Wed, 10 Aug 2016 14:59:21 +0200
8152172: PPC64: Support AES intrinsics
Summary: Add support for AES intrinsics on PPC64.
Reviewed-by: kvn, mdoerr, simonis, zmajo
Contributed-by: Hiroshi H Horii <horii@jp.ibm.com>
1 /*
2 * Copyright (c) 2001, 2015, 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 */
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
28 #include "gc_implementation/g1/heapRegion.hpp"
29 #include "gc_interface/collectedHeap.hpp"
30 #include "memory/barrierSet.hpp"
31 #include "memory/cardTableModRefBS.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/graphKit.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "runtime/deoptimization.hpp"
41 #include "runtime/sharedRuntime.hpp"
43 //----------------------------GraphKit-----------------------------------------
44 // Main utility constructor.
45 GraphKit::GraphKit(JVMState* jvms)
46 : Phase(Phase::Parser),
47 _env(C->env()),
48 _gvn(*C->initial_gvn())
49 {
50 _exceptions = jvms->map()->next_exception();
51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
52 set_jvms(jvms);
53 }
55 // Private constructor for parser.
56 GraphKit::GraphKit()
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn())
60 {
61 _exceptions = NULL;
62 set_map(NULL);
63 debug_only(_sp = -99);
64 debug_only(set_bci(-99));
65 }
69 //---------------------------clean_stack---------------------------------------
70 // Clear away rubbish from the stack area of the JVM state.
71 // This destroys any arguments that may be waiting on the stack.
72 void GraphKit::clean_stack(int from_sp) {
73 SafePointNode* map = this->map();
74 JVMState* jvms = this->jvms();
75 int stk_size = jvms->stk_size();
76 int stkoff = jvms->stkoff();
77 Node* top = this->top();
78 for (int i = from_sp; i < stk_size; i++) {
79 if (map->in(stkoff + i) != top) {
80 map->set_req(stkoff + i, top);
81 }
82 }
83 }
86 //--------------------------------sync_jvms-----------------------------------
87 // Make sure our current jvms agrees with our parse state.
88 JVMState* GraphKit::sync_jvms() const {
89 JVMState* jvms = this->jvms();
90 jvms->set_bci(bci()); // Record the new bci in the JVMState
91 jvms->set_sp(sp()); // Record the new sp in the JVMState
92 assert(jvms_in_sync(), "jvms is now in sync");
93 return jvms;
94 }
96 //--------------------------------sync_jvms_for_reexecute---------------------
97 // Make sure our current jvms agrees with our parse state. This version
98 // uses the reexecute_sp for reexecuting bytecodes.
99 JVMState* GraphKit::sync_jvms_for_reexecute() {
100 JVMState* jvms = this->jvms();
101 jvms->set_bci(bci()); // Record the new bci in the JVMState
102 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
103 return jvms;
104 }
106 #ifdef ASSERT
107 bool GraphKit::jvms_in_sync() const {
108 Parse* parse = is_Parse();
109 if (parse == NULL) {
110 if (bci() != jvms()->bci()) return false;
111 if (sp() != (int)jvms()->sp()) return false;
112 return true;
113 }
114 if (jvms()->method() != parse->method()) return false;
115 if (jvms()->bci() != parse->bci()) return false;
116 int jvms_sp = jvms()->sp();
117 if (jvms_sp != parse->sp()) return false;
118 int jvms_depth = jvms()->depth();
119 if (jvms_depth != parse->depth()) return false;
120 return true;
121 }
123 // Local helper checks for special internal merge points
124 // used to accumulate and merge exception states.
125 // They are marked by the region's in(0) edge being the map itself.
126 // Such merge points must never "escape" into the parser at large,
127 // until they have been handed to gvn.transform.
128 static bool is_hidden_merge(Node* reg) {
129 if (reg == NULL) return false;
130 if (reg->is_Phi()) {
131 reg = reg->in(0);
132 if (reg == NULL) return false;
133 }
134 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
135 }
137 void GraphKit::verify_map() const {
138 if (map() == NULL) return; // null map is OK
139 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
140 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
141 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
142 }
144 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
145 assert(ex_map->next_exception() == NULL, "not already part of a chain");
146 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
147 }
148 #endif
150 //---------------------------stop_and_kill_map---------------------------------
151 // Set _map to NULL, signalling a stop to further bytecode execution.
152 // First smash the current map's control to a constant, to mark it dead.
153 void GraphKit::stop_and_kill_map() {
154 SafePointNode* dead_map = stop();
155 if (dead_map != NULL) {
156 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
157 assert(dead_map->is_killed(), "must be so marked");
158 }
159 }
162 //--------------------------------stopped--------------------------------------
163 // Tell if _map is NULL, or control is top.
164 bool GraphKit::stopped() {
165 if (map() == NULL) return true;
166 else if (control() == top()) return true;
167 else return false;
168 }
171 //-----------------------------has_ex_handler----------------------------------
172 // Tell if this method or any caller method has exception handlers.
173 bool GraphKit::has_ex_handler() {
174 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
175 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
176 return true;
177 }
178 }
179 return false;
180 }
182 //------------------------------save_ex_oop------------------------------------
183 // Save an exception without blowing stack contents or other JVM state.
184 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
185 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
186 ex_map->add_req(ex_oop);
187 debug_only(verify_exception_state(ex_map));
188 }
190 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
191 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
192 Node* ex_oop = ex_map->in(ex_map->req()-1);
193 if (clear_it) ex_map->del_req(ex_map->req()-1);
194 return ex_oop;
195 }
197 //-----------------------------saved_ex_oop------------------------------------
198 // Recover a saved exception from its map.
199 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
200 return common_saved_ex_oop(ex_map, false);
201 }
203 //--------------------------clear_saved_ex_oop---------------------------------
204 // Erase a previously saved exception from its map.
205 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
206 return common_saved_ex_oop(ex_map, true);
207 }
209 #ifdef ASSERT
210 //---------------------------has_saved_ex_oop----------------------------------
211 // Erase a previously saved exception from its map.
212 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
213 return ex_map->req() == ex_map->jvms()->endoff()+1;
214 }
215 #endif
217 //-------------------------make_exception_state--------------------------------
218 // Turn the current JVM state into an exception state, appending the ex_oop.
219 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
220 sync_jvms();
221 SafePointNode* ex_map = stop(); // do not manipulate this map any more
222 set_saved_ex_oop(ex_map, ex_oop);
223 return ex_map;
224 }
227 //--------------------------add_exception_state--------------------------------
228 // Add an exception to my list of exceptions.
229 void GraphKit::add_exception_state(SafePointNode* ex_map) {
230 if (ex_map == NULL || ex_map->control() == top()) {
231 return;
232 }
233 #ifdef ASSERT
234 verify_exception_state(ex_map);
235 if (has_exceptions()) {
236 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
237 }
238 #endif
240 // If there is already an exception of exactly this type, merge with it.
241 // In particular, null-checks and other low-level exceptions common up here.
242 Node* ex_oop = saved_ex_oop(ex_map);
243 const Type* ex_type = _gvn.type(ex_oop);
244 if (ex_oop == top()) {
245 // No action needed.
246 return;
247 }
248 assert(ex_type->isa_instptr(), "exception must be an instance");
249 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
250 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
251 // We check sp also because call bytecodes can generate exceptions
252 // both before and after arguments are popped!
253 if (ex_type2 == ex_type
254 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
255 combine_exception_states(ex_map, e2);
256 return;
257 }
258 }
260 // No pre-existing exception of the same type. Chain it on the list.
261 push_exception_state(ex_map);
262 }
264 //-----------------------add_exception_states_from-----------------------------
265 void GraphKit::add_exception_states_from(JVMState* jvms) {
266 SafePointNode* ex_map = jvms->map()->next_exception();
267 if (ex_map != NULL) {
268 jvms->map()->set_next_exception(NULL);
269 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
270 next_map = ex_map->next_exception();
271 ex_map->set_next_exception(NULL);
272 add_exception_state(ex_map);
273 }
274 }
275 }
277 //-----------------------transfer_exceptions_into_jvms-------------------------
278 JVMState* GraphKit::transfer_exceptions_into_jvms() {
279 if (map() == NULL) {
280 // We need a JVMS to carry the exceptions, but the map has gone away.
281 // Create a scratch JVMS, cloned from any of the exception states...
282 if (has_exceptions()) {
283 _map = _exceptions;
284 _map = clone_map();
285 _map->set_next_exception(NULL);
286 clear_saved_ex_oop(_map);
287 debug_only(verify_map());
288 } else {
289 // ...or created from scratch
290 JVMState* jvms = new (C) JVMState(_method, NULL);
291 jvms->set_bci(_bci);
292 jvms->set_sp(_sp);
293 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms));
294 set_jvms(jvms);
295 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
296 set_all_memory(top());
297 while (map()->req() < jvms->endoff()) map()->add_req(top());
298 }
299 // (This is a kludge, in case you didn't notice.)
300 set_control(top());
301 }
302 JVMState* jvms = sync_jvms();
303 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
304 jvms->map()->set_next_exception(_exceptions);
305 _exceptions = NULL; // done with this set of exceptions
306 return jvms;
307 }
309 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
310 assert(is_hidden_merge(dstphi), "must be a special merge node");
311 assert(is_hidden_merge(srcphi), "must be a special merge node");
312 uint limit = srcphi->req();
313 for (uint i = PhiNode::Input; i < limit; i++) {
314 dstphi->add_req(srcphi->in(i));
315 }
316 }
317 static inline void add_one_req(Node* dstphi, Node* src) {
318 assert(is_hidden_merge(dstphi), "must be a special merge node");
319 assert(!is_hidden_merge(src), "must not be a special merge node");
320 dstphi->add_req(src);
321 }
323 //-----------------------combine_exception_states------------------------------
324 // This helper function combines exception states by building phis on a
325 // specially marked state-merging region. These regions and phis are
326 // untransformed, and can build up gradually. The region is marked by
327 // having a control input of its exception map, rather than NULL. Such
328 // regions do not appear except in this function, and in use_exception_state.
329 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
330 if (failing()) return; // dying anyway...
331 JVMState* ex_jvms = ex_map->_jvms;
332 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
333 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
334 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
335 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
336 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
337 assert(ex_map->req() == phi_map->req(), "matching maps");
338 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
339 Node* hidden_merge_mark = root();
340 Node* region = phi_map->control();
341 MergeMemNode* phi_mem = phi_map->merged_memory();
342 MergeMemNode* ex_mem = ex_map->merged_memory();
343 if (region->in(0) != hidden_merge_mark) {
344 // The control input is not (yet) a specially-marked region in phi_map.
345 // Make it so, and build some phis.
346 region = new (C) RegionNode(2);
347 _gvn.set_type(region, Type::CONTROL);
348 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
349 region->init_req(1, phi_map->control());
350 phi_map->set_control(region);
351 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
352 record_for_igvn(io_phi);
353 _gvn.set_type(io_phi, Type::ABIO);
354 phi_map->set_i_o(io_phi);
355 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
356 Node* m = mms.memory();
357 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
358 record_for_igvn(m_phi);
359 _gvn.set_type(m_phi, Type::MEMORY);
360 mms.set_memory(m_phi);
361 }
362 }
364 // Either or both of phi_map and ex_map might already be converted into phis.
365 Node* ex_control = ex_map->control();
366 // if there is special marking on ex_map also, we add multiple edges from src
367 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
368 // how wide was the destination phi_map, originally?
369 uint orig_width = region->req();
371 if (add_multiple) {
372 add_n_reqs(region, ex_control);
373 add_n_reqs(phi_map->i_o(), ex_map->i_o());
374 } else {
375 // ex_map has no merges, so we just add single edges everywhere
376 add_one_req(region, ex_control);
377 add_one_req(phi_map->i_o(), ex_map->i_o());
378 }
379 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
380 if (mms.is_empty()) {
381 // get a copy of the base memory, and patch some inputs into it
382 const TypePtr* adr_type = mms.adr_type(C);
383 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
384 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
385 mms.set_memory(phi);
386 // Prepare to append interesting stuff onto the newly sliced phi:
387 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
388 }
389 // Append stuff from ex_map:
390 if (add_multiple) {
391 add_n_reqs(mms.memory(), mms.memory2());
392 } else {
393 add_one_req(mms.memory(), mms.memory2());
394 }
395 }
396 uint limit = ex_map->req();
397 for (uint i = TypeFunc::Parms; i < limit; i++) {
398 // Skip everything in the JVMS after tos. (The ex_oop follows.)
399 if (i == tos) i = ex_jvms->monoff();
400 Node* src = ex_map->in(i);
401 Node* dst = phi_map->in(i);
402 if (src != dst) {
403 PhiNode* phi;
404 if (dst->in(0) != region) {
405 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
406 record_for_igvn(phi);
407 _gvn.set_type(phi, phi->type());
408 phi_map->set_req(i, dst);
409 // Prepare to append interesting stuff onto the new phi:
410 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
411 } else {
412 assert(dst->is_Phi(), "nobody else uses a hidden region");
413 phi = dst->as_Phi();
414 }
415 if (add_multiple && src->in(0) == ex_control) {
416 // Both are phis.
417 add_n_reqs(dst, src);
418 } else {
419 while (dst->req() < region->req()) add_one_req(dst, src);
420 }
421 const Type* srctype = _gvn.type(src);
422 if (phi->type() != srctype) {
423 const Type* dsttype = phi->type()->meet_speculative(srctype);
424 if (phi->type() != dsttype) {
425 phi->set_type(dsttype);
426 _gvn.set_type(phi, dsttype);
427 }
428 }
429 }
430 }
431 phi_map->merge_replaced_nodes_with(ex_map);
432 }
434 //--------------------------use_exception_state--------------------------------
435 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
436 if (failing()) { stop(); return top(); }
437 Node* region = phi_map->control();
438 Node* hidden_merge_mark = root();
439 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
440 Node* ex_oop = clear_saved_ex_oop(phi_map);
441 if (region->in(0) == hidden_merge_mark) {
442 // Special marking for internal ex-states. Process the phis now.
443 region->set_req(0, region); // now it's an ordinary region
444 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
445 // Note: Setting the jvms also sets the bci and sp.
446 set_control(_gvn.transform(region));
447 uint tos = jvms()->stkoff() + sp();
448 for (uint i = 1; i < tos; i++) {
449 Node* x = phi_map->in(i);
450 if (x->in(0) == region) {
451 assert(x->is_Phi(), "expected a special phi");
452 phi_map->set_req(i, _gvn.transform(x));
453 }
454 }
455 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
456 Node* x = mms.memory();
457 if (x->in(0) == region) {
458 assert(x->is_Phi(), "nobody else uses a hidden region");
459 mms.set_memory(_gvn.transform(x));
460 }
461 }
462 if (ex_oop->in(0) == region) {
463 assert(ex_oop->is_Phi(), "expected a special phi");
464 ex_oop = _gvn.transform(ex_oop);
465 }
466 } else {
467 set_jvms(phi_map->jvms());
468 }
470 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
471 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
472 return ex_oop;
473 }
475 //---------------------------------java_bc-------------------------------------
476 Bytecodes::Code GraphKit::java_bc() const {
477 ciMethod* method = this->method();
478 int bci = this->bci();
479 if (method != NULL && bci != InvocationEntryBci)
480 return method->java_code_at_bci(bci);
481 else
482 return Bytecodes::_illegal;
483 }
485 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
486 bool must_throw) {
487 // if the exception capability is set, then we will generate code
488 // to check the JavaThread.should_post_on_exceptions flag to see
489 // if we actually need to report exception events (for this
490 // thread). If we don't need to report exception events, we will
491 // take the normal fast path provided by add_exception_events. If
492 // exception event reporting is enabled for this thread, we will
493 // take the uncommon_trap in the BuildCutout below.
495 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
496 Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
497 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
498 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
500 // Test the should_post_on_exceptions_flag vs. 0
501 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
502 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
504 // Branch to slow_path if should_post_on_exceptions_flag was true
505 { BuildCutout unless(this, tst, PROB_MAX);
506 // Do not try anything fancy if we're notifying the VM on every throw.
507 // Cf. case Bytecodes::_athrow in parse2.cpp.
508 uncommon_trap(reason, Deoptimization::Action_none,
509 (ciKlass*)NULL, (char*)NULL, must_throw);
510 }
512 }
514 //------------------------------builtin_throw----------------------------------
515 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
516 bool must_throw = true;
518 if (env()->jvmti_can_post_on_exceptions()) {
519 // check if we must post exception events, take uncommon trap if so
520 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
521 // here if should_post_on_exceptions is false
522 // continue on with the normal codegen
523 }
525 // If this particular condition has not yet happened at this
526 // bytecode, then use the uncommon trap mechanism, and allow for
527 // a future recompilation if several traps occur here.
