1 /*
2 * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2024, Alibaba Group Holding Limited. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciInstanceKlass.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "classfile/systemDictionary.hpp"
31 #include "classfile/vmIntrinsics.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "gc/shared/tlab_globals.hpp"
36 #include "memory/allocation.inline.hpp"
37 #include "memory/resourceArea.hpp"
38 #include "oops/flatArrayKlass.hpp"
39 #include "oops/objArrayKlass.hpp"
40 #include "opto/addnode.hpp"
41 #include "opto/arraycopynode.hpp"
42 #include "opto/callnode.hpp"
43 #include "opto/cfgnode.hpp"
44 #include "opto/compile.hpp"
45 #include "opto/connode.hpp"
46 #include "opto/convertnode.hpp"
47 #include "opto/inlinetypenode.hpp"
48 #include "opto/loopnode.hpp"
49 #include "opto/machnode.hpp"
50 #include "opto/matcher.hpp"
51 #include "opto/memnode.hpp"
52 #include "opto/mempointer.hpp"
53 #include "opto/mulnode.hpp"
54 #include "opto/narrowptrnode.hpp"
55 #include "opto/opcodes.hpp"
56 #include "opto/phaseX.hpp"
57 #include "opto/regalloc.hpp"
58 #include "opto/regmask.hpp"
59 #include "opto/rootnode.hpp"
60 #include "opto/traceMergeStoresTag.hpp"
61 #include "opto/vectornode.hpp"
62 #include "runtime/arguments.hpp"
63 #include "utilities/align.hpp"
64 #include "utilities/copy.hpp"
65 #include "utilities/globalDefinitions.hpp"
66 #include "utilities/macros.hpp"
67 #include "utilities/powerOfTwo.hpp"
68 #include "utilities/vmError.hpp"
69
70 // Portions of code courtesy of Clifford Click
71
72 // Optimization - Graph Style
73
74 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
75
76 //=============================================================================
77 uint MemNode::size_of() const { return sizeof(*this); }
78
79 const TypePtr *MemNode::adr_type() const {
80 Node* adr = in(Address);
81 if (adr == nullptr) return nullptr; // node is dead
82 const TypePtr* cross_check = nullptr;
83 DEBUG_ONLY(cross_check = _adr_type);
84 return calculate_adr_type(adr->bottom_type(), cross_check);
85 }
86
87 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
88 if (adr != nullptr) {
89 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
90 return true;
91 }
92 }
93 return false;
94 }
95
96 #ifndef PRODUCT
97 void MemNode::dump_spec(outputStream *st) const {
98 if (in(Address) == nullptr) {
99 // node is dead
100 return;
101 }
102 #ifndef ASSERT
103 // fake the missing field
104 const TypePtr* _adr_type = in(Address)->bottom_type()->isa_ptr();
105 #endif
106 dump_adr_type(_adr_type, st);
107
108 Compile* C = Compile::current();
109 if (C->alias_type(_adr_type)->is_volatile()) {
110 st->print(" Volatile!");
111 }
112 if (_unaligned_access) {
113 st->print(" unaligned");
114 }
115 if (_mismatched_access) {
116 st->print(" mismatched");
117 }
118 if (_unsafe_access) {
119 st->print(" unsafe");
120 }
121 st->print(" barrier(0x%x)", _barrier_data);
122 }
123
124 void MemNode::dump_adr_type(const TypePtr* adr_type, outputStream* st) {
125 st->print(" @");
126 if (adr_type == nullptr) {
127 st->print("null");
128 } else {
129 adr_type->dump_on(st);
130 Compile* C = Compile::current();
131 Compile::AliasType* atp = nullptr;
132 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
133 if (atp == nullptr)
134 st->print(", idx=?\?;");
135 else if (atp->index() == Compile::AliasIdxBot)
136 st->print(", idx=Bot;");
137 else if (atp->index() == Compile::AliasIdxTop)
138 st->print(", idx=Top;");
139 else if (atp->index() == Compile::AliasIdxRaw)
140 st->print(", idx=Raw;");
141 else {
142 ciField* field = atp->field();
143 if (field) {
144 st->print(", name=");
145 field->print_name_on(st);
146 }
147 st->print(", idx=%d;", atp->index());
148 }
149 }
150 }
151
152 extern void print_alias_types();
153
154 #endif
155
156 // Find the memory output corresponding to the fall-through path of a call
157 static Node* find_call_fallthrough_mem_output(CallNode* call) {
158 ResourceMark rm;
159 CallProjections* projs = call->extract_projections(false, false);
160 Node* res = projs->fallthrough_memproj;
161 assert(res != nullptr, "must have a fallthrough mem output");
162 return res;
163 }
164
165 // Try to find a better memory input for a load from a strict final field
166 static Node* try_optimize_strict_final_load_memory(PhaseGVN* phase, Node* adr, ProjNode*& base_local) {
167 intptr_t offset = 0;
168 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
169 if (base == nullptr) {
170 return nullptr;
171 }
172
173 Node* base_uncasted = base->uncast();
174 if (base_uncasted->is_Proj()) {
175 Node* multi = base_uncasted->in(0);
176 if (multi->is_top()) {
177 // The pointer dies, make the memory die, too
178 return multi;
179 } else if (multi->is_Allocate()) {
180 base_local = base_uncasted->as_Proj();
181 return nullptr;
182 } else if (multi->is_Call()) {
183 if (!multi->is_CallJava() || multi->as_CallJava()->method() == nullptr || !multi->as_CallJava()->method()->return_value_is_larval()) {
184 // The oop is returned from a call, the memory can be the fallthrough output of the call
185 return find_call_fallthrough_mem_output(multi->as_Call());
186 }
187 } else if (multi->is_Start()) {
188 // The oop is a parameter
189 if (base_uncasted->as_Proj()->_con == TypeFunc::Parms && phase->C->method()->receiver_maybe_larval()) {
190 // The receiver of a constructor is similar to the result of an AllocateNode
191 base_local = base_uncasted->as_Proj();
192 return nullptr;
193 } else {
194 // Use the start memory otherwise
195 return multi->as_Start()->proj_out(TypeFunc::Memory);
196 }
197 }
198 }
199
200 return nullptr;
201 }
202
203 // Whether a call can modify a strict final field, given that the object is allocated inside the
204 // current compilation unit, or is the first parameter when the compilation root is a constructor.
205 // This is equivalent to asking whether 'call' is a constructor invocation and the class declaring
206 // the target method is a subclass of the class declaring 'field'.
207 static bool call_can_modify_local_object(ciField* field, CallNode* call) {
208 if (!call->is_CallJava()) {
209 return false;
210 }
211
212 ciMethod* target = call->as_CallJava()->method();
213 if (target == nullptr) {
214 return false;
215 } else if (target->intrinsic_id() == vmIntrinsicID::_linkToSpecial) {
216 // linkToSpecial can be used to call a constructor, used in the construction of objects in the
217 // reflection API
218 return true;
219 } else if (!target->is_object_constructor()) {
220 return false;
221 }
222
223 // If 'field' is declared in a class that is a subclass of the one declaring the constructor,
224 // then the field is set inside the constructor, else the field must be set before the
225 // constructor invocation. E.g. A field Super.x will be set during the execution of Sub::<init>,
226 // while a field Sub.y must be set before Super::<init> is invoked.
227 // We can try to be more heroic and decide if the receiver of the constructor invocation is the
228 // object from which we are loading from. This, however, may be problematic as deciding if 2
229 // nodes are definitely different may not be trivial, especially if the graph is not canonical.
230 // As a result, it is made more conservative for now.
231 assert(call->req() > TypeFunc::Parms, "constructor must have at least 1 argument");
232 return target->holder()->is_subclass_of(field->holder());
233 }
234
235 Node* MemNode::optimize_simple_memory_chain(Node* mchain, const TypeOopPtr* t_oop, Node* load, PhaseGVN* phase) {
236 assert(t_oop != nullptr, "sanity");
237 bool is_known_instance = t_oop->is_known_instance_field();
238 bool is_strict_final_load = false;
239
240 // After macro expansion, an allocation may become a call, changing the memory input to the
241 // memory output of that call would be illegal. As a result, disallow this transformation after
242 // macro expansion.
243 if (phase->is_IterGVN() && phase->C->allow_macro_nodes() && load != nullptr && load->is_Load() && !load->as_Load()->is_mismatched_access()) {
244 is_strict_final_load = t_oop->is_ptr_to_strict_final_field();
245 #ifdef ASSERT
246 if ((t_oop->is_inlinetypeptr() && t_oop->inline_klass()->contains_field_offset(t_oop->offset())) || t_oop->is_ptr_to_boxed_value()) {
247 assert(is_strict_final_load, "sanity check for basic cases");
248 }
249 #endif // ASSERT
250 }
251
252 if (!is_known_instance && !is_strict_final_load) {
253 return mchain;
254 }
255
256 Node* result = mchain;
257 ProjNode* base_local = nullptr;
258
259 ciField* field = nullptr;
260 if (is_strict_final_load) {
261 field = phase->C->alias_type(t_oop)->field();
262 assert(field != nullptr, "must point to a field");
263
264 Node* adr = load->in(MemNode::Address);
265 assert(phase->type(adr) == t_oop, "inconsistent type");
266 Node* tmp = try_optimize_strict_final_load_memory(phase, adr, base_local);
267 if (tmp != nullptr) {
268 result = tmp;
269 }
270 }
271
272 uint instance_id = t_oop->instance_id();
273 Node* start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
274 Node* prev = nullptr;
275 while (prev != result) {
276 prev = result;
277 if (result == start_mem) {
278 // start_mem is the earliest memory possible
279 break;
280 }
281
282 // skip over a call which does not affect this memory slice
283 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
284 Node* proj_in = result->in(0);
285 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
286 // This is the allocation that creates the object from which we are loading from
287 break;
288 } else if (proj_in->is_Call()) {
289 // ArrayCopyNodes processed here as well
290 CallNode* call = proj_in->as_Call();
291 if (!call->may_modify(t_oop, phase)) {
292 result = call->in(TypeFunc::Memory);
293 } else if (is_strict_final_load && base_local != nullptr && !call_can_modify_local_object(field, call)) {
294 result = call->in(TypeFunc::Memory);
295 }
296 } else if (proj_in->Opcode() == Op_Tuple) {
297 // The call will be folded, skip over it.
298 break;
299 } else if (proj_in->is_Initialize()) {
300 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
301 // Stop if this is the initialization for the object instance which
302 // contains this memory slice, otherwise skip over it.
303 if ((alloc == nullptr) || (alloc->_idx == instance_id)) {
304 break;
305 }
306 if (is_known_instance) {
307 result = proj_in->in(TypeFunc::Memory);
308 } else if (is_strict_final_load) {
309 Node* klass = alloc->in(AllocateNode::KlassNode);
310 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
311 if (tklass->klass_is_exact() && !tklass->exact_klass()->is_subclass_of(t_oop->is_instptr()->instance_klass())) {
312 // Allocation of an unrelated type, must be another object
313 result = proj_in->in(TypeFunc::Memory);
314 } else if (base_local != nullptr && (base_local->is_Parm() || base_local->in(0) != alloc)) {
315 // Allocation of another object
316 result = proj_in->in(TypeFunc::Memory);
317 }
318 }
319 } else if (proj_in->is_MemBar()) {
320 ArrayCopyNode* ac = nullptr;
321 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
322 break;
323 }
324 result = proj_in->in(TypeFunc::Memory);
325 } else if (proj_in->is_LoadFlat() || proj_in->is_StoreFlat()) {
326 bool mismatched = proj_in->is_LoadFlat() ? proj_in->as_LoadFlat()->is_mismatched() : proj_in->as_StoreFlat()->is_mismatched();
327 if (is_strict_final_load || (is_known_instance && !mismatched)) {
328 // LoadFlat and StoreFlat cannot happen to strict final fields
329 // LoadFlat and StoreFlat to known instances are removed at the end of EA unless mismatched: this one is unrelated
330 result = proj_in->in(TypeFunc::Memory);
331 }
332 } else if (proj_in->is_top()) {
333 break; // dead code
334 } else {
335 assert(false, "unexpected projection of %s", proj_in->Name());
336 }
337 } else if (result->is_ClearArray()) {
338 if (!is_known_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
339 // Can not bypass initialization of the instance
340 // we are looking for.
341 break;
342 }
343 // Otherwise skip it (the call updated 'result' value).
344 } else if (result->is_MergeMem()) {
345 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty);
346 }
347 }
348 return result;
349 }
350
351 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
352 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
353 if (t_oop == nullptr)
354 return mchain; // don't try to optimize non-oop types
355 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
356 bool is_instance = t_oop->is_known_instance_field();
357 PhaseIterGVN *igvn = phase->is_IterGVN();
358 if (is_instance && igvn != nullptr && result->is_Phi()) {
359 PhiNode *mphi = result->as_Phi();
360 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
361 const TypePtr *t = mphi->adr_type();
362 bool do_split = false;
363 // In the following cases, Load memory input can be further optimized based on
364 // its precise address type
365 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
366 do_split = true;
367 } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
368 const TypeOopPtr* mem_t =
369 t->is_oopptr()->cast_to_exactness(true)
370 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
371 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
372 if (t_oop->isa_aryptr()) {
373 mem_t = mem_t->is_aryptr()
374 ->cast_to_stable(t_oop->is_aryptr()->is_stable())
375 ->cast_to_size(t_oop->is_aryptr()->size())
376 ->cast_to_not_flat(t_oop->is_aryptr()->is_not_flat())
377 ->cast_to_not_null_free(t_oop->is_aryptr()->is_not_null_free())
378 ->with_offset(t_oop->is_aryptr()->offset())
379 ->is_aryptr();
380 }
381 do_split = mem_t == t_oop;
382 }
383 if (do_split) {
384 // clone the Phi with our address type
385 result = mphi->split_out_instance(t_adr, igvn);
386 } else {
387 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
388 }
389 }
390 return result;
391 }
392
393 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
394 uint alias_idx = phase->C->get_alias_index(tp);
395 Node *mem = mmem;
396 #ifdef ASSERT
397 {
398 // Check that current type is consistent with the alias index used during graph construction
399 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
400 bool consistent = adr_check == nullptr || adr_check->empty() ||
401 phase->C->must_alias(adr_check, alias_idx );
402 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
403 if( !consistent && adr_check != nullptr && !adr_check->empty() &&
404 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
405 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
406 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
407 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
408 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
409 // don't assert if it is dead code.
