2
* Copyright (c) 2001, 2024, Oracle and/or its affiliates. All rights reserved.
3
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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.
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
25
#include "precompiled.hpp"
26
#include "ci/ciUtilities.hpp"
27
#include "classfile/javaClasses.hpp"
28
#include "ci/ciObjArray.hpp"
29
#include "asm/register.hpp"
30
#include "compiler/compileLog.hpp"
31
#include "gc/shared/barrierSet.hpp"
32
#include "gc/shared/c2/barrierSetC2.hpp"
33
#include "interpreter/interpreter.hpp"
34
#include "memory/resourceArea.hpp"
35
#include "opto/addnode.hpp"
36
#include "opto/castnode.hpp"
37
#include "opto/convertnode.hpp"
38
#include "opto/graphKit.hpp"
39
#include "opto/idealKit.hpp"
40
#include "opto/intrinsicnode.hpp"
41
#include "opto/locknode.hpp"
42
#include "opto/machnode.hpp"
43
#include "opto/opaquenode.hpp"
44
#include "opto/parse.hpp"
45
#include "opto/rootnode.hpp"
46
#include "opto/runtime.hpp"
47
#include "opto/subtypenode.hpp"
48
#include "runtime/deoptimization.hpp"
49
#include "runtime/sharedRuntime.hpp"
50
#include "utilities/bitMap.inline.hpp"
51
#include "utilities/powerOfTwo.hpp"
52
#include "utilities/growableArray.hpp"
54
//----------------------------GraphKit-----------------------------------------
55
// Main utility constructor.
56
GraphKit::GraphKit(JVMState* jvms)
57
: Phase(Phase::Parser),
59
_gvn(*C->initial_gvn()),
60
_barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
62
_exceptions = jvms->map()->next_exception();
63
if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
67
// Private constructor for parser.
69
: Phase(Phase::Parser),
71
_gvn(*C->initial_gvn()),
72
_barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
74
_exceptions = nullptr;
76
debug_only(_sp = -99);
77
debug_only(set_bci(-99));
82
//---------------------------clean_stack---------------------------------------
83
// Clear away rubbish from the stack area of the JVM state.
84
// This destroys any arguments that may be waiting on the stack.
85
void GraphKit::clean_stack(int from_sp) {
86
SafePointNode* map = this->map();
87
JVMState* jvms = this->jvms();
88
int stk_size = jvms->stk_size();
89
int stkoff = jvms->stkoff();
90
Node* top = this->top();
91
for (int i = from_sp; i < stk_size; i++) {
92
if (map->in(stkoff + i) != top) {
93
map->set_req(stkoff + i, top);
99
//--------------------------------sync_jvms-----------------------------------
100
// Make sure our current jvms agrees with our parse state.
101
JVMState* GraphKit::sync_jvms() const {
102
JVMState* jvms = this->jvms();
103
jvms->set_bci(bci()); // Record the new bci in the JVMState
104
jvms->set_sp(sp()); // Record the new sp in the JVMState
105
assert(jvms_in_sync(), "jvms is now in sync");
109
//--------------------------------sync_jvms_for_reexecute---------------------
110
// Make sure our current jvms agrees with our parse state. This version
111
// uses the reexecute_sp for reexecuting bytecodes.
112
JVMState* GraphKit::sync_jvms_for_reexecute() {
113
JVMState* jvms = this->jvms();
114
jvms->set_bci(bci()); // Record the new bci in the JVMState
115
jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
120
bool GraphKit::jvms_in_sync() const {
121
Parse* parse = is_Parse();
122
if (parse == nullptr) {
123
if (bci() != jvms()->bci()) return false;
124
if (sp() != (int)jvms()->sp()) return false;
127
if (jvms()->method() != parse->method()) return false;
128
if (jvms()->bci() != parse->bci()) return false;
129
int jvms_sp = jvms()->sp();
130
if (jvms_sp != parse->sp()) return false;
131
int jvms_depth = jvms()->depth();
132
if (jvms_depth != parse->depth()) return false;
136
// Local helper checks for special internal merge points
137
// used to accumulate and merge exception states.
138
// They are marked by the region's in(0) edge being the map itself.
139
// Such merge points must never "escape" into the parser at large,
140
// until they have been handed to gvn.transform.
141
static bool is_hidden_merge(Node* reg) {
142
if (reg == nullptr) return false;
145
if (reg == nullptr) return false;
147
return reg->is_Region() && reg->in(0) != nullptr && reg->in(0)->is_Root();
150
void GraphKit::verify_map() const {
151
if (map() == nullptr) return; // null map is OK
152
assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
153
assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
154
assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
157
void GraphKit::verify_exception_state(SafePointNode* ex_map) {
158
assert(ex_map->next_exception() == nullptr, "not already part of a chain");
159
assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
163
//---------------------------stop_and_kill_map---------------------------------
164
// Set _map to null, signalling a stop to further bytecode execution.
165
// First smash the current map's control to a constant, to mark it dead.
166
void GraphKit::stop_and_kill_map() {
167
SafePointNode* dead_map = stop();
168
if (dead_map != nullptr) {
169
dead_map->disconnect_inputs(C); // Mark the map as killed.
170
assert(dead_map->is_killed(), "must be so marked");
175
//--------------------------------stopped--------------------------------------
176
// Tell if _map is null, or control is top.
177
bool GraphKit::stopped() {
178
if (map() == nullptr) return true;
179
else if (control() == top()) return true;
184
//-----------------------------has_exception_handler----------------------------------
185
// Tell if this method or any caller method has exception handlers.
186
bool GraphKit::has_exception_handler() {
187
for (JVMState* jvmsp = jvms(); jvmsp != nullptr; jvmsp = jvmsp->caller()) {
188
if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
195
//------------------------------save_ex_oop------------------------------------
196
// Save an exception without blowing stack contents or other JVM state.
197
void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
198
assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
199
ex_map->add_req(ex_oop);
200
debug_only(verify_exception_state(ex_map));
203
inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
204
assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
205
Node* ex_oop = ex_map->in(ex_map->req()-1);
206
if (clear_it) ex_map->del_req(ex_map->req()-1);
210
//-----------------------------saved_ex_oop------------------------------------
211
// Recover a saved exception from its map.
212
Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
213
return common_saved_ex_oop(ex_map, false);
216
//--------------------------clear_saved_ex_oop---------------------------------
217
// Erase a previously saved exception from its map.
218
Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
219
return common_saved_ex_oop(ex_map, true);
223
//---------------------------has_saved_ex_oop----------------------------------
224
// Erase a previously saved exception from its map.
225
bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
226
return ex_map->req() == ex_map->jvms()->endoff()+1;
230
//-------------------------make_exception_state--------------------------------
231
// Turn the current JVM state into an exception state, appending the ex_oop.
232
SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
234
SafePointNode* ex_map = stop(); // do not manipulate this map any more
235
set_saved_ex_oop(ex_map, ex_oop);
240
//--------------------------add_exception_state--------------------------------
241
// Add an exception to my list of exceptions.
242
void GraphKit::add_exception_state(SafePointNode* ex_map) {
243
if (ex_map == nullptr || ex_map->control() == top()) {
247
verify_exception_state(ex_map);
248
if (has_exceptions()) {
249
assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
253
// If there is already an exception of exactly this type, merge with it.
254
// In particular, null-checks and other low-level exceptions common up here.
255
Node* ex_oop = saved_ex_oop(ex_map);
256
const Type* ex_type = _gvn.type(ex_oop);
257
if (ex_oop == top()) {
261
assert(ex_type->isa_instptr(), "exception must be an instance");
262
for (SafePointNode* e2 = _exceptions; e2 != nullptr; e2 = e2->next_exception()) {
263
const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
264
// We check sp also because call bytecodes can generate exceptions
265
// both before and after arguments are popped!
266
if (ex_type2 == ex_type
267
&& e2->_jvms->sp() == ex_map->_jvms->sp()) {
268
combine_exception_states(ex_map, e2);
273
// No pre-existing exception of the same type. Chain it on the list.
274
push_exception_state(ex_map);
277
//-----------------------add_exception_states_from-----------------------------
278
void GraphKit::add_exception_states_from(JVMState* jvms) {
279
SafePointNode* ex_map = jvms->map()->next_exception();
280
if (ex_map != nullptr) {
281
jvms->map()->set_next_exception(nullptr);
282
for (SafePointNode* next_map; ex_map != nullptr; ex_map = next_map) {
283
next_map = ex_map->next_exception();
284
ex_map->set_next_exception(nullptr);
285
add_exception_state(ex_map);
290
//-----------------------transfer_exceptions_into_jvms-------------------------
291
JVMState* GraphKit::transfer_exceptions_into_jvms() {
292
if (map() == nullptr) {
293
// We need a JVMS to carry the exceptions, but the map has gone away.
294
// Create a scratch JVMS, cloned from any of the exception states...
295
if (has_exceptions()) {
298
_map->set_next_exception(nullptr);
299
clear_saved_ex_oop(_map);
300
debug_only(verify_map());
302
// ...or created from scratch
303
JVMState* jvms = new (C) JVMState(_method, nullptr);
306
jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
308
for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
309
set_all_memory(top());
310
while (map()->req() < jvms->endoff()) map()->add_req(top());
312
// (This is a kludge, in case you didn't notice.)
315
JVMState* jvms = sync_jvms();
316
assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
317
jvms->map()->set_next_exception(_exceptions);
318
_exceptions = nullptr; // done with this set of exceptions
322
static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
323
assert(is_hidden_merge(dstphi), "must be a special merge node");
324
assert(is_hidden_merge(srcphi), "must be a special merge node");
325
uint limit = srcphi->req();
326
for (uint i = PhiNode::Input; i < limit; i++) {
327
dstphi->add_req(srcphi->in(i));
330
static inline void add_one_req(Node* dstphi, Node* src) {
331
assert(is_hidden_merge(dstphi), "must be a special merge node");
332
assert(!is_hidden_merge(src), "must not be a special merge node");
333
dstphi->add_req(src);
336
//-----------------------combine_exception_states------------------------------
337
// This helper function combines exception states by building phis on a
338
// specially marked state-merging region. These regions and phis are
339
// untransformed, and can build up gradually. The region is marked by
340
// having a control input of its exception map, rather than null. Such
341
// regions do not appear except in this function, and in use_exception_state.
342
void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
343
if (failing()) return; // dying anyway...
344
JVMState* ex_jvms = ex_map->_jvms;
345
assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
346
assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
347
assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
348
assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
349
assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
350
assert(ex_map->req() == phi_map->req(), "matching maps");
351
uint tos = ex_jvms->stkoff() + ex_jvms->sp();
352
Node* hidden_merge_mark = root();
353
Node* region = phi_map->control();
354
MergeMemNode* phi_mem = phi_map->merged_memory();
355
MergeMemNode* ex_mem = ex_map->merged_memory();
356
if (region->in(0) != hidden_merge_mark) {
357
// The control input is not (yet) a specially-marked region in phi_map.
358
// Make it so, and build some phis.
359
region = new RegionNode(2);
360
_gvn.set_type(region, Type::CONTROL);
361
region->set_req(0, hidden_merge_mark); // marks an internal ex-state
362
region->init_req(1, phi_map->control());
363
phi_map->set_control(region);
364
Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
365
record_for_igvn(io_phi);
366
_gvn.set_type(io_phi, Type::ABIO);
367
phi_map->set_i_o(io_phi);
368
for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
369
Node* m = mms.memory();
370
Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
371
record_for_igvn(m_phi);
372
_gvn.set_type(m_phi, Type::MEMORY);
373
mms.set_memory(m_phi);
377
// Either or both of phi_map and ex_map might already be converted into phis.
378
Node* ex_control = ex_map->control();
379
// if there is special marking on ex_map also, we add multiple edges from src
380
bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
381
// how wide was the destination phi_map, originally?
382
uint orig_width = region->req();
385
add_n_reqs(region, ex_control);
386
add_n_reqs(phi_map->i_o(), ex_map->i_o());
388
// ex_map has no merges, so we just add single edges everywhere
389
add_one_req(region, ex_control);
390
add_one_req(phi_map->i_o(), ex_map->i_o());
392
for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
393
if (mms.is_empty()) {
394
// get a copy of the base memory, and patch some inputs into it
395
const TypePtr* adr_type = mms.adr_type(C);
396
Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
397
assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
399
// Prepare to append interesting stuff onto the newly sliced phi:
400
while (phi->req() > orig_width) phi->del_req(phi->req()-1);
402
// Append stuff from ex_map:
404
add_n_reqs(mms.memory(), mms.memory2());
406
add_one_req(mms.memory(), mms.memory2());
409
uint limit = ex_map->req();
410
for (uint i = TypeFunc::Parms; i < limit; i++) {
411
// Skip everything in the JVMS after tos. (The ex_oop follows.)
412
if (i == tos) i = ex_jvms->monoff();
413
Node* src = ex_map->in(i);
414
Node* dst = phi_map->in(i);
417
if (dst->in(0) != region) {
418
dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
419
record_for_igvn(phi);
420
_gvn.set_type(phi, phi->type());
421
phi_map->set_req(i, dst);
422
// Prepare to append interesting stuff onto the new phi:
423
while (dst->req() > orig_width) dst->del_req(dst->req()-1);
425
assert(dst->is_Phi(), "nobody else uses a hidden region");
428
if (add_multiple && src->in(0) == ex_control) {
430
add_n_reqs(dst, src);
432
while (dst->req() < region->req()) add_one_req(dst, src);
434
const Type* srctype = _gvn.type(src);
435
if (phi->type() != srctype) {
436
const Type* dsttype = phi->type()->meet_speculative(srctype);
437
if (phi->type() != dsttype) {
438
phi->set_type(dsttype);
439
_gvn.set_type(phi, dsttype);
444
phi_map->merge_replaced_nodes_with(ex_map);
447
//--------------------------use_exception_state--------------------------------
448
Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
449
if (failing()) { stop(); return top(); }
450
Node* region = phi_map->control();
451
Node* hidden_merge_mark = root();
452
assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
453
Node* ex_oop = clear_saved_ex_oop(phi_map);
454
if (region->in(0) == hidden_merge_mark) {
455
// Special marking for internal ex-states. Process the phis now.
456
region->set_req(0, region); // now it's an ordinary region
457
set_jvms(phi_map->jvms()); // ...so now we can use it as a map
458
// Note: Setting the jvms also sets the bci and sp.
459
set_control(_gvn.transform(region));
460
uint tos = jvms()->stkoff() + sp();
461
for (uint i = 1; i < tos; i++) {
462
Node* x = phi_map->in(i);
463
if (x->in(0) == region) {
464
assert(x->is_Phi(), "expected a special phi");
465
phi_map->set_req(i, _gvn.transform(x));
468
for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
469
Node* x = mms.memory();
470
if (x->in(0) == region) {
471
assert(x->is_Phi(), "nobody else uses a hidden region");
472
mms.set_memory(_gvn.transform(x));
475
if (ex_oop->in(0) == region) {
476
assert(ex_oop->is_Phi(), "expected a special phi");
477
ex_oop = _gvn.transform(ex_oop);
480
set_jvms(phi_map->jvms());
483
assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
484
assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
488
//---------------------------------java_bc-------------------------------------
489
Bytecodes::Code GraphKit::java_bc() const {
490
ciMethod* method = this->method();
491
int bci = this->bci();
492
if (method != nullptr && bci != InvocationEntryBci)
493
return method->java_code_at_bci(bci);
495
return Bytecodes::_illegal;
498
void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
500
// if the exception capability is set, then we will generate code
501
// to check the JavaThread.should_post_on_exceptions flag to see
502
// if we actually need to report exception events (for this
503
// thread). If we don't need to report exception events, we will
504
// take the normal fast path provided by add_exception_events. If
505
// exception event reporting is enabled for this thread, we will
506
// take the uncommon_trap in the BuildCutout below.
508
// first must access the should_post_on_exceptions_flag in this thread's JavaThread
509
Node* jthread = _gvn.transform(new ThreadLocalNode());
510
Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
511
Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
513
// Test the should_post_on_exceptions_flag vs. 0
514
Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
515
Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
517
// Branch to slow_path if should_post_on_exceptions_flag was true
518
{ BuildCutout unless(this, tst, PROB_MAX);
519
// Do not try anything fancy if we're notifying the VM on every throw.
