fixing this remaining case of 4292742 requires some flavor of
569 // escape analysis. Leave that for the future.
570 ciInstance* ex_obj = NULL;
571 switch (reason) {
572 case Deoptimization::Reason_null_check:
573 ex_obj = env()->NullPointerException_instance();
574 break;
575 case Deoptimization::Reason_div0_check:
576 ex_obj = env()->ArithmeticException_instance();
577 break;
578 case Deoptimization::Reason_range_check:
579 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
580 break;
581 case Deoptimization::Reason_class_check:
582 if (java_bc() == Bytecodes::_aastore) {
583 ex_obj = env()->ArrayStoreException_instance();
584 } else {
585 ex_obj = env()->ClassCastException_instance();
586 }
587 break;
588 default:
589 break;
590 }
591 if (failing()) { stop(); return; } // exception allocation might fail
592 if (ex_obj != NULL) {
593 // Cheat with a preallocated exception object.
594 if (C->log() != NULL)
595 C->log()->elem("hot_throw preallocated='1' reason='%s'",
596 Deoptimization::trap_reason_name(reason));
597 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
598 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
599
600 // Clear the detail message of the preallocated exception object.
601 // Weblogic sometimes mutates the detail message of exceptions
602 // using reflection.
603 int offset = java_lang_Throwable::get_detailMessage_offset();
604 const TypePtr* adr_typ = ex_con->add_offset(offset);
605
606 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
607 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
608 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
609
610 add_exception_state(make_exception_state(ex_node));
611 return;
612 }
613 }
614
615 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
616 // It won't be much cheaper than bailing to the interp., since we'll
617 // have to pass up all the debug-info, and the runtime will have to
618 // create the stack trace.
619
620 // Usual case: Bail to interpreter.
621 // Reserve the right to recompile if we haven't seen anything yet.
622
623 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
624 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
625 if (treat_throw_as_hot
626 && (method()->method_data()->trap_recompiled_at(bci(), m)
627 || C->too_many_traps(reason))) {
628 // We cannot afford to take more traps here. Suffer in the interpreter.
629 if (C->log() != NULL)
630 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
631 Deoptimization::trap_reason_name(reason),
632 C->trap_count(reason));
633 action = Deoptimization::Action_none;
634 }
635
636 // "must_throw" prunes the JVM state to include only the stack, if there
637 // are no local exception handlers. This should cut down on register
638 // allocation time and code size, by drastically reducing the number
639 // of in-edges on the call to the uncommon trap.
640
641 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
642 }
643
644
645 //----------------------------PreserveJVMState---------------------------------
PreserveJVMState(GraphKit * kit,bool clone_map)646 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
647 debug_only(kit->verify_map());
648 _kit = kit;
649 _map = kit->map(); // preserve the map
650 _sp = kit->sp();
651 kit->set_map(clone_map ? kit->clone_map() : NULL);
652 #ifdef ASSERT
653 _bci = kit->bci();
654 Parse* parser = kit->is_Parse();
655 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
656 _block = block;
657 #endif
658 }
~PreserveJVMState()659 PreserveJVMState::~PreserveJVMState() {
660 GraphKit* kit = _kit;
661 #ifdef ASSERT
662 assert(kit->bci() == _bci, "bci must not shift");
663 Parse* parser = kit->is_Parse();
664 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
665 assert(block == _block, "block must not shift");
666 #endif
667 kit->set_map(_map);
668 kit->set_sp(_sp);
669 }
670
671
672 //-----------------------------BuildCutout-------------------------------------
BuildCutout(GraphKit * kit,Node * p,float prob,float cnt)673 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
674 : PreserveJVMState(kit)
675 {
676 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
677 SafePointNode* outer_map = _map; // preserved map is caller's
678 SafePointNode* inner_map = kit->map();
679 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
680 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
681 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
682 }
~BuildCutout()683 BuildCutout::~BuildCutout() {
684 GraphKit* kit = _kit;
685 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
686 }
687
688 //---------------------------PreserveReexecuteState----------------------------
PreserveReexecuteState(GraphKit * kit)689 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
690 assert(!kit->stopped(), "must call stopped() before");
691 _kit = kit;
692 _sp = kit->sp();
693 _reexecute = kit->jvms()->_reexecute;
694 }
~PreserveReexecuteState()695 PreserveReexecuteState::~PreserveReexecuteState() {
696 if (_kit->stopped()) return;
697 _kit->jvms()->_reexecute = _reexecute;
698 _kit->set_sp(_sp);
699 }
700
701 //------------------------------clone_map--------------------------------------
702 // Implementation of PreserveJVMState
703 //
704 // Only clone_map(...) here. If this function is only used in the
705 // PreserveJVMState class we may want to get rid of this extra
706 // function eventually and do it all there.
707
clone_map()708 SafePointNode* GraphKit::clone_map() {
709 if (map() == NULL) return NULL;
710
711 // Clone the memory edge first
712 Node* mem = MergeMemNode::make(map()->memory());
713 gvn().set_type_bottom(mem);
714
715 SafePointNode *clonemap = (SafePointNode*)map()->clone();
716 JVMState* jvms = this->jvms();
717 JVMState* clonejvms = jvms->clone_shallow(C);
718 clonemap->set_memory(mem);
719 clonemap->set_jvms(clonejvms);
720 clonejvms->set_map(clonemap);
721 record_for_igvn(clonemap);
722 gvn().set_type_bottom(clonemap);
723 return clonemap;
724 }
725
726
727 //-----------------------------set_map_clone-----------------------------------
set_map_clone(SafePointNode * m)728 void GraphKit::set_map_clone(SafePointNode* m) {
729 _map = m;
730 _map = clone_map();
731 _map->set_next_exception(NULL);
732 debug_only(verify_map());
733 }
734
735
736 //----------------------------kill_dead_locals---------------------------------
737 // Detect any locals which are known to be dead, and force them to top.
kill_dead_locals()738 void GraphKit::kill_dead_locals() {
739 // Consult the liveness information for the locals. If any
740 // of them are unused, then they can be replaced by top(). This
741 // should help register allocation time and cut down on the size
742 // of the deoptimization information.
743
744 // This call is made from many of the bytecode handling
745 // subroutines called from the Big Switch in do_one_bytecode.
746 // Every bytecode which might include a slow path is responsible
747 // for killing its dead locals. The more consistent we
748 // are about killing deads, the fewer useless phis will be
749 // constructed for them at various merge points.
750
751 // bci can be -1 (InvocationEntryBci). We return the entry
752 // liveness for the method.
753
754 if (method() == NULL || method()->code_size() == 0) {
755 // We are building a graph for a call to a native method.
756 // All locals are live.
757 return;
758 }
759
760 ResourceMark rm;
761
762 // Consult the liveness information for the locals. If any
763 // of them are unused, then they can be replaced by top(). This
764 // should help register allocation time and cut down on the size
765 // of the deoptimization information.
766 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
767
768 int len = (int)live_locals.size();
769 assert(len <= jvms()->loc_size(), "too many live locals");
770 for (int local = 0; local < len; local++) {
771 if (!live_locals.at(local)) {
772 set_local(local, top());
773 }
774 }
775 }
776
777 #ifdef ASSERT
778 //-------------------------dead_locals_are_killed------------------------------
779 // Return true if all dead locals are set to top in the map.
780 // Used to assert "clean" debug info at various points.
dead_locals_are_killed()781 bool GraphKit::dead_locals_are_killed() {
782 if (method() == NULL || method()->code_size() == 0) {
783 // No locals need to be dead, so all is as it should be.
784 return true;
785 }
786
787 // Make sure somebody called kill_dead_locals upstream.
788 ResourceMark rm;
789 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
790 if (jvms->loc_size() == 0) continue; // no locals to consult
791 SafePointNode* map = jvms->map();
792 ciMethod* method = jvms->method();
793 int bci = jvms->bci();
794 if (jvms == this->jvms()) {
795 bci = this->bci(); // it might not yet be synched
796 }
797 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
798 int len = (int)live_locals.size();
799 if (!live_locals.is_valid() || len == 0)
800 // This method is trivial, or is poisoned by a breakpoint.
801 return true;
802 assert(len == jvms->loc_size(), "live map consistent with locals map");
803 for (int local = 0; local < len; local++) {
804 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
805 if (PrintMiscellaneous && (Verbose || WizardMode)) {
806 tty->print_cr("Zombie local %d: ", local);
807 jvms->dump();
808 }
809 return false;
810 }
811 }
812 }
813 return true;
814 }
815
816 #endif //ASSERT
817
818 // Helper function for enforcing certain bytecodes to reexecute if
819 // deoptimization happens
should_reexecute_implied_by_bytecode(JVMState * jvms,bool is_anewarray)820 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
821 ciMethod* cur_method = jvms->method();
822 int cur_bci = jvms->bci();
823 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
824 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
825 return Interpreter::bytecode_should_reexecute(code) ||
826 (is_anewarray && code == Bytecodes::_multianewarray);
827 // Reexecute _multianewarray bytecode which was replaced with
828 // sequence of [a]newarray. See Parse::do_multianewarray().
829 //
830 // Note: interpreter should not have it set since this optimization
831 // is limited by dimensions and guarded by flag so in some cases
832 // multianewarray() runtime calls will be generated and
833 // the bytecode should not be reexecutes (stack will not be reset).
834 } else
835 return false;
836 }
837
838 // Helper function for adding JVMState and debug information to node
add_safepoint_edges(SafePointNode * call,bool must_throw)839 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
840 // Add the safepoint edges to the call (or other safepoint).
841
842 // Make sure dead locals are set to top. This
843 // should help register allocation time and cut down on the size
844 // of the deoptimization information.
845 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
846
847 // Walk the inline list to fill in the correct set of JVMState's
848 // Also fill in the associated edges for each JVMState.
849
850 // If the bytecode needs to be reexecuted we need to put
851 // the arguments back on the stack.
852 const bool should_reexecute = jvms()->should_reexecute();
853 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
854
855 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
856 // undefined if the bci is different. This is normal for Parse but it
857 // should not happen for LibraryCallKit because only one bci is processed.
858 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
859 "in LibraryCallKit the reexecute bit should not change");
860
861 // If we are guaranteed to throw, we can prune everything but the
862 // input to the current bytecode.
863 bool can_prune_locals = false;
864 uint stack_slots_not_pruned = 0;
865 int inputs = 0, depth = 0;
866 if (must_throw) {
867 assert(method() == youngest_jvms->method(), "sanity");
868 if (compute_stack_effects(inputs, depth)) {
869 can_prune_locals = true;
870 stack_slots_not_pruned = inputs;
871 }
872 }
873
874 if (env()->should_retain_local_variables()) {
875 // At any safepoint, this method can get breakpointed, which would
876 // then require an immediate deoptimization.
877 can_prune_locals = false; // do not prune locals
878 stack_slots_not_pruned = 0;
879 }
880
881 // do not scribble on the input jvms
882 JVMState* out_jvms = youngest_jvms->clone_deep(C);
883 call->set_jvms(out_jvms); // Start jvms list for call node
884
885 // For a known set of bytecodes, the interpreter should reexecute them if
886 // deoptimization happens. We set the reexecute state for them here
887 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
888 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
889 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
890 }
891
892 // Presize the call:
893 DEBUG_ONLY(uint non_debug_edges = call->req());
894 call->add_req_batch(top(), youngest_jvms->debug_depth());
895 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
896
897 // Set up edges so that the call looks like this:
898 // Call [state:] ctl io mem fptr retadr
899 // [parms:] parm0 ... parmN
900 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
901 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
902 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
903 // Note that caller debug info precedes callee debug info.
904
905 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
906 uint debug_ptr = call->req();
907
908 // Loop over the map input edges associated with jvms, add them
909 // to the call node, & reset all offsets to match call node array.
910 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
911 uint debug_end = debug_ptr;
912 uint debug_start = debug_ptr - in_jvms->debug_size();
913 debug_ptr = debug_start; // back up the ptr
914
915 uint p = debug_start; // walks forward in [debug_start, debug_end)
916 uint j, k, l;
917 SafePointNode* in_map = in_jvms->map();
918 out_jvms->set_map(call);
919
920 if (can_prune_locals) {
921 assert(in_jvms->method() == out_jvms->method(), "sanity");
922 // If the current throw can reach an exception handler in this JVMS,
923 // then we must keep everything live that can reach that handler.
924 // As a quick and dirty approximation, we look for any handlers at all.
925 if (in_jvms->method()->has_exception_handlers()) {
926 can_prune_locals = false;
927 }
928 }
929
930 // Add the Locals
931 k = in_jvms->locoff();
932 l = in_jvms->loc_size();
933 out_jvms->set_locoff(p);
934 if (!can_prune_locals) {
935 for (j = 0; j < l; j++)
936 call->set_req(p++, in_map->in(k+j));
937 } else {
938 p += l; // already set to top above by add_req_batch
939 }
940
941 // Add the Expression Stack
942 k = in_jvms->stkoff();
943 l = in_jvms->sp();
944 out_jvms->set_stkoff(p);
945 if (!can_prune_locals) {
946 for (j = 0; j < l; j++)
947 call->set_req(p++, in_map->in(k+j));
948 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
949 // Divide stack into {S0,...,S1}, where S0 is set to top.
950 uint s1 = stack_slots_not_pruned;
951 stack_slots_not_pruned = 0; // for next iteration
952 if (s1 > l) s1 = l;
953 uint s0 = l - s1;
954 p += s0; // skip the tops preinstalled by add_req_batch
955 for (j = s0; j < l; j++)
956 call->set_req(p++, in_map->in(k+j));
957 } else {
958 p += l; // already set to top above by add_req_batch
959 }
960
961 // Add the Monitors
962 k = in_jvms->monoff();
963 l = in_jvms->mon_size();
964 out_jvms->set_monoff(p);
965 for (j = 0; j < l; j++)
966 call->set_req(p++, in_map->in(k+j));
967
968 // Copy any scalar object fields.
969 k = in_jvms->scloff();
970 l = in_jvms->scl_size();
971 out_jvms->set_scloff(p);
972 for (j = 0; j < l; j++)
973 call->set_req(p++, in_map->in(k+j));
974
975 // Finish the new jvms.
976 out_jvms->set_endoff(p);
977
978 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
979 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
980 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
981 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
982 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
983 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
984
985 // Update the two tail pointers in parallel.