410 consistent = true;
411 }
412 if( !consistent ) {
413 st->print("alias_idx==%d, adr_check==", alias_idx);
414 if( adr_check == nullptr ) {
415 st->print("null");
416 } else {
417 adr_check->dump();
418 }
419 st->cr();
420 print_alias_types();
421 assert(consistent, "adr_check must match alias idx");
422 }
423 }
424 #endif
425 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
426 // means an array I have not precisely typed yet. Do not do any
427 // alias stuff with it any time soon.
428 const TypeOopPtr *toop = tp->isa_oopptr();
429 if (tp->base() != Type::AnyPtr &&
430 !(toop &&
431 toop->isa_instptr() &&
432 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
433 toop->offset() == Type::OffsetBot)) {
434 // IGVN _delay_transform may be set to true and if that is the case and mmem
435 // is already a registered node then the validation inside transform will
436 // complain.
437 Node* m = mmem;
438 PhaseIterGVN* igvn = phase->is_IterGVN();
439 if (igvn == nullptr || !igvn->delay_transform()) {
440 // compress paths and change unreachable cycles to TOP
441 // If not, we can update the input infinitely along a MergeMem cycle
442 // Equivalent code in PhiNode::Ideal
443 m = phase->transform(mmem);
444 }
445 // If transformed to a MergeMem, get the desired slice
446 // Otherwise the returned node represents memory for every slice
447 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
448 // Update input if it is progress over what we have now
449 }
450 return mem;
451 }
452
453 //--------------------------Ideal_common---------------------------------------
454 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
455 // Unhook non-raw memories from complete (macro-expanded) initializations.
456 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
457 // If our control input is a dead region, kill all below the region
458 Node *ctl = in(MemNode::Control);
459 if (ctl && remove_dead_region(phase, can_reshape))
460 return this;
461 ctl = in(MemNode::Control);
462 // Don't bother trying to transform a dead node
463 if (ctl && ctl->is_top()) return NodeSentinel;
464
465 PhaseIterGVN *igvn = phase->is_IterGVN();
466 // Wait if control on the worklist.
467 if (ctl && can_reshape && igvn != nullptr) {
468 Node* bol = nullptr;
469 Node* cmp = nullptr;
470 if (ctl->in(0)->is_If()) {
471 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
472 bol = ctl->in(0)->in(1);
473 if (bol->is_Bool())
474 cmp = ctl->in(0)->in(1)->in(1);
475 }
476 if (igvn->_worklist.member(ctl) ||
477 (bol != nullptr && igvn->_worklist.member(bol)) ||
478 (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
479 // This control path may be dead.
480 // Delay this memory node transformation until the control is processed.
481 igvn->_worklist.push(this);
482 return NodeSentinel; // caller will return null
483 }
484 }
485 // Ignore if memory is dead, or self-loop
486 Node *mem = in(MemNode::Memory);
487 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
488 assert(mem != this, "dead loop in MemNode::Ideal");
489
490 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
491 // This memory slice may be dead.
492 // Delay this mem node transformation until the memory is processed.
493 igvn->_worklist.push(this);
494 return NodeSentinel; // caller will return null
495 }
496
497 Node *address = in(MemNode::Address);
498 const Type *t_adr = phase->type(address);
499 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null
500
501 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
502 // Unsafe off-heap access with zero address. Remove access and other control users
503 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
504 assert(ctl != nullptr, "unsafe accesses should be control dependent");
505 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
506 Node* u = ctl->fast_out(i);
507 if (u != ctl) {
508 igvn->rehash_node_delayed(u);
509 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
510 --i, imax -= nb;
511 }
512 }
513 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
514 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
515 phase->C->root()->add_req(halt);
516 return this;
517 }
518
519 if (can_reshape && igvn != nullptr &&
520 (igvn->_worklist.member(address) ||
521 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
522 // The address's base and type may change when the address is processed.
523 // Delay this mem node transformation until the address is processed.
524 igvn->_worklist.push(this);
525 return NodeSentinel; // caller will return null
526 }
527
528 // Do NOT remove or optimize the next lines: ensure a new alias index
529 // is allocated for an oop pointer type before Escape Analysis.
530 // Note: C++ will not remove it since the call has side effect.
531 if (t_adr->isa_oopptr()) {
532 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
533 }
534
535 Node* base = nullptr;
536 if (address->is_AddP()) {
537 base = address->in(AddPNode::Base);
538 }
539 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
540 !t_adr->isa_rawptr()) {
541 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
542 // Skip this node optimization if its address has TOP base.
543 return NodeSentinel; // caller will return null
544 }
545
546 // Avoid independent memory operations
547 Node* old_mem = mem;
548
549 // The code which unhooks non-raw memories from complete (macro-expanded)
550 // initializations was removed. After macro-expansion all stores caught
551 // by Initialize node became raw stores and there is no information
552 // which memory slices they modify. So it is unsafe to move any memory
553 // operation above these stores. Also in most cases hooked non-raw memories
554 // were already unhooked by using information from detect_ptr_independence()
555 // and find_previous_store().
556
557 if (mem->is_MergeMem()) {
558 MergeMemNode* mmem = mem->as_MergeMem();
559 const TypePtr *tp = t_adr->is_ptr();
560
561 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
562 }
563
564 if (mem != old_mem) {
565 set_req_X(MemNode::Memory, mem, phase);
566 if (phase->type(mem) == Type::TOP) return NodeSentinel;
567 return this;
568 }
569
570 // let the subclass continue analyzing...
571 return nullptr;
572 }
573
574 // Helper function for proving some simple control dominations.
575 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
576 // Already assumes that 'dom' is available at 'sub', and that 'sub'
577 // is not a constant (dominated by the method's StartNode).
578 // Used by MemNode::find_previous_store to prove that the
579 // control input of a memory operation predates (dominates)
580 // an allocation it wants to look past.
581 // Returns 'DomResult::Dominate' if all control inputs of 'dom'
582 // dominate 'sub', 'DomResult::NotDominate' if not,
583 // and 'DomResult::EncounteredDeadCode' if we can't decide due to
584 // dead code, but at the end of IGVN, we know the definite result
585 // once the dead code is cleaned up.
586 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub, PhaseGVN* phase) {
587 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) {
588 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
589 }
590
591 // Check 'dom'. Skip Proj and CatchProj nodes.
592 dom = dom->find_exact_control(dom);
593 if (dom == nullptr || dom->is_top()) {
594 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
595 }
596
597 if (dom == sub) {
598 // For the case when, for example, 'sub' is Initialize and the original
599 // 'dom' is Proj node of the 'sub'.
600 return DomResult::NotDominate;
601 }
602
603 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) {
604 return DomResult::Dominate;
605 }
606
607 // 'dom' dominates 'sub' if its control edge and control edges
608 // of all its inputs dominate or equal to sub's control edge.
609
610 // Currently 'sub' is either Allocate, Initialize or Start nodes.
611 // Or Region for the check in LoadNode::Ideal();
612 // 'sub' should have sub->in(0) != nullptr.
613 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
614 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
615
616 // Get control edge of 'sub'.
617 Node* orig_sub = sub;
618 sub = sub->find_exact_control(sub->in(0));
619 if (sub == nullptr || sub->is_top()) {
620 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
621 }
622
623 assert(sub->is_CFG(), "expecting control");
624
625 if (sub == dom) {
626 return DomResult::Dominate;
627 }
628
629 if (sub->is_Start() || sub->is_Root()) {
630 return DomResult::NotDominate;
631 }
632
633 {
634 // Check all control edges of 'dom'.
635
636 ResourceMark rm;
637 Node_List nlist;
638 Unique_Node_List dom_list;
639
640 dom_list.push(dom);
641 bool only_dominating_controls = false;
642
643 for (uint next = 0; next < dom_list.size(); next++) {
644 Node* n = dom_list.at(next);
645 if (n == orig_sub) {
646 return DomResult::NotDominate; // One of dom's inputs dominated by sub.
647 }
648 if (!n->is_CFG() && n->pinned()) {
649 // Check only own control edge for pinned non-control nodes.
650 n = n->find_exact_control(n->in(0));
651 if (n == nullptr || n->is_top()) {
652 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
653 }
654 assert(n->is_CFG(), "expecting control");
655 dom_list.push(n);
656 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
657 only_dominating_controls = true;
658 } else if (n->is_CFG()) {
659 DomResult dom_result = n->dominates(sub, nlist);
660 if (dom_result == DomResult::Dominate) {
661 only_dominating_controls = true;
662 } else {
663 return dom_result;
664 }
665 } else {
666 if (n->Value(phase) == Type::TOP) {
667 return DomResult::EncounteredDeadCode;
668 }
669 // First, own control edge.
670 Node* m = n->find_exact_control(n->in(0));
671 if (m != nullptr) {
672 if (m->is_top()) {
673 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
674 }
675 dom_list.push(m);
676 }
677 // Now, the rest of edges.
678 uint cnt = n->req();
679 for (uint i = 1; i < cnt; i++) {
680 m = n->find_exact_control(n->in(i));
681 if (m == nullptr || m->is_top()) {
682 continue;
683 }
684 dom_list.push(m);
685 }
686 }
687 }
688 return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate;
689 }
690 }
691
692 //---------------------detect_ptr_independence---------------------------------
693 // Used by MemNode::find_previous_store to prove that two base
694 // pointers are never equal.
695 // The pointers are accompanied by their associated allocations,
696 // if any, which have been previously discovered by the caller.
697 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
698 Node* p2, AllocateNode* a2,
699 PhaseGVN* phase) {
700 // Trivial case: Non-overlapping values. Be careful, we can cast a raw pointer to an oop (e.g. in
701 // the allocation pattern) so joining the types only works if both are oops. join may also give
702 // an incorrect result when both pointers are nullable and the result is supposed to be
703 // TypePtr::NULL_PTR, so we exclude that case.
704 const Type* p1_type = p1->bottom_type();
705 const Type* p2_type = p2->bottom_type();
706 if (p1_type != p2_type && p1_type->isa_oopptr() && p2_type->isa_oopptr() &&
707 (!p1_type->maybe_null() || !p2_type->maybe_null()) &&
708 p1_type->join(p2_type)->empty()) {
709 return true;
710 }
711
712 // Attempt to prove that these two pointers cannot be aliased.
713 // They may both manifestly be allocations, and they should differ.
714 // Or, if they are not both allocations, they can be distinct constants.
715 // Otherwise, one is an allocation and the other a pre-existing value.
716 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
717 return (p1 != p2) && p1->is_Con() && p2->is_Con();
718 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
719 return (a1 != a2);
720 } else if (a1 != nullptr) { // one allocation a1
721 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
722 return all_controls_dominate(p2->uncast(), a1, phase);
723 } else { //(a2 != null) // one allocation a2
724 return all_controls_dominate(p1->uncast(), a2, phase);
725 }
726 return false;
727 }
728
729 // Find an arraycopy ac that produces the memory state represented by parameter mem.
730 // Return ac if
731 // (a) can_see_stored_value=true and ac must have set the value for this load or if
732 // (b) can_see_stored_value=false and ac could have set the value for this load or if
733 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
734 // In case (c) change the parameter mem to the memory input of ac to skip it
735 // when searching stored value.
736 // Otherwise return null.
737 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
738 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
739 if (ac != nullptr) {
740 Node* ld_addp = in(MemNode::Address);
741 Node* src = ac->in(ArrayCopyNode::Src);
742 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
743
744 // This is a load from a cloned array. The corresponding arraycopy ac must
745 // have set the value for the load and we can return ac but only if the load
746 // is known to be within bounds. This is checked below.
747 // TODO 8350865: Support flat arrays in LoadNode::find_previous_arraycopy
748 if (ary_t != nullptr && ary_t->is_not_flat() && ld_addp->is_AddP()) {
749 Node* ld_offs = ld_addp->in(AddPNode::Offset);
750 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
751 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
752 jlong elemsize = type2aelembytes(ary_elem);
753
754 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
755 const TypeInt* sizetype = ary_t->size();
756
757 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
758 // The load is known to be within bounds. It receives its value from ac.
759 return ac;
760 }
761 // The load is known to be out-of-bounds.
762 }
763 // The load could be out-of-bounds. It must not be hoisted but must remain
764 // dependent on the runtime range check. This is achieved by returning null.