520
// Cf. case Bytecodes::_athrow in parse2.cpp.
521
uncommon_trap(reason, Deoptimization::Action_none,
522
(ciKlass*)nullptr, (char*)nullptr, must_throw);
527
//------------------------------builtin_throw----------------------------------
528
void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) {
529
bool must_throw = true;
531
// If this particular condition has not yet happened at this
532
// bytecode, then use the uncommon trap mechanism, and allow for
533
// a future recompilation if several traps occur here.
534
// If the throw is hot, try to use a more complicated inline mechanism
535
// which keeps execution inside the compiled code.
536
bool treat_throw_as_hot = false;
537
ciMethodData* md = method()->method_data();
540
if (too_many_traps(reason)) {
541
treat_throw_as_hot = true;
543
// (If there is no MDO at all, assume it is early in
544
// execution, and that any deopts are part of the
545
// startup transient, and don't need to be remembered.)
547
// Also, if there is a local exception handler, treat all throws
548
// as hot if there has been at least one in this method.
549
if (C->trap_count(reason) != 0
550
&& method()->method_data()->trap_count(reason) != 0
551
&& has_exception_handler()) {
552
treat_throw_as_hot = true;
556
// If this throw happens frequently, an uncommon trap might cause
557
// a performance pothole. If there is a local exception handler,
558
// and if this particular bytecode appears to be deoptimizing often,
559
// let us handle the throw inline, with a preconstructed instance.
560
// Note: If the deopt count has blown up, the uncommon trap
561
// runtime is going to flush this nmethod, not matter what.
562
if (treat_throw_as_hot && method()->can_omit_stack_trace()) {
563
// If the throw is local, we use a pre-existing instance and
564
// punt on the backtrace. This would lead to a missing backtrace
565
// (a repeat of 4292742) if the backtrace object is ever asked
566
// for its backtrace.
567
// Fixing this remaining case of 4292742 requires some flavor of
568
// escape analysis. Leave that for the future.
569
ciInstance* ex_obj = nullptr;
571
case Deoptimization::Reason_null_check:
572
ex_obj = env()->NullPointerException_instance();
574
case Deoptimization::Reason_div0_check:
575
ex_obj = env()->ArithmeticException_instance();
577
case Deoptimization::Reason_range_check:
578
ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
580
case Deoptimization::Reason_class_check:
581
ex_obj = env()->ClassCastException_instance();
583
case Deoptimization::Reason_array_check:
584
ex_obj = env()->ArrayStoreException_instance();
589
if (failing()) { stop(); return; } // exception allocation might fail
590
if (ex_obj != nullptr) {
591
if (env()->jvmti_can_post_on_exceptions()) {
592
// check if we must post exception events, take uncommon trap if so
593
uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
594
// here if should_post_on_exceptions is false
595
// continue on with the normal codegen
598
// Cheat with a preallocated exception object.
599
if (C->log() != nullptr)
600
C->log()->elem("hot_throw preallocated='1' reason='%s'",
601
Deoptimization::trap_reason_name(reason));
602
const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
603
Node* ex_node = _gvn.transform(ConNode::make(ex_con));
605
// Clear the detail message of the preallocated exception object.
606
// Weblogic sometimes mutates the detail message of exceptions
608
int offset = java_lang_Throwable::get_detailMessage_offset();
609
const TypePtr* adr_typ = ex_con->add_offset(offset);
611
Node *adr = basic_plus_adr(ex_node, ex_node, offset);
612
const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
613
Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
615
if (!method()->has_exception_handlers()) {
616
// We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway.
617
// This prevents issues with exception handling and late inlining.
622
add_exception_state(make_exception_state(ex_node));
627
// %%% Maybe add entry to OptoRuntime which directly throws the exc.?
628
// It won't be much cheaper than bailing to the interp., since we'll
629
// have to pass up all the debug-info, and the runtime will have to
630
// create the stack trace.
632
// Usual case: Bail to interpreter.
633
// Reserve the right to recompile if we haven't seen anything yet.
635
ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr;
636
Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
637
if (treat_throw_as_hot
638
&& (method()->method_data()->trap_recompiled_at(bci(), m)
639
|| C->too_many_traps(reason))) {
640
// We cannot afford to take more traps here. Suffer in the interpreter.
641
if (C->log() != nullptr)
642
C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
643
Deoptimization::trap_reason_name(reason),
644
C->trap_count(reason));
645
action = Deoptimization::Action_none;
648
// "must_throw" prunes the JVM state to include only the stack, if there
649
// are no local exception handlers. This should cut down on register
650
// allocation time and code size, by drastically reducing the number
651
// of in-edges on the call to the uncommon trap.
653
uncommon_trap(reason, action, (ciKlass*)nullptr, (char*)nullptr, must_throw);
657
//----------------------------PreserveJVMState---------------------------------
658
PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
659
debug_only(kit->verify_map());
661
_map = kit->map(); // preserve the map
663
kit->set_map(clone_map ? kit->clone_map() : nullptr);
666
Parse* parser = kit->is_Parse();
667
int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
671
PreserveJVMState::~PreserveJVMState() {
672
GraphKit* kit = _kit;
674
assert(kit->bci() == _bci, "bci must not shift");
675
Parse* parser = kit->is_Parse();
676
int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
677
assert(block == _block, "block must not shift");
684
//-----------------------------BuildCutout-------------------------------------
685
BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
686
: PreserveJVMState(kit)
688
assert(p->is_Con() || p->is_Bool(), "test must be a bool");
689
SafePointNode* outer_map = _map; // preserved map is caller's
690
SafePointNode* inner_map = kit->map();
691
IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
692
outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
693
inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
695
BuildCutout::~BuildCutout() {
696
GraphKit* kit = _kit;
697
assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
700
//---------------------------PreserveReexecuteState----------------------------
701
PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
702
assert(!kit->stopped(), "must call stopped() before");
705
_reexecute = kit->jvms()->_reexecute;
707
PreserveReexecuteState::~PreserveReexecuteState() {
708
if (_kit->stopped()) return;
709
_kit->jvms()->_reexecute = _reexecute;
713
//------------------------------clone_map--------------------------------------
714
// Implementation of PreserveJVMState
716
// Only clone_map(...) here. If this function is only used in the
717
// PreserveJVMState class we may want to get rid of this extra
718
// function eventually and do it all there.
720
SafePointNode* GraphKit::clone_map() {
721
if (map() == nullptr) return nullptr;
723
// Clone the memory edge first
724
Node* mem = MergeMemNode::make(map()->memory());
725
gvn().set_type_bottom(mem);
727
SafePointNode *clonemap = (SafePointNode*)map()->clone();
728
JVMState* jvms = this->jvms();
729
JVMState* clonejvms = jvms->clone_shallow(C);
730
clonemap->set_memory(mem);
731
clonemap->set_jvms(clonejvms);
732
clonejvms->set_map(clonemap);
733
record_for_igvn(clonemap);
734
gvn().set_type_bottom(clonemap);
738
//-----------------------------destruct_map_clone------------------------------
740
// Order of destruct is important to increase the likelyhood that memory can be re-used. We need
741
// to destruct/free/delete in the exact opposite order as clone_map().
742
void GraphKit::destruct_map_clone(SafePointNode* sfp) {
743
if (sfp == nullptr) return;
745
Node* mem = sfp->memory();
746
JVMState* jvms = sfp->jvms();
748
if (jvms != nullptr) {
752
remove_for_igvn(sfp);
753
gvn().clear_type(sfp);
754
sfp->destruct(&_gvn);
756
if (mem != nullptr) {
757
gvn().clear_type(mem);
758
mem->destruct(&_gvn);
762
//-----------------------------set_map_clone-----------------------------------
763
void GraphKit::set_map_clone(SafePointNode* m) {
766
_map->set_next_exception(nullptr);
767
debug_only(verify_map());
771
//----------------------------kill_dead_locals---------------------------------
772
// Detect any locals which are known to be dead, and force them to top.
773
void GraphKit::kill_dead_locals() {
774
// Consult the liveness information for the locals. If any
775
// of them are unused, then they can be replaced by top(). This
776
// should help register allocation time and cut down on the size
777
// of the deoptimization information.
779
// This call is made from many of the bytecode handling
780
// subroutines called from the Big Switch in do_one_bytecode.
781
// Every bytecode which might include a slow path is responsible
782
// for killing its dead locals. The more consistent we
783
// are about killing deads, the fewer useless phis will be
784
// constructed for them at various merge points.
786
// bci can be -1 (InvocationEntryBci). We return the entry
787
// liveness for the method.
789
if (method() == nullptr || method()->code_size() == 0) {
790
// We are building a graph for a call to a native method.
791
// All locals are live.
797
// Consult the liveness information for the locals. If any
798
// of them are unused, then they can be replaced by top(). This
799
// should help register allocation time and cut down on the size
800
// of the deoptimization information.
801
MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
803
int len = (int)live_locals.size();
804
assert(len <= jvms()->loc_size(), "too many live locals");
805
for (int local = 0; local < len; local++) {
806
if (!live_locals.at(local)) {
807
set_local(local, top());
813
//-------------------------dead_locals_are_killed------------------------------
814
// Return true if all dead locals are set to top in the map.
815
// Used to assert "clean" debug info at various points.
816
bool GraphKit::dead_locals_are_killed() {
817
if (method() == nullptr || method()->code_size() == 0) {
818
// No locals need to be dead, so all is as it should be.
822
// Make sure somebody called kill_dead_locals upstream.
824
for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
825
if (jvms->loc_size() == 0) continue; // no locals to consult
826
SafePointNode* map = jvms->map();
827
ciMethod* method = jvms->method();
828
int bci = jvms->bci();
829
if (jvms == this->jvms()) {
830
bci = this->bci(); // it might not yet be synched
832
MethodLivenessResult live_locals = method->liveness_at_bci(bci);
833
int len = (int)live_locals.size();
834
if (!live_locals.is_valid() || len == 0)
835
// This method is trivial, or is poisoned by a breakpoint.
837
assert(len == jvms->loc_size(), "live map consistent with locals map");
838
for (int local = 0; local < len; local++) {
839
if (!live_locals.at(local) && map->local(jvms, local) != top()) {
840
if (PrintMiscellaneous && (Verbose || WizardMode)) {
841
tty->print_cr("Zombie local %d: ", local);
853
// Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
854
static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
855
ciMethod* cur_method = jvms->method();
856
int cur_bci = jvms->bci();
857
if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
858
Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
859
return Interpreter::bytecode_should_reexecute(code) ||
860
(is_anewarray && code == Bytecodes::_multianewarray);
861
// Reexecute _multianewarray bytecode which was replaced with
862
// sequence of [a]newarray. See Parse::do_multianewarray().
864
// Note: interpreter should not have it set since this optimization
865
// is limited by dimensions and guarded by flag so in some cases
866
// multianewarray() runtime calls will be generated and
867
// the bytecode should not be reexecutes (stack will not be reset).
873
// Helper function for adding JVMState and debug information to node
874
void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
875
// Add the safepoint edges to the call (or other safepoint).
877
// Make sure dead locals are set to top. This
878
// should help register allocation time and cut down on the size
879
// of the deoptimization information.
880
assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
882
// Walk the inline list to fill in the correct set of JVMState's
883
// Also fill in the associated edges for each JVMState.
885
// If the bytecode needs to be reexecuted we need to put
886
// the arguments back on the stack.
887
const bool should_reexecute = jvms()->should_reexecute();
888
JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
890
// NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
891
// undefined if the bci is different. This is normal for Parse but it
892
// should not happen for LibraryCallKit because only one bci is processed.
893
assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
894
"in LibraryCallKit the reexecute bit should not change");
896
// If we are guaranteed to throw, we can prune everything but the
897
// input to the current bytecode.
898
bool can_prune_locals = false;
899
uint stack_slots_not_pruned = 0;
900
int inputs = 0, depth = 0;
902
assert(method() == youngest_jvms->method(), "sanity");
903
if (compute_stack_effects(inputs, depth)) {
904
can_prune_locals = true;
905
stack_slots_not_pruned = inputs;
909
if (env()->should_retain_local_variables()) {
910
// At any safepoint, this method can get breakpointed, which would
911
// then require an immediate deoptimization.
912
can_prune_locals = false; // do not prune locals
913
stack_slots_not_pruned = 0;
916
// do not scribble on the input jvms
917
JVMState* out_jvms = youngest_jvms->clone_deep(C);
918
call->set_jvms(out_jvms); // Start jvms list for call node
920
// For a known set of bytecodes, the interpreter should reexecute them if
921
// deoptimization happens. We set the reexecute state for them here
922
if (out_jvms->is_reexecute_undefined() && //don't change if already specified
923
should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
925
int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
926
assert(method() == youngest_jvms->method(), "sanity");
927
assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
928
assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
930
out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
934
DEBUG_ONLY(uint non_debug_edges = call->req());
935
call->add_req_batch(top(), youngest_jvms->debug_depth());
936
assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
938
// Set up edges so that the call looks like this:
939
// Call [state:] ctl io mem fptr retadr
940
// [parms:] parm0 ... parmN
941
// [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
942
// [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
943
// [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
944
// Note that caller debug info precedes callee debug info.
946
// Fill pointer walks backwards from "young:" to "root:" in the diagram above:
947
uint debug_ptr = call->req();
949
// Loop over the map input edges associated with jvms, add them
950
// to the call node, & reset all offsets to match call node array.
951
for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
952
uint debug_end = debug_ptr;
953
uint debug_start = debug_ptr - in_jvms->debug_size();
954
debug_ptr = debug_start; // back up the ptr
956
uint p = debug_start; // walks forward in [debug_start, debug_end)
958
SafePointNode* in_map = in_jvms->map();
959
out_jvms->set_map(call);
961
if (can_prune_locals) {
962
assert(in_jvms->method() == out_jvms->method(), "sanity");
963
// If the current throw can reach an exception handler in this JVMS,
964
// then we must keep everything live that can reach that handler.
965
// As a quick and dirty approximation, we look for any handlers at all.
966
if (in_jvms->method()->has_exception_handlers()) {
967
can_prune_locals = false;
972
k = in_jvms->locoff();
973
l = in_jvms->loc_size();
974
out_jvms->set_locoff(p);
975
if (!can_prune_locals) {
976
for (j = 0; j < l; j++)
977
call->set_req(p++, in_map->in(k+j));
979
p += l; // already set to top above by add_req_batch
982
// Add the Expression Stack
983
k = in_jvms->stkoff();
985
out_jvms->set_stkoff(p);
986
if (!can_prune_locals) {
987
for (j = 0; j < l; j++)
988
call->set_req(p++, in_map->in(k+j));
989
} else if (can_prune_locals && stack_slots_not_pruned != 0) {
990
// Divide stack into {S0,...,S1}, where S0 is set to top.
991
uint s1 = stack_slots_not_pruned;
992
stack_slots_not_pruned = 0; // for next iteration
995
p += s0; // skip the tops preinstalled by add_req_batch
996
for (j = s0; j < l; j++)
997
call->set_req(p++, in_map->in(k+j));
999
p += l; // already set to top above by add_req_batch
1003
k = in_jvms->monoff();
1004
l = in_jvms->mon_size();
1005
out_jvms->set_monoff(p);
1006
for (j = 0; j < l; j++)
1007
call->set_req(p++, in_map->in(k+j));
1009
// Copy any scalar object fields.
1010
k = in_jvms->scloff();
1011
l = in_jvms->scl_size();
1012
out_jvms->set_scloff(p);
1013
for (j = 0; j < l; j++)
1014
call->set_req(p++, in_map->in(k+j));
1016
// Finish the new jvms.
1017
out_jvms->set_endoff(p);
1019
assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1020
assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1021
assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1022
assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1023
assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1024
assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1026
// Update the two tail pointers in parallel.