528 // If the throw is hot, try to use a more complicated inline mechanism
529 // which keeps execution inside the compiled code.
530 bool treat_throw_as_hot = false;
531 ciMethodData* md = method()->method_data();
533 if (ProfileTraps) {
534 if (too_many_traps(reason)) {
535 treat_throw_as_hot = true;
536 }
537 // (If there is no MDO at all, assume it is early in
538 // execution, and that any deopts are part of the
539 // startup transient, and don't need to be remembered.)
541 // Also, if there is a local exception handler, treat all throws
542 // as hot if there has been at least one in this method.
543 if (C->trap_count(reason) != 0
544 && method()->method_data()->trap_count(reason) != 0
545 && has_ex_handler()) {
546 treat_throw_as_hot = true;
547 }
548 }
550 // If this throw happens frequently, an uncommon trap might cause
551 // a performance pothole. If there is a local exception handler,
552 // and if this particular bytecode appears to be deoptimizing often,
553 // let us handle the throw inline, with a preconstructed instance.
554 // Note: If the deopt count has blown up, the uncommon trap
555 // runtime is going to flush this nmethod, not matter what.
556 if (treat_throw_as_hot
557 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
558 // If the throw is local, we use a pre-existing instance and
559 // punt on the backtrace. This would lead to a missing backtrace
560 // (a repeat of 4292742) if the backtrace object is ever asked
561 // for its backtrace.
562 // Fixing this remaining case of 4292742 requires some flavor of
563 // escape analysis. Leave that for the future.
564 ciInstance* ex_obj = NULL;
565 switch (reason) {
566 case Deoptimization::Reason_null_check:
567 ex_obj = env()->NullPointerException_instance();
568 break;
569 case Deoptimization::Reason_div0_check:
570 ex_obj = env()->ArithmeticException_instance();
571 break;
572 case Deoptimization::Reason_range_check:
573 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
574 break;
575 case Deoptimization::Reason_class_check:
576 if (java_bc() == Bytecodes::_aastore) {
577 ex_obj = env()->ArrayStoreException_instance();
578 } else {
579 ex_obj = env()->ClassCastException_instance();
580 }
581 break;
582 }
583 if (failing()) { stop(); return; } // exception allocation might fail
584 if (ex_obj != NULL) {
585 // Cheat with a preallocated exception object.
586 if (C->log() != NULL)
587 C->log()->elem("hot_throw preallocated='1' reason='%s'",
588 Deoptimization::trap_reason_name(reason));
589 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
590 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
592 // Clear the detail message of the preallocated exception object.
593 // Weblogic sometimes mutates the detail message of exceptions
594 // using reflection.
595 int offset = java_lang_Throwable::get_detailMessage_offset();
596 const TypePtr* adr_typ = ex_con->add_offset(offset);
598 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
599 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
600 // Conservatively release stores of object references.
601 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
603 add_exception_state(make_exception_state(ex_node));
604 return;
605 }
606 }
608 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
609 // It won't be much cheaper than bailing to the interp., since we'll
610 // have to pass up all the debug-info, and the runtime will have to
611 // create the stack trace.
613 // Usual case: Bail to interpreter.
614 // Reserve the right to recompile if we haven't seen anything yet.
616 assert(!Deoptimization::reason_is_speculate(reason), "unsupported");
617 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
618 if (treat_throw_as_hot
619 && (method()->method_data()->trap_recompiled_at(bci(), NULL)
620 || C->too_many_traps(reason))) {
621 // We cannot afford to take more traps here. Suffer in the interpreter.
622 if (C->log() != NULL)
623 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
624 Deoptimization::trap_reason_name(reason),
625 C->trap_count(reason));
626 action = Deoptimization::Action_none;
627 }
629 // "must_throw" prunes the JVM state to include only the stack, if there
630 // are no local exception handlers. This should cut down on register
631 // allocation time and code size, by drastically reducing the number
632 // of in-edges on the call to the uncommon trap.
634 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
635 }
638 //----------------------------PreserveJVMState---------------------------------
639 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
640 debug_only(kit->verify_map());
641 _kit = kit;
642 _map = kit->map(); // preserve the map
643 _sp = kit->sp();
644 kit->set_map(clone_map ? kit->clone_map() : NULL);
645 #ifdef ASSERT
646 _bci = kit->bci();
647 Parse* parser = kit->is_Parse();
648 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
649 _block = block;
650 #endif
651 }
652 PreserveJVMState::~PreserveJVMState() {
653 GraphKit* kit = _kit;
654 #ifdef ASSERT
655 assert(kit->bci() == _bci, "bci must not shift");
656 Parse* parser = kit->is_Parse();
657 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
658 assert(block == _block, "block must not shift");
659 #endif
660 kit->set_map(_map);
661 kit->set_sp(_sp);
662 }
665 //-----------------------------BuildCutout-------------------------------------
666 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
667 : PreserveJVMState(kit)
668 {
669 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
670 SafePointNode* outer_map = _map; // preserved map is caller's
671 SafePointNode* inner_map = kit->map();
672 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
673 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
674 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
675 }
676 BuildCutout::~BuildCutout() {
677 GraphKit* kit = _kit;
678 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
679 }
681 //---------------------------PreserveReexecuteState----------------------------
682 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
683 assert(!kit->stopped(), "must call stopped() before");
684 _kit = kit;
685 _sp = kit->sp();
686 _reexecute = kit->jvms()->_reexecute;
687 }
688 PreserveReexecuteState::~PreserveReexecuteState() {
689 if (_kit->stopped()) return;
690 _kit->jvms()->_reexecute = _reexecute;
691 _kit->set_sp(_sp);
692 }
694 //------------------------------clone_map--------------------------------------
695 // Implementation of PreserveJVMState
696 //
697 // Only clone_map(...) here. If this function is only used in the
698 // PreserveJVMState class we may want to get rid of this extra
699 // function eventually and do it all there.
701 SafePointNode* GraphKit::clone_map() {
702 if (map() == NULL) return NULL;
704 // Clone the memory edge first
705 Node* mem = MergeMemNode::make(C, map()->memory());
706 gvn().set_type_bottom(mem);
708 SafePointNode *clonemap = (SafePointNode*)map()->clone();
709 JVMState* jvms = this->jvms();
710 JVMState* clonejvms = jvms->clone_shallow(C);
711 clonemap->set_memory(mem);
712 clonemap->set_jvms(clonejvms);
713 clonejvms->set_map(clonemap);
714 record_for_igvn(clonemap);
715 gvn().set_type_bottom(clonemap);
716 return clonemap;
717 }
720 //-----------------------------set_map_clone-----------------------------------
721 void GraphKit::set_map_clone(SafePointNode* m) {
722 _map = m;
723 _map = clone_map();
724 _map->set_next_exception(NULL);
725 debug_only(verify_map());
726 }
729 //----------------------------kill_dead_locals---------------------------------
730 // Detect any locals which are known to be dead, and force them to top.
731 void GraphKit::kill_dead_locals() {
732 // Consult the liveness information for the locals. If any
733 // of them are unused, then they can be replaced by top(). This
734 // should help register allocation time and cut down on the size
735 // of the deoptimization information.
737 // This call is made from many of the bytecode handling
738 // subroutines called from the Big Switch in do_one_bytecode.
739 // Every bytecode which might include a slow path is responsible
740 // for killing its dead locals. The more consistent we
741 // are about killing deads, the fewer useless phis will be
742 // constructed for them at various merge points.
744 // bci can be -1 (InvocationEntryBci). We return the entry
745 // liveness for the method.
747 if (method() == NULL || method()->code_size() == 0) {
748 // We are building a graph for a call to a native method.
749 // All locals are live.
750 return;
751 }
753 ResourceMark rm;
755 // Consult the liveness information for the locals. If any
756 // of them are unused, then they can be replaced by top(). This
757 // should help register allocation time and cut down on the size
758 // of the deoptimization information.
759 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
761 int len = (int)live_locals.size();
762 assert(len <= jvms()->loc_size(), "too many live locals");
763 for (int local = 0; local < len; local++) {
764 if (!live_locals.at(local)) {
765 set_local(local, top());
766 }
767 }
768 }
770 #ifdef ASSERT
771 //-------------------------dead_locals_are_killed------------------------------
772 // Return true if all dead locals are set to top in the map.
773 // Used to assert "clean" debug info at various points.
774 bool GraphKit::dead_locals_are_killed() {
775 if (method() == NULL || method()->code_size() == 0) {
776 // No locals need to be dead, so all is as it should be.
777 return true;
778 }
780 // Make sure somebody called kill_dead_locals upstream.
781 ResourceMark rm;
782 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
783 if (jvms->loc_size() == 0) continue; // no locals to consult
784 SafePointNode* map = jvms->map();
785 ciMethod* method = jvms->method();
786 int bci = jvms->bci();
787 if (jvms == this->jvms()) {
788 bci = this->bci(); // it might not yet be synched
789 }
790 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
791 int len = (int)live_locals.size();
792 if (!live_locals.is_valid() || len == 0)
793 // This method is trivial, or is poisoned by a breakpoint.
794 return true;
795 assert(len == jvms->loc_size(), "live map consistent with locals map");
796 for (int local = 0; local < len; local++) {
797 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
798 if (PrintMiscellaneous && (Verbose || WizardMode)) {
799 tty->print_cr("Zombie local %d: ", local);
800 jvms->dump();
801 }
802 return false;
803 }
804 }
805 }
806 return true;
807 }
809 #endif //ASSERT
811 // Helper function for enforcing certain bytecodes to reexecute if
812 // deoptimization happens
813 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
814 ciMethod* cur_method = jvms->method();
815 int cur_bci = jvms->bci();
816 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
817 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
818 return Interpreter::bytecode_should_reexecute(code) ||
819 is_anewarray && code == Bytecodes::_multianewarray;
820 // Reexecute _multianewarray bytecode which was replaced with
821 // sequence of [a]newarray. See Parse::do_multianewarray().
822 //
823 // Note: interpreter should not have it set since this optimization
824 // is limited by dimensions and guarded by flag so in some cases
825 // multianewarray() runtime calls will be generated and
826 // the bytecode should not be reexecutes (stack will not be reset).
827 } else
828 return false;
829 }
831 // Helper function for adding JVMState and debug information to node
832 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
833 // Add the safepoint edges to the call (or other safepoint).
835 // Make sure dead locals are set to top. This
836 // should help register allocation time and cut down on the size
837 // of the deoptimization information.
838 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
840 // Walk the inline list to fill in the correct set of JVMState's
841 // Also fill in the associated edges for each JVMState.
843 // If the bytecode needs to be reexecuted we need to put
844 // the arguments back on the stack.
845 const bool should_reexecute = jvms()->should_reexecute();
846 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
848 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
849 // undefined if the bci is different. This is normal for Parse but it
850 // should not happen for LibraryCallKit because only one bci is processed.
851 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
852 "in LibraryCallKit the reexecute bit should not change");
854 // If we are guaranteed to throw, we can prune everything but the
855 // input to the current bytecode.
856 bool can_prune_locals = false;
857 uint stack_slots_not_pruned = 0;
858 int inputs = 0, depth = 0;
859 if (must_throw) {
860 assert(method() == youngest_jvms->method(), "sanity");
861 if (compute_stack_effects(inputs, depth)) {
862 can_prune_locals = true;
863 stack_slots_not_pruned = inputs;
864 }
865 }
867 if (env()->jvmti_can_access_local_variables()) {
868 // At any safepoint, this method can get breakpointed, which would
869 // then require an immediate deoptimization.
870 can_prune_locals = false; // do not prune locals
871 stack_slots_not_pruned = 0;
872 }
874 // do not scribble on the input jvms
875 JVMState* out_jvms = youngest_jvms->clone_deep(C);
876 call->set_jvms(out_jvms); // Start jvms list for call node
878 // For a known set of bytecodes, the interpreter should reexecute them if
879 // deoptimization happens. We set the reexecute state for them here
880 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
881 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
882 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
883 }
885 // Presize the call:
886 DEBUG_ONLY(uint non_debug_edges = call->req());
887 call->add_req_batch(top(), youngest_jvms->debug_depth());
888 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
890 // Set up edges so that the call looks like this:
891 // Call [state:] ctl io mem fptr retadr
892 // [parms:] parm0 ... parmN
893 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
894 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
895 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
896 // Note that caller debug info precedes callee debug info.
898 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
899 uint debug_ptr = call->req();
901 // Loop over the map input edges associated with jvms, add them
902 // to the call node, & reset all offsets to match call node array.
903 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
904 uint debug_end = debug_ptr;
905 uint debug_start = debug_ptr - in_jvms->debug_size();
906 debug_ptr = debug_start; // back up the ptr
908 uint p = debug_start; // walks forward in [debug_start, debug_end)
909 uint j, k, l;
910 SafePointNode* in_map = in_jvms->map();
911 out_jvms->set_map(call);
913 if (can_prune_locals) {
914 assert(in_jvms->method() == out_jvms->method(), "sanity");
915 // If the current throw can reach an exception handler in this JVMS,
916 // then we must keep everything live that can reach that handler.
917 // As a quick and dirty approximation, we look for any handlers at all.
918 if (in_jvms->method()->has_exception_handlers()) {
919 can_prune_locals = false;
920 }
921 }
923 // Add the Locals
924 k = in_jvms->locoff();
925 l = in_jvms->loc_size();
926 out_jvms->set_locoff(p);
927 if (!can_prune_locals) {
928 for (j = 0; j < l; j++)
929 call->set_req(p++, in_map->in(k+j));
930 } else {
931 p += l; // already set to top above by add_req_batch
932 }
934 // Add the Expression Stack
935 k = in_jvms->stkoff();
936 l = in_jvms->sp();
937 out_jvms->set_stkoff(p);
938 if (!can_prune_locals) {
939 for (j = 0; j < l; j++)
940 call->set_req(p++, in_map->in(k+j));
941 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
942 // Divide stack into {S0,...,S1}, where S0 is set to top.
943 uint s1 = stack_slots_not_pruned;
944 stack_slots_not_pruned = 0; // for next iteration
945 if (s1 > l) s1 = l;
946 uint s0 = l - s1;
947 p += s0; // skip the tops preinstalled by add_req_batch
948 for (j = s0; j < l; j++)
949 call->set_req(p++, in_map->in(k+j));
950 } else {
951 p += l; // already set to top above by add_req_batch
952 }
954 // Add the Monitors
955 k = in_jvms->monoff();
956 l = in_jvms->mon_size();
957 out_jvms->set_monoff(p);
958 for (j = 0; j < l; j++)
959 call->set_req(p++, in_map->in(k+j));
961 // Copy any scalar object fields.
962 k = in_jvms->scloff();
963 l = in_jvms->scl_size();
964 out_jvms->set_scloff(p);
965 for (j = 0; j < l; j++)
966 call->set_req(p++, in_map->in(k+j));
968 // Finish the new jvms.
969 out_jvms->set_endoff(p);
971 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
972 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
973 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
974 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
975 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
976 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
978 // Update the two tail pointers in parallel.