765 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
766 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
767
768 if (ac->is_arraycopy_validated() ||
769 ac->is_copyof_validated() ||
770 ac->is_copyofrange_validated()) {
771 Node* ld_addp = in(MemNode::Address);
772 if (ld_addp->is_AddP()) {
773 Node* ld_base = ld_addp->in(AddPNode::Address);
774 Node* ld_offs = ld_addp->in(AddPNode::Offset);
775
776 Node* dest = ac->in(ArrayCopyNode::Dest);
777
778 if (dest == ld_base) {
779 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
780 assert(!ld_offs_t->empty(), "dead reference should be checked already");
781 // Take into account vector or unsafe access size
782 jlong ld_size_in_bytes = (jlong)memory_size();
783 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
784 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
785 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
786 return ac;
787 }
788 if (!can_see_stored_value) {
789 mem = ac->in(TypeFunc::Memory);
790 return ac;
791 }
792 }
793 }
794 }
795 }
796 return nullptr;
797 }
798
799 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
800 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
801 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
802 if (ld_alloc != nullptr) {
803 // Check if there is an array copy for a clone
804 Node* mb = mem->in(0);
805 ArrayCopyNode* ac = nullptr;
806 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
807 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
808 ac = mb->in(0)->in(0)->as_ArrayCopy();
809 } else {
810 // Step over GC barrier when ReduceInitialCardMarks is disabled
811 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
812 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
813
814 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
815 ac = control_proj_ac->in(0)->as_ArrayCopy();
816 }
817 }
818
819 if (ac != nullptr && ac->is_clonebasic()) {
820 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
821 if (alloc != nullptr && alloc == ld_alloc) {
822 return ac;
823 }
824 }
825 }
826 }
827 return nullptr;
828 }
829
830 // The logic for reordering loads and stores uses four steps:
831 // (a) Walk carefully past stores and initializations which we
832 // can prove are independent of this load.
833 // (b) Observe that the next memory state makes an exact match
834 // with self (load or store), and locate the relevant store.
835 // (c) Ensure that, if we were to wire self directly to the store,
836 // the optimizer would fold it up somehow.
837 // (d) Do the rewiring, and return, depending on some other part of
838 // the optimizer to fold up the load.
839 // This routine handles steps (a) and (b). Steps (c) and (d) are
840 // specific to loads and stores, so they are handled by the callers.
841 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
842 //
843 Node* MemNode::find_previous_store(PhaseGVN* phase) {
844 AccessAnalyzer analyzer(phase, this);
845
846 Node* mem = in(MemNode::Memory); // start searching here...
847 int cnt = 50; // Cycle limiter
848 for (;; cnt--) {
849 // While we can dance past unrelated stores...
850 if (phase->type(mem) == Type::TOP) {
851 // Encounter a dead node
852 return phase->C->top();
853 } else if (cnt <= 0) {
854 // Caught in cycle or a complicated dance?
855 return nullptr;
856 } else if (mem->is_Phi()) {
857 return nullptr;
858 }
859
860 AccessAnalyzer::AccessIndependence independence = analyzer.detect_access_independence(mem);
861 if (independence.independent) {
862 // (a) advance through the independent store
863 mem = independence.mem;
864 assert(mem != nullptr, "must not be nullptr");
865 } else {
866 // (b) found the store that this access observes if this is not null
867 // Otherwise, give up if it is null
868 return independence.mem;
869 }
870 }
871
872 return nullptr; // bail out
873 }
874
875 //----------------------calculate_adr_type-------------------------------------
876 // Helper function. Notices when the given type of address hits top or bottom.
877 // Also, asserts a cross-check of the type against the expected address type.
878 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
879 if (t == Type::TOP) return nullptr; // does not touch memory any more?
880 #ifdef ASSERT
881 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
882 #endif
883 const TypePtr* tp = t->isa_ptr();
884 if (tp == nullptr) {
885 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
886 return TypePtr::BOTTOM; // touches lots of memory
887 } else {
888 #ifdef ASSERT
889 // %%%% [phh] We don't check the alias index if cross_check is
890 // TypeRawPtr::BOTTOM. Needs to be investigated.
891 if (cross_check != nullptr &&
892 cross_check != TypePtr::BOTTOM &&
893 cross_check != TypeRawPtr::BOTTOM) {
894 // Recheck the alias index, to see if it has changed (due to a bug).
895 Compile* C = Compile::current();
896 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
897 "must stay in the original alias category");
898 // The type of the address must be contained in the adr_type,
899 // disregarding "null"-ness.
900 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
901 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
902 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
903 "real address must not escape from expected memory type");
904 }
905 #endif
906 return tp;
907 }
908 }
909
910 uint8_t MemNode::barrier_data(const Node* n) {
911 if (n->is_LoadStore()) {
912 return n->as_LoadStore()->barrier_data();
913 } else if (n->is_Mem()) {
914 return n->as_Mem()->barrier_data();
915 }
916 return 0;
917 }
918
919 AccessAnalyzer::AccessAnalyzer(PhaseGVN* phase, MemNode* n)
920 : _phase(phase), _n(n), _memory_size(n->memory_size()), _alias_idx(-1) {
921 Node* adr = _n->in(MemNode::Address);
922 _offset = 0;
923 _base = AddPNode::Ideal_base_and_offset(adr, _phase, _offset);
924 _maybe_raw = MemNode::check_if_adr_maybe_raw(adr);
925 _alloc = AllocateNode::Ideal_allocation(_base);
926 _adr_type = _n->adr_type();
927
928 if (_adr_type != nullptr && _adr_type->base() != TypePtr::AnyPtr) {
929 // Avoid the cases that will upset Compile::get_alias_index
930 _alias_idx = _phase->C->get_alias_index(_adr_type);
931 assert(_alias_idx != Compile::AliasIdxTop, "must not be a dead node");
932 assert(_alias_idx != Compile::AliasIdxBot || !phase->C->do_aliasing(), "must not be a very wide access");
933 }
934 }
935
936 // Decide whether the memory accessed by '_n' and 'other' may overlap. This function may be used
937 // when we want to walk the memory graph to fold a load, or when we want to hoist a load above a
938 // loop when there are no stores that may overlap with the load inside the loop.
939 AccessAnalyzer::AccessIndependence AccessAnalyzer::detect_access_independence(Node* other) const {
940 assert(_phase->type(other) == Type::MEMORY, "must be a memory node %s", other->Name());
941 assert(!other->is_Phi(), "caller must handle Phi");
942
943 if (_adr_type == nullptr) {
944 // This means the access is dead
945 return {false, _phase->C->top()};
946 } else if (_adr_type->base() == TypePtr::AnyPtr) {
947 // An example for this case is an access into the memory address 0 performed using Unsafe
948 assert(_adr_type->ptr() == TypePtr::Null, "MemNode should never access a wide memory");
949 return {false, nullptr};
950 }
951
952 if (_offset == Type::OffsetBot) {
953 // cannot unalias unless there are precise offsets
954 return {false, nullptr};
955 }
956
957 const TypeOopPtr* adr_oop_type = _adr_type->isa_oopptr();
958 Node* prev = other;
959 if (other->is_Store()) {
960 Node* st_adr = other->in(MemNode::Address);
961 intptr_t st_offset = 0;
962 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, _phase, st_offset);
963 if (st_base == nullptr) {
964 // inscrutable pointer
965 return {false, nullptr};
966 }
967
968 // If the bases are the same and the offsets are the same, it seems that this is the exact
969 // store we are looking for, the caller will check if the type of the store matches using
970 // MemNode::can_see_stored_value
971 if (st_base == _base && st_offset == _offset) {
972 return {false, other};
973 }
974
975 // If it is provable that the memory accessed by 'other' does not overlap the memory accessed
976 // by '_n', we may walk past 'other'.
977 // For raw accesses, 2 accesses are independent if they have the same base and the offsets
978 // say that they do not overlap.
979 // For heap accesses, 2 accesses are independent if either the bases are provably different
980 // at runtime or the offsets say that the accesses do not overlap.
981 if ((_maybe_raw || MemNode::check_if_adr_maybe_raw(st_adr)) && st_base != _base) {
982 // Raw accesses can only be provably independent if they have the same base
983 return {false, nullptr};
984 }
985
986 // If the offsets say that the accesses do not overlap, then it is provable that 'other' and
987 // '_n' do not overlap. For example, a LoadI from Object+8 is independent from a StoreL into
988 // Object+12, no matter what the bases are.
989 if (st_offset != _offset && st_offset != Type::OffsetBot) {
990 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
991 assert(other->as_Store()->memory_size() <= MAX_STORE, "");
992 if (st_offset >= _offset + _memory_size ||
993 st_offset <= _offset - MAX_STORE ||
994 st_offset <= _offset - other->as_Store()->memory_size()) {
995 // Success: The offsets are provably independent.
996 // (You may ask, why not just test st_offset != offset and be done?
997 // The answer is that stores of different sizes can co-exist
998 // in the same sequence of RawMem effects. We sometimes initialize
999 // a whole 'tile' of array elements with a single jint or jlong.)
1000 return {true, other->in(MemNode::Memory)};
1001 }
1002 }
1003
1004 // Same base and overlapping offsets, it seems provable that the accesses overlap, give up
1005 if (st_base == _base) {
1006 return {false, nullptr};
1007 }
1008
1009 // Try to prove that 2 different base nodes at compile time are different values at runtime
1010 bool known_independent = false;
1011 if (MemNode::detect_ptr_independence(_base, _alloc, st_base, AllocateNode::Ideal_allocation(st_base), _phase)) {
1012 known_independent = true;
1013 }
1014
1015 if (known_independent) {
1016 return {true, other->in(MemNode::Memory)};
1017 }
1018 } else if (other->is_Proj() && other->in(0)->is_Initialize()) {
1019 InitializeNode* st_init = other->in(0)->as_Initialize();
1020 AllocateNode* st_alloc = st_init->allocation();
1021 if (st_alloc == nullptr) {
1022 // Something degenerated
1023 return {false, nullptr};
1024 }
1025 bool known_identical = false;
1026 bool known_independent = false;
1027 if (_alloc == st_alloc) {
1028 known_identical = true;
1029 } else if (_alloc != nullptr) {
1030 known_independent = true;
1031 } else if (MemNode::all_controls_dominate(_base->uncast(), st_alloc, _phase)) {
1032 known_independent = true;
1033 }
1034
1035 if (known_independent) {
1036 // The bases are provably independent: Either they are
1037 // manifestly distinct allocations, or else the control
1038 // of _base dominates the store's allocation.
1039 if (_alias_idx == Compile::AliasIdxRaw) {
1040 other = st_alloc->in(TypeFunc::Memory);
1041 } else {
1042 other = st_init->memory(_alias_idx);
1043 }
1044 return {true, other};
1045 }
1046
1047 // If we are not looking at a store initializing the same
1048 // allocation we are loading from, we lose.
1049 if (known_identical) {
1050 // From caller, can_see_stored_value will consult find_captured_store.
1051 return {false, other};
1052 }
1053
1054 } else if (_n->find_previous_arraycopy(_phase, _alloc, other, false) != nullptr) {
1055 // Find an arraycopy that may or may not affect the MemNode
1056 return {prev != other, other};
1057 } else if (other->is_MergeMem()) {
1058 return {true, other->as_MergeMem()->memory_at(_alias_idx)};
1059 } else if (adr_oop_type != nullptr && adr_oop_type->is_known_instance_field()) {
1060 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
1061 if (other->is_Proj() && other->in(0)->is_Call()) {
1062 // ArrayCopyNodes processed here as well.
1063 CallNode* call = other->in(0)->as_Call();
1064 if (!call->may_modify(adr_oop_type, _phase)) {
1065 return {true, call->in(TypeFunc::Memory)};
1066 }
1067 } else if (other->is_Proj() && other->in(0)->is_MemBar()) {
1068 ArrayCopyNode* ac = nullptr;
1069 if (!ArrayCopyNode::may_modify(adr_oop_type, other->in(0)->as_MemBar(), _phase, ac)) {
1070 return {true, other->in(0)->in(TypeFunc::Memory)};
1071 }
1072 } else if (other->is_ClearArray()) {
1073 if (ClearArrayNode::step_through(&other, (uint)adr_oop_type->instance_id(), _phase)) {
1074 // (the call updated 'other' value)
1075 return {true, other};
1076 } else {
1077 // Can not bypass initialization of the instance
1078 // we are looking for.
1079 return {false, other};
1080 }
1081 }
1082 }
1083
1084 return {false, nullptr};
1085 }
1086
1087 //=============================================================================
1088 // Should LoadNode::Ideal() attempt to remove control edges?