1027
out_jvms = out_jvms->caller();
1028
in_jvms = in_jvms->caller();
1031
assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1033
// Test the correctness of JVMState::debug_xxx accessors:
1034
assert(call->jvms()->debug_start() == non_debug_edges, "");
1035
assert(call->jvms()->debug_end() == call->req(), "");
1036
assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1039
bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1040
Bytecodes::Code code = java_bc();
1041
if (code == Bytecodes::_wide) {
1042
code = method()->java_code_at_bci(bci() + 1);
1045
if (code != Bytecodes::_illegal) {
1046
depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1049
auto rsize = [&]() {
1050
assert(code != Bytecodes::_illegal, "code is illegal!");
1051
BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1052
return (rtype < T_CONFLICT) ? type2size[rtype] : 0;
1056
case Bytecodes::_illegal:
1059
case Bytecodes::_ldc:
1060
case Bytecodes::_ldc_w:
1061
case Bytecodes::_ldc2_w:
1065
case Bytecodes::_dup: inputs = 1; break;
1066
case Bytecodes::_dup_x1: inputs = 2; break;
1067
case Bytecodes::_dup_x2: inputs = 3; break;
1068
case Bytecodes::_dup2: inputs = 2; break;
1069
case Bytecodes::_dup2_x1: inputs = 3; break;
1070
case Bytecodes::_dup2_x2: inputs = 4; break;
1071
case Bytecodes::_swap: inputs = 2; break;
1072
case Bytecodes::_arraylength: inputs = 1; break;
1074
case Bytecodes::_getstatic:
1075
case Bytecodes::_putstatic:
1076
case Bytecodes::_getfield:
1077
case Bytecodes::_putfield:
1079
bool ignored_will_link;
1080
ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1081
int size = field->type()->size();
1082
bool is_get = (depth >= 0), is_static = (depth & 1);
1083
inputs = (is_static ? 0 : 1);
1085
depth = size - inputs;
1087
inputs += size; // putxxx pops the value from the stack
1093
case Bytecodes::_invokevirtual:
1094
case Bytecodes::_invokespecial:
1095
case Bytecodes::_invokestatic:
1096
case Bytecodes::_invokedynamic:
1097
case Bytecodes::_invokeinterface:
1099
bool ignored_will_link;
1100
ciSignature* declared_signature = nullptr;
1101
ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1102
assert(declared_signature != nullptr, "cannot be null");
1103
inputs = declared_signature->arg_size_for_bc(code);
1104
int size = declared_signature->return_type()->size();
1105
depth = size - inputs;
1109
case Bytecodes::_multianewarray:
1111
ciBytecodeStream iter(method());
1112
iter.reset_to_bci(bci());
1114
inputs = iter.get_dimensions();
1115
assert(rsize() == 1, "");
1120
case Bytecodes::_ireturn:
1121
case Bytecodes::_lreturn:
1122
case Bytecodes::_freturn:
1123
case Bytecodes::_dreturn:
1124
case Bytecodes::_areturn:
1125
assert(rsize() == -depth, "");
1129
case Bytecodes::_jsr:
1130
case Bytecodes::_jsr_w:
1132
depth = 1; // S.B. depth=1, not zero
1136
// bytecode produces a typed result
1137
inputs = rsize() - depth;
1138
assert(inputs >= 0, "");
1144
int outputs = depth + inputs;
1145
assert(outputs >= 0, "sanity");
1147
case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1148
case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1149
case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1150
case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1151
case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1161
//------------------------------basic_plus_adr---------------------------------
1162
Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1163
// short-circuit a common case
1164
if (offset == intcon(0)) return ptr;
1165
return _gvn.transform( new AddPNode(base, ptr, offset) );
1168
Node* GraphKit::ConvI2L(Node* offset) {
1169
// short-circuit a common case
1170
jint offset_con = find_int_con(offset, Type::OffsetBot);
1171
if (offset_con != Type::OffsetBot) {
1172
return longcon((jlong) offset_con);
1174
return _gvn.transform( new ConvI2LNode(offset));
1177
Node* GraphKit::ConvI2UL(Node* offset) {
1178
juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1179
if (offset_con != (juint) Type::OffsetBot) {
1180
return longcon((julong) offset_con);
1182
Node* conv = _gvn.transform( new ConvI2LNode(offset));
1183
Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1184
return _gvn.transform( new AndLNode(conv, mask) );
1187
Node* GraphKit::ConvL2I(Node* offset) {
1188
// short-circuit a common case
1189
jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1190
if (offset_con != (jlong)Type::OffsetBot) {
1191
return intcon((int) offset_con);
1193
return _gvn.transform( new ConvL2INode(offset));
1196
//-------------------------load_object_klass-----------------------------------
1197
Node* GraphKit::load_object_klass(Node* obj) {
1198
// Special-case a fresh allocation to avoid building nodes:
1199
Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1200
if (akls != nullptr) return akls;
1201
Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1202
return _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1205
//-------------------------load_array_length-----------------------------------
1206
Node* GraphKit::load_array_length(Node* array) {
1207
// Special-case a fresh allocation to avoid building nodes:
1208
AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1210
if (alloc == nullptr) {
1211
Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1212
alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1214
alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1219
Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1220
const TypeOopPtr* oop_type,
1221
bool replace_length_in_map) {
1222
Node* length = alloc->Ideal_length();
1223
if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1224
Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1225
if (ccast != length) {
1226
// do not transform ccast here, it might convert to top node for
1227
// negative array length and break assumptions in parsing stage.
1228
_gvn.set_type_bottom(ccast);
1229
record_for_igvn(ccast);
1230
if (replace_length_in_map) {
1231
replace_in_map(length, ccast);
1239
//------------------------------do_null_check----------------------------------
1240
// Helper function to do a null pointer check. Returned value is
1241
// the incoming address with null casted away. You are allowed to use the
1242
// not-null value only if you are control dependent on the test.
1244
extern uint explicit_null_checks_inserted,
1245
explicit_null_checks_elided;
1247
Node* GraphKit::null_check_common(Node* value, BasicType type,
1248
// optional arguments for variations:
1250
Node* *null_control,
1252
assert(!assert_null || null_control == nullptr, "not both at once");
1253
if (stopped()) return top();
1254
NOT_PRODUCT(explicit_null_checks_inserted++);
1256
// Construct null check
1257
Node *chk = nullptr;
1259
case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1260
case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1261
case T_ARRAY : // fall through
1262
type = T_OBJECT; // simplify further tests
1264
const Type *t = _gvn.type( value );
1266
const TypeOopPtr* tp = t->isa_oopptr();
1267
if (tp != nullptr && !tp->is_loaded()
1268
// Only for do_null_check, not any of its siblings:
1269
&& !assert_null && null_control == nullptr) {
1270
// Usually, any field access or invocation on an unloaded oop type
1271
// will simply fail to link, since the statically linked class is
1272
// likely also to be unloaded. However, in -Xcomp mode, sometimes
1273
// the static class is loaded but the sharper oop type is not.
1274
// Rather than checking for this obscure case in lots of places,
1275
// we simply observe that a null check on an unloaded class
1276
// will always be followed by a nonsense operation, so we
1277
// can just issue the uncommon trap here.
1278
// Our access to the unloaded class will only be correct
1279
// after it has been loaded and initialized, which requires
1280
// a trip through the interpreter.
1281
ciKlass* klass = tp->unloaded_klass();
1283
if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1285
uncommon_trap(Deoptimization::Reason_unloaded,
1286
Deoptimization::Action_reinterpret,
1292
// See if the type is contained in NULL_PTR.
1293
// If so, then the value is already null.
1294
if (t->higher_equal(TypePtr::NULL_PTR)) {
1295
NOT_PRODUCT(explicit_null_checks_elided++);
1296
return value; // Elided null assert quickly!
1299
// See if mixing in the null pointer changes type.
1300
// If so, then the null pointer was not allowed in the original
1301
// type. In other words, "value" was not-null.
1302
if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1303
// same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1304
NOT_PRODUCT(explicit_null_checks_elided++);
1305
return value; // Elided null check quickly!
1308
chk = new CmpPNode( value, null() );
1313
fatal("unexpected type: %s", type2name(type));
1315
assert(chk != nullptr, "sanity check");
1316
chk = _gvn.transform(chk);
1318
BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1319
BoolNode *btst = new BoolNode( chk, btest);
1320
Node *tst = _gvn.transform( btst );
1323
// if peephole optimizations occurred, a prior test existed.
1324
// If a prior test existed, maybe it dominates as we can avoid this test.
1325
if (tst != btst && type == T_OBJECT) {
1326
// At this point we want to scan up the CFG to see if we can
1327
// find an identical test (and so avoid this test altogether).
1328
Node *cfg = control();
1330
while( depth < 16 ) { // Limit search depth for speed
1331
if( cfg->Opcode() == Op_IfTrue &&
1332
cfg->in(0)->in(1) == tst ) {
1333
// Found prior test. Use "cast_not_null" to construct an identical
1334
// CastPP (and hence hash to) as already exists for the prior test.
1335
// Return that casted value.
1337
replace_in_map(value, null());
1338
return null(); // do not issue the redundant test
1340
Node *oldcontrol = control();
1342
Node *res = cast_not_null(value);
1343
set_control(oldcontrol);
1344
NOT_PRODUCT(explicit_null_checks_elided++);
1347
cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1348
if (cfg == nullptr) break; // Quit at region nodes
1354
// Branch to failure if null
1355
float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1356
Deoptimization::DeoptReason reason;
1358
reason = Deoptimization::reason_null_assert(speculative);
1359
} else if (type == T_OBJECT) {
1360
reason = Deoptimization::reason_null_check(speculative);
1362
reason = Deoptimization::Reason_div0_check;
1364
// %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1365
// ciMethodData::has_trap_at will return a conservative -1 if any
1366
// must-be-null assertion has failed. This could cause performance
1367
// problems for a method after its first do_null_assert failure.
1368
// Consider using 'Reason_class_check' instead?
1370
// To cause an implicit null check, we set the not-null probability
1371
// to the maximum (PROB_MAX). For an explicit check the probability
1372
// is set to a smaller value.
1373
if (null_control != nullptr || too_many_traps(reason)) {
1374
// probability is less likely
1375
ok_prob = PROB_LIKELY_MAG(3);
1376
} else if (!assert_null &&
1377
(ImplicitNullCheckThreshold > 0) &&
1378
method() != nullptr &&
1379
(method()->method_data()->trap_count(reason)
1380
>= (uint)ImplicitNullCheckThreshold)) {
1381
ok_prob = PROB_LIKELY_MAG(3);
1384
if (null_control != nullptr) {
1385
IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1386
Node* null_true = _gvn.transform( new IfFalseNode(iff));
1387
set_control( _gvn.transform( new IfTrueNode(iff)));
1389
if (null_true == top()) {
1390
explicit_null_checks_elided++;
1393
(*null_control) = null_true;
1395
BuildCutout unless(this, tst, ok_prob);
1396
// Check for optimizer eliding test at parse time
1398
// Failure not possible; do not bother making uncommon trap.
1399
NOT_PRODUCT(explicit_null_checks_elided++);
1400
} else if (assert_null) {
1401
uncommon_trap(reason,
1402
Deoptimization::Action_make_not_entrant,
1403
nullptr, "assert_null");
1405
replace_in_map(value, zerocon(type));
1406
builtin_throw(reason);
1410
// Must throw exception, fall-thru not possible?
1412
return top(); // No result
1416
// Cast obj to null on this path.
1417
replace_in_map(value, zerocon(type));
1418
return zerocon(type);
1421
// Cast obj to not-null on this path, if there is no null_control.
1422
// (If there is a null_control, a non-null value may come back to haunt us.)
1423
if (type == T_OBJECT) {
1424
Node* cast = cast_not_null(value, false);
1425
if (null_control == nullptr || (*null_control) == top())
1426
replace_in_map(value, cast);
1434
//------------------------------cast_not_null----------------------------------
1435
// Cast obj to not-null on this path
1436
Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1437
const Type *t = _gvn.type(obj);
1438
const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1439
// Object is already not-null?
1440
if( t == t_not_null ) return obj;
1442
Node* cast = new CastPPNode(control(), obj,t_not_null);
1443
cast = _gvn.transform( cast );
1445
// Scan for instances of 'obj' in the current JVM mapping.
1446
// These instances are known to be not-null after the test.
1447
if (do_replace_in_map)
1448
replace_in_map(obj, cast);
1450
return cast; // Return casted value
1453
// Sometimes in intrinsics, we implicitly know an object is not null
1454
// (there's no actual null check) so we can cast it to not null. In
1455
// the course of optimizations, the input to the cast can become null.
1456
// In that case that data path will die and we need the control path
1457
// to become dead as well to keep the graph consistent. So we have to
1458
// add a check for null for which one branch can't be taken. It uses
1459
// an Opaque4 node that will cause the check to be removed after loop
1460
// opts so the test goes away and the compiled code doesn't execute a
1462
Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1463
if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1466
Node* chk = _gvn.transform(new CmpPNode(value, null()));
1467
Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1468
Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1469
IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1470
_gvn.set_type(iff, iff->Value(&_gvn));
1471
if (!tst->is_Con()) {
1472
record_for_igvn(iff);
1474
Node *if_f = _gvn.transform(new IfFalseNode(iff));
1475
Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1476
Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1477
C->root()->add_req(halt);
1478
Node *if_t = _gvn.transform(new IfTrueNode(iff));
1480
return cast_not_null(value, do_replace_in_map);
1484
//--------------------------replace_in_map-------------------------------------
1485
void GraphKit::replace_in_map(Node* old, Node* neww) {
1490
map()->replace_edge(old, neww);
1492
// Note: This operation potentially replaces any edge
1493
// on the map. This includes locals, stack, and monitors
1494
// of the current (innermost) JVM state.
1496
// don't let inconsistent types from profiling escape this
1499
const Type* told = _gvn.type(old);
1500
const Type* tnew = _gvn.type(neww);
1502
if (!tnew->higher_equal(told)) {
1506
map()->record_replaced_node(old, neww);
1510
//=============================================================================
1511
//--------------------------------memory---------------------------------------
1512
Node* GraphKit::memory(uint alias_idx) {
1513
MergeMemNode* mem = merged_memory();
1514
Node* p = mem->memory_at(alias_idx);
1515
assert(p != mem->empty_memory(), "empty");
1516
_gvn.set_type(p, Type::MEMORY); // must be mapped
1520
//-----------------------------reset_memory------------------------------------
1521
Node* GraphKit::reset_memory() {
1522
Node* mem = map()->memory();
1523
// do not use this node for any more parsing!
1524
debug_only( map()->set_memory((Node*)nullptr) );
1525
return _gvn.transform( mem );
1528
//------------------------------set_all_memory---------------------------------
1529
void GraphKit::set_all_memory(Node* newmem) {
1530
Node* mergemem = MergeMemNode::make(newmem);
1531
gvn().set_type_bottom(mergemem);
1532
map()->set_memory(mergemem);
1535
//------------------------------set_all_memory_call----------------------------
1536
void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1537
Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1538
set_all_memory(newmem);
1541
//=============================================================================
1543
// parser factory methods for MemNodes
1545
// These are layered on top of the factory methods in LoadNode and StoreNode,
1546
// and integrate with the parser's memory state and _gvn engine.
1549
// factory methods in "int adr_idx"
1550
Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1553
LoadNode::ControlDependency control_dependency,
1554
bool require_atomic_access,
1558
uint8_t barrier_data) {
1559
assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1560
const TypePtr* adr_type = nullptr; // debug-mode-only argument
1561
debug_only(adr_type = C->get_adr_type(adr_idx));
1562
Node* mem = memory(adr_idx);
1563
Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1564
ld = _gvn.transform(ld);
1565
if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1566
// Improve graph before escape analysis and boxing elimination.
1567
record_for_igvn(ld);
1568
if (ld->is_DecodeN()) {
1569
// Also record the actual load (LoadN) in case ld is DecodeN. In some
1570
// rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1571
// a Phi). Recording such cases is still perfectly sound, but may be
1572
// unnecessary and result in some minor IGVN overhead.