979 out_jvms = out_jvms->caller();
980 in_jvms = in_jvms->caller();
981 }
983 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
985 // Test the correctness of JVMState::debug_xxx accessors:
986 assert(call->jvms()->debug_start() == non_debug_edges, "");
987 assert(call->jvms()->debug_end() == call->req(), "");
988 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
989 }
991 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
992 Bytecodes::Code code = java_bc();
993 if (code == Bytecodes::_wide) {
994 code = method()->java_code_at_bci(bci() + 1);
995 }
997 BasicType rtype = T_ILLEGAL;
998 int rsize = 0;
1000 if (code != Bytecodes::_illegal) {
1001 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1002 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1003 if (rtype < T_CONFLICT)
1004 rsize = type2size[rtype];
1005 }
1007 switch (code) {
1008 case Bytecodes::_illegal:
1009 return false;
1011 case Bytecodes::_ldc:
1012 case Bytecodes::_ldc_w:
1013 case Bytecodes::_ldc2_w:
1014 inputs = 0;
1015 break;
1017 case Bytecodes::_dup: inputs = 1; break;
1018 case Bytecodes::_dup_x1: inputs = 2; break;
1019 case Bytecodes::_dup_x2: inputs = 3; break;
1020 case Bytecodes::_dup2: inputs = 2; break;
1021 case Bytecodes::_dup2_x1: inputs = 3; break;
1022 case Bytecodes::_dup2_x2: inputs = 4; break;
1023 case Bytecodes::_swap: inputs = 2; break;
1024 case Bytecodes::_arraylength: inputs = 1; break;
1026 case Bytecodes::_getstatic:
1027 case Bytecodes::_putstatic:
1028 case Bytecodes::_getfield:
1029 case Bytecodes::_putfield:
1030 {
1031 bool ignored_will_link;
1032 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1033 int size = field->type()->size();
1034 bool is_get = (depth >= 0), is_static = (depth & 1);
1035 inputs = (is_static ? 0 : 1);
1036 if (is_get) {
1037 depth = size - inputs;
1038 } else {
1039 inputs += size; // putxxx pops the value from the stack
1040 depth = - inputs;
1041 }
1042 }
1043 break;
1045 case Bytecodes::_invokevirtual:
1046 case Bytecodes::_invokespecial:
1047 case Bytecodes::_invokestatic:
1048 case Bytecodes::_invokedynamic:
1049 case Bytecodes::_invokeinterface:
1050 {
1051 bool ignored_will_link;
1052 ciSignature* declared_signature = NULL;
1053 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1054 assert(declared_signature != NULL, "cannot be null");
1055 inputs = declared_signature->arg_size_for_bc(code);
1056 int size = declared_signature->return_type()->size();
1057 depth = size - inputs;
1058 }
1059 break;
1061 case Bytecodes::_multianewarray:
1062 {
1063 ciBytecodeStream iter(method());
1064 iter.reset_to_bci(bci());
1065 iter.next();
1066 inputs = iter.get_dimensions();
1067 assert(rsize == 1, "");
1068 depth = rsize - inputs;
1069 }
1070 break;
1072 case Bytecodes::_ireturn:
1073 case Bytecodes::_lreturn:
1074 case Bytecodes::_freturn:
1075 case Bytecodes::_dreturn:
1076 case Bytecodes::_areturn:
1077 assert(rsize = -depth, "");
1078 inputs = rsize;
1079 break;
1081 case Bytecodes::_jsr:
1082 case Bytecodes::_jsr_w:
1083 inputs = 0;
1084 depth = 1; // S.B. depth=1, not zero
1085 break;
1087 default:
1088 // bytecode produces a typed result
1089 inputs = rsize - depth;
1090 assert(inputs >= 0, "");
1091 break;
1092 }
1094 #ifdef ASSERT
1095 // spot check
1096 int outputs = depth + inputs;
1097 assert(outputs >= 0, "sanity");
1098 switch (code) {
1099 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1100 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1101 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1102 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1103 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1104 }
1105 #endif //ASSERT
1107 return true;
1108 }
1112 //------------------------------basic_plus_adr---------------------------------
1113 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1114 // short-circuit a common case
1115 if (offset == intcon(0)) return ptr;
1116 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1117 }
1119 Node* GraphKit::ConvI2L(Node* offset) {
1120 // short-circuit a common case
1121 jint offset_con = find_int_con(offset, Type::OffsetBot);
1122 if (offset_con != Type::OffsetBot) {
1123 return longcon((jlong) offset_con);
1124 }
1125 return _gvn.transform( new (C) ConvI2LNode(offset));
1126 }
1128 Node* GraphKit::ConvI2UL(Node* offset) {
1129 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1130 if (offset_con != (juint) Type::OffsetBot) {
1131 return longcon((julong) offset_con);
1132 }
1133 Node* conv = _gvn.transform( new (C) ConvI2LNode(offset));
1134 Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) );
1135 return _gvn.transform( new (C) AndLNode(conv, mask) );
1136 }
1138 Node* GraphKit::ConvL2I(Node* offset) {
1139 // short-circuit a common case
1140 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1141 if (offset_con != (jlong)Type::OffsetBot) {
1142 return intcon((int) offset_con);
1143 }
1144 return _gvn.transform( new (C) ConvL2INode(offset));
1145 }
1147 //-------------------------load_object_klass-----------------------------------
1148 Node* GraphKit::load_object_klass(Node* obj) {
1149 // Special-case a fresh allocation to avoid building nodes:
1150 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1151 if (akls != NULL) return akls;
1152 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1153 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1154 }
1156 //-------------------------load_array_length-----------------------------------
1157 Node* GraphKit::load_array_length(Node* array) {
1158 // Special-case a fresh allocation to avoid building nodes:
1159 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1160 Node *alen;
1161 if (alloc == NULL) {
1162 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1163 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1164 } else {
1165 alen = alloc->Ideal_length();
1166 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1167 if (ccast != alen) {
1168 alen = _gvn.transform(ccast);
1169 }
1170 }
1171 return alen;
1172 }
1174 //------------------------------do_null_check----------------------------------
1175 // Helper function to do a NULL pointer check. Returned value is
1176 // the incoming address with NULL casted away. You are allowed to use the
1177 // not-null value only if you are control dependent on the test.
1178 extern int explicit_null_checks_inserted,
1179 explicit_null_checks_elided;
1180 Node* GraphKit::null_check_common(Node* value, BasicType type,
1181 // optional arguments for variations:
1182 bool assert_null,
1183 Node* *null_control) {
1184 assert(!assert_null || null_control == NULL, "not both at once");
1185 if (stopped()) return top();
1186 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1187 // For some performance testing, we may wish to suppress null checking.
1188 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1189 return value;
1190 }
1191 explicit_null_checks_inserted++;
1193 // Construct NULL check
1194 Node *chk = NULL;
1195 switch(type) {
1196 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1197 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1198 case T_ARRAY : // fall through
1199 type = T_OBJECT; // simplify further tests
1200 case T_OBJECT : {
1201 const Type *t = _gvn.type( value );
1203 const TypeOopPtr* tp = t->isa_oopptr();
1204 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1205 // Only for do_null_check, not any of its siblings:
1206 && !assert_null && null_control == NULL) {
1207 // Usually, any field access or invocation on an unloaded oop type
1208 // will simply fail to link, since the statically linked class is
1209 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1210 // the static class is loaded but the sharper oop type is not.
1211 // Rather than checking for this obscure case in lots of places,
1212 // we simply observe that a null check on an unloaded class
1213 // will always be followed by a nonsense operation, so we
1214 // can just issue the uncommon trap here.
1215 // Our access to the unloaded class will only be correct
1216 // after it has been loaded and initialized, which requires
1217 // a trip through the interpreter.
1218 #ifndef PRODUCT
1219 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1220 #endif
1221 uncommon_trap(Deoptimization::Reason_unloaded,
1222 Deoptimization::Action_reinterpret,
1223 tp->klass(), "!loaded");
1224 return top();
1225 }
1227 if (assert_null) {
1228 // See if the type is contained in NULL_PTR.
1229 // If so, then the value is already null.
1230 if (t->higher_equal(TypePtr::NULL_PTR)) {
1231 explicit_null_checks_elided++;
1232 return value; // Elided null assert quickly!
1233 }
1234 } else {
1235 // See if mixing in the NULL pointer changes type.
1236 // If so, then the NULL pointer was not allowed in the original
1237 // type. In other words, "value" was not-null.
1238 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1239 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1240 explicit_null_checks_elided++;
1241 return value; // Elided null check quickly!
1242 }
1243 }
1244 chk = new (C) CmpPNode( value, null() );
1245 break;
1246 }
1248 default:
1249 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1250 }
1251 assert(chk != NULL, "sanity check");
1252 chk = _gvn.transform(chk);
1254 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1255 BoolNode *btst = new (C) BoolNode( chk, btest);
1256 Node *tst = _gvn.transform( btst );
1258 //-----------
1259 // if peephole optimizations occurred, a prior test existed.
1260 // If a prior test existed, maybe it dominates as we can avoid this test.
1261 if (tst != btst && type == T_OBJECT) {
1262 // At this point we want to scan up the CFG to see if we can
1263 // find an identical test (and so avoid this test altogether).
1264 Node *cfg = control();
1265 int depth = 0;
1266 while( depth < 16 ) { // Limit search depth for speed
1267 if( cfg->Opcode() == Op_IfTrue &&
1268 cfg->in(0)->in(1) == tst ) {
1269 // Found prior test. Use "cast_not_null" to construct an identical
1270 // CastPP (and hence hash to) as already exists for the prior test.
1271 // Return that casted value.
1272 if (assert_null) {
1273 replace_in_map(value, null());
1274 return null(); // do not issue the redundant test
1275 }
1276 Node *oldcontrol = control();
1277 set_control(cfg);
1278 Node *res = cast_not_null(value);
1279 set_control(oldcontrol);
1280 explicit_null_checks_elided++;
1281 return res;
1282 }
1283 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1284 if (cfg == NULL) break; // Quit at region nodes
1285 depth++;
1286 }
1287 }
1289 //-----------
1290 // Branch to failure if null
1291 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1292 Deoptimization::DeoptReason reason;
1293 if (assert_null)
1294 reason = Deoptimization::Reason_null_assert;
1295 else if (type == T_OBJECT)
1296 reason = Deoptimization::Reason_null_check;
1297 else
1298 reason = Deoptimization::Reason_div0_check;
1300 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1301 // ciMethodData::has_trap_at will return a conservative -1 if any
1302 // must-be-null assertion has failed. This could cause performance
1303 // problems for a method after its first do_null_assert failure.
1304 // Consider using 'Reason_class_check' instead?
1306 // To cause an implicit null check, we set the not-null probability
1307 // to the maximum (PROB_MAX). For an explicit check the probability
1308 // is set to a smaller value.
1309 if (null_control != NULL || too_many_traps(reason)) {
1310 // probability is less likely
1311 ok_prob = PROB_LIKELY_MAG(3);
1312 } else if (!assert_null &&
1313 (ImplicitNullCheckThreshold > 0) &&
1314 method() != NULL &&
1315 (method()->method_data()->trap_count(reason)
1316 >= (uint)ImplicitNullCheckThreshold)) {
1317 ok_prob = PROB_LIKELY_MAG(3);
1318 }
1320 if (null_control != NULL) {
1321 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1322 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1323 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1324 if (null_true == top())
1325 explicit_null_checks_elided++;
1326 (*null_control) = null_true;
1327 } else {
1328 BuildCutout unless(this, tst, ok_prob);
1329 // Check for optimizer eliding test at parse time
1330 if (stopped()) {
1331 // Failure not possible; do not bother making uncommon trap.
1332 explicit_null_checks_elided++;
1333 } else if (assert_null) {
1334 uncommon_trap(reason,
1335 Deoptimization::Action_make_not_entrant,
1336 NULL, "assert_null");
1337 } else {
1338 replace_in_map(value, zerocon(type));
1339 builtin_throw(reason);
1340 }
1341 }
1343 // Must throw exception, fall-thru not possible?
1344 if (stopped()) {
1345 return top(); // No result
1346 }
1348 if (assert_null) {
1349 // Cast obj to null on this path.
1350 replace_in_map(value, zerocon(type));
1351 return zerocon(type);
1352 }
1354 // Cast obj to not-null on this path, if there is no null_control.
1355 // (If there is a null_control, a non-null value may come back to haunt us.)
1356 if (type == T_OBJECT) {
1357 Node* cast = cast_not_null(value, false);
1358 if (null_control == NULL || (*null_control) == top())
1359 replace_in_map(value, cast);
1360 value = cast;
1361 }
1363 return value;
1364 }
1367 //------------------------------cast_not_null----------------------------------
1368 // Cast obj to not-null on this path
1369 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1370 const Type *t = _gvn.type(obj);
1371 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1372 // Object is already not-null?
1373 if( t == t_not_null ) return obj;
1375 Node *cast = new (C) CastPPNode(obj,t_not_null);
1376 cast->init_req(0, control());
1377 cast = _gvn.transform( cast );
1379 // Scan for instances of 'obj' in the current JVM mapping.
1380 // These instances are known to be not-null after the test.
1381 if (do_replace_in_map)
1382 replace_in_map(obj, cast);
1384 return cast; // Return casted value
1385 }
1388 //--------------------------replace_in_map-------------------------------------
1389 void GraphKit::replace_in_map(Node* old, Node* neww) {
1390 if (old == neww) {
1391 return;
1392 }
1394 map()->replace_edge(old, neww);
1396 // Note: This operation potentially replaces any edge
1397 // on the map. This includes locals, stack, and monitors
1398 // of the current (innermost) JVM state.
1400 // don't let inconsistent types from profiling escape this
1401 // method
1403 const Type* told = _gvn.type(old);
1404 const Type* tnew = _gvn.type(neww);
1406 if (!tnew->higher_equal(told)) {
1407 return;
1408 }
1410 map()->record_replaced_node(old, neww);
1411 }
1414 //=============================================================================
1415 //--------------------------------memory---------------------------------------
1416 Node* GraphKit::memory(uint alias_idx) {
1417 MergeMemNode* mem = merged_memory();
1418 Node* p = mem->memory_at(alias_idx);
1419 _gvn.set_type(p, Type::MEMORY); // must be mapped
1420 return p;
1421 }
1423 //-----------------------------reset_memory------------------------------------
1424 Node* GraphKit::reset_memory() {
1425 Node* mem = map()->memory();
1426 // do not use this node for any more parsing!
1427 debug_only( map()->set_memory((Node*)NULL) );
1428 return _gvn.transform( mem );
1429 }
1431 //------------------------------set_all_memory---------------------------------
1432 void GraphKit::set_all_memory(Node* newmem) {
1433 Node* mergemem = MergeMemNode::make(C, newmem);
1434 gvn().set_type_bottom(mergemem);
1435 map()->set_memory(mergemem);
1436 }
1438 //------------------------------set_all_memory_call----------------------------
1439 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1440 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1441 set_all_memory(newmem);
1442 }
1444 //=============================================================================
1445 //
1446 // parser factory methods for MemNodes
1447 //
1448 // These are layered on top of the factory methods in LoadNode and StoreNode,
1449 // and integrate with the parser's memory state and _gvn engine.
1450 //
1452 // factory methods in "int adr_idx"
1453 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1454 int adr_idx,
1455 MemNode::MemOrd mo, LoadNode::ControlDependency control_dependency, bool require_atomic_access) {
1456 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1457 const TypePtr* adr_type = NULL; // debug-mode-only argument
1458 debug_only(adr_type = C->get_adr_type(adr_idx));
1459 Node* mem = memory(adr_idx);
1460 Node* ld;
1461 if (require_atomic_access && bt == T_LONG) {
1462 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency);
1463 } else if (require_atomic_access && bt == T_DOUBLE) {
1464 ld = LoadDNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency);
1465 } else {
1466 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency);
1467 }
1468 ld = _gvn.transform(ld);
1469 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1470 // Improve graph before escape analysis and boxing elimination.
1471 record_for_igvn(ld);
1472 }
1473 return ld;
1474 }
1476 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1477 int adr_idx,
1478 MemNode::MemOrd mo,
1479 bool require_atomic_access) {
1480 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1481 const TypePtr* adr_type = NULL;
1482 debug_only(adr_type = C->get_adr_type(adr_idx));
1483 Node *mem = memory(adr_idx);
1484 Node* st;
1485 if (require_atomic_access && bt == T_LONG) {
1486 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1487 } else if (require_atomic_access && bt == T_DOUBLE) {
1488 st = StoreDNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1489 } else {
1490 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1491 }
1492 st = _gvn.transform(st);
1493 set_memory(st, adr_idx);
1494 // Back-to-back stores can only remove intermediate store with DU info
1495 // so push on worklist for optimizer.
1496 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1497 record_for_igvn(st);
1499 return st;
1500 }
1503 void GraphKit::pre_barrier(bool do_load,
1504 Node* ctl,
1505 Node* obj,
1506 Node* adr,
1507 uint adr_idx,
1508 Node* val,
1509 const TypeOopPtr* val_type,
1510 Node* pre_val,
1511 BasicType bt) {
1513 BarrierSet* bs = Universe::heap()->barrier_set();
1514 set_control(ctl);
1515 switch (bs->kind()) {
1516 case BarrierSet::G1SATBCT:
1517 case BarrierSet::G1SATBCTLogging:
1518 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1519 break;
1521 case BarrierSet::CardTableModRef:
1522 case BarrierSet::CardTableExtension:
1523 case BarrierSet::ModRef:
1524 break;
1526 case BarrierSet::Other:
1527 default :
1528 ShouldNotReachHere();
1530 }
1531 }
1533 bool GraphKit::can_move_pre_barrier() const {
1534 BarrierSet* bs = Universe::heap()->barrier_set();
1535 switch (bs->kind()) {
1536 case BarrierSet::G1SATBCT:
1537 case BarrierSet::G1SATBCTLogging:
1538 return true; // Can move it if no safepoint
1540 case BarrierSet::CardTableModRef:
1541 case BarrierSet::CardTableExtension:
1542 case BarrierSet::ModRef:
1543 return true; // There is no pre-barrier
1545 case BarrierSet::Other:
1546 default :
1547 ShouldNotReachHere();
1548 }
1549 return false;
1550 }
1552 void GraphKit::post_barrier(Node* ctl,
1553 Node* store,
1554 Node* obj,
1555 Node* adr,
1556 uint adr_idx,
1557 Node* val,
1558 BasicType bt,
1559 bool use_precise) {
1560 BarrierSet* bs = Universe::heap()->barrier_set();
1561 set_control(ctl);
1562 switch (bs->kind()) {
1563 case BarrierSet::G1SATBCT:
1564 case BarrierSet::G1SATBCTLogging:
1565 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1566 break;
1568 case BarrierSet::CardTableModRef:
1569 case BarrierSet::CardTableExtension:
1570 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1571 break;
1573 case BarrierSet::ModRef:
1574 break;
1576 case BarrierSet::Other:
1577 default :
1578 ShouldNotReachHere();
1580 }
1581 }
1583 Node* GraphKit::store_oop(Node* ctl,
1584 Node* obj,
1585 Node* adr,
1586 const TypePtr* adr_type,
1587 Node* val,
1588 const TypeOopPtr* val_type,
1589 BasicType bt,
1590 bool use_precise,
1591 MemNode::MemOrd mo) {
1592 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1593 // could be delayed during Parse (for example, in adjust_map_after_if()).