1089 bool LoadNode::can_remove_control() const {
1090 return !has_pinned_control_dependency();
1091 }
1092 uint LoadNode::size_of() const { return sizeof(*this); }
1093 bool LoadNode::cmp(const Node &n) const {
1094 LoadNode& load = (LoadNode &)n;
1095 return Type::equals(_type, load._type) &&
1096 _control_dependency == load._control_dependency &&
1097 _mo == load._mo;
1098 }
1099 const Type *LoadNode::bottom_type() const { return _type; }
1100 uint LoadNode::ideal_reg() const {
1101 return _type->ideal_reg();
1102 }
1103
1104 #ifndef PRODUCT
1105 void LoadNode::dump_spec(outputStream *st) const {
1106 MemNode::dump_spec(st);
1107 if( !Verbose && !WizardMode ) {
1108 // standard dump does this in Verbose and WizardMode
1109 st->print(" #"); _type->dump_on(st);
1110 }
1111 if (in(0) != nullptr && !depends_only_on_test()) {
1112 st->print(" (does not depend only on test, ");
1113 if (control_dependency() == UnknownControl) {
1114 st->print("unknown control");
1115 } else if (control_dependency() == Pinned) {
1116 st->print("pinned");
1117 } else if (adr_type() == TypeRawPtr::BOTTOM) {
1118 st->print("raw access");
1119 } else {
1120 st->print("unknown reason");
1121 }
1122 st->print(")");
1123 }
1124 }
1125 #endif
1126
1127 #ifdef ASSERT
1128 //----------------------------is_immutable_value-------------------------------
1129 // Helper function to allow a raw load without control edge for some cases
1130 bool LoadNode::is_immutable_value(Node* adr) {
1131 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
1132 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
1133
1134 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
1135 int offsets[] = {
1136 in_bytes(JavaThread::osthread_offset()),
1137 in_bytes(JavaThread::threadObj_offset()),
1138 in_bytes(JavaThread::vthread_offset()),
1139 in_bytes(JavaThread::scopedValueCache_offset()),
1140 };
1141
1142 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
1143 if (offset == offsets[i]) {
1144 return true;
1145 }
1146 }
1147 }
1148
1149 return false;
1150 }
1151 #endif
1152
1153 //----------------------------LoadNode::make-----------------------------------
1154 // Polymorphic factory method:
1155 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
1156 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
1157 Compile* C = gvn.C;
1158 assert(adr->is_top() || C->get_alias_index(gvn.type(adr)->is_ptr(), true) == C->get_alias_index(adr_type, true), "adr and adr_type must agree");
1159
1160 // sanity check the alias category against the created node type
1161 assert(!(adr_type->isa_oopptr() &&
1162 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
1163 "use LoadKlassNode instead");
1164 assert(!(adr_type->isa_aryptr() &&
1165 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
1166 "use LoadRangeNode instead");
1167 // Check control edge of raw loads
1168 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
1169 // oop will be recorded in oop map if load crosses safepoint
1170 rt->isa_oopptr() || is_immutable_value(adr),
1171 "raw memory operations should have control edge");
1172 LoadNode* load = nullptr;
1173 switch (bt) {
1174 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1175 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1176 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1177 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1178 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1179 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
1180 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
1181 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
1182 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
1183 case T_ARRAY:
1184 case T_OBJECT:
1185 case T_NARROWOOP:
1186 #ifdef _LP64
1187 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1188 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
1189 } else
1190 #endif
1191 {
1192 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
1193 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
1194 }
1195 break;
1196 default:
1197 guarantee(false, "unexpected basic type %s", type2name(bt));
1198 break;
1199 }
1200 assert(load != nullptr, "LoadNode should have been created");
1201 if (unaligned) {
1202 load->set_unaligned_access();
1203 }
1204 if (mismatched) {
1205 load->set_mismatched_access();
1206 }
1207 if (unsafe) {
1208 load->set_unsafe_access();
1209 }
1210 load->set_barrier_data(barrier_data);
1211 if (load->Opcode() == Op_LoadN) {
1212 Node* ld = gvn.transform(load);
1213 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1214 }
1215
1216 return load;
1217 }
1218
1219 //------------------------------hash-------------------------------------------
1220 uint LoadNode::hash() const {
1221 // unroll addition of interesting fields
1222 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1223 }
1224
1225 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1226 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1227 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1228 bool is_stable_ary = FoldStableValues &&
1229 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1230 tp->isa_aryptr()->is_stable();
1231
1232 return (eliminate_boxing && non_volatile) || is_stable_ary || tp->is_inlinetypeptr();
1233 }
1234
1235 return false;
1236 }
1237
1238 // Is the value loaded previously stored by an arraycopy? If so return
1239 // a load node that reads from the source array so we may be able to
1240 // optimize out the ArrayCopy node later.
1241 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1242 Node* ld_adr = in(MemNode::Address);
1243 intptr_t ld_off = 0;
1244 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1245 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1246 if (ac != nullptr) {
1247 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1248
1249 Node* mem = ac->in(TypeFunc::Memory);
1250 Node* ctl = ac->in(0);
1251 Node* src = ac->in(ArrayCopyNode::Src);
1252
1253 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1254 return nullptr;
1255 }
1256
1257 // load depends on the tests that validate the arraycopy
1258 LoadNode* ld = clone_pinned();
1259 Node* addp = in(MemNode::Address)->clone();
1260 if (ac->as_ArrayCopy()->is_clonebasic()) {
1261 assert(ld_alloc != nullptr, "need an alloc");
1262 assert(addp->is_AddP(), "address must be addp");
1263 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1264 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1265 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1266 addp->set_req(AddPNode::Base, src);
1267 addp->set_req(AddPNode::Address, src);
1268 } else {
1269 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1270 ac->as_ArrayCopy()->is_copyof_validated() ||
1271 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1272 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1273 addp->set_req(AddPNode::Base, src);
1274 addp->set_req(AddPNode::Address, src);
1275
1276 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1277 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1278 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1279
1280 uint shift = ary_t->is_flat() ? ary_t->flat_log_elem_size() : exact_log2(type2aelembytes(ary_elem));
1281
1282 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1283 #ifdef _LP64
1284 diff = phase->transform(new ConvI2LNode(diff));
1285 #endif
1286 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1287
1288 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1289 addp->set_req(AddPNode::Offset, offset);
1290 }
1291 addp = phase->transform(addp);
1292 #ifdef ASSERT
1293 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1294 ld->_adr_type = adr_type;
1295 #endif
1296 ld->set_req(MemNode::Address, addp);
1297 ld->set_req(0, ctl);
1298 ld->set_req(MemNode::Memory, mem);
1299 return ld;
1300 }
1301 return nullptr;
1302 }
1303
1304 static Node* see_through_inline_type(PhaseValues* phase, const LoadNode* load, Node* base, int offset) {
1305 if (load->is_mismatched_access() || base == nullptr) {
1306 return nullptr;
1307 }
1308
1309 InlineTypeNode* vt = base->isa_InlineType();
1310 if (vt == nullptr || offset < vt->type()->inline_klass()->payload_offset()) {
1311 return nullptr;
1312 }
1313
1314 Node* value = vt->field_value_by_offset(offset, true);
1315 assert(value != nullptr, "must see some value");
1316 return value;
1317 }
1318
1319 // This routine exists to make sure this set of tests is done the same
1320 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1321 // will change the graph shape in a way which makes memory alive twice at the
1322 // same time (uses the Oracle model of aliasing), then some
1323 // LoadXNode::Identity will fold things back to the equivalence-class model
1324 // of aliasing.
1325 // This method may find an unencoded node instead of the corresponding encoded one.
1326 Node* LoadNode::can_see_stored_value_through_membars(Node* st, PhaseValues* phase) const {
1327 Node* ld_adr = in(MemNode::Address);
1328 intptr_t ld_off = 0;
1329 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1330 // Try to see through an InlineTypeNode
1331 Node* value = see_through_inline_type(phase, this, ld_base, ld_off);
1332 if (value != nullptr) {
1333 return value;
1334 }
1335
1336 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1337 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1338
1339 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1340 uint alias_idx = atp->index();
1341 Node* result = nullptr;
1342 Node* current = st;
1343 // Skip through chains of MemBarNodes checking the MergeMems for new states for the slice of
1344 // this load. Stop once any other kind of node is encountered.
1345 //
1346 // In principle, folding a load is moving it up until it meets a matching store.
1347 //
1348 // store(ptr, v); store(ptr, v); store(ptr, v);
1349 // membar1; -> membar1; -> load(ptr);
1350 // membar2; load(ptr); membar1;
1351 // load(ptr); membar2; membar2;
1352 //
1353 // So, we can decide which kinds of barriers we can walk past. It is not safe to step over
1354 // MemBarCPUOrder, even if the memory is not rewritable, because alias info above them may be
1355 // inaccurate (e.g., due to mixed/mismatched unsafe accesses).
1356 bool is_final_mem = !atp->is_rewritable();
1357 while (current->is_Proj()) {
1358 int opc = current->in(0)->Opcode();
1359 if ((is_final_mem && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock || opc == Op_LoadFence)) ||
1360 opc == Op_MemBarRelease ||
1361 opc == Op_StoreFence ||
1362 opc == Op_MemBarReleaseLock ||
1363 opc == Op_MemBarStoreStore ||
1364 opc == Op_StoreStoreFence) {
1365 Node* mem = current->in(0)->in(TypeFunc::Memory);
1366 if (mem->is_MergeMem()) {
1367 MergeMemNode* merge = mem->as_MergeMem();
1368 Node* new_st = merge->memory_at(alias_idx);
1369 if (new_st == merge->base_memory()) {
1370 // Keep searching
1371 current = new_st;
1372 continue;
1373 }
1374 // Save the new memory state for the slice and fall through
1375 // to exit.
1376 result = new_st;
1377 }
1378 }
1379 break;
1380 }
1381 if (result != nullptr) {
1382 st = result;
1383 }
1384 }
1385
1386 Node* res = can_see_stored_value(st, phase);
1387 // TODO: reimplement assert, see: JDK-8386157
1388 //assert(res == nullptr || is_java_primitive(value_basic_type()) || res->bottom_type()->higher_equal(type()), "the fold is unsafe");
1389 return res;
1390 }
1391
1392 // If st is a store to the same location as this, return the stored value
1393 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1394 Node* ld_adr = in(MemNode::Address);
1395 intptr_t ld_off = 0;
1396 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1397 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1398 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1399
1400 // Loop around twice in the case Load -> Initialize -> Store.
1401 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1402 for (int trip = 0; trip <= 1; trip++) {
1403
1404 if (st->is_Store()) {
1405 Node* st_adr = st->in(MemNode::Address);
1406 if (st_adr != ld_adr) {
1407 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1408 intptr_t st_off = 0;
1409 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1410 if (ld_base == nullptr) return nullptr;
1411 if (st_base == nullptr) return nullptr;
1412 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1413 if (ld_off != st_off) return nullptr;
1414 if (ld_off == Type::OffsetBot) return nullptr;
1415 // Same base, same offset.
1416 // Possible improvement for arrays: check index value instead of absolute offset.
1417
1418 // At this point we have proven something like this setup:
1419 // B = << base >>
1420 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1421 // S = StoreQ(AddP( B , #Off), V)
1422 // (Actually, we haven't yet proven the Q's are the same.)
1423 // In other words, we are loading from a casted version of
1424 // the same pointer-and-offset that we stored to.
1425 // Casted version may carry a dependency and it is respected.
1426 // Thus, we are able to replace L by V.
1427 }
1428 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1429 if (store_Opcode() != st->Opcode()) {
1430 return nullptr;
1431 }
1432 // LoadVector/StoreVector needs additional check to ensure the types match.
1433 if (st->is_StoreVector()) {
1434 if ((Opcode() != Op_LoadVector && Opcode() != Op_StoreVector) || st->Opcode() != Op_StoreVector) {
1435 // Some kind of masked access or gather/scatter
1436 return nullptr;
1437 }
1438
1439 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1440 const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1441 if (in_vt != out_vt) {
1442 return nullptr;
1443 }
1444 }
1445
1446 // Even if we can see the store, we cannot fold the load if the store is not type safe (e.g.
1447 // store a j.l.Object into an array of j.l.String) because folding makes the compiler lose the
1448 // type information that the uses of this node may need. This is only necessary for pointers, we
1449 // can see the stored value of a LoadS even if it is an int because LoadSNode::Ideal will do the
1450 // necessary truncation.
1451 // The same phenomenon is not an issue for StoreNodes because they don't use res.
1452 Node* res = st->in(MemNode::ValueIn);
1453 if (is_Store() || is_java_primitive(value_basic_type()) || res->bottom_type()->higher_equal(bottom_type())) {
1454 return res;
1455 }
1456
1457 // There are some cases in which the Type of the load is narrower than the Type of the value
1458 // that is stored into that location. The most common case is array polymorphism, when the
1459 // type of an array element depends on the type of the array. In addition, there are some
1460 // corner cases, the first one is concurrent class loading, when CHA can result in a narrower
1461 // Type than what is declared only after the child class is loaded, and the second case is
1462 // unsafe accesses when we do not check for type safety. See JDK-8388184.
1463 return nullptr;
1464 }
1465
1466 // A load from a freshly-created object always returns zero.
1467 // (This can happen after LoadNode::Ideal resets the load's memory input
1468 // to find_captured_store, which returned InitializeNode::zero_memory.)
1469 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1470 (st->in(0) == ld_alloc) &&
1471 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1472 // return a zero value for the load's basic type
1473 // (This is one of the few places where a generic PhaseTransform
1474 // can create new nodes. Think of it as lazily manifesting
1475 // virtually pre-existing constants.)
1476 Node* init_value = ld_alloc->in(AllocateNode::InitValue);
1477 if (init_value != nullptr) {
1478 const TypeAryPtr* ld_adr_type = phase->type(ld_adr)->isa_aryptr();
1479 if (ld_adr_type == nullptr) {
1480 return nullptr;
1481 }
1482
1483 // We know that this is not a flat array, the load should return the whole oop
1484 if (ld_adr_type->is_not_flat()) {
1485 return init_value;
1486 }
1487
1488 // If this is a flat array, try to see through init_value
1489 if (init_value->is_EncodeP()) {
1490 init_value = init_value->in(1);
1491 }
1492 if (!init_value->is_InlineType() || ld_adr_type->field_offset() == Type::Offset::bottom) {
1493 return nullptr;
1494 }
1495
1496 ciInlineKlass* vk = phase->type(init_value)->inline_klass();
1497 int field_offset_in_payload = ld_adr_type->field_offset().get();
1498 if (field_offset_in_payload == vk->null_marker_offset_in_payload()) {
1499 return init_value->as_InlineType()->get_null_marker();
1500 } else {
1501 return init_value->as_InlineType()->field_value_by_offset(field_offset_in_payload + vk->payload_offset(), true);
1502 }
1503 }
1504 assert(ld_alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
1505 if (value_basic_type() != T_VOID) {
1506 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1507 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1508 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1509 // by the ArrayCopyNode.