1573
record_for_igvn(ld->in(1));
1579
Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1582
bool require_atomic_access,
1587
assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1588
const TypePtr* adr_type = nullptr;
1589
debug_only(adr_type = C->get_adr_type(adr_idx));
1590
Node *mem = memory(adr_idx);
1591
Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1593
st->as_Store()->set_unaligned_access();
1596
st->as_Store()->set_mismatched_access();
1599
st->as_Store()->set_unsafe_access();
1601
st->as_Store()->set_barrier_data(barrier_data);
1602
st = _gvn.transform(st);
1603
set_memory(st, adr_idx);
1604
// Back-to-back stores can only remove intermediate store with DU info
1605
// so push on worklist for optimizer.
1606
if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1607
record_for_igvn(st);
1612
Node* GraphKit::access_store_at(Node* obj,
1614
const TypePtr* adr_type,
1616
const Type* val_type,
1618
DecoratorSet decorators) {
1619
// Transformation of a value which could be null pointer (CastPP #null)
1620
// could be delayed during Parse (for example, in adjust_map_after_if()).
1621
// Execute transformation here to avoid barrier generation in such case.
1622
if (_gvn.type(val) == TypePtr::NULL_PTR) {
1623
val = _gvn.makecon(TypePtr::NULL_PTR);
1627
return top(); // Dead path ?
1630
assert(val != nullptr, "not dead path");
1632
C2AccessValuePtr addr(adr, adr_type);
1633
C2AccessValue value(val, val_type);
1634
C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1635
if (access.is_raw()) {
1636
return _barrier_set->BarrierSetC2::store_at(access, value);
1638
return _barrier_set->store_at(access, value);
1642
Node* GraphKit::access_load_at(Node* obj, // containing obj
1643
Node* adr, // actual address to store val at
1644
const TypePtr* adr_type,
1645
const Type* val_type,
1647
DecoratorSet decorators) {
1649
return top(); // Dead path ?
1652
C2AccessValuePtr addr(adr, adr_type);
1653
C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1654
if (access.is_raw()) {
1655
return _barrier_set->BarrierSetC2::load_at(access, val_type);
1657
return _barrier_set->load_at(access, val_type);
1661
Node* GraphKit::access_load(Node* adr, // actual address to load val at
1662
const Type* val_type,
1664
DecoratorSet decorators) {
1666
return top(); // Dead path ?
1669
C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1670
C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1671
if (access.is_raw()) {
1672
return _barrier_set->BarrierSetC2::load_at(access, val_type);
1674
return _barrier_set->load_at(access, val_type);
1678
Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1680
const TypePtr* adr_type,
1684
const Type* value_type,
1686
DecoratorSet decorators) {
1687
C2AccessValuePtr addr(adr, adr_type);
1688
C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1689
bt, obj, addr, alias_idx);
1690
if (access.is_raw()) {
1691
return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1693
return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1697
Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1699
const TypePtr* adr_type,
1703
const Type* value_type,
1705
DecoratorSet decorators) {
1706
C2AccessValuePtr addr(adr, adr_type);
1707
C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1708
bt, obj, addr, alias_idx);
1709
if (access.is_raw()) {
1710
return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1712
return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1716
Node* GraphKit::access_atomic_xchg_at(Node* obj,
1718
const TypePtr* adr_type,
1721
const Type* value_type,
1723
DecoratorSet decorators) {
1724
C2AccessValuePtr addr(adr, adr_type);
1725
C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1726
bt, obj, addr, alias_idx);
1727
if (access.is_raw()) {
1728
return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1730
return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1734
Node* GraphKit::access_atomic_add_at(Node* obj,
1736
const TypePtr* adr_type,
1739
const Type* value_type,
1741
DecoratorSet decorators) {
1742
C2AccessValuePtr addr(adr, adr_type);
1743
C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1744
if (access.is_raw()) {
1745
return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1747
return _barrier_set->atomic_add_at(access, new_val, value_type);
1751
void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1752
return _barrier_set->clone(this, src, dst, size, is_array);
1755
//-------------------------array_element_address-------------------------
1756
Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1757
const TypeInt* sizetype, Node* ctrl) {
1758
uint shift = exact_log2(type2aelembytes(elembt));
1759
uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1761
// short-circuit a common case (saves lots of confusing waste motion)
1762
jint idx_con = find_int_con(idx, -1);
1764
intptr_t offset = header + ((intptr_t)idx_con << shift);
1765
return basic_plus_adr(ary, offset);
1768
// must be correct type for alignment purposes
1769
Node* base = basic_plus_adr(ary, header);
1770
idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1771
Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1772
return basic_plus_adr(ary, base, scale);
1775
//-------------------------load_array_element-------------------------
1776
Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1777
const Type* elemtype = arytype->elem();
1778
BasicType elembt = elemtype->array_element_basic_type();
1779
Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1780
if (elembt == T_NARROWOOP) {
1781
elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1783
Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1784
IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1788
//-------------------------set_arguments_for_java_call-------------------------
1789
// Arguments (pre-popped from the stack) are taken from the JVMS.
1790
void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1791
// Add the call arguments:
1792
uint nargs = call->method()->arg_size();
1793
for (uint i = 0; i < nargs; i++) {
1794
Node* arg = argument(i);
1795
call->init_req(i + TypeFunc::Parms, arg);
1799
//---------------------------set_edges_for_java_call---------------------------
1800
// Connect a newly created call into the current JVMS.
1801
// A return value node (if any) is returned from set_edges_for_java_call.
1802
void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1804
// Add the predefined inputs:
1805
call->init_req( TypeFunc::Control, control() );
1806
call->init_req( TypeFunc::I_O , i_o() );
1807
call->init_req( TypeFunc::Memory , reset_memory() );
1808
call->init_req( TypeFunc::FramePtr, frameptr() );
1809
call->init_req( TypeFunc::ReturnAdr, top() );
1811
add_safepoint_edges(call, must_throw);
1813
Node* xcall = _gvn.transform(call);
1815
if (xcall == top()) {
1819
assert(xcall == call, "call identity is stable");
1821
// Re-use the current map to produce the result.
1823
set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1824
set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1825
set_all_memory_call(xcall, separate_io_proj);
1827
//return xcall; // no need, caller already has it
1830
Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1831
if (stopped()) return top(); // maybe the call folded up?
1833
// Capture the return value, if any.
1835
if (call->method() == nullptr ||
1836
call->method()->return_type()->basic_type() == T_VOID)
1838
else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1840
// Note: Since any out-of-line call can produce an exception,
1841
// we always insert an I_O projection from the call into the result.
1843
make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1845
if (separate_io_proj) {
1846
// The caller requested separate projections be used by the fall
1847
// through and exceptional paths, so replace the projections for
1848
// the fall through path.
1849
set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1850
set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1855
//--------------------set_predefined_input_for_runtime_call--------------------
1856
// Reading and setting the memory state is way conservative here.
1857
// The real problem is that I am not doing real Type analysis on memory,
1858
// so I cannot distinguish card mark stores from other stores. Across a GC
1859
// point the Store Barrier and the card mark memory has to agree. I cannot
1860
// have a card mark store and its barrier split across the GC point from
1861
// either above or below. Here I get that to happen by reading ALL of memory.
1862
// A better answer would be to separate out card marks from other memory.
1863
// For now, return the input memory state, so that it can be reused
1864
// after the call, if this call has restricted memory effects.
1865
Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1866
// Set fixed predefined input arguments
1867
Node* memory = reset_memory();
1868
Node* m = narrow_mem == nullptr ? memory : narrow_mem;
1869
call->init_req( TypeFunc::Control, control() );
1870
call->init_req( TypeFunc::I_O, top() ); // does no i/o
1871
call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1872
call->init_req( TypeFunc::FramePtr, frameptr() );
1873
call->init_req( TypeFunc::ReturnAdr, top() );
1877
//-------------------set_predefined_output_for_runtime_call--------------------
1878
// Set control and memory (not i_o) from the call.
1879
// If keep_mem is not null, use it for the output state,
1880
// except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1881
// If hook_mem is null, this call produces no memory effects at all.
1882
// If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1883
// then only that memory slice is taken from the call.
1884
// In the last case, we must put an appropriate memory barrier before
1885
// the call, so as to create the correct anti-dependencies on loads
1886
// preceding the call.
1887
void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1889
const TypePtr* hook_mem) {
1891
set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1893
// First clone the existing memory state
1894
set_all_memory(keep_mem);
1895
if (hook_mem != nullptr) {
1896
// Make memory for the call
1897
Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1898
// Set the RawPtr memory state only. This covers all the heap top/GC stuff
1899
// We also use hook_mem to extract specific effects from arraycopy stubs.
1900
set_memory(mem, hook_mem);
1902
// ...else the call has NO memory effects.
1904
// Make sure the call advertises its memory effects precisely.
1905
// This lets us build accurate anti-dependences in gcm.cpp.
1906
assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1907
"call node must be constructed correctly");
1909
assert(hook_mem == nullptr, "");
1910
// This is not a "slow path" call; all memory comes from the call.
1911
set_all_memory_call(call);
1915
// Keep track of MergeMems feeding into other MergeMems
1916
static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
1917
if (!mem->is_MergeMem()) {
1920
for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
1921
Node* use = i.get();
1922
if (use->is_MergeMem()) {
1928
// Replace the call with the current state of the kit.
1929
void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1930
JVMState* ejvms = nullptr;
1931
if (has_exceptions()) {
1932
ejvms = transfer_exceptions_into_jvms();
1935
ReplacedNodes replaced_nodes = map()->replaced_nodes();
1936
ReplacedNodes replaced_nodes_exception;
1937
Node* ex_ctl = top();
1939
SafePointNode* final_state = stop();
1941
// Find all the needed outputs of this call
1942
CallProjections callprojs;
1943
call->extract_projections(&callprojs, true);
1945
Unique_Node_List wl;
1946
Node* init_mem = call->in(TypeFunc::Memory);
1947
Node* final_mem = final_state->in(TypeFunc::Memory);
1948
Node* final_ctl = final_state->in(TypeFunc::Control);
1949
Node* final_io = final_state->in(TypeFunc::I_O);
1951
// Replace all the old call edges with the edges from the inlining result
1952
if (callprojs.fallthrough_catchproj != nullptr) {
1953
C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1955
if (callprojs.fallthrough_memproj != nullptr) {
1956
if (final_mem->is_MergeMem()) {
1957
// Parser's exits MergeMem was not transformed but may be optimized
1958
final_mem = _gvn.transform(final_mem);
1960
C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1961
add_mergemem_users_to_worklist(wl, final_mem);
1963
if (callprojs.fallthrough_ioproj != nullptr) {
1964
C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1967
// Replace the result with the new result if it exists and is used
1968
if (callprojs.resproj != nullptr && result != nullptr) {
1969
C->gvn_replace_by(callprojs.resproj, result);
1972
if (ejvms == nullptr) {
1973
// No exception edges to simply kill off those paths
1974
if (callprojs.catchall_catchproj != nullptr) {
1975
C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1977
if (callprojs.catchall_memproj != nullptr) {
1978
C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1980
if (callprojs.catchall_ioproj != nullptr) {
1981
C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1983
// Replace the old exception object with top
1984
if (callprojs.exobj != nullptr) {
1985
C->gvn_replace_by(callprojs.exobj, C->top());
1988
GraphKit ekit(ejvms);
1990
// Load my combined exception state into the kit, with all phis transformed:
1991
SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1992
replaced_nodes_exception = ex_map->replaced_nodes();
1994
Node* ex_oop = ekit.use_exception_state(ex_map);
1996
if (callprojs.catchall_catchproj != nullptr) {
1997
C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1998
ex_ctl = ekit.control();
2000
if (callprojs.catchall_memproj != nullptr) {
2001
Node* ex_mem = ekit.reset_memory();
2002
C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
2003
add_mergemem_users_to_worklist(wl, ex_mem);
2005
if (callprojs.catchall_ioproj != nullptr) {
2006
C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
2009
// Replace the old exception object with the newly created one
2010
if (callprojs.exobj != nullptr) {
2011
C->gvn_replace_by(callprojs.exobj, ex_oop);
2015
// Disconnect the call from the graph
2016
call->disconnect_inputs(C);
2017
C->gvn_replace_by(call, C->top());
2019
// Clean up any MergeMems that feed other MergeMems since the
2020
// optimizer doesn't like that.
2021
while (wl.size() > 0) {
2022
_gvn.transform(wl.pop());
2025
if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2026
replaced_nodes.apply(C, final_ctl);
2028
if (!ex_ctl->is_top() && do_replaced_nodes) {
2029
replaced_nodes_exception.apply(C, ex_ctl);
2034
//------------------------------increment_counter------------------------------
2035
// for statistics: increment a VM counter by 1
2037
void GraphKit::increment_counter(address counter_addr) {
2038
Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2039
increment_counter(adr1);
2042
void GraphKit::increment_counter(Node* counter_addr) {
2043
int adr_type = Compile::AliasIdxRaw;
2044
Node* ctrl = control();
2045
Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered);
2046
Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2047
store_to_memory(ctrl, counter_addr, incr, T_LONG, adr_type, MemNode::unordered);
2051
//------------------------------uncommon_trap----------------------------------
2052
// Bail out to the interpreter in mid-method. Implemented by calling the
2053
// uncommon_trap blob. This helper function inserts a runtime call with the
2055
Node* GraphKit::uncommon_trap(int trap_request,
2056
ciKlass* klass, const char* comment,
2058
bool keep_exact_action) {
2059
if (failing()) stop();
2060
if (stopped()) return nullptr; // trap reachable?
2062
// Note: If ProfileTraps is true, and if a deopt. actually
2063
// occurs here, the runtime will make sure an MDO exists. There is
2064
// no need to call method()->ensure_method_data() at this point.
2066
// Set the stack pointer to the right value for reexecution:
2067
set_sp(reexecute_sp());
2071
// Make sure the stack has at least enough depth to execute
2072
// the current bytecode.
2073
int inputs, ignored_depth;
2074
if (compute_stack_effects(inputs, ignored_depth)) {
2075
assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2076
Bytecodes::name(java_bc()), sp(), inputs);
2081
Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2082
Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2085
case Deoptimization::Action_maybe_recompile:
2086
case Deoptimization::Action_reinterpret:
2087
// Temporary fix for 6529811 to allow virtual calls to be sure they
2088
// get the chance to go from mono->bi->mega
2089
if (!keep_exact_action &&
2090
Deoptimization::trap_request_index(trap_request) < 0 &&
2091
too_many_recompiles(reason)) {
2092
// This BCI is causing too many recompilations.
2093
if (C->log() != nullptr) {
2094
C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2095
Deoptimization::trap_reason_name(reason),
2096
Deoptimization::trap_action_name(action));
2098
action = Deoptimization::Action_none;
2099
trap_request = Deoptimization::make_trap_request(reason, action);
2101
C->set_trap_can_recompile(true);
2104
case Deoptimization::Action_make_not_entrant:
2105
C->set_trap_can_recompile(true);
2107
case Deoptimization::Action_none:
2108
case Deoptimization::Action_make_not_compilable:
2112
fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2117
if (TraceOptoParse) {
2119
tty->print_cr("Uncommon trap %s at bci:%d",
2120
Deoptimization::format_trap_request(buf, sizeof(buf),
2121
trap_request), bci());
2124
CompileLog* log = C->log();
2125
if (log != nullptr) {
2126
int kid = (klass == nullptr)? -1: log->identify(klass);
2127
log->begin_elem("uncommon_trap bci='%d'", bci());
2129
log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2131
if (kid >= 0) log->print(" klass='%d'", kid);
2132
if (comment != nullptr) log->print(" comment='%s'", comment);
2136
// Make sure any guarding test views this path as very unlikely
2137
Node *i0 = control()->in(0);
2138
if (i0 != nullptr && i0->is_If()) { // Found a guarding if test?
2139
IfNode *iff = i0->as_If();
2140
float f = iff->_prob; // Get prob
2141
if (control()->Opcode() == Op_IfTrue) {
2142
if (f > PROB_UNLIKELY_MAG(4))
2143
iff->_prob = PROB_MIN;
2145
if (f < PROB_LIKELY_MAG(4))
2146
iff->_prob = PROB_MAX;
2150
// Clear out dead values from the debug info.