986 out_jvms = out_jvms->caller();
987 in_jvms = in_jvms->caller();
988 }
989
990 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
991
992 // Test the correctness of JVMState::debug_xxx accessors:
993 assert(call->jvms()->debug_start() == non_debug_edges, "");
994 assert(call->jvms()->debug_end() == call->req(), "");
995 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
996 }
997
compute_stack_effects(int & inputs,int & depth)998 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
999 Bytecodes::Code code = java_bc();
1000 if (code == Bytecodes::_wide) {
1001 code = method()->java_code_at_bci(bci() + 1);
1002 }
1003
1004 BasicType rtype = T_ILLEGAL;
1005 int rsize = 0;
1006
1007 if (code != Bytecodes::_illegal) {
1008 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1009 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1010 if (rtype < T_CONFLICT)
1011 rsize = type2size[rtype];
1012 }
1013
1014 switch (code) {
1015 case Bytecodes::_illegal:
1016 return false;
1017
1018 case Bytecodes::_ldc:
1019 case Bytecodes::_ldc_w:
1020 case Bytecodes::_ldc2_w:
1021 inputs = 0;
1022 break;
1023
1024 case Bytecodes::_dup: inputs = 1; break;
1025 case Bytecodes::_dup_x1: inputs = 2; break;
1026 case Bytecodes::_dup_x2: inputs = 3; break;
1027 case Bytecodes::_dup2: inputs = 2; break;
1028 case Bytecodes::_dup2_x1: inputs = 3; break;
1029 case Bytecodes::_dup2_x2: inputs = 4; break;
1030 case Bytecodes::_swap: inputs = 2; break;
1031 case Bytecodes::_arraylength: inputs = 1; break;
1032
1033 case Bytecodes::_getstatic:
1034 case Bytecodes::_putstatic:
1035 case Bytecodes::_getfield:
1036 case Bytecodes::_putfield:
1037 {
1038 bool ignored_will_link;
1039 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1040 int size = field->type()->size();
1041 bool is_get = (depth >= 0), is_static = (depth & 1);
1042 inputs = (is_static ? 0 : 1);
1043 if (is_get) {
1044 depth = size - inputs;
1045 } else {
1046 inputs += size; // putxxx pops the value from the stack
1047 depth = - inputs;
1048 }
1049 }
1050 break;
1051
1052 case Bytecodes::_invokevirtual:
1053 case Bytecodes::_invokespecial:
1054 case Bytecodes::_invokestatic:
1055 case Bytecodes::_invokedynamic:
1056 case Bytecodes::_invokeinterface:
1057 {
1058 bool ignored_will_link;
1059 ciSignature* declared_signature = NULL;
1060 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1061 assert(declared_signature != NULL, "cannot be null");
1062 inputs = declared_signature->arg_size_for_bc(code);
1063 int size = declared_signature->return_type()->size();
1064 depth = size - inputs;
1065 }
1066 break;
1067
1068 case Bytecodes::_multianewarray:
1069 {
1070 ciBytecodeStream iter(method());
1071 iter.reset_to_bci(bci());
1072 iter.next();
1073 inputs = iter.get_dimensions();
1074 assert(rsize == 1, "");
1075 depth = rsize - inputs;
1076 }
1077 break;
1078
1079 case Bytecodes::_ireturn:
1080 case Bytecodes::_lreturn:
1081 case Bytecodes::_freturn:
1082 case Bytecodes::_dreturn:
1083 case Bytecodes::_areturn:
1084 assert(rsize == -depth, "");
1085 inputs = rsize;
1086 break;
1087
1088 case Bytecodes::_jsr:
1089 case Bytecodes::_jsr_w:
1090 inputs = 0;
1091 depth = 1; // S.B. depth=1, not zero
1092 break;
1093
1094 default:
1095 // bytecode produces a typed result
1096 inputs = rsize - depth;
1097 assert(inputs >= 0, "");
1098 break;
1099 }
1100
1101 #ifdef ASSERT
1102 // spot check
1103 int outputs = depth + inputs;
1104 assert(outputs >= 0, "sanity");
1105 switch (code) {
1106 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1107 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1108 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1109 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1110 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1111 default: break;
1112 }
1113 #endif //ASSERT
1114
1115 return true;
1116 }
1117
1118
1119
1120 //------------------------------basic_plus_adr---------------------------------
basic_plus_adr(Node * base,Node * ptr,Node * offset)1121 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1122 // short-circuit a common case
1123 if (offset == intcon(0)) return ptr;
1124 return _gvn.transform( new AddPNode(base, ptr, offset) );
1125 }
1126
ConvI2L(Node * offset)1127 Node* GraphKit::ConvI2L(Node* offset) {
1128 // short-circuit a common case
1129 jint offset_con = find_int_con(offset, Type::OffsetBot);
1130 if (offset_con != Type::OffsetBot) {
1131 return longcon((jlong) offset_con);
1132 }
1133 return _gvn.transform( new ConvI2LNode(offset));
1134 }
1135
ConvI2UL(Node * offset)1136 Node* GraphKit::ConvI2UL(Node* offset) {
1137 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1138 if (offset_con != (juint) Type::OffsetBot) {
1139 return longcon((julong) offset_con);
1140 }
1141 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1142 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1143 return _gvn.transform( new AndLNode(conv, mask) );
1144 }
1145
ConvL2I(Node * offset)1146 Node* GraphKit::ConvL2I(Node* offset) {
1147 // short-circuit a common case
1148 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1149 if (offset_con != (jlong)Type::OffsetBot) {
1150 return intcon((int) offset_con);
1151 }
1152 return _gvn.transform( new ConvL2INode(offset));
1153 }
1154
1155 //-------------------------load_object_klass-----------------------------------
load_object_klass(Node * obj)1156 Node* GraphKit::load_object_klass(Node* obj) {
1157 // Special-case a fresh allocation to avoid building nodes:
1158 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1159 if (akls != NULL) return akls;
1160 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1161 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1162 }
1163
1164 //-------------------------load_array_length-----------------------------------
load_array_length(Node * array)1165 Node* GraphKit::load_array_length(Node* array) {
1166 // Special-case a fresh allocation to avoid building nodes:
1167 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1168 Node *alen;
1169 if (alloc == NULL) {
1170 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1171 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1172 } else {
1173 alen = alloc->Ideal_length();
1174 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1175 if (ccast != alen) {
1176 alen = _gvn.transform(ccast);
1177 }
1178 }
1179 return alen;
1180 }
1181
1182 //------------------------------do_null_check----------------------------------
1183 // Helper function to do a NULL pointer check. Returned value is
1184 // the incoming address with NULL casted away. You are allowed to use the
1185 // not-null value only if you are control dependent on the test.
1186 #ifndef PRODUCT
1187 extern int explicit_null_checks_inserted,
1188 explicit_null_checks_elided;
1189 #endif
null_check_common(Node * value,BasicType type,bool assert_null,Node ** null_control,bool speculative)1190 Node* GraphKit::null_check_common(Node* value, BasicType type,
1191 // optional arguments for variations:
1192 bool assert_null,
1193 Node* *null_control,
1194 bool speculative) {
1195 assert(!assert_null || null_control == NULL, "not both at once");
1196 if (stopped()) return top();
1197 NOT_PRODUCT(explicit_null_checks_inserted++);
1198
1199 // Construct NULL check
1200 Node *chk = NULL;
1201 switch(type) {
1202 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1203 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1204 case T_ARRAY : // fall through
1205 type = T_OBJECT; // simplify further tests
1206 case T_OBJECT : {
1207 const Type *t = _gvn.type( value );
1208
1209 const TypeOopPtr* tp = t->isa_oopptr();
1210 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1211 // Only for do_null_check, not any of its siblings:
1212 && !assert_null && null_control == NULL) {
1213 // Usually, any field access or invocation on an unloaded oop type
1214 // will simply fail to link, since the statically linked class is
1215 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1216 // the static class is loaded but the sharper oop type is not.
1217 // Rather than checking for this obscure case in lots of places,
1218 // we simply observe that a null check on an unloaded class
1219 // will always be followed by a nonsense operation, so we
1220 // can just issue the uncommon trap here.
1221 // Our access to the unloaded class will only be correct
1222 // after it has been loaded and initialized, which requires
1223 // a trip through the interpreter.
1224 #ifndef PRODUCT
1225 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1226 #endif
1227 uncommon_trap(Deoptimization::Reason_unloaded,
1228 Deoptimization::Action_reinterpret,
1229 tp->klass(), "!loaded");
1230 return top();
1231 }
1232
1233 if (assert_null) {
1234 // See if the type is contained in NULL_PTR.
1235 // If so, then the value is already null.
1236 if (t->higher_equal(TypePtr::NULL_PTR)) {
1237 NOT_PRODUCT(explicit_null_checks_elided++);
1238 return value; // Elided null assert quickly!
1239 }
1240 } else {
1241 // See if mixing in the NULL pointer changes type.
1242 // If so, then the NULL pointer was not allowed in the original
1243 // type. In other words, "value" was not-null.
1244 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1245 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1246 NOT_PRODUCT(explicit_null_checks_elided++);
1247 return value; // Elided null check quickly!
1248 }
1249 }
1250 chk = new CmpPNode( value, null() );
1251 break;
1252 }
1253
1254 default:
1255 fatal("unexpected type: %s", type2name(type));
1256 }
1257 assert(chk != NULL, "sanity check");
1258 chk = _gvn.transform(chk);
1259
1260 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1261 BoolNode *btst = new BoolNode( chk, btest);
1262 Node *tst = _gvn.transform( btst );
1263
1264 //-----------
1265 // if peephole optimizations occurred, a prior test existed.
1266 // If a prior test existed, maybe it dominates as we can avoid this test.
1267 if (tst != btst && type == T_OBJECT) {
1268 // At this point we want to scan up the CFG to see if we can
1269 // find an identical test (and so avoid this test altogether).
1270 Node *cfg = control();
1271 int depth = 0;
1272 while( depth < 16 ) { // Limit search depth for speed
1273 if( cfg->Opcode() == Op_IfTrue &&
1274 cfg->in(0)->in(1) == tst ) {
1275 // Found prior test. Use "cast_not_null" to construct an identical
1276 // CastPP (and hence hash to) as already exists for the prior test.
1277 // Return that casted value.
1278 if (assert_null) {
1279 replace_in_map(value, null());
1280 return null(); // do not issue the redundant test
1281 }
1282 Node *oldcontrol = control();
1283 set_control(cfg);
1284 Node *res = cast_not_null(value);
1285 set_control(oldcontrol);
1286 NOT_PRODUCT(explicit_null_checks_elided++);
1287 return res;
1288 }
1289 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1290 if (cfg == NULL) break; // Quit at region nodes
1291 depth++;
1292 }
1293 }
1294
1295 //-----------
1296 // Branch to failure if null
1297 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1298 Deoptimization::DeoptReason reason;
1299 if (assert_null) {
1300 reason = Deoptimization::reason_null_assert(speculative);
1301 } else if (type == T_OBJECT) {
1302 reason = Deoptimization::reason_null_check(speculative);
1303 } else {
1304 reason = Deoptimization::Reason_div0_check;
1305 }
1306 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1307 // ciMethodData::has_trap_at will return a conservative -1 if any
1308 // must-be-null assertion has failed. This could cause performance
1309 // problems for a method after its first do_null_assert failure.
1310 // Consider using 'Reason_class_check' instead?
1311
1312 // To cause an implicit null check, we set the not-null probability
1313 // to the maximum (PROB_MAX). For an explicit check the probability
1314 // is set to a smaller value.
1315 if (null_control != NULL || too_many_traps(reason)) {
1316 // probability is less likely
1317 ok_prob = PROB_LIKELY_MAG(3);
1318 } else if (!assert_null &&
1319 (ImplicitNullCheckThreshold > 0) &&
1320 method() != NULL &&
1321 (method()->method_data()->trap_count(reason)
1322 >= (uint)ImplicitNullCheckThreshold)) {
1323 ok_prob = PROB_LIKELY_MAG(3);
1324 }
1325
1326 if (null_control != NULL) {
1327 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1328 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1329 set_control( _gvn.transform( new IfTrueNode(iff)));
1330 #ifndef PRODUCT
1331 if (null_true == top()) {
1332 explicit_null_checks_elided++;
1333 }
1334 #endif
1335 (*null_control) = null_true;
1336 } else {
1337 BuildCutout unless(this, tst, ok_prob);
1338 // Check for optimizer eliding test at parse time
1339 if (stopped()) {
1340 // Failure not possible; do not bother making uncommon trap.
1341 NOT_PRODUCT(explicit_null_checks_elided++);
1342 } else if (assert_null) {
1343 uncommon_trap(reason,
1344 Deoptimization::Action_make_not_entrant,
1345 NULL, "assert_null");
1346 } else {
1347 replace_in_map(value, zerocon(type));
1348 builtin_throw(reason);
1349 }
1350 }
1351
1352 // Must throw exception, fall-thru not possible?
1353 if (stopped()) {
1354 return top(); // No result
1355 }
1356
1357 if (assert_null) {
1358 // Cast obj to null on this path.
1359 replace_in_map(value, zerocon(type));
1360 return zerocon(type);
1361 }
1362
1363 // Cast obj to not-null on this path, if there is no null_control.
1364 // (If there is a null_control, a non-null value may come back to haunt us.)
1365 if (type == T_OBJECT) {
1366 Node* cast = cast_not_null(value, false);
1367 if (null_control == NULL || (*null_control) == top())
1368 replace_in_map(value, cast);
1369 value = cast;
1370 }
1371
1372 return value;
1373 }
1374
1375
1376 //------------------------------cast_not_null----------------------------------
1377 // Cast obj to not-null on this path
cast_not_null(Node * obj,bool do_replace_in_map)1378 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1379 const Type *t = _gvn.type(obj);
1380 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1381 // Object is already not-null?
1382 if( t == t_not_null ) return obj;
1383
1384 Node *cast = new CastPPNode(obj,t_not_null);
1385 cast->init_req(0, control());
1386 cast = _gvn.transform( cast );
1387
1388 // Scan for instances of 'obj' in the current JVM mapping.
1389 // These instances are known to be not-null after the test.
1390 if (do_replace_in_map)
1391 replace_in_map(obj, cast);
1392
1393 return cast; // Return casted value
1394 }
1395
1396 // Sometimes in intrinsics, we implicitly know an object is not null
1397 // (there's no actual null check) so we can cast it to not null. In
1398 // the course of optimizations, the input to the cast can become null.
1399 // In that case that data path will die and we need the control path
1400 // to become dead as well to keep the graph consistent. So we have to
1401 // add a check for null for which one branch can't be taken. It uses
1402 // an Opaque4 node that will cause the check to be removed after loop
1403 // opts so the test goes away and the compiled code doesn't execute a
1404 // useless check.
must_be_not_null(Node * value,bool do_replace_in_map)1405 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1406 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1407 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1408 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1409 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1410 _gvn.set_type(iff, iff->Value(&_gvn));
1411 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1412 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1413 Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1414 C->root()->add_req(halt);
1415 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1416 set_control(if_t);
1417 return cast_not_null(value, do_replace_in_map);
1418 }
1419
1420
1421 //--------------------------replace_in_map-------------------------------------
replace_in_map(Node * old,Node * neww)1422 void GraphKit::replace_in_map(Node* old, Node* neww) {
1423 if (old == neww) {
1424 return;
1425 }
1426
1427 map()->replace_edge(old, neww);
1428
1429 // Note: This operation potentially replaces any edge
1430 // on the map. This includes locals, stack, and monitors
1431 // of the current (innermost) JVM state.
1432
1433 // don't let inconsistent types from profiling escape this
1434 // method
1435
1436 const Type* told = _gvn.type(old);
1437 const Type* tnew = _gvn.type(neww);
1438
1439 if (!tnew->higher_equal(told)) {
1440 return;
1441 }
1442
1443 map()->record_replaced_node(old, neww);
1444 }
1445
1446
1447 //=============================================================================
1448 //--------------------------------memory---------------------------------------
memory(uint alias_idx)1449 Node* GraphKit::memory(uint alias_idx) {
1450 MergeMemNode* mem = merged_memory();
1451 Node* p = mem->memory_at(alias_idx);
1452 _gvn.set_type(p, Type::MEMORY); // must be mapped
1453 return p;
1454 }
1455
1456 //-----------------------------reset_memory------------------------------------
reset_memory()1457 Node* GraphKit::reset_memory() {
1458 Node* mem = map()->memory();
1459 // do not use this node for any more parsing!
1460 debug_only( map()->set_memory((Node*)NULL) );
1461 return _gvn.transform( mem );
1462 }
1463
1464 //------------------------------set_all_memory---------------------------------
set_all_memory(Node * newmem)1465 void GraphKit::set_all_memory(Node* newmem) {
1466 Node* mergemem = MergeMemNode::make(newmem);
1467 gvn().set_type_bottom(mergemem);
1468 map()->set_memory(mergemem);
1469 }
1470
1471 //------------------------------set_all_memory_call----------------------------
set_all_memory_call(Node * call,bool separate_io_proj)1472 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1473 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1474 set_all_memory(newmem);
1475 }
1476
1477 //=============================================================================
1478 //
1479 // parser factory methods for MemNodes
1480 //
1481 // These are layered on top of the factory methods in LoadNode and StoreNode,
1482 // and integrate with the parser's memory state and _gvn engine.
1483 //
1484
1485 // 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)1486 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1487 int adr_idx,
1488 MemNode::MemOrd mo,
1489 LoadNode::ControlDependency control_dependency,
1490 bool require_atomic_access,
1491 bool unaligned,
1492 bool mismatched) {
1493 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1494 const TypePtr* adr_type = NULL; // debug-mode-only argument
1495 debug_only(adr_type = C->get_adr_type(adr_idx));
1496 Node* mem = memory(adr_idx);
1497 Node* ld;
1498 if (require_atomic_access && bt == T_LONG) {
1499 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1500 } else if (require_atomic_access && bt == T_DOUBLE) {
1501 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1502 } else {
1503 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1504 }
1505 ld = _gvn.transform(ld);
1506 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1507 // Improve graph before escape analysis and boxing elimination.
1508 record_for_igvn(ld);
1509 }
1510 return ld;
1511 }
1512
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)1513 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1514 int adr_idx,
1515 MemNode::MemOrd mo,
1516 bool require_atomic_access,
1517 bool unaligned,
1518 bool mismatched) {
1519 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1520 const TypePtr* adr_type = NULL;
1521 debug_only(adr_type = C->get_adr_type(adr_idx));
1522 Node *mem = memory(adr_idx);
1523 Node* st;
1524 if (require_atomic_access && bt == T_LONG) {
1525 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1526 } else if (require_atomic_access && bt == T_DOUBLE) {
1527 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1528 } else {
1529 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1530 }
1531 if (unaligned) {
1532 st->as_Store()->set_unaligned_access();
1533 }
1534 if (mismatched) {
1535 st->as_Store()->set_mismatched_access();
1536 }
1537 st = _gvn.transform(st);
1538 set_memory(st, adr_idx);
1539 // Back-to-back stores can only remove intermediate store with DU info
1540 // so push on worklist for optimizer.
1541 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1542 record_for_igvn(st);
1543
1544 return st;
1545 }
1546
access_store_at(Node * obj,Node * adr,const TypePtr * adr_type,Node * val,const Type * val_type,BasicType bt,DecoratorSet decorators)1547 Node* GraphKit::access_store_at(Node* obj,
1548 Node* adr,
1549 const TypePtr* adr_type,
1550 Node* val,
1551 const Type* val_type,
1552 BasicType bt,
1553 DecoratorSet decorators) {
1554 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1555 // could be delayed during Parse (for example, in adjust_map_after_if()).
1556 // Execute transformation here to avoid barrier generation in such case.
1557 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1558 val = _gvn.makecon(TypePtr::NULL_PTR);
1559 }
1560
1561 if (stopped()) {
1562 return top(); // Dead path ?
1563 }
1564
1565 assert(val != NULL, "not dead path");
1566
1567 C2AccessValuePtr addr(adr, adr_type);
1568 C2AccessValue value(val, val_type);
1569 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1570 if (access.is_raw()) {
1571 return _barrier_set->BarrierSetC2::store_at(access, value);
1572 } else {
1573 return _barrier_set->store_at(access, value);
1574 }
1575 }
1576
access_load_at(Node * obj,Node * adr,const TypePtr * adr_type,const Type * val_type,BasicType bt,DecoratorSet decorators)1577 Node* GraphKit::access_load_at(Node* obj, // containing obj
1578 Node* adr, // actual adress to store val at
1579 const TypePtr* adr_type,
1580 const Type* val_type,
1581 BasicType bt,
1582 DecoratorSet decorators) {
1583 if (stopped()) {
1584 return top(); // Dead path ?
1585 }
1586
1587 C2AccessValuePtr addr(adr, adr_type);
1588 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1589 if (access.is_raw()) {
1590 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1591 } else {
1592 return _barrier_set->load_at(access, val_type);
1593 }
1594 }
1595
access_load(Node * adr,const Type * val_type,BasicType bt,DecoratorSet decorators)1596 Node* GraphKit::access_load(Node* adr, // actual adress to load val at
1597 const Type* val_type,
1598 BasicType bt,
1599 DecoratorSet decorators) {
1600 if (stopped()) {
1601 return top(); // Dead path ?
1602 }
1603
1604 C2AccessValuePtr addr(adr, NULL);
1605 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1606 if (access.is_raw()) {
1607 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1608 } else {
1609 return _barrier_set->load_at(access, val_type);
1610 }
1611 }
1612
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)1613 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1614 Node* adr,
1615 const TypePtr* adr_type,
1616 int alias_idx,
1617 Node* expected_val,
1618 Node* new_val,
1619 const Type* value_type,
1620 BasicType bt,
1621 DecoratorSet decorators) {
1622 C2AccessValuePtr addr(adr, adr_type);
1623 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1624 bt, obj, addr, alias_idx);
1625 if (access.is_raw()) {
1626 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1627 } else {
1628 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1629 }
1630 }
1631
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)1632 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1633 Node* adr,
1634 const TypePtr* adr_type,
1635 int alias_idx,
1636 Node* expected_val,
1637 Node* new_val,
1638 const Type* value_type,
1639 BasicType bt,
1640 DecoratorSet decorators) {
1641 C2AccessValuePtr addr(adr, adr_type);
1642 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1643 bt, obj, addr, alias_idx);
1644 if (access.is_raw()) {
1645 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1646 } else {
1647 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1648 }
1649 }
1650
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)1651 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1652 Node* adr,
1653 const TypePtr* adr_type,
1654 int alias_idx,
1655 Node* new_val,
1656 const Type* value_type,
1657 BasicType bt,
1658 DecoratorSet decorators) {
1659 C2AccessValuePtr addr(adr, adr_type);
1660 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1661 bt, obj, addr, alias_idx);
1662 if (access.is_raw()) {
1663 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1664 } else {
1665 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1666 }
1667 }
1668
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)1669 Node* GraphKit::access_atomic_add_at(Node* obj,
1670 Node* adr,
1671 const TypePtr* adr_type,
1672 int alias_idx,
1673 Node* new_val,
1674 const Type* value_type,
1675 BasicType bt,
1676 DecoratorSet decorators) {
1677 C2AccessValuePtr addr(adr, adr_type);
1678 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1679 if (access.is_raw()) {
1680 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1681 } else {
1682 return _barrier_set->atomic_add_at(access, new_val, value_type);
1683 }
1684 }
1685
access_clone(Node * src,Node * dst,Node * size,bool is_array)1686 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1687 return _barrier_set->clone(this, src, dst, size, is_array);
1688 }
1689
access_resolve(Node * n,DecoratorSet decorators)1690 Node* GraphKit::access_resolve(Node* n, DecoratorSet decorators) {
1691 // Use stronger ACCESS_WRITE|ACCESS_READ by default.
1692 if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) {
1693 decorators |= ACCESS_READ | ACCESS_WRITE;
1694 }
1695 return _barrier_set->resolve(this, n, decorators);
1696 }
1697
1698 //-------------------------array_element_address-------------------------
array_element_address(Node * ary,Node * idx,BasicType elembt,const TypeInt * sizetype,Node * ctrl)1699 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1700 const TypeInt* sizetype, Node* ctrl) {
1701 uint shift = exact_log2(type2aelembytes(elembt));
1702 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1703
1704 // short-circuit a common case (saves lots of confusing waste motion)
1705 jint idx_con = find_int_con(idx, -1);
1706 if (idx_con >= 0) {
1707 intptr_t offset = header + ((intptr_t)idx_con << shift);
1708 return basic_plus_adr(ary, offset);
1709 }
1710
1711 // must be correct type for alignment purposes
1712 Node* base = basic_plus_adr(ary, header);
1713 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1714 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1715 return basic_plus_adr(ary, base, scale);
1716 }
1717
1718 //-------------------------load_array_element-------------------------
load_array_element(Node * ctl,Node * ary,Node * idx,const TypeAryPtr * arytype)1719 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1720 const Type* elemtype = arytype->elem();
1721 BasicType elembt = elemtype->array_element_basic_type();
1722 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1723 if (elembt == T_NARROWOOP) {
1724 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1725 }
1726 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1727 return ld;
1728 }
1729
1730 //-------------------------set_arguments_for_java_call-------------------------
1731 // Arguments (pre-popped from the stack) are taken from the JVMS.
set_arguments_for_java_call(CallJavaNode * call)1732 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1733 // Add the call arguments:
1734 uint nargs = call->method()->arg_size();
1735 for (uint i = 0; i < nargs; i++) {
1736 Node* arg = argument(i);
1737 call->init_req(i + TypeFunc::Parms, arg);
1738 }
1739 }
1740
1741 //---------------------------set_edges_for_java_call---------------------------
1742 // Connect a newly created call into the current JVMS.
1743 // 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)1744 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1745
1746 // Add the predefined inputs:
1747 call->init_req( TypeFunc::Control, control() );
1748 call->init_req( TypeFunc::I_O , i_o() );
1749 call->init_req( TypeFunc::Memory , reset_memory() );
1750 call->init_req( TypeFunc::FramePtr, frameptr() );
1751 call->init_req( TypeFunc::ReturnAdr, top() );
1752
1753 add_safepoint_edges(call, must_throw);
1754
1755 Node* xcall = _gvn.transform(call);
1756
1757 if (xcall == top()) {
1758 set_control(top());
1759 return;
1760 }
1761 assert(xcall == call, "call identity is stable");
1762
1763 // Re-use the current map to produce the result.
1764
1765 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1766 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1767 set_all_memory_call(xcall, separate_io_proj);
1768
1769 //return xcall; // no need, caller already has it
1770 }
1771
set_results_for_java_call(CallJavaNode * call,bool separate_io_proj,bool deoptimize)1772 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1773 if (stopped()) return top(); // maybe the call folded up?
1774
1775 // Capture the return value, if any.
1776 Node* ret;
1777 if (call->method() == NULL ||
1778 call->method()->return_type()->basic_type() == T_VOID)
1779 ret = top();
1780 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1781
1782 // Note: Since any out-of-line call can produce an exception,
1783 // we always insert an I_O projection from the call into the result.
1784
1785 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1786
1787 if (separate_io_proj) {
1788 // The caller requested separate projections be used by the fall
1789 // through and exceptional paths, so replace the projections for
1790 // the fall through path.
1791 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1792 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1793 }
1794 return ret;
1795 }
1796
1797 //--------------------set_predefined_input_for_runtime_call--------------------
1798 // Reading and setting the memory state is way conservative here.
1799 // The real problem is that I am not doing real Type analysis on memory,
1800 // so I cannot distinguish card mark stores from other stores. Across a GC
1801 // point the Store Barrier and the card mark memory has to agree. I cannot
1802 // have a card mark store and its barrier split across the GC point from
1803 // either above or below. Here I get that to happen by reading ALL of memory.
1804 // A better answer would be to separate out card marks from other memory.
1805 // For now, return the input memory state, so that it can be reused
1806 // after the call, if this call has restricted memory effects.
set_predefined_input_for_runtime_call(SafePointNode * call,Node * narrow_mem)1807 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1808 // Set fixed predefined input arguments
1809 Node* memory = reset_memory();
1810 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1811 call->init_req( TypeFunc::Control, control() );
1812 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1813 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1814 call->init_req( TypeFunc::FramePtr, frameptr() );
1815 call->init_req( TypeFunc::ReturnAdr, top() );
1816 return memory;
1817 }
1818
1819 //-------------------set_predefined_output_for_runtime_call--------------------
1820 // Set control and memory (not i_o) from the call.
1821 // If keep_mem is not NULL, use it for the output state,
1822 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1823 // If hook_mem is NULL, this call produces no memory effects at all.
1824 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1825 // then only that memory slice is taken from the call.
1826 // In the last case, we must put an appropriate memory barrier before
1827 // the call, so as to create the correct anti-dependencies on loads
1828 // preceding the call.
set_predefined_output_for_runtime_call(Node * call,Node * keep_mem,const TypePtr * hook_mem)1829 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1830 Node* keep_mem,
1831 const TypePtr* hook_mem) {
1832 // no i/o
1833 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1834 if (keep_mem) {
1835 // First clone the existing memory state
1836 set_all_memory(keep_mem);
1837 if (hook_mem != NULL) {
1838 // Make memory for the call
1839 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1840 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1841 // We also use hook_mem to extract specific effects from arraycopy stubs.
1842 set_memory(mem, hook_mem);
1843 }
1844 // ...else the call has NO memory effects.
1845
1846 // Make sure the call advertises its memory effects precisely.
1847 // This lets us build accurate anti-dependences in gcm.cpp.
1848 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1849 "call node must be constructed correctly");
1850 } else {
1851 assert(hook_mem == NULL, "");
1852 // This is not a "slow path" call; all memory comes from the call.
1853 set_all_memory_call(call);
1854 }
1855 }
1856
1857
1858 // Replace the call with the current state of the kit.
replace_call(CallNode * call,Node * result,bool do_replaced_nodes)1859 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1860 JVMState* ejvms = NULL;
1861 if (has_exceptions()) {
1862 ejvms = transfer_exceptions_into_jvms();
1863 }
1864
1865 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1866 ReplacedNodes replaced_nodes_exception;
1867 Node* ex_ctl = top();
1868
1869 SafePointNode* final_state = stop();
1870
1871 // Find all the needed outputs of this call
1872 CallProjections callprojs;
1873 call->extract_projections(&callprojs, true);
1874
1875 Node* init_mem = call->in(TypeFunc::Memory);
1876 Node* final_mem = final_state->in(TypeFunc::Memory);
1877 Node* final_ctl = final_state->in(TypeFunc::Control);
1878 Node* final_io = final_state->in(TypeFunc::I_O);
1879
1880 // Replace all the old call edges with the edges from the inlining result
1881 if (callprojs.fallthrough_catchproj != NULL) {
1882 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1883 }
1884 if (callprojs.fallthrough_memproj != NULL) {
1885 if (final_mem->is_MergeMem()) {
1886 // Parser's exits MergeMem was not transformed but may be optimized
1887 final_mem = _gvn.transform(final_mem);
1888 }
1889 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1890 }
1891 if (callprojs.fallthrough_ioproj != NULL) {
1892 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1893 }
1894
1895 // Replace the result with the new result if it exists and is used
1896 if (callprojs.resproj != NULL && result != NULL) {
1897 C->gvn_replace_by(callprojs.resproj, result);
1898 }
1899
1900 if (ejvms == NULL) {
1901 // No exception edges to simply kill off those paths
1902 if (callprojs.catchall_catchproj != NULL) {
1903 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1904 }
1905 if (callprojs.catchall_memproj != NULL) {
1906 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1907 }
1908 if (callprojs.catchall_ioproj != NULL) {
1909 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1910 }
1911 // Replace the old exception object with top
1912 if (callprojs.exobj != NULL) {
1913 C->gvn_replace_by(callprojs.exobj, C->top());
1914 }
1915 } else {
1916 GraphKit ekit(ejvms);
1917
1918 // Load my combined exception state into the kit, with all phis transformed:
1919 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1920 replaced_nodes_exception = ex_map->replaced_nodes();
1921
1922 Node* ex_oop = ekit.use_exception_state(ex_map);
1923
1924 if (callprojs.catchall_catchproj != NULL) {
1925 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1926 ex_ctl = ekit.control();
1927 }
1928 if (callprojs.catchall_memproj != NULL) {
1929 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1930 }
1931 if (callprojs.catchall_ioproj != NULL) {
1932 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1933 }
1934
1935 // Replace the old exception object with the newly created one
1936 if (callprojs.exobj != NULL) {
1937 C->gvn_replace_by(callprojs.exobj, ex_oop);
1938 }
1939 }
1940
1941 // Disconnect the call from the graph
1942 call->disconnect_inputs(NULL, C);
1943 C->gvn_replace_by(call, C->top());
1944
1945 // Clean up any MergeMems that feed other MergeMems since the
1946 // optimizer doesn't like that.
1947 if (final_mem->is_MergeMem()) {
1948 Node_List wl;
1949 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1950 Node* m = i.get();
1951 if (m->is_MergeMem() && !wl.contains(m)) {
1952 wl.push(m);
1953 }
1954 }
1955 while (wl.size() > 0) {
1956 _gvn.transform(wl.pop());
1957 }
1958 }
1959
1960 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1961 replaced_nodes.apply(C, final_ctl);
1962 }
1963 if (!ex_ctl->is_top() && do_replaced_nodes) {
1964 replaced_nodes_exception.apply(C, ex_ctl);
1965 }
1966 }
1967
1968
1969 //------------------------------increment_counter------------------------------
1970 // for statistics: increment a VM counter by 1
1971
increment_counter(address counter_addr)1972 void GraphKit::increment_counter(address counter_addr) {
1973 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1974 increment_counter(adr1);
1975 }
1976
increment_counter(Node * counter_addr)1977 void GraphKit::increment_counter(Node* counter_addr) {
1978 int adr_type = Compile::AliasIdxRaw;
1979 Node* ctrl = control();
1980 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1981 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1982 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1983 }
1984
1985
1986 //------------------------------uncommon_trap----------------------------------
1987 // Bail out to the interpreter in mid-method. Implemented by calling the
1988 // uncommon_trap blob. This helper function inserts a runtime call with the
1989 // right debug info.
uncommon_trap(int trap_request,ciKlass * klass,const char * comment,bool must_throw,bool keep_exact_action)1990 void GraphKit::uncommon_trap(int trap_request,
1991 ciKlass* klass, const char* comment,
1992 bool must_throw,
1993 bool keep_exact_action) {
1994 if (failing()) stop();
1995 if (stopped()) return; // trap reachable?
1996
1997 // Note: If ProfileTraps is true, and if a deopt. actually
1998 // occurs here, the runtime will make sure an MDO exists. There is
1999 // no need to call method()->ensure_method_data() at this point.
2000
2001 // Set the stack pointer to the right value for reexecution:
2002 set_sp(reexecute_sp());
2003
2004 #ifdef ASSERT
2005 if (!must_throw) {
2006 // Make sure the stack has at least enough depth to execute
2007 // the current bytecode.
2008 int inputs, ignored_depth;
2009 if (compute_stack_effects(inputs, ignored_depth)) {
2010 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2011 Bytecodes::name(java_bc()), sp(), inputs);
2012 }
2013 }
2014 #endif
2015
2016 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2017 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2018
2019 switch (action) {
2020 case Deoptimization::Action_maybe_recompile:
2021 case Deoptimization::Action_reinterpret:
2022 // Temporary fix for 6529811 to allow virtual calls to be sure they
2023 // get the chance to go from mono->bi->mega
2024 if (!keep_exact_action &&
2025 Deoptimization::trap_request_index(trap_request) < 0 &&
2026 too_many_recompiles(reason)) {
2027 // This BCI is causing too many recompilations.
2028 if (C->log() != NULL) {
2029 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2030 Deoptimization::trap_reason_name(reason),
2031 Deoptimization::trap_action_name(action));
2032 }
2033 action = Deoptimization::Action_none;
2034 trap_request = Deoptimization::make_trap_request(reason, action);
2035 } else {
2036 C->set_trap_can_recompile(true);
2037 }
2038 break;
2039 case Deoptimization::Action_make_not_entrant:
2040 C->set_trap_can_recompile(true);
2041 break;
2042 case Deoptimization::Action_none:
2043 case Deoptimization::Action_make_not_compilable:
2044 break;
2045 default:
2046 #ifdef ASSERT
2047 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2048 #endif
2049 break;
2050 }
2051
2052 if (TraceOptoParse) {
2053 char buf[100];
2054 tty->print_cr("Uncommon trap %s at bci:%d",
2055 Deoptimization::format_trap_request(buf, sizeof(buf),
2056 trap_request), bci());
2057 }
2058
2059 CompileLog* log = C->log();
2060 if (log != NULL) {
2061 int kid = (klass == NULL)? -1: log->identify(klass);
2062 log->begin_elem("uncommon_trap bci='%d'", bci());
2063 char buf[100];
2064 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2065 trap_request));
2066 if (kid >= 0) log->print(" klass='%d'", kid);
2067 if (comment != NULL) log->print(" comment='%s'", comment);
2068 log->end_elem();
2069 }
2070
2071 // Make sure any guarding test views this path as very unlikely
2072 Node *i0 = control()->in(0);
2073 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2074 IfNode *iff = i0->as_If();
2075 float f = iff->_prob; // Get prob
2076 if (control()->Opcode() == Op_IfTrue) {
2077 if (f > PROB_UNLIKELY_MAG(4))
2078 iff->_prob = PROB_MIN;
2079 } else {
2080 if (f < PROB_LIKELY_MAG(4))
2081 iff->_prob = PROB_MAX;
2082 }
2083 }
2084
2085 // Clear out dead values from the debug info.
2086 kill_dead_locals();
2087
2088 // Now insert the uncommon trap subroutine call
2089 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2090 const TypePtr* no_memory_effects = NULL;
2091 // Pass the index of the class to be loaded
2092 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2093 (must_throw ? RC_MUST_THROW : 0),
2094 OptoRuntime::uncommon_trap_Type(),
2095 call_addr, "uncommon_trap", no_memory_effects,
2096 intcon(trap_request));
2097 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2098 "must extract request correctly from the graph");
2099 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2100
2101 call->set_req(TypeFunc::ReturnAdr, returnadr());
2102 // The debug info is the only real input to this call.
2103
2104 // Halt-and-catch fire here. The above call should never return!
2105 HaltNode* halt = new HaltNode(control(), frameptr());
2106 _gvn.set_type_bottom(halt);
2107 root()->add_req(halt);
2108
2109 stop_and_kill_map();
2110 }
2111
2112
2113 //--------------------------just_allocated_object------------------------------
2114 // Report the object that was just allocated.
2115 // It must be the case that there are no intervening safepoints.
2116 // We use this to determine if an object is so "fresh" that
2117 // it does not require card marks.
just_allocated_object(Node * current_control)2118 Node* GraphKit::just_allocated_object(Node* current_control) {
2119 Node* ctrl = current_control;
2120 // Object::<init> is invoked after allocation, most of invoke nodes
2121 // will be reduced, but a region node is kept in parse time, we check
2122 // the pattern and skip the region node if it degraded to a copy.
2123 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 &&
2124 ctrl->as_Region()->is_copy()) {
2125 ctrl = ctrl->as_Region()->is_copy();
2126 }
2127 if (C->recent_alloc_ctl() == ctrl) {
2128 return C->recent_alloc_obj();
2129 }
2130 return NULL;
2131 }
2132
2133
round_double_arguments(ciMethod * dest_method)2134 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2135 // (Note: TypeFunc::make has a cache that makes this fast.)
2136 const TypeFunc* tf = TypeFunc::make(dest_method);
2137 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2138 for (int j = 0; j < nargs; j++) {
2139 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2140 if( targ->basic_type() == T_DOUBLE ) {
2141 // If any parameters are doubles, they must be rounded before
2142 // the call, dstore_rounding does gvn.transform
2143 Node *arg = argument(j);
2144 arg = dstore_rounding(arg);
2145 set_argument(j, arg);
2146 }
2147 }
2148 }
2149
2150 /**
2151 * Record profiling data exact_kls for Node n with the type system so
2152 * that it can propagate it (speculation)
2153 *
2154 * @param n node that the type applies to
2155 * @param exact_kls type from profiling
2156 * @param maybe_null did profiling see null?
2157 *
2158 * @return node with improved type
2159 */
record_profile_for_speculation(Node * n,ciKlass * exact_kls,ProfilePtrKind ptr_kind)2160 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2161 const Type* current_type = _gvn.type(n);
2162 assert(UseTypeSpeculation, "type speculation must be on");
2163
2164 const TypePtr* speculative = current_type->speculative();
2165
2166 // Should the klass from the profile be recorded in the speculative type?
2167 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2168 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2169 const TypeOopPtr* xtype = tklass->as_instance_type();
2170 assert(xtype->klass_is_exact(), "Should be exact");
2171 // Any reason to believe n is not null (from this profiling or a previous one)?
2172 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2173 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2174 // record the new speculative type's depth
2175 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2176 speculative = speculative->with_inline_depth(jvms()->depth());
2177 } else if (current_type->would_improve_ptr(ptr_kind)) {
2178 // Profiling report that null was never seen so we can change the
2179 // speculative type to non null ptr.
2180 if (ptr_kind == ProfileAlwaysNull) {
2181 speculative = TypePtr::NULL_PTR;
2182 } else {
2183 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2184 const TypePtr* ptr = TypePtr::NOTNULL;
2185 if (speculative != NULL) {
2186 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2187 } else {
2188 speculative = ptr;
2189 }
2190 }
2191 }
2192
2193 if (speculative != current_type->speculative()) {
2194 // Build a type with a speculative type (what we think we know
2195 // about the type but will need a guard when we use it)
2196 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2197 // We're changing the type, we need a new CheckCast node to carry
2198 // the new type. The new type depends on the control: what
2199 // profiling tells us is only valid from here as far as we can
2200 // tell.
2201 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2202 cast = _gvn.transform(cast);
2203 replace_in_map(n, cast);
2204 n = cast;
2205 }
2206
2207 return n;
2208 }
2209
2210 /**
2211 * Record profiling data from receiver profiling at an invoke with the
2212 * type system so that it can propagate it (speculation)
2213 *
2214 * @param n receiver node
2215 *
2216 * @return node with improved type
2217 */
record_profiled_receiver_for_speculation(Node * n)2218 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2219 if (!UseTypeSpeculation) {
2220 return n;
2221 }
2222 ciKlass* exact_kls = profile_has_unique_klass();
2223 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2224 if ((java_bc() == Bytecodes::_checkcast ||
2225 java_bc() == Bytecodes::_instanceof ||
2226 java_bc() == Bytecodes::_aastore) &&
2227 method()->method_data()->is_mature()) {
2228 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2229 if (data != NULL) {
2230 if (!data->as_BitData()->null_seen()) {
2231 ptr_kind = ProfileNeverNull;
2232 } else {
2233 assert(data->is_ReceiverTypeData(), "bad profile data type");
2234 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2235 uint i = 0;
2236 for (; i < call->row_limit(); i++) {
2237 ciKlass* receiver = call->receiver(i);
2238 if (receiver != NULL) {
2239 break;
2240 }
2241 }
2242 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2243 }
2244 }
2245 }
2246 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2247 }
2248
2249 /**
2250 * Record profiling data from argument profiling at an invoke with the
2251 * type system so that it can propagate it (speculation)
2252 *
2253 * @param dest_method target method for the call
2254 * @param bc what invoke bytecode is this?
2255 */
record_profiled_arguments_for_speculation(ciMethod * dest_method,Bytecodes::Code bc)2256 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2257 if (!UseTypeSpeculation) {
2258 return;
2259 }
2260 const TypeFunc* tf = TypeFunc::make(dest_method);
2261 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2262 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2263 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2264 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2265 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2266 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2267 ciKlass* better_type = NULL;
2268 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2269 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2270 }
2271 i++;
2272 }
2273 }
2274 }
2275
2276 /**
2277 * Record profiling data from parameter profiling at an invoke with
2278 * the type system so that it can propagate it (speculation)
2279 */
record_profiled_parameters_for_speculation()2280 void GraphKit::record_profiled_parameters_for_speculation() {
2281 if (!UseTypeSpeculation) {
2282 return;
2283 }
2284 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2285 if (_gvn.type(local(i))->isa_oopptr()) {
2286 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2287 ciKlass* better_type = NULL;
2288 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2289 record_profile_for_speculation(local(i), better_type, ptr_kind);
2290 }
2291 j++;
2292 }
2293 }
2294 }
2295
2296 /**
2297 * Record profiling data from return value profiling at an invoke with
2298 * the type system so that it can propagate it (speculation)
2299 */
record_profiled_return_for_speculation()2300 void GraphKit::record_profiled_return_for_speculation() {
2301 if (!UseTypeSpeculation) {
2302 return;
2303 }
2304 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2305 ciKlass* better_type = NULL;
2306 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2307 // If profiling reports a single type for the return value,
2308 // feed it to the type system so it can propagate it as a
2309 // speculative type
2310 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2311 }
2312 }
2313
round_double_result(ciMethod * dest_method)2314 void GraphKit::round_double_result(ciMethod* dest_method) {
2315 // A non-strict method may return a double value which has an extended
2316 // exponent, but this must not be visible in a caller which is 'strict'
2317 // If a strict caller invokes a non-strict callee, round a double result
2318
2319 BasicType result_type = dest_method->return_type()->basic_type();
2320 assert( method() != NULL, "must have caller context");
2321 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2322 // Destination method's return value is on top of stack
2323 // dstore_rounding() does gvn.transform
2324 Node *result = pop_pair();
2325 result = dstore_rounding(result);
2326 push_pair(result);
2327 }
2328 }
2329
2330 // rounding for strict float precision conformance
precision_rounding(Node * n)2331 Node* GraphKit::precision_rounding(Node* n) {
2332 return UseStrictFP && _method->flags().is_strict()
2333 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2334 ? _gvn.transform( new RoundFloatNode(0, n) )
2335 : n;
2336 }
2337
2338 // rounding for strict double precision conformance
dprecision_rounding(Node * n)2339 Node* GraphKit::dprecision_rounding(Node *n) {
2340 return UseStrictFP && _method->flags().is_strict()
2341 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2342 ? _gvn.transform( new RoundDoubleNode(0, n) )
2343 : n;
2344 }
2345
2346 // rounding for non-strict double stores
dstore_rounding(Node * n)2347 Node* GraphKit::dstore_rounding(Node* n) {
2348 return Matcher::strict_fp_requires_explicit_rounding
2349 && UseSSE <= 1
2350 ? _gvn.transform( new RoundDoubleNode(0, n) )
2351 : n;
2352 }
2353
2354 //=============================================================================
2355 // Generate a fast path/slow path idiom. Graph looks like:
2356 // [foo] indicates that 'foo' is a parameter
2357 //
2358 // [in] NULL
2359 // \ /
2360 // CmpP
2361 // Bool ne
2362 // If
2363 // / \
2364 // True False-<2>
2365 // / |
2366 // / cast_not_null
2367 // Load | | ^
2368 // [fast_test] | |
2369 // gvn to opt_test | |
2370 // / \ | <1>
2371 // True False |
2372 // | \\ |
2373 // [slow_call] \[fast_result]
2374 // Ctl Val \ \
2375 // | \ \
2376 // Catch <1> \ \
2377 // / \ ^ \ \
2378 // Ex No_Ex | \ \
2379 // | \ \ | \ <2> \
2380 // ... \ [slow_res] | | \ [null_result]
2381 // \ \--+--+--- | |
2382 // \ | / \ | /
2383 // --------Region Phi
2384 //
2385 //=============================================================================
2386 // Code is structured as a series of driver functions all called 'do_XXX' that
2387 // call a set of helper functions. Helper functions first, then drivers.
2388
2389 //------------------------------null_check_oop---------------------------------
2390 // Null check oop. Set null-path control into Region in slot 3.
2391 // 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)2392 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2393 bool never_see_null,
2394 bool safe_for_replace,
2395 bool speculative) {
2396 // Initial NULL check taken path
2397 (*null_control) = top();
2398 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2399
2400 // Generate uncommon_trap:
2401 if (never_see_null && (*null_control) != top()) {
2402 // If we see an unexpected null at a check-cast we record it and force a
2403 // recompile; the offending check-cast will be compiled to handle NULLs.
2404 // If we see more than one offending BCI, then all checkcasts in the
2405 // method will be compiled to handle NULLs.
2406 PreserveJVMState pjvms(this);
2407 set_control(*null_control);
2408 replace_in_map(value, null());
2409 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2410 uncommon_trap(reason,
2411 Deoptimization::Action_make_not_entrant);
2412 (*null_control) = top(); // NULL path is dead
2413 }
2414 if ((*null_control) == top() && safe_for_replace) {
2415 replace_in_map(value, cast);
2416 }
2417
2418 // Cast away null-ness on the result
2419 return cast;
2420 }
2421
2422 //------------------------------opt_iff----------------------------------------
2423 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2424 // Return slow-path control.
opt_iff(Node * region,Node * iff)2425 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2426 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2427
2428 // Fast path taken; set region slot 2
2429 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2430 region->init_req(2,fast_taken); // Capture fast-control
2431
2432 // Fast path not-taken, i.e. slow path
2433 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2434 return slow_taken;
2435 }
2436
2437 //-----------------------------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)2438 Node* GraphKit::make_runtime_call(int flags,
2439 const TypeFunc* call_type, address call_addr,
2440 const char* call_name,
2441 const TypePtr* adr_type,
2442 // The following parms are all optional.
2443 // The first NULL ends the list.
2444 Node* parm0, Node* parm1,
2445 Node* parm2, Node* parm3,
2446 Node* parm4, Node* parm5,
2447 Node* parm6, Node* parm7) {
2448 assert(call_addr != NULL, "must not call NULL targets");
2449
2450 // Slow-path call
2451 bool is_leaf = !(flags & RC_NO_LEAF);
2452 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2453 if (call_name == NULL) {
2454 assert(!is_leaf, "must supply name for leaf");
2455 call_name = OptoRuntime::stub_name(call_addr);
2456 }
2457 CallNode* call;
2458 if (!is_leaf) {
2459 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2460 bci(), adr_type);
2461 } else if (flags & RC_NO_FP) {
2462 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2463 } else {
2464 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2465 }
2466
2467 // The following is similar to set_edges_for_java_call,
2468 // except that the memory effects of the call are restricted to AliasIdxRaw.
2469
2470 // Slow path call has no side-effects, uses few values
2471 bool wide_in = !(flags & RC_NARROW_MEM);
2472 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2473
2474 Node* prev_mem = NULL;
2475 if (wide_in) {
2476 prev_mem = set_predefined_input_for_runtime_call(call);
2477 } else {
2478 assert(!wide_out, "narrow in => narrow out");
2479 Node* narrow_mem = memory(adr_type);
2480 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2481 }
2482
2483 // Hook each parm in order. Stop looking at the first NULL.
2484 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2485 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2486 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2487 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2488 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2489 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2490 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2491 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2492 /* close each nested if ===> */ } } } } } } } }
2493 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2494
2495 if (!is_leaf) {
2496 // Non-leaves can block and take safepoints:
2497 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2498 }
2499 // Non-leaves can throw exceptions:
2500 if (has_io) {
2501 call->set_req(TypeFunc::I_O, i_o());
2502 }
2503
2504 if (flags & RC_UNCOMMON) {
2505 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2506 // (An "if" probability corresponds roughly to an unconditional count.
2507 // Sort of.)
2508 call->set_cnt(PROB_UNLIKELY_MAG(4));
2509 }
2510
2511 Node* c = _gvn.transform(call);
2512 assert(c == call, "cannot disappear");
2513
2514 if (wide_out) {
2515 // Slow path call has full side-effects.
2516 set_predefined_output_for_runtime_call(call);
2517 } else {
2518 // Slow path call has few side-effects, and/or sets few values.
2519 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2520 }
2521
2522 if (has_io) {
2523 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2524 }
2525 return call;
2526
2527 }
2528
2529 //------------------------------merge_memory-----------------------------------
2530 // Merge memory from one path into the current memory state.
merge_memory(Node * new_mem,Node * region,int new_path)2531 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2532 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2533 Node* old_slice = mms.force_memory();
2534 Node* new_slice = mms.memory2();
2535 if (old_slice != new_slice) {
2536 PhiNode* phi;
2537 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2538 if (mms.is_empty()) {
2539 // clone base memory Phi's inputs for this memory slice
2540 assert(old_slice == mms.base_memory(), "sanity");
2541 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2542 _gvn.set_type(phi, Type::MEMORY);
2543 for (uint i = 1; i < phi->req(); i++) {
2544 phi->init_req(i, old_slice->in(i));
2545 }
2546 } else {
2547 phi = old_slice->as_Phi(); // Phi was generated already
2548 }
2549 } else {
2550 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2551 _gvn.set_type(phi, Type::MEMORY);
2552 }
2553 phi->set_req(new_path, new_slice);
2554 mms.set_memory(phi);
2555 }
2556 }
2557 }
2558
2559 //------------------------------make_slow_call_ex------------------------------
2560 // Make the exception handler hookups for the slow call
make_slow_call_ex(Node * call,ciInstanceKlass * ex_klass,bool separate_io_proj,bool deoptimize)2561 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2562 if (stopped()) return;
2563
2564 // Make a catch node with just two handlers: fall-through and catch-all
2565 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2566 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2567 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2568 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2569
2570 { PreserveJVMState pjvms(this);
2571 set_control(excp);
2572 set_i_o(i_o);
2573
2574 if (excp != top()) {
2575 if (deoptimize) {
2576 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2577 uncommon_trap(Deoptimization::Reason_unhandled,
2578 Deoptimization::Action_none);
2579 } else {
2580 // Create an exception state also.
2581 // Use an exact type if the caller has a specific exception.
2582 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2583 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2584 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2585 }
2586 }
2587 }
2588
2589 // Get the no-exception control from the CatchNode.
2590 set_control(norm);
2591 }
2592
gen_subtype_check_compare(Node * ctrl,Node * in1,Node * in2,BoolTest::mask test,float p,PhaseGVN * gvn,BasicType bt)2593 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2594 Node* cmp = NULL;
2595 switch(bt) {
2596 case T_INT: cmp = new CmpINode(in1, in2); break;
2597 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2598 default: fatal("unexpected comparison type %s", type2name(bt));
2599 }
2600 gvn->transform(cmp);
2601 Node* bol = gvn->transform(new BoolNode(cmp, test));
2602 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2603 gvn->transform(iff);
2604 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2605 return iff;
2606 }
2607
2608
2609 //-------------------------------gen_subtype_check-----------------------------
2610 // Generate a subtyping check. Takes as input the subtype and supertype.
2611 // Returns 2 values: sets the default control() to the true path and returns
2612 // the false path. Only reads invariant memory; sets no (visible) memory.
2613 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2614 // but that's not exposed to the optimizer. This call also doesn't take in an
2615 // Object; if you wish to check an Object you need to load the Object's class
2616 // prior to coming here.
gen_subtype_check(Node * subklass,Node * superklass,Node ** ctrl,MergeMemNode * mem,PhaseGVN * gvn)2617 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2618 Compile* C = gvn->C;
2619
2620 if ((*ctrl)->is_top()) {
2621 return C->top();
2622 }
2623
2624 // Fast check for identical types, perhaps identical constants.
2625 // The types can even be identical non-constants, in cases
2626 // involving Array.newInstance, Object.clone, etc.
2627 if (subklass == superklass)
2628 return C->top(); // false path is dead; no test needed.
2629
2630 if (gvn->type(superklass)->singleton()) {
2631 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2632 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2633
2634 // In the common case of an exact superklass, try to fold up the
2635 // test before generating code. You may ask, why not just generate
2636 // the code and then let it fold up? The answer is that the generated
2637 // code will necessarily include null checks, which do not always
2638 // completely fold away. If they are also needless, then they turn
2639 // into a performance loss. Example:
2640 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2641 // Here, the type of 'fa' is often exact, so the store check
2642 // of fa[1]=x will fold up, without testing the nullness of x.
2643 switch (C->static_subtype_check(superk, subk)) {
2644 case Compile::SSC_always_false:
2645 {
2646 Node* always_fail = *ctrl;
2647 *ctrl = gvn->C->top();
2648 return always_fail;
2649 }
2650 case Compile::SSC_always_true:
2651 return C->top();
2652 case Compile::SSC_easy_test:
2653 {
2654 // Just do a direct pointer compare and be done.
2655 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2656 *ctrl = gvn->transform(new IfTrueNode(iff));
2657 return gvn->transform(new IfFalseNode(iff));
2658 }
2659 case Compile::SSC_full_test:
2660 break;
2661 default:
2662 ShouldNotReachHere();
2663 }
2664 }
2665
2666 // %%% Possible further optimization: Even if the superklass is not exact,
2667 // if the subklass is the unique subtype of the superklass, the check
2668 // will always succeed. We could leave a dependency behind to ensure this.
2669
2670 // First load the super-klass's check-offset
2671 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2672 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2673 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2674 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2675 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2676
2677 // Load from the sub-klass's super-class display list, or a 1-word cache of
2678 // the secondary superclass list, or a failing value with a sentinel offset
2679 // if the super-klass is an interface or exceptionally deep in the Java
2680 // hierarchy and we have to scan the secondary superclass list the hard way.
2681 // Worst-case type is a little odd: NULL is allowed as a result (usually
2682 // klass loads can never produce a NULL).
2683 Node *chk_off_X = chk_off;
2684 #ifdef _LP64
2685 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2686 #endif
2687 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2688 // For some types like interfaces the following loadKlass is from a 1-word
2689 // cache which is mutable so can't use immutable memory. Other
2690 // types load from the super-class display table which is immutable.
2691 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2692 Node *kmem = might_be_cache ? m : C->immutable_memory();
2693 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2694
2695 // Compile speed common case: ARE a subtype and we canNOT fail
2696 if( superklass == nkls )
2697 return C->top(); // false path is dead; no test needed.
2698
2699 // See if we get an immediate positive hit. Happens roughly 83% of the
2700 // time. Test to see if the value loaded just previously from the subklass
2701 // is exactly the superklass.
2702 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2703 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2704 *ctrl = gvn->transform(new IfFalseNode(iff1));
2705
2706 // Compile speed common case: Check for being deterministic right now. If
2707 // chk_off is a constant and not equal to cacheoff then we are NOT a
2708 // subklass. In this case we need exactly the 1 test above and we can
2709 // return those results immediately.
2710 if (!might_be_cache) {
2711 Node* not_subtype_ctrl = *ctrl;
2712 *ctrl = iftrue1; // We need exactly the 1 test above
2713 return not_subtype_ctrl;
2714 }
2715
2716 // Gather the various success & failures here
2717 RegionNode *r_ok_subtype = new RegionNode(4);
2718 gvn->record_for_igvn(r_ok_subtype);
2719 RegionNode *r_not_subtype = new RegionNode(3);
2720 gvn->record_for_igvn(r_not_subtype);
2721
2722 r_ok_subtype->init_req(1, iftrue1);
2723
2724 // Check for immediate negative hit. Happens roughly 11% of the time (which
2725 // is roughly 63% of the remaining cases). Test to see if the loaded
2726 // check-offset points into the subklass display list or the 1-element
2727 // cache. If it points to the display (and NOT the cache) and the display
2728 // missed then it's not a subtype.
2729 Node *cacheoff = gvn->intcon(cacheoff_con);
2730 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2731 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2732 *ctrl = gvn->transform(new IfFalseNode(iff2));
2733
2734 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2735 // No performance impact (too rare) but allows sharing of secondary arrays
2736 // which has some footprint reduction.
2737 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2738 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2739 *ctrl = gvn->transform(new IfFalseNode(iff3));
2740
2741 // -- Roads not taken here: --
2742 // We could also have chosen to perform the self-check at the beginning
2743 // of this code sequence, as the assembler does. This would not pay off
2744 // the same way, since the optimizer, unlike the assembler, can perform
2745 // static type analysis to fold away many successful self-checks.
2746 // Non-foldable self checks work better here in second position, because
2747 // the initial primary superclass check subsumes a self-check for most
2748 // types. An exception would be a secondary type like array-of-interface,
2749 // which does not appear in its own primary supertype display.
2750 // Finally, we could have chosen to move the self-check into the
2751 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2752 // dependent manner. But it is worthwhile to have the check here,
2753 // where it can be perhaps be optimized. The cost in code space is
2754 // small (register compare, branch).
2755
2756 // Now do a linear scan of the secondary super-klass array. Again, no real
2757 // performance impact (too rare) but it's gotta be done.
2758 // Since the code is rarely used, there is no penalty for moving it
2759 // out of line, and it can only improve I-cache density.
2760 // The decision to inline or out-of-line this final check is platform
2761 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2762 Node* psc = gvn->transform(
2763 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2764
2765 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2766 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2767 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2768
2769 // Return false path; set default control to true path.
2770 *ctrl = gvn->transform(r_ok_subtype);
2771 return gvn->transform(r_not_subtype);
2772 }
2773
2774 // Profile-driven exact type check:
type_check_receiver(Node * receiver,ciKlass * klass,float prob,Node ** casted_receiver)2775 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2776 float prob,
2777 Node* *casted_receiver) {
2778 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2779 Node* recv_klass = load_object_klass(receiver);
2780 Node* want_klass = makecon(tklass);
2781 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2782 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2783 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2784 set_control( _gvn.transform( new IfTrueNode (iff) ));
2785 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2786
2787 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2788 assert(recv_xtype->klass_is_exact(), "");
2789
2790 // Subsume downstream occurrences of receiver with a cast to
2791 // recv_xtype, since now we know what the type will be.
2792 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2793 (*casted_receiver) = _gvn.transform(cast);
2794 // (User must make the replace_in_map call.)
2795
2796 return fail;
2797 }
2798
2799
2800 //------------------------------seems_never_null-------------------------------
2801 // Use null_seen information if it is available from the profile.
2802 // If we see an unexpected null at a type check we record it and force a
2803 // recompile; the offending check will be recompiled to handle NULLs.
2804 // If we see several offending BCIs, then all checks in the
2805 // method will be recompiled.
seems_never_null(Node * obj,ciProfileData * data,bool & speculating)2806 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2807 speculating = !_gvn.type(obj)->speculative_maybe_null();
2808 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2809 if (UncommonNullCast // Cutout for this technique
2810 && obj != null() // And not the -Xcomp stupid case?
2811 && !too_many_traps(reason)
2812 ) {
2813 if (speculating) {
2814 return true;
2815 }
2816 if (data == NULL)
2817 // Edge case: no mature data. Be optimistic here.
2818 return true;
2819 // If the profile has not seen a null, assume it won't happen.
2820 assert(java_bc() == Bytecodes::_checkcast ||
2821 java_bc() == Bytecodes::_instanceof ||
2822 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2823 return !data->as_BitData()->null_seen();
2824 }
2825 speculating = false;
2826 return false;
2827 }
2828
2829 //------------------------maybe_cast_profiled_receiver-------------------------
2830 // If the profile has seen exactly one type, narrow to exactly that type.
2831 // 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)2832 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2833 ciKlass* require_klass,
2834 ciKlass* spec_klass,
2835 bool safe_for_replace) {
2836 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2837
2838 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2839
2840 // Make sure we haven't already deoptimized from this tactic.
2841 if (too_many_traps(reason) || too_many_recompiles(reason))
2842 return NULL;
2843
2844 // (No, this isn't a call, but it's enough like a virtual call
2845 // to use the same ciMethod accessor to get the profile info...)
2846 // If we have a speculative type use it instead of profiling (which
2847 // may not help us)
2848 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2849 if (exact_kls != NULL) {// no cast failures here
2850 if (require_klass == NULL ||
2851 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2852 // If we narrow the type to match what the type profile sees or
2853 // the speculative type, we can then remove the rest of the
2854 // cast.
2855 // This is a win, even if the exact_kls is very specific,
2856 // because downstream operations, such as method calls,
2857 // will often benefit from the sharper type.
2858 Node* exact_obj = not_null_obj; // will get updated in place...
2859 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2860 &exact_obj);
2861 { PreserveJVMState pjvms(this);
2862 set_control(slow_ctl);
2863 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2864 }
2865 if (safe_for_replace) {
2866 replace_in_map(not_null_obj, exact_obj);
2867 }
2868 return exact_obj;
2869 }
2870 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2871 }
2872
2873 return NULL;
2874 }
2875
2876 /**
2877 * Cast obj to type and emit guard unless we had too many traps here
2878 * already
2879 *
2880 * @param obj node being casted
2881 * @param type type to cast the node to
2882 * @param not_null true if we know node cannot be null
2883 */
maybe_cast_profiled_obj(Node * obj,ciKlass * type,bool not_null)2884 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2885 ciKlass* type,
2886 bool not_null) {
2887 if (stopped()) {
2888 return obj;
2889 }
2890
2891 // type == NULL if profiling tells us this object is always null
2892 if (type != NULL) {
2893 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2894 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2895
2896 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2897 !too_many_traps(class_reason) &&
2898 !too_many_recompiles(class_reason)) {
2899 Node* not_null_obj = NULL;
2900 // not_null is true if we know the object is not null and
2901 // there's no need for a null check
2902 if (!not_null) {
2903 Node* null_ctl = top();
2904 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2905 assert(null_ctl->is_top(), "no null control here");
2906 } else {
2907 not_null_obj = obj;
2908 }
2909
2910 Node* exact_obj = not_null_obj;
2911 ciKlass* exact_kls = type;
2912 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2913 &exact_obj);
2914 {
2915 PreserveJVMState pjvms(this);
2916 set_control(slow_ctl);
2917 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2918 }
2919 replace_in_map(not_null_obj, exact_obj);
2920 obj = exact_obj;
2921 }
2922 } else {
2923 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2924 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2925 Node* exact_obj = null_assert(obj);
2926 replace_in_map(obj, exact_obj);
2927 obj = exact_obj;
2928 }
2929 }
2930 return obj;
2931 }
2932
2933 //-------------------------------gen_instanceof--------------------------------
2934 // Generate an instance-of idiom. Used by both the instance-of bytecode
2935 // and the reflective instance-of call.
gen_instanceof(Node * obj,Node * superklass,bool safe_for_replace)2936 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2937 kill_dead_locals(); // Benefit all the uncommon traps
2938 assert( !stopped(), "dead parse path should be checked in callers" );
2939 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2940 "must check for not-null not-dead klass in callers");
2941
2942 // Make the merge point
2943 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2944 RegionNode* region = new RegionNode(PATH_LIMIT);
2945 Node* phi = new PhiNode(region, TypeInt::BOOL);
2946 C->set_has_split_ifs(true); // Has chance for split-if optimization
2947
2948 ciProfileData* data = NULL;
2949 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2950 data = method()->method_data()->bci_to_data(bci());
2951 }
2952 bool speculative_not_null = false;
2953 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2954 && seems_never_null(obj, data, speculative_not_null));
2955
2956 // Null check; get casted pointer; set region slot 3
2957 Node* null_ctl = top();
2958 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2959
2960 // If not_null_obj is dead, only null-path is taken
2961 if (stopped()) { // Doing instance-of on a NULL?
2962 set_control(null_ctl);
2963 return intcon(0);
2964 }
2965 region->init_req(_null_path, null_ctl);
2966 phi ->init_req(_null_path, intcon(0)); // Set null path value
2967 if (null_ctl == top()) {
2968 // Do this eagerly, so that pattern matches like is_diamond_phi
2969 // will work even during parsing.
2970 assert(_null_path == PATH_LIMIT-1, "delete last");
2971 region->del_req(_null_path);
2972 phi ->del_req(_null_path);
2973 }
2974
2975 // Do we know the type check always succeed?
2976 bool known_statically = false;
2977 if (_gvn.type(superklass)->singleton()) {
2978 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2979 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2980 if (subk != NULL && subk->is_loaded()) {
2981 int static_res = C->static_subtype_check(superk, subk);
2982 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2983 }
2984 }
2985
2986 if (!known_statically) {
2987 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2988 // We may not have profiling here or it may not help us. If we
2989 // have a speculative type use it to perform an exact cast.
2990 ciKlass* spec_obj_type = obj_type->speculative_type();
2991 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2992 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2993 if (stopped()) { // Profile disagrees with this path.
2994 set_control(null_ctl); // Null is the only remaining possibility.
2995 return intcon(0);
2996 }
2997 if (cast_obj != NULL) {
2998 not_null_obj = cast_obj;
2999 }
3000 }
3001 }
3002
3003 // Load the object's klass
3004 Node* obj_klass = load_object_klass(not_null_obj);
3005
3006 // Generate the subtype check
3007 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3008
3009 // Plug in the success path to the general merge in slot 1.
3010 region->init_req(_obj_path, control());
3011 phi ->init_req(_obj_path, intcon(1));
3012
3013 // Plug in the failing path to the general merge in slot 2.
3014 region->init_req(_fail_path, not_subtype_ctrl);
3015 phi ->init_req(_fail_path, intcon(0));
3016
3017 // Return final merged results
3018 set_control( _gvn.transform(region) );
3019 record_for_igvn(region);
3020
3021 // If we know the type check always succeeds then we don't use the
3022 // profiling data at this bytecode. Don't lose it, feed it to the
3023 // type system as a speculative type.
3024 if (safe_for_replace) {
3025 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3026 replace_in_map(obj, casted_obj);
3027 }
3028
3029 return _gvn.transform(phi);
3030 }
3031
3032 //-------------------------------gen_checkcast---------------------------------
3033 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3034 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3035 // uncommon-trap paths work. Adjust stack after this call.
3036 // If failure_control is supplied and not null, it is filled in with
3037 // the control edge for the cast failure. Otherwise, an appropriate
3038 // uncommon trap or exception is thrown.
gen_checkcast(Node * obj,Node * superklass,Node ** failure_control)3039 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3040 Node* *failure_control) {
3041 kill_dead_locals(); // Benefit all the uncommon traps
3042 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3043 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3044
3045 // Fast cutout: Check the case that the cast is vacuously true.
3046 // This detects the common cases where the test will short-circuit
3047 // away completely. We do this before we perform the null check,
3048 // because if the test is going to turn into zero code, we don't
3049 // want a residual null check left around. (Causes a slowdown,
3050 // for example, in some objArray manipulations, such as a[i]=a[j].)
3051 if (tk->singleton()) {
3052 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3053 if (objtp != NULL && objtp->klass() != NULL) {
3054 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3055 case Compile::SSC_always_true:
3056 // If we know the type check always succeed then we don't use
3057 // the profiling data at this bytecode. Don't lose it, feed it
3058 // to the type system as a speculative type.
3059 return record_profiled_receiver_for_speculation(obj);
3060 case Compile::SSC_always_false:
3061 // It needs a null check because a null will *pass* the cast check.
3062 // A non-null value will always produce an exception.
3063 return null_assert(obj);
3064 }
3065 }
3066 }
3067
3068 ciProfileData* data = NULL;
3069 bool safe_for_replace = false;
3070 if (failure_control == NULL) { // use MDO in regular case only
3071 assert(java_bc() == Bytecodes::_aastore ||
3072 java_bc() == Bytecodes::_checkcast,
3073 "interpreter profiles type checks only for these BCs");
3074 data = method()->method_data()->bci_to_data(bci());
3075 safe_for_replace = true;
3076 }
3077
3078 // Make the merge point
3079 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3080 RegionNode* region = new RegionNode(PATH_LIMIT);
3081 Node* phi = new PhiNode(region, toop);
3082 C->set_has_split_ifs(true); // Has chance for split-if optimization
3083
3084 // Use null-cast information if it is available
3085 bool speculative_not_null = false;
3086 bool never_see_null = ((failure_control == NULL) // regular case only
3087 && seems_never_null(obj, data, speculative_not_null));
3088
3089 // Null check; get casted pointer; set region slot 3
3090 Node* null_ctl = top();
3091 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3092
3093 // If not_null_obj is dead, only null-path is taken
3094 if (stopped()) { // Doing instance-of on a NULL?
3095 set_control(null_ctl);
3096 return null();
3097 }
3098 region->init_req(_null_path, null_ctl);
3099 phi ->init_req(_null_path, null()); // Set null path value
3100 if (null_ctl == top()) {
3101 // Do this eagerly, so that pattern matches like is_diamond_phi
3102 // will work even during parsing.
3103 assert(_null_path == PATH_LIMIT-1, "delete last");
3104 region->del_req(_null_path);
3105 phi ->del_req(_null_path);
3106 }
3107
3108 Node* cast_obj = NULL;
3109 if (tk->klass_is_exact()) {
3110 // The following optimization tries to statically cast the speculative type of the object
3111 // (for example obtained during profiling) to the type of the superklass and then do a
3112 // dynamic check that the type of the object is what we expect. To work correctly
3113 // for checkcast and aastore the type of superklass should be exact.
3114 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3115 // We may not have profiling here or it may not help us. If we have
3116 // a speculative type use it to perform an exact cast.
3117 ciKlass* spec_obj_type = obj_type->speculative_type();
3118 if (spec_obj_type != NULL || data != NULL) {
3119 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3120 if (cast_obj != NULL) {
3121 if (failure_control != NULL) // failure is now impossible
3122 (*failure_control) = top();
3123 // adjust the type of the phi to the exact klass:
3124 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3125 }
3126 }
3127 }
3128
3129 if (cast_obj == NULL) {
3130 // Load the object's klass
3131 Node* obj_klass = load_object_klass(not_null_obj);
3132
3133 // Generate the subtype check
3134 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3135
3136 // Plug in success path into the merge
3137 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3138 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3139 if (failure_control == NULL) {
3140 if (not_subtype_ctrl != top()) { // If failure is possible
3141 PreserveJVMState pjvms(this);
3142 set_control(not_subtype_ctrl);
3143 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3144 }
3145 } else {
3146 (*failure_control) = not_subtype_ctrl;
3147 }
3148 }
3149
3150 region->init_req(_obj_path, control());
3151 phi ->init_req(_obj_path, cast_obj);
3152
3153 // A merge of NULL or Casted-NotNull obj
3154 Node* res = _gvn.transform(phi);
3155
3156 // Note I do NOT always 'replace_in_map(obj,result)' here.
3157 // if( tk->klass()->can_be_primary_super() )
3158 // This means that if I successfully store an Object into an array-of-String
3159 // I 'forget' that the Object is really now known to be a String. I have to
3160 // do this because we don't have true union types for interfaces - if I store
3161 // a Baz into an array-of-Interface and then tell the optimizer it's an
3162 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3163 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3164 // replace_in_map( obj, res );
3165
3166 // Return final merged results
3167 set_control( _gvn.transform(region) );
3168 record_for_igvn(region);
3169
3170 return record_profiled_receiver_for_speculation(res);
3171 }
3172
3173 //------------------------------next_monitor-----------------------------------
3174 // What number should be given to the next monitor?
next_monitor()3175 int GraphKit::next_monitor() {
3176 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3177 int next = current + C->sync_stack_slots();
3178 // Keep the toplevel high water mark current:
3179 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3180 return current;
3181 }
3182
3183 //------------------------------insert_mem_bar---------------------------------
3184 // Memory barrier to avoid floating things around
3185 // The membar serves as a pinch point between both control and all memory slices.
insert_mem_bar(int opcode,Node * precedent)3186 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3187 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3188 mb->init_req(TypeFunc::Control, control());
3189 mb->init_req(TypeFunc::Memory, reset_memory());
3190 Node* membar = _gvn.transform(mb);
3191 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3192 set_all_memory_call(membar);
3193 return membar;
3194 }
3195
3196 //-------------------------insert_mem_bar_volatile----------------------------
3197 // Memory barrier to avoid floating things around
3198 // The membar serves as a pinch point between both control and memory(alias_idx).
3199 // If you want to make a pinch point on all memory slices, do not use this
3200 // function (even with AliasIdxBot); use insert_mem_bar() instead.
insert_mem_bar_volatile(int opcode,int alias_idx,Node * precedent)3201 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3202 // When Parse::do_put_xxx updates a volatile field, it appends a series
3203 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3204 // The first membar is on the same memory slice as the field store opcode.
3205 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3206 // All the other membars (for other volatile slices, including AliasIdxBot,
3207 // which stands for all unknown volatile slices) are control-dependent
3208 // on the first membar. This prevents later volatile loads or stores
3209 // from sliding up past the just-emitted store.
3210
3211 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3212 mb->set_req(TypeFunc::Control,control());
3213 if (alias_idx == Compile::AliasIdxBot) {
3214 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3215 } else {
3216 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3217 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3218 }
3219 Node* membar = _gvn.transform(mb);
3220 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3221 if (alias_idx == Compile::AliasIdxBot) {
3222 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3223 } else {
3224 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3225 }
3226 return membar;
3227 }
3228
3229 //------------------------------shared_lock------------------------------------
3230 // Emit locking code.
shared_lock(Node * obj)3231 FastLockNode* GraphKit::shared_lock(Node* obj) {
3232 // bci is either a monitorenter bc or InvocationEntryBci
3233 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3234 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3235
3236 if( !GenerateSynchronizationCode )
3237 return NULL; // Not locking things?
3238 if (stopped()) // Dead monitor?
3239 return NULL;
3240
3241 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3242
3243 obj = access_resolve(obj, ACCESS_READ | ACCESS_WRITE);
3244
3245 // Box the stack location
3246 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3247 Node* mem = reset_memory();
3248
3249 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3250 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3251 // Create the counters for this fast lock.
3252 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3253 }
3254
3255 // Create the rtm counters for this fast lock if needed.
3256 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3257
3258 // Add monitor to debug info for the slow path. If we block inside the
3259 // slow path and de-opt, we need the monitor hanging around
3260 map()->push_monitor( flock );
3261
3262 const TypeFunc *tf = LockNode::lock_type();
3263 LockNode *lock = new LockNode(C, tf);
3264
3265 lock->init_req( TypeFunc::Control, control() );
3266 lock->init_req( TypeFunc::Memory , mem );
3267 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3268 lock->init_req( TypeFunc::FramePtr, frameptr() );
3269 lock->init_req( TypeFunc::ReturnAdr, top() );
3270
3271 lock->init_req(TypeFunc::Parms + 0, obj);
3272 lock->init_req(TypeFunc::Parms + 1, box);
3273 lock->init_req(TypeFunc::Parms + 2, flock);
3274 add_safepoint_edges(lock);
3275
3276 lock = _gvn.transform( lock )->as_Lock();
3277
3278 // lock has no side-effects, sets few values
3279 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3280
3281 insert_mem_bar(Op_MemBarAcquireLock);
3282
3283 // Add this to the worklist so that the lock can be eliminated
3284 record_for_igvn(lock);
3285
3286 #ifndef PRODUCT
3287 if (PrintLockStatistics) {
3288 // Update the counter for this lock. Don't bother using an atomic
3289 // operation since we don't require absolute accuracy.
3290 lock->create_lock_counter(map()->jvms());
3291 increment_counter(lock->counter()->addr());
3292 }
3293 #endif
3294
3295 return flock;
3296 }
3297
3298
3299 //------------------------------shared_unlock----------------------------------
3300 // Emit unlocking code.
shared_unlock(Node * box,Node * obj)3301 void GraphKit::shared_unlock(Node* box, Node* obj) {
3302 // bci is either a monitorenter bc or InvocationEntryBci
3303 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3304 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3305
3306 if( !GenerateSynchronizationCode )
3307 return;
3308 if (stopped()) { // Dead monitor?
3309 map()->pop_monitor(); // Kill monitor from debug info
3310 return;
3311 }
3312
3313 // Memory barrier to avoid floating things down past the locked region
3314 insert_mem_bar(Op_MemBarReleaseLock);
3315
3316 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3317 UnlockNode *unlock = new UnlockNode(C, tf);
3318 #ifdef ASSERT
3319 unlock->set_dbg_jvms(sync_jvms());
3320 #endif
3321 uint raw_idx = Compile::AliasIdxRaw;
3322 unlock->init_req( TypeFunc::Control, control() );
3323 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3324 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3325 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3326 unlock->init_req( TypeFunc::ReturnAdr, top() );
3327
3328 unlock->init_req(TypeFunc::Parms + 0, obj);
3329 unlock->init_req(TypeFunc::Parms + 1, box);
3330 unlock = _gvn.transform(unlock)->as_Unlock();
3331
3332 Node* mem = reset_memory();
3333
3334 // unlock has no side-effects, sets few values
3335 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3336
3337 // Kill monitor from debug info
3338 map()->pop_monitor( );
3339 }
3340
3341 //-------------------------------get_layout_helper-----------------------------
3342 // If the given klass is a constant or known to be an array,
3343 // fetch the constant layout helper value into constant_value
3344 // and return (Node*)NULL. Otherwise, load the non-constant
3345 // layout helper value, and return the node which represents it.
3346 // This two-faced routine is useful because allocation sites
3347 // almost always feature constant types.
get_layout_helper(Node * klass_node,jint & constant_value)3348 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3349 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3350 if (!StressReflectiveCode && inst_klass != NULL) {
3351 ciKlass* klass = inst_klass->klass();
3352 bool xklass = inst_klass->klass_is_exact();
3353 if (xklass || klass->is_array_klass()) {
3354 jint lhelper = klass->layout_helper();
3355 if (lhelper != Klass::_lh_neutral_value) {
3356 constant_value = lhelper;
3357 return (Node*) NULL;
3358 }
3359 }
3360 }
3361 constant_value = Klass::_lh_neutral_value; // put in a known value
3362 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3363 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3364 }
3365
3366 // We just put in an allocate/initialize with a big raw-memory effect.
3367 // 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)3368 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3369 MergeMemNode* init_in_merge,
3370 Node* init_out_raw) {
3371 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3372 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3373
3374 Node* prevmem = kit.memory(alias_idx);
3375 init_in_merge->set_memory_at(alias_idx, prevmem);
3376 kit.set_memory(init_out_raw, alias_idx);
3377 }
3378
3379 //---------------------------set_output_for_allocation-------------------------
set_output_for_allocation(AllocateNode * alloc,const TypeOopPtr * oop_type,bool deoptimize_on_exception)3380 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3381 const TypeOopPtr* oop_type,
3382 bool deoptimize_on_exception) {
3383 int rawidx = Compile::AliasIdxRaw;
3384 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3385 add_safepoint_edges(alloc);
3386 Node* allocx = _gvn.transform(alloc);
3387 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3388 // create memory projection for i_o
3389 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3390 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3391
3392 // create a memory projection as for the normal control path
3393 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3394 set_memory(malloc, rawidx);
3395
3396 // a normal slow-call doesn't change i_o, but an allocation does
3397 // we create a separate i_o projection for the normal control path
3398 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3399 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3400
3401 // put in an initialization barrier
3402 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3403 rawoop)->as_Initialize();
3404 assert(alloc->initialization() == init, "2-way macro link must work");
3405 assert(init ->allocation() == alloc, "2-way macro link must work");
3406 {
3407 // Extract memory strands which may participate in the new object's
3408 // initialization, and source them from the new InitializeNode.
3409 // This will allow us to observe initializations when they occur,
3410 // and link them properly (as a group) to the InitializeNode.
3411 assert(init->in(InitializeNode::Memory) == malloc, "");
3412 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3413 init->set_req(InitializeNode::Memory, minit_in);
3414 record_for_igvn(minit_in); // fold it up later, if possible
3415 Node* minit_out = memory(rawidx);
3416 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3417 if (oop_type->isa_aryptr()) {
3418 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3419 int elemidx = C->get_alias_index(telemref);
3420 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3421 } else if (oop_type->isa_instptr()) {
3422 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3423 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3424 ciField* field = ik->nonstatic_field_at(i);
3425 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3426 continue; // do not bother to track really large numbers of fields
3427 // Find (or create) the alias category for this field:
3428 int fieldidx = C->alias_type(field)->index();
3429 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3430 }
3431 }
3432 }
3433
3434 // Cast raw oop to the real thing...
3435 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3436 javaoop = _gvn.transform(javaoop);
3437 C->set_recent_alloc(control(), javaoop);
3438 assert(just_allocated_object(control()) == javaoop, "just allocated");
3439
3440 #ifdef ASSERT
3441 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3442 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3443 "Ideal_allocation works");
3444 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3445 "Ideal_allocation works");
3446 if (alloc->is_AllocateArray()) {
3447 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3448 "Ideal_allocation works");
3449 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3450 "Ideal_allocation works");
3451 } else {
3452 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3453 }
3454 }
3455 #endif //ASSERT
3456
3457 return javaoop;
3458 }
3459
3460 //---------------------------new_instance--------------------------------------
3461 // This routine takes a klass_node which may be constant (for a static type)
3462 // or may be non-constant (for reflective code). It will work equally well
3463 // for either, and the graph will fold nicely if the optimizer later reduces
3464 // the type to a constant.
3465 // The optional arguments are for specialized use by intrinsics:
3466 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3467 // - If 'return_size_val', report the the total object size to the caller.
3468 // - 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)3469 Node* GraphKit::new_instance(Node* klass_node,
3470 Node* extra_slow_test,
3471 Node* *return_size_val,
3472 bool deoptimize_on_exception) {
3473 // Compute size in doublewords
3474 // The size is always an integral number of doublewords, represented
3475 // as a positive bytewise size stored in the klass's layout_helper.
3476 // The layout_helper also encodes (in a low bit) the need for a slow path.
3477 jint layout_con = Klass::_lh_neutral_value;
3478 Node* layout_val = get_layout_helper(klass_node, layout_con);
3479 int layout_is_con = (layout_val == NULL);
3480
3481 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3482 // Generate the initial go-slow test. It's either ALWAYS (return a
3483 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3484 // case) a computed value derived from the layout_helper.
3485 Node* initial_slow_test = NULL;
3486 if (layout_is_con) {
3487 assert(!StressReflectiveCode, "stress mode does not use these paths");
3488 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3489 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3490 } else { // reflective case
3491 // This reflective path is used by Unsafe.allocateInstance.
3492 // (It may be stress-tested by specifying StressReflectiveCode.)
3493 // Basically, we want to get into the VM is there's an illegal argument.
3494 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3495 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3496 if (extra_slow_test != intcon(0)) {
3497 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3498 }
3499 // (Macro-expander will further convert this to a Bool, if necessary.)
3500 }
3501
3502 // Find the size in bytes. This is easy; it's the layout_helper.
3503 // The size value must be valid even if the slow path is taken.
3504 Node* size = NULL;
3505 if (layout_is_con) {
3506 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3507 } else { // reflective case
3508 // This reflective path is used by clone and Unsafe.allocateInstance.
3509 size = ConvI2X(layout_val);
3510
3511 // Clear the low bits to extract layout_helper_size_in_bytes:
3512 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3513 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3514 size = _gvn.transform( new AndXNode(size, mask) );
3515 }
3516 if (return_size_val != NULL) {
3517 (*return_size_val) = size;
3518 }
3519
3520 // This is a precise notnull oop of the klass.
3521 // (Actually, it need not be precise if this is a reflective allocation.)
3522 // It's what we cast the result to.
3523 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3524 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3525 const TypeOopPtr* oop_type = tklass->as_instance_type();
3526
3527 // Now generate allocation code
3528
3529 // The entire memory state is needed for slow path of the allocation
3530 // since GC and deoptimization can happened.
3531 Node *mem = reset_memory();
3532 set_all_memory(mem); // Create new memory state
3533
3534 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3535 control(), mem, i_o(),
3536 size, klass_node,
3537 initial_slow_test);
3538
3539 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3540 }
3541
3542 //-------------------------------new_array-------------------------------------
3543 // helper for both newarray and anewarray
3544 // The 'length' parameter is (obviously) the length of the array.
3545 // 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)3546 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3547 Node* length, // number of array elements
3548 int nargs, // number of arguments to push back for uncommon trap
3549 Node* *return_size_val,
3550 bool deoptimize_on_exception) {
3551 jint layout_con = Klass::_lh_neutral_value;
3552 Node* layout_val = get_layout_helper(klass_node, layout_con);
3553 int layout_is_con = (layout_val == NULL);
3554
3555 if (!layout_is_con && !StressReflectiveCode &&
3556 !too_many_traps(Deoptimization::Reason_class_check)) {
3557 // This is a reflective array creation site.
3558 // Optimistically assume that it is a subtype of Object[],
3559 // so that we can fold up all the address arithmetic.
3560 layout_con = Klass::array_layout_helper(T_OBJECT);
3561 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3562 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3563 { BuildCutout unless(this, bol_lh, PROB_MAX);
3564 inc_sp(nargs);
3565 uncommon_trap(Deoptimization::Reason_class_check,
3566 Deoptimization::Action_maybe_recompile);
3567 }
3568 layout_val = NULL;
3569 layout_is_con = true;
3570 }
3571
3572 // Generate the initial go-slow test. Make sure we do not overflow
3573 // if length is huge (near 2Gig) or negative! We do not need
3574 // exact double-words here, just a close approximation of needed
3575 // double-words. We can't add any offset or rounding bits, lest we
3576 // take a size -1 of bytes and make it positive. Use an unsigned
3577 // compare, so negative sizes look hugely positive.
3578 int fast_size_limit = FastAllocateSizeLimit;
3579 if (layout_is_con) {
3580 assert(!StressReflectiveCode, "stress mode does not use these paths");
3581 // Increase the size limit if we have exact knowledge of array type.
3582 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3583 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3584 }
3585
3586 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3587 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3588
3589 // --- Size Computation ---
3590 // array_size = round_to_heap(array_header + (length << elem_shift));
3591 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3592 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3593 // The rounding mask is strength-reduced, if possible.
3594 int round_mask = MinObjAlignmentInBytes - 1;
3595 Node* header_size = NULL;
3596 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3597 // (T_BYTE has the weakest alignment and size restrictions...)
3598 if (layout_is_con) {
3599 int hsize = Klass::layout_helper_header_size(layout_con);
3600 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3601 BasicType etype = Klass::layout_helper_element_type(layout_con);
3602 if ((round_mask & ~right_n_bits(eshift)) == 0)
3603 round_mask = 0; // strength-reduce it if it goes away completely
3604 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3605 assert(header_size_min <= hsize, "generic minimum is smallest");
3606 header_size_min = hsize;
3607 header_size = intcon(hsize + round_mask);
3608 } else {
3609 Node* hss = intcon(Klass::_lh_header_size_shift);
3610 Node* hsm = intcon(Klass::_lh_header_size_mask);
3611 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3612 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3613 Node* mask = intcon(round_mask);
3614 header_size = _gvn.transform( new AddINode(hsize, mask) );
3615 }
3616
3617 Node* elem_shift = NULL;
3618 if (layout_is_con) {
3619 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3620 if (eshift != 0)
3621 elem_shift = intcon(eshift);
3622 } else {
3623 // There is no need to mask or shift this value.
3624 // The semantics of LShiftINode include an implicit mask to 0x1F.
3625 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3626 elem_shift = layout_val;
3627 }
3628
3629 // Transition to native address size for all offset calculations:
3630 Node* lengthx = ConvI2X(length);
3631 Node* headerx = ConvI2X(header_size);
3632 #ifdef _LP64
3633 { const TypeInt* tilen = _gvn.find_int_type(length);
3634 if (tilen != NULL && tilen->_lo < 0) {
3635 // Add a manual constraint to a positive range. Cf. array_element_address.
3636 jint size_max = fast_size_limit;
3637 if (size_max > tilen->_hi) size_max = tilen->_hi;
3638 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3639
3640 // Only do a narrow I2L conversion if the range check passed.
3641 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3642 _gvn.transform(iff);
3643 RegionNode* region = new RegionNode(3);
3644 _gvn.set_type(region, Type::CONTROL);
3645 lengthx = new PhiNode(region, TypeLong::LONG);
3646 _gvn.set_type(lengthx, TypeLong::LONG);
3647
3648 // Range check passed. Use ConvI2L node with narrow type.
3649 Node* passed = IfFalse(iff);
3650 region->init_req(1, passed);
3651 // Make I2L conversion control dependent to prevent it from
3652 // floating above the range check during loop optimizations.
3653 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3654
3655 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3656 region->init_req(2, IfTrue(iff));
3657 lengthx->init_req(2, ConvI2X(length));
3658
3659 set_control(region);
3660 record_for_igvn(region);
3661 record_for_igvn(lengthx);
3662 }
3663 }
3664 #endif
3665
3666 // Combine header size (plus rounding) and body size. Then round down.
3667 // This computation cannot overflow, because it is used only in two
3668 // places, one where the length is sharply limited, and the other
3669 // after a successful allocation.
3670 Node* abody = lengthx;
3671 if (elem_shift != NULL)
3672 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3673 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3674 if (round_mask != 0) {
3675 Node* mask = MakeConX(~round_mask);
3676 size = _gvn.transform( new AndXNode(size, mask) );
3677 }
3678 // else if round_mask == 0, the size computation is self-rounding
3679
3680 if (return_size_val != NULL) {
3681 // This is the size
3682 (*return_size_val) = size;
3683 }
3684
3685 // Now generate allocation code
3686
3687 // The entire memory state is needed for slow path of the allocation
3688 // since GC and deoptimization can happened.
3689 Node *mem = reset_memory();
3690 set_all_memory(mem); // Create new memory state
3691
3692 if (initial_slow_test->is_Bool()) {
3693 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3694 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3695 }
3696
3697 // Create the AllocateArrayNode and its result projections
3698 AllocateArrayNode* alloc
3699 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3700 control(), mem, i_o(),
3701 size, klass_node,
3702 initial_slow_test,
3703 length);
3704
3705 // Cast to correct type. Note that the klass_node may be constant or not,
3706 // and in the latter case the actual array type will be inexact also.
3707 // (This happens via a non-constant argument to inline_native_newArray.)
3708 // In any case, the value of klass_node provides the desired array type.
3709 const TypeInt* length_type = _gvn.find_int_type(length);
3710 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3711 if (ary_type->isa_aryptr() && length_type != NULL) {
3712 // Try to get a better type than POS for the size
3713 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3714 }
3715
3716 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3717
3718 // Cast length on remaining path to be as narrow as possible
3719 if (map()->find_edge(length) >= 0) {
3720 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3721 if (ccast != length) {
3722 _gvn.set_type_bottom(ccast);
3723 record_for_igvn(ccast);
3724 replace_in_map(length, ccast);
3725 }
3726 }
3727
3728 return javaoop;
3729 }
3730
3731 // The following "Ideal_foo" functions are placed here because they recognize
3732 // the graph shapes created by the functions immediately above.
3733
3734 //---------------------------Ideal_allocation----------------------------------
3735 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
Ideal_allocation(Node * ptr,PhaseTransform * phase)3736 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3737 if (ptr == NULL) { // reduce dumb test in callers
3738 return NULL;
3739 }
3740
3741 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
3742 ptr = bs->step_over_gc_barrier(ptr);
3743
3744 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3745 ptr = ptr->in(1);
3746 if (ptr == NULL) return NULL;
3747 }
3748 // Return NULL for allocations with several casts:
3749 // j.l.reflect.Array.newInstance(jobject, jint)
3750 // Object.clone()
3751 // to keep more precise type from last cast.
3752 if (ptr->is_Proj()) {
3753 Node* allo = ptr->in(0);
3754 if (allo != NULL && allo->is_Allocate()) {
3755 return allo->as_Allocate();
3756 }
3757 }
3758 // Report failure to match.
3759 return NULL;
3760 }
3761
3762 // Fancy version which also strips off an offset (and reports it to caller).
Ideal_allocation(Node * ptr,PhaseTransform * phase,intptr_t & offset)3763 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3764 intptr_t& offset) {
3765 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3766 if (base == NULL) return NULL;
3767 return Ideal_allocation(base, phase);
3768 }
3769
3770 // Trace Initialize <- Proj[Parm] <- Allocate
allocation()3771 AllocateNode* InitializeNode::allocation() {
3772 Node* rawoop = in(InitializeNode::RawAddress);
3773 if (rawoop->is_Proj()) {
3774 Node* alloc = rawoop->in(0);
3775 if (alloc->is_Allocate()) {
3776 return alloc->as_Allocate();
3777 }
3778 }
3779 return NULL;
3780 }
3781
3782 // Trace Allocate -> Proj[Parm] -> Initialize
initialization()3783 InitializeNode* AllocateNode::initialization() {
3784 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
3785 if (rawoop == NULL) return NULL;
3786 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3787 Node* init = rawoop->fast_out(i);
3788 if (init->is_Initialize()) {
3789 assert(init->as_Initialize()->allocation() == this, "2-way link");
3790 return init->as_Initialize();
3791 }
3792 }
3793 return NULL;
3794 }
3795
3796 //----------------------------- loop predicates ---------------------------
3797
3798 //------------------------------add_predicate_impl----------------------------
add_predicate_impl(Deoptimization::DeoptReason reason,int nargs)3799 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3800 // Too many traps seen?
3801 if (too_many_traps(reason)) {
3802 #ifdef ASSERT
3803 if (TraceLoopPredicate) {
3804 int tc = C->trap_count(reason);
3805 tty->print("too many traps=%s tcount=%d in ",
3806 Deoptimization::trap_reason_name(reason), tc);
3807 method()->print(); // which method has too many predicate traps
3808 tty->cr();
3809 }
3810 #endif
3811 // We cannot afford to take more traps here,
3812 // do not generate predicate.
3813 return;
3814 }
3815
3816 Node *cont = _gvn.intcon(1);
3817 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
3818 Node *bol = _gvn.transform(new Conv2BNode(opq));
3819 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3820 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3821 C->add_predicate_opaq(opq);
3822 {
3823 PreserveJVMState pjvms(this);
3824 set_control(iffalse);
3825 inc_sp(nargs);
3826 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3827 }
3828 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3829 set_control(iftrue);
3830 }
3831
3832 //------------------------------add_predicate---------------------------------
add_predicate(int nargs)3833 void GraphKit::add_predicate(int nargs) {
3834 if (UseLoopPredicate) {
3835 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3836 }
3837 if (UseProfiledLoopPredicate) {
3838 add_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
3839 }
3840 // loop's limit check predicate should be near the loop.
3841 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3842 }
3843
sync_kit(IdealKit & ideal)3844 void GraphKit::sync_kit(IdealKit& ideal) {
3845 set_all_memory(ideal.merged_memory());
3846 set_i_o(ideal.i_o());
3847 set_control(ideal.ctrl());
3848 }
3849
final_sync(IdealKit & ideal)3850 void GraphKit::final_sync(IdealKit& ideal) {
3851 // Final sync IdealKit and graphKit.
3852 sync_kit(ideal);
3853 }
3854
load_String_length(Node * str,bool set_ctrl)3855 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
3856 Node* len = load_array_length(load_String_value(str, set_ctrl));
3857 Node* coder = load_String_coder(str, set_ctrl);
3858 // Divide length by 2 if coder is UTF16
3859 return _gvn.transform(new RShiftINode(len, coder));
3860 }
3861
load_String_value(Node * str,bool set_ctrl)3862 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
3863 int value_offset = java_lang_String::value_offset_in_bytes();
3864 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3865 false, NULL, 0);
3866 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3867 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
3868 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
3869 ciTypeArrayKlass::make(T_BYTE), true, 0);
3870 Node* p = basic_plus_adr(str, str, value_offset);
3871 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
3872 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
3873 // String.value field is known to be @Stable.
3874 if (UseImplicitStableValues) {
3875 load = cast_array_to_stable(load, value_type);
3876 }
3877 return load;
3878 }
3879
load_String_coder(Node * str,bool set_ctrl)3880 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
3881 if (!CompactStrings) {
3882 return intcon(java_lang_String::CODER_UTF16);
3883 }
3884 int coder_offset = java_lang_String::coder_offset_in_bytes();
3885 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3886 false, NULL, 0);
3887 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
3888 int coder_field_idx = C->get_alias_index(coder_field_type);
3889
3890 Node* p = basic_plus_adr(str, str, coder_offset);
3891 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
3892 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
3893 return load;
3894 }
3895
store_String_value(Node * str,Node * value)3896 void GraphKit::store_String_value(Node* str, Node* value) {
3897 int value_offset = java_lang_String::value_offset_in_bytes();
3898 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3899 false, NULL, 0);
3900 const TypePtr* value_field_type = string_type->add_offset(value_offset);
3901
3902 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
3903 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
3904 }
3905
store_String_coder(Node * str,Node * value)3906 void GraphKit::store_String_coder(Node* str, Node* value) {
3907 int coder_offset = java_lang_String::coder_offset_in_bytes();
3908 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
3909 false, NULL, 0);
3910 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
3911
3912 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
3913 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
3914 }
3915
3916 // Capture src and dst memory state with a MergeMemNode
capture_memory(const TypePtr * src_type,const TypePtr * dst_type)3917 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
3918 if (src_type == dst_type) {
3919 // Types are equal, we don't need a MergeMemNode
3920 return memory(src_type);
3921 }
3922 MergeMemNode* merge = MergeMemNode::make(map()->memory());
3923 record_for_igvn(merge); // fold it up later, if possible
3924 int src_idx = C->get_alias_index(src_type);
3925 int dst_idx = C->get_alias_index(dst_type);
3926 merge->set_memory_at(src_idx, memory(src_idx));
3927 merge->set_memory_at(dst_idx, memory(dst_idx));
3928 return merge;
3929 }
3930
compress_string(Node * src,const TypeAryPtr * src_type,Node * dst,Node * count)3931 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
3932 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
3933 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
3934 // If input and output memory types differ, capture both states to preserve
3935 // the dependency between preceding and subsequent loads/stores.
3936 // For example, the following program:
3937 // StoreB
3938 // compress_string
3939 // LoadB
3940 // has this memory graph (use->def):
3941 // LoadB -> compress_string -> CharMem
3942 // ... -> StoreB -> ByteMem
3943 // The intrinsic hides the dependency between LoadB and StoreB, causing
3944 // the load to read from memory not containing the result of the StoreB.
3945 // The correct memory graph should look like this:
3946 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
3947 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
3948 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
3949 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
3950 set_memory(res_mem, TypeAryPtr::BYTES);
3951 return str;
3952 }
3953
inflate_string(Node * src,Node * dst,const TypeAryPtr * dst_type,Node * count)3954 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
3955 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
3956 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
3957 // Capture src and dst memory (see comment in 'compress_string').
3958 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
3959 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
3960 set_memory(_gvn.transform(str), dst_type);
3961 }
3962
inflate_string_slow(Node * src,Node * dst,Node * start,Node * count)3963 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
3964 /**
3965 * int i_char = start;
3966 * for (int i_byte = 0; i_byte < count; i_byte++) {
3967 * dst[i_char++] = (char)(src[i_byte] & 0xff);
3968 * }
3969 */
3970 src = access_resolve(src, ACCESS_READ);
3971 dst = access_resolve(dst, ACCESS_WRITE);
3972 add_predicate();
3973 RegionNode* head = new RegionNode(3);
3974 head->init_req(1, control());
3975 gvn().set_type(head, Type::CONTROL);
3976 record_for_igvn(head);
3977
3978 Node* i_byte = new PhiNode(head, TypeInt::INT);
3979 i_byte->init_req(1, intcon(0));
3980 gvn().set_type(i_byte, TypeInt::INT);
3981 record_for_igvn(i_byte);
3982
3983 Node* i_char = new PhiNode(head, TypeInt::INT);
3984 i_char->init_req(1, start);
3985 gvn().set_type(i_char, TypeInt::INT);
3986 record_for_igvn(i_char);
3987
3988 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
3989 gvn().set_type(mem, Type::MEMORY);
3990 record_for_igvn(mem);
3991 set_control(head);
3992 set_memory(mem, TypeAryPtr::BYTES);
3993 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
3994 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
3995 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
3996 false, false, true /* mismatched */);
3997
3998 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
3999 head->init_req(2, IfTrue(iff));
4000 mem->init_req(2, st);
4001 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4002 i_char->init_req(2, AddI(i_char, intcon(2)));
4003
4004 set_control(IfFalse(iff));
4005 set_memory(st, TypeAryPtr::BYTES);
4006 }
4007
make_constant_from_field(ciField * field,Node * obj)4008 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4009 if (!field->is_constant()) {
4010 return NULL; // Field not marked as constant.
4011 }
4012 ciInstance* holder = NULL;
4013 if (!field->is_static()) {
4014 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4015 if (const_oop != NULL && const_oop->is_instance()) {
4016 holder = const_oop->as_instance();
4017 }
4018 }
4019 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4020 /*is_unsigned_load=*/false);
4021 if (con_type != NULL) {
4022 return makecon(con_type);
4023 }
4024 return NULL;
4025 }
4026
cast_array_to_stable(Node * ary,const TypeAryPtr * ary_type)4027 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4028 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4029 // assumption of CCP analysis.
4030 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4031 }
4032