1510 return phase->zerocon(value_basic_type());
1511 }
1512 } else {
1513 // TODO: materialize all-zero vector constant
1514 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1515 }
1516 }
1517
1518 // A load from an initialization barrier can match a captured store.
1519 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1520 InitializeNode* init = st->in(0)->as_Initialize();
1521 AllocateNode* alloc = init->allocation();
1522 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1523 // examine a captured store value
1524 st = init->find_captured_store(ld_off, memory_size(), phase);
1525 if (st != nullptr) {
1526 continue; // take one more trip around
1527 }
1528 }
1529 }
1530
1531 // Load boxed value from result of valueOf() call is input parameter.
1532 if (this->is_Load() && ld_adr->is_AddP() &&
1533 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1534 intptr_t ignore = 0;
1535 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1536 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1537 base = bs->step_over_gc_barrier(base);
1538 if (base != nullptr && base->is_Proj() &&
1539 base->as_Proj()->_con == TypeFunc::Parms &&
1540 base->in(0)->is_CallStaticJava() &&
1541 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1542 return base->in(0)->in(TypeFunc::Parms);
1543 }
1544 }
1545
1546 break;
1547 }
1548
1549 return nullptr;
1550 }
1551
1552 //----------------------is_instance_field_load_with_local_phi------------------
1553 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1554 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1555 in(Address)->is_AddP() ) {
1556 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1557 // Only known instances and immutable fields
1558 if( t_oop != nullptr &&
1559 (t_oop->is_ptr_to_strict_final_field() ||
1560 t_oop->is_known_instance_field()) &&
1561 t_oop->offset() != Type::OffsetBot &&
1562 t_oop->offset() != Type::OffsetTop) {
1563 return true;
1564 }
1565 }
1566 return false;
1567 }
1568
1569 //------------------------------Identity---------------------------------------
1570 // Loads are identity if previous store is to same address
1571 Node* LoadNode::Identity(PhaseGVN* phase) {
1572 // If the previous store-maker is the right kind of Store, and the store is
1573 // to the same address, then we are equal to the value stored.
1574 Node* mem = in(Memory);
1575 Node* value = can_see_stored_value_through_membars(mem, phase);
1576 if( value ) {
1577 // byte, short & char stores truncate naturally.
1578 // A load has to load the truncated value which requires
1579 // some sort of masking operation and that requires an
1580 // Ideal call instead of an Identity call.
1581 if (memory_size() < BytesPerInt) {
1582 // If the input to the store does not fit with the load's result type,
1583 // it must be truncated via an Ideal call.
1584 if (!phase->type(value)->higher_equal(phase->type(this)))
1585 return this;
1586 }
1587
1588 if (phase->type(value)->isa_ptr() && phase->type(this)->isa_narrowoop()) {
1589 return this;
1590 }
1591 // (This works even when value is a Con, but LoadNode::Value
1592 // usually runs first, producing the singleton type of the Con.)
1593 if (!has_pinned_control_dependency() || value->is_Con()) {
1594 return value;
1595 } else {
1596 return this;
1597 }
1598 }
1599
1600 if (has_pinned_control_dependency()) {
1601 return this;
1602 }
1603 // Search for an existing data phi which was generated before for the same
1604 // instance's field to avoid infinite generation of phis in a loop.
1605 Node *region = mem->in(0);
1606 if (is_instance_field_load_with_local_phi(region)) {
1607 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1608 int this_index = phase->C->get_alias_index(addr_t);
1609 int this_offset = addr_t->offset();
1610 int this_iid = addr_t->instance_id();
1611 if (!addr_t->is_known_instance() &&
1612 addr_t->is_ptr_to_strict_final_field()) {
1613 // Use _idx of address base (could be Phi node) for immutable fields in unknown instances
1614 intptr_t ignore = 0;
1615 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1616 if (base == nullptr) {
1617 return this;
1618 }
1619 this_iid = base->_idx;
1620 }
1621 const Type* this_type = bottom_type();
1622 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1623 Node* phi = region->fast_out(i);
1624 if (phi->is_Phi() && phi != mem &&
1625 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1626 return phi;
1627 }
1628 }
1629 }
1630
1631 return this;
1632 }
1633
1634 // Construct an equivalent unsigned load.
1635 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1636 BasicType bt = T_ILLEGAL;
1637 const Type* rt = nullptr;
1638 switch (Opcode()) {
1639 case Op_LoadUB: return this;
1640 case Op_LoadUS: return this;
1641 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1642 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1643 default:
1644 assert(false, "no unsigned variant: %s", Name());
1645 return nullptr;
1646 }
1647 const Type* mem_t = gvn.type(in(MemNode::Address));
1648 if (mem_t == Type::TOP) {
1649 return gvn.C->top();
1650 }
1651 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1652 mem_t->is_ptr(), rt, bt, _mo, _control_dependency,
1653 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1654 }
1655
1656 // Construct an equivalent signed load.
1657 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1658 BasicType bt = T_ILLEGAL;
1659 const Type* rt = nullptr;
1660 switch (Opcode()) {
1661 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1662 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1663 case Op_LoadB: // fall through
1664 case Op_LoadS: // fall through
1665 case Op_LoadI: // fall through
1666 case Op_LoadL: return this;
1667 default:
1668 assert(false, "no signed variant: %s", Name());
1669 return nullptr;
1670 }
1671 const Type* mem_t = gvn.type(in(MemNode::Address));
1672 if (mem_t == Type::TOP) {
1673 return gvn.C->top();
1674 }
1675 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1676 mem_t->is_ptr(), rt, bt, _mo, _control_dependency,
1677 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1678 }
1679
1680 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1681 BasicType bt = rt->basic_type();
1682 switch (Opcode()) {
1683 case Op_LoadI: return (bt == T_FLOAT);
1684 case Op_LoadL: return (bt == T_DOUBLE);
1685 case Op_LoadF: return (bt == T_INT);
1686 case Op_LoadD: return (bt == T_LONG);
1687
1688 default: return false;
1689 }
1690 }
1691
1692 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1693 BasicType bt = rt->basic_type();
1694 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1695 bool is_mismatched = is_mismatched_access();
1696 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1697 if (raw_type == nullptr) {
1698 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1699 }
1700 const int op = Opcode();
1701 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1702 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1703 const Type* mem_t = gvn.type(in(MemNode::Address));
1704 if (mem_t == Type::TOP) {
1705 return gvn.C->top();
1706 }
1707 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1708 mem_t->is_ptr(), rt, bt, _mo, _control_dependency,
1709 require_atomic_access, is_unaligned_access(), is_mismatched);
1710 }
1711
1712 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1713 BasicType bt = vt->basic_type();
1714 switch (Opcode()) {
1715 case Op_StoreI: return (bt == T_FLOAT);
1716 case Op_StoreL: return (bt == T_DOUBLE);
1717 case Op_StoreF: return (bt == T_INT);
1718 case Op_StoreD: return (bt == T_LONG);
1719
1720 default: return false;
1721 }
1722 }
1723
1724 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1725 BasicType bt = vt->basic_type();
1726 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1727 const int op = Opcode();
1728 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1729 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1730 const Type* mem_t = gvn.type(in(MemNode::Address));
1731 if (mem_t == Type::TOP) {
1732 return gvn.C->top();
1733 }
1734 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1735 mem_t->is_ptr(), val, bt, _mo, require_atomic_access);
1736
1737 bool is_mismatched = is_mismatched_access();
1738 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1739 if (raw_type == nullptr) {
1740 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1741 }
1742 if (is_mismatched) {
1743 st->set_mismatched_access();
1744 }
1745 return st;
1746 }
1747
1748 // We're loading from an object which has autobox behaviour.
1749 // If this object is result of a valueOf call we'll have a phi
1750 // merging a newly allocated object and a load from the cache.
1751 // We want to replace this load with the original incoming
1752 // argument to the valueOf call.
1753 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1754 assert(igvn->C->eliminate_boxing(), "sanity");
1755 intptr_t ignore = 0;
1756 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1757 if ((base == nullptr) || base->is_Phi()) {
1758 // Push the loads from the phi that comes from valueOf up
1759 // through it to allow elimination of the loads and the recovery
1760 // of the original value. It is done in split_through_phi().
1761 return nullptr;
1762 } else if (base->is_Load() ||
1763 (base->is_DecodeN() && base->in(1)->is_Load())) {
1764 // Eliminate the load of boxed value for integer types from the cache
1765 // array by deriving the value from the index into the array.
1766 // Capture the offset of the load and then reverse the computation.
1767
1768 // Get LoadN node which loads a boxing object from 'cache' array.
1769 if (base->is_DecodeN()) {
1770 base = base->in(1);
1771 }
1772 if (!base->in(Address)->is_AddP()) {
1773 return nullptr; // Complex address
1774 }
1775 AddPNode* address = base->in(Address)->as_AddP();
1776 Node* cache_base = address->in(AddPNode::Base);
1777 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1778 // Get ConP node which is static 'cache' field.
1779 cache_base = cache_base->in(1);
1780 }
1781 if ((cache_base != nullptr) && cache_base->is_Con()) {
1782 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1783 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1784 Node* elements[4];
1785 int shift = exact_log2(type2aelembytes(T_OBJECT));
1786 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1787 if (count > 0 && elements[0]->is_Con() &&
1788 (count == 1 ||
1789 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1790 elements[1]->in(2) == igvn->intcon(shift)))) {
1791 ciObjArray* array = base_type->const_oop()->as_obj_array();
1792 // Fetch the box object cache[0] at the base of the array and get its value
1793 ciInstance* box = array->obj_at(0)->as_instance();
1794 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1795 assert(ik->is_box_klass(), "sanity");
1796 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1797 if (ik->nof_nonstatic_fields() == 1) {
1798 // This should be true nonstatic_field_at requires calling
1799 // nof_nonstatic_fields so check it anyway
1800 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1801 BasicType bt = c.basic_type();
1802 // Only integer types have boxing cache.
1803 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1804 bt == T_BYTE || bt == T_SHORT ||
1805 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1806 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1807 if (cache_low != (int)cache_low) {
1808 return nullptr; // should not happen since cache is array indexed by value
1809 }
1810 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1811 if (offset != (int)offset) {
1812 return nullptr; // should not happen since cache is array indexed by value
1813 }
1814 // Add up all the offsets making of the address of the load
1815 Node* result = elements[0];
1816 for (int i = 1; i < count; i++) {
1817 result = igvn->transform(new AddXNode(result, elements[i]));
1818 }
1819 // Remove the constant offset from the address and then
1820 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1821 // remove the scaling of the offset to recover the original index.
1822 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1823 // Peel the shift off directly but wrap it in a dummy node
1824 // since Ideal can't return existing nodes
1825 igvn->_worklist.push(result); // remove dead node later
1826 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1827 } else if (result->is_Add() && result->in(2)->is_Con() &&
1828 result->in(1)->Opcode() == Op_LShiftX &&
1829 result->in(1)->in(2) == igvn->intcon(shift)) {
1830 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1831 // but for boxing cache access we know that X<<Z will not overflow
1832 // (there is range check) so we do this optimizatrion by hand here.
1833 igvn->_worklist.push(result); // remove dead node later
1834 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1835 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1836 } else {
1837 result = new RShiftXNode(result, igvn->intcon(shift));
1838 }
1839 #ifdef _LP64
1840 if (bt != T_LONG) {
1841 result = new ConvL2INode(igvn->transform(result));
1842 }
1843 #else
1844 if (bt == T_LONG) {
1845 result = new ConvI2LNode(igvn->transform(result));
1846 }
1847 #endif
1848 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1849 // Need to preserve unboxing load type if it is unsigned.
1850 switch(this->Opcode()) {
1851 case Op_LoadUB:
1852 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1853 break;
1854 case Op_LoadUS:
1855 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1856 break;
1857 }
1858 return result;
1859 }
1860 }
1861 }
1862 }
1863 }
1864 return nullptr;
1865 }
1866
1867 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1868 Node* region = phi->in(0);
1869 if (region == nullptr) {
1870 return false; // Wait stable graph
1871 }
1872 uint cnt = phi->req();
1873 for (uint i = 1; i < cnt; i++) {
1874 Node* rc = region->in(i);
1875 if (rc == nullptr || phase->type(rc) == Type::TOP)
1876 return false; // Wait stable graph
1877 Node* in = phi->in(i);
1878 if (in == nullptr || phase->type(in) == Type::TOP)
1879 return false; // Wait stable graph
1880 }
1881 return true;
1882 }
1883
1884 //------------------------------split_through_phi------------------------------
1885 // Check whether a call to 'split_through_phi' would split this load through the
1886 // Phi *base*. This method is essentially a copy of the validations performed
1887 // by 'split_through_phi'. The first use of this method was in EA code as part
1888 // of simplification of allocation merges.
1889 // Some differences from original method (split_through_phi):
1890 // - If base->is_CastPP(): base = base->in(1)
1891 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1892 Node* mem = in(Memory);
1893 Node* address = in(Address);
1894 intptr_t ignore = 0;
1895 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1896
1897 if (base == nullptr) {
1898 return false;
1899 }
1900
1901 if (base->is_CastPP()) {
1902 base = base->in(1);
1903 }
1904
1905 if (req() > 3 || !base->is_Phi()) {
1906 return false;
1907 }
1908
1909 if (!mem->is_Phi()) {
1910 if (!MemNode::all_controls_dominate(mem, base->in(0), phase)) {
1911 return false;
1912 }
1913 } else if (base->in(0) != mem->in(0)) {
1914 if (!MemNode::all_controls_dominate(mem, base->in(0), phase)) {
1915 return false;
1916 }
1917 }
1918
1919 return true;
1920 }
1921
1922 //------------------------------split_through_phi------------------------------
1923 // Split instance or boxed field load through Phi.