1594 // Execute transformation here to avoid barrier generation in such case.
1595 if (_gvn.type(val) == TypePtr::NULL_PTR)
1596 val = _gvn.makecon(TypePtr::NULL_PTR);
1598 set_control(ctl);
1599 if (stopped()) return top(); // Dead path ?
1601 assert(bt == T_OBJECT, "sanity");
1602 assert(val != NULL, "not dead path");
1603 uint adr_idx = C->get_alias_index(adr_type);
1604 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1606 pre_barrier(true /* do_load */,
1607 control(), obj, adr, adr_idx, val, val_type,
1608 NULL /* pre_val */,
1609 bt);
1611 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo);
1612 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1613 return store;
1614 }
1616 // Could be an array or object we don't know at compile time (unsafe ref.)
1617 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1618 Node* obj, // containing obj
1619 Node* adr, // actual adress to store val at
1620 const TypePtr* adr_type,
1621 Node* val,
1622 BasicType bt,
1623 MemNode::MemOrd mo) {
1624 Compile::AliasType* at = C->alias_type(adr_type);
1625 const TypeOopPtr* val_type = NULL;
1626 if (adr_type->isa_instptr()) {
1627 if (at->field() != NULL) {
1628 // known field. This code is a copy of the do_put_xxx logic.
1629 ciField* field = at->field();
1630 if (!field->type()->is_loaded()) {
1631 val_type = TypeInstPtr::BOTTOM;
1632 } else {
1633 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1634 }
1635 }
1636 } else if (adr_type->isa_aryptr()) {
1637 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1638 }
1639 if (val_type == NULL) {
1640 val_type = TypeInstPtr::BOTTOM;
1641 }
1642 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo);
1643 }
1646 //-------------------------array_element_address-------------------------
1647 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1648 const TypeInt* sizetype, Node* ctrl) {
1649 uint shift = exact_log2(type2aelembytes(elembt));
1650 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1652 // short-circuit a common case (saves lots of confusing waste motion)
1653 jint idx_con = find_int_con(idx, -1);
1654 if (idx_con >= 0) {
1655 intptr_t offset = header + ((intptr_t)idx_con << shift);
1656 return basic_plus_adr(ary, offset);
1657 }
1659 // must be correct type for alignment purposes
1660 Node* base = basic_plus_adr(ary, header);
1661 #ifdef _LP64
1662 // The scaled index operand to AddP must be a clean 64-bit value.
1663 // Java allows a 32-bit int to be incremented to a negative
1664 // value, which appears in a 64-bit register as a large
1665 // positive number. Using that large positive number as an
1666 // operand in pointer arithmetic has bad consequences.
1667 // On the other hand, 32-bit overflow is rare, and the possibility
1668 // can often be excluded, if we annotate the ConvI2L node with
1669 // a type assertion that its value is known to be a small positive
1670 // number. (The prior range check has ensured this.)
1671 // This assertion is used by ConvI2LNode::Ideal.
1672 int index_max = max_jint - 1; // array size is max_jint, index is one less
1673 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1674 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
1675 idx = C->constrained_convI2L(&_gvn, idx, iidxtype, ctrl);
1676 #endif
1677 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1678 return basic_plus_adr(ary, base, scale);
1679 }
1681 //-------------------------load_array_element-------------------------
1682 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1683 const Type* elemtype = arytype->elem();
1684 BasicType elembt = elemtype->array_element_basic_type();
1685 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1686 if (elembt == T_NARROWOOP) {
1687 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1688 }
1689 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1690 return ld;
1691 }
1693 //-------------------------set_arguments_for_java_call-------------------------
1694 // Arguments (pre-popped from the stack) are taken from the JVMS.
1695 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1696 // Add the call arguments:
1697 uint nargs = call->method()->arg_size();
1698 for (uint i = 0; i < nargs; i++) {
1699 Node* arg = argument(i);
1700 call->init_req(i + TypeFunc::Parms, arg);
1701 }
1702 }
1704 //---------------------------set_edges_for_java_call---------------------------
1705 // Connect a newly created call into the current JVMS.
1706 // A return value node (if any) is returned from set_edges_for_java_call.
1707 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1709 // Add the predefined inputs:
1710 call->init_req( TypeFunc::Control, control() );
1711 call->init_req( TypeFunc::I_O , i_o() );
1712 call->init_req( TypeFunc::Memory , reset_memory() );
1713 call->init_req( TypeFunc::FramePtr, frameptr() );
1714 call->init_req( TypeFunc::ReturnAdr, top() );
1716 add_safepoint_edges(call, must_throw);
1718 Node* xcall = _gvn.transform(call);
1720 if (xcall == top()) {
1721 set_control(top());
1722 return;
1723 }
1724 assert(xcall == call, "call identity is stable");
1726 // Re-use the current map to produce the result.
1728 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1729 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1730 set_all_memory_call(xcall, separate_io_proj);
1732 //return xcall; // no need, caller already has it
1733 }
1735 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1736 if (stopped()) return top(); // maybe the call folded up?
1738 // Capture the return value, if any.
1739 Node* ret;
1740 if (call->method() == NULL ||
1741 call->method()->return_type()->basic_type() == T_VOID)
1742 ret = top();
1743 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1745 // Note: Since any out-of-line call can produce an exception,
1746 // we always insert an I_O projection from the call into the result.
1748 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1750 if (separate_io_proj) {
1751 // The caller requested separate projections be used by the fall
1752 // through and exceptional paths, so replace the projections for
1753 // the fall through path.
1754 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1755 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1756 }
1757 return ret;
1758 }
1760 //--------------------set_predefined_input_for_runtime_call--------------------
1761 // Reading and setting the memory state is way conservative here.
1762 // The real problem is that I am not doing real Type analysis on memory,
1763 // so I cannot distinguish card mark stores from other stores. Across a GC
1764 // point the Store Barrier and the card mark memory has to agree. I cannot
1765 // have a card mark store and its barrier split across the GC point from
1766 // either above or below. Here I get that to happen by reading ALL of memory.
1767 // A better answer would be to separate out card marks from other memory.
1768 // For now, return the input memory state, so that it can be reused
1769 // after the call, if this call has restricted memory effects.
1770 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1771 // Set fixed predefined input arguments
1772 Node* memory = reset_memory();
1773 call->init_req( TypeFunc::Control, control() );
1774 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1775 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1776 call->init_req( TypeFunc::FramePtr, frameptr() );
1777 call->init_req( TypeFunc::ReturnAdr, top() );
1778 return memory;
1779 }
1781 //-------------------set_predefined_output_for_runtime_call--------------------
1782 // Set control and memory (not i_o) from the call.
1783 // If keep_mem is not NULL, use it for the output state,
1784 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1785 // If hook_mem is NULL, this call produces no memory effects at all.
1786 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1787 // then only that memory slice is taken from the call.
1788 // In the last case, we must put an appropriate memory barrier before
1789 // the call, so as to create the correct anti-dependencies on loads
1790 // preceding the call.
1791 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1792 Node* keep_mem,
1793 const TypePtr* hook_mem) {
1794 // no i/o
1795 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1796 if (keep_mem) {
1797 // First clone the existing memory state
1798 set_all_memory(keep_mem);
1799 if (hook_mem != NULL) {
1800 // Make memory for the call
1801 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1802 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1803 // We also use hook_mem to extract specific effects from arraycopy stubs.
1804 set_memory(mem, hook_mem);
1805 }
1806 // ...else the call has NO memory effects.
1808 // Make sure the call advertises its memory effects precisely.
1809 // This lets us build accurate anti-dependences in gcm.cpp.
1810 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1811 "call node must be constructed correctly");
1812 } else {
1813 assert(hook_mem == NULL, "");
1814 // This is not a "slow path" call; all memory comes from the call.
1815 set_all_memory_call(call);
1816 }
1817 }
1820 // Replace the call with the current state of the kit.
1821 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1822 JVMState* ejvms = NULL;
1823 if (has_exceptions()) {
1824 ejvms = transfer_exceptions_into_jvms();
1825 }
1827 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1828 ReplacedNodes replaced_nodes_exception;
1829 Node* ex_ctl = top();
1831 SafePointNode* final_state = stop();
1833 // Find all the needed outputs of this call
1834 CallProjections callprojs;
1835 call->extract_projections(&callprojs, true);
1837 Node* init_mem = call->in(TypeFunc::Memory);
1838 Node* final_mem = final_state->in(TypeFunc::Memory);
1839 Node* final_ctl = final_state->in(TypeFunc::Control);
1840 Node* final_io = final_state->in(TypeFunc::I_O);
1842 // Replace all the old call edges with the edges from the inlining result
1843 if (callprojs.fallthrough_catchproj != NULL) {
1844 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1845 }
1846 if (callprojs.fallthrough_memproj != NULL) {
1847 if (final_mem->is_MergeMem()) {
1848 // Parser's exits MergeMem was not transformed but may be optimized
1849 final_mem = _gvn.transform(final_mem);
1850 }
1851 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1852 }
1853 if (callprojs.fallthrough_ioproj != NULL) {
1854 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1855 }
1857 // Replace the result with the new result if it exists and is used
1858 if (callprojs.resproj != NULL && result != NULL) {
1859 C->gvn_replace_by(callprojs.resproj, result);
1860 }
1862 if (ejvms == NULL) {
1863 // No exception edges to simply kill off those paths
1864 if (callprojs.catchall_catchproj != NULL) {
1865 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1866 }
1867 if (callprojs.catchall_memproj != NULL) {
1868 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1869 }
1870 if (callprojs.catchall_ioproj != NULL) {
1871 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1872 }
1873 // Replace the old exception object with top
1874 if (callprojs.exobj != NULL) {
1875 C->gvn_replace_by(callprojs.exobj, C->top());
1876 }
1877 } else {
1878 GraphKit ekit(ejvms);
1880 // Load my combined exception state into the kit, with all phis transformed:
1881 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1882 replaced_nodes_exception = ex_map->replaced_nodes();
1884 Node* ex_oop = ekit.use_exception_state(ex_map);
1886 if (callprojs.catchall_catchproj != NULL) {
1887 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1888 ex_ctl = ekit.control();
1889 }
1890 if (callprojs.catchall_memproj != NULL) {
1891 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1892 }
1893 if (callprojs.catchall_ioproj != NULL) {
1894 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1895 }
1897 // Replace the old exception object with the newly created one
1898 if (callprojs.exobj != NULL) {
1899 C->gvn_replace_by(callprojs.exobj, ex_oop);
1900 }
1901 }
1903 // Disconnect the call from the graph
1904 call->disconnect_inputs(NULL, C);
1905 C->gvn_replace_by(call, C->top());
1907 // Clean up any MergeMems that feed other MergeMems since the
1908 // optimizer doesn't like that.
1909 if (final_mem->is_MergeMem()) {
1910 Node_List wl;
1911 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1912 Node* m = i.get();
1913 if (m->is_MergeMem() && !wl.contains(m)) {
1914 wl.push(m);
1915 }
1916 }
1917 while (wl.size() > 0) {
1918 _gvn.transform(wl.pop());
1919 }
1920 }
1922 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1923 replaced_nodes.apply(C, final_ctl);
1924 }
1925 if (!ex_ctl->is_top() && do_replaced_nodes) {
1926 replaced_nodes_exception.apply(C, ex_ctl);
1927 }
1928 }
1931 //------------------------------increment_counter------------------------------
1932 // for statistics: increment a VM counter by 1
1934 void GraphKit::increment_counter(address counter_addr) {
1935 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1936 increment_counter(adr1);
1937 }
1939 void GraphKit::increment_counter(Node* counter_addr) {
1940 int adr_type = Compile::AliasIdxRaw;
1941 Node* ctrl = control();
1942 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1943 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1944 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1945 }
1948 //------------------------------uncommon_trap----------------------------------
1949 // Bail out to the interpreter in mid-method. Implemented by calling the
1950 // uncommon_trap blob. This helper function inserts a runtime call with the
1951 // right debug info.
1952 void GraphKit::uncommon_trap(int trap_request,
1953 ciKlass* klass, const char* comment,
1954 bool must_throw,
1955 bool keep_exact_action) {
1956 if (failing()) stop();
1957 if (stopped()) return; // trap reachable?
1959 // Note: If ProfileTraps is true, and if a deopt. actually
1960 // occurs here, the runtime will make sure an MDO exists. There is
1961 // no need to call method()->ensure_method_data() at this point.
1963 // Set the stack pointer to the right value for reexecution:
1964 set_sp(reexecute_sp());
1966 #ifdef ASSERT
1967 if (!must_throw) {
1968 // Make sure the stack has at least enough depth to execute
1969 // the current bytecode.
1970 int inputs, ignored_depth;
1971 if (compute_stack_effects(inputs, ignored_depth)) {
1972 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1973 Bytecodes::name(java_bc()), sp(), inputs));
1974 }
1975 }
1976 #endif
1978 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1979 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1981 switch (action) {
1982 case Deoptimization::Action_maybe_recompile:
1983 case Deoptimization::Action_reinterpret:
1984 // Temporary fix for 6529811 to allow virtual calls to be sure they
1985 // get the chance to go from mono->bi->mega
1986 if (!keep_exact_action &&
1987 Deoptimization::trap_request_index(trap_request) < 0 &&
1988 too_many_recompiles(reason)) {
1989 // This BCI is causing too many recompilations.
1990 if (C->log() != NULL) {
1991 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
1992 Deoptimization::trap_reason_name(reason),
1993 Deoptimization::trap_action_name(action));
1994 }
1995 action = Deoptimization::Action_none;
1996 trap_request = Deoptimization::make_trap_request(reason, action);
1997 } else {
1998 C->set_trap_can_recompile(true);
1999 }
2000 break;
2001 case Deoptimization::Action_make_not_entrant:
2002 C->set_trap_can_recompile(true);
2003 break;
2004 #ifdef ASSERT
2005 case Deoptimization::Action_none:
2006 case Deoptimization::Action_make_not_compilable:
2007 break;
2008 default:
2009 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
2010 break;
2011 #endif
2012 }
2014 if (TraceOptoParse) {
2015 char buf[100];
2016 tty->print_cr("Uncommon trap %s at bci:%d",
2017 Deoptimization::format_trap_request(buf, sizeof(buf),
2018 trap_request), bci());
2019 }
2021 CompileLog* log = C->log();
2022 if (log != NULL) {
2023 int kid = (klass == NULL)? -1: log->identify(klass);
2024 log->begin_elem("uncommon_trap bci='%d'", bci());
2025 char buf[100];
2026 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2027 trap_request));
2028 if (kid >= 0) log->print(" klass='%d'", kid);
2029 if (comment != NULL) log->print(" comment='%s'", comment);
2030 log->end_elem();
2031 }
2033 // Make sure any guarding test views this path as very unlikely
2034 Node *i0 = control()->in(0);
2035 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2036 IfNode *iff = i0->as_If();
2037 float f = iff->_prob; // Get prob
2038 if (control()->Opcode() == Op_IfTrue) {
2039 if (f > PROB_UNLIKELY_MAG(4))
2040 iff->_prob = PROB_MIN;
2041 } else {
2042 if (f < PROB_LIKELY_MAG(4))
2043 iff->_prob = PROB_MAX;
2044 }
2045 }
2047 // Clear out dead values from the debug info.
2048 kill_dead_locals();
2050 // Now insert the uncommon trap subroutine call
2051 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2052 const TypePtr* no_memory_effects = NULL;
2053 // Pass the index of the class to be loaded
2054 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2055 (must_throw ? RC_MUST_THROW : 0),
2056 OptoRuntime::uncommon_trap_Type(),
2057 call_addr, "uncommon_trap", no_memory_effects,
2058 intcon(trap_request));
2059 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2060 "must extract request correctly from the graph");
2061 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2063 call->set_req(TypeFunc::ReturnAdr, returnadr());
2064 // The debug info is the only real input to this call.
2066 // Halt-and-catch fire here. The above call should never return!
2067 HaltNode* halt = new(C) HaltNode(control(), frameptr());
2068 _gvn.set_type_bottom(halt);
2069 root()->add_req(halt);
2071 stop_and_kill_map();
2072 }
2075 //--------------------------just_allocated_object------------------------------
2076 // Report the object that was just allocated.
2077 // It must be the case that there are no intervening safepoints.
2078 // We use this to determine if an object is so "fresh" that
2079 // it does not require card marks.
2080 Node* GraphKit::just_allocated_object(Node* current_control) {
2081 if (C->recent_alloc_ctl() == current_control)
2082 return C->recent_alloc_obj();
2083 return NULL;
2084 }
2087 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2088 // (Note: TypeFunc::make has a cache that makes this fast.)
2089 const TypeFunc* tf = TypeFunc::make(dest_method);
2090 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2091 for (int j = 0; j < nargs; j++) {
2092 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2093 if( targ->basic_type() == T_DOUBLE ) {
2094 // If any parameters are doubles, they must be rounded before
2095 // the call, dstore_rounding does gvn.transform
2096 Node *arg = argument(j);
2097 arg = dstore_rounding(arg);
2098 set_argument(j, arg);
2099 }
2100 }
2101 }
2103 /**
2104 * Record profiling data exact_kls for Node n with the type system so
2105 * that it can propagate it (speculation)
2106 *
2107 * @param n node that the type applies to
2108 * @param exact_kls type from profiling
2109 *
2110 * @return node with improved type
2111 */
2112 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) {
2113 const Type* current_type = _gvn.type(n);
2114 assert(UseTypeSpeculation, "type speculation must be on");
2116 const TypeOopPtr* speculative = current_type->speculative();
2118 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2119 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2120 const TypeOopPtr* xtype = tklass->as_instance_type();
2121 assert(xtype->klass_is_exact(), "Should be exact");
2122 // record the new speculative type's depth
2123 speculative = xtype->with_inline_depth(jvms()->depth());
2124 }
2126 if (speculative != current_type->speculative()) {
2127 // Build a type with a speculative type (what we think we know
2128 // about the type but will need a guard when we use it)
2129 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2130 // We're changing the type, we need a new CheckCast node to carry
2131 // the new type. The new type depends on the control: what
2132 // profiling tells us is only valid from here as far as we can
2133 // tell.
2134 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2135 cast = _gvn.transform(cast);
2136 replace_in_map(n, cast);
2137 n = cast;
2138 }
2140 return n;
2141 }
2143 /**
2144 * Record profiling data from receiver profiling at an invoke with the
2145 * type system so that it can propagate it (speculation)
2146 *
2147 * @param n receiver node
2148 *
2149 * @return node with improved type
2150 */
2151 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2152 if (!UseTypeSpeculation) {
2153 return n;
2154 }
2155 ciKlass* exact_kls = profile_has_unique_klass();
2156 return record_profile_for_speculation(n, exact_kls);
2157 }
2159 /**
2160 * Record profiling data from argument profiling at an invoke with the
2161 * type system so that it can propagate it (speculation)
2162 *
2163 * @param dest_method target method for the call
2164 * @param bc what invoke bytecode is this?
2165 */
2166 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2167 if (!UseTypeSpeculation) {
2168 return;
2169 }
2170 const TypeFunc* tf = TypeFunc::make(dest_method);
2171 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2172 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2173 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2174 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2175 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2176 ciKlass* better_type = method()->argument_profiled_type(bci(), i);
2177 if (better_type != NULL) {
2178 record_profile_for_speculation(argument(j), better_type);
2179 }
2180 i++;
2181 }
2182 }
2183 }
2185 /**
2186 * Record profiling data from parameter profiling at an invoke with
2187 * the type system so that it can propagate it (speculation)
2188 */
2189 void GraphKit::record_profiled_parameters_for_speculation() {
2190 if (!UseTypeSpeculation) {
2191 return;
2192 }
2193 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2194 if (_gvn.type(local(i))->isa_oopptr()) {
2195 ciKlass* better_type = method()->parameter_profiled_type(j);
2196 if (better_type != NULL) {
2197 record_profile_for_speculation(local(i), better_type);
2198 }
2199 j++;
2200 }
2201 }
2202 }
2204 void GraphKit::round_double_result(ciMethod* dest_method) {
2205 // A non-strict method may return a double value which has an extended
2206 // exponent, but this must not be visible in a caller which is 'strict'
2207 // If a strict caller invokes a non-strict callee, round a double result
2209 BasicType result_type = dest_method->return_type()->basic_type();
2210 assert( method() != NULL, "must have caller context");
2211 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2212 // Destination method's return value is on top of stack
2213 // dstore_rounding() does gvn.transform
2214 Node *result = pop_pair();
2215 result = dstore_rounding(result);
2216 push_pair(result);
2217 }
2218 }
2220 // rounding for strict float precision conformance
2221 Node* GraphKit::precision_rounding(Node* n) {
2222 return UseStrictFP && _method->flags().is_strict()
2223 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2224 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2225 : n;
2226 }
2228 // rounding for strict double precision conformance
2229 Node* GraphKit::dprecision_rounding(Node *n) {
2230 return UseStrictFP && _method->flags().is_strict()
2231 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2232 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2233 : n;
2234 }
2236 // rounding for non-strict double stores
2237 Node* GraphKit::dstore_rounding(Node* n) {
2238 return Matcher::strict_fp_requires_explicit_rounding
2239 && UseSSE <= 1
2240 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2241 : n;
2242 }
2244 //=============================================================================
2245 // Generate a fast path/slow path idiom. Graph looks like:
2246 // [foo] indicates that 'foo' is a parameter
2247 //
2248 // [in] NULL
2249 // \ /
2250 // CmpP
2251 // Bool ne
2252 // If
2253 // / \
2254 // True False-<2>
2255 // / |
2256 // / cast_not_null
2257 // Load | | ^
2258 // [fast_test] | |
2259 // gvn to opt_test | |
2260 // / \ | <1>
2261 // True False |
2262 // | \\ |
2263 // [slow_call] \[fast_result]
2264 // Ctl Val \ \
2265 // | \ \
2266 // Catch <1> \ \
2267 // / \ ^ \ \
2268 // Ex No_Ex | \ \
2269 // | \ \ | \ <2> \
2270 // ... \ [slow_res] | | \ [null_result]
2271 // \ \--+--+--- | |
2272 // \ | / \ | /
2273 // --------Region Phi
2274 //
2275 //=============================================================================
2276 // Code is structured as a series of driver functions all called 'do_XXX' that
2277 // call a set of helper functions. Helper functions first, then drivers.
2279 //------------------------------null_check_oop---------------------------------
2280 // Null check oop. Set null-path control into Region in slot 3.
2281 // Make a cast-not-nullness use the other not-null control. Return cast.
2282 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2283 bool never_see_null, bool safe_for_replace) {
2284 // Initial NULL check taken path
2285 (*null_control) = top();
2286 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2288 // Generate uncommon_trap:
2289 if (never_see_null && (*null_control) != top()) {
2290 // If we see an unexpected null at a check-cast we record it and force a
2291 // recompile; the offending check-cast will be compiled to handle NULLs.
2292 // If we see more than one offending BCI, then all checkcasts in the
2293 // method will be compiled to handle NULLs.
2294 PreserveJVMState pjvms(this);
2295 set_control(*null_control);
2296 replace_in_map(value, null());
2297 uncommon_trap(Deoptimization::Reason_null_check,
2298 Deoptimization::Action_make_not_entrant);
2299 (*null_control) = top(); // NULL path is dead
2300 }
2301 if ((*null_control) == top() && safe_for_replace) {
2302 replace_in_map(value, cast);
2303 }
2305 // Cast away null-ness on the result
2306 return cast;
2307 }
2309 //------------------------------opt_iff----------------------------------------
2310 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2311 // Return slow-path control.
2312 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2313 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2315 // Fast path taken; set region slot 2
2316 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2317 region->init_req(2,fast_taken); // Capture fast-control
2319 // Fast path not-taken, i.e. slow path
2320 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2321 return slow_taken;
2322 }
2324 //-----------------------------make_runtime_call-------------------------------
2325 Node* GraphKit::make_runtime_call(int flags,
2326 const TypeFunc* call_type, address call_addr,
2327 const char* call_name,
2328 const TypePtr* adr_type,
2329 // The following parms are all optional.
2330 // The first NULL ends the list.
2331 Node* parm0, Node* parm1,
2332 Node* parm2, Node* parm3,
2333 Node* parm4, Node* parm5,
2334 Node* parm6, Node* parm7) {
2335 // Slow-path call
2336 bool is_leaf = !(flags & RC_NO_LEAF);
2337 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2338 if (call_name == NULL) {
2339 assert(!is_leaf, "must supply name for leaf");
2340 call_name = OptoRuntime::stub_name(call_addr);
2341 }
2342 CallNode* call;
2343 if (!is_leaf) {
2344 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2345 bci(), adr_type);
2346 } else if (flags & RC_NO_FP) {
2347 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2348 } else {
2349 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2350 }
2352 // The following is similar to set_edges_for_java_call,
2353 // except that the memory effects of the call are restricted to AliasIdxRaw.
2355 // Slow path call has no side-effects, uses few values
2356 bool wide_in = !(flags & RC_NARROW_MEM);
2357 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2359 Node* prev_mem = NULL;
2360 if (wide_in) {
2361 prev_mem = set_predefined_input_for_runtime_call(call);
2362 } else {
2363 assert(!wide_out, "narrow in => narrow out");
2364 Node* narrow_mem = memory(adr_type);
2365 prev_mem = reset_memory();
2366 map()->set_memory(narrow_mem);
2367 set_predefined_input_for_runtime_call(call);
2368 }
2370 // Hook each parm in order. Stop looking at the first NULL.
2371 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2372 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2373 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2374 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2375 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2376 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2377 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2378 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2379 /* close each nested if ===> */ } } } } } } } }
2380 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2382 if (!is_leaf) {
2383 // Non-leaves can block and take safepoints:
2384 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2385 }
2386 // Non-leaves can throw exceptions:
2387 if (has_io) {
2388 call->set_req(TypeFunc::I_O, i_o());
2389 }
2391 if (flags & RC_UNCOMMON) {
2392 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2393 // (An "if" probability corresponds roughly to an unconditional count.
2394 // Sort of.)
2395 call->set_cnt(PROB_UNLIKELY_MAG(4));
2396 }
2398 Node* c = _gvn.transform(call);
2399 assert(c == call, "cannot disappear");
2401 if (wide_out) {
2402 // Slow path call has full side-effects.
2403 set_predefined_output_for_runtime_call(call);
2404 } else {
2405 // Slow path call has few side-effects, and/or sets few values.
2406 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2407 }
2409 if (has_io) {
2410 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2411 }
2412 return call;
2414 }
2416 //------------------------------merge_memory-----------------------------------
2417 // Merge memory from one path into the current memory state.
2418 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2419 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2420 Node* old_slice = mms.force_memory();
2421 Node* new_slice = mms.memory2();
2422 if (old_slice != new_slice) {
2423 PhiNode* phi;
2424 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2425 if (mms.is_empty()) {
2426 // clone base memory Phi's inputs for this memory slice
2427 assert(old_slice == mms.base_memory(), "sanity");
2428 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2429 _gvn.set_type(phi, Type::MEMORY);
2430 for (uint i = 1; i < phi->req(); i++) {
2431 phi->init_req(i, old_slice->in(i));
2432 }
2433 } else {
2434 phi = old_slice->as_Phi(); // Phi was generated already
2435 }
2436 } else {
2437 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2438 _gvn.set_type(phi, Type::MEMORY);
2439 }
2440 phi->set_req(new_path, new_slice);
2441 mms.set_memory(phi);
2442 }
2443 }
2444 }
2446 //------------------------------make_slow_call_ex------------------------------
2447 // Make the exception handler hookups for the slow call
2448 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2449 if (stopped()) return;
2451 // Make a catch node with just two handlers: fall-through and catch-all
2452 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2453 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2454 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2455 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2457 { PreserveJVMState pjvms(this);
2458 set_control(excp);
2459 set_i_o(i_o);
2461 if (excp != top()) {
2462 if (deoptimize) {
2463 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2464 uncommon_trap(Deoptimization::Reason_unhandled,
2465 Deoptimization::Action_none);
2466 } else {
2467 // Create an exception state also.
2468 // Use an exact type if the caller has specified a specific exception.
2469 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2470 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2471 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2472 }
2473 }
2474 }
2476 // Get the no-exception control from the CatchNode.
2477 set_control(norm);
2478 }
2481 //-------------------------------gen_subtype_check-----------------------------
2482 // Generate a subtyping check. Takes as input the subtype and supertype.
2483 // Returns 2 values: sets the default control() to the true path and returns
2484 // the false path. Only reads invariant memory; sets no (visible) memory.
2485 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2486 // but that's not exposed to the optimizer. This call also doesn't take in an
2487 // Object; if you wish to check an Object you need to load the Object's class
2488 // prior to coming here.
2489 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2490 // Fast check for identical types, perhaps identical constants.
2491 // The types can even be identical non-constants, in cases
2492 // involving Array.newInstance, Object.clone, etc.
2493 if (subklass == superklass)
2494 return top(); // false path is dead; no test needed.
2496 if (_gvn.type(superklass)->singleton()) {
2497 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2498 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2500 // In the common case of an exact superklass, try to fold up the
2501 // test before generating code. You may ask, why not just generate
2502 // the code and then let it fold up? The answer is that the generated
2503 // code will necessarily include null checks, which do not always
2504 // completely fold away. If they are also needless, then they turn
2505 // into a performance loss. Example:
2506 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2507 // Here, the type of 'fa' is often exact, so the store check
2508 // of fa[1]=x will fold up, without testing the nullness of x.
2509 switch (static_subtype_check(superk, subk)) {
2510 case SSC_always_false:
2511 {
2512 Node* always_fail = control();
2513 set_control(top());
2514 return always_fail;
2515 }
2516 case SSC_always_true:
2517 return top();
2518 case SSC_easy_test:
2519 {
2520 // Just do a direct pointer compare and be done.
2521 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2522 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2523 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2524 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2525 return _gvn.transform( new(C) IfFalseNode(iff) );
2526 }
2527 case SSC_full_test:
2528 break;
2529 default:
2530 ShouldNotReachHere();
2531 }
2532 }
2534 // %%% Possible further optimization: Even if the superklass is not exact,
2535 // if the subklass is the unique subtype of the superklass, the check
2536 // will always succeed. We could leave a dependency behind to ensure this.
2538 // First load the super-klass's check-offset
2539 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2540 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(),
2541 TypeInt::INT, MemNode::unordered));
2542 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2543 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2545 // Load from the sub-klass's super-class display list, or a 1-word cache of
2546 // the secondary superclass list, or a failing value with a sentinel offset
2547 // if the super-klass is an interface or exceptionally deep in the Java
2548 // hierarchy and we have to scan the secondary superclass list the hard way.
2549 // Worst-case type is a little odd: NULL is allowed as a result (usually
2550 // klass loads can never produce a NULL).
2551 Node *chk_off_X = ConvI2X(chk_off);
2552 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2553 // For some types like interfaces the following loadKlass is from a 1-word
2554 // cache which is mutable so can't use immutable memory. Other
2555 // types load from the super-class display table which is immutable.
2556 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2557 Node* nkls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2559 // Compile speed common case: ARE a subtype and we canNOT fail
2560 if( superklass == nkls )
2561 return top(); // false path is dead; no test needed.
2563 // See if we get an immediate positive hit. Happens roughly 83% of the
2564 // time. Test to see if the value loaded just previously from the subklass
2565 // is exactly the superklass.
2566 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2567 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2568 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2569 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2570 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2572 // Compile speed common case: Check for being deterministic right now. If
2573 // chk_off is a constant and not equal to cacheoff then we are NOT a
2574 // subklass. In this case we need exactly the 1 test above and we can
2575 // return those results immediately.
2576 if (!might_be_cache) {
2577 Node* not_subtype_ctrl = control();
2578 set_control(iftrue1); // We need exactly the 1 test above
2579 return not_subtype_ctrl;
2580 }
2582 // Gather the various success & failures here
2583 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2584 record_for_igvn(r_ok_subtype);
2585 RegionNode *r_not_subtype = new (C) RegionNode(3);
2586 record_for_igvn(r_not_subtype);
2588 r_ok_subtype->init_req(1, iftrue1);
2590 // Check for immediate negative hit. Happens roughly 11% of the time (which
2591 // is roughly 63% of the remaining cases). Test to see if the loaded
2592 // check-offset points into the subklass display list or the 1-element
2593 // cache. If it points to the display (and NOT the cache) and the display
2594 // missed then it's not a subtype.
2595 Node *cacheoff = _gvn.intcon(cacheoff_con);
2596 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2597 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2598 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2599 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2600 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2602 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2603 // No performance impact (too rare) but allows sharing of secondary arrays
2604 // which has some footprint reduction.
2605 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2606 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2607 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2608 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2609 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2611 // -- Roads not taken here: --
2612 // We could also have chosen to perform the self-check at the beginning
2613 // of this code sequence, as the assembler does. This would not pay off
2614 // the same way, since the optimizer, unlike the assembler, can perform
2615 // static type analysis to fold away many successful self-checks.
2616 // Non-foldable self checks work better here in second position, because
2617 // the initial primary superclass check subsumes a self-check for most
2618 // types. An exception would be a secondary type like array-of-interface,
2619 // which does not appear in its own primary supertype display.
2620 // Finally, we could have chosen to move the self-check into the
2621 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2622 // dependent manner. But it is worthwhile to have the check here,
2623 // where it can be perhaps be optimized. The cost in code space is
2624 // small (register compare, branch).
2626 // Now do a linear scan of the secondary super-klass array. Again, no real
2627 // performance impact (too rare) but it's gotta be done.
2628 // Since the code is rarely used, there is no penalty for moving it
2629 // out of line, and it can only improve I-cache density.
2630 // The decision to inline or out-of-line this final check is platform
2631 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2632 Node* psc = _gvn.transform(
2633 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2635 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2636 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2637 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2638 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2639 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2641 // Return false path; set default control to true path.
2642 set_control( _gvn.transform(r_ok_subtype) );
2643 return _gvn.transform(r_not_subtype);
2644 }
2646 //----------------------------static_subtype_check-----------------------------
2647 // Shortcut important common cases when superklass is exact:
2648 // (0) superklass is java.lang.Object (can occur in reflective code)
2649 // (1) subklass is already limited to a subtype of superklass => always ok
2650 // (2) subklass does not overlap with superklass => always fail
2651 // (3) superklass has NO subtypes and we can check with a simple compare.
2652 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2653 if (StressReflectiveCode) {
2654 return SSC_full_test; // Let caller generate the general case.
2655 }
2657 if (superk == env()->Object_klass()) {
2658 return SSC_always_true; // (0) this test cannot fail
2659 }
2661 ciType* superelem = superk;
2662 if (superelem->is_array_klass())
2663 superelem = superelem->as_array_klass()->base_element_type();
2665 if (!subk->is_interface()) { // cannot trust static interface types yet
2666 if (subk->is_subtype_of(superk)) {
2667 return SSC_always_true; // (1) false path dead; no dynamic test needed
2668 }
2669 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2670 !superk->is_subtype_of(subk)) {
2671 return SSC_always_false;
2672 }
2673 }
2675 // If casting to an instance klass, it must have no subtypes
2676 if (superk->is_interface()) {
2677 // Cannot trust interfaces yet.
2678 // %%% S.B. superk->nof_implementors() == 1
2679 } else if (superelem->is_instance_klass()) {
2680 ciInstanceKlass* ik = superelem->as_instance_klass();
2681 if (!ik->has_subklass() && !ik->is_interface()) {
2682 if (!ik->is_final()) {
2683 // Add a dependency if there is a chance of a later subclass.
2684 C->dependencies()->assert_leaf_type(ik);
2685 }
2686 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2687 }
2688 } else {
2689 // A primitive array type has no subtypes.
2690 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2691 }
2693 return SSC_full_test;
2694 }
2696 // Profile-driven exact type check:
2697 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2698 float prob,
2699 Node* *casted_receiver) {
2700 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2701 Node* recv_klass = load_object_klass(receiver);
2702 Node* want_klass = makecon(tklass);
2703 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2704 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2705 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2706 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2707 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2709 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2710 assert(recv_xtype->klass_is_exact(), "");
2712 // Subsume downstream occurrences of receiver with a cast to
2713 // recv_xtype, since now we know what the type will be.
2714 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2715 (*casted_receiver) = _gvn.transform(cast);
2716 // (User must make the replace_in_map call.)
2718 return fail;
2719 }
2722 //------------------------------seems_never_null-------------------------------
2723 // Use null_seen information if it is available from the profile.
2724 // If we see an unexpected null at a type check we record it and force a
2725 // recompile; the offending check will be recompiled to handle NULLs.
2726 // If we see several offending BCIs, then all checks in the
2727 // method will be recompiled.
2728 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2729 if (UncommonNullCast // Cutout for this technique
2730 && obj != null() // And not the -Xcomp stupid case?
2731 && !too_many_traps(Deoptimization::Reason_null_check)
2732 ) {
2733 if (data == NULL)
2734 // Edge case: no mature data. Be optimistic here.
2735 return true;
2736 // If the profile has not seen a null, assume it won't happen.
2737 assert(java_bc() == Bytecodes::_checkcast ||
2738 java_bc() == Bytecodes::_instanceof ||
2739 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2740 return !data->as_BitData()->null_seen();
2741 }
2742 return false;
2743 }
2745 //------------------------maybe_cast_profiled_receiver-------------------------
2746 // If the profile has seen exactly one type, narrow to exactly that type.
2747 // Subsequent type checks will always fold up.
2748 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2749 ciKlass* require_klass,
2750 ciKlass* spec_klass,
2751 bool safe_for_replace) {
2752 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2754 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check;
2756 // Make sure we haven't already deoptimized from this tactic.
2757 if (too_many_traps(reason) || too_many_recompiles(reason))
2758 return NULL;
2760 // (No, this isn't a call, but it's enough like a virtual call
2761 // to use the same ciMethod accessor to get the profile info...)
2762 // If we have a speculative type use it instead of profiling (which
2763 // may not help us)
2764 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2765 if (exact_kls != NULL) {// no cast failures here
2766 if (require_klass == NULL ||
2767 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2768 // If we narrow the type to match what the type profile sees or
2769 // the speculative type, we can then remove the rest of the
2770 // cast.
2771 // This is a win, even if the exact_kls is very specific,
2772 // because downstream operations, such as method calls,
2773 // will often benefit from the sharper type.
2774 Node* exact_obj = not_null_obj; // will get updated in place...
2775 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2776 &exact_obj);
2777 { PreserveJVMState pjvms(this);
2778 set_control(slow_ctl);
2779 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2780 }
2781 if (safe_for_replace) {
2782 replace_in_map(not_null_obj, exact_obj);
2783 }
2784 return exact_obj;
2785 }
2786 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2787 }
2789 return NULL;
2790 }
2792 /**
2793 * Cast obj to type and emit guard unless we had too many traps here
2794 * already
2795 *
2796 * @param obj node being casted
2797 * @param type type to cast the node to
2798 * @param not_null true if we know node cannot be null
2799 */
2800 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2801 ciKlass* type,
2802 bool not_null) {
2803 // type == NULL if profiling tells us this object is always null
2804 if (type != NULL) {
2805 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2806 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check;
2807 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2808 !too_many_traps(class_reason) && !too_many_recompiles(class_reason)) {
2809 Node* not_null_obj = NULL;
2810 // not_null is true if we know the object is not null and
2811 // there's no need for a null check
2812 if (!not_null) {
2813 Node* null_ctl = top();
2814 not_null_obj = null_check_oop(obj, &null_ctl, true, true);
2815 assert(null_ctl->is_top(), "no null control here");
2816 } else {
2817 not_null_obj = obj;
2818 }
2820 Node* exact_obj = not_null_obj;
2821 ciKlass* exact_kls = type;
2822 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2823 &exact_obj);
2824 {
2825 PreserveJVMState pjvms(this);
2826 set_control(slow_ctl);
2827 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2828 }
2829 replace_in_map(not_null_obj, exact_obj);
2830 obj = exact_obj;
2831 }
2832 } else {
2833 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2834 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2835 Node* exact_obj = null_assert(obj);
2836 replace_in_map(obj, exact_obj);
2837 obj = exact_obj;
2838 }
2839 }
2840 return obj;
2841 }
2843 //-------------------------------gen_instanceof--------------------------------
2844 // Generate an instance-of idiom. Used by both the instance-of bytecode
2845 // and the reflective instance-of call.
2846 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2847 kill_dead_locals(); // Benefit all the uncommon traps
2848 assert( !stopped(), "dead parse path should be checked in callers" );
2849 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2850 "must check for not-null not-dead klass in callers");
2852 // Make the merge point
2853 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2854 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2855 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2856 C->set_has_split_ifs(true); // Has chance for split-if optimization
2858 ciProfileData* data = NULL;
2859 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2860 data = method()->method_data()->bci_to_data(bci());
2861 }
2862 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2863 && seems_never_null(obj, data));
2865 // Null check; get casted pointer; set region slot 3
2866 Node* null_ctl = top();
2867 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2869 // If not_null_obj is dead, only null-path is taken
2870 if (stopped()) { // Doing instance-of on a NULL?
2871 set_control(null_ctl);
2872 return intcon(0);
2873 }
2874 region->init_req(_null_path, null_ctl);
2875 phi ->init_req(_null_path, intcon(0)); // Set null path value
2876 if (null_ctl == top()) {
2877 // Do this eagerly, so that pattern matches like is_diamond_phi
2878 // will work even during parsing.
2879 assert(_null_path == PATH_LIMIT-1, "delete last");
2880 region->del_req(_null_path);
2881 phi ->del_req(_null_path);
2882 }
2884 // Do we know the type check always succeed?
2885 bool known_statically = false;
2886 if (_gvn.type(superklass)->singleton()) {
2887 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2888 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2889 if (subk != NULL && subk->is_loaded()) {
2890 int static_res = static_subtype_check(superk, subk);
2891 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false);
2892 }
2893 }
2895 if (known_statically && UseTypeSpeculation) {
2896 // If we know the type check always succeeds then we don't use the
2897 // profiling data at this bytecode. Don't lose it, feed it to the
2898 // type system as a speculative type.
2899 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2900 } else {
2901 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2902 // We may not have profiling here or it may not help us. If we
2903 // have a speculative type use it to perform an exact cast.
2904 ciKlass* spec_obj_type = obj_type->speculative_type();
2905 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2906 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2907 if (stopped()) { // Profile disagrees with this path.
2908 set_control(null_ctl); // Null is the only remaining possibility.
2909 return intcon(0);
2910 }
2911 if (cast_obj != NULL) {
2912 not_null_obj = cast_obj;
2913 }
2914 }
2915 }
2917 // Load the object's klass
2918 Node* obj_klass = load_object_klass(not_null_obj);
2920 // Generate the subtype check
2921 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2923 // Plug in the success path to the general merge in slot 1.
2924 region->init_req(_obj_path, control());
2925 phi ->init_req(_obj_path, intcon(1));
2927 // Plug in the failing path to the general merge in slot 2.
2928 region->init_req(_fail_path, not_subtype_ctrl);
2929 phi ->init_req(_fail_path, intcon(0));
2931 // Return final merged results
2932 set_control( _gvn.transform(region) );
2933 record_for_igvn(region);
2934 return _gvn.transform(phi);
2935 }
2937 //-------------------------------gen_checkcast---------------------------------
2938 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2939 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2940 // uncommon-trap paths work. Adjust stack after this call.
2941 // If failure_control is supplied and not null, it is filled in with
2942 // the control edge for the cast failure. Otherwise, an appropriate
2943 // uncommon trap or exception is thrown.
2944 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2945 Node* *failure_control) {
2946 kill_dead_locals(); // Benefit all the uncommon traps
2947 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2948 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2950 // Fast cutout: Check the case that the cast is vacuously true.
2951 // This detects the common cases where the test will short-circuit
2952 // away completely. We do this before we perform the null check,
2953 // because if the test is going to turn into zero code, we don't
2954 // want a residual null check left around. (Causes a slowdown,
2955 // for example, in some objArray manipulations, such as a[i]=a[j].)
2956 if (tk->singleton()) {
2957 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2958 if (objtp != NULL && objtp->klass() != NULL) {
2959 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2960 case SSC_always_true:
2961 // If we know the type check always succeed then we don't use
2962 // the profiling data at this bytecode. Don't lose it, feed it
2963 // to the type system as a speculative type.
2964 return record_profiled_receiver_for_speculation(obj);
2965 case SSC_always_false:
2966 // It needs a null check because a null will *pass* the cast check.
2967 // A non-null value will always produce an exception.
2968 return null_assert(obj);
2969 }
2970 }
2971 }
2973 ciProfileData* data = NULL;
2974 bool safe_for_replace = false;
2975 if (failure_control == NULL) { // use MDO in regular case only
2976 assert(java_bc() == Bytecodes::_aastore ||
2977 java_bc() == Bytecodes::_checkcast,
2978 "interpreter profiles type checks only for these BCs");
2979 data = method()->method_data()->bci_to_data(bci());
2980 safe_for_replace = true;
2981 }
2983 // Make the merge point
2984 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2985 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
2986 Node* phi = new (C) PhiNode(region, toop);
2987 C->set_has_split_ifs(true); // Has chance for split-if optimization
2989 // Use null-cast information if it is available
2990 bool never_see_null = ((failure_control == NULL) // regular case only
2991 && seems_never_null(obj, data));
2993 // Null check; get casted pointer; set region slot 3
2994 Node* null_ctl = top();
2995 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2997 // If not_null_obj is dead, only null-path is taken
2998 if (stopped()) { // Doing instance-of on a NULL?
2999 set_control(null_ctl);
3000 return null();
3001 }
3002 region->init_req(_null_path, null_ctl);
3003 phi ->init_req(_null_path, null()); // Set null path value
3004 if (null_ctl == top()) {
3005 // Do this eagerly, so that pattern matches like is_diamond_phi
3006 // will work even during parsing.
3007 assert(_null_path == PATH_LIMIT-1, "delete last");
3008 region->del_req(_null_path);
3009 phi ->del_req(_null_path);
3010 }
3012 Node* cast_obj = NULL;
3013 if (tk->klass_is_exact()) {
3014 // The following optimization tries to statically cast the speculative type of the object
3015 // (for example obtained during profiling) to the type of the superklass and then do a
3016 // dynamic check that the type of the object is what we expect. To work correctly
3017 // for checkcast and aastore the type of superklass should be exact.
3018 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3019 // We may not have profiling here or it may not help us. If we have
3020 // a speculative type use it to perform an exact cast.
3021 ciKlass* spec_obj_type = obj_type->speculative_type();
3022 if (spec_obj_type != NULL ||
3023 (data != NULL &&
3024 // Counter has never been decremented (due to cast failure).
3025 // ...This is a reasonable thing to expect. It is true of
3026 // all casts inserted by javac to implement generic types.
3027 data->as_CounterData()->count() >= 0)) {
3028 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3029 if (cast_obj != NULL) {
3030 if (failure_control != NULL) // failure is now impossible
3031 (*failure_control) = top();
3032 // adjust the type of the phi to the exact klass:
3033 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3034 }
3035 }
3036 }
3038 if (cast_obj == NULL) {
3039 // Load the object's klass
3040 Node* obj_klass = load_object_klass(not_null_obj);
3042 // Generate the subtype check
3043 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3045 // Plug in success path into the merge
3046 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
3047 not_null_obj, toop));
3048 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3049 if (failure_control == NULL) {
3050 if (not_subtype_ctrl != top()) { // If failure is possible
3051 PreserveJVMState pjvms(this);
3052 set_control(not_subtype_ctrl);
3053 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3054 }
3055 } else {
3056 (*failure_control) = not_subtype_ctrl;
3057 }
3058 }
3060 region->init_req(_obj_path, control());
3061 phi ->init_req(_obj_path, cast_obj);
3063 // A merge of NULL or Casted-NotNull obj
3064 Node* res = _gvn.transform(phi);
3066 // Note I do NOT always 'replace_in_map(obj,result)' here.
3067 // if( tk->klass()->can_be_primary_super() )
3068 // This means that if I successfully store an Object into an array-of-String
3069 // I 'forget' that the Object is really now known to be a String. I have to
3070 // do this because we don't have true union types for interfaces - if I store
3071 // a Baz into an array-of-Interface and then tell the optimizer it's an
3072 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3073 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3074 // replace_in_map( obj, res );
3076 // Return final merged results
3077 set_control( _gvn.transform(region) );
3078 record_for_igvn(region);
3079 return res;
3080 }
3082 //------------------------------next_monitor-----------------------------------
3083 // What number should be given to the next monitor?
3084 int GraphKit::next_monitor() {
3085 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3086 int next = current + C->sync_stack_slots();
3087 // Keep the toplevel high water mark current:
3088 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3089 return current;
3090 }
3092 //------------------------------insert_mem_bar---------------------------------
3093 // Memory barrier to avoid floating things around
3094 // The membar serves as a pinch point between both control and all memory slices.
3095 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3096 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3097 mb->init_req(TypeFunc::Control, control());
3098 mb->init_req(TypeFunc::Memory, reset_memory());
3099 Node* membar = _gvn.transform(mb);
3100 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3101 set_all_memory_call(membar);
3102 return membar;
3103 }
3105 //-------------------------insert_mem_bar_volatile----------------------------
3106 // Memory barrier to avoid floating things around
3107 // The membar serves as a pinch point between both control and memory(alias_idx).
3108 // If you want to make a pinch point on all memory slices, do not use this
3109 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3110 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3111 // When Parse::do_put_xxx updates a volatile field, it appends a series
3112 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3113 // The first membar is on the same memory slice as the field store opcode.
3114 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3115 // All the other membars (for other volatile slices, including AliasIdxBot,
3116 // which stands for all unknown volatile slices) are control-dependent
3117 // on the first membar. This prevents later volatile loads or stores
3118 // from sliding up past the just-emitted store.
3120 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3121 mb->set_req(TypeFunc::Control,control());
3122 if (alias_idx == Compile::AliasIdxBot) {
3123 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3124 } else {
3125 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3126 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3127 }
3128 Node* membar = _gvn.transform(mb);
3129 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3130 if (alias_idx == Compile::AliasIdxBot) {
3131 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
3132 } else {
3133 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
3134 }
3135 return membar;
3136 }
3138 //------------------------------shared_lock------------------------------------
3139 // Emit locking code.
3140 FastLockNode* GraphKit::shared_lock(Node* obj) {
3141 // bci is either a monitorenter bc or InvocationEntryBci
3142 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3143 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3145 if( !GenerateSynchronizationCode )
3146 return NULL; // Not locking things?
3147 if (stopped()) // Dead monitor?
3148 return NULL;
3150 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3152 // Box the stack location
3153 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
3154 Node* mem = reset_memory();
3156 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
3157 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3158 // Create the counters for this fast lock.
3159 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3160 }
3162 // Create the rtm counters for this fast lock if needed.
3163 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3165 // Add monitor to debug info for the slow path. If we block inside the
3166 // slow path and de-opt, we need the monitor hanging around
3167 map()->push_monitor( flock );
3169 const TypeFunc *tf = LockNode::lock_type();
3170 LockNode *lock = new (C) LockNode(C, tf);
3172 lock->init_req( TypeFunc::Control, control() );
3173 lock->init_req( TypeFunc::Memory , mem );
3174 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3175 lock->init_req( TypeFunc::FramePtr, frameptr() );
3176 lock->init_req( TypeFunc::ReturnAdr, top() );
3178 lock->init_req(TypeFunc::Parms + 0, obj);
3179 lock->init_req(TypeFunc::Parms + 1, box);
3180 lock->init_req(TypeFunc::Parms + 2, flock);
3181 add_safepoint_edges(lock);
3183 lock = _gvn.transform( lock )->as_Lock();
3185 // lock has no side-effects, sets few values
3186 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3188 insert_mem_bar(Op_MemBarAcquireLock);
3190 // Add this to the worklist so that the lock can be eliminated
3191 record_for_igvn(lock);
3193 #ifndef PRODUCT
3194 if (PrintLockStatistics) {
3195 // Update the counter for this lock. Don't bother using an atomic
3196 // operation since we don't require absolute accuracy.
3197 lock->create_lock_counter(map()->jvms());
3198 increment_counter(lock->counter()->addr());
3199 }
3200 #endif
3202 return flock;
3203 }
3206 //------------------------------shared_unlock----------------------------------
3207 // Emit unlocking code.
3208 void GraphKit::shared_unlock(Node* box, Node* obj) {
3209 // bci is either a monitorenter bc or InvocationEntryBci
3210 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3211 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3213 if( !GenerateSynchronizationCode )
3214 return;
3215 if (stopped()) { // Dead monitor?
3216 map()->pop_monitor(); // Kill monitor from debug info
3217 return;
3218 }
3220 // Memory barrier to avoid floating things down past the locked region
3221 insert_mem_bar(Op_MemBarReleaseLock);
3223 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3224 UnlockNode *unlock = new (C) UnlockNode(C, tf);
3225 #ifdef ASSERT
3226 unlock->set_dbg_jvms(sync_jvms());
3227 #endif
3228 uint raw_idx = Compile::AliasIdxRaw;
3229 unlock->init_req( TypeFunc::Control, control() );
3230 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3231 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3232 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3233 unlock->init_req( TypeFunc::ReturnAdr, top() );
3235 unlock->init_req(TypeFunc::Parms + 0, obj);
3236 unlock->init_req(TypeFunc::Parms + 1, box);
3237 unlock = _gvn.transform(unlock)->as_Unlock();
3239 Node* mem = reset_memory();
3241 // unlock has no side-effects, sets few values
3242 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3244 // Kill monitor from debug info
3245 map()->pop_monitor( );
3246 }
3248 //-------------------------------get_layout_helper-----------------------------
3249 // If the given klass is a constant or known to be an array,
3250 // fetch the constant layout helper value into constant_value
3251 // and return (Node*)NULL. Otherwise, load the non-constant
3252 // layout helper value, and return the node which represents it.
3253 // This two-faced routine is useful because allocation sites
3254 // almost always feature constant types.
3255 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3256 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3257 if (!StressReflectiveCode && inst_klass != NULL) {
3258 ciKlass* klass = inst_klass->klass();
3259 bool xklass = inst_klass->klass_is_exact();
3260 if (xklass || klass->is_array_klass()) {
3261 jint lhelper = klass->layout_helper();
3262 if (lhelper != Klass::_lh_neutral_value) {
3263 constant_value = lhelper;
3264 return (Node*) NULL;
3265 }
3266 }
3267 }
3268 constant_value = Klass::_lh_neutral_value; // put in a known value
3269 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3270 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3271 }
3273 // We just put in an allocate/initialize with a big raw-memory effect.
3274 // Hook selected additional alias categories on the initialization.
3275 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3276 MergeMemNode* init_in_merge,
3277 Node* init_out_raw) {
3278 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3279 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3281 Node* prevmem = kit.memory(alias_idx);
3282 init_in_merge->set_memory_at(alias_idx, prevmem);
3283 kit.set_memory(init_out_raw, alias_idx);
3284 }
3286 //---------------------------set_output_for_allocation-------------------------
3287 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3288 const TypeOopPtr* oop_type,
3289 bool deoptimize_on_exception) {
3290 int rawidx = Compile::AliasIdxRaw;
3291 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3292 add_safepoint_edges(alloc);
3293 Node* allocx = _gvn.transform(alloc);
3294 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3295 // create memory projection for i_o
3296 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3297 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3299 // create a memory projection as for the normal control path
3300 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3301 set_memory(malloc, rawidx);
3303 // a normal slow-call doesn't change i_o, but an allocation does
3304 // we create a separate i_o projection for the normal control path
3305 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3306 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3308 // put in an initialization barrier
3309 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3310 rawoop)->as_Initialize();
3311 assert(alloc->initialization() == init, "2-way macro link must work");
3312 assert(init ->allocation() == alloc, "2-way macro link must work");
3313 {
3314 // Extract memory strands which may participate in the new object's
3315 // initialization, and source them from the new InitializeNode.
3316 // This will allow us to observe initializations when they occur,
3317 // and link them properly (as a group) to the InitializeNode.
3318 assert(init->in(InitializeNode::Memory) == malloc, "");
3319 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3320 init->set_req(InitializeNode::Memory, minit_in);
3321 record_for_igvn(minit_in); // fold it up later, if possible
3322 Node* minit_out = memory(rawidx);
3323 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3324 if (oop_type->isa_aryptr()) {
3325 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3326 int elemidx = C->get_alias_index(telemref);
3327 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3328 } else if (oop_type->isa_instptr()) {
3329 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3330 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3331 ciField* field = ik->nonstatic_field_at(i);
3332 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3333 continue; // do not bother to track really large numbers of fields
3334 // Find (or create) the alias category for this field:
3335 int fieldidx = C->alias_type(field)->index();
3336 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3337 }
3338 }
3339 }
3341 // Cast raw oop to the real thing...
3342 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3343 javaoop = _gvn.transform(javaoop);
3344 C->set_recent_alloc(control(), javaoop);
3345 assert(just_allocated_object(control()) == javaoop, "just allocated");
3347 #ifdef ASSERT
3348 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3349 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3350 "Ideal_allocation works");
3351 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3352 "Ideal_allocation works");
3353 if (alloc->is_AllocateArray()) {
3354 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3355 "Ideal_allocation works");
3356 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3357 "Ideal_allocation works");
3358 } else {
3359 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3360 }
3361 }
3362 #endif //ASSERT
3364 return javaoop;
3365 }
3367 //---------------------------new_instance--------------------------------------
3368 // This routine takes a klass_node which may be constant (for a static type)
3369 // or may be non-constant (for reflective code). It will work equally well
3370 // for either, and the graph will fold nicely if the optimizer later reduces
3371 // the type to a constant.
3372 // The optional arguments are for specialized use by intrinsics:
3373 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3374 // - If 'return_size_val', report the the total object size to the caller.
3375 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3376 Node* GraphKit::new_instance(Node* klass_node,
3377 Node* extra_slow_test,
3378 Node* *return_size_val,
3379 bool deoptimize_on_exception) {
3380 // Compute size in doublewords
3381 // The size is always an integral number of doublewords, represented
3382 // as a positive bytewise size stored in the klass's layout_helper.
3383 // The layout_helper also encodes (in a low bit) the need for a slow path.
3384 jint layout_con = Klass::_lh_neutral_value;
3385 Node* layout_val = get_layout_helper(klass_node, layout_con);
3386 int layout_is_con = (layout_val == NULL);
3388 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3389 // Generate the initial go-slow test. It's either ALWAYS (return a
3390 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3391 // case) a computed value derived from the layout_helper.
3392 Node* initial_slow_test = NULL;
3393 if (layout_is_con) {
3394 assert(!StressReflectiveCode, "stress mode does not use these paths");
3395 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3396 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3398 } else { // reflective case
3399 // This reflective path is used by Unsafe.allocateInstance.
3400 // (It may be stress-tested by specifying StressReflectiveCode.)
3401 // Basically, we want to get into the VM is there's an illegal argument.
3402 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3403 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3404 if (extra_slow_test != intcon(0)) {
3405 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3406 }
3407 // (Macro-expander will further convert this to a Bool, if necessary.)
3408 }
3410 // Find the size in bytes. This is easy; it's the layout_helper.
3411 // The size value must be valid even if the slow path is taken.
3412 Node* size = NULL;
3413 if (layout_is_con) {
3414 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3415 } else { // reflective case
3416 // This reflective path is used by clone and Unsafe.allocateInstance.
3417 size = ConvI2X(layout_val);
3419 // Clear the low bits to extract layout_helper_size_in_bytes:
3420 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3421 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3422 size = _gvn.transform( new (C) AndXNode(size, mask) );
3423 }
3424 if (return_size_val != NULL) {
3425 (*return_size_val) = size;
3426 }
3428 // This is a precise notnull oop of the klass.
3429 // (Actually, it need not be precise if this is a reflective allocation.)
3430 // It's what we cast the result to.
3431 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3432 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3433 const TypeOopPtr* oop_type = tklass->as_instance_type();
3435 // Now generate allocation code
3437 // The entire memory state is needed for slow path of the allocation
3438 // since GC and deoptimization can happened.
3439 Node *mem = reset_memory();
3440 set_all_memory(mem); // Create new memory state
3442 AllocateNode* alloc
3443 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3444 control(), mem, i_o(),
3445 size, klass_node,
3446 initial_slow_test);
3448 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3449 }
3451 //-------------------------------new_array-------------------------------------
3452 // helper for both newarray and anewarray
3453 // The 'length' parameter is (obviously) the length of the array.
3454 // See comments on new_instance for the meaning of the other arguments.
3455 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3456 Node* length, // number of array elements
3457 int nargs, // number of arguments to push back for uncommon trap
3458 Node* *return_size_val,
3459 bool deoptimize_on_exception) {
3460 jint layout_con = Klass::_lh_neutral_value;
3461 Node* layout_val = get_layout_helper(klass_node, layout_con);
3462 int layout_is_con = (layout_val == NULL);
3464 if (!layout_is_con && !StressReflectiveCode &&
3465 !too_many_traps(Deoptimization::Reason_class_check)) {
3466 // This is a reflective array creation site.
3467 // Optimistically assume that it is a subtype of Object[],
3468 // so that we can fold up all the address arithmetic.
3469 layout_con = Klass::array_layout_helper(T_OBJECT);
3470 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3471 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3472 { BuildCutout unless(this, bol_lh, PROB_MAX);
3473 inc_sp(nargs);
3474 uncommon_trap(Deoptimization::Reason_class_check,
3475 Deoptimization::Action_maybe_recompile);
3476 }
3477 layout_val = NULL;
3478 layout_is_con = true;
3479 }
3481 // Generate the initial go-slow test. Make sure we do not overflow
3482 // if length is huge (near 2Gig) or negative! We do not need
3483 // exact double-words here, just a close approximation of needed
3484 // double-words. We can't add any offset or rounding bits, lest we
3485 // take a size -1 of bytes and make it positive. Use an unsigned
3486 // compare, so negative sizes look hugely positive.
3487 int fast_size_limit = FastAllocateSizeLimit;
3488 if (layout_is_con) {
3489 assert(!StressReflectiveCode, "stress mode does not use these paths");
3490 // Increase the size limit if we have exact knowledge of array type.
3491 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3492 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3493 }
3495 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3496 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3498 // --- Size Computation ---
3499 // array_size = round_to_heap(array_header + (length << elem_shift));
3500 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3501 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3502 // The rounding mask is strength-reduced, if possible.
3503 int round_mask = MinObjAlignmentInBytes - 1;
3504 Node* header_size = NULL;
3505 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3506 // (T_BYTE has the weakest alignment and size restrictions...)
3507 if (layout_is_con) {
3508 int hsize = Klass::layout_helper_header_size(layout_con);
3509 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3510 BasicType etype = Klass::layout_helper_element_type(layout_con);
3511 if ((round_mask & ~right_n_bits(eshift)) == 0)
3512 round_mask = 0; // strength-reduce it if it goes away completely
3513 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3514 assert(header_size_min <= hsize, "generic minimum is smallest");
3515 header_size_min = hsize;
3516 header_size = intcon(hsize + round_mask);
3517 } else {
3518 Node* hss = intcon(Klass::_lh_header_size_shift);
3519 Node* hsm = intcon(Klass::_lh_header_size_mask);
3520 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3521 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3522 Node* mask = intcon(round_mask);
3523 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3524 }
3526 Node* elem_shift = NULL;
3527 if (layout_is_con) {
3528 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3529 if (eshift != 0)
3530 elem_shift = intcon(eshift);
3531 } else {
3532 // There is no need to mask or shift this value.
3533 // The semantics of LShiftINode include an implicit mask to 0x1F.
3534 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3535 elem_shift = layout_val;
3536 }
3538 // Transition to native address size for all offset calculations:
3539 Node* lengthx = ConvI2X(length);
3540 Node* headerx = ConvI2X(header_size);
3541 #ifdef _LP64
3542 { const TypeInt* tilen = _gvn.find_int_type(length);
3543 if (tilen != NULL && tilen->_lo < 0) {
3544 // Add a manual constraint to a positive range. Cf. array_element_address.
3545 jlong size_max = fast_size_limit;
3546 if (size_max > tilen->_hi) size_max = tilen->_hi;
3547 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3549 // Only do a narrow I2L conversion if the range check passed.
3550 IfNode* iff = new (C) IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3551 _gvn.transform(iff);
3552 RegionNode* region = new (C) RegionNode(3);
3553 _gvn.set_type(region, Type::CONTROL);
3554 lengthx = new (C) PhiNode(region, TypeLong::LONG);
3555 _gvn.set_type(lengthx, TypeLong::LONG);
3557 // Range check passed. Use ConvI2L node with narrow type.
3558 Node* passed = IfFalse(iff);
3559 region->init_req(1, passed);
3560 // Make I2L conversion control dependent to prevent it from
3561 // floating above the range check during loop optimizations.
3562 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3564 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3565 region->init_req(2, IfTrue(iff));
3566 lengthx->init_req(2, ConvI2X(length));
3568 set_control(region);
3569 record_for_igvn(region);
3570 record_for_igvn(lengthx);
3571 }
3572 }
3573 #endif
3575 // Combine header size (plus rounding) and body size. Then round down.
3576 // This computation cannot overflow, because it is used only in two
3577 // places, one where the length is sharply limited, and the other
3578 // after a successful allocation.
3579 Node* abody = lengthx;
3580 if (elem_shift != NULL)
3581 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3582 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3583 if (round_mask != 0) {
3584 Node* mask = MakeConX(~round_mask);
3585 size = _gvn.transform( new(C) AndXNode(size, mask) );
3586 }
3587 // else if round_mask == 0, the size computation is self-rounding
3589 if (return_size_val != NULL) {
3590 // This is the size
3591 (*return_size_val) = size;
3592 }
3594 // Now generate allocation code
3596 // The entire memory state is needed for slow path of the allocation
3597 // since GC and deoptimization can happened.
3598 Node *mem = reset_memory();
3599 set_all_memory(mem); // Create new memory state
3601 if (initial_slow_test->is_Bool()) {
3602 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3603 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3604 }
3606 // Create the AllocateArrayNode and its result projections
3607 AllocateArrayNode* alloc
3608 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3609 control(), mem, i_o(),
3610 size, klass_node,
3611 initial_slow_test,
3612 length);
3614 // Cast to correct type. Note that the klass_node may be constant or not,
3615 // and in the latter case the actual array type will be inexact also.
3616 // (This happens via a non-constant argument to inline_native_newArray.)
3617 // In any case, the value of klass_node provides the desired array type.
3618 const TypeInt* length_type = _gvn.find_int_type(length);
3619 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3620 if (ary_type->isa_aryptr() && length_type != NULL) {
3621 // Try to get a better type than POS for the size
3622 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3623 }
3625 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3627 // Cast length on remaining path to be as narrow as possible
3628 if (map()->find_edge(length) >= 0) {
3629 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3630 if (ccast != length) {
3631 _gvn.set_type_bottom(ccast);
3632 record_for_igvn(ccast);
3633 replace_in_map(length, ccast);
3634 }
3635 }
3637 return javaoop;
3638 }
3640 // The following "Ideal_foo" functions are placed here because they recognize
3641 // the graph shapes created by the functions immediately above.
3643 //---------------------------Ideal_allocation----------------------------------
3644 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3645 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3646 if (ptr == NULL) { // reduce dumb test in callers
3647 return NULL;
3648 }
3649 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3650 ptr = ptr->in(1);
3651 if (ptr == NULL) return NULL;
3652 }
3653 // Return NULL for allocations with several casts:
3654 // j.l.reflect.Array.newInstance(jobject, jint)
3655 // Object.clone()
3656 // to keep more precise type from last cast.
3657 if (ptr->is_Proj()) {
3658 Node* allo = ptr->in(0);
3659 if (allo != NULL && allo->is_Allocate()) {
3660 return allo->as_Allocate();
3661 }
3662 }
3663 // Report failure to match.
3664 return NULL;
3665 }
3667 // Fancy version which also strips off an offset (and reports it to caller).
3668 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3669 intptr_t& offset) {
3670 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3671 if (base == NULL) return NULL;
3672 return Ideal_allocation(base, phase);
3673 }
3675 // Trace Initialize <- Proj[Parm] <- Allocate
3676 AllocateNode* InitializeNode::allocation() {
3677 Node* rawoop = in(InitializeNode::RawAddress);
3678 if (rawoop->is_Proj()) {
3679 Node* alloc = rawoop->in(0);
3680 if (alloc->is_Allocate()) {
3681 return alloc->as_Allocate();
3682 }
3683 }
3684 return NULL;
3685 }
3687 // Trace Allocate -> Proj[Parm] -> Initialize
3688 InitializeNode* AllocateNode::initialization() {
3689 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3690 if (rawoop == NULL) return NULL;
3691 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3692 Node* init = rawoop->fast_out(i);
3693 if (init->is_Initialize()) {
3694 assert(init->as_Initialize()->allocation() == this, "2-way link");
3695 return init->as_Initialize();
3696 }
3697 }
3698 return NULL;
3699 }
3701 //----------------------------- loop predicates ---------------------------
3703 //------------------------------add_predicate_impl----------------------------
3704 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3705 // Too many traps seen?
3706 if (too_many_traps(reason)) {
3707 #ifdef ASSERT
3708 if (TraceLoopPredicate) {
3709 int tc = C->trap_count(reason);
3710 tty->print("too many traps=%s tcount=%d in ",
3711 Deoptimization::trap_reason_name(reason), tc);
3712 method()->print(); // which method has too many predicate traps
3713 tty->cr();
3714 }
3715 #endif
3716 // We cannot afford to take more traps here,
3717 // do not generate predicate.
3718 return;
3719 }
3721 Node *cont = _gvn.intcon(1);
3722 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3723 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3724 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3725 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3726 C->add_predicate_opaq(opq);
3727 {
3728 PreserveJVMState pjvms(this);
3729 set_control(iffalse);
3730 inc_sp(nargs);
3731 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3732 }
3733 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3734 set_control(iftrue);
3735 }
3737 //------------------------------add_predicate---------------------------------
3738 void GraphKit::add_predicate(int nargs) {
3739 if (UseLoopPredicate) {
3740 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3741 }
3742 // loop's limit check predicate should be near the loop.
3743 if (LoopLimitCheck) {
3744 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3745 }
3746 }
3748 //----------------------------- store barriers ----------------------------
3749 #define __ ideal.
3751 void GraphKit::sync_kit(IdealKit& ideal) {
3752 set_all_memory(__ merged_memory());
3753 set_i_o(__ i_o());
3754 set_control(__ ctrl());
3755 }
3757 void GraphKit::final_sync(IdealKit& ideal) {
3758 // Final sync IdealKit and graphKit.
3759 sync_kit(ideal);
3760 }
3762 // vanilla/CMS post barrier
3763 // Insert a write-barrier store. This is to let generational GC work; we have
3764 // to flag all oop-stores before the next GC point.
3765 void GraphKit::write_barrier_post(Node* oop_store,
3766 Node* obj,
3767 Node* adr,
3768 uint adr_idx,
3769 Node* val,
3770 bool use_precise) {
3771 // No store check needed if we're storing a NULL or an old object
3772 // (latter case is probably a string constant). The concurrent
3773 // mark sweep garbage collector, however, needs to have all nonNull
3774 // oop updates flagged via card-marks.
3775 if (val != NULL && val->is_Con()) {
3776 // must be either an oop or NULL
3777 const Type* t = val->bottom_type();
3778 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3779 // stores of null never (?) need barriers
3780 return;
3781 }
3783 if (use_ReduceInitialCardMarks()
3784 && obj == just_allocated_object(control())) {
3785 // We can skip marks on a freshly-allocated object in Eden.
3786 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3787 // That routine informs GC to take appropriate compensating steps,
3788 // upon a slow-path allocation, so as to make this card-mark
3789 // elision safe.
3790 return;
3791 }
3793 if (!use_precise) {
3794 // All card marks for a (non-array) instance are in one place:
3795 adr = obj;
3796 }
3797 // (Else it's an array (or unknown), and we want more precise card marks.)
3798 assert(adr != NULL, "");
3800 IdealKit ideal(this, true);
3802 // Convert the pointer to an int prior to doing math on it
3803 Node* cast = __ CastPX(__ ctrl(), adr);
3805 // Divide by card size
3806 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3807 "Only one we handle so far.");
3808 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3810 // Combine card table base and card offset
3811 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3813 // Get the alias_index for raw card-mark memory
3814 int adr_type = Compile::AliasIdxRaw;
3815 Node* zero = __ ConI(0); // Dirty card value
3816 BasicType bt = T_BYTE;
3818 if (UseCondCardMark) {
3819 // The classic GC reference write barrier is typically implemented
3820 // as a store into the global card mark table. Unfortunately
3821 // unconditional stores can result in false sharing and excessive
3822 // coherence traffic as well as false transactional aborts.
3823 // UseCondCardMark enables MP "polite" conditional card mark
3824 // stores. In theory we could relax the load from ctrl() to
3825 // no_ctrl, but that doesn't buy much latitude.
3826 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3827 __ if_then(card_val, BoolTest::ne, zero);
3828 }
3830 // Smash zero into card
3831 if( !UseConcMarkSweepGC ) {
3832 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release);
3833 } else {
3834 // Specialized path for CM store barrier
3835 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3836 }
3838 if (UseCondCardMark) {
3839 __ end_if();
3840 }
3842 // Final sync IdealKit and GraphKit.
3843 final_sync(ideal);
3844 }
3846 // G1 pre/post barriers
3847 void GraphKit::g1_write_barrier_pre(bool do_load,
3848 Node* obj,
3849 Node* adr,
3850 uint alias_idx,
3851 Node* val,
3852 const TypeOopPtr* val_type,
3853 Node* pre_val,
3854 BasicType bt) {
3856 // Some sanity checks
3857 // Note: val is unused in this routine.
3859 if (do_load) {
3860 // We need to generate the load of the previous value
3861 assert(obj != NULL, "must have a base");
3862 assert(adr != NULL, "where are loading from?");
3863 assert(pre_val == NULL, "loaded already?");
3864 assert(val_type != NULL, "need a type");
3865 } else {
3866 // In this case both val_type and alias_idx are unused.
3867 assert(pre_val != NULL, "must be loaded already");
3868 // Nothing to be done if pre_val is null.
3869 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3870 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3871 }
3872 assert(bt == T_OBJECT, "or we shouldn't be here");
3874 IdealKit ideal(this, true);
3876 Node* tls = __ thread(); // ThreadLocalStorage
3878 Node* no_ctrl = NULL;
3879 Node* no_base = __ top();
3880 Node* zero = __ ConI(0);
3881 Node* zeroX = __ ConX(0);
3883 float likely = PROB_LIKELY(0.999);
3884 float unlikely = PROB_UNLIKELY(0.999);
3886 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3887 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3889 // Offsets into the thread
3890 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3891 PtrQueue::byte_offset_of_active());
3892 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3893 PtrQueue::byte_offset_of_index());
3894 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3895 PtrQueue::byte_offset_of_buf());
3897 // Now the actual pointers into the thread
3898 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3899 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3900 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3902 // Now some of the values
3903 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3905 // if (!marking)
3906 __ if_then(marking, BoolTest::ne, zero, unlikely); {
3907 BasicType index_bt = TypeX_X->basic_type();
3908 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3909 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3911 if (do_load) {
3912 // load original value
3913 // alias_idx correct??
3914 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3915 }
3917 // if (pre_val != NULL)
3918 __ if_then(pre_val, BoolTest::ne, null()); {
3919 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3921 // is the queue for this thread full?
3922 __ if_then(index, BoolTest::ne, zeroX, likely); {
3924 // decrement the index
3925 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3927 // Now get the buffer location we will log the previous value into and store it
3928 Node *log_addr = __ AddP(no_base, buffer, next_index);
3929 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
3930 // update the index
3931 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
3933 } __ else_(); {
3935 // logging buffer is full, call the runtime
3936 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3937 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3938 } __ end_if(); // (!index)
3939 } __ end_if(); // (pre_val != NULL)
3940 } __ end_if(); // (!marking)
3942 // Final sync IdealKit and GraphKit.
3943 final_sync(ideal);
3944 }
3946 //
3947 // Update the card table and add card address to the queue
3948 //
3949 void GraphKit::g1_mark_card(IdealKit& ideal,
3950 Node* card_adr,
3951 Node* oop_store,
3952 uint oop_alias_idx,
3953 Node* index,
3954 Node* index_adr,
3955 Node* buffer,
3956 const TypeFunc* tf) {
3958 Node* zero = __ ConI(0);
3959 Node* zeroX = __ ConX(0);
3960 Node* no_base = __ top();
3961 BasicType card_bt = T_BYTE;
3962 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3963 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3965 // Now do the queue work
3966 __ if_then(index, BoolTest::ne, zeroX); {
3968 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3969 Node* log_addr = __ AddP(no_base, buffer, next_index);
3971 // Order, see storeCM.
3972 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
3973 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
3975 } __ else_(); {
3976 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3977 } __ end_if();
3979 }
3981 void GraphKit::g1_write_barrier_post(Node* oop_store,
3982 Node* obj,
3983 Node* adr,
3984 uint alias_idx,
3985 Node* val,
3986 BasicType bt,
3987 bool use_precise) {
3988 // If we are writing a NULL then we need no post barrier
3990 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3991 // Must be NULL
3992 const Type* t = val->bottom_type();
3993 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3994 // No post barrier if writing NULLx
3995 return;
3996 }
3998 if (!use_precise) {
3999 // All card marks for a (non-array) instance are in one place:
4000 adr = obj;
4001 }
4002 // (Else it's an array (or unknown), and we want more precise card marks.)
4003 assert(adr != NULL, "");
4005 IdealKit ideal(this, true);
4007 Node* tls = __ thread(); // ThreadLocalStorage
4009 Node* no_base = __ top();
4010 float likely = PROB_LIKELY(0.999);
4011 float unlikely = PROB_UNLIKELY(0.999);
4012 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
4013 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
4014 Node* zeroX = __ ConX(0);
4016 // Get the alias_index for raw card-mark memory
4017 const TypePtr* card_type = TypeRawPtr::BOTTOM;
4019 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4021 // Offsets into the thread
4022 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4023 PtrQueue::byte_offset_of_index());
4024 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4025 PtrQueue::byte_offset_of_buf());
4027 // Pointers into the thread
4029 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4030 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4032 // Now some values
4033 // Use ctrl to avoid hoisting these values past a safepoint, which could
4034 // potentially reset these fields in the JavaThread.
4035 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4036 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4038 // Convert the store obj pointer to an int prior to doing math on it
4039 // Must use ctrl to prevent "integerized oop" existing across safepoint
4040 Node* cast = __ CastPX(__ ctrl(), adr);
4042 // Divide pointer by card size
4043 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4045 // Combine card table base and card offset
4046 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4048 // If we know the value being stored does it cross regions?
4050 if (val != NULL) {
4051 // Does the store cause us to cross regions?
4053 // Should be able to do an unsigned compare of region_size instead of
4054 // and extra shift. Do we have an unsigned compare??
4055 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4056 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4058 // if (xor_res == 0) same region so skip
4059 __ if_then(xor_res, BoolTest::ne, zeroX); {
4061 // No barrier if we are storing a NULL
4062 __ if_then(val, BoolTest::ne, null(), unlikely); {
4064 // Ok must mark the card if not already dirty
4066 // load the original value of the card
4067 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4069 __ if_then(card_val, BoolTest::ne, young_card); {
4070 sync_kit(ideal);
4071 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4072 insert_mem_bar(Op_MemBarVolatile, oop_store);
4073 __ sync_kit(this);
4075 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4076 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4077 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4078 } __ end_if();
4079 } __ end_if();
4080 } __ end_if();
4081 } __ end_if();
4082 } else {
4083 // Object.clone() instrinsic uses this path.
4084 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4085 }
4087 // Final sync IdealKit and GraphKit.
4088 final_sync(ideal);
4089 }
4090 #undef __
4094 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
4095 if (java_lang_String::has_offset_field()) {
4096 int offset_offset = java_lang_String::offset_offset_in_bytes();
4097 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4098 false, NULL, 0);
4099 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4100 int offset_field_idx = C->get_alias_index(offset_field_type);
4101 return make_load(ctrl,
4102 basic_plus_adr(str, str, offset_offset),
4103 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered);
4104 } else {
4105 return intcon(0);
4106 }
4107 }
4109 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4110 if (java_lang_String::has_count_field()) {
4111 int count_offset = java_lang_String::count_offset_in_bytes();
4112 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4113 false, NULL, 0);
4114 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4115 int count_field_idx = C->get_alias_index(count_field_type);
4116 return make_load(ctrl,
4117 basic_plus_adr(str, str, count_offset),
4118 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered);
4119 } else {
4120 return load_array_length(load_String_value(ctrl, str));
4121 }
4122 }
4124 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4125 int value_offset = java_lang_String::value_offset_in_bytes();
4126 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4127 false, NULL, 0);
4128 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4129 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4130 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
4131 ciTypeArrayKlass::make(T_CHAR), true, 0);
4132 int value_field_idx = C->get_alias_index(value_field_type);
4133 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4134 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4135 // String.value field is known to be @Stable.
4136 if (UseImplicitStableValues) {
4137 load = cast_array_to_stable(load, value_type);
4138 }
4139 return load;
4140 }
4142 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
4143 int offset_offset = java_lang_String::offset_offset_in_bytes();
4144 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4145 false, NULL, 0);
4146 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4147 int offset_field_idx = C->get_alias_index(offset_field_type);
4148 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
4149 value, T_INT, offset_field_idx, MemNode::unordered);
4150 }
4152 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4153 int value_offset = java_lang_String::value_offset_in_bytes();
4154 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4155 false, NULL, 0);
4156 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4158 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4159 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered);
4160 }
4162 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
4163 int count_offset = java_lang_String::count_offset_in_bytes();
4164 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4165 false, NULL, 0);
4166 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4167 int count_field_idx = C->get_alias_index(count_field_type);
4168 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
4169 value, T_INT, count_field_idx, MemNode::unordered);
4170 }
4172 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4173 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4174 // assumption of CCP analysis.
4175 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
4176 }