2153
// Now insert the uncommon trap subroutine call
2154
address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2155
const TypePtr* no_memory_effects = nullptr;
2156
// Pass the index of the class to be loaded
2157
Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2158
(must_throw ? RC_MUST_THROW : 0),
2159
OptoRuntime::uncommon_trap_Type(),
2160
call_addr, "uncommon_trap", no_memory_effects,
2161
intcon(trap_request));
2162
assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2163
"must extract request correctly from the graph");
2164
assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2166
call->set_req(TypeFunc::ReturnAdr, returnadr());
2167
// The debug info is the only real input to this call.
2169
// Halt-and-catch fire here. The above call should never return!
2170
HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2171
PRODUCT_ONLY(COMMA /*reachable*/false));
2172
_gvn.set_type_bottom(halt);
2173
root()->add_req(halt);
2175
stop_and_kill_map();
2180
//--------------------------just_allocated_object------------------------------
2181
// Report the object that was just allocated.
2182
// It must be the case that there are no intervening safepoints.
2183
// We use this to determine if an object is so "fresh" that
2184
// it does not require card marks.
2185
Node* GraphKit::just_allocated_object(Node* current_control) {
2186
Node* ctrl = current_control;
2187
// Object::<init> is invoked after allocation, most of invoke nodes
2188
// will be reduced, but a region node is kept in parse time, we check
2189
// the pattern and skip the region node if it degraded to a copy.
2190
if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 &&
2191
ctrl->as_Region()->is_copy()) {
2192
ctrl = ctrl->as_Region()->is_copy();
2194
if (C->recent_alloc_ctl() == ctrl) {
2195
return C->recent_alloc_obj();
2202
* Record profiling data exact_kls for Node n with the type system so
2203
* that it can propagate it (speculation)
2205
* @param n node that the type applies to
2206
* @param exact_kls type from profiling
2207
* @param maybe_null did profiling see null?
2209
* @return node with improved type
2211
Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2212
const Type* current_type = _gvn.type(n);
2213
assert(UseTypeSpeculation, "type speculation must be on");
2215
const TypePtr* speculative = current_type->speculative();
2217
// Should the klass from the profile be recorded in the speculative type?
2218
if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2219
const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2220
const TypeOopPtr* xtype = tklass->as_instance_type();
2221
assert(xtype->klass_is_exact(), "Should be exact");
2222
// Any reason to believe n is not null (from this profiling or a previous one)?
2223
assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2224
const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2225
// record the new speculative type's depth
2226
speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2227
speculative = speculative->with_inline_depth(jvms()->depth());
2228
} else if (current_type->would_improve_ptr(ptr_kind)) {
2229
// Profiling report that null was never seen so we can change the
2230
// speculative type to non null ptr.
2231
if (ptr_kind == ProfileAlwaysNull) {
2232
speculative = TypePtr::NULL_PTR;
2234
assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2235
const TypePtr* ptr = TypePtr::NOTNULL;
2236
if (speculative != nullptr) {
2237
speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2244
if (speculative != current_type->speculative()) {
2245
// Build a type with a speculative type (what we think we know
2246
// about the type but will need a guard when we use it)
2247
const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2248
// We're changing the type, we need a new CheckCast node to carry
2249
// the new type. The new type depends on the control: what
2250
// profiling tells us is only valid from here as far as we can
2252
Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2253
cast = _gvn.transform(cast);
2254
replace_in_map(n, cast);
2262
* Record profiling data from receiver profiling at an invoke with the
2263
* type system so that it can propagate it (speculation)
2265
* @param n receiver node
2267
* @return node with improved type
2269
Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2270
if (!UseTypeSpeculation) {
2273
ciKlass* exact_kls = profile_has_unique_klass();
2274
ProfilePtrKind ptr_kind = ProfileMaybeNull;
2275
if ((java_bc() == Bytecodes::_checkcast ||
2276
java_bc() == Bytecodes::_instanceof ||
2277
java_bc() == Bytecodes::_aastore) &&
2278
method()->method_data()->is_mature()) {
2279
ciProfileData* data = method()->method_data()->bci_to_data(bci());
2280
if (data != nullptr) {
2281
if (!data->as_BitData()->null_seen()) {
2282
ptr_kind = ProfileNeverNull;
2284
assert(data->is_ReceiverTypeData(), "bad profile data type");
2285
ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2287
for (; i < call->row_limit(); i++) {
2288
ciKlass* receiver = call->receiver(i);
2289
if (receiver != nullptr) {
2293
ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2297
return record_profile_for_speculation(n, exact_kls, ptr_kind);
2301
* Record profiling data from argument profiling at an invoke with the
2302
* type system so that it can propagate it (speculation)
2304
* @param dest_method target method for the call
2305
* @param bc what invoke bytecode is this?
2307
void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2308
if (!UseTypeSpeculation) {
2311
const TypeFunc* tf = TypeFunc::make(dest_method);
2312
int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2313
int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2314
for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2315
const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2316
if (is_reference_type(targ->basic_type())) {
2317
ProfilePtrKind ptr_kind = ProfileMaybeNull;
2318
ciKlass* better_type = nullptr;
2319
if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2320
record_profile_for_speculation(argument(j), better_type, ptr_kind);
2328
* Record profiling data from parameter profiling at an invoke with
2329
* the type system so that it can propagate it (speculation)
2331
void GraphKit::record_profiled_parameters_for_speculation() {
2332
if (!UseTypeSpeculation) {
2335
for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2336
if (_gvn.type(local(i))->isa_oopptr()) {
2337
ProfilePtrKind ptr_kind = ProfileMaybeNull;
2338
ciKlass* better_type = nullptr;
2339
if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2340
record_profile_for_speculation(local(i), better_type, ptr_kind);
2348
* Record profiling data from return value profiling at an invoke with
2349
* the type system so that it can propagate it (speculation)
2351
void GraphKit::record_profiled_return_for_speculation() {
2352
if (!UseTypeSpeculation) {
2355
ProfilePtrKind ptr_kind = ProfileMaybeNull;
2356
ciKlass* better_type = nullptr;
2357
if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2358
// If profiling reports a single type for the return value,
2359
// feed it to the type system so it can propagate it as a
2361
record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2365
void GraphKit::round_double_arguments(ciMethod* dest_method) {
2366
if (Matcher::strict_fp_requires_explicit_rounding) {
2367
// (Note: TypeFunc::make has a cache that makes this fast.)
2368
const TypeFunc* tf = TypeFunc::make(dest_method);
2369
int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2370
for (int j = 0; j < nargs; j++) {
2371
const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2372
if (targ->basic_type() == T_DOUBLE) {
2373
// If any parameters are doubles, they must be rounded before
2374
// the call, dprecision_rounding does gvn.transform
2375
Node *arg = argument(j);
2376
arg = dprecision_rounding(arg);
2377
set_argument(j, arg);
2383
// rounding for strict float precision conformance
2384
Node* GraphKit::precision_rounding(Node* n) {
2385
if (Matcher::strict_fp_requires_explicit_rounding) {
2388
return _gvn.transform(new RoundFloatNode(nullptr, n));
2397
// rounding for strict double precision conformance
2398
Node* GraphKit::dprecision_rounding(Node *n) {
2399
if (Matcher::strict_fp_requires_explicit_rounding) {
2402
return _gvn.transform(new RoundDoubleNode(nullptr, n));
2411
//=============================================================================
2412
// Generate a fast path/slow path idiom. Graph looks like:
2413
// [foo] indicates that 'foo' is a parameter
2426
// gvn to opt_test | |
2430
// [slow_call] \[fast_result]
2437
// ... \ [slow_res] | | \ [null_result]
2440
// --------Region Phi
2442
//=============================================================================
2443
// Code is structured as a series of driver functions all called 'do_XXX' that
2444
// call a set of helper functions. Helper functions first, then drivers.
2446
//------------------------------null_check_oop---------------------------------
2447
// Null check oop. Set null-path control into Region in slot 3.
2448
// Make a cast-not-nullness use the other not-null control. Return cast.
2449
Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2450
bool never_see_null,
2451
bool safe_for_replace,
2453
// Initial null check taken path
2454
(*null_control) = top();
2455
Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2457
// Generate uncommon_trap:
2458
if (never_see_null && (*null_control) != top()) {
2459
// If we see an unexpected null at a check-cast we record it and force a
2460
// recompile; the offending check-cast will be compiled to handle nulls.
2461
// If we see more than one offending BCI, then all checkcasts in the
2462
// method will be compiled to handle nulls.
2463
PreserveJVMState pjvms(this);
2464
set_control(*null_control);
2465
replace_in_map(value, null());
2466
Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2467
uncommon_trap(reason,
2468
Deoptimization::Action_make_not_entrant);
2469
(*null_control) = top(); // null path is dead
2471
if ((*null_control) == top() && safe_for_replace) {
2472
replace_in_map(value, cast);
2475
// Cast away null-ness on the result
2479
//------------------------------opt_iff----------------------------------------
2480
// Optimize the fast-check IfNode. Set the fast-path region slot 2.
2481
// Return slow-path control.
2482
Node* GraphKit::opt_iff(Node* region, Node* iff) {
2483
IfNode *opt_iff = _gvn.transform(iff)->as_If();
2485
// Fast path taken; set region slot 2
2486
Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2487
region->init_req(2,fast_taken); // Capture fast-control
2489
// Fast path not-taken, i.e. slow path
2490
Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2494
//-----------------------------make_runtime_call-------------------------------
2495
Node* GraphKit::make_runtime_call(int flags,
2496
const TypeFunc* call_type, address call_addr,
2497
const char* call_name,
2498
const TypePtr* adr_type,
2499
// The following parms are all optional.
2500
// The first null ends the list.
2501
Node* parm0, Node* parm1,
2502
Node* parm2, Node* parm3,
2503
Node* parm4, Node* parm5,
2504
Node* parm6, Node* parm7) {
2505
assert(call_addr != nullptr, "must not call null targets");
2508
bool is_leaf = !(flags & RC_NO_LEAF);
2509
bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2510
if (call_name == nullptr) {
2511
assert(!is_leaf, "must supply name for leaf");
2512
call_name = OptoRuntime::stub_name(call_addr);
2516
call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2517
} else if (flags & RC_NO_FP) {
2518
call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2519
} else if (flags & RC_VECTOR){
2520
uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2521
call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2523
call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2526
// The following is similar to set_edges_for_java_call,
2527
// except that the memory effects of the call are restricted to AliasIdxRaw.
2529
// Slow path call has no side-effects, uses few values
2530
bool wide_in = !(flags & RC_NARROW_MEM);
2531
bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2533
Node* prev_mem = nullptr;
2535
prev_mem = set_predefined_input_for_runtime_call(call);
2537
assert(!wide_out, "narrow in => narrow out");
2538
Node* narrow_mem = memory(adr_type);
2539
prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2542
// Hook each parm in order. Stop looking at the first null.
2543
if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2544
if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2545
if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2546
if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2547
if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2548
if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2549
if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2550
if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2551
/* close each nested if ===> */ } } } } } } } }
2552
assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2555
// Non-leaves can block and take safepoints:
2556
add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2558
// Non-leaves can throw exceptions:
2560
call->set_req(TypeFunc::I_O, i_o());
2563
if (flags & RC_UNCOMMON) {
2564
// Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2565
// (An "if" probability corresponds roughly to an unconditional count.
2567
call->set_cnt(PROB_UNLIKELY_MAG(4));
2570
Node* c = _gvn.transform(call);
2571
assert(c == call, "cannot disappear");
2574
// Slow path call has full side-effects.
2575
set_predefined_output_for_runtime_call(call);
2577
// Slow path call has few side-effects, and/or sets few values.
2578
set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2582
set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2589
Node* GraphKit::sign_extend_byte(Node* in) {
2590
Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2591
return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2595
Node* GraphKit::sign_extend_short(Node* in) {
2596
Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2597
return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2600
//------------------------------merge_memory-----------------------------------
2601
// Merge memory from one path into the current memory state.
2602
void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2603
for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2604
Node* old_slice = mms.force_memory();
2605
Node* new_slice = mms.memory2();
2606
if (old_slice != new_slice) {
2608
if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2609
if (mms.is_empty()) {
2610
// clone base memory Phi's inputs for this memory slice
2611
assert(old_slice == mms.base_memory(), "sanity");
2612
phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2613
_gvn.set_type(phi, Type::MEMORY);
2614
for (uint i = 1; i < phi->req(); i++) {
2615
phi->init_req(i, old_slice->in(i));
2618
phi = old_slice->as_Phi(); // Phi was generated already
2621
phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2622
_gvn.set_type(phi, Type::MEMORY);
2624
phi->set_req(new_path, new_slice);
2625
mms.set_memory(phi);
2630
//------------------------------make_slow_call_ex------------------------------
2631
// Make the exception handler hookups for the slow call
2632
void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2633
if (stopped()) return;
2635
// Make a catch node with just two handlers: fall-through and catch-all
2636
Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2637
Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2638
Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2639
_gvn.set_type_bottom(norm);
2640
C->record_for_igvn(norm);
2641
Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2643
{ PreserveJVMState pjvms(this);
2647
if (excp != top()) {
2649
// Deoptimize if an exception is caught. Don't construct exception state in this case.
2650
uncommon_trap(Deoptimization::Reason_unhandled,
2651
Deoptimization::Action_none);
2653
// Create an exception state also.
2654
// Use an exact type if the caller has a specific exception.
2655
const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2656
Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2657
add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2662
// Get the no-exception control from the CatchNode.
2666
static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2667
Node* cmp = nullptr;
2669
case T_INT: cmp = new CmpINode(in1, in2); break;
2670
case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2671
default: fatal("unexpected comparison type %s", type2name(bt));
2673
cmp = gvn.transform(cmp);
2674
Node* bol = gvn.transform(new BoolNode(cmp, test));
2675
IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2677
if (!bol->is_Con()) gvn.record_for_igvn(iff);
2681
//-------------------------------gen_subtype_check-----------------------------
2682
// Generate a subtyping check. Takes as input the subtype and supertype.
2683
// Returns 2 values: sets the default control() to the true path and returns
2684
// the false path. Only reads invariant memory; sets no (visible) memory.
2685
// The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2686
// but that's not exposed to the optimizer. This call also doesn't take in an
2687
// Object; if you wish to check an Object you need to load the Object's class
2688
// prior to coming here.
2689
Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2690
ciMethod* method, int bci) {
2692
if ((*ctrl)->is_top()) {
2696
// Fast check for identical types, perhaps identical constants.
2697
// The types can even be identical non-constants, in cases
2698
// involving Array.newInstance, Object.clone, etc.
2699
if (subklass == superklass)
2700
return C->top(); // false path is dead; no test needed.
2702
if (gvn.type(superklass)->singleton()) {
2703
const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2704
const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2706
// In the common case of an exact superklass, try to fold up the
2707
// test before generating code. You may ask, why not just generate
2708
// the code and then let it fold up? The answer is that the generated
2709
// code will necessarily include null checks, which do not always
2710
// completely fold away. If they are also needless, then they turn
2711
// into a performance loss. Example:
2712
// Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2713
// Here, the type of 'fa' is often exact, so the store check
2714
// of fa[1]=x will fold up, without testing the nullness of x.
2716
// At macro expansion, we would have already folded the SubTypeCheckNode
2717
// being expanded here because we always perform the static sub type
2718
// check in SubTypeCheckNode::sub() regardless of whether
2719
// StressReflectiveCode is set or not. We can therefore skip this
2720
// static check when StressReflectiveCode is on.
2721
switch (C->static_subtype_check(superk, subk)) {
2722
case Compile::SSC_always_false:
2724
Node* always_fail = *ctrl;
2725
*ctrl = gvn.C->top();
2728
case Compile::SSC_always_true:
2730
case Compile::SSC_easy_test:
2732
// Just do a direct pointer compare and be done.
2733
IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2734
*ctrl = gvn.transform(new IfTrueNode(iff));
2735
return gvn.transform(new IfFalseNode(iff));
2737
case Compile::SSC_full_test:
2740
ShouldNotReachHere();
2744
// %%% Possible further optimization: Even if the superklass is not exact,
2745
// if the subklass is the unique subtype of the superklass, the check
2746
// will always succeed. We could leave a dependency behind to ensure this.
2748
// First load the super-klass's check-offset
2749
Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2750
Node* m = C->immutable_memory();
2751
Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2752
int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2753
const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2754
int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
2755
bool might_be_cache = (chk_off_con == cacheoff_con);
2757
// Load from the sub-klass's super-class display list, or a 1-word cache of
2758
// the secondary superclass list, or a failing value with a sentinel offset
2759
// if the super-klass is an interface or exceptionally deep in the Java
2760
// hierarchy and we have to scan the secondary superclass list the hard way.
2761
// Worst-case type is a little odd: null is allowed as a result (usually
2762
// klass loads can never produce a null).
2763
Node *chk_off_X = chk_off;
2765
chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2767
Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2768
// For some types like interfaces the following loadKlass is from a 1-word
2769
// cache which is mutable so can't use immutable memory. Other
2770
// types load from the super-class display table which is immutable.
2771
Node *kmem = C->immutable_memory();
2772
// secondary_super_cache is not immutable but can be treated as such because:
2773
// - no ideal node writes to it in a way that could cause an
2774
// incorrect/missed optimization of the following Load.
2775
// - it's a cache so, worse case, not reading the latest value
2776
// wouldn't cause incorrect execution
2777
if (might_be_cache && mem != nullptr) {
2778
kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
2780
Node *nkls = gvn.transform(LoadKlassNode::make(gvn, nullptr, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
2782
// Compile speed common case: ARE a subtype and we canNOT fail
2783
if (superklass == nkls) {
2784
return C->top(); // false path is dead; no test needed.
2787
// Gather the various success & failures here
2788
RegionNode* r_not_subtype = new RegionNode(3);
2789
gvn.record_for_igvn(r_not_subtype);
2790
RegionNode* r_ok_subtype = new RegionNode(4);
2791
gvn.record_for_igvn(r_ok_subtype);
2793
// If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2794
// SubTypeCheck node
2795
if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2796
ciCallProfile profile = method->call_profile_at_bci(bci);
2797
float total_prob = 0;
2798
for (int i = 0; profile.has_receiver(i); ++i) {
2799
float prob = profile.receiver_prob(i);
2802
if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2803
const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2804
for (int i = 0; profile.has_receiver(i); ++i) {
2805
ciKlass* klass = profile.receiver(i);
2806
const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2807
Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2808
if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2811
float prob = profile.receiver_prob(i);
2812
ConNode* klass_node = gvn.makecon(klass_t);
2813
IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2814
Node* iftrue = gvn.transform(new IfTrueNode(iff));
2816
if (result == Compile::SSC_always_true) {
2817
r_ok_subtype->add_req(iftrue);
2819
assert(result == Compile::SSC_always_false, "");
2820
r_not_subtype->add_req(iftrue);
2822
*ctrl = gvn.transform(new IfFalseNode(iff));
2827
// See if we get an immediate positive hit. Happens roughly 83% of the
2828
// time. Test to see if the value loaded just previously from the subklass
2829
// is exactly the superklass.
2830
IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2831
Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
2832
*ctrl = gvn.transform(new IfFalseNode(iff1));
2834
// Compile speed common case: Check for being deterministic right now. If
2835
// chk_off is a constant and not equal to cacheoff then we are NOT a
2836
// subklass. In this case we need exactly the 1 test above and we can
2837
// return those results immediately.
2838
if (!might_be_cache) {
2839
Node* not_subtype_ctrl = *ctrl;
2840
*ctrl = iftrue1; // We need exactly the 1 test above
2841
PhaseIterGVN* igvn = gvn.is_IterGVN();
2842
if (igvn != nullptr) {
2843
igvn->remove_globally_dead_node(r_ok_subtype);
2844
igvn->remove_globally_dead_node(r_not_subtype);
2846
return not_subtype_ctrl;
2849
r_ok_subtype->init_req(1, iftrue1);
2851
// Check for immediate negative hit. Happens roughly 11% of the time (which
2852
// is roughly 63% of the remaining cases). Test to see if the loaded
2853
// check-offset points into the subklass display list or the 1-element
2854
// cache. If it points to the display (and NOT the cache) and the display
2855
// missed then it's not a subtype.
2856
Node *cacheoff = gvn.intcon(cacheoff_con);
2857
IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2858
r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2859
*ctrl = gvn.transform(new IfFalseNode(iff2));
2861
// Check for self. Very rare to get here, but it is taken 1/3 the time.
2862
// No performance impact (too rare) but allows sharing of secondary arrays
2863
// which has some footprint reduction.
2864
IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2865
r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2866
*ctrl = gvn.transform(new IfFalseNode(iff3));
2868
// -- Roads not taken here: --
2869
// We could also have chosen to perform the self-check at the beginning
2870
// of this code sequence, as the assembler does. This would not pay off
2871
// the same way, since the optimizer, unlike the assembler, can perform
2872
// static type analysis to fold away many successful self-checks.
2873
// Non-foldable self checks work better here in second position, because
2874
// the initial primary superclass check subsumes a self-check for most
2875
// types. An exception would be a secondary type like array-of-interface,
2876
// which does not appear in its own primary supertype display.
2877
// Finally, we could have chosen to move the self-check into the
2878
// PartialSubtypeCheckNode, and from there out-of-line in a platform
2879
// dependent manner. But it is worthwhile to have the check here,
2880
// where it can be perhaps be optimized. The cost in code space is
2881
// small (register compare, branch).
2883
// Now do a linear scan of the secondary super-klass array. Again, no real
2884
// performance impact (too rare) but it's gotta be done.
2885
// Since the code is rarely used, there is no penalty for moving it
2886
// out of line, and it can only improve I-cache density.
2887
// The decision to inline or out-of-line this final check is platform
2888
// dependent, and is found in the AD file definition of PartialSubtypeCheck.
2889
Node* psc = gvn.transform(
2890
new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2892
IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2893
r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2894
r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2896
// Return false path; set default control to true path.
2897
*ctrl = gvn.transform(r_ok_subtype);
2898
return gvn.transform(r_not_subtype);
2901
Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2902
bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
2903
if (expand_subtype_check) {
2904
MergeMemNode* mem = merged_memory();
2905
Node* ctrl = control();
2906
Node* subklass = obj_or_subklass;
2907
if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2908
subklass = load_object_klass(obj_or_subklass);
2911
Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
2916
Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
2917
Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2918
IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2919
set_control(_gvn.transform(new IfTrueNode(iff)));
2920
return _gvn.transform(new IfFalseNode(iff));
2923
// Profile-driven exact type check:
2924
Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2926
Node* *casted_receiver) {
2927
assert(!klass->is_interface(), "no exact type check on interfaces");
2929
const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
2930
Node* recv_klass = load_object_klass(receiver);
2931
Node* want_klass = makecon(tklass);
2932
Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
2933
Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
2934
IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2935
set_control( _gvn.transform(new IfTrueNode (iff)));
2936
Node* fail = _gvn.transform(new IfFalseNode(iff));
2939
const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2940
const TypeOopPtr* recvx_type = tklass->as_instance_type();
2941
assert(recvx_type->klass_is_exact(), "");
2943
if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
2944
// Subsume downstream occurrences of receiver with a cast to
2945
// recv_xtype, since now we know what the type will be.
2946
Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
2947
(*casted_receiver) = _gvn.transform(cast);
2948
assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
2949
// (User must make the replace_in_map call.)
2956
//------------------------------subtype_check_receiver-------------------------
2957
Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2958
Node** casted_receiver) {
2959
const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
2960
Node* want_klass = makecon(tklass);
2962
Node* slow_ctl = gen_subtype_check(receiver, want_klass);
2964
// Ignore interface type information until interface types are properly tracked.
2965
if (!stopped() && !klass->is_interface()) {
2966
const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2967
const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2968
if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
2969
Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2970
(*casted_receiver) = _gvn.transform(cast);
2977
//------------------------------seems_never_null-------------------------------
2978
// Use null_seen information if it is available from the profile.
2979
// If we see an unexpected null at a type check we record it and force a
2980
// recompile; the offending check will be recompiled to handle nulls.
2981
// If we see several offending BCIs, then all checks in the
2982
// method will be recompiled.
2983
bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2984
speculating = !_gvn.type(obj)->speculative_maybe_null();
2985
Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2986
if (UncommonNullCast // Cutout for this technique
2987
&& obj != null() // And not the -Xcomp stupid case?
2988
&& !too_many_traps(reason)
2993
if (data == nullptr)
2994
// Edge case: no mature data. Be optimistic here.
2996
// If the profile has not seen a null, assume it won't happen.
2997
assert(java_bc() == Bytecodes::_checkcast ||
2998
java_bc() == Bytecodes::_instanceof ||
2999
java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3000
return !data->as_BitData()->null_seen();
3002
speculating = false;
3006
void GraphKit::guard_klass_being_initialized(Node* klass) {
3007
int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3008
Node* adr = basic_plus_adr(top(), klass, init_state_off);
3009
Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3010
adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3011
T_BYTE, MemNode::unordered);
3012
init_state = _gvn.transform(init_state);
3014
Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3016
Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3017
Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3019
{ BuildCutout unless(this, tst, PROB_MAX);
3020
uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3024
void GraphKit::guard_init_thread(Node* klass) {
3025
int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3026
Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3028
Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3029
adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3030
T_ADDRESS, MemNode::unordered);
3031
init_thread = _gvn.transform(init_thread);
3033
Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3035
Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3036
Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3038
{ BuildCutout unless(this, tst, PROB_MAX);
3039
uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3043
void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3044
if (ik->is_being_initialized()) {
3045
if (C->needs_clinit_barrier(ik, context)) {
3046
Node* klass = makecon(TypeKlassPtr::make(ik));
3047
guard_klass_being_initialized(klass);
3048
guard_init_thread(klass);
3049
insert_mem_bar(Op_MemBarCPUOrder);
3051
} else if (ik->is_initialized()) {
3052
return; // no barrier needed
3054
uncommon_trap(Deoptimization::Reason_uninitialized,
3055
Deoptimization::Action_reinterpret,
3060
//------------------------maybe_cast_profiled_receiver-------------------------
3061
// If the profile has seen exactly one type, narrow to exactly that type.
3062
// Subsequent type checks will always fold up.
3063
Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3064
const TypeKlassPtr* require_klass,
3065
ciKlass* spec_klass,
3066
bool safe_for_replace) {
3067
if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3069
Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3071
// Make sure we haven't already deoptimized from this tactic.
3072
if (too_many_traps_or_recompiles(reason))
3075
// (No, this isn't a call, but it's enough like a virtual call
3076
// to use the same ciMethod accessor to get the profile info...)
3077
// If we have a speculative type use it instead of profiling (which
3079
ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;
3080
if (exact_kls != nullptr) {// no cast failures here
3081
if (require_klass == nullptr ||
3082
C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3083
// If we narrow the type to match what the type profile sees or
3084
// the speculative type, we can then remove the rest of the
3086
// This is a win, even if the exact_kls is very specific,
3087
// because downstream operations, such as method calls,
3088
// will often benefit from the sharper type.
3089
Node* exact_obj = not_null_obj; // will get updated in place...
3090
Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3092
{ PreserveJVMState pjvms(this);
3093
set_control(slow_ctl);
3094
uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3096
if (safe_for_replace) {
3097
replace_in_map(not_null_obj, exact_obj);
3101
// assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3108
* Cast obj to type and emit guard unless we had too many traps here
3111
* @param obj node being casted
3112
* @param type type to cast the node to
3113
* @param not_null true if we know node cannot be null
3115
Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3122
// type is null if profiling tells us this object is always null
3123
if (type != nullptr) {
3124
Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3125
Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3127
if (!too_many_traps_or_recompiles(null_reason) &&
3128
!too_many_traps_or_recompiles(class_reason)) {
3129
Node* not_null_obj = nullptr;
3130
// not_null is true if we know the object is not null and
3131
// there's no need for a null check
3133
Node* null_ctl = top();
3134
not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3135
assert(null_ctl->is_top(), "no null control here");
3140
Node* exact_obj = not_null_obj;
3141
ciKlass* exact_kls = type;
3142
Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3145
PreserveJVMState pjvms(this);
3146
set_control(slow_ctl);
3147
uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3149
replace_in_map(not_null_obj, exact_obj);
3153
if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3154
Node* exact_obj = null_assert(obj);
3155
replace_in_map(obj, exact_obj);
3162
//-------------------------------gen_instanceof--------------------------------
3163
// Generate an instance-of idiom. Used by both the instance-of bytecode
3164
// and the reflective instance-of call.
3165
Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3166
kill_dead_locals(); // Benefit all the uncommon traps
3167
assert( !stopped(), "dead parse path should be checked in callers" );
3168
assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3169
"must check for not-null not-dead klass in callers");
3171
// Make the merge point
3172
enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3173
RegionNode* region = new RegionNode(PATH_LIMIT);
3174
Node* phi = new PhiNode(region, TypeInt::BOOL);
3175
C->set_has_split_ifs(true); // Has chance for split-if optimization
3177
ciProfileData* data = nullptr;
3178
if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3179
data = method()->method_data()->bci_to_data(bci());
3181
bool speculative_not_null = false;
3182
bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3183
&& seems_never_null(obj, data, speculative_not_null));
3185
// Null check; get casted pointer; set region slot 3
3186
Node* null_ctl = top();
3187
Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3189
// If not_null_obj is dead, only null-path is taken
3190
if (stopped()) { // Doing instance-of on a null?
3191
set_control(null_ctl);
3194
region->init_req(_null_path, null_ctl);
3195
phi ->init_req(_null_path, intcon(0)); // Set null path value
3196
if (null_ctl == top()) {
3197
// Do this eagerly, so that pattern matches like is_diamond_phi
3198
// will work even during parsing.
3199
assert(_null_path == PATH_LIMIT-1, "delete last");
3200
region->del_req(_null_path);
3201
phi ->del_req(_null_path);
3204
// Do we know the type check always succeed?
3205
bool known_statically = false;
3206
if (_gvn.type(superklass)->singleton()) {
3207
const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3208
const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3209
if (subk->is_loaded()) {
3210
int static_res = C->static_subtype_check(superk, subk);
3211
known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3215
if (!known_statically) {
3216
const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3217
// We may not have profiling here or it may not help us. If we
3218
// have a speculative type use it to perform an exact cast.
3219
ciKlass* spec_obj_type = obj_type->speculative_type();
3220
if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3221
Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3222
if (stopped()) { // Profile disagrees with this path.
3223
set_control(null_ctl); // Null is the only remaining possibility.
3226
if (cast_obj != nullptr) {
3227
not_null_obj = cast_obj;
3232
// Generate the subtype check
3233
Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3235
// Plug in the success path to the general merge in slot 1.
3236
region->init_req(_obj_path, control());
3237
phi ->init_req(_obj_path, intcon(1));
3239
// Plug in the failing path to the general merge in slot 2.
3240
region->init_req(_fail_path, not_subtype_ctrl);
3241
phi ->init_req(_fail_path, intcon(0));
3243
// Return final merged results
3244
set_control( _gvn.transform(region) );
3245
record_for_igvn(region);
3247
// If we know the type check always succeeds then we don't use the
3248
// profiling data at this bytecode. Don't lose it, feed it to the
3249
// type system as a speculative type.
3250
if (safe_for_replace) {
3251
Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3252
replace_in_map(obj, casted_obj);
3255
return _gvn.transform(phi);
3258
//-------------------------------gen_checkcast---------------------------------
3259
// Generate a checkcast idiom. Used by both the checkcast bytecode and the
3260
// array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3261
// uncommon-trap paths work. Adjust stack after this call.
3262
// If failure_control is supplied and not null, it is filled in with
3263
// the control edge for the cast failure. Otherwise, an appropriate
3264
// uncommon trap or exception is thrown.
3265
Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3266
Node* *failure_control) {
3267
kill_dead_locals(); // Benefit all the uncommon traps
3268
const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3269
const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3270
const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3272
// Fast cutout: Check the case that the cast is vacuously true.
3273
// This detects the common cases where the test will short-circuit
3274
// away completely. We do this before we perform the null check,
3275
// because if the test is going to turn into zero code, we don't
3276
// want a residual null check left around. (Causes a slowdown,
3277
// for example, in some objArray manipulations, such as a[i]=a[j].)
3278
if (improved_klass_ptr_type->singleton()) {
3279
const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3280
if (objtp != nullptr) {
3281
switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) {
3282
case Compile::SSC_always_true:
3283
// If we know the type check always succeed then we don't use
3284
// the profiling data at this bytecode. Don't lose it, feed it
3285
// to the type system as a speculative type.
3286
return record_profiled_receiver_for_speculation(obj);
3287
case Compile::SSC_always_false:
3288
// It needs a null check because a null will *pass* the cast check.
3289
// A non-null value will always produce an exception.
3290
if (!objtp->maybe_null()) {
3291
bool is_aastore = (java_bc() == Bytecodes::_aastore);
3292
Deoptimization::DeoptReason reason = is_aastore ?
3293
Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3294
builtin_throw(reason);
3296
} else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3297
return null_assert(obj);
3299
break; // Fall through to full check
3306
ciProfileData* data = nullptr;
3307
bool safe_for_replace = false;
3308
if (failure_control == nullptr) { // use MDO in regular case only
3309
assert(java_bc() == Bytecodes::_aastore ||
3310
java_bc() == Bytecodes::_checkcast,
3311
"interpreter profiles type checks only for these BCs");
3312
data = method()->method_data()->bci_to_data(bci());
3313
safe_for_replace = true;
3316
// Make the merge point
3317
enum { _obj_path = 1, _null_path, PATH_LIMIT };
3318
RegionNode* region = new RegionNode(PATH_LIMIT);
3319
Node* phi = new PhiNode(region, toop);
3320
C->set_has_split_ifs(true); // Has chance for split-if optimization
3322
// Use null-cast information if it is available
3323
bool speculative_not_null = false;
3324
bool never_see_null = ((failure_control == nullptr) // regular case only
3325
&& seems_never_null(obj, data, speculative_not_null));
3327
// Null check; get casted pointer; set region slot 3
3328
Node* null_ctl = top();
3329
Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3331
// If not_null_obj is dead, only null-path is taken
3332
if (stopped()) { // Doing instance-of on a null?
3333
set_control(null_ctl);
3336
region->init_req(_null_path, null_ctl);
3337
phi ->init_req(_null_path, null()); // Set null path value
3338
if (null_ctl == top()) {
3339
// Do this eagerly, so that pattern matches like is_diamond_phi
3340
// will work even during parsing.
3341
assert(_null_path == PATH_LIMIT-1, "delete last");
3342
region->del_req(_null_path);
3343
phi ->del_req(_null_path);
3346
Node* cast_obj = nullptr;
3347
if (improved_klass_ptr_type->klass_is_exact()) {
3348
// The following optimization tries to statically cast the speculative type of the object
3349
// (for example obtained during profiling) to the type of the superklass and then do a
3350
// dynamic check that the type of the object is what we expect. To work correctly
3351
// for checkcast and aastore the type of superklass should be exact.
3352
const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3353
// We may not have profiling here or it may not help us. If we have
3354
// a speculative type use it to perform an exact cast.
3355
ciKlass* spec_obj_type = obj_type->speculative_type();
3356
if (spec_obj_type != nullptr || data != nullptr) {
3357
cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3358
if (cast_obj != nullptr) {
3359
if (failure_control != nullptr) // failure is now impossible
3360
(*failure_control) = top();
3361
// adjust the type of the phi to the exact klass:
3362
phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3367
if (cast_obj == nullptr) {
3368
// Generate the subtype check
3369
Node* improved_superklass = superklass;
3370
if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3371
improved_superklass = makecon(improved_klass_ptr_type);
3373
Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3375
// Plug in success path into the merge
3376
cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3377
// Failure path ends in uncommon trap (or may be dead - failure impossible)
3378
if (failure_control == nullptr) {
3379
if (not_subtype_ctrl != top()) { // If failure is possible
3380
PreserveJVMState pjvms(this);
3381
set_control(not_subtype_ctrl);
3382
bool is_aastore = (java_bc() == Bytecodes::_aastore);
3383
Deoptimization::DeoptReason reason = is_aastore ?
3384
Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3385
builtin_throw(reason);
3388
(*failure_control) = not_subtype_ctrl;
3392
region->init_req(_obj_path, control());
3393
phi ->init_req(_obj_path, cast_obj);
3395
// A merge of null or Casted-NotNull obj
3396
Node* res = _gvn.transform(phi);
3398
// Note I do NOT always 'replace_in_map(obj,result)' here.
3399
// if( tk->klass()->can_be_primary_super() )
3400
// This means that if I successfully store an Object into an array-of-String
3401
// I 'forget' that the Object is really now known to be a String. I have to
3402
// do this because we don't have true union types for interfaces - if I store
3403
// a Baz into an array-of-Interface and then tell the optimizer it's an
3404
// Interface, I forget that it's also a Baz and cannot do Baz-like field
3405
// references to it. FIX THIS WHEN UNION TYPES APPEAR!
3406
// replace_in_map( obj, res );
3408
// Return final merged results
3409
set_control( _gvn.transform(region) );
3410
record_for_igvn(region);
3412
return record_profiled_receiver_for_speculation(res);
3415
//------------------------------next_monitor-----------------------------------
3416
// What number should be given to the next monitor?
3417
int GraphKit::next_monitor() {
3418
int current = jvms()->monitor_depth()* C->sync_stack_slots();
3419
int next = current + C->sync_stack_slots();
3420
// Keep the toplevel high water mark current:
3421
if (C->fixed_slots() < next) C->set_fixed_slots(next);
3425
//------------------------------insert_mem_bar---------------------------------
3426
// Memory barrier to avoid floating things around
3427
// The membar serves as a pinch point between both control and all memory slices.
3428
Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3429
MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3430
mb->init_req(TypeFunc::Control, control());
3431
mb->init_req(TypeFunc::Memory, reset_memory());
3432
Node* membar = _gvn.transform(mb);
3433
set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3434
set_all_memory_call(membar);
3438
//-------------------------insert_mem_bar_volatile----------------------------
3439
// Memory barrier to avoid floating things around
3440
// The membar serves as a pinch point between both control and memory(alias_idx).
3441
// If you want to make a pinch point on all memory slices, do not use this
3442
// function (even with AliasIdxBot); use insert_mem_bar() instead.
3443
Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3444
// When Parse::do_put_xxx updates a volatile field, it appends a series
3445
// of MemBarVolatile nodes, one for *each* volatile field alias category.
3446
// The first membar is on the same memory slice as the field store opcode.
3447
// This forces the membar to follow the store. (Bug 6500685 broke this.)
3448
// All the other membars (for other volatile slices, including AliasIdxBot,
3449
// which stands for all unknown volatile slices) are control-dependent
3450
// on the first membar. This prevents later volatile loads or stores
3451
// from sliding up past the just-emitted store.
3453
MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3454
mb->set_req(TypeFunc::Control,control());
3455
if (alias_idx == Compile::AliasIdxBot) {
3456
mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3458
assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3459
mb->set_req(TypeFunc::Memory, memory(alias_idx));
3461
Node* membar = _gvn.transform(mb);
3462
set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3463
if (alias_idx == Compile::AliasIdxBot) {
3464
merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3466
set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3471
//------------------------------shared_lock------------------------------------
3472
// Emit locking code.
3473
FastLockNode* GraphKit::shared_lock(Node* obj) {
3474
// bci is either a monitorenter bc or InvocationEntryBci
3475
// %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3476
assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3478
if( !GenerateSynchronizationCode )
3479
return nullptr; // Not locking things?
3480
if (stopped()) // Dead monitor?
3483
assert(dead_locals_are_killed(), "should kill locals before sync. point");
3485
// Box the stack location
3486
Node* box = new BoxLockNode(next_monitor());
3487
// Check for bailout after new BoxLockNode
3488
if (failing()) { return nullptr; }
3489
box = _gvn.transform(box);
3490
Node* mem = reset_memory();
3492
FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3494
// Add monitor to debug info for the slow path. If we block inside the
3495
// slow path and de-opt, we need the monitor hanging around
3496
map()->push_monitor( flock );
3498
const TypeFunc *tf = LockNode::lock_type();
3499
LockNode *lock = new LockNode(C, tf);
3501
lock->init_req( TypeFunc::Control, control() );
3502
lock->init_req( TypeFunc::Memory , mem );
3503
lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3504
lock->init_req( TypeFunc::FramePtr, frameptr() );
3505
lock->init_req( TypeFunc::ReturnAdr, top() );
3507
lock->init_req(TypeFunc::Parms + 0, obj);
3508
lock->init_req(TypeFunc::Parms + 1, box);
3509
lock->init_req(TypeFunc::Parms + 2, flock);
3510
add_safepoint_edges(lock);
3512
lock = _gvn.transform( lock )->as_Lock();
3514
// lock has no side-effects, sets few values
3515
set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3517
insert_mem_bar(Op_MemBarAcquireLock);
3519
// Add this to the worklist so that the lock can be eliminated
3520
record_for_igvn(lock);
3523
if (PrintLockStatistics) {
3524
// Update the counter for this lock. Don't bother using an atomic
3525
// operation since we don't require absolute accuracy.
3526
lock->create_lock_counter(map()->jvms());
3527
increment_counter(lock->counter()->addr());
3535
//------------------------------shared_unlock----------------------------------
3536
// Emit unlocking code.
3537
void GraphKit::shared_unlock(Node* box, Node* obj) {
3538
// bci is either a monitorenter bc or InvocationEntryBci
3539
// %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3540
assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3542
if( !GenerateSynchronizationCode )
3544
if (stopped()) { // Dead monitor?
3545
map()->pop_monitor(); // Kill monitor from debug info
3549
// Memory barrier to avoid floating things down past the locked region
3550
insert_mem_bar(Op_MemBarReleaseLock);
3552
const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3553
UnlockNode *unlock = new UnlockNode(C, tf);
3555
unlock->set_dbg_jvms(sync_jvms());
3557
uint raw_idx = Compile::AliasIdxRaw;
3558
unlock->init_req( TypeFunc::Control, control() );
3559
unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3560
unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3561
unlock->init_req( TypeFunc::FramePtr, frameptr() );
3562
unlock->init_req( TypeFunc::ReturnAdr, top() );
3564
unlock->init_req(TypeFunc::Parms + 0, obj);
3565
unlock->init_req(TypeFunc::Parms + 1, box);
3566
unlock = _gvn.transform(unlock)->as_Unlock();
3568
Node* mem = reset_memory();
3570
// unlock has no side-effects, sets few values
3571
set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3573
// Kill monitor from debug info
3574
map()->pop_monitor( );
3577
//-------------------------------get_layout_helper-----------------------------
3578
// If the given klass is a constant or known to be an array,
3579
// fetch the constant layout helper value into constant_value
3580
// and return null. Otherwise, load the non-constant
3581
// layout helper value, and return the node which represents it.
3582
// This two-faced routine is useful because allocation sites
3583
// almost always feature constant types.
3584
Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3585
const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3586
if (!StressReflectiveCode && klass_t != nullptr) {
3587
bool xklass = klass_t->klass_is_exact();
3588
if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {
3590
if (klass_t->isa_aryklassptr()) {
3591
BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
3592
if (is_reference_type(elem, true)) {
3595
lhelper = Klass::array_layout_helper(elem);
3597
lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3599
if (lhelper != Klass::_lh_neutral_value) {
3600
constant_value = lhelper;
3601
return (Node*) nullptr;
3605
constant_value = Klass::_lh_neutral_value; // put in a known value
3606
Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3607
return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3610
// We just put in an allocate/initialize with a big raw-memory effect.
3611
// Hook selected additional alias categories on the initialization.
3612
static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3613
MergeMemNode* init_in_merge,
3614
Node* init_out_raw) {
3615
DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3616
assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3618
Node* prevmem = kit.memory(alias_idx);
3619
init_in_merge->set_memory_at(alias_idx, prevmem);
3620
kit.set_memory(init_out_raw, alias_idx);
3623
//---------------------------set_output_for_allocation-------------------------
3624
Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3625
const TypeOopPtr* oop_type,
3626
bool deoptimize_on_exception) {
3627
int rawidx = Compile::AliasIdxRaw;
3628
alloc->set_req( TypeFunc::FramePtr, frameptr() );
3629
add_safepoint_edges(alloc);
3630
Node* allocx = _gvn.transform(alloc);
3631
set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3632
// create memory projection for i_o
3633
set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3634
make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3636
// create a memory projection as for the normal control path
3637
Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3638
set_memory(malloc, rawidx);
3640
// a normal slow-call doesn't change i_o, but an allocation does
3641
// we create a separate i_o projection for the normal control path
3642
set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3643
Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3645
// put in an initialization barrier
3646
InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3647
rawoop)->as_Initialize();
3648
assert(alloc->initialization() == init, "2-way macro link must work");
3649
assert(init ->allocation() == alloc, "2-way macro link must work");
3651
// Extract memory strands which may participate in the new object's
3652
// initialization, and source them from the new InitializeNode.
3653
// This will allow us to observe initializations when they occur,
3654
// and link them properly (as a group) to the InitializeNode.
3655
assert(init->in(InitializeNode::Memory) == malloc, "");
3656
MergeMemNode* minit_in = MergeMemNode::make(malloc);
3657
init->set_req(InitializeNode::Memory, minit_in);
3658
record_for_igvn(minit_in); // fold it up later, if possible
3659
Node* minit_out = memory(rawidx);
3660
assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3661
// Add an edge in the MergeMem for the header fields so an access
3662
// to one of those has correct memory state
3663
set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3664
set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3665
if (oop_type->isa_aryptr()) {
3666
const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3667
int elemidx = C->get_alias_index(telemref);
3668
hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3669
} else if (oop_type->isa_instptr()) {
3670
ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3671
for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3672
ciField* field = ik->nonstatic_field_at(i);
3673
if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3674
continue; // do not bother to track really large numbers of fields
3675
// Find (or create) the alias category for this field:
3676
int fieldidx = C->alias_type(field)->index();
3677
hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3682
// Cast raw oop to the real thing...
3683
Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3684
javaoop = _gvn.transform(javaoop);
3685
C->set_recent_alloc(control(), javaoop);
3686
assert(just_allocated_object(control()) == javaoop, "just allocated");
3689
{ // Verify that the AllocateNode::Ideal_allocation recognizers work:
3690
assert(AllocateNode::Ideal_allocation(rawoop) == alloc,
3691
"Ideal_allocation works");
3692
assert(AllocateNode::Ideal_allocation(javaoop) == alloc,
3693
"Ideal_allocation works");
3694
if (alloc->is_AllocateArray()) {
3695
assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(),
3696
"Ideal_allocation works");
3697
assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(),
3698
"Ideal_allocation works");
3700
assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3708
//---------------------------new_instance--------------------------------------
3709
// This routine takes a klass_node which may be constant (for a static type)
3710
// or may be non-constant (for reflective code). It will work equally well
3711
// for either, and the graph will fold nicely if the optimizer later reduces
3712
// the type to a constant.
3713
// The optional arguments are for specialized use by intrinsics:
3714
// - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3715
// - If 'return_size_val', report the total object size to the caller.
3716
// - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3717
Node* GraphKit::new_instance(Node* klass_node,
3718
Node* extra_slow_test,
3719
Node* *return_size_val,
3720
bool deoptimize_on_exception) {
3721
// Compute size in doublewords
3722
// The size is always an integral number of doublewords, represented
3723
// as a positive bytewise size stored in the klass's layout_helper.
3724
// The layout_helper also encodes (in a low bit) the need for a slow path.
3725
jint layout_con = Klass::_lh_neutral_value;
3726
Node* layout_val = get_layout_helper(klass_node, layout_con);
3727
int layout_is_con = (layout_val == nullptr);
3729
if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
3730
// Generate the initial go-slow test. It's either ALWAYS (return a
3731
// Node for 1) or NEVER (return a null) or perhaps (in the reflective
3732
// case) a computed value derived from the layout_helper.
3733
Node* initial_slow_test = nullptr;
3734
if (layout_is_con) {
3735
assert(!StressReflectiveCode, "stress mode does not use these paths");
3736
bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3737
initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3738
} else { // reflective case
3739
// This reflective path is used by Unsafe.allocateInstance.
3740
// (It may be stress-tested by specifying StressReflectiveCode.)
3741
// Basically, we want to get into the VM is there's an illegal argument.
3742
Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3743
initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3744
if (extra_slow_test != intcon(0)) {
3745
initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3747
// (Macro-expander will further convert this to a Bool, if necessary.)
3750
// Find the size in bytes. This is easy; it's the layout_helper.
3751
// The size value must be valid even if the slow path is taken.
3752
Node* size = nullptr;
3753
if (layout_is_con) {
3754
size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3755
} else { // reflective case
3756
// This reflective path is used by clone and Unsafe.allocateInstance.
3757
size = ConvI2X(layout_val);
3759
// Clear the low bits to extract layout_helper_size_in_bytes:
3760
assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3761
Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3762
size = _gvn.transform( new AndXNode(size, mask) );
3764
if (return_size_val != nullptr) {
3765
(*return_size_val) = size;
3768
// This is a precise notnull oop of the klass.
3769
// (Actually, it need not be precise if this is a reflective allocation.)
3770
// It's what we cast the result to.
3771
const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3772
if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
3773
const TypeOopPtr* oop_type = tklass->as_instance_type();
3775
// Now generate allocation code
3777
// The entire memory state is needed for slow path of the allocation
3778
// since GC and deoptimization can happened.
3779
Node *mem = reset_memory();
3780
set_all_memory(mem); // Create new memory state
3782
AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3783
control(), mem, i_o(),
3787
return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3790
//-------------------------------new_array-------------------------------------
3791
// helper for both newarray and anewarray
3792
// The 'length' parameter is (obviously) the length of the array.
3793
// The optional arguments are for specialized use by intrinsics:
3794
// - If 'return_size_val', report the non-padded array size (sum of header size
3795
// and array body) to the caller.
3796
// - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3797
Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3798
Node* length, // number of array elements
3799
int nargs, // number of arguments to push back for uncommon trap
3800
Node* *return_size_val,
3801
bool deoptimize_on_exception) {
3802
jint layout_con = Klass::_lh_neutral_value;
3803
Node* layout_val = get_layout_helper(klass_node, layout_con);
3804
int layout_is_con = (layout_val == nullptr);
3806
if (!layout_is_con && !StressReflectiveCode &&
3807
!too_many_traps(Deoptimization::Reason_class_check)) {
3808
// This is a reflective array creation site.
3809
// Optimistically assume that it is a subtype of Object[],
3810
// so that we can fold up all the address arithmetic.
3811
layout_con = Klass::array_layout_helper(T_OBJECT);
3812
Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3813
Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3814
{ BuildCutout unless(this, bol_lh, PROB_MAX);
3816
uncommon_trap(Deoptimization::Reason_class_check,
3817
Deoptimization::Action_maybe_recompile);
3819
layout_val = nullptr;
3820
layout_is_con = true;
3823
// Generate the initial go-slow test. Make sure we do not overflow
3824
// if length is huge (near 2Gig) or negative! We do not need
3825
// exact double-words here, just a close approximation of needed
3826
// double-words. We can't add any offset or rounding bits, lest we
3827
// take a size -1 of bytes and make it positive. Use an unsigned
3828
// compare, so negative sizes look hugely positive.
3829
int fast_size_limit = FastAllocateSizeLimit;
3830
if (layout_is_con) {
3831
assert(!StressReflectiveCode, "stress mode does not use these paths");
3832
// Increase the size limit if we have exact knowledge of array type.
3833
int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3834
fast_size_limit <<= (LogBytesPerLong - log2_esize);
3837
Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3838
Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3840
// --- Size Computation ---
3841
// array_size = round_to_heap(array_header + (length << elem_shift));
3842
// where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3843
// and align_to(x, y) == ((x + y-1) & ~(y-1))
3844
// The rounding mask is strength-reduced, if possible.
3845
int round_mask = MinObjAlignmentInBytes - 1;
3846
Node* header_size = nullptr;
3847
// (T_BYTE has the weakest alignment and size restrictions...)
3848
if (layout_is_con) {
3849
int hsize = Klass::layout_helper_header_size(layout_con);
3850
int eshift = Klass::layout_helper_log2_element_size(layout_con);
3851
if ((round_mask & ~right_n_bits(eshift)) == 0)
3852
round_mask = 0; // strength-reduce it if it goes away completely
3853
assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3854
int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3855
assert(header_size_min <= hsize, "generic minimum is smallest");
3856
header_size = intcon(hsize);
3858
Node* hss = intcon(Klass::_lh_header_size_shift);
3859
Node* hsm = intcon(Klass::_lh_header_size_mask);
3860
header_size = _gvn.transform(new URShiftINode(layout_val, hss));
3861
header_size = _gvn.transform(new AndINode(header_size, hsm));
3864
Node* elem_shift = nullptr;
3865
if (layout_is_con) {
3866
int eshift = Klass::layout_helper_log2_element_size(layout_con);
3868
elem_shift = intcon(eshift);
3870
// There is no need to mask or shift this value.
3871
// The semantics of LShiftINode include an implicit mask to 0x1F.
3872
assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3873
elem_shift = layout_val;
3876
// Transition to native address size for all offset calculations:
3877
Node* lengthx = ConvI2X(length);
3878
Node* headerx = ConvI2X(header_size);
3880
{ const TypeInt* tilen = _gvn.find_int_type(length);
3881
if (tilen != nullptr && tilen->_lo < 0) {
3882
// Add a manual constraint to a positive range. Cf. array_element_address.
3883
jint size_max = fast_size_limit;
3884
if (size_max > tilen->_hi) size_max = tilen->_hi;
3885
const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3887
// Only do a narrow I2L conversion if the range check passed.
3888
IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3889
_gvn.transform(iff);
3890
RegionNode* region = new RegionNode(3);
3891
_gvn.set_type(region, Type::CONTROL);
3892
lengthx = new PhiNode(region, TypeLong::LONG);
3893
_gvn.set_type(lengthx, TypeLong::LONG);
3895
// Range check passed. Use ConvI2L node with narrow type.
3896
Node* passed = IfFalse(iff);
3897
region->init_req(1, passed);
3898
// Make I2L conversion control dependent to prevent it from
3899
// floating above the range check during loop optimizations.
3900
lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3902
// Range check failed. Use ConvI2L with wide type because length may be invalid.
3903
region->init_req(2, IfTrue(iff));
3904
lengthx->init_req(2, ConvI2X(length));
3906
set_control(region);
3907
record_for_igvn(region);
3908
record_for_igvn(lengthx);
3913
// Combine header size and body size for the array copy part, then align (if
3914
// necessary) for the allocation part. This computation cannot overflow,
3915
// because it is used only in two places, one where the length is sharply
3916
// limited, and the other after a successful allocation.
3917
Node* abody = lengthx;
3918
if (elem_shift != nullptr) {
3919
abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift));
3921
Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
3923
if (return_size_val != nullptr) {
3925
(*return_size_val) = non_rounded_size;
3928
Node* size = non_rounded_size;
3929
if (round_mask != 0) {
3930
Node* mask1 = MakeConX(round_mask);
3931
size = _gvn.transform(new AddXNode(size, mask1));
3932
Node* mask2 = MakeConX(~round_mask);
3933
size = _gvn.transform(new AndXNode(size, mask2));
3935
// else if round_mask == 0, the size computation is self-rounding
3937
// Now generate allocation code
3939
// The entire memory state is needed for slow path of the allocation
3940
// since GC and deoptimization can happened.
3941
Node *mem = reset_memory();
3942
set_all_memory(mem); // Create new memory state
3944
if (initial_slow_test->is_Bool()) {
3945
// Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3946
initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3949
const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3950
Node* valid_length_test = _gvn.intcon(1);
3951
if (ary_type->isa_aryptr()) {
3952
BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
3953
jint max = TypeAryPtr::max_array_length(bt);
3954
Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
3955
valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
3958
// Create the AllocateArrayNode and its result projections
3959
AllocateArrayNode* alloc
3960
= new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3961
control(), mem, i_o(),
3964
length, valid_length_test);
3966
// Cast to correct type. Note that the klass_node may be constant or not,
3967
// and in the latter case the actual array type will be inexact also.
3968
// (This happens via a non-constant argument to inline_native_newArray.)
3969
// In any case, the value of klass_node provides the desired array type.
3970
const TypeInt* length_type = _gvn.find_int_type(length);
3971
if (ary_type->isa_aryptr() && length_type != nullptr) {
3972
// Try to get a better type than POS for the size
3973
ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3976
Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3978
array_ideal_length(alloc, ary_type, true);
3982
// The following "Ideal_foo" functions are placed here because they recognize
3983
// the graph shapes created by the functions immediately above.
3985
//---------------------------Ideal_allocation----------------------------------
3986
// Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3987
AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) {
3988
if (ptr == nullptr) { // reduce dumb test in callers
3992
BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
3993
ptr = bs->step_over_gc_barrier(ptr);
3995
if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3997
if (ptr == nullptr) return nullptr;
3999
// Return null for allocations with several casts:
4000
// j.l.reflect.Array.newInstance(jobject, jint)
4002
// to keep more precise type from last cast.
4003
if (ptr->is_Proj()) {
4004
Node* allo = ptr->in(0);
4005
if (allo != nullptr && allo->is_Allocate()) {
4006
return allo->as_Allocate();
4009
// Report failure to match.
4013
// Fancy version which also strips off an offset (and reports it to caller).
4014
AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase,
4016
Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4017
if (base == nullptr) return nullptr;
4018
return Ideal_allocation(base);
4021
// Trace Initialize <- Proj[Parm] <- Allocate
4022
AllocateNode* InitializeNode::allocation() {
4023
Node* rawoop = in(InitializeNode::RawAddress);
4024
if (rawoop->is_Proj()) {
4025
Node* alloc = rawoop->in(0);
4026
if (alloc->is_Allocate()) {
4027
return alloc->as_Allocate();
4033
// Trace Allocate -> Proj[Parm] -> Initialize
4034
InitializeNode* AllocateNode::initialization() {
4035
ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4036
if (rawoop == nullptr) return nullptr;
4037
for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4038
Node* init = rawoop->fast_out(i);
4039
if (init->is_Initialize()) {
4040
assert(init->as_Initialize()->allocation() == this, "2-way link");
4041
return init->as_Initialize();
4047
// Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
4048
void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) {
4049
// Too many traps seen?
4050
if (too_many_traps(reason)) {
4052
if (TraceLoopPredicate) {
4053
int tc = C->trap_count(reason);
4054
tty->print("too many traps=%s tcount=%d in ",
4055
Deoptimization::trap_reason_name(reason), tc);
4056
method()->print(); // which method has too many predicate traps
4060
// We cannot afford to take more traps here,
4061
// do not generate Parse Predicate.
4065
ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn);
4066
_gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn));
4067
Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate));
4069
PreserveJVMState pjvms(this);
4070
set_control(if_false);
4072
uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4074
Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate));
4075
set_control(if_true);
4078
// Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All
4079
// Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate.
4080
void GraphKit::add_parse_predicates(int nargs) {
4081
if (UseLoopPredicate) {
4082
add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4084
if (UseProfiledLoopPredicate) {
4085
add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4087
// Loop Limit Check Predicate should be near the loop.
4088
add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4091
void GraphKit::sync_kit(IdealKit& ideal) {
4092
set_all_memory(ideal.merged_memory());
4093
set_i_o(ideal.i_o());
4094
set_control(ideal.ctrl());
4097
void GraphKit::final_sync(IdealKit& ideal) {
4098
// Final sync IdealKit and graphKit.
4102
Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4103
Node* len = load_array_length(load_String_value(str, set_ctrl));
4104
Node* coder = load_String_coder(str, set_ctrl);
4105
// Divide length by 2 if coder is UTF16
4106
return _gvn.transform(new RShiftINode(len, coder));
4109
Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4110
int value_offset = java_lang_String::value_offset();
4111
const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4113
const TypePtr* value_field_type = string_type->add_offset(value_offset);
4114
const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4115
TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4116
ciTypeArrayKlass::make(T_BYTE), true, 0);
4117
Node* p = basic_plus_adr(str, str, value_offset);
4118
Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4119
IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4123
Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4124
if (!CompactStrings) {
4125
return intcon(java_lang_String::CODER_UTF16);
4127
int coder_offset = java_lang_String::coder_offset();
4128
const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4130
const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4132
Node* p = basic_plus_adr(str, str, coder_offset);
4133
Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4134
IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4138
void GraphKit::store_String_value(Node* str, Node* value) {
4139
int value_offset = java_lang_String::value_offset();
4140
const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4142
const TypePtr* value_field_type = string_type->add_offset(value_offset);
4144
access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4145
value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4148
void GraphKit::store_String_coder(Node* str, Node* value) {
4149
int coder_offset = java_lang_String::coder_offset();
4150
const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4152
const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4154
access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4155
value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4158
// Capture src and dst memory state with a MergeMemNode
4159
Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4160
if (src_type == dst_type) {
4161
// Types are equal, we don't need a MergeMemNode
4162
return memory(src_type);
4164
MergeMemNode* merge = MergeMemNode::make(map()->memory());
4165
record_for_igvn(merge); // fold it up later, if possible
4166
int src_idx = C->get_alias_index(src_type);
4167
int dst_idx = C->get_alias_index(dst_type);
4168
merge->set_memory_at(src_idx, memory(src_idx));
4169
merge->set_memory_at(dst_idx, memory(dst_idx));
4173
Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4174
assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4175
assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4176
// If input and output memory types differ, capture both states to preserve
4177
// the dependency between preceding and subsequent loads/stores.
4178
// For example, the following program:
4182
// has this memory graph (use->def):
4183
// LoadB -> compress_string -> CharMem
4184
// ... -> StoreB -> ByteMem
4185
// The intrinsic hides the dependency between LoadB and StoreB, causing
4186
// the load to read from memory not containing the result of the StoreB.
4187
// The correct memory graph should look like this:
4188
// LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4189
Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4190
StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4191
Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4192
set_memory(res_mem, TypeAryPtr::BYTES);
4196
void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4197
assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4198
assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4199
// Capture src and dst memory (see comment in 'compress_string').
4200
Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4201
StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4202
set_memory(_gvn.transform(str), dst_type);
4205
void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4207
* int i_char = start;
4208
* for (int i_byte = 0; i_byte < count; i_byte++) {
4209
* dst[i_char++] = (char)(src[i_byte] & 0xff);
4212
add_parse_predicates();
4213
C->set_has_loops(true);
4215
RegionNode* head = new RegionNode(3);
4216
head->init_req(1, control());
4217
gvn().set_type(head, Type::CONTROL);
4218
record_for_igvn(head);
4220
Node* i_byte = new PhiNode(head, TypeInt::INT);
4221
i_byte->init_req(1, intcon(0));
4222
gvn().set_type(i_byte, TypeInt::INT);
4223
record_for_igvn(i_byte);
4225
Node* i_char = new PhiNode(head, TypeInt::INT);
4226
i_char->init_req(1, start);
4227
gvn().set_type(i_char, TypeInt::INT);
4228
record_for_igvn(i_char);
4230
Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4231
gvn().set_type(mem, Type::MEMORY);
4232
record_for_igvn(mem);
4234
set_memory(mem, TypeAryPtr::BYTES);
4235
Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4236
Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4237
AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4238
false, false, true /* mismatched */);
4240
IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4241
head->init_req(2, IfTrue(iff));
4242
mem->init_req(2, st);
4243
i_byte->init_req(2, AddI(i_byte, intcon(1)));
4244
i_char->init_req(2, AddI(i_char, intcon(2)));
4246
set_control(IfFalse(iff));
4247
set_memory(st, TypeAryPtr::BYTES);
4250
Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4251
if (!field->is_constant()) {
4252
return nullptr; // Field not marked as constant.
4254
ciInstance* holder = nullptr;
4255
if (!field->is_static()) {
4256
ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4257
if (const_oop != nullptr && const_oop->is_instance()) {
4258
holder = const_oop->as_instance();
4261
const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4262
/*is_unsigned_load=*/false);
4263
if (con_type != nullptr) {
4264
return makecon(con_type);
4269
Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4270
const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr();
4271
const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4272
if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4273
const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4274
Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));