1924 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1925 if (req() > 3) {
1926 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1927 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1928 return nullptr;
1929 }
1930 Node* mem = in(Memory);
1931 Node* address = in(Address);
1932 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1933
1934 assert((t_oop != nullptr) &&
1935 (ignore_missing_instance_id ||
1936 t_oop->is_known_instance_field() ||
1937 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1938
1939 Compile* C = phase->C;
1940 intptr_t ignore = 0;
1941 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1942 bool base_is_phi = (base != nullptr) && base->is_Phi();
1943 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1944 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1945 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1946
1947 if (!((mem->is_Phi() || base_is_phi) &&
1948 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1949 return nullptr; // Neither memory or base are Phi
1950 }
1951
1952 if (mem->is_Phi()) {
1953 if (!stable_phi(mem->as_Phi(), phase)) {
1954 return nullptr; // Wait stable graph
1955 }
1956 uint cnt = mem->req();
1957 // Check for loop invariant memory.
1958 if (cnt == 3) {
1959 for (uint i = 1; i < cnt; i++) {
1960 Node* in = mem->in(i);
1961 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1962 if (m == mem) {
1963 if (i == 1) {
1964 // if the first edge was a loop, check second edge too.
1965 // If both are replaceable - we are in an infinite loop
1966 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1967 if (n == mem) {
1968 break;
1969 }
1970 }
1971 set_req(Memory, mem->in(cnt - i));
1972 return this; // made change
1973 }
1974 }
1975 }
1976 }
1977 if (base_is_phi) {
1978 if (!stable_phi(base->as_Phi(), phase)) {
1979 return nullptr; // Wait stable graph
1980 }
1981 uint cnt = base->req();
1982 // Check for loop invariant memory.
1983 if (cnt == 3) {
1984 for (uint i = 1; i < cnt; i++) {
1985 if (base->in(i) == base) {
1986 return nullptr; // Wait stable graph
1987 }
1988 }
1989 }
1990 }
1991
1992 // Split through Phi (see original code in loopopts.cpp).
1993 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1994
1995 // Do nothing here if Identity will find a value
1996 // (to avoid infinite chain of value phis generation).
1997 if (this != Identity(phase)) {
1998 return nullptr;
1999 }
2000
2001 // Select Region to split through.
2002 Node* region;
2003 DomResult dom_result = DomResult::Dominate;
2004 if (!base_is_phi) {
2005 assert(mem->is_Phi(), "sanity");
2006 region = mem->in(0);
2007 // Skip if the region dominates some control edge of the address.
2008 // We will check `dom_result` later.
2009 dom_result = MemNode::maybe_all_controls_dominate(address, region, phase);
2010 } else if (!mem->is_Phi()) {
2011 assert(base_is_phi, "sanity");
2012 region = base->in(0);
2013 // Skip if the region dominates some control edge of the memory.
2014 // We will check `dom_result` later.
2015 dom_result = MemNode::maybe_all_controls_dominate(mem, region, phase);
2016 } else if (base->in(0) != mem->in(0)) {
2017 assert(base_is_phi && mem->is_Phi(), "sanity");
2018 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0), phase);
2019 if (dom_result == DomResult::Dominate) {
2020 region = base->in(0);
2021 } else {
2022 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0), phase);
2023 if (dom_result == DomResult::Dominate) {
2024 region = mem->in(0);
2025 }
2026 // Otherwise we encountered a complex graph.
2027 }
2028 } else {
2029 assert(base->in(0) == mem->in(0), "sanity");
2030 region = mem->in(0);
2031 }
2032
2033 PhaseIterGVN* igvn = phase->is_IterGVN();
2034 if (dom_result != DomResult::Dominate) {
2035 if (dom_result == DomResult::EncounteredDeadCode) {
2036 // There is some dead code which eventually will be removed in IGVN.
2037 // Once this is the case, we get an unambiguous dominance result.
2038 // Push the node to the worklist again until the dead code is removed.
2039 igvn->_worklist.push(this);
2040 }
2041 return nullptr;
2042 }
2043
2044 Node* phi = nullptr;
2045 const Type* this_type = this->bottom_type();
2046 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
2047 int this_index = C->get_alias_index(t_oop);
2048 int this_offset = t_oop->offset();
2049 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
2050 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
2051 } else if (ignore_missing_instance_id) {
2052 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
2053 } else {
2054 return nullptr;
2055 }
2056
2057 for (uint i = 1; i < region->req(); i++) {
2058 Node* x;
2059 Node* the_clone = nullptr;
2060 Node* in = region->in(i);
2061 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
2062 in != nullptr && in->is_OuterStripMinedLoop()) {
2063 // No node should go in the outer strip mined loop
2064 in = in->in(LoopNode::EntryControl);
2065 }
2066 if (in == nullptr || in == C->top()) {
2067 x = C->top(); // Dead path? Use a dead data op
2068 } else {
2069 x = this->clone(); // Else clone up the data op
2070 the_clone = x; // Remember for possible deletion.
2071 // Alter data node to use pre-phi inputs
2072 if (this->in(0) == region) {
2073 x->set_req(0, in);
2074 } else {
2075 x->set_req(0, nullptr);
2076 }
2077 if (mem->is_Phi() && (mem->in(0) == region)) {
2078 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
2079 }
2080 if (address->is_Phi() && address->in(0) == region) {
2081 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
2082 }
2083 if (base_is_phi && (base->in(0) == region)) {
2084 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
2085 Node* adr_x = phase->transform(AddPNode::make_with_base(base_x, address->in(AddPNode::Offset)));
2086 x->set_req(Address, adr_x);
2087 }
2088 }
2089 // Check for a 'win' on some paths
2090 const Type *t = x->Value(igvn);
2091
2092 bool singleton = t->singleton();
2093
2094 // See comments in PhaseIdealLoop::split_thru_phi().
2095 if (singleton && t == Type::TOP) {
2096 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
2097 }
2098
2099 if (singleton) {
2100 x = igvn->makecon(t);
2101 } else {
2102 // We now call Identity to try to simplify the cloned node.
2103 // Note that some Identity methods call phase->type(this).
2104 // Make sure that the type array is big enough for
2105 // our new node, even though we may throw the node away.
2106 // (This tweaking with igvn only works because x is a new node.)
2107 igvn->set_type(x, t);
2108 // If x is a TypeNode, capture any more-precise type permanently into Node
2109 // otherwise it will be not updated during igvn->transform since
2110 // igvn->type(x) is set to x->Value() already.
2111 x->raise_bottom_type(t);
2112 Node* y = x->Identity(igvn);
2113 if (y != x) {
2114 x = y;
2115 } else {
2116 y = igvn->hash_find_insert(x);
2117 if (y) {
2118 x = y;
2119 } else {
2120 // Else x is a new node we are keeping
2121 // We do not need register_new_node_with_optimizer
2122 // because set_type has already been called.
2123 igvn->_worklist.push(x);
2124 }
2125 }
2126 }
2127 if (x != the_clone && the_clone != nullptr) {
2128 igvn->remove_dead_node(the_clone, PhaseIterGVN::NodeOrigin::Speculative);
2129 }
2130 phi->set_req(i, x);
2131 }
2132 // Record Phi
2133 igvn->register_new_node_with_optimizer(phi);
2134 return phi;
2135 }
2136
2137 AllocateNode* LoadNode::is_new_object_mark_load() const {
2138 if (Opcode() == Op_LoadX) {
2139 Node* address = in(MemNode::Address);
2140 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
2141 Node* mem = in(MemNode::Memory);
2142 if (alloc != nullptr && mem->is_Proj() &&
2143 mem->in(0) != nullptr &&
2144 mem->in(0) == alloc->initialization() &&
2145 alloc->initialization()->proj_out_or_null(0) != nullptr) {
2146 return alloc;
2147 }
2148 }
2149 return nullptr;
2150 }
2151
2152
2153 //------------------------------Ideal------------------------------------------
2154 // If the load is from Field memory and the pointer is non-null, it might be possible to
2155 // zero out the control input.
2156 // If the offset is constant and the base is an object allocation,
2157 // try to hook me up to the exact initializing store.
2158 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2159 if (has_pinned_control_dependency()) {
2160 return nullptr;
2161 }
2162 Node* p = MemNode::Ideal_common(phase, can_reshape);
2163 if (p) return (p == NodeSentinel) ? nullptr : p;
2164
2165 Node* ctrl = in(MemNode::Control);
2166 Node* address = in(MemNode::Address);
2167
2168 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
2169 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
2170
2171 // Skip up past a SafePoint control. Cannot do this for Stores because
2172 // pointer stores & cardmarks must stay on the same side of a SafePoint.
2173 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
2174 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
2175 !addr_mark &&
2176 (depends_only_on_test() || has_unknown_control_dependency())) {
2177 ctrl = ctrl->in(0);
2178 set_req(MemNode::Control,ctrl);
2179 return this;
2180 }
2181
2182 intptr_t ignore = 0;
2183 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
2184 if (base != nullptr
2185 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
2186 // Check for useless control edge in some common special cases
2187 if (in(MemNode::Control) != nullptr
2188 // TODO 8350865 Can we re-enable this?
2189 && !(phase->type(address)->is_inlinetypeptr() && is_mismatched_access())
2190 && can_remove_control()
2191 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
2192 && all_controls_dominate(base, phase->C->start(), phase)) {
2193 // A method-invariant, non-null address (constant or 'this' argument).
2194 set_req(MemNode::Control, nullptr);
2195 return this;
2196 }
2197 }
2198
2199 Node* mem = in(MemNode::Memory);
2200 const TypePtr *addr_t = phase->type(address)->isa_ptr();
2201
2202 if (can_reshape && (addr_t != nullptr)) {
2203 // try to optimize our memory input
2204 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
2205 if (opt_mem != mem) {
2206 set_req_X(MemNode::Memory, opt_mem, phase);
2207 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
2208 return this;
2209 }
2210 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
2211 if ((t_oop != nullptr) &&
2212 (t_oop->is_known_instance_field() ||
2213 t_oop->is_ptr_to_boxed_value())) {
2214 PhaseIterGVN *igvn = phase->is_IterGVN();
2215 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
2216 if (igvn->_worklist.member(opt_mem)) {
2217 // Delay this transformation until memory Phi is processed.
2218 igvn->_worklist.push(this);
2219 return nullptr;
2220 }
2221 // Split instance field load through Phi.
2222 Node* result = split_through_phi(phase);
2223 if (result != nullptr) return result;
2224
2225 if (t_oop->is_ptr_to_boxed_value()) {
2226 Node* result = eliminate_autobox(igvn);
2227 if (result != nullptr) return result;
2228 }
2229 }
2230 }
2231
2232 // Is there a dominating load that loads the same value? Leave
2233 // anything that is not a load of a field/array element (like
2234 // barriers etc.) alone
2235 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
2236 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2237 Node *use = mem->fast_out(i);
2238 if (use != this &&
2239 use->Opcode() == Opcode() &&
2240 use->in(0) != nullptr &&
2241 use->in(0) != in(0) &&
2242 use->in(Address) == in(Address)) {
2243 Node* ctl = in(0);
2244 for (int i = 0; i < 10 && ctl != nullptr; i++) {
2245 ctl = IfNode::up_one_dom(ctl);
2246 if (ctl == use->in(0)) {
2247 set_req(0, use->in(0));
2248 return this;
2249 }
2250 }
2251 }
2252 }
2253 }
2254
2255 // Check for prior store with a different base or offset; make Load
2256 // independent. Skip through any number of them. Bail out if the stores
2257 // are in an endless dead cycle and report no progress. This is a key
2258 // transform for Reflection. However, if after skipping through the Stores
2259 // we can't then fold up against a prior store do NOT do the transform as
2260 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
2261 // array memory alive twice: once for the hoisted Load and again after the
2262 // bypassed Store. This situation only works if EVERYBODY who does
2263 // anti-dependence work knows how to bypass. I.e. we need all
2264 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
2265 // the alias index stuff. So instead, peek through Stores and IFF we can
2266 // fold up, do so.
2267 Node* prev_mem = find_previous_store(phase);
2268 if (prev_mem != nullptr && prev_mem->is_top()) {
2269 // find_previous_store returns top when the access is dead
2270 return prev_mem;
2271 }
2272 if (prev_mem != nullptr) {
2273 Node* value = can_see_arraycopy_value(prev_mem, phase);
2274 if (value != nullptr) {
2275 return value;
2276 }
2277 }
2278 // Steps (a), (b): Walk past independent stores to find an exact match.
2279 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
2280 // (c) See if we can fold up on the spot, but don't fold up here.
2281 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
2282 // just return a prior value, which is done by Identity calls.
2283 if (can_see_stored_value_through_membars(prev_mem, phase)) {
2284 // Make ready for step (d):
2285 set_req_X(MemNode::Memory, prev_mem, phase);
2286 return this;
2287 }
2288 }
2289
2290 if (!can_reshape) {
2291 phase->record_for_igvn(this);
2292 }
2293
2294 return nullptr;
2295 }
2296
2297 // Helper to recognize certain Klass fields which are invariant across
2298 // some group of array types (e.g., int[] or all T[] where T < Object).
2299 const Type*
2300 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
2301 ciKlass* klass) const {
2302 assert(!UseCompactObjectHeaders || tkls->offset() != in_bytes(Klass::prototype_header_offset()),
2303 "must not happen");
2304
2305 if (tkls->isa_instklassptr() && tkls->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
2306 // The field is InstanceKlass::_access_flags. Return its (constant) value.
2307 assert(Opcode() == Op_LoadUS, "must load an unsigned short from _access_flags");
2308 ciInstanceKlass* iklass = tkls->is_instklassptr()->instance_klass();
2309 return TypeInt::make(iklass->access_flags());
2310 }
2311 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
2312 // The field is Klass::_misc_flags. Return its (constant) value.
2313 assert(Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
2314 return TypeInt::make(klass->misc_flags());
2315 }
2316 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
2317 // The field is Klass::_layout_helper. Return its constant value if known.
2318 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2319 return TypeInt::make(klass->layout_helper());
2320 }
2321
2322 // No match.
2323 return nullptr;
2324 }
2325
2326 //------------------------------Value-----------------------------------------
2327 const Type* LoadNode::Value(PhaseGVN* phase) const {
2328 // Either input is TOP ==> the result is TOP
2329 Node* mem = in(MemNode::Memory);
2330 const Type *t1 = phase->type(mem);
2331 if (t1 == Type::TOP) return Type::TOP;
2332 Node* adr = in(MemNode::Address);
2333 const TypePtr* tp = phase->type(adr)->isa_ptr();
2334 if (tp == nullptr || tp->empty()) return Type::TOP;
2335 int off = tp->offset();
2336 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2337 Compile* C = phase->C;
2338
2339 // If load can see a previous constant store, use that.
2340 Node* value = can_see_stored_value_through_membars(mem, phase);
2341 if (value != nullptr && value->is_Con()) {
2342 if (phase->type(value)->isa_ptr() && _type->isa_narrowoop()) {
2343 return phase->type(value)->make_narrowoop();
2344 } else {
2345 assert(value->bottom_type()->higher_equal(_type), "sanity");
2346 return phase->type(value);
2347 }
2348 }
2349 // Try to guess loaded type from pointer type
2350 if (tp->isa_aryptr()) {
2351 const TypeAryPtr* ary = tp->is_aryptr();
2352 const Type* t = ary->elem();
2353
2354 // Determine whether the reference is beyond the header or not, by comparing
2355 // the offset against the offset of the start of the array's data.
2356 // Different array types begin at slightly different offsets (12 vs. 16).
2357 // We choose T_BYTE as an example base type that is least restrictive
2358 // as to alignment, which will therefore produce the smallest
2359 // possible base offset.
2360 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2361 const bool off_beyond_header = (off >= min_base_off);
2362
2363 // Try to constant-fold a stable array element.
2364 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2365 // Make sure the reference is not into the header and the offset is constant
2366 ciObject* aobj = ary->const_oop();
2367 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2368 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2369 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off, ary->field_offset().get(),
2370 stable_dimension,
2371 value_basic_type(), is_unsigned());
2372 if (con_type != nullptr) {
2373 return con_type;
2374 }
2375 }
2376 }
2377
2378 // Don't do this for integer types. There is only potential profit if
2379 // the element type t is lower than _type; that is, for int types, if _type is
2380 // more restrictive than t. This only happens here if one is short and the other
2381 // char (both 16 bits), and in those cases we've made an intentional decision
2382 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2383 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2384 //
2385 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2386 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2387 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2388 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2389 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2390 // In fact, that could have been the original type of p1, and p1 could have
2391 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2392 // expression (LShiftL quux 3) independently optimized to the constant 8.
2393 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2394 && (_type->isa_vect() == nullptr)
2395 && !ary->is_flat()
2396 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2397 // t might actually be lower than _type, if _type is a unique
2398 // concrete subclass of abstract class t.
2399 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2400 const Type* jt = t->join_speculative(_type);
2401 // In any case, do not allow the join, per se, to empty out the type.
2402 if (jt->empty() && !t->empty()) {
2403 // This can happen if a interface-typed array narrows to a class type.
2404 jt = _type;
2405 }
2406 #ifdef ASSERT
2407 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2408 // The pointers in the autobox arrays are always non-null
2409 Node* base = adr->in(AddPNode::Base);
2410 if ((base != nullptr) && base->is_DecodeN()) {
2411 // Get LoadN node which loads IntegerCache.cache field
2412 base = base->in(1);
2413 }
2414 if ((base != nullptr) && base->is_Con()) {
2415 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2416 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2417 // It could be narrow oop
2418 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2419 }
2420 }
2421 }
2422 #endif
2423 return jt;
2424 }
2425 }
2426 } else if (tp->base() == Type::InstPtr) {
2427 assert( off != Type::OffsetBot ||
2428 // arrays can be cast to Objects
2429 !tp->isa_instptr() ||
2430 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2431 // Default value load
2432 tp->is_instptr()->instance_klass() == ciEnv::current()->Class_klass() ||
2433 // unsafe field access may not have a constant offset
2434 is_unsafe_access(),
2435 "Field accesses must be precise" );
2436 // For oop loads, we expect the _type to be precise.
2437
2438 const TypeInstPtr* tinst = tp->is_instptr();
2439 BasicType bt = value_basic_type();
2440
2441 // Fold loads of the field map
2442 if (tinst != nullptr) {
2443 ciInstanceKlass* ik = tinst->instance_klass();
2444 int offset = tinst->offset();
2445 if (ik == phase->C->env()->Class_klass()) {
2446 ciType* t = tinst->java_mirror_type();
2447 if (t != nullptr && t->is_inlinetype() && offset == t->as_inline_klass()->field_map_offset()) {
2448 ciConstant map = t->as_inline_klass()->get_field_map();
2449 bool is_narrow_oop = (bt == T_NARROWOOP);
2450 return Type::make_from_constant(map, true, 1, is_narrow_oop);
2451 }
2452 }
2453 }
2454
2455 // Optimize loads from constant fields.
2456 ciObject* const_oop = tinst->const_oop();
2457 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2458 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
2459 if (con_type != nullptr) {
2460 return con_type;
2461 }
2462 }
2463 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2464 assert(off != Type::OffsetBot ||
2465 !tp->isa_instklassptr() ||
2466 // arrays can be cast to Objects
2467 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2468 // also allow array-loading from the primary supertype
2469 // array during subtype checks
2470 Opcode() == Op_LoadKlass,
2471 "Field accesses must be precise");
2472 // For klass/static loads, we expect the _type to be precise
2473 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2474 /* With mirrors being an indirect in the Klass*
2475 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2476 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2477 *
2478 * So check the type and klass of the node before the LoadP.
2479 */
2480 Node* adr2 = adr->in(MemNode::Address);
2481 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2482 if (tkls != nullptr && !StressReflectiveCode) {
2483 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2484 ciKlass* klass = tkls->exact_klass();
2485 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2486 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2487 return TypeInstPtr::make(klass->java_mirror());
2488 }
2489 }
2490 }
2491
2492 const TypeKlassPtr *tkls = tp->isa_klassptr();
2493 if (tkls != nullptr) {
2494 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2495 ciKlass* klass = tkls->exact_klass();
2496 // We are loading a field from a Klass metaobject whose identity
2497 // is known at compile time (the type is "exact" or "precise").
2498 // Check for fields we know are maintained as constants by the VM.
2499 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2500 // The field is Klass::_super_check_offset. Return its (constant) value.
2501 // (Folds up type checking code.)
2502 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2503 return TypeInt::make(klass->super_check_offset());
2504 }
2505 if (klass->is_inlinetype() && tkls->offset() == in_bytes(InstanceKlass::acmp_maps_offset_offset())) {
2506 return TypeInt::make(klass->as_inline_klass()->field_map_offset());
2507 }
2508 if (klass->is_obj_array_klass() && tkls->offset() == in_bytes(ObjArrayKlass::next_refined_array_klass_offset())) {
2509 // Fold loads from LibraryCallKit::load_default_refined_array_klass
2510 return tkls->is_aryklassptr()->cast_to_refined_array_klass_ptr();
2511 }
2512 if (klass->is_array_klass() && tkls->offset() == in_bytes(ObjArrayKlass::properties_offset())) {
2513 assert(klass->is_type_array_klass() || tkls->is_aryklassptr()->is_refined_type(), "Must be a refined array klass pointer");
2514 return TypeInt::make((jint)klass->as_array_klass()->properties().value());
2515 }
2516 if (klass->is_flat_array_klass() && tkls->offset() == in_bytes(FlatArrayKlass::layout_kind_offset())) {
2517 assert(Opcode() == Op_LoadI, "must load an int from _layout_kind");
2518 return TypeInt::make(static_cast<jint>(klass->as_flat_array_klass()->layout_kind()));
2519 }
2520 if (UseCompactObjectHeaders && tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2521 // The field is Klass::_prototype_header. Return its (constant) value.
2522 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2523 return TypeX::make(klass->prototype_header());
2524 }
2525 // Compute index into primary_supers array
2526 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2527 // Check for overflowing; use unsigned compare to handle the negative case.
2528 if( depth < ciKlass::primary_super_limit() ) {
2529 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2530 // (Folds up type checking code.)
2531 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2532 ciKlass *ss = klass->super_of_depth(depth);
2533 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2534 }
2535 const Type* aift = load_array_final_field(tkls, klass);
2536 if (aift != nullptr) return aift;
2537 }
2538
2539 // We can still check if we are loading from the primary_supers array at a
2540 // shallow enough depth. Even though the klass is not exact, entries less
2541 // than or equal to its super depth are correct.
2542 if (tkls->is_loaded()) {
2543 ciKlass* klass = nullptr;
2544 if (tkls->isa_instklassptr()) {
2545 klass = tkls->is_instklassptr()->instance_klass();
2546 } else {
2547 int dims;
2548 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2549 if (inner->isa_instklassptr()) {
2550 klass = inner->is_instklassptr()->instance_klass();
2551 klass = ciObjArrayKlass::make(klass, dims);
2552 }
2553 }
2554 if (klass != nullptr) {
2555 // Compute index into primary_supers array
2556 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2557 // Check for overflowing; use unsigned compare to handle the negative case.
2558 if (depth < ciKlass::primary_super_limit() &&
2559 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2560 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2561 // (Folds up type checking code.)
2562 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2563 ciKlass *ss = klass->super_of_depth(depth);
2564 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2565 }
2566 }
2567 }
2568
2569 // If the type is enough to determine that the thing is not an array,
2570 // we can give the layout_helper a positive interval type.
2571 // This will help short-circuit some reflective code.
2572 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2573 tkls->isa_instklassptr() && // not directly typed as an array
2574 !tkls->is_instklassptr()->might_be_an_array() // not the supertype of all T[] (java.lang.Object) or has an interface that is not Serializable or Cloneable
2575 ) {
2576 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2577 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2578 // The key property of this type is that it folds up tests
2579 // for array-ness, since it proves that the layout_helper is positive.
2580 // Thus, a generic value like the basic object layout helper works fine.
2581 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2582 }
2583 }
2584
2585 // If we are loading from a freshly-allocated object/array, produce a zero.
2586 // Things to check:
2587 // 1. Load is beyond the header: headers are not guaranteed to be zero
2588 // 2. Load is not vectorized: vectors have no zero constant
2589 // 3. Load has no matching store, i.e. the input is the initial memory state
2590 const TypeOopPtr* tinst = tp->isa_oopptr();
2591 bool is_not_header = (tinst != nullptr) && tinst->is_known_instance_field();
2592 bool is_not_vect = (_type->isa_vect() == nullptr);
2593 if (is_not_header && is_not_vect) {
2594 Node* mem = in(MemNode::Memory);
2595 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2596 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2597 // TODO 8350865 Scalar replacement does not work well for flat arrays.
2598 // Escape Analysis assumes that arrays are always zeroed during allocation which is not true for null-free arrays
2599 // ConnectionGraph::split_unique_types will re-wire the memory of loads from such arrays around the allocation
2600 // TestArrays::test6 and test152 and TestBasicFunctionality::test20 are affected by this.
2601 if (tp->isa_aryptr() && tp->is_aryptr()->is_flat() && tp->is_aryptr()->is_null_free()) {
2602 intptr_t offset = 0;
2603 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2604 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
2605 if (alloc != nullptr && alloc->is_AllocateArray() && alloc->in(AllocateNode::InitValue) != nullptr) {
2606 return _type;
2607 }
2608 }
2609 return Type::get_zero_type(_type->basic_type());
2610 }
2611 }
2612 if (!UseCompactObjectHeaders) {
2613 Node* alloc = is_new_object_mark_load();
2614 if (alloc != nullptr) {
2615 if (Arguments::is_valhalla_enabled()) {
2616 // The mark word may contain property bits (inline, flat, null-free)
2617 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2618 const TypeKlassPtr* tkls = phase->type(klass_node)->isa_klassptr();
2619 if (tkls != nullptr && tkls->is_loaded() && tkls->klass_is_exact()) {
2620 return TypeX::make(tkls->exact_klass()->prototype_header());
2621 }
2622 } else {
2623 return TypeX::make(markWord::prototype().value());
2624 }
2625 }
2626 }
2627
2628 return _type;
2629 }
2630
2631 //------------------------------match_edge-------------------------------------
2632 // Do we Match on this edge index or not? Match only the address.
2633 uint LoadNode::match_edge(uint idx) const {
2634 return idx == MemNode::Address;
2635 }
2636
2637 //--------------------------LoadBNode::Ideal--------------------------------------
2638 //
2639 // If the previous store is to the same address as this load,
2640 // and the value stored was larger than a byte, replace this load
2641 // with the value stored truncated to a byte. If no truncation is
2642 // needed, the replacement is done in LoadNode::Identity().
2643 //
2644 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2645 Node* mem = in(MemNode::Memory);
2646 Node* value = can_see_stored_value_through_membars(mem, phase);
2647 if (value != nullptr) {
2648 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2649 if (narrow != value) {
2650 return narrow;
2651 }
2652 }
2653 // Identity call will handle the case where truncation is not needed.
2654 return LoadNode::Ideal(phase, can_reshape);
2655 }
2656
2657 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2658 Node* mem = in(MemNode::Memory);
2659 Node* value = can_see_stored_value_through_membars(mem, phase);
2660 if (value != nullptr && value->is_Con() &&
2661 !value->bottom_type()->higher_equal(_type)) {
2662 // If the input to the store does not fit with the load's result type,
2663 // it must be truncated. We can't delay until Ideal call since
2664 // a singleton Value is needed for split_thru_phi optimization.
2665 int con = value->get_int();
2666 return TypeInt::make((con << 24) >> 24);
2667 }
2668 return LoadNode::Value(phase);
2669 }
2670
2671 //--------------------------LoadUBNode::Ideal-------------------------------------
2672 //
2673 // If the previous store is to the same address as this load,
2674 // and the value stored was larger than a byte, replace this load
2675 // with the value stored truncated to a byte. If no truncation is
2676 // needed, the replacement is done in LoadNode::Identity().
2677 //
2678 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2679 Node* mem = in(MemNode::Memory);
2680 Node* value = can_see_stored_value_through_membars(mem, phase);
2681 if (value != nullptr) {
2682 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2683 if (narrow != value) {
2684 return narrow;
2685 }
2686 }
2687 // Identity call will handle the case where truncation is not needed.
2688 return LoadNode::Ideal(phase, can_reshape);
2689 }
2690
2691 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2692 Node* mem = in(MemNode::Memory);
2693 Node* value = can_see_stored_value_through_membars(mem, phase);
2694 if (value != nullptr && value->is_Con() &&
2695 !value->bottom_type()->higher_equal(_type)) {
2696 // If the input to the store does not fit with the load's result type,
2697 // it must be truncated. We can't delay until Ideal call since
2698 // a singleton Value is needed for split_thru_phi optimization.
2699 int con = value->get_int();
2700 return TypeInt::make(con & 0xFF);
2701 }
2702 return LoadNode::Value(phase);
2703 }
2704
2705 //--------------------------LoadUSNode::Ideal-------------------------------------
2706 //
2707 // If the previous store is to the same address as this load,
2708 // and the value stored was larger than a char, replace this load
2709 // with the value stored truncated to a char. If no truncation is
2710 // needed, the replacement is done in LoadNode::Identity().
2711 //
2712 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2713 Node* mem = in(MemNode::Memory);
2714 Node* value = can_see_stored_value_through_membars(mem, phase);
2715 if (value != nullptr) {
2716 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2717 if (narrow != value) {
2718 return narrow;
2719 }
2720 }
2721 // Identity call will handle the case where truncation is not needed.
2722 return LoadNode::Ideal(phase, can_reshape);
2723 }
2724
2725 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2726 Node* mem = in(MemNode::Memory);
2727 Node* value = can_see_stored_value_through_membars(mem, phase);
2728 if (value != nullptr && value->is_Con() &&
2729 !value->bottom_type()->higher_equal(_type)) {
2730 // If the input to the store does not fit with the load's result type,
2731 // it must be truncated. We can't delay until Ideal call since
2732 // a singleton Value is needed for split_thru_phi optimization.
2733 int con = value->get_int();
2734 return TypeInt::make(con & 0xFFFF);
2735 }
2736 return LoadNode::Value(phase);
2737 }
2738
2739 //--------------------------LoadSNode::Ideal--------------------------------------
2740 //
2741 // If the previous store is to the same address as this load,
2742 // and the value stored was larger than a short, replace this load
2743 // with the value stored truncated to a short. If no truncation is
2744 // needed, the replacement is done in LoadNode::Identity().
2745 //
2746 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2747 Node* mem = in(MemNode::Memory);
2748 Node* value = can_see_stored_value_through_membars(mem, phase);
2749 if (value != nullptr) {
2750 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2751 if (narrow != value) {
2752 return narrow;
2753 }
2754 }
2755 // Identity call will handle the case where truncation is not needed.
2756 return LoadNode::Ideal(phase, can_reshape);
2757 }
2758
2759 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2760 Node* mem = in(MemNode::Memory);
2761 Node* value = can_see_stored_value_through_membars(mem, phase);
2762 if (value != nullptr && value->is_Con() &&
2763 !value->bottom_type()->higher_equal(_type)) {
2764 // If the input to the store does not fit with the load's result type,
2765 // it must be truncated. We can't delay until Ideal call since
2766 // a singleton Value is needed for split_thru_phi optimization.
2767 int con = value->get_int();
2768 return TypeInt::make((con << 16) >> 16);
2769 }
2770 return LoadNode::Value(phase);
2771 }
2772
2773 Node* LoadNNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2774 // Can see the corresponding value, may need to add an EncodeP
2775 Node* value = can_see_stored_value_through_membars(in(Memory), phase);
2776 if (value != nullptr && phase->type(value)->isa_ptr() && type()->isa_narrowoop()) {
2777 return new EncodePNode(value, type());
2778 }
2779
2780 // Identity call will handle the case where EncodeP is unnecessary
2781 return LoadNode::Ideal(phase, can_reshape);
2782 }
2783
2784 //=============================================================================
2785 //----------------------------LoadKlassNode::make------------------------------
2786 // Polymorphic factory method:
2787 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2788 // sanity check the alias category against the created node type
2789 const TypePtr* adr_type = adr->bottom_type()->isa_ptr();
2790 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2791 #ifdef _LP64
2792 if (adr_type->is_ptr_to_narrowklass()) {
2793 Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2794 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2795 }
2796 #endif
2797 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2798 return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered);
2799 }
2800
2801 //------------------------------Value------------------------------------------
2802 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2803 return klass_value_common(phase);
2804 }
2805
2806 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2807 // Either input is TOP ==> the result is TOP
2808 const Type *t1 = phase->type( in(MemNode::Memory) );
2809 if (t1 == Type::TOP) return Type::TOP;
2810 Node *adr = in(MemNode::Address);
2811 const Type *t2 = phase->type( adr );
2812 if (t2 == Type::TOP) return Type::TOP;
2813 const TypePtr *tp = t2->is_ptr();
2814 if (TypePtr::above_centerline(tp->ptr()) ||
2815 tp->ptr() == TypePtr::Null) return Type::TOP;
2816
2817 // Return a more precise klass, if possible
2818 const TypeInstPtr *tinst = tp->isa_instptr();
2819 if (tinst != nullptr) {
2820 ciInstanceKlass* ik = tinst->instance_klass();
2821 int offset = tinst->offset();
2822 if (ik == phase->C->env()->Class_klass()
2823 && (offset == java_lang_Class::klass_offset() ||
2824 offset == java_lang_Class::array_klass_offset())) {
2825 // We are loading a special hidden field from a Class mirror object,
2826 // the field which points to the VM's Klass metaobject.
2827 ciType* t = tinst->java_mirror_type();
2828 // java_mirror_type returns non-null for compile-time Class constants.
2829 if (t != nullptr) {
2830 // constant oop => constant klass
2831 if (offset == java_lang_Class::array_klass_offset()) {
2832 if (t->is_void()) {
2833 // We cannot create a void array. Since void is a primitive type return null
2834 // klass. Users of this result need to do a null check on the returned klass.
2835 return TypePtr::NULL_PTR;
2836 }
2837 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2838 }
2839 if (!t->is_klass()) {
2840 // a primitive Class (e.g., int.class) has null for a klass field
2841 return TypePtr::NULL_PTR;
2842 }
2843 // Fold up the load of the hidden field
2844 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2845 }
2846 // non-constant mirror, so we can't tell what's going on
2847 }
2848 if (!tinst->is_loaded())
2849 return _type; // Bail out if not loaded
2850 if (offset == oopDesc::klass_offset_in_bytes()) {
2851 return tinst->as_klass_type(true);
2852 }
2853 }
2854
2855 // Check for loading klass from an array
2856 const TypeAryPtr* tary = tp->isa_aryptr();
2857 if (tary != nullptr &&
2858 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2859 return tary->as_klass_type(true)->is_aryklassptr();
2860 }
2861
2862 // Check for loading klass from an array klass
2863 const TypeKlassPtr *tkls = tp->isa_klassptr();
2864 if (tkls != nullptr && !StressReflectiveCode) {
2865 if (!tkls->is_loaded())
2866 return _type; // Bail out if not loaded
2867 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2868 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2869 // // Always returning precise element type is incorrect,
2870 // // e.g., element type could be object and array may contain strings
2871 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2872
2873 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2874 // according to the element type's subclassing.
2875 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2876 }
2877 if (tkls->isa_aryklassptr() != nullptr && tkls->klass_is_exact() &&
2878 !tkls->exact_klass()->is_type_array_klass() &&
2879 tkls->offset() == in_bytes(Klass::super_offset())) {
2880 // We are loading the super klass of a refined array klass, return the non-refined klass pointer
2881 assert(tkls->is_aryklassptr()->is_refined_type(), "Must be a refined array klass pointer");
2882 return tkls->is_aryklassptr()->with_offset(0)->cast_to_non_refined();
2883 }
2884 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2885 tkls->offset() == in_bytes(Klass::super_offset())) {
2886 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2887 // The field is Klass::_super. Return its (constant) value.
2888 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2889 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2890 }
2891 }
2892
2893 if (tkls != nullptr && !UseSecondarySupersCache
2894 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2895 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2896 return TypePtr::NULL_PTR;
2897 }
2898
2899 // Bailout case
2900 return LoadNode::Value(phase);
2901 }
2902
2903 //------------------------------Identity---------------------------------------
2904 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2905 // Also feed through the klass in Allocate(...klass...)._klass.
2906 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2907 return klass_identity_common(phase);
2908 }
2909
2910 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2911 Node* x = LoadNode::Identity(phase);
2912 if (x != this) return x;
2913
2914 // Take apart the address into an oop and offset.
2915 // Return 'this' if we cannot.
2916 Node* adr = in(MemNode::Address);
2917 intptr_t offset = 0;
2918 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2919 if (base == nullptr) return this;
2920 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2921 if (toop == nullptr) return this;
2922
2923 // Step over potential GC barrier for OopHandle resolve
2924 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2925 if (bs->is_gc_barrier_node(base)) {
2926 base = bs->step_over_gc_barrier(base);
2927 }
2928
2929 // We can fetch the klass directly through an AllocateNode.
2930 // This works even if the klass is not constant (clone or newArray).
2931 if (offset == oopDesc::klass_offset_in_bytes()) {
2932 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2933 if (allocated_klass != nullptr) {
2934 return allocated_klass;
2935 }
2936 }
2937
2938 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2939 // See inline_native_Class_query for occurrences of these patterns.
2940 // Java Example: x.getClass().isAssignableFrom(y)
2941 //
2942 // This improves reflective code, often making the Class
2943 // mirror go completely dead. (Current exception: Class
2944 // mirrors may appear in debug info, but we could clean them out by
2945 // introducing a new debug info operator for Klass.java_mirror).
2946 //
2947 // This optimization does not apply to arrays because if k is not a
2948 // constant, it was obtained via load_klass which returns the refined type
2949 // and '.java_mirror.as_klass' should return the Java type instead.
2950
2951 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2952 && offset == java_lang_Class::klass_offset()) {
2953 if (base->is_Load()) {
2954 Node* base2 = base->in(MemNode::Address);
2955 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2956 Node* adr2 = base2->in(MemNode::Address);
2957 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2958 if (tkls != nullptr && !tkls->empty()
2959 && ((tkls->isa_instklassptr() && !tkls->is_instklassptr()->might_be_an_array()))
2960 && adr2->is_AddP()) {
2961 int mirror_field = in_bytes(Klass::java_mirror_offset());
2962 if (tkls->offset() == mirror_field) {
2963 #ifdef ASSERT
2964 const TypeKlassPtr* tkls2 = phase->type(adr2->in(AddPNode::Address))->is_klassptr();
2965 assert(tkls2->offset() == 0, "not a load of java_mirror");
2966 #endif
2967 assert(adr2->in(AddPNode::Base)->is_top(), "not an off heap load");
2968 assert(adr2->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(Klass::java_mirror_offset()), "incorrect offset");
2969 return adr2->in(AddPNode::Address);
2970 }
2971 }
2972 }
2973 }
2974 }
2975
2976 return this;
2977 }
2978
2979 LoadNode* LoadNode::clone_pinned() const {
2980 LoadNode* ld = clone()->as_Load();
2981 ld->_control_dependency = UnknownControl;
2982 return ld;
2983 }
2984
2985 // Pin a LoadNode if it carries a dependency on its control input. There are cases when the node
2986 // does not actually have any dependency on its control input. For example, if we have a LoadNode
2987 // being used only outside a loop but it must be scheduled inside the loop, we can clone the node
2988 // for each of its use so that all the clones can be scheduled outside the loop. Then, to prevent
2989 // the clones from being GVN-ed again, we add a control input for each of them at the loop exit. In
2990 // those case, since there is not a dependency between the node and its control input, we do not
2991 // need to pin it.
2992 LoadNode* LoadNode::pin_node_under_control_impl() const {
2993 const TypePtr* adr_type = this->adr_type();
2994 if (adr_type != nullptr && adr_type->isa_aryptr()) {
2995 // Only array accesses have dependencies on their control input
2996 return clone_pinