1 /*
2 * Copyright (c) 2001, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "ci/ciNativeEntryPoint.hpp"
29 #include "ci/ciObjArray.hpp"
30 #include "asm/register.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "gc/shared/barrierSet.hpp"
33 #include "gc/shared/c2/barrierSetC2.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "memory/resourceArea.hpp"
36 #include "opto/addnode.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/graphKit.hpp"
40 #include "opto/idealKit.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/machnode.hpp"
44 #include "opto/opaquenode.hpp"
45 #include "opto/parse.hpp"
46 #include "opto/rootnode.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/subtypenode.hpp"
49 #include "runtime/deoptimization.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #include "utilities/bitMap.inline.hpp"
52 #include "utilities/powerOfTwo.hpp"
53 #include "utilities/growableArray.hpp"
54
55 //----------------------------GraphKit-----------------------------------------
56 // Main utility constructor.
GraphKit(JVMState * jvms)57 GraphKit::GraphKit(JVMState* jvms)
58 : Phase(Phase::Parser),
59 _env(C->env()),
60 _gvn(*C->initial_gvn()),
61 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
62 {
63 _exceptions = jvms->map()->next_exception();
64 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
65 set_jvms(jvms);
66 }
67
68 // Private constructor for parser.
GraphKit()69 GraphKit::GraphKit()
70 : Phase(Phase::Parser),
71 _env(C->env()),
72 _gvn(*C->initial_gvn()),
73 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
74 {
75 _exceptions = NULL;
76 set_map(NULL);
77 debug_only(_sp = -99);
78 debug_only(set_bci(-99));
79 }
80
81
82
83 //---------------------------clean_stack---------------------------------------
84 // Clear away rubbish from the stack area of the JVM state.
85 // This destroys any arguments that may be waiting on the stack.
clean_stack(int from_sp)86 void GraphKit::clean_stack(int from_sp) {
87 SafePointNode* map = this->map();
88 JVMState* jvms = this->jvms();
89 int stk_size = jvms->stk_size();
90 int stkoff = jvms->stkoff();
91 Node* top = this->top();
92 for (int i = from_sp; i < stk_size; i++) {
93 if (map->in(stkoff + i) != top) {
94 map->set_req(stkoff + i, top);
95 }
96 }
97 }
98
99
100 //--------------------------------sync_jvms-----------------------------------
101 // Make sure our current jvms agrees with our parse state.
sync_jvms() const102 JVMState* GraphKit::sync_jvms() const {
103 JVMState* jvms = this->jvms();
104 jvms->set_bci(bci()); // Record the new bci in the JVMState
105 jvms->set_sp(sp()); // Record the new sp in the JVMState
106 assert(jvms_in_sync(), "jvms is now in sync");
107 return jvms;
108 }
109
110 //--------------------------------sync_jvms_for_reexecute---------------------
111 // Make sure our current jvms agrees with our parse state. This version
112 // uses the reexecute_sp for reexecuting bytecodes.
sync_jvms_for_reexecute()113 JVMState* GraphKit::sync_jvms_for_reexecute() {
114 JVMState* jvms = this->jvms();
115 jvms->set_bci(bci()); // Record the new bci in the JVMState
116 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
117 return jvms;
118 }
119
120 #ifdef ASSERT
jvms_in_sync() const121 bool GraphKit::jvms_in_sync() const {
122 Parse* parse = is_Parse();
123 if (parse == NULL) {
124 if (bci() != jvms()->bci()) return false;
125 if (sp() != (int)jvms()->sp()) return false;
126 return true;
127 }
128 if (jvms()->method() != parse->method()) return false;
129 if (jvms()->bci() != parse->bci()) return false;
130 int jvms_sp = jvms()->sp();
131 if (jvms_sp != parse->sp()) return false;
132 int jvms_depth = jvms()->depth();
133 if (jvms_depth != parse->depth()) return false;
134 return true;
135 }
136
137 // Local helper checks for special internal merge points
138 // used to accumulate and merge exception states.
139 // They are marked by the region's in(0) edge being the map itself.
140 // Such merge points must never "escape" into the parser at large,
141 // until they have been handed to gvn.transform.
is_hidden_merge(Node * reg)142 static bool is_hidden_merge(Node* reg) {
143 if (reg == NULL) return false;
144 if (reg->is_Phi()) {
145 reg = reg->in(0);
146 if (reg == NULL) return false;
147 }
148 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
149 }
150
verify_map() const151 void GraphKit::verify_map() const {
152 if (map() == NULL) return; // null map is OK
153 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
154 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
155 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
156 }
157
verify_exception_state(SafePointNode * ex_map)158 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
159 assert(ex_map->next_exception() == NULL, "not already part of a chain");
160 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
161 }
162 #endif
163
164 //---------------------------stop_and_kill_map---------------------------------
165 // Set _map to NULL, signalling a stop to further bytecode execution.
166 // First smash the current map's control to a constant, to mark it dead.
stop_and_kill_map()167 void GraphKit::stop_and_kill_map() {
168 SafePointNode* dead_map = stop();
169 if (dead_map != NULL) {
170 dead_map->disconnect_inputs(C); // Mark the map as killed.
171 assert(dead_map->is_killed(), "must be so marked");
172 }
173 }
174
175
176 //--------------------------------stopped--------------------------------------
177 // Tell if _map is NULL, or control is top.
stopped()178 bool GraphKit::stopped() {
179 if (map() == NULL) return true;
180 else if (control() == top()) return true;
181 else return false;
182 }
183
184
185 //-----------------------------has_ex_handler----------------------------------
186 // Tell if this method or any caller method has exception handlers.
has_ex_handler()187 bool GraphKit::has_ex_handler() {
188 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
189 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
190 return true;
191 }
192 }
193 return false;
194 }
195
196 //------------------------------save_ex_oop------------------------------------
197 // Save an exception without blowing stack contents or other JVM state.
set_saved_ex_oop(SafePointNode * ex_map,Node * ex_oop)198 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
199 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
200 ex_map->add_req(ex_oop);
201 debug_only(verify_exception_state(ex_map));
202 }
203
common_saved_ex_oop(SafePointNode * ex_map,bool clear_it)204 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
205 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
206 Node* ex_oop = ex_map->in(ex_map->req()-1);
207 if (clear_it) ex_map->del_req(ex_map->req()-1);
208 return ex_oop;
209 }
210
211 //-----------------------------saved_ex_oop------------------------------------
212 // Recover a saved exception from its map.
saved_ex_oop(SafePointNode * ex_map)213 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
214 return common_saved_ex_oop(ex_map, false);
215 }
216
217 //--------------------------clear_saved_ex_oop---------------------------------
218 // Erase a previously saved exception from its map.
clear_saved_ex_oop(SafePointNode * ex_map)219 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
220 return common_saved_ex_oop(ex_map, true);
221 }
222
223 #ifdef ASSERT
224 //---------------------------has_saved_ex_oop----------------------------------
225 // Erase a previously saved exception from its map.
has_saved_ex_oop(SafePointNode * ex_map)226 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
227 return ex_map->req() == ex_map->jvms()->endoff()+1;
228 }
229 #endif
230
231 //-------------------------make_exception_state--------------------------------
232 // Turn the current JVM state into an exception state, appending the ex_oop.
make_exception_state(Node * ex_oop)233 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
234 sync_jvms();
235 SafePointNode* ex_map = stop(); // do not manipulate this map any more
236 set_saved_ex_oop(ex_map, ex_oop);
237 return ex_map;
238 }
239
240
241 //--------------------------add_exception_state--------------------------------
242 // Add an exception to my list of exceptions.
add_exception_state(SafePointNode * ex_map)243 void GraphKit::add_exception_state(SafePointNode* ex_map) {
244 if (ex_map == NULL || ex_map->control() == top()) {
245 return;
246 }
247 #ifdef ASSERT
248 verify_exception_state(ex_map);
249 if (has_exceptions()) {
250 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
251 }
252 #endif
253
254 // If there is already an exception of exactly this type, merge with it.
255 // In particular, null-checks and other low-level exceptions common up here.
256 Node* ex_oop = saved_ex_oop(ex_map);
257 const Type* ex_type = _gvn.type(ex_oop);
258 if (ex_oop == top()) {
259 // No action needed.
260 return;
261 }
262 assert(ex_type->isa_instptr(), "exception must be an instance");
263 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
264 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
265 // We check sp also because call bytecodes can generate exceptions
266 // both before and after arguments are popped!
267 if (ex_type2 == ex_type
268 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
269 combine_exception_states(ex_map, e2);
270 return;
271 }
272 }
273
274 // No pre-existing exception of the same type. Chain it on the list.
275 push_exception_state(ex_map);
276 }
277
278 //-----------------------add_exception_states_from-----------------------------
add_exception_states_from(JVMState * jvms)279 void GraphKit::add_exception_states_from(JVMState* jvms) {
280 SafePointNode* ex_map = jvms->map()->next_exception();
281 if (ex_map != NULL) {
282 jvms->map()->set_next_exception(NULL);
283 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
284 next_map = ex_map->next_exception();
285 ex_map->set_next_exception(NULL);
286 add_exception_state(ex_map);
287 }
288 }
289 }
290
291 //-----------------------transfer_exceptions_into_jvms-------------------------
transfer_exceptions_into_jvms()292 JVMState* GraphKit::transfer_exceptions_into_jvms() {
293 if (map() == NULL) {
294 // We need a JVMS to carry the exceptions, but the map has gone away.
295 // Create a scratch JVMS, cloned from any of the exception states...
296 if (has_exceptions()) {
297 _map = _exceptions;
298 _map = clone_map();
299 _map->set_next_exception(NULL);
300 clear_saved_ex_oop(_map);
301 debug_only(verify_map());
302 } else {
303 // ...or created from scratch
304 JVMState* jvms = new (C) JVMState(_method, NULL);
305 jvms->set_bci(_bci);
306 jvms->set_sp(_sp);
307 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
308 set_jvms(jvms);
309 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
310 set_all_memory(top());
311 while (map()->req() < jvms->endoff()) map()->add_req(top());
312 }
313 // (This is a kludge, in case you didn't notice.)
314 set_control(top());
315 }
316 JVMState* jvms = sync_jvms();
317 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
318 jvms->map()->set_next_exception(_exceptions);
319 _exceptions = NULL; // done with this set of exceptions
320 return jvms;
321 }
322
add_n_reqs(Node * dstphi,Node * srcphi)323 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
324 assert(is_hidden_merge(dstphi), "must be a special merge node");
325 assert(is_hidden_merge(srcphi), "must be a special merge node");
326 uint limit = srcphi->req();
327 for (uint i = PhiNode::Input; i < limit; i++) {
328 dstphi->add_req(srcphi->in(i));
329 }
330 }
add_one_req(Node * dstphi,Node * src)331 static inline void add_one_req(Node* dstphi, Node* src) {
332 assert(is_hidden_merge(dstphi), "must be a special merge node");
333 assert(!is_hidden_merge(src), "must not be a special merge node");
334 dstphi->add_req(src);
335 }
336
337 //-----------------------combine_exception_states------------------------------
338 // This helper function combines exception states by building phis on a
339 // specially marked state-merging region. These regions and phis are
340 // untransformed, and can build up gradually. The region is marked by
341 // having a control input of its exception map, rather than NULL. Such
342 // regions do not appear except in this function, and in use_exception_state.
combine_exception_states(SafePointNode * ex_map,SafePointNode * phi_map)343 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
344 if (failing()) return; // dying anyway...
345 JVMState* ex_jvms = ex_map->_jvms;
346 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
347 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
348 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
349 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
350 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
351 assert(ex_map->req() == phi_map->req(), "matching maps");
352 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
353 Node* hidden_merge_mark = root();
354 Node* region = phi_map->control();
355 MergeMemNode* phi_mem = phi_map->merged_memory();
356 MergeMemNode* ex_mem = ex_map->merged_memory();
357 if (region->in(0) != hidden_merge_mark) {
358 // The control input is not (yet) a specially-marked region in phi_map.
359 // Make it so, and build some phis.
360 region = new RegionNode(2);
361 _gvn.set_type(region, Type::CONTROL);
362 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
363 region->init_req(1, phi_map->control());
364 phi_map->set_control(region);
365 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
366 record_for_igvn(io_phi);
367 _gvn.set_type(io_phi, Type::ABIO);
368 phi_map->set_i_o(io_phi);
369 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
370 Node* m = mms.memory();
371 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
372 record_for_igvn(m_phi);
373 _gvn.set_type(m_phi, Type::MEMORY);
374 mms.set_memory(m_phi);
375 }
376 }
377
378 // Either or both of phi_map and ex_map might already be converted into phis.
379 Node* ex_control = ex_map->control();
380 // if there is special marking on ex_map also, we add multiple edges from src
381 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
382 // how wide was the destination phi_map, originally?
383 uint orig_width = region->req();
384
385 if (add_multiple) {
386 add_n_reqs(region, ex_control);
387 add_n_reqs(phi_map->i_o(), ex_map->i_o());
388 } else {
389 // ex_map has no merges, so we just add single edges everywhere
390 add_one_req(region, ex_control);
391 add_one_req(phi_map->i_o(), ex_map->i_o());
392 }
393 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
394 if (mms.is_empty()) {
395 // get a copy of the base memory, and patch some inputs into it
396 const TypePtr* adr_type = mms.adr_type(C);
397 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
398 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
399 mms.set_memory(phi);
400 // Prepare to append interesting stuff onto the newly sliced phi:
401 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
402 }
403 // Append stuff from ex_map:
404 if (add_multiple) {
405 add_n_reqs(mms.memory(), mms.memory2());
406 } else {
407 add_one_req(mms.memory(), mms.memory2());
408 }
409 }
410 uint limit = ex_map->req();
411 for (uint i = TypeFunc::Parms; i < limit; i++) {
412 // Skip everything in the JVMS after tos. (The ex_oop follows.)
413 if (i == tos) i = ex_jvms->monoff();
414 Node* src = ex_map->in(i);
415 Node* dst = phi_map->in(i);
416 if (src != dst) {
417 PhiNode* phi;
418 if (dst->in(0) != region) {
419 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
420 record_for_igvn(phi);
421 _gvn.set_type(phi, phi->type());
422 phi_map->set_req(i, dst);
423 // Prepare to append interesting stuff onto the new phi:
424 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
425 } else {
426 assert(dst->is_Phi(), "nobody else uses a hidden region");
427 phi = dst->as_Phi();
428 }
429 if (add_multiple && src->in(0) == ex_control) {
430 // Both are phis.
431 add_n_reqs(dst, src);
432 } else {
433 while (dst->req() < region->req()) add_one_req(dst, src);
434 }
435 const Type* srctype = _gvn.type(src);
436 if (phi->type() != srctype) {
437 const Type* dsttype = phi->type()->meet_speculative(srctype);
438 if (phi->type() != dsttype) {
439 phi->set_type(dsttype);
440 _gvn.set_type(phi, dsttype);
441 }
442 }
443 }
444 }
445 phi_map->merge_replaced_nodes_with(ex_map);
446 }
447
448 //--------------------------use_exception_state--------------------------------
use_exception_state(SafePointNode * phi_map)449 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
450 if (failing()) { stop(); return top(); }
451 Node* region = phi_map->control();
452 Node* hidden_merge_mark = root();
453 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
454 Node* ex_oop = clear_saved_ex_oop(phi_map);
455 if (region->in(0) == hidden_merge_mark) {
456 // Special marking for internal ex-states. Process the phis now.
457 region->set_req(0, region); // now it's an ordinary region
458 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
459 // Note: Setting the jvms also sets the bci and sp.
460 set_control(_gvn.transform(region));
461 uint tos = jvms()->stkoff() + sp();
462 for (uint i = 1; i < tos; i++) {
463 Node* x = phi_map->in(i);
464 if (x->in(0) == region) {
465 assert(x->is_Phi(), "expected a special phi");
466 phi_map->set_req(i, _gvn.transform(x));
467 }
468 }
469 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
470 Node* x = mms.memory();
471 if (x->in(0) == region) {
472 assert(x->is_Phi(), "nobody else uses a hidden region");
473 mms.set_memory(_gvn.transform(x));
474 }
475 }
476 if (ex_oop->in(0) == region) {
477 assert(ex_oop->is_Phi(), "expected a special phi");
478 ex_oop = _gvn.transform(ex_oop);
479 }
480 } else {
481 set_jvms(phi_map->jvms());
482 }
483
484 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
485 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
486 return ex_oop;
487 }
488
489 //---------------------------------java_bc-------------------------------------
java_bc() const490 Bytecodes::Code GraphKit::java_bc() const {
491 ciMethod* method = this->method();
492 int bci = this->bci();
493 if (method != NULL && bci != InvocationEntryBci)
494 return method->java_code_at_bci(bci);
495 else
496 return Bytecodes::_illegal;
497 }
498
uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,bool must_throw)499 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
500 bool must_throw) {
501 // if the exception capability is set, then we will generate code
502 // to check the JavaThread.should_post_on_exceptions flag to see
503 // if we actually need to report exception events (for this
504 // thread). If we don't need to report exception events, we will
505 // take the normal fast path provided by add_exception_events. If
506 // exception event reporting is enabled for this thread, we will
507 // take the uncommon_trap in the BuildCutout below.
508
509 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
510 Node* jthread = _gvn.transform(new ThreadLocalNode());
511 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
512 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
513
514 // Test the should_post_on_exceptions_flag vs. 0
515 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
516 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
517
518 // Branch to slow_path if should_post_on_exceptions_flag was true
519 { BuildCutout unless(this, tst, PROB_MAX);
520 // Do not try anything fancy if we're notifying the VM on every throw.
521 // Cf. case Bytecodes::_athrow in parse2.cpp.
522 uncommon_trap(reason, Deoptimization::Action_none,
523 (ciKlass*)NULL, (char*)NULL, must_throw);
524 }
525
526 }
527
528 //------------------------------builtin_throw----------------------------------
builtin_throw(Deoptimization::DeoptReason reason,Node * arg)529 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
530 bool must_throw = true;
531
532 if (env()->jvmti_can_post_on_exceptions()) {
533 // check if we must post exception events, take uncommon trap if so
534 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
535 // here if should_post_on_exceptions is false
536 // continue on with the normal codegen
537 }
538
539 // If this particular condition has not yet happened at this
540 // bytecode, then use the uncommon trap mechanism, and allow for
541 // a future recompilation if several traps occur here.
542 // If the throw is hot, try to use a more complicated inline mechanism
543 // which keeps execution inside the compiled code.
544 bool treat_throw_as_hot = false;
545 ciMethodData* md = method()->method_data();
546
547 if (ProfileTraps) {
548 if (too_many_traps(reason)) {
549 treat_throw_as_hot = true;
550 }
551 // (If there is no MDO at all, assume it is early in
552 // execution, and that any deopts are part of the
553 // startup transient, and don't need to be remembered.)
554
555 // Also, if there is a local exception handler, treat all throws
556 // as hot if there has been at least one in this method.
557 if (C->trap_count(reason) != 0
558 && method()->method_data()->trap_count(reason) != 0
559 && has_ex_handler()) {
560 treat_throw_as_hot = true;
561 }
562 }
563
564 // If this throw happens frequently, an uncommon trap might cause
565 // a performance pothole. If there is a local exception handler,
566 // and if this particular bytecode appears to be deoptimizing often,
567 // let us handle the throw inline, with a preconstructed instance.
568 // Note: If the deopt count has blown up, the uncommon trap
569 // runtime is going to flush this nmethod, not matter what.
570 if (treat_throw_as_hot
571 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
572 // If the throw is local, we use a pre-existing instance and
573 // punt on the backtrace. This would lead to a missing backtrace
574 // (a repeat of 4292742) if the backtrace object is ever asked
575 // for its backtrace.
576 // Fixing this remaining case of 4292742 requires some flavor of
577 // escape analysis. Leave that for the future.
578 ciInstance* ex_obj = NULL;
579 switch (reason) {
580 case Deoptimization::Reason_null_check:
581 ex_obj = env()->NullPointerException_instance();
582 break;
583 case Deoptimization::Reason_div0_check:
584 ex_obj = env()->ArithmeticException_instance();
585 break;
586 case Deoptimization::Reason_range_check:
587 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
588 break;
589 case Deoptimization::Reason_class_check:
590 if (java_bc() == Bytecodes::_aastore) {
591 ex_obj = env()->ArrayStoreException_instance();
592 } else {
593 ex_obj = env()->ClassCastException_instance();
594 }
595 break;
596 default:
597 break;
598 }
599 if (failing()) { stop(); return; } // exception allocation might fail
600 if (ex_obj != NULL) {
601 // Cheat with a preallocated exception object.
602 if (C->log() != NULL)
603 C->log()->elem("hot_throw preallocated='1' reason='%s'",
604 Deoptimization::trap_reason_name(reason));
605 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
606 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
607
608 // Clear the detail message of the preallocated exception object.
609 // Weblogic sometimes mutates the detail message of exceptions
610 // using reflection.
611 int offset = java_lang_Throwable::get_detailMessage_offset();
612 const TypePtr* adr_typ = ex_con->add_offset(offset);
613
614 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
615 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
616 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
617
618 add_exception_state(make_exception_state(ex_node));
619 return;
620 }
621 }
622
623 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
624 // It won't be much cheaper than bailing to the interp., since we'll
625 // have to pass up all the debug-info, and the runtime will have to
626 // create the stack trace.
627
628 // Usual case: Bail to interpreter.
629 // Reserve the right to recompile if we haven't seen anything yet.
630
631 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
632 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
633 if (treat_throw_as_hot
634 && (method()->method_data()->trap_recompiled_at(bci(), m)
635 || C->too_many_traps(reason))) {
636 // We cannot afford to take more traps here. Suffer in the interpreter.
637 if (C->log() != NULL)
638 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
639 Deoptimization::trap_reason_name(reason),
640 C->trap_count(reason));
641 action = Deoptimization::Action_none;
642 }
643
644 // "must_throw" prunes the JVM state to include only the stack, if there
645 // are no local exception handlers. This should cut down on register
646 // allocation time and code size, by drastically reducing the number
647 // of in-edges on the call to the uncommon trap.
648
649 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
650 }
651
652
653 //----------------------------PreserveJVMState---------------------------------
PreserveJVMState(GraphKit * kit,bool clone_map)654 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
655 debug_only(kit->verify_map());
656 _kit = kit;
657 _map = kit->map(); // preserve the map
658 _sp = kit->sp();
659 kit->set_map(clone_map ? kit->clone_map() : NULL);
660 #ifdef ASSERT
661 _bci = kit->bci();
662 Parse* parser = kit->is_Parse();
663 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
664 _block = block;
665 #endif
666 }
~PreserveJVMState()667 PreserveJVMState::~PreserveJVMState() {
668 GraphKit* kit = _kit;
669 #ifdef ASSERT
670 assert(kit->bci() == _bci, "bci must not shift");
671 Parse* parser = kit->is_Parse();
672 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
673 assert(block == _block, "block must not shift");
674 #endif
675 kit->set_map(_map);
676 kit->set_sp(_sp);
677 }
678
679
680 //-----------------------------BuildCutout-------------------------------------
BuildCutout(GraphKit * kit,Node * p,float prob,float cnt)681 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
682 : PreserveJVMState(kit)
683 {
684 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
685 SafePointNode* outer_map = _map; // preserved map is caller's
686 SafePointNode* inner_map = kit->map();
687 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
688 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
689 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
690 }
~BuildCutout()691 BuildCutout::~BuildCutout() {
692 GraphKit* kit = _kit;
693 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
694 }
695
696 //---------------------------PreserveReexecuteState----------------------------
PreserveReexecuteState(GraphKit * kit)697 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
698 assert(!kit->stopped(), "must call stopped() before");
699 _kit = kit;
700 _sp = kit->sp();
701 _reexecute = kit->jvms()->_reexecute;
702 }
~PreserveReexecuteState()703 PreserveReexecuteState::~PreserveReexecuteState() {
704 if (_kit->stopped()) return;
705 _kit->jvms()->_reexecute = _reexecute;
706 _kit->set_sp(_sp);
707 }
708
709 //------------------------------clone_map--------------------------------------
710 // Implementation of PreserveJVMState
711 //
712 // Only clone_map(...) here. If this function is only used in the
713 // PreserveJVMState class we may want to get rid of this extra
714 // function eventually and do it all there.
715
clone_map()716 SafePointNode* GraphKit::clone_map() {
717 if (map() == NULL) return NULL;
718
719 // Clone the memory edge first
720 Node* mem = MergeMemNode::make(map()->memory());
721 gvn().set_type_bottom(mem);
722
723 SafePointNode *clonemap = (SafePointNode*)map()->clone();
724 JVMState* jvms = this->jvms();
725 JVMState* clonejvms = jvms->clone_shallow(C);
726 clonemap->set_memory(mem);
727 clonemap->set_jvms(clonejvms);
728 clonejvms->set_map(clonemap);
729 record_for_igvn(clonemap);
730 gvn().set_type_bottom(clonemap);
731 return clonemap;
732 }
733
734
735 //-----------------------------set_map_clone-----------------------------------
set_map_clone(SafePointNode * m)736 void GraphKit::set_map_clone(SafePointNode* m) {
737 _map = m;
738 _map = clone_map();
739 _map->set_next_exception(NULL);
740 debug_only(verify_map());
741 }
742
743
744 //----------------------------kill_dead_locals---------------------------------
745 // Detect any locals which are known to be dead, and force them to top.
kill_dead_locals()746 void GraphKit::kill_dead_locals() {
747 // Consult the liveness information for the locals. If any
748 // of them are unused, then they can be replaced by top(). This
749 // should help register allocation time and cut down on the size
750 // of the deoptimization information.
751
752 // This call is made from many of the bytecode handling
753 // subroutines called from the Big Switch in do_one_bytecode.
754 // Every bytecode which might include a slow path is responsible
755 // for killing its dead locals. The more consistent we
756 // are about killing deads, the fewer useless phis will be
757 // constructed for them at various merge points.
758
759 // bci can be -1 (InvocationEntryBci). We return the entry
760 // liveness for the method.
761
762 if (method() == NULL || method()->code_size() == 0) {
763 // We are building a graph for a call to a native method.
764 // All locals are live.
765 return;
766 }
767
768 ResourceMark rm;
769
770 // Consult the liveness information for the locals. If any
771 // of them are unused, then they can be replaced by top(). This
772 // should help register allocation time and cut down on the size
773 // of the deoptimization information.
774 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
775
776 int len = (int)live_locals.size();
777 assert(len <= jvms()->loc_size(), "too many live locals");
778 for (int local = 0; local < len; local++) {
779 if (!live_locals.at(local)) {
780 set_local(local, top());
781 }
782 }
783 }
784
785 #ifdef ASSERT
786 //-------------------------dead_locals_are_killed------------------------------
787 // Return true if all dead locals are set to top in the map.
788 // Used to assert "clean" debug info at various points.
dead_locals_are_killed()789 bool GraphKit::dead_locals_are_killed() {
790 if (method() == NULL || method()->code_size() == 0) {
791 // No locals need to be dead, so all is as it should be.
792 return true;
793 }
794
795 // Make sure somebody called kill_dead_locals upstream.
796 ResourceMark rm;
797 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
798 if (jvms->loc_size() == 0) continue; // no locals to consult
799 SafePointNode* map = jvms->map();
800 ciMethod* method = jvms->method();
801 int bci = jvms->bci();
802 if (jvms == this->jvms()) {
803 bci = this->bci(); // it might not yet be synched
804 }
805 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
806 int len = (int)live_locals.size();
807 if (!live_locals.is_valid() || len == 0)
808 // This method is trivial, or is poisoned by a breakpoint.
809 return true;
810 assert(len == jvms->loc_size(), "live map consistent with locals map");
811 for (int local = 0; local < len; local++) {
812 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
813 if (PrintMiscellaneous && (Verbose || WizardMode)) {
814 tty->print_cr("Zombie local %d: ", local);
815 jvms->dump();
816 }
817 return false;
818 }
819 }
820 }
821 return true;
822 }
823
824 #endif //ASSERT
825
826 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
should_reexecute_implied_by_bytecode(JVMState * jvms,bool is_anewarray)827 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
828 ciMethod* cur_method = jvms->method();
829 int cur_bci = jvms->bci();
830 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
831 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
832 return Interpreter::bytecode_should_reexecute(code) ||
833 (is_anewarray && code == Bytecodes::_multianewarray);
834 // Reexecute _multianewarray bytecode which was replaced with
835 // sequence of [a]newarray. See Parse::do_multianewarray().
836 //
837 // Note: interpreter should not have it set since this optimization
838 // is limited by dimensions and guarded by flag so in some cases
839 // multianewarray() runtime calls will be generated and
840 // the bytecode should not be reexecutes (stack will not be reset).
841 } else {
842 return false;
843 }
844 }
845
846 // Helper function for adding JVMState and debug information to node
add_safepoint_edges(SafePointNode * call,bool must_throw)847 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
848 // Add the safepoint edges to the call (or other safepoint).
849
850 // Make sure dead locals are set to top. This
851 // should help register allocation time and cut down on the size
852 // of the deoptimization information.
853 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
854
855 // Walk the inline list to fill in the correct set of JVMState's
856 // Also fill in the associated edges for each JVMState.
857
858 // If the bytecode needs to be reexecuted we need to put
859 // the arguments back on the stack.
860 const bool should_reexecute = jvms()->should_reexecute();
861 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
862
863 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
864 // undefined if the bci is different. This is normal for Parse but it
865 // should not happen for LibraryCallKit because only one bci is processed.
866 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
867 "in LibraryCallKit the reexecute bit should not change");
868
869 // If we are guaranteed to throw, we can prune everything but the
870 // input to the current bytecode.
871 bool can_prune_locals = false;
872 uint stack_slots_not_pruned = 0;
873 int inputs = 0, depth = 0;
874 if (must_throw) {
875 assert(method() == youngest_jvms->method(), "sanity");
876 if (compute_stack_effects(inputs, depth)) {
877 can_prune_locals = true;
878 stack_slots_not_pruned = inputs;
879 }
880 }
881
882 if (env()->should_retain_local_variables()) {
883 // At any safepoint, this method can get breakpointed, which would
884 // then require an immediate deoptimization.
885 can_prune_locals = false; // do not prune locals
886 stack_slots_not_pruned = 0;
887 }
888
889 // do not scribble on the input jvms
890 JVMState* out_jvms = youngest_jvms->clone_deep(C);
891 call->set_jvms(out_jvms); // Start jvms list for call node
892
893 // For a known set of bytecodes, the interpreter should reexecute them if
894 // deoptimization happens. We set the reexecute state for them here
895 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
896 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
897 #ifdef ASSERT
898 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
899 assert(method() == youngest_jvms->method(), "sanity");
900 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
901 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
902 #endif // ASSERT
903 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
904 }
905
906 // Presize the call:
907 DEBUG_ONLY(uint non_debug_edges = call->req());
908 call->add_req_batch(top(), youngest_jvms->debug_depth());
909 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
910
911 // Set up edges so that the call looks like this:
912 // Call [state:] ctl io mem fptr retadr
913 // [parms:] parm0 ... parmN
914 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
915 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
916 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
917 // Note that caller debug info precedes callee debug info.
918
919 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
920 uint debug_ptr = call->req();
921
922 // Loop over the map input edges associated with jvms, add them
923 // to the call node, & reset all offsets to match call node array.
924 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
925 uint debug_end = debug_ptr;
926 uint debug_start = debug_ptr - in_jvms->debug_size();
927 debug_ptr = debug_start; // back up the ptr
928
929 uint p = debug_start; // walks forward in [debug_start, debug_end)
930 uint j, k, l;
931 SafePointNode* in_map = in_jvms->map();
932 out_jvms->set_map(call);
933
934 if (can_prune_locals) {
935 assert(in_jvms->method() == out_jvms->method(), "sanity");
936 // If the current throw can reach an exception handler in this JVMS,
937 // then we must keep everything live that can reach that handler.
938 // As a quick and dirty approximation, we look for any handlers at all.
939 if (in_jvms->method()->has_exception_handlers()) {
940 can_prune_locals = false;
941 }
942 }
943
944 // Add the Locals
945 k = in_jvms->locoff();
946 l = in_jvms->loc_size();
947 out_jvms->set_locoff(p);
948 if (!can_prune_locals) {
949 for (j = 0; j < l; j++)
950 call->set_req(p++, in_map->in(k+j));
951 } else {
952 p += l; // already set to top above by add_req_batch
953 }
954
955 // Add the Expression Stack
956 k = in_jvms->stkoff();
957 l = in_jvms->sp();
958 out_jvms->set_stkoff(p);
959 if (!can_prune_locals) {
960 for (j = 0; j < l; j++)
961 call->set_req(p++, in_map->in(k+j));
962 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
963 // Divide stack into {S0,...,S1}, where S0 is set to top.
964 uint s1 = stack_slots_not_pruned;
965 stack_slots_not_pruned = 0; // for next iteration
966 if (s1 > l) s1 = l;
967 uint s0 = l - s1;
968 p += s0; // skip the tops preinstalled by add_req_batch
969 for (j = s0; j < l; j++)
970 call->set_req(p++, in_map->in(k+j));
971 } else {
972 p += l; // already set to top above by add_req_batch
973 }
974
975 // Add the Monitors
976 k = in_jvms->monoff();
977 l = in_jvms->mon_size();
978 out_jvms->set_monoff(p);
979 for (j = 0; j < l; j++)
980 call->set_req(p++, in_map->in(k+j));
981
982 // Copy any scalar object fields.
983 k = in_jvms->scloff();
984 l = in_jvms->scl_size();
985 out_jvms->set_scloff(p);
986 for (j = 0; j < l; j++)
987 call->set_req(p++, in_map->in(k+j));
988
989 // Finish the new jvms.
990 out_jvms->set_endoff(p);
991
992 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
993 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
994 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
995 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
996 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
997 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
998
999 // Update the two tail pointers in parallel.
1000 out_jvms = out_jvms->caller();
1001 in_jvms = in_jvms->caller();
1002 }
1003
1004 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1005
1006 // Test the correctness of JVMState::debug_xxx accessors:
1007 assert(call->jvms()->debug_start() == non_debug_edges, "");
1008 assert(call->jvms()->debug_end() == call->req(), "");
1009 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1010 }
1011
compute_stack_effects(int & inputs,int & depth)1012 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1013 Bytecodes::Code code = java_bc();
1014 if (code == Bytecodes::_wide) {
1015 code = method()->java_code_at_bci(bci() + 1);
1016 }
1017
1018 BasicType rtype = T_ILLEGAL;
1019 int rsize = 0;
1020
1021 if (code != Bytecodes::_illegal) {
1022 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1023 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1024 if (rtype < T_CONFLICT)
1025 rsize = type2size[rtype];
1026 }
1027
1028 switch (code) {
1029 case Bytecodes::_illegal:
1030 return false;
1031
1032 case Bytecodes::_ldc:
1033 case Bytecodes::_ldc_w:
1034 case Bytecodes::_ldc2_w:
1035 inputs = 0;
1036 break;
1037
1038 case Bytecodes::_dup: inputs = 1; break;
1039 case Bytecodes::_dup_x1: inputs = 2; break;
1040 case Bytecodes::_dup_x2: inputs = 3; break;
1041 case Bytecodes::_dup2: inputs = 2; break;
1042 case Bytecodes::_dup2_x1: inputs = 3; break;
1043 case Bytecodes::_dup2_x2: inputs = 4; break;
1044 case Bytecodes::_swap: inputs = 2; break;
1045 case Bytecodes::_arraylength: inputs = 1; break;
1046
1047 case Bytecodes::_getstatic:
1048 case Bytecodes::_putstatic:
1049 case Bytecodes::_getfield:
1050 case Bytecodes::_putfield:
1051 {
1052 bool ignored_will_link;
1053 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1054 int size = field->type()->size();
1055 bool is_get = (depth >= 0), is_static = (depth & 1);
1056 inputs = (is_static ? 0 : 1);
1057 if (is_get) {
1058 depth = size - inputs;
1059 } else {
1060 inputs += size; // putxxx pops the value from the stack
1061 depth = - inputs;
1062 }
1063 }
1064 break;
1065
1066 case Bytecodes::_invokevirtual:
1067 case Bytecodes::_invokespecial:
1068 case Bytecodes::_invokestatic:
1069 case Bytecodes::_invokedynamic:
1070 case Bytecodes::_invokeinterface:
1071 {
1072 bool ignored_will_link;
1073 ciSignature* declared_signature = NULL;
1074 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1075 assert(declared_signature != NULL, "cannot be null");
1076 inputs = declared_signature->arg_size_for_bc(code);
1077 int size = declared_signature->return_type()->size();
1078 depth = size - inputs;
1079 }
1080 break;
1081
1082 case Bytecodes::_multianewarray:
1083 {
1084 ciBytecodeStream iter(method());
1085 iter.reset_to_bci(bci());
1086 iter.next();
1087 inputs = iter.get_dimensions();
1088 assert(rsize == 1, "");
1089 depth = rsize - inputs;
1090 }
1091 break;
1092
1093 case Bytecodes::_ireturn:
1094 case Bytecodes::_lreturn:
1095 case Bytecodes::_freturn:
1096 case Bytecodes::_dreturn:
1097 case Bytecodes::_areturn:
1098 assert(rsize == -depth, "");
1099 inputs = rsize;
1100 break;
1101
1102 case Bytecodes::_jsr:
1103 case Bytecodes::_jsr_w:
1104 inputs = 0;
1105 depth = 1; // S.B. depth=1, not zero
1106 break;
1107
1108 default:
1109 // bytecode produces a typed result
1110 inputs = rsize - depth;
1111 assert(inputs >= 0, "");
1112 break;
1113 }
1114
1115 #ifdef ASSERT
1116 // spot check
1117 int outputs = depth + inputs;
1118 assert(outputs >= 0, "sanity");
1119 switch (code) {
1120 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1121 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1122 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1123 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1124 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1125 default: break;
1126 }
1127 #endif //ASSERT
1128
1129 return true;
1130 }
1131
1132
1133
1134 //------------------------------basic_plus_adr---------------------------------
basic_plus_adr(Node * base,Node * ptr,Node * offset)1135 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1136 // short-circuit a common case
1137 if (offset == intcon(0)) return ptr;
1138 return _gvn.transform( new AddPNode(base, ptr, offset) );
1139 }
1140
ConvI2L(Node * offset)1141 Node* GraphKit::ConvI2L(Node* offset) {
1142 // short-circuit a common case
1143 jint offset_con = find_int_con(offset, Type::OffsetBot);
1144 if (offset_con != Type::OffsetBot) {
1145 return longcon((jlong) offset_con);
1146 }
1147 return _gvn.transform( new ConvI2LNode(offset));
1148 }
1149
ConvI2UL(Node * offset)1150 Node* GraphKit::ConvI2UL(Node* offset) {
1151 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1152 if (offset_con != (juint) Type::OffsetBot) {
1153 return longcon((julong) offset_con);
1154 }
1155 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1156 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1157 return _gvn.transform( new AndLNode(conv, mask) );
1158 }
1159
ConvL2I(Node * offset)1160 Node* GraphKit::ConvL2I(Node* offset) {
1161 // short-circuit a common case
1162 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1163 if (offset_con != (jlong)Type::OffsetBot) {
1164 return intcon((int) offset_con);
1165 }
1166 return _gvn.transform( new ConvL2INode(offset));
1167 }
1168
1169 //-------------------------load_object_klass-----------------------------------
load_object_klass(Node * obj)1170 Node* GraphKit::load_object_klass(Node* obj) {
1171 // Special-case a fresh allocation to avoid building nodes:
1172 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1173 if (akls != NULL) return akls;
1174 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1175 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1176 }
1177
1178 //-------------------------load_array_length-----------------------------------
load_array_length(Node * array)1179 Node* GraphKit::load_array_length(Node* array) {
1180 // Special-case a fresh allocation to avoid building nodes:
1181 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1182 Node *alen;
1183 if (alloc == NULL) {
1184 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1185 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1186 } else {
1187 alen = alloc->Ideal_length();
1188 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1189 if (ccast != alen) {
1190 alen = _gvn.transform(ccast);
1191 }
1192 }
1193 return alen;
1194 }
1195
1196 //------------------------------do_null_check----------------------------------
1197 // Helper function to do a NULL pointer check. Returned value is
1198 // the incoming address with NULL casted away. You are allowed to use the
1199 // not-null value only if you are control dependent on the test.
1200 #ifndef PRODUCT
1201 extern int explicit_null_checks_inserted,
1202 explicit_null_checks_elided;
1203 #endif
null_check_common(Node * value,BasicType type,bool assert_null,Node ** null_control,bool speculative)1204 Node* GraphKit::null_check_common(Node* value, BasicType type,
1205 // optional arguments for variations:
1206 bool assert_null,
1207 Node* *null_control,
1208 bool speculative) {
1209 assert(!assert_null || null_control == NULL, "not both at once");
1210 if (stopped()) return top();
1211 NOT_PRODUCT(explicit_null_checks_inserted++);
1212
1213 // Construct NULL check
1214 Node *chk = NULL;
1215 switch(type) {
1216 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1217 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1218 case T_ARRAY : // fall through
1219 type = T_OBJECT; // simplify further tests
1220 case T_OBJECT : {
1221 const Type *t = _gvn.type( value );
1222
1223 const TypeOopPtr* tp = t->isa_oopptr();
1224 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1225 // Only for do_null_check, not any of its siblings:
1226 && !assert_null && null_control == NULL) {
1227 // Usually, any field access or invocation on an unloaded oop type
1228 // will simply fail to link, since the statically linked class is
1229 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1230 // the static class is loaded but the sharper oop type is not.
1231 // Rather than checking for this obscure case in lots of places,
1232 // we simply observe that a null check on an unloaded class
1233 // will always be followed by a nonsense operation, so we
1234 // can just issue the uncommon trap here.
1235 // Our access to the unloaded class will only be correct
1236 // after it has been loaded and initialized, which requires
1237 // a trip through the interpreter.
1238 #ifndef PRODUCT
1239 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1240 #endif
1241 uncommon_trap(Deoptimization::Reason_unloaded,
1242 Deoptimization::Action_reinterpret,
1243 tp->klass(), "!loaded");
1244 return top();
1245 }
1246
1247 if (assert_null) {
1248 // See if the type is contained in NULL_PTR.
1249 // If so, then the value is already null.
1250 if (t->higher_equal(TypePtr::NULL_PTR)) {
1251 NOT_PRODUCT(explicit_null_checks_elided++);
1252 return value; // Elided null assert quickly!
1253 }
1254 } else {
1255 // See if mixing in the NULL pointer changes type.
1256 // If so, then the NULL pointer was not allowed in the original
1257 // type. In other words, "value" was not-null.
1258 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1259 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1260 NOT_PRODUCT(explicit_null_checks_elided++);
1261 return value; // Elided null check quickly!
1262 }
1263 }
1264 chk = new CmpPNode( value, null() );
1265 break;
1266 }
1267
1268 default:
1269 fatal("unexpected type: %s", type2name(type));
1270 }
1271 assert(chk != NULL, "sanity check");
1272 chk = _gvn.transform(chk);
1273
1274 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1275 BoolNode *btst = new BoolNode( chk, btest);
1276 Node *tst = _gvn.transform( btst );
1277
1278 //-----------
1279 // if peephole optimizations occurred, a prior test existed.
1280 // If a prior test existed, maybe it dominates as we can avoid this test.
1281 if (tst != btst && type == T_OBJECT) {
1282 // At this point we want to scan up the CFG to see if we can
1283 // find an identical test (and so avoid this test altogether).
1284 Node *cfg = control();
1285 int depth = 0;
1286 while( depth < 16 ) { // Limit search depth for speed
1287 if( cfg->Opcode() == Op_IfTrue &&
1288 cfg->in(0)->in(1) == tst ) {
1289 // Found prior test. Use "cast_not_null" to construct an identical
1290 // CastPP (and hence hash to) as already exists for the prior test.
1291 // Return that casted value.
1292 if (assert_null) {
1293 replace_in_map(value, null());
1294 return null(); // do not issue the redundant test
1295 }
1296 Node *oldcontrol = control();
1297 set_control(cfg);
1298 Node *res = cast_not_null(value);
1299 set_control(oldcontrol);
1300 NOT_PRODUCT(explicit_null_checks_elided++);
1301 return res;
1302 }
1303 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1304 if (cfg == NULL) break; // Quit at region nodes
1305 depth++;
1306 }
1307 }
1308
1309 //-----------
1310 // Branch to failure if null
1311 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1312 Deoptimization::DeoptReason reason;
1313 if (assert_null) {
1314 reason = Deoptimization::reason_null_assert(speculative);
1315 } else if (type == T_OBJECT) {
1316 reason = Deoptimization::reason_null_check(speculative);
1317 } else {
1318 reason = Deoptimization::Reason_div0_check;
1319 }
1320 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1321 // ciMethodData::has_trap_at will return a conservative -1 if any
1322 // must-be-null assertion has failed. This could cause performance
1323 // problems for a method after its first do_null_assert failure.
1324 // Consider using 'Reason_class_check' instead?
1325
1326 // To cause an implicit null check, we set the not-null probability
1327 // to the maximum (PROB_MAX). For an explicit check the probability
1328 // is set to a smaller value.
1329 if (null_control != NULL || too_many_traps(reason)) {
1330 // probability is less likely
1331 ok_prob = PROB_LIKELY_MAG(3);
1332 } else if (!assert_null &&
1333 (ImplicitNullCheckThreshold > 0) &&
1334 method() != NULL &&
1335 (method()->method_data()->trap_count(reason)
1336 >= (uint)ImplicitNullCheckThreshold)) {
1337 ok_prob = PROB_LIKELY_MAG(3);
1338 }
1339
1340 if (null_control != NULL) {
1341 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1342 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1343 set_control( _gvn.transform( new IfTrueNode(iff)));
1344 #ifndef PRODUCT
1345 if (null_true == top()) {
1346 explicit_null_checks_elided++;
1347 }
1348 #endif
1349 (*null_control) = null_true;
1350 } else {
1351 BuildCutout unless(this, tst, ok_prob);
1352 // Check for optimizer eliding test at parse time
1353 if (stopped()) {
1354 // Failure not possible; do not bother making uncommon trap.
1355 NOT_PRODUCT(explicit_null_checks_elided++);
1356 } else if (assert_null) {
1357 uncommon_trap(reason,
1358 Deoptimization::Action_make_not_entrant,
1359 NULL, "assert_null");
1360 } else {
1361 replace_in_map(value, zerocon(type));
1362 builtin_throw(reason);
1363 }
1364 }
1365
1366 // Must throw exception, fall-thru not possible?
1367 if (stopped()) {
1368 return top(); // No result
1369 }
1370
1371 if (assert_null) {
1372 // Cast obj to null on this path.
1373 replace_in_map(value, zerocon(type));
1374 return zerocon(type);
1375 }
1376
1377 // Cast obj to not-null on this path, if there is no null_control.
1378 // (If there is a null_control, a non-null value may come back to haunt us.)
1379 if (type == T_OBJECT) {
1380 Node* cast = cast_not_null(value, false);
1381 if (null_control == NULL || (*null_control) == top())
1382 replace_in_map(value, cast);
1383 value = cast;
1384 }
1385
1386 return value;
1387 }
1388
1389
1390 //------------------------------cast_not_null----------------------------------
1391 // Cast obj to not-null on this path
cast_not_null(Node * obj,bool do_replace_in_map)1392 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1393 const Type *t = _gvn.type(obj);
1394 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1395 // Object is already not-null?
1396 if( t == t_not_null ) return obj;
1397
1398 Node *cast = new CastPPNode(obj,t_not_null);
1399 cast->init_req(0, control());
1400 cast = _gvn.transform( cast );
1401
1402 // Scan for instances of 'obj' in the current JVM mapping.
1403 // These instances are known to be not-null after the test.
1404 if (do_replace_in_map)
1405 replace_in_map(obj, cast);
1406
1407 return cast; // Return casted value
1408 }
1409
1410 // Sometimes in intrinsics, we implicitly know an object is not null
1411 // (there's no actual null check) so we can cast it to not null. In
1412 // the course of optimizations, the input to the cast can become null.
1413 // In that case that data path will die and we need the control path
1414 // to become dead as well to keep the graph consistent. So we have to
1415 // add a check for null for which one branch can't be taken. It uses
1416 // an Opaque4 node that will cause the check to be removed after loop
1417 // opts so the test goes away and the compiled code doesn't execute a
1418 // useless check.
must_be_not_null(Node * value,bool do_replace_in_map)1419 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1420 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1421 return value;
1422 }
1423 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1424 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1425 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1426 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1427 _gvn.set_type(iff, iff->Value(&_gvn));
1428 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1429 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1430 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1431 C->root()->add_req(halt);
1432 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1433 set_control(if_t);
1434 return cast_not_null(value, do_replace_in_map);
1435 }
1436
1437
1438 //--------------------------replace_in_map-------------------------------------
replace_in_map(Node * old,Node * neww)1439 void GraphKit::replace_in_map(Node* old, Node* neww) {
1440 if (old == neww) {
1441 return;
1442 }
1443
1444 map()->replace_edge(old, neww);
1445
1446 // Note: This operation potentially replaces any edge
1447 // on the map. This includes locals, stack, and monitors
1448 // of the current (innermost) JVM state.
1449
1450 // don't let inconsistent types from profiling escape this
1451 // method
1452
1453 const Type* told = _gvn.type(old);
1454 const Type* tnew = _gvn.type(neww);
1455
1456 if (!tnew->higher_equal(told)) {
1457 return;
1458 }
1459
1460 map()->record_replaced_node(old, neww);
1461 }
1462
1463
1464 //=============================================================================
1465 //--------------------------------memory---------------------------------------
memory(uint alias_idx)1466 Node* GraphKit::memory(uint alias_idx) {
1467 MergeMemNode* mem = merged_memory();
1468 Node* p = mem->memory_at(alias_idx);
1469 assert(p != mem->empty_memory(), "empty");
1470 _gvn.set_type(p, Type::MEMORY); // must be mapped
1471 return p;
1472 }
1473
1474 //-----------------------------reset_memory------------------------------------
reset_memory()1475 Node* GraphKit::reset_memory() {
1476 Node* mem = map()->memory();
1477 // do not use this node for any more parsing!
1478 debug_only( map()->set_memory((Node*)NULL) );
1479 return _gvn.transform( mem );
1480 }
1481
1482 //------------------------------set_all_memory---------------------------------
set_all_memory(Node * newmem)1483 void GraphKit::set_all_memory(Node* newmem) {
1484 Node* mergemem = MergeMemNode::make(newmem);
1485 gvn().set_type_bottom(mergemem);
1486 map()->set_memory(mergemem);
1487 }
1488
1489 //------------------------------set_all_memory_call----------------------------
set_all_memory_call(Node * call,bool separate_io_proj)1490 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1491 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1492 set_all_memory(newmem);
1493 }
1494
1495 //=============================================================================
1496 //
1497 // parser factory methods for MemNodes
1498 //
1499 // These are layered on top of the factory methods in LoadNode and StoreNode,
1500 // and integrate with the parser's memory state and _gvn engine.
1501 //
1502
1503 // factory methods in "int adr_idx"
make_load(Node * ctl,Node * adr,const Type * t,BasicType bt,int adr_idx,MemNode::MemOrd mo,LoadNode::ControlDependency control_dependency,bool require_atomic_access,bool unaligned,bool mismatched,bool unsafe,uint8_t barrier_data)1504 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1505 int adr_idx,
1506 MemNode::MemOrd mo,
1507 LoadNode::ControlDependency control_dependency,
1508 bool require_atomic_access,
1509 bool unaligned,
1510 bool mismatched,
1511 bool unsafe,
1512 uint8_t barrier_data) {
1513 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1514 const TypePtr* adr_type = NULL; // debug-mode-only argument
1515 debug_only(adr_type = C->get_adr_type(adr_idx));
1516 Node* mem = memory(adr_idx);
1517 Node* ld;
1518 if (require_atomic_access && bt == T_LONG) {
1519 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1520 } else if (require_atomic_access && bt == T_DOUBLE) {
1521 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1522 } else {
1523 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1524 }
1525 ld = _gvn.transform(ld);
1526 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1527 // Improve graph before escape analysis and boxing elimination.
1528 record_for_igvn(ld);
1529 }
1530 return ld;
1531 }
1532
store_to_memory(Node * ctl,Node * adr,Node * val,BasicType bt,int adr_idx,MemNode::MemOrd mo,bool require_atomic_access,bool unaligned,bool mismatched,bool unsafe)1533 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1534 int adr_idx,
1535 MemNode::MemOrd mo,
1536 bool require_atomic_access,
1537 bool unaligned,
1538 bool mismatched,
1539 bool unsafe) {
1540 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1541 const TypePtr* adr_type = NULL;
1542 debug_only(adr_type = C->get_adr_type(adr_idx));
1543 Node *mem = memory(adr_idx);
1544 Node* st;
1545 if (require_atomic_access && bt == T_LONG) {
1546 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1547 } else if (require_atomic_access && bt == T_DOUBLE) {
1548 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1549 } else {
1550 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1551 }
1552 if (unaligned) {
1553 st->as_Store()->set_unaligned_access();
1554 }
1555 if (mismatched) {
1556 st->as_Store()->set_mismatched_access();
1557 }
1558 if (unsafe) {
1559 st->as_Store()->set_unsafe_access();
1560 }
1561 st = _gvn.transform(st);
1562 set_memory(st, adr_idx);
1563 // Back-to-back stores can only remove intermediate store with DU info
1564 // so push on worklist for optimizer.
1565 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1566 record_for_igvn(st);
1567
1568 return st;
1569 }
1570
access_store_at(Node * obj,Node * adr,const TypePtr * adr_type,Node * val,const Type * val_type,BasicType bt,DecoratorSet decorators)1571 Node* GraphKit::access_store_at(Node* obj,
1572 Node* adr,
1573 const TypePtr* adr_type,
1574 Node* val,
1575 const Type* val_type,
1576 BasicType bt,
1577 DecoratorSet decorators) {
1578 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1579 // could be delayed during Parse (for example, in adjust_map_after_if()).
1580 // Execute transformation here to avoid barrier generation in such case.
1581 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1582 val = _gvn.makecon(TypePtr::NULL_PTR);
1583 }
1584
1585 if (stopped()) {
1586 return top(); // Dead path ?
1587 }
1588
1589 assert(val != NULL, "not dead path");
1590
1591 C2AccessValuePtr addr(adr, adr_type);
1592 C2AccessValue value(val, val_type);
1593 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1594 if (access.is_raw()) {
1595 return _barrier_set->BarrierSetC2::store_at(access, value);
1596 } else {
1597 return _barrier_set->store_at(access, value);
1598 }
1599 }
1600
access_load_at(Node * obj,Node * adr,const TypePtr * adr_type,const Type * val_type,BasicType bt,DecoratorSet decorators)1601 Node* GraphKit::access_load_at(Node* obj, // containing obj
1602 Node* adr, // actual adress to store val at
1603 const TypePtr* adr_type,
1604 const Type* val_type,
1605 BasicType bt,
1606 DecoratorSet decorators) {
1607 if (stopped()) {
1608 return top(); // Dead path ?
1609 }
1610
1611 C2AccessValuePtr addr(adr, adr_type);
1612 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1613 if (access.is_raw()) {
1614 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1615 } else {
1616 return _barrier_set->load_at(access, val_type);
1617 }
1618 }
1619
access_load(Node * adr,const Type * val_type,BasicType bt,DecoratorSet decorators)1620 Node* GraphKit::access_load(Node* adr, // actual adress to load val at
1621 const Type* val_type,
1622 BasicType bt,
1623 DecoratorSet decorators) {
1624 if (stopped()) {
1625 return top(); // Dead path ?
1626 }
1627
1628 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1629 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1630 if (access.is_raw()) {
1631 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1632 } else {
1633 return _barrier_set->load_at(access, val_type);
1634 }
1635 }
1636
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)1637 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1638 Node* adr,
1639 const TypePtr* adr_type,
1640 int alias_idx,
1641 Node* expected_val,
1642 Node* new_val,
1643 const Type* value_type,
1644 BasicType bt,
1645 DecoratorSet decorators) {
1646 C2AccessValuePtr addr(adr, adr_type);
1647 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1648 bt, obj, addr, alias_idx);
1649 if (access.is_raw()) {
1650 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1651 } else {
1652 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1653 }
1654 }
1655
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)1656 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1657 Node* adr,
1658 const TypePtr* adr_type,
1659 int alias_idx,
1660 Node* expected_val,
1661 Node* new_val,
1662 const Type* value_type,
1663 BasicType bt,
1664 DecoratorSet decorators) {
1665 C2AccessValuePtr addr(adr, adr_type);
1666 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1667 bt, obj, addr, alias_idx);
1668 if (access.is_raw()) {
1669 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1670 } else {
1671 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1672 }
1673 }
1674
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)1675 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1676 Node* adr,
1677 const TypePtr* adr_type,
1678 int alias_idx,
1679 Node* new_val,
1680 const Type* value_type,
1681 BasicType bt,
1682 DecoratorSet decorators) {
1683 C2AccessValuePtr addr(adr, adr_type);
1684 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1685 bt, obj, addr, alias_idx);
1686 if (access.is_raw()) {
1687 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1688 } else {
1689 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1690 }
1691 }
1692
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)1693 Node* GraphKit::access_atomic_add_at(Node* obj,
1694 Node* adr,
1695 const TypePtr* adr_type,
1696 int alias_idx,
1697 Node* new_val,
1698 const Type* value_type,
1699 BasicType bt,
1700 DecoratorSet decorators) {
1701 C2AccessValuePtr addr(adr, adr_type);
1702 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1703 if (access.is_raw()) {
1704 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1705 } else {
1706 return _barrier_set->atomic_add_at(access, new_val, value_type);
1707 }
1708 }
1709
access_clone(Node * src,Node * dst,Node * size,bool is_array)1710 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1711 return _barrier_set->clone(this, src, dst, size, is_array);
1712 }
1713
1714 //-------------------------array_element_address-------------------------
array_element_address(Node * ary,Node * idx,BasicType elembt,const TypeInt * sizetype,Node * ctrl)1715 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1716 const TypeInt* sizetype, Node* ctrl) {
1717 uint shift = exact_log2(type2aelembytes(elembt));
1718 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1719
1720 // short-circuit a common case (saves lots of confusing waste motion)
1721 jint idx_con = find_int_con(idx, -1);
1722 if (idx_con >= 0) {
1723 intptr_t offset = header + ((intptr_t)idx_con << shift);
1724 return basic_plus_adr(ary, offset);
1725 }
1726
1727 // must be correct type for alignment purposes
1728 Node* base = basic_plus_adr(ary, header);
1729 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1730 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1731 return basic_plus_adr(ary, base, scale);
1732 }
1733
1734 //-------------------------load_array_element-------------------------
load_array_element(Node * ctl,Node * ary,Node * idx,const TypeAryPtr * arytype)1735 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1736 const Type* elemtype = arytype->elem();
1737 BasicType elembt = elemtype->array_element_basic_type();
1738 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1739 if (elembt == T_NARROWOOP) {
1740 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1741 }
1742 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1743 return ld;
1744 }
1745
1746 //-------------------------set_arguments_for_java_call-------------------------
1747 // Arguments (pre-popped from the stack) are taken from the JVMS.
set_arguments_for_java_call(CallJavaNode * call)1748 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1749 // Add the call arguments:
1750 uint nargs = call->method()->arg_size();
1751 for (uint i = 0; i < nargs; i++) {
1752 Node* arg = argument(i);
1753 call->init_req(i + TypeFunc::Parms, arg);
1754 }
1755 }
1756
1757 //---------------------------set_edges_for_java_call---------------------------
1758 // Connect a newly created call into the current JVMS.
1759 // 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)1760 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1761
1762 // Add the predefined inputs:
1763 call->init_req( TypeFunc::Control, control() );
1764 call->init_req( TypeFunc::I_O , i_o() );
1765 call->init_req( TypeFunc::Memory , reset_memory() );
1766 call->init_req( TypeFunc::FramePtr, frameptr() );
1767 call->init_req( TypeFunc::ReturnAdr, top() );
1768
1769 add_safepoint_edges(call, must_throw);
1770
1771 Node* xcall = _gvn.transform(call);
1772
1773 if (xcall == top()) {
1774 set_control(top());
1775 return;
1776 }
1777 assert(xcall == call, "call identity is stable");
1778
1779 // Re-use the current map to produce the result.
1780
1781 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1782 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1783 set_all_memory_call(xcall, separate_io_proj);
1784
1785 //return xcall; // no need, caller already has it
1786 }
1787
set_results_for_java_call(CallJavaNode * call,bool separate_io_proj,bool deoptimize)1788 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1789 if (stopped()) return top(); // maybe the call folded up?
1790
1791 // Capture the return value, if any.
1792 Node* ret;
1793 if (call->method() == NULL ||
1794 call->method()->return_type()->basic_type() == T_VOID)
1795 ret = top();
1796 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1797
1798 // Note: Since any out-of-line call can produce an exception,
1799 // we always insert an I_O projection from the call into the result.
1800
1801 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1802
1803 if (separate_io_proj) {
1804 // The caller requested separate projections be used by the fall
1805 // through and exceptional paths, so replace the projections for
1806 // the fall through path.
1807 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1808 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1809 }
1810 return ret;
1811 }
1812
1813 //--------------------set_predefined_input_for_runtime_call--------------------
1814 // Reading and setting the memory state is way conservative here.
1815 // The real problem is that I am not doing real Type analysis on memory,
1816 // so I cannot distinguish card mark stores from other stores. Across a GC
1817 // point the Store Barrier and the card mark memory has to agree. I cannot
1818 // have a card mark store and its barrier split across the GC point from
1819 // either above or below. Here I get that to happen by reading ALL of memory.
1820 // A better answer would be to separate out card marks from other memory.
1821 // For now, return the input memory state, so that it can be reused
1822 // after the call, if this call has restricted memory effects.
set_predefined_input_for_runtime_call(SafePointNode * call,Node * narrow_mem)1823 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1824 // Set fixed predefined input arguments
1825 Node* memory = reset_memory();
1826 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1827 call->init_req( TypeFunc::Control, control() );
1828 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1829 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1830 call->init_req( TypeFunc::FramePtr, frameptr() );
1831 call->init_req( TypeFunc::ReturnAdr, top() );
1832 return memory;
1833 }
1834
1835 //-------------------set_predefined_output_for_runtime_call--------------------
1836 // Set control and memory (not i_o) from the call.
1837 // If keep_mem is not NULL, use it for the output state,
1838 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1839 // If hook_mem is NULL, this call produces no memory effects at all.
1840 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1841 // then only that memory slice is taken from the call.
1842 // In the last case, we must put an appropriate memory barrier before
1843 // the call, so as to create the correct anti-dependencies on loads
1844 // preceding the call.
set_predefined_output_for_runtime_call(Node * call,Node * keep_mem,const TypePtr * hook_mem)1845 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1846 Node* keep_mem,
1847 const TypePtr* hook_mem) {
1848 // no i/o
1849 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1850 if (keep_mem) {
1851 // First clone the existing memory state
1852 set_all_memory(keep_mem);
1853 if (hook_mem != NULL) {
1854 // Make memory for the call
1855 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1856 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1857 // We also use hook_mem to extract specific effects from arraycopy stubs.
1858 set_memory(mem, hook_mem);
1859 }
1860 // ...else the call has NO memory effects.
1861
1862 // Make sure the call advertises its memory effects precisely.
1863 // This lets us build accurate anti-dependences in gcm.cpp.
1864 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1865 "call node must be constructed correctly");
1866 } else {
1867 assert(hook_mem == NULL, "");
1868 // This is not a "slow path" call; all memory comes from the call.
1869 set_all_memory_call(call);
1870 }
1871 }
1872
1873 // Keep track of MergeMems feeding into other MergeMems
add_mergemem_users_to_worklist(Unique_Node_List & wl,Node * mem)1874 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
1875 if (!mem->is_MergeMem()) {
1876 return;
1877 }
1878 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
1879 Node* use = i.get();
1880 if (use->is_MergeMem()) {
1881 wl.push(use);
1882 }
1883 }
1884 }
1885
1886 // Replace the call with the current state of the kit.
replace_call(CallNode * call,Node * result,bool do_replaced_nodes)1887 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1888 JVMState* ejvms = NULL;
1889 if (has_exceptions()) {
1890 ejvms = transfer_exceptions_into_jvms();
1891 }
1892
1893 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1894 ReplacedNodes replaced_nodes_exception;
1895 Node* ex_ctl = top();
1896
1897 SafePointNode* final_state = stop();
1898
1899 // Find all the needed outputs of this call
1900 CallProjections callprojs;
1901 call->extract_projections(&callprojs, true);
1902
1903 Unique_Node_List wl;
1904 Node* init_mem = call->in(TypeFunc::Memory);
1905 Node* final_mem = final_state->in(TypeFunc::Memory);
1906 Node* final_ctl = final_state->in(TypeFunc::Control);
1907 Node* final_io = final_state->in(TypeFunc::I_O);
1908
1909 // Replace all the old call edges with the edges from the inlining result
1910 if (callprojs.fallthrough_catchproj != NULL) {
1911 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1912 }
1913 if (callprojs.fallthrough_memproj != NULL) {
1914 if (final_mem->is_MergeMem()) {
1915 // Parser's exits MergeMem was not transformed but may be optimized
1916 final_mem = _gvn.transform(final_mem);
1917 }
1918 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1919 add_mergemem_users_to_worklist(wl, final_mem);
1920 }
1921 if (callprojs.fallthrough_ioproj != NULL) {
1922 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1923 }
1924
1925 // Replace the result with the new result if it exists and is used
1926 if (callprojs.resproj != NULL && result != NULL) {
1927 C->gvn_replace_by(callprojs.resproj, result);
1928 }
1929
1930 if (ejvms == NULL) {
1931 // No exception edges to simply kill off those paths
1932 if (callprojs.catchall_catchproj != NULL) {
1933 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1934 }
1935 if (callprojs.catchall_memproj != NULL) {
1936 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1937 }
1938 if (callprojs.catchall_ioproj != NULL) {
1939 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1940 }
1941 // Replace the old exception object with top
1942 if (callprojs.exobj != NULL) {
1943 C->gvn_replace_by(callprojs.exobj, C->top());
1944 }
1945 } else {
1946 GraphKit ekit(ejvms);
1947
1948 // Load my combined exception state into the kit, with all phis transformed:
1949 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1950 replaced_nodes_exception = ex_map->replaced_nodes();
1951
1952 Node* ex_oop = ekit.use_exception_state(ex_map);
1953
1954 if (callprojs.catchall_catchproj != NULL) {
1955 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1956 ex_ctl = ekit.control();
1957 }
1958 if (callprojs.catchall_memproj != NULL) {
1959 Node* ex_mem = ekit.reset_memory();
1960 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
1961 add_mergemem_users_to_worklist(wl, ex_mem);
1962 }
1963 if (callprojs.catchall_ioproj != NULL) {
1964 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1965 }
1966
1967 // Replace the old exception object with the newly created one
1968 if (callprojs.exobj != NULL) {
1969 C->gvn_replace_by(callprojs.exobj, ex_oop);
1970 }
1971 }
1972
1973 // Disconnect the call from the graph
1974 call->disconnect_inputs(C);
1975 C->gvn_replace_by(call, C->top());
1976
1977 // Clean up any MergeMems that feed other MergeMems since the
1978 // optimizer doesn't like that.
1979 while (wl.size() > 0) {
1980 _gvn.transform(wl.pop());
1981 }
1982
1983 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1984 replaced_nodes.apply(C, final_ctl);
1985 }
1986 if (!ex_ctl->is_top() && do_replaced_nodes) {
1987 replaced_nodes_exception.apply(C, ex_ctl);
1988 }
1989 }
1990
1991
1992 //------------------------------increment_counter------------------------------
1993 // for statistics: increment a VM counter by 1
1994
increment_counter(address counter_addr)1995 void GraphKit::increment_counter(address counter_addr) {
1996 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1997 increment_counter(adr1);
1998 }
1999
increment_counter(Node * counter_addr)2000 void GraphKit::increment_counter(Node* counter_addr) {
2001 int adr_type = Compile::AliasIdxRaw;
2002 Node* ctrl = control();
2003 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2004 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
2005 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
2006 }
2007
2008
2009 //------------------------------uncommon_trap----------------------------------
2010 // Bail out to the interpreter in mid-method. Implemented by calling the
2011 // uncommon_trap blob. This helper function inserts a runtime call with the
2012 // right debug info.
uncommon_trap(int trap_request,ciKlass * klass,const char * comment,bool must_throw,bool keep_exact_action)2013 void GraphKit::uncommon_trap(int trap_request,
2014 ciKlass* klass, const char* comment,
2015 bool must_throw,
2016 bool keep_exact_action) {
2017 if (failing()) stop();
2018 if (stopped()) return; // trap reachable?
2019
2020 // Note: If ProfileTraps is true, and if a deopt. actually
2021 // occurs here, the runtime will make sure an MDO exists. There is
2022 // no need to call method()->ensure_method_data() at this point.
2023
2024 // Set the stack pointer to the right value for reexecution:
2025 set_sp(reexecute_sp());
2026
2027 #ifdef ASSERT
2028 if (!must_throw) {
2029 // Make sure the stack has at least enough depth to execute
2030 // the current bytecode.
2031 int inputs, ignored_depth;
2032 if (compute_stack_effects(inputs, ignored_depth)) {
2033 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2034 Bytecodes::name(java_bc()), sp(), inputs);
2035 }
2036 }
2037 #endif
2038
2039 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2040 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2041
2042 switch (action) {
2043 case Deoptimization::Action_maybe_recompile:
2044 case Deoptimization::Action_reinterpret:
2045 // Temporary fix for 6529811 to allow virtual calls to be sure they
2046 // get the chance to go from mono->bi->mega
2047 if (!keep_exact_action &&
2048 Deoptimization::trap_request_index(trap_request) < 0 &&
2049 too_many_recompiles(reason)) {
2050 // This BCI is causing too many recompilations.
2051 if (C->log() != NULL) {
2052 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2053 Deoptimization::trap_reason_name(reason),
2054 Deoptimization::trap_action_name(action));
2055 }
2056 action = Deoptimization::Action_none;
2057 trap_request = Deoptimization::make_trap_request(reason, action);
2058 } else {
2059 C->set_trap_can_recompile(true);
2060 }
2061 break;
2062 case Deoptimization::Action_make_not_entrant:
2063 C->set_trap_can_recompile(true);
2064 break;
2065 case Deoptimization::Action_none:
2066 case Deoptimization::Action_make_not_compilable:
2067 break;
2068 default:
2069 #ifdef ASSERT
2070 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2071 #endif
2072 break;
2073 }
2074
2075 if (TraceOptoParse) {
2076 char buf[100];
2077 tty->print_cr("Uncommon trap %s at bci:%d",
2078 Deoptimization::format_trap_request(buf, sizeof(buf),
2079 trap_request), bci());
2080 }
2081
2082 CompileLog* log = C->log();
2083 if (log != NULL) {
2084 int kid = (klass == NULL)? -1: log->identify(klass);
2085 log->begin_elem("uncommon_trap bci='%d'", bci());
2086 char buf[100];
2087 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2088 trap_request));
2089 if (kid >= 0) log->print(" klass='%d'", kid);
2090 if (comment != NULL) log->print(" comment='%s'", comment);
2091 log->end_elem();
2092 }
2093
2094 // Make sure any guarding test views this path as very unlikely
2095 Node *i0 = control()->in(0);
2096 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2097 IfNode *iff = i0->as_If();
2098 float f = iff->_prob; // Get prob
2099 if (control()->Opcode() == Op_IfTrue) {
2100 if (f > PROB_UNLIKELY_MAG(4))
2101 iff->_prob = PROB_MIN;
2102 } else {
2103 if (f < PROB_LIKELY_MAG(4))
2104 iff->_prob = PROB_MAX;
2105 }
2106 }
2107
2108 // Clear out dead values from the debug info.
2109 kill_dead_locals();
2110
2111 // Now insert the uncommon trap subroutine call
2112 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2113 const TypePtr* no_memory_effects = NULL;
2114 // Pass the index of the class to be loaded
2115 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2116 (must_throw ? RC_MUST_THROW : 0),
2117 OptoRuntime::uncommon_trap_Type(),
2118 call_addr, "uncommon_trap", no_memory_effects,
2119 intcon(trap_request));
2120 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2121 "must extract request correctly from the graph");
2122 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2123
2124 call->set_req(TypeFunc::ReturnAdr, returnadr());
2125 // The debug info is the only real input to this call.
2126
2127 // Halt-and-catch fire here. The above call should never return!
2128 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2129 PRODUCT_ONLY(COMMA /*reachable*/false));
2130 _gvn.set_type_bottom(halt);
2131 root()->add_req(halt);
2132
2133 stop_and_kill_map();
2134 }
2135
2136
2137 //--------------------------just_allocated_object------------------------------
2138 // Report the object that was just allocated.
2139 // It must be the case that there are no intervening safepoints.
2140 // We use this to determine if an object is so "fresh" that
2141 // it does not require card marks.
just_allocated_object(Node * current_control)2142 Node* GraphKit::just_allocated_object(Node* current_control) {
2143 Node* ctrl = current_control;
2144 // Object::<init> is invoked after allocation, most of invoke nodes
2145 // will be reduced, but a region node is kept in parse time, we check
2146 // the pattern and skip the region node if it degraded to a copy.
2147 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 &&
2148 ctrl->as_Region()->is_copy()) {
2149 ctrl = ctrl->as_Region()->is_copy();
2150 }
2151 if (C->recent_alloc_ctl() == ctrl) {
2152 return C->recent_alloc_obj();
2153 }
2154 return NULL;
2155 }
2156
2157
2158 /**
2159 * Record profiling data exact_kls for Node n with the type system so
2160 * that it can propagate it (speculation)
2161 *
2162 * @param n node that the type applies to
2163 * @param exact_kls type from profiling
2164 * @param maybe_null did profiling see null?
2165 *
2166 * @return node with improved type
2167 */
record_profile_for_speculation(Node * n,ciKlass * exact_kls,ProfilePtrKind ptr_kind)2168 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2169 const Type* current_type = _gvn.type(n);
2170 assert(UseTypeSpeculation, "type speculation must be on");
2171
2172 const TypePtr* speculative = current_type->speculative();
2173
2174 // Should the klass from the profile be recorded in the speculative type?
2175 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2176 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2177 const TypeOopPtr* xtype = tklass->as_instance_type();
2178 assert(xtype->klass_is_exact(), "Should be exact");
2179 // Any reason to believe n is not null (from this profiling or a previous one)?
2180 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2181 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2182 // record the new speculative type's depth
2183 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2184 speculative = speculative->with_inline_depth(jvms()->depth());
2185 } else if (current_type->would_improve_ptr(ptr_kind)) {
2186 // Profiling report that null was never seen so we can change the
2187 // speculative type to non null ptr.
2188 if (ptr_kind == ProfileAlwaysNull) {
2189 speculative = TypePtr::NULL_PTR;
2190 } else {
2191 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2192 const TypePtr* ptr = TypePtr::NOTNULL;
2193 if (speculative != NULL) {
2194 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2195 } else {
2196 speculative = ptr;
2197 }
2198 }
2199 }
2200
2201 if (speculative != current_type->speculative()) {
2202 // Build a type with a speculative type (what we think we know
2203 // about the type but will need a guard when we use it)
2204 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2205 // We're changing the type, we need a new CheckCast node to carry
2206 // the new type. The new type depends on the control: what
2207 // profiling tells us is only valid from here as far as we can
2208 // tell.
2209 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2210 cast = _gvn.transform(cast);
2211 replace_in_map(n, cast);
2212 n = cast;
2213 }
2214
2215 return n;
2216 }
2217
2218 /**
2219 * Record profiling data from receiver profiling at an invoke with the
2220 * type system so that it can propagate it (speculation)
2221 *
2222 * @param n receiver node
2223 *
2224 * @return node with improved type
2225 */
record_profiled_receiver_for_speculation(Node * n)2226 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2227 if (!UseTypeSpeculation) {
2228 return n;
2229 }
2230 ciKlass* exact_kls = profile_has_unique_klass();
2231 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2232 if ((java_bc() == Bytecodes::_checkcast ||
2233 java_bc() == Bytecodes::_instanceof ||
2234 java_bc() == Bytecodes::_aastore) &&
2235 method()->method_data()->is_mature()) {
2236 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2237 if (data != NULL) {
2238 if (!data->as_BitData()->null_seen()) {
2239 ptr_kind = ProfileNeverNull;
2240 } else {
2241 assert(data->is_ReceiverTypeData(), "bad profile data type");
2242 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2243 uint i = 0;
2244 for (; i < call->row_limit(); i++) {
2245 ciKlass* receiver = call->receiver(i);
2246 if (receiver != NULL) {
2247 break;
2248 }
2249 }
2250 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2251 }
2252 }
2253 }
2254 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2255 }
2256
2257 /**
2258 * Record profiling data from argument profiling at an invoke with the
2259 * type system so that it can propagate it (speculation)
2260 *
2261 * @param dest_method target method for the call
2262 * @param bc what invoke bytecode is this?
2263 */
record_profiled_arguments_for_speculation(ciMethod * dest_method,Bytecodes::Code bc)2264 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2265 if (!UseTypeSpeculation) {
2266 return;
2267 }
2268 const TypeFunc* tf = TypeFunc::make(dest_method);
2269 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2270 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2271 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2272 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2273 if (is_reference_type(targ->basic_type())) {
2274 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2275 ciKlass* better_type = NULL;
2276 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2277 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2278 }
2279 i++;
2280 }
2281 }
2282 }
2283
2284 /**
2285 * Record profiling data from parameter profiling at an invoke with
2286 * the type system so that it can propagate it (speculation)
2287 */
record_profiled_parameters_for_speculation()2288 void GraphKit::record_profiled_parameters_for_speculation() {
2289 if (!UseTypeSpeculation) {
2290 return;
2291 }
2292 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2293 if (_gvn.type(local(i))->isa_oopptr()) {
2294 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2295 ciKlass* better_type = NULL;
2296 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2297 record_profile_for_speculation(local(i), better_type, ptr_kind);
2298 }
2299 j++;
2300 }
2301 }
2302 }
2303
2304 /**
2305 * Record profiling data from return value profiling at an invoke with
2306 * the type system so that it can propagate it (speculation)
2307 */
record_profiled_return_for_speculation()2308 void GraphKit::record_profiled_return_for_speculation() {
2309 if (!UseTypeSpeculation) {
2310 return;
2311 }
2312 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2313 ciKlass* better_type = NULL;
2314 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2315 // If profiling reports a single type for the return value,
2316 // feed it to the type system so it can propagate it as a
2317 // speculative type
2318 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2319 }
2320 }
2321
round_double_result(ciMethod * dest_method)2322 void GraphKit::round_double_result(ciMethod* dest_method) {
2323 if (Matcher::strict_fp_requires_explicit_rounding) {
2324 // If a strict caller invokes a non-strict callee, round a double result.
2325 // A non-strict method may return a double value which has an extended exponent,
2326 // but this must not be visible in a caller which is strict.
2327 BasicType result_type = dest_method->return_type()->basic_type();
2328 assert(method() != NULL, "must have caller context");
2329 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2330 // Destination method's return value is on top of stack
2331 // dstore_rounding() does gvn.transform
2332 Node *result = pop_pair();
2333 result = dstore_rounding(result);
2334 push_pair(result);
2335 }
2336 }
2337 }
2338
round_double_arguments(ciMethod * dest_method)2339 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2340 if (Matcher::strict_fp_requires_explicit_rounding) {
2341 // (Note: TypeFunc::make has a cache that makes this fast.)
2342 const TypeFunc* tf = TypeFunc::make(dest_method);
2343 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2344 for (int j = 0; j < nargs; j++) {
2345 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2346 if (targ->basic_type() == T_DOUBLE) {
2347 // If any parameters are doubles, they must be rounded before
2348 // the call, dstore_rounding does gvn.transform
2349 Node *arg = argument(j);
2350 arg = dstore_rounding(arg);
2351 set_argument(j, arg);
2352 }
2353 }
2354 }
2355 }
2356
2357 // rounding for strict float precision conformance
precision_rounding(Node * n)2358 Node* GraphKit::precision_rounding(Node* n) {
2359 if (Matcher::strict_fp_requires_explicit_rounding) {
2360 #ifdef IA32
2361 if (_method->flags().is_strict() && UseSSE == 0) {
2362 return _gvn.transform(new RoundFloatNode(0, n));
2363 }
2364 #else
2365 Unimplemented();
2366 #endif // IA32
2367 }
2368 return n;
2369 }
2370
2371 // rounding for strict double precision conformance
dprecision_rounding(Node * n)2372 Node* GraphKit::dprecision_rounding(Node *n) {
2373 if (Matcher::strict_fp_requires_explicit_rounding) {
2374 #ifdef IA32
2375 if (_method->flags().is_strict() && UseSSE < 2) {
2376 return _gvn.transform(new RoundDoubleNode(0, n));
2377 }
2378 #else
2379 Unimplemented();
2380 #endif // IA32
2381 }
2382 return n;
2383 }
2384
2385 // rounding for non-strict double stores
dstore_rounding(Node * n)2386 Node* GraphKit::dstore_rounding(Node* n) {
2387 if (Matcher::strict_fp_requires_explicit_rounding) {
2388 #ifdef IA32
2389 if (UseSSE < 2) {
2390 return _gvn.transform(new RoundDoubleNode(0, n));
2391 }
2392 #else
2393 Unimplemented();
2394 #endif // IA32
2395 }
2396 return n;
2397 }
2398
2399 //=============================================================================
2400 // Generate a fast path/slow path idiom. Graph looks like:
2401 // [foo] indicates that 'foo' is a parameter
2402 //
2403 // [in] NULL
2404 // \ /
2405 // CmpP
2406 // Bool ne
2407 // If
2408 // / \
2409 // True False-<2>
2410 // / |
2411 // / cast_not_null
2412 // Load | | ^
2413 // [fast_test] | |
2414 // gvn to opt_test | |
2415 // / \ | <1>
2416 // True False |
2417 // | \\ |
2418 // [slow_call] \[fast_result]
2419 // Ctl Val \ \
2420 // | \ \
2421 // Catch <1> \ \
2422 // / \ ^ \ \
2423 // Ex No_Ex | \ \
2424 // | \ \ | \ <2> \
2425 // ... \ [slow_res] | | \ [null_result]
2426 // \ \--+--+--- | |
2427 // \ | / \ | /
2428 // --------Region Phi
2429 //
2430 //=============================================================================
2431 // Code is structured as a series of driver functions all called 'do_XXX' that
2432 // call a set of helper functions. Helper functions first, then drivers.
2433
2434 //------------------------------null_check_oop---------------------------------
2435 // Null check oop. Set null-path control into Region in slot 3.
2436 // 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)2437 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2438 bool never_see_null,
2439 bool safe_for_replace,
2440 bool speculative) {
2441 // Initial NULL check taken path
2442 (*null_control) = top();
2443 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2444
2445 // Generate uncommon_trap:
2446 if (never_see_null && (*null_control) != top()) {
2447 // If we see an unexpected null at a check-cast we record it and force a
2448 // recompile; the offending check-cast will be compiled to handle NULLs.
2449 // If we see more than one offending BCI, then all checkcasts in the
2450 // method will be compiled to handle NULLs.
2451 PreserveJVMState pjvms(this);
2452 set_control(*null_control);
2453 replace_in_map(value, null());
2454 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2455 uncommon_trap(reason,
2456 Deoptimization::Action_make_not_entrant);
2457 (*null_control) = top(); // NULL path is dead
2458 }
2459 if ((*null_control) == top() && safe_for_replace) {
2460 replace_in_map(value, cast);
2461 }
2462
2463 // Cast away null-ness on the result
2464 return cast;
2465 }
2466
2467 //------------------------------opt_iff----------------------------------------
2468 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2469 // Return slow-path control.
opt_iff(Node * region,Node * iff)2470 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2471 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2472
2473 // Fast path taken; set region slot 2
2474 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2475 region->init_req(2,fast_taken); // Capture fast-control
2476
2477 // Fast path not-taken, i.e. slow path
2478 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2479 return slow_taken;
2480 }
2481
2482 //-----------------------------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)2483 Node* GraphKit::make_runtime_call(int flags,
2484 const TypeFunc* call_type, address call_addr,
2485 const char* call_name,
2486 const TypePtr* adr_type,
2487 // The following parms are all optional.
2488 // The first NULL ends the list.
2489 Node* parm0, Node* parm1,
2490 Node* parm2, Node* parm3,
2491 Node* parm4, Node* parm5,
2492 Node* parm6, Node* parm7) {
2493 assert(call_addr != NULL, "must not call NULL targets");
2494
2495 // Slow-path call
2496 bool is_leaf = !(flags & RC_NO_LEAF);
2497 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2498 if (call_name == NULL) {
2499 assert(!is_leaf, "must supply name for leaf");
2500 call_name = OptoRuntime::stub_name(call_addr);
2501 }
2502 CallNode* call;
2503 if (!is_leaf) {
2504 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2505 bci(), adr_type);
2506 } else if (flags & RC_NO_FP) {
2507 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2508 } else {
2509 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2510 }
2511
2512 // The following is similar to set_edges_for_java_call,
2513 // except that the memory effects of the call are restricted to AliasIdxRaw.
2514
2515 // Slow path call has no side-effects, uses few values
2516 bool wide_in = !(flags & RC_NARROW_MEM);
2517 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2518
2519 Node* prev_mem = NULL;
2520 if (wide_in) {
2521 prev_mem = set_predefined_input_for_runtime_call(call);
2522 } else {
2523 assert(!wide_out, "narrow in => narrow out");
2524 Node* narrow_mem = memory(adr_type);
2525 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2526 }
2527
2528 // Hook each parm in order. Stop looking at the first NULL.
2529 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2530 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2531 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2532 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2533 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2534 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2535 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2536 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2537 /* close each nested if ===> */ } } } } } } } }
2538 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2539
2540 if (!is_leaf) {
2541 // Non-leaves can block and take safepoints:
2542 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2543 }
2544 // Non-leaves can throw exceptions:
2545 if (has_io) {
2546 call->set_req(TypeFunc::I_O, i_o());
2547 }
2548
2549 if (flags & RC_UNCOMMON) {
2550 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2551 // (An "if" probability corresponds roughly to an unconditional count.
2552 // Sort of.)
2553 call->set_cnt(PROB_UNLIKELY_MAG(4));
2554 }
2555
2556 Node* c = _gvn.transform(call);
2557 assert(c == call, "cannot disappear");
2558
2559 if (wide_out) {
2560 // Slow path call has full side-effects.
2561 set_predefined_output_for_runtime_call(call);
2562 } else {
2563 // Slow path call has few side-effects, and/or sets few values.
2564 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2565 }
2566
2567 if (has_io) {
2568 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2569 }
2570 return call;
2571
2572 }
2573
2574 // i2b
sign_extend_byte(Node * in)2575 Node* GraphKit::sign_extend_byte(Node* in) {
2576 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2577 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2578 }
2579
2580 // i2s
sign_extend_short(Node * in)2581 Node* GraphKit::sign_extend_short(Node* in) {
2582 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2583 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2584 }
2585
2586 //-----------------------------make_native_call-------------------------------
make_native_call(const TypeFunc * call_type,uint nargs,ciNativeEntryPoint * nep)2587 Node* GraphKit::make_native_call(const TypeFunc* call_type, uint nargs, ciNativeEntryPoint* nep) {
2588 uint n_filtered_args = nargs - 2; // -fallback, -nep;
2589 ResourceMark rm;
2590 Node** argument_nodes = NEW_RESOURCE_ARRAY(Node*, n_filtered_args);
2591 const Type** arg_types = TypeTuple::fields(n_filtered_args);
2592 GrowableArray<VMReg> arg_regs(C->comp_arena(), n_filtered_args, n_filtered_args, VMRegImpl::Bad());
2593
2594 VMReg* argRegs = nep->argMoves();
2595 {
2596 for (uint vm_arg_pos = 0, java_arg_read_pos = 0;
2597 vm_arg_pos < n_filtered_args; vm_arg_pos++) {
2598 uint vm_unfiltered_arg_pos = vm_arg_pos + 1; // +1 to skip fallback handle argument
2599 Node* node = argument(vm_unfiltered_arg_pos);
2600 const Type* type = call_type->domain()->field_at(TypeFunc::Parms + vm_unfiltered_arg_pos);
2601 VMReg reg = type == Type::HALF
2602 ? VMRegImpl::Bad()
2603 : argRegs[java_arg_read_pos++];
2604
2605 argument_nodes[vm_arg_pos] = node;
2606 arg_types[TypeFunc::Parms + vm_arg_pos] = type;
2607 arg_regs.at_put(vm_arg_pos, reg);
2608 }
2609 }
2610
2611 uint n_returns = call_type->range()->cnt() - TypeFunc::Parms;
2612 GrowableArray<VMReg> ret_regs(C->comp_arena(), n_returns, n_returns, VMRegImpl::Bad());
2613 const Type** ret_types = TypeTuple::fields(n_returns);
2614
2615 VMReg* retRegs = nep->returnMoves();
2616 {
2617 for (uint vm_ret_pos = 0, java_ret_read_pos = 0;
2618 vm_ret_pos < n_returns; vm_ret_pos++) { // 0 or 1
2619 const Type* type = call_type->range()->field_at(TypeFunc::Parms + vm_ret_pos);
2620 VMReg reg = type == Type::HALF
2621 ? VMRegImpl::Bad()
2622 : retRegs[java_ret_read_pos++];
2623
2624 ret_regs.at_put(vm_ret_pos, reg);
2625 ret_types[TypeFunc::Parms + vm_ret_pos] = type;
2626 }
2627 }
2628
2629 const TypeFunc* new_call_type = TypeFunc::make(
2630 TypeTuple::make(TypeFunc::Parms + n_filtered_args, arg_types),
2631 TypeTuple::make(TypeFunc::Parms + n_returns, ret_types)
2632 );
2633
2634 address call_addr = nep->entry_point();
2635 if (nep->need_transition()) {
2636 BufferBlob* invoker = SharedRuntime::make_native_invoker(call_addr,
2637 nep->shadow_space(),
2638 arg_regs, ret_regs);
2639 if (invoker == NULL) {
2640 C->record_failure("native invoker not implemented on this platform");
2641 return NULL;
2642 }
2643 C->add_native_invoker(invoker);
2644 call_addr = invoker->code_begin();
2645 }
2646 assert(call_addr != NULL, "sanity");
2647
2648 CallNativeNode* call = new CallNativeNode(new_call_type, call_addr, nep->name(), TypePtr::BOTTOM,
2649 arg_regs,
2650 ret_regs,
2651 nep->shadow_space(),
2652 nep->need_transition());
2653
2654 if (call->_need_transition) {
2655 add_safepoint_edges(call);
2656 }
2657
2658 set_predefined_input_for_runtime_call(call);
2659
2660 for (uint i = 0; i < n_filtered_args; i++) {
2661 call->init_req(i + TypeFunc::Parms, argument_nodes[i]);
2662 }
2663
2664 Node* c = gvn().transform(call);
2665 assert(c == call, "cannot disappear");
2666
2667 set_predefined_output_for_runtime_call(call);
2668
2669 Node* ret;
2670 if (method() == NULL || method()->return_type()->basic_type() == T_VOID) {
2671 ret = top();
2672 } else {
2673 ret = gvn().transform(new ProjNode(call, TypeFunc::Parms));
2674 // Unpack native results if needed
2675 // Need this method type since it's unerased
2676 switch (nep->method_type()->rtype()->basic_type()) {
2677 case T_CHAR:
2678 ret = _gvn.transform(new AndINode(ret, _gvn.intcon(0xFFFF)));
2679 break;
2680 case T_BYTE:
2681 ret = sign_extend_byte(ret);
2682 break;
2683 case T_SHORT:
2684 ret = sign_extend_short(ret);
2685 break;
2686 default: // do nothing
2687 break;
2688 }
2689 }
2690
2691 push_node(method()->return_type()->basic_type(), ret);
2692
2693 return call;
2694 }
2695
2696 //------------------------------merge_memory-----------------------------------
2697 // Merge memory from one path into the current memory state.
merge_memory(Node * new_mem,Node * region,int new_path)2698 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2699 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2700 Node* old_slice = mms.force_memory();
2701 Node* new_slice = mms.memory2();
2702 if (old_slice != new_slice) {
2703 PhiNode* phi;
2704 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2705 if (mms.is_empty()) {
2706 // clone base memory Phi's inputs for this memory slice
2707 assert(old_slice == mms.base_memory(), "sanity");
2708 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2709 _gvn.set_type(phi, Type::MEMORY);
2710 for (uint i = 1; i < phi->req(); i++) {
2711 phi->init_req(i, old_slice->in(i));
2712 }
2713 } else {
2714 phi = old_slice->as_Phi(); // Phi was generated already
2715 }
2716 } else {
2717 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2718 _gvn.set_type(phi, Type::MEMORY);
2719 }
2720 phi->set_req(new_path, new_slice);
2721 mms.set_memory(phi);
2722 }
2723 }
2724 }
2725
2726 //------------------------------make_slow_call_ex------------------------------
2727 // Make the exception handler hookups for the slow call
make_slow_call_ex(Node * call,ciInstanceKlass * ex_klass,bool separate_io_proj,bool deoptimize)2728 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2729 if (stopped()) return;
2730
2731 // Make a catch node with just two handlers: fall-through and catch-all
2732 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2733 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2734 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2735 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2736
2737 { PreserveJVMState pjvms(this);
2738 set_control(excp);
2739 set_i_o(i_o);
2740
2741 if (excp != top()) {
2742 if (deoptimize) {
2743 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2744 uncommon_trap(Deoptimization::Reason_unhandled,
2745 Deoptimization::Action_none);
2746 } else {
2747 // Create an exception state also.
2748 // Use an exact type if the caller has a specific exception.
2749 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2750 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2751 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2752 }
2753 }
2754 }
2755
2756 // Get the no-exception control from the CatchNode.
2757 set_control(norm);
2758 }
2759
gen_subtype_check_compare(Node * ctrl,Node * in1,Node * in2,BoolTest::mask test,float p,PhaseGVN & gvn,BasicType bt)2760 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2761 Node* cmp = NULL;
2762 switch(bt) {
2763 case T_INT: cmp = new CmpINode(in1, in2); break;
2764 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2765 default: fatal("unexpected comparison type %s", type2name(bt));
2766 }
2767 gvn.transform(cmp);
2768 Node* bol = gvn.transform(new BoolNode(cmp, test));
2769 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2770 gvn.transform(iff);
2771 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2772 return iff;
2773 }
2774
2775 //-------------------------------gen_subtype_check-----------------------------
2776 // Generate a subtyping check. Takes as input the subtype and supertype.
2777 // Returns 2 values: sets the default control() to the true path and returns
2778 // the false path. Only reads invariant memory; sets no (visible) memory.
2779 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2780 // but that's not exposed to the optimizer. This call also doesn't take in an
2781 // Object; if you wish to check an Object you need to load the Object's class
2782 // prior to coming here.
gen_subtype_check(Node * subklass,Node * superklass,Node ** ctrl,Node * mem,PhaseGVN & gvn)2783 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn) {
2784 Compile* C = gvn.C;
2785 if ((*ctrl)->is_top()) {
2786 return C->top();
2787 }
2788
2789 // Fast check for identical types, perhaps identical constants.
2790 // The types can even be identical non-constants, in cases
2791 // involving Array.newInstance, Object.clone, etc.
2792 if (subklass == superklass)
2793 return C->top(); // false path is dead; no test needed.
2794
2795 if (gvn.type(superklass)->singleton()) {
2796 ciKlass* superk = gvn.type(superklass)->is_klassptr()->klass();
2797 ciKlass* subk = gvn.type(subklass)->is_klassptr()->klass();
2798
2799 // In the common case of an exact superklass, try to fold up the
2800 // test before generating code. You may ask, why not just generate
2801 // the code and then let it fold up? The answer is that the generated
2802 // code will necessarily include null checks, which do not always
2803 // completely fold away. If they are also needless, then they turn
2804 // into a performance loss. Example:
2805 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2806 // Here, the type of 'fa' is often exact, so the store check
2807 // of fa[1]=x will fold up, without testing the nullness of x.
2808 switch (C->static_subtype_check(superk, subk)) {
2809 case Compile::SSC_always_false:
2810 {
2811 Node* always_fail = *ctrl;
2812 *ctrl = gvn.C->top();
2813 return always_fail;
2814 }
2815 case Compile::SSC_always_true:
2816 return C->top();
2817 case Compile::SSC_easy_test:
2818 {
2819 // Just do a direct pointer compare and be done.
2820 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2821 *ctrl = gvn.transform(new IfTrueNode(iff));
2822 return gvn.transform(new IfFalseNode(iff));
2823 }
2824 case Compile::SSC_full_test:
2825 break;
2826 default:
2827 ShouldNotReachHere();
2828 }
2829 }
2830
2831 // %%% Possible further optimization: Even if the superklass is not exact,
2832 // if the subklass is the unique subtype of the superklass, the check
2833 // will always succeed. We could leave a dependency behind to ensure this.
2834
2835 // First load the super-klass's check-offset
2836 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2837 Node* m = C->immutable_memory();
2838 Node *chk_off = gvn.transform(new LoadINode(NULL, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2839 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2840 bool might_be_cache = (gvn.find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2841
2842 // Load from the sub-klass's super-class display list, or a 1-word cache of
2843 // the secondary superclass list, or a failing value with a sentinel offset
2844 // if the super-klass is an interface or exceptionally deep in the Java
2845 // hierarchy and we have to scan the secondary superclass list the hard way.
2846 // Worst-case type is a little odd: NULL is allowed as a result (usually
2847 // klass loads can never produce a NULL).
2848 Node *chk_off_X = chk_off;
2849 #ifdef _LP64
2850 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2851 #endif
2852 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2853 // For some types like interfaces the following loadKlass is from a 1-word
2854 // cache which is mutable so can't use immutable memory. Other
2855 // types load from the super-class display table which is immutable.
2856 Node *kmem = C->immutable_memory();
2857 // secondary_super_cache is not immutable but can be treated as such because:
2858 // - no ideal node writes to it in a way that could cause an
2859 // incorrect/missed optimization of the following Load.
2860 // - it's a cache so, worse case, not reading the latest value
2861 // wouldn't cause incorrect execution
2862 if (might_be_cache && mem != NULL) {
2863 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
2864 }
2865 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, NULL, kmem, p2, gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2866
2867 // Compile speed common case: ARE a subtype and we canNOT fail
2868 if( superklass == nkls )
2869 return C->top(); // false path is dead; no test needed.
2870
2871 // See if we get an immediate positive hit. Happens roughly 83% of the
2872 // time. Test to see if the value loaded just previously from the subklass
2873 // is exactly the superklass.
2874 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2875 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
2876 *ctrl = gvn.transform(new IfFalseNode(iff1));
2877
2878 // Compile speed common case: Check for being deterministic right now. If
2879 // chk_off is a constant and not equal to cacheoff then we are NOT a
2880 // subklass. In this case we need exactly the 1 test above and we can
2881 // return those results immediately.
2882 if (!might_be_cache) {
2883 Node* not_subtype_ctrl = *ctrl;
2884 *ctrl = iftrue1; // We need exactly the 1 test above
2885 return not_subtype_ctrl;
2886 }
2887
2888 // Gather the various success & failures here
2889 RegionNode *r_ok_subtype = new RegionNode(4);
2890 gvn.record_for_igvn(r_ok_subtype);
2891 RegionNode *r_not_subtype = new RegionNode(3);
2892 gvn.record_for_igvn(r_not_subtype);
2893
2894 r_ok_subtype->init_req(1, iftrue1);
2895
2896 // Check for immediate negative hit. Happens roughly 11% of the time (which
2897 // is roughly 63% of the remaining cases). Test to see if the loaded
2898 // check-offset points into the subklass display list or the 1-element
2899 // cache. If it points to the display (and NOT the cache) and the display
2900 // missed then it's not a subtype.
2901 Node *cacheoff = gvn.intcon(cacheoff_con);
2902 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2903 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2904 *ctrl = gvn.transform(new IfFalseNode(iff2));
2905
2906 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2907 // No performance impact (too rare) but allows sharing of secondary arrays
2908 // which has some footprint reduction.
2909 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2910 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2911 *ctrl = gvn.transform(new IfFalseNode(iff3));
2912
2913 // -- Roads not taken here: --
2914 // We could also have chosen to perform the self-check at the beginning
2915 // of this code sequence, as the assembler does. This would not pay off
2916 // the same way, since the optimizer, unlike the assembler, can perform
2917 // static type analysis to fold away many successful self-checks.
2918 // Non-foldable self checks work better here in second position, because
2919 // the initial primary superclass check subsumes a self-check for most
2920 // types. An exception would be a secondary type like array-of-interface,
2921 // which does not appear in its own primary supertype display.
2922 // Finally, we could have chosen to move the self-check into the
2923 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2924 // dependent manner. But it is worthwhile to have the check here,
2925 // where it can be perhaps be optimized. The cost in code space is
2926 // small (register compare, branch).
2927
2928 // Now do a linear scan of the secondary super-klass array. Again, no real
2929 // performance impact (too rare) but it's gotta be done.
2930 // Since the code is rarely used, there is no penalty for moving it
2931 // out of line, and it can only improve I-cache density.
2932 // The decision to inline or out-of-line this final check is platform
2933 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2934 Node* psc = gvn.transform(
2935 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2936
2937 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2938 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2939 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2940
2941 // Return false path; set default control to true path.
2942 *ctrl = gvn.transform(r_ok_subtype);
2943 return gvn.transform(r_not_subtype);
2944 }
2945
gen_subtype_check(Node * obj_or_subklass,Node * superklass)2946 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2947 if (ExpandSubTypeCheckAtParseTime) {
2948 MergeMemNode* mem = merged_memory();
2949 Node* ctrl = control();
2950 Node* subklass = obj_or_subklass;
2951 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2952 subklass = load_object_klass(obj_or_subklass);
2953 }
2954
2955 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn);
2956 set_control(ctrl);
2957 return n;
2958 }
2959
2960 const TypePtr* adr_type = TypeKlassPtr::make(TypePtr::NotNull, C->env()->Object_klass(), Type::OffsetBot);
2961 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass));
2962 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2963 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2964 set_control(_gvn.transform(new IfTrueNode(iff)));
2965 return _gvn.transform(new IfFalseNode(iff));
2966 }
2967
2968 // Profile-driven exact type check:
type_check_receiver(Node * receiver,ciKlass * klass,float prob,Node ** casted_receiver)2969 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2970 float prob,
2971 Node* *casted_receiver) {
2972 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2973 Node* recv_klass = load_object_klass(receiver);
2974 Node* want_klass = makecon(tklass);
2975 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2976 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2977 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2978 set_control( _gvn.transform( new IfTrueNode (iff) ));
2979 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2980
2981 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2982 assert(recv_xtype->klass_is_exact(), "");
2983
2984 // Subsume downstream occurrences of receiver with a cast to
2985 // recv_xtype, since now we know what the type will be.
2986 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2987 (*casted_receiver) = _gvn.transform(cast);
2988 // (User must make the replace_in_map call.)
2989
2990 return fail;
2991 }
2992
2993 //------------------------------subtype_check_receiver-------------------------
subtype_check_receiver(Node * receiver,ciKlass * klass,Node ** casted_receiver)2994 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2995 Node** casted_receiver) {
2996 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2997 Node* want_klass = makecon(tklass);
2998
2999 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3000
3001 // Cast receiver after successful check
3002 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3003 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
3004 (*casted_receiver) = _gvn.transform(cast);
3005
3006 return slow_ctl;
3007 }
3008
3009 //------------------------------seems_never_null-------------------------------
3010 // Use null_seen information if it is available from the profile.
3011 // If we see an unexpected null at a type check we record it and force a
3012 // recompile; the offending check will be recompiled to handle NULLs.
3013 // If we see several offending BCIs, then all checks in the
3014 // method will be recompiled.
seems_never_null(Node * obj,ciProfileData * data,bool & speculating)3015 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3016 speculating = !_gvn.type(obj)->speculative_maybe_null();
3017 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3018 if (UncommonNullCast // Cutout for this technique
3019 && obj != null() // And not the -Xcomp stupid case?
3020 && !too_many_traps(reason)
3021 ) {
3022 if (speculating) {
3023 return true;
3024 }
3025 if (data == NULL)
3026 // Edge case: no mature data. Be optimistic here.
3027 return true;
3028 // If the profile has not seen a null, assume it won't happen.
3029 assert(java_bc() == Bytecodes::_checkcast ||
3030 java_bc() == Bytecodes::_instanceof ||
3031 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3032 return !data->as_BitData()->null_seen();
3033 }
3034 speculating = false;
3035 return false;
3036 }
3037
guard_klass_being_initialized(Node * klass)3038 void GraphKit::guard_klass_being_initialized(Node* klass) {
3039 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3040 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3041 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
3042 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3043 T_BYTE, MemNode::unordered);
3044 init_state = _gvn.transform(init_state);
3045
3046 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3047
3048 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3049 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3050
3051 { BuildCutout unless(this, tst, PROB_MAX);
3052 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3053 }
3054 }
3055
guard_init_thread(Node * klass)3056 void GraphKit::guard_init_thread(Node* klass) {
3057 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3058 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3059
3060 Node* init_thread = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
3061 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3062 T_ADDRESS, MemNode::unordered);
3063 init_thread = _gvn.transform(init_thread);
3064
3065 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3066
3067 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3068 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3069
3070 { BuildCutout unless(this, tst, PROB_MAX);
3071 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3072 }
3073 }
3074
clinit_barrier(ciInstanceKlass * ik,ciMethod * context)3075 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3076 if (ik->is_being_initialized()) {
3077 if (C->needs_clinit_barrier(ik, context)) {
3078 Node* klass = makecon(TypeKlassPtr::make(ik));
3079 guard_klass_being_initialized(klass);
3080 guard_init_thread(klass);
3081 insert_mem_bar(Op_MemBarCPUOrder);
3082 }
3083 } else if (ik->is_initialized()) {
3084 return; // no barrier needed
3085 } else {
3086 uncommon_trap(Deoptimization::Reason_uninitialized,
3087 Deoptimization::Action_reinterpret,
3088 NULL);
3089 }
3090 }
3091
3092 //------------------------maybe_cast_profiled_receiver-------------------------
3093 // If the profile has seen exactly one type, narrow to exactly that type.
3094 // 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)3095 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3096 ciKlass* require_klass,
3097 ciKlass* spec_klass,
3098 bool safe_for_replace) {
3099 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
3100
3101 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
3102
3103 // Make sure we haven't already deoptimized from this tactic.
3104 if (too_many_traps_or_recompiles(reason))
3105 return NULL;
3106
3107 // (No, this isn't a call, but it's enough like a virtual call
3108 // to use the same ciMethod accessor to get the profile info...)
3109 // If we have a speculative type use it instead of profiling (which
3110 // may not help us)
3111 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
3112 if (exact_kls != NULL) {// no cast failures here
3113 if (require_klass == NULL ||
3114 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
3115 // If we narrow the type to match what the type profile sees or
3116 // the speculative type, we can then remove the rest of the
3117 // cast.
3118 // This is a win, even if the exact_kls is very specific,
3119 // because downstream operations, such as method calls,
3120 // will often benefit from the sharper type.
3121 Node* exact_obj = not_null_obj; // will get updated in place...
3122 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3123 &exact_obj);
3124 { PreserveJVMState pjvms(this);
3125 set_control(slow_ctl);
3126 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3127 }
3128 if (safe_for_replace) {
3129 replace_in_map(not_null_obj, exact_obj);
3130 }
3131 return exact_obj;
3132 }
3133 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3134 }
3135
3136 return NULL;
3137 }
3138
3139 /**
3140 * Cast obj to type and emit guard unless we had too many traps here
3141 * already
3142 *
3143 * @param obj node being casted
3144 * @param type type to cast the node to
3145 * @param not_null true if we know node cannot be null
3146 */
maybe_cast_profiled_obj(Node * obj,ciKlass * type,bool not_null)3147 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3148 ciKlass* type,
3149 bool not_null) {
3150 if (stopped()) {
3151 return obj;
3152 }
3153
3154 // type == NULL if profiling tells us this object is always null
3155 if (type != NULL) {
3156 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3157 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3158
3159 if (!too_many_traps_or_recompiles(null_reason) &&
3160 !too_many_traps_or_recompiles(class_reason)) {
3161 Node* not_null_obj = NULL;
3162 // not_null is true if we know the object is not null and
3163 // there's no need for a null check
3164 if (!not_null) {
3165 Node* null_ctl = top();
3166 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3167 assert(null_ctl->is_top(), "no null control here");
3168 } else {
3169 not_null_obj = obj;
3170 }
3171
3172 Node* exact_obj = not_null_obj;
3173 ciKlass* exact_kls = type;
3174 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3175 &exact_obj);
3176 {
3177 PreserveJVMState pjvms(this);
3178 set_control(slow_ctl);
3179 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3180 }
3181 replace_in_map(not_null_obj, exact_obj);
3182 obj = exact_obj;
3183 }
3184 } else {
3185 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3186 Node* exact_obj = null_assert(obj);
3187 replace_in_map(obj, exact_obj);
3188 obj = exact_obj;
3189 }
3190 }
3191 return obj;
3192 }
3193
3194 //-------------------------------gen_instanceof--------------------------------
3195 // Generate an instance-of idiom. Used by both the instance-of bytecode
3196 // and the reflective instance-of call.
gen_instanceof(Node * obj,Node * superklass,bool safe_for_replace)3197 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3198 kill_dead_locals(); // Benefit all the uncommon traps
3199 assert( !stopped(), "dead parse path should be checked in callers" );
3200 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3201 "must check for not-null not-dead klass in callers");
3202
3203 // Make the merge point
3204 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3205 RegionNode* region = new RegionNode(PATH_LIMIT);
3206 Node* phi = new PhiNode(region, TypeInt::BOOL);
3207 C->set_has_split_ifs(true); // Has chance for split-if optimization
3208
3209 ciProfileData* data = NULL;
3210 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3211 data = method()->method_data()->bci_to_data(bci());
3212 }
3213 bool speculative_not_null = false;
3214 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3215 && seems_never_null(obj, data, speculative_not_null));
3216
3217 // Null check; get casted pointer; set region slot 3
3218 Node* null_ctl = top();
3219 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3220
3221 // If not_null_obj is dead, only null-path is taken
3222 if (stopped()) { // Doing instance-of on a NULL?
3223 set_control(null_ctl);
3224 return intcon(0);
3225 }
3226 region->init_req(_null_path, null_ctl);
3227 phi ->init_req(_null_path, intcon(0)); // Set null path value
3228 if (null_ctl == top()) {
3229 // Do this eagerly, so that pattern matches like is_diamond_phi
3230 // will work even during parsing.
3231 assert(_null_path == PATH_LIMIT-1, "delete last");
3232 region->del_req(_null_path);
3233 phi ->del_req(_null_path);
3234 }
3235
3236 // Do we know the type check always succeed?
3237 bool known_statically = false;
3238 if (_gvn.type(superklass)->singleton()) {
3239 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3240 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3241 if (subk != NULL && subk->is_loaded()) {
3242 int static_res = C->static_subtype_check(superk, subk);
3243 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3244 }
3245 }
3246
3247 if (!known_statically) {
3248 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3249 // We may not have profiling here or it may not help us. If we
3250 // have a speculative type use it to perform an exact cast.
3251 ciKlass* spec_obj_type = obj_type->speculative_type();
3252 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3253 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3254 if (stopped()) { // Profile disagrees with this path.
3255 set_control(null_ctl); // Null is the only remaining possibility.
3256 return intcon(0);
3257 }
3258 if (cast_obj != NULL) {
3259 not_null_obj = cast_obj;
3260 }
3261 }
3262 }
3263
3264 // Generate the subtype check
3265 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3266
3267 // Plug in the success path to the general merge in slot 1.
3268 region->init_req(_obj_path, control());
3269 phi ->init_req(_obj_path, intcon(1));
3270
3271 // Plug in the failing path to the general merge in slot 2.
3272 region->init_req(_fail_path, not_subtype_ctrl);
3273 phi ->init_req(_fail_path, intcon(0));
3274
3275 // Return final merged results
3276 set_control( _gvn.transform(region) );
3277 record_for_igvn(region);
3278
3279 // If we know the type check always succeeds then we don't use the
3280 // profiling data at this bytecode. Don't lose it, feed it to the
3281 // type system as a speculative type.
3282 if (safe_for_replace) {
3283 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3284 replace_in_map(obj, casted_obj);
3285 }
3286
3287 return _gvn.transform(phi);
3288 }
3289
3290 //-------------------------------gen_checkcast---------------------------------
3291 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3292 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3293 // uncommon-trap paths work. Adjust stack after this call.
3294 // If failure_control is supplied and not null, it is filled in with
3295 // the control edge for the cast failure. Otherwise, an appropriate
3296 // uncommon trap or exception is thrown.
gen_checkcast(Node * obj,Node * superklass,Node ** failure_control)3297 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3298 Node* *failure_control) {
3299 kill_dead_locals(); // Benefit all the uncommon traps
3300 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3301 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3302
3303 // Fast cutout: Check the case that the cast is vacuously true.
3304 // This detects the common cases where the test will short-circuit
3305 // away completely. We do this before we perform the null check,
3306 // because if the test is going to turn into zero code, we don't
3307 // want a residual null check left around. (Causes a slowdown,
3308 // for example, in some objArray manipulations, such as a[i]=a[j].)
3309 if (tk->singleton()) {
3310 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3311 if (objtp != NULL && objtp->klass() != NULL) {
3312 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3313 case Compile::SSC_always_true:
3314 // If we know the type check always succeed then we don't use
3315 // the profiling data at this bytecode. Don't lose it, feed it
3316 // to the type system as a speculative type.
3317 return record_profiled_receiver_for_speculation(obj);
3318 case Compile::SSC_always_false:
3319 // It needs a null check because a null will *pass* the cast check.
3320 // A non-null value will always produce an exception.
3321 if (!objtp->maybe_null()) {
3322 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(objtp->klass())));
3323 return top();
3324 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3325 return null_assert(obj);
3326 }
3327 break; // Fall through to full check
3328 }
3329 }
3330 }
3331
3332 ciProfileData* data = NULL;
3333 bool safe_for_replace = false;
3334 if (failure_control == NULL) { // use MDO in regular case only
3335 assert(java_bc() == Bytecodes::_aastore ||
3336 java_bc() == Bytecodes::_checkcast,
3337 "interpreter profiles type checks only for these BCs");
3338 data = method()->method_data()->bci_to_data(bci());
3339 safe_for_replace = true;
3340 }
3341
3342 // Make the merge point
3343 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3344 RegionNode* region = new RegionNode(PATH_LIMIT);
3345 Node* phi = new PhiNode(region, toop);
3346 C->set_has_split_ifs(true); // Has chance for split-if optimization
3347
3348 // Use null-cast information if it is available
3349 bool speculative_not_null = false;
3350 bool never_see_null = ((failure_control == NULL) // regular case only
3351 && seems_never_null(obj, data, speculative_not_null));
3352
3353 // Null check; get casted pointer; set region slot 3
3354 Node* null_ctl = top();
3355 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3356
3357 // If not_null_obj is dead, only null-path is taken
3358 if (stopped()) { // Doing instance-of on a NULL?
3359 set_control(null_ctl);
3360 return null();
3361 }
3362 region->init_req(_null_path, null_ctl);
3363 phi ->init_req(_null_path, null()); // Set null path value
3364 if (null_ctl == top()) {
3365 // Do this eagerly, so that pattern matches like is_diamond_phi
3366 // will work even during parsing.
3367 assert(_null_path == PATH_LIMIT-1, "delete last");
3368 region->del_req(_null_path);
3369 phi ->del_req(_null_path);
3370 }
3371
3372 Node* cast_obj = NULL;
3373 if (tk->klass_is_exact()) {
3374 // The following optimization tries to statically cast the speculative type of the object
3375 // (for example obtained during profiling) to the type of the superklass and then do a
3376 // dynamic check that the type of the object is what we expect. To work correctly
3377 // for checkcast and aastore the type of superklass should be exact.
3378 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3379 // We may not have profiling here or it may not help us. If we have
3380 // a speculative type use it to perform an exact cast.
3381 ciKlass* spec_obj_type = obj_type->speculative_type();
3382 if (spec_obj_type != NULL || data != NULL) {
3383 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3384 if (cast_obj != NULL) {
3385 if (failure_control != NULL) // failure is now impossible
3386 (*failure_control) = top();
3387 // adjust the type of the phi to the exact klass:
3388 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3389 }
3390 }
3391 }
3392
3393 if (cast_obj == NULL) {
3394 // Generate the subtype check
3395 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass );
3396
3397 // Plug in success path into the merge
3398 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3399 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3400 if (failure_control == NULL) {
3401 if (not_subtype_ctrl != top()) { // If failure is possible
3402 PreserveJVMState pjvms(this);
3403 set_control(not_subtype_ctrl);
3404 builtin_throw(Deoptimization::Reason_class_check, load_object_klass(not_null_obj));
3405 }
3406 } else {
3407 (*failure_control) = not_subtype_ctrl;
3408 }
3409 }
3410
3411 region->init_req(_obj_path, control());
3412 phi ->init_req(_obj_path, cast_obj);
3413
3414 // A merge of NULL or Casted-NotNull obj
3415 Node* res = _gvn.transform(phi);
3416
3417 // Note I do NOT always 'replace_in_map(obj,result)' here.
3418 // if( tk->klass()->can_be_primary_super() )
3419 // This means that if I successfully store an Object into an array-of-String
3420 // I 'forget' that the Object is really now known to be a String. I have to
3421 // do this because we don't have true union types for interfaces - if I store
3422 // a Baz into an array-of-Interface and then tell the optimizer it's an
3423 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3424 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3425 // replace_in_map( obj, res );
3426
3427 // Return final merged results
3428 set_control( _gvn.transform(region) );
3429 record_for_igvn(region);
3430
3431 return record_profiled_receiver_for_speculation(res);
3432 }
3433
3434 //------------------------------next_monitor-----------------------------------
3435 // What number should be given to the next monitor?
next_monitor()3436 int GraphKit::next_monitor() {
3437 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3438 int next = current + C->sync_stack_slots();
3439 // Keep the toplevel high water mark current:
3440 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3441 return current;
3442 }
3443
3444 //------------------------------insert_mem_bar---------------------------------
3445 // Memory barrier to avoid floating things around
3446 // The membar serves as a pinch point between both control and all memory slices.
insert_mem_bar(int opcode,Node * precedent)3447 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3448 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3449 mb->init_req(TypeFunc::Control, control());
3450 mb->init_req(TypeFunc::Memory, reset_memory());
3451 Node* membar = _gvn.transform(mb);
3452 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3453 set_all_memory_call(membar);
3454 return membar;
3455 }
3456
3457 //-------------------------insert_mem_bar_volatile----------------------------
3458 // Memory barrier to avoid floating things around
3459 // The membar serves as a pinch point between both control and memory(alias_idx).
3460 // If you want to make a pinch point on all memory slices, do not use this
3461 // function (even with AliasIdxBot); use insert_mem_bar() instead.
insert_mem_bar_volatile(int opcode,int alias_idx,Node * precedent)3462 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3463 // When Parse::do_put_xxx updates a volatile field, it appends a series
3464 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3465 // The first membar is on the same memory slice as the field store opcode.
3466 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3467 // All the other membars (for other volatile slices, including AliasIdxBot,
3468 // which stands for all unknown volatile slices) are control-dependent
3469 // on the first membar. This prevents later volatile loads or stores
3470 // from sliding up past the just-emitted store.
3471
3472 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3473 mb->set_req(TypeFunc::Control,control());
3474 if (alias_idx == Compile::AliasIdxBot) {
3475 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3476 } else {
3477 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3478 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3479 }
3480 Node* membar = _gvn.transform(mb);
3481 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3482 if (alias_idx == Compile::AliasIdxBot) {
3483 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3484 } else {
3485 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3486 }
3487 return membar;
3488 }
3489
3490 //------------------------------shared_lock------------------------------------
3491 // Emit locking code.
shared_lock(Node * obj)3492 FastLockNode* GraphKit::shared_lock(Node* obj) {
3493 // bci is either a monitorenter bc or InvocationEntryBci
3494 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3495 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3496
3497 if( !GenerateSynchronizationCode )
3498 return NULL; // Not locking things?
3499 if (stopped()) // Dead monitor?
3500 return NULL;
3501
3502 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3503
3504 // Box the stack location
3505 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3506 Node* mem = reset_memory();
3507
3508 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3509 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3510 // Create the counters for this fast lock.
3511 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3512 }
3513
3514 // Create the rtm counters for this fast lock if needed.
3515 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3516
3517 // Add monitor to debug info for the slow path. If we block inside the
3518 // slow path and de-opt, we need the monitor hanging around
3519 map()->push_monitor( flock );
3520
3521 const TypeFunc *tf = LockNode::lock_type();
3522 LockNode *lock = new LockNode(C, tf);
3523
3524 lock->init_req( TypeFunc::Control, control() );
3525 lock->init_req( TypeFunc::Memory , mem );
3526 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3527 lock->init_req( TypeFunc::FramePtr, frameptr() );
3528 lock->init_req( TypeFunc::ReturnAdr, top() );
3529
3530 lock->init_req(TypeFunc::Parms + 0, obj);
3531 lock->init_req(TypeFunc::Parms + 1, box);
3532 lock->init_req(TypeFunc::Parms + 2, flock);
3533 add_safepoint_edges(lock);
3534
3535 lock = _gvn.transform( lock )->as_Lock();
3536
3537 // lock has no side-effects, sets few values
3538 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3539
3540 insert_mem_bar(Op_MemBarAcquireLock);
3541
3542 // Add this to the worklist so that the lock can be eliminated
3543 record_for_igvn(lock);
3544
3545 #ifndef PRODUCT
3546 if (PrintLockStatistics) {
3547 // Update the counter for this lock. Don't bother using an atomic
3548 // operation since we don't require absolute accuracy.
3549 lock->create_lock_counter(map()->jvms());
3550 increment_counter(lock->counter()->addr());
3551 }
3552 #endif
3553
3554 return flock;
3555 }
3556
3557
3558 //------------------------------shared_unlock----------------------------------
3559 // Emit unlocking code.
shared_unlock(Node * box,Node * obj)3560 void GraphKit::shared_unlock(Node* box, Node* obj) {
3561 // bci is either a monitorenter bc or InvocationEntryBci
3562 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3563 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3564
3565 if( !GenerateSynchronizationCode )
3566 return;
3567 if (stopped()) { // Dead monitor?
3568 map()->pop_monitor(); // Kill monitor from debug info
3569 return;
3570 }
3571
3572 // Memory barrier to avoid floating things down past the locked region
3573 insert_mem_bar(Op_MemBarReleaseLock);
3574
3575 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3576 UnlockNode *unlock = new UnlockNode(C, tf);
3577 #ifdef ASSERT
3578 unlock->set_dbg_jvms(sync_jvms());
3579 #endif
3580 uint raw_idx = Compile::AliasIdxRaw;
3581 unlock->init_req( TypeFunc::Control, control() );
3582 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3583 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3584 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3585 unlock->init_req( TypeFunc::ReturnAdr, top() );
3586
3587 unlock->init_req(TypeFunc::Parms + 0, obj);
3588 unlock->init_req(TypeFunc::Parms + 1, box);
3589 unlock = _gvn.transform(unlock)->as_Unlock();
3590
3591 Node* mem = reset_memory();
3592
3593 // unlock has no side-effects, sets few values
3594 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3595
3596 // Kill monitor from debug info
3597 map()->pop_monitor( );
3598 }
3599
3600 //-------------------------------get_layout_helper-----------------------------
3601 // If the given klass is a constant or known to be an array,
3602 // fetch the constant layout helper value into constant_value
3603 // and return (Node*)NULL. Otherwise, load the non-constant
3604 // layout helper value, and return the node which represents it.
3605 // This two-faced routine is useful because allocation sites
3606 // almost always feature constant types.
get_layout_helper(Node * klass_node,jint & constant_value)3607 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3608 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3609 if (!StressReflectiveCode && inst_klass != NULL) {
3610 ciKlass* klass = inst_klass->klass();
3611 bool xklass = inst_klass->klass_is_exact();
3612 if (xklass || klass->is_array_klass()) {
3613 jint lhelper = klass->layout_helper();
3614 if (lhelper != Klass::_lh_neutral_value) {
3615 constant_value = lhelper;
3616 return (Node*) NULL;
3617 }
3618 }
3619 }
3620 constant_value = Klass::_lh_neutral_value; // put in a known value
3621 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3622 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3623 }
3624
3625 // We just put in an allocate/initialize with a big raw-memory effect.
3626 // 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)3627 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3628 MergeMemNode* init_in_merge,
3629 Node* init_out_raw) {
3630 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3631 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3632
3633 Node* prevmem = kit.memory(alias_idx);
3634 init_in_merge->set_memory_at(alias_idx, prevmem);
3635 kit.set_memory(init_out_raw, alias_idx);
3636 }
3637
3638 //---------------------------set_output_for_allocation-------------------------
set_output_for_allocation(AllocateNode * alloc,const TypeOopPtr * oop_type,bool deoptimize_on_exception)3639 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3640 const TypeOopPtr* oop_type,
3641 bool deoptimize_on_exception) {
3642 int rawidx = Compile::AliasIdxRaw;
3643 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3644 add_safepoint_edges(alloc);
3645 Node* allocx = _gvn.transform(alloc);
3646 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3647 // create memory projection for i_o
3648 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3649 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3650
3651 // create a memory projection as for the normal control path
3652 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3653 set_memory(malloc, rawidx);
3654
3655 // a normal slow-call doesn't change i_o, but an allocation does
3656 // we create a separate i_o projection for the normal control path
3657 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3658 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3659
3660 // put in an initialization barrier
3661 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3662 rawoop)->as_Initialize();
3663 assert(alloc->initialization() == init, "2-way macro link must work");
3664 assert(init ->allocation() == alloc, "2-way macro link must work");
3665 {
3666 // Extract memory strands which may participate in the new object's
3667 // initialization, and source them from the new InitializeNode.
3668 // This will allow us to observe initializations when they occur,
3669 // and link them properly (as a group) to the InitializeNode.
3670 assert(init->in(InitializeNode::Memory) == malloc, "");
3671 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3672 init->set_req(InitializeNode::Memory, minit_in);
3673 record_for_igvn(minit_in); // fold it up later, if possible
3674 Node* minit_out = memory(rawidx);
3675 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3676 // Add an edge in the MergeMem for the header fields so an access
3677 // to one of those has correct memory state
3678 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3679 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3680 if (oop_type->isa_aryptr()) {
3681 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3682 int elemidx = C->get_alias_index(telemref);
3683 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3684 } else if (oop_type->isa_instptr()) {
3685 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3686 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3687 ciField* field = ik->nonstatic_field_at(i);
3688 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3689 continue; // do not bother to track really large numbers of fields
3690 // Find (or create) the alias category for this field:
3691 int fieldidx = C->alias_type(field)->index();
3692 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3693 }
3694 }
3695 }
3696
3697 // Cast raw oop to the real thing...
3698 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3699 javaoop = _gvn.transform(javaoop);
3700 C->set_recent_alloc(control(), javaoop);
3701 assert(just_allocated_object(control()) == javaoop, "just allocated");
3702
3703 #ifdef ASSERT
3704 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3705 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3706 "Ideal_allocation works");
3707 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3708 "Ideal_allocation works");
3709 if (alloc->is_AllocateArray()) {
3710 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3711 "Ideal_allocation works");
3712 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3713 "Ideal_allocation works");
3714 } else {
3715 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3716 }
3717 }
3718 #endif //ASSERT
3719
3720 return javaoop;
3721 }
3722
3723 //---------------------------new_instance--------------------------------------
3724 // This routine takes a klass_node which may be constant (for a static type)
3725 // or may be non-constant (for reflective code). It will work equally well
3726 // for either, and the graph will fold nicely if the optimizer later reduces
3727 // the type to a constant.
3728 // The optional arguments are for specialized use by intrinsics:
3729 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3730 // - If 'return_size_val', report the the total object size to the caller.
3731 // - 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)3732 Node* GraphKit::new_instance(Node* klass_node,
3733 Node* extra_slow_test,
3734 Node* *return_size_val,
3735 bool deoptimize_on_exception) {
3736 // Compute size in doublewords
3737 // The size is always an integral number of doublewords, represented
3738 // as a positive bytewise size stored in the klass's layout_helper.
3739 // The layout_helper also encodes (in a low bit) the need for a slow path.
3740 jint layout_con = Klass::_lh_neutral_value;
3741 Node* layout_val = get_layout_helper(klass_node, layout_con);
3742 int layout_is_con = (layout_val == NULL);
3743
3744 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3745 // Generate the initial go-slow test. It's either ALWAYS (return a
3746 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3747 // case) a computed value derived from the layout_helper.
3748 Node* initial_slow_test = NULL;
3749 if (layout_is_con) {
3750 assert(!StressReflectiveCode, "stress mode does not use these paths");
3751 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3752 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3753 } else { // reflective case
3754 // This reflective path is used by Unsafe.allocateInstance.
3755 // (It may be stress-tested by specifying StressReflectiveCode.)
3756 // Basically, we want to get into the VM is there's an illegal argument.
3757 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3758 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3759 if (extra_slow_test != intcon(0)) {
3760 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3761 }
3762 // (Macro-expander will further convert this to a Bool, if necessary.)
3763 }
3764
3765 // Find the size in bytes. This is easy; it's the layout_helper.
3766 // The size value must be valid even if the slow path is taken.
3767 Node* size = NULL;
3768 if (layout_is_con) {
3769 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3770 } else { // reflective case
3771 // This reflective path is used by clone and Unsafe.allocateInstance.
3772 size = ConvI2X(layout_val);
3773
3774 // Clear the low bits to extract layout_helper_size_in_bytes:
3775 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3776 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3777 size = _gvn.transform( new AndXNode(size, mask) );
3778 }
3779 if (return_size_val != NULL) {
3780 (*return_size_val) = size;
3781 }
3782
3783 // This is a precise notnull oop of the klass.
3784 // (Actually, it need not be precise if this is a reflective allocation.)
3785 // It's what we cast the result to.
3786 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3787 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3788 const TypeOopPtr* oop_type = tklass->as_instance_type();
3789
3790 // Now generate allocation code
3791
3792 // The entire memory state is needed for slow path of the allocation
3793 // since GC and deoptimization can happened.
3794 Node *mem = reset_memory();
3795 set_all_memory(mem); // Create new memory state
3796
3797 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3798 control(), mem, i_o(),
3799 size, klass_node,
3800 initial_slow_test);
3801
3802 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3803 }
3804
3805 //-------------------------------new_array-------------------------------------
3806 // helper for both newarray and anewarray
3807 // The 'length' parameter is (obviously) the length of the array.
3808 // 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)3809 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3810 Node* length, // number of array elements
3811 int nargs, // number of arguments to push back for uncommon trap
3812 Node* *return_size_val,
3813 bool deoptimize_on_exception) {
3814 jint layout_con = Klass::_lh_neutral_value;
3815 Node* layout_val = get_layout_helper(klass_node, layout_con);
3816 int layout_is_con = (layout_val == NULL);
3817
3818 if (!layout_is_con && !StressReflectiveCode &&
3819 !too_many_traps(Deoptimization::Reason_class_check)) {
3820 // This is a reflective array creation site.
3821 // Optimistically assume that it is a subtype of Object[],
3822 // so that we can fold up all the address arithmetic.
3823 layout_con = Klass::array_layout_helper(T_OBJECT);
3824 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3825 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3826 { BuildCutout unless(this, bol_lh, PROB_MAX);
3827 inc_sp(nargs);
3828 uncommon_trap(Deoptimization::Reason_class_check,
3829 Deoptimization::Action_maybe_recompile);
3830 }
3831 layout_val = NULL;
3832 layout_is_con = true;
3833 }
3834
3835 // Generate the initial go-slow test. Make sure we do not overflow
3836 // if length is huge (near 2Gig) or negative! We do not need
3837 // exact double-words here, just a close approximation of needed
3838 // double-words. We can't add any offset or rounding bits, lest we
3839 // take a size -1 of bytes and make it positive. Use an unsigned
3840 // compare, so negative sizes look hugely positive.
3841 int fast_size_limit = FastAllocateSizeLimit;
3842 if (layout_is_con) {
3843 assert(!StressReflectiveCode, "stress mode does not use these paths");
3844 // Increase the size limit if we have exact knowledge of array type.
3845 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3846 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3847 }
3848
3849 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3850 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3851
3852 // --- Size Computation ---
3853 // array_size = round_to_heap(array_header + (length << elem_shift));
3854 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3855 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3856 // The rounding mask is strength-reduced, if possible.
3857 int round_mask = MinObjAlignmentInBytes - 1;
3858 Node* header_size = NULL;
3859 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3860 // (T_BYTE has the weakest alignment and size restrictions...)
3861 if (layout_is_con) {
3862 int hsize = Klass::layout_helper_header_size(layout_con);
3863 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3864 BasicType etype = Klass::layout_helper_element_type(layout_con);
3865 if ((round_mask & ~right_n_bits(eshift)) == 0)
3866 round_mask = 0; // strength-reduce it if it goes away completely
3867 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3868 assert(header_size_min <= hsize, "generic minimum is smallest");
3869 header_size_min = hsize;
3870 header_size = intcon(hsize + round_mask);
3871 } else {
3872 Node* hss = intcon(Klass::_lh_header_size_shift);
3873 Node* hsm = intcon(Klass::_lh_header_size_mask);
3874 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3875 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3876 Node* mask = intcon(round_mask);
3877 header_size = _gvn.transform( new AddINode(hsize, mask) );
3878 }
3879
3880 Node* elem_shift = NULL;
3881 if (layout_is_con) {
3882 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3883 if (eshift != 0)
3884 elem_shift = intcon(eshift);
3885 } else {
3886 // There is no need to mask or shift this value.
3887 // The semantics of LShiftINode include an implicit mask to 0x1F.
3888 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3889 elem_shift = layout_val;
3890 }
3891
3892 // Transition to native address size for all offset calculations:
3893 Node* lengthx = ConvI2X(length);
3894 Node* headerx = ConvI2X(header_size);
3895 #ifdef _LP64
3896 { const TypeInt* tilen = _gvn.find_int_type(length);
3897 if (tilen != NULL && tilen->_lo < 0) {
3898 // Add a manual constraint to a positive range. Cf. array_element_address.
3899 jint size_max = fast_size_limit;
3900 if (size_max > tilen->_hi) size_max = tilen->_hi;
3901 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3902
3903 // Only do a narrow I2L conversion if the range check passed.
3904 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3905 _gvn.transform(iff);
3906 RegionNode* region = new RegionNode(3);
3907 _gvn.set_type(region, Type::CONTROL);
3908 lengthx = new PhiNode(region, TypeLong::LONG);
3909 _gvn.set_type(lengthx, TypeLong::LONG);
3910
3911 // Range check passed. Use ConvI2L node with narrow type.
3912 Node* passed = IfFalse(iff);
3913 region->init_req(1, passed);
3914 // Make I2L conversion control dependent to prevent it from
3915 // floating above the range check during loop optimizations.
3916 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3917
3918 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3919 region->init_req(2, IfTrue(iff));
3920 lengthx->init_req(2, ConvI2X(length));
3921
3922 set_control(region);
3923 record_for_igvn(region);
3924 record_for_igvn(lengthx);
3925 }
3926 }
3927 #endif
3928
3929 // Combine header size (plus rounding) and body size. Then round down.
3930 // This computation cannot overflow, because it is used only in two
3931 // places, one where the length is sharply limited, and the other
3932 // after a successful allocation.
3933 Node* abody = lengthx;
3934 if (elem_shift != NULL)
3935 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3936 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3937 if (round_mask != 0) {
3938 Node* mask = MakeConX(~round_mask);
3939 size = _gvn.transform( new AndXNode(size, mask) );
3940 }
3941 // else if round_mask == 0, the size computation is self-rounding
3942
3943 if (return_size_val != NULL) {
3944 // This is the size
3945 (*return_size_val) = size;
3946 }
3947
3948 // Now generate allocation code
3949
3950 // The entire memory state is needed for slow path of the allocation
3951 // since GC and deoptimization can happened.
3952 Node *mem = reset_memory();
3953 set_all_memory(mem); // Create new memory state
3954
3955 if (initial_slow_test->is_Bool()) {
3956 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3957 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3958 }
3959
3960 // Create the AllocateArrayNode and its result projections
3961 AllocateArrayNode* alloc
3962 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3963 control(), mem, i_o(),
3964 size, klass_node,
3965 initial_slow_test,
3966 length);
3967
3968 // Cast to correct type. Note that the klass_node may be constant or not,
3969 // and in the latter case the actual array type will be inexact also.
3970 // (This happens via a non-constant argument to inline_native_newArray.)
3971 // In any case, the value of klass_node provides the desired array type.
3972 const TypeInt* length_type = _gvn.find_int_type(length);
3973 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3974 if (ary_type->isa_aryptr() && length_type != NULL) {
3975 // Try to get a better type than POS for the size
3976 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3977 }
3978
3979 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3980
3981 // Cast length on remaining path to be as narrow as possible
3982 if (map()->find_edge(length) >= 0) {
3983 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3984 if (ccast != length) {
3985 _gvn.set_type_bottom(ccast);
3986 record_for_igvn(ccast);
3987 replace_in_map(length, ccast);
3988 }
3989 }
3990
3991 return javaoop;
3992 }
3993
3994 // The following "Ideal_foo" functions are placed here because they recognize
3995 // the graph shapes created by the functions immediately above.
3996
3997 //---------------------------Ideal_allocation----------------------------------
3998 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
Ideal_allocation(Node * ptr,PhaseTransform * phase)3999 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
4000 if (ptr == NULL) { // reduce dumb test in callers
4001 return NULL;
4002 }
4003
4004 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4005 ptr = bs->step_over_gc_barrier(ptr);
4006
4007 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4008 ptr = ptr->in(1);
4009 if (ptr == NULL) return NULL;
4010 }
4011 // Return NULL for allocations with several casts:
4012 // j.l.reflect.Array.newInstance(jobject, jint)
4013 // Object.clone()
4014 // to keep more precise type from last cast.
4015 if (ptr->is_Proj()) {
4016 Node* allo = ptr->in(0);
4017 if (allo != NULL && allo->is_Allocate()) {
4018 return allo->as_Allocate();
4019 }
4020 }
4021 // Report failure to match.
4022 return NULL;
4023 }
4024
4025 // Fancy version which also strips off an offset (and reports it to caller).
Ideal_allocation(Node * ptr,PhaseTransform * phase,intptr_t & offset)4026 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
4027 intptr_t& offset) {
4028 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4029 if (base == NULL) return NULL;
4030 return Ideal_allocation(base, phase);
4031 }
4032
4033 // Trace Initialize <- Proj[Parm] <- Allocate
allocation()4034 AllocateNode* InitializeNode::allocation() {
4035 Node* rawoop = in(InitializeNode::RawAddress);
4036 if (rawoop->is_Proj()) {
4037 Node* alloc = rawoop->in(0);
4038 if (alloc->is_Allocate()) {
4039 return alloc->as_Allocate();
4040 }
4041 }
4042 return NULL;
4043 }
4044
4045 // Trace Allocate -> Proj[Parm] -> Initialize
initialization()4046 InitializeNode* AllocateNode::initialization() {
4047 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4048 if (rawoop == NULL) return NULL;
4049 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4050 Node* init = rawoop->fast_out(i);
4051 if (init->is_Initialize()) {
4052 assert(init->as_Initialize()->allocation() == this, "2-way link");
4053 return init->as_Initialize();
4054 }
4055 }
4056 return NULL;
4057 }
4058
4059 //----------------------------- loop predicates ---------------------------
4060
4061 //------------------------------add_predicate_impl----------------------------
add_empty_predicate_impl(Deoptimization::DeoptReason reason,int nargs)4062 void GraphKit::add_empty_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
4063 // Too many traps seen?
4064 if (too_many_traps(reason)) {
4065 #ifdef ASSERT
4066 if (TraceLoopPredicate) {
4067 int tc = C->trap_count(reason);
4068 tty->print("too many traps=%s tcount=%d in ",
4069 Deoptimization::trap_reason_name(reason), tc);
4070 method()->print(); // which method has too many predicate traps
4071 tty->cr();
4072 }
4073 #endif
4074 // We cannot afford to take more traps here,
4075 // do not generate predicate.
4076 return;
4077 }
4078
4079 Node *cont = _gvn.intcon(1);
4080 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
4081 Node *bol = _gvn.transform(new Conv2BNode(opq));
4082 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4083 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
4084 C->add_predicate_opaq(opq);
4085 {
4086 PreserveJVMState pjvms(this);
4087 set_control(iffalse);
4088 inc_sp(nargs);
4089 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4090 }
4091 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
4092 set_control(iftrue);
4093 }
4094
4095 //------------------------------add_predicate---------------------------------
add_empty_predicates(int nargs)4096 void GraphKit::add_empty_predicates(int nargs) {
4097 // These loop predicates remain empty. All concrete loop predicates are inserted above the corresponding
4098 // empty loop predicate later by 'PhaseIdealLoop::create_new_if_for_predicate'. All concrete loop predicates of
4099 // a specific kind (normal, profile or limit check) share the same uncommon trap as the empty loop predicate.
4100 if (UseLoopPredicate) {
4101 add_empty_predicate_impl(Deoptimization::Reason_predicate, nargs);
4102 }
4103 if (UseProfiledLoopPredicate) {
4104 add_empty_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
4105 }
4106 // loop's limit check predicate should be near the loop.
4107 add_empty_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
4108 }
4109
sync_kit(IdealKit & ideal)4110 void GraphKit::sync_kit(IdealKit& ideal) {
4111 set_all_memory(ideal.merged_memory());
4112 set_i_o(ideal.i_o());
4113 set_control(ideal.ctrl());
4114 }
4115
final_sync(IdealKit & ideal)4116 void GraphKit::final_sync(IdealKit& ideal) {
4117 // Final sync IdealKit and graphKit.
4118 sync_kit(ideal);
4119 }
4120
load_String_length(Node * str,bool set_ctrl)4121 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4122 Node* len = load_array_length(load_String_value(str, set_ctrl));
4123 Node* coder = load_String_coder(str, set_ctrl);
4124 // Divide length by 2 if coder is UTF16
4125 return _gvn.transform(new RShiftINode(len, coder));
4126 }
4127
load_String_value(Node * str,bool set_ctrl)4128 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4129 int value_offset = java_lang_String::value_offset();
4130 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4131 false, NULL, 0);
4132 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4133 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4134 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4135 ciTypeArrayKlass::make(T_BYTE), true, 0);
4136 Node* p = basic_plus_adr(str, str, value_offset);
4137 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4138 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4139 return load;
4140 }
4141
load_String_coder(Node * str,bool set_ctrl)4142 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4143 if (!CompactStrings) {
4144 return intcon(java_lang_String::CODER_UTF16);
4145 }
4146 int coder_offset = java_lang_String::coder_offset();
4147 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4148 false, NULL, 0);
4149 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4150
4151 Node* p = basic_plus_adr(str, str, coder_offset);
4152 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4153 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4154 return load;
4155 }
4156
store_String_value(Node * str,Node * value)4157 void GraphKit::store_String_value(Node* str, Node* value) {
4158 int value_offset = java_lang_String::value_offset();
4159 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4160 false, NULL, 0);
4161 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4162
4163 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4164 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4165 }
4166
store_String_coder(Node * str,Node * value)4167 void GraphKit::store_String_coder(Node* str, Node* value) {
4168 int coder_offset = java_lang_String::coder_offset();
4169 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4170 false, NULL, 0);
4171 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4172
4173 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4174 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4175 }
4176
4177 // Capture src and dst memory state with a MergeMemNode
capture_memory(const TypePtr * src_type,const TypePtr * dst_type)4178 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4179 if (src_type == dst_type) {
4180 // Types are equal, we don't need a MergeMemNode
4181 return memory(src_type);
4182 }
4183 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4184 record_for_igvn(merge); // fold it up later, if possible
4185 int src_idx = C->get_alias_index(src_type);
4186 int dst_idx = C->get_alias_index(dst_type);
4187 merge->set_memory_at(src_idx, memory(src_idx));
4188 merge->set_memory_at(dst_idx, memory(dst_idx));
4189 return merge;
4190 }
4191
compress_string(Node * src,const TypeAryPtr * src_type,Node * dst,Node * count)4192 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4193 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4194 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4195 // If input and output memory types differ, capture both states to preserve
4196 // the dependency between preceding and subsequent loads/stores.
4197 // For example, the following program:
4198 // StoreB
4199 // compress_string
4200 // LoadB
4201 // has this memory graph (use->def):
4202 // LoadB -> compress_string -> CharMem
4203 // ... -> StoreB -> ByteMem
4204 // The intrinsic hides the dependency between LoadB and StoreB, causing
4205 // the load to read from memory not containing the result of the StoreB.
4206 // The correct memory graph should look like this:
4207 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4208 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4209 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4210 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4211 set_memory(res_mem, TypeAryPtr::BYTES);
4212 return str;
4213 }
4214
inflate_string(Node * src,Node * dst,const TypeAryPtr * dst_type,Node * count)4215 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4216 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4217 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4218 // Capture src and dst memory (see comment in 'compress_string').
4219 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4220 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4221 set_memory(_gvn.transform(str), dst_type);
4222 }
4223
inflate_string_slow(Node * src,Node * dst,Node * start,Node * count)4224 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4225 /**
4226 * int i_char = start;
4227 * for (int i_byte = 0; i_byte < count; i_byte++) {
4228 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4229 * }
4230 */
4231 add_empty_predicates();
4232 RegionNode* head = new RegionNode(3);
4233 head->init_req(1, control());
4234 gvn().set_type(head, Type::CONTROL);
4235 record_for_igvn(head);
4236
4237 Node* i_byte = new PhiNode(head, TypeInt::INT);
4238 i_byte->init_req(1, intcon(0));
4239 gvn().set_type(i_byte, TypeInt::INT);
4240 record_for_igvn(i_byte);
4241
4242 Node* i_char = new PhiNode(head, TypeInt::INT);
4243 i_char->init_req(1, start);
4244 gvn().set_type(i_char, TypeInt::INT);
4245 record_for_igvn(i_char);
4246
4247 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4248 gvn().set_type(mem, Type::MEMORY);
4249 record_for_igvn(mem);
4250 set_control(head);
4251 set_memory(mem, TypeAryPtr::BYTES);
4252 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4253 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4254 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4255 false, false, true /* mismatched */);
4256
4257 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4258 head->init_req(2, IfTrue(iff));
4259 mem->init_req(2, st);
4260 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4261 i_char->init_req(2, AddI(i_char, intcon(2)));
4262
4263 set_control(IfFalse(iff));
4264 set_memory(st, TypeAryPtr::BYTES);
4265 }
4266
make_constant_from_field(ciField * field,Node * obj)4267 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4268 if (!field->is_constant()) {
4269 return NULL; // Field not marked as constant.
4270 }
4271 ciInstance* holder = NULL;
4272 if (!field->is_static()) {
4273 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4274 if (const_oop != NULL && const_oop->is_instance()) {
4275 holder = const_oop->as_instance();
4276 }
4277 }
4278 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4279 /*is_unsigned_load=*/false);
4280 if (con_type != NULL) {
4281 return makecon(con_type);
4282 }
4283 return NULL;
4284 }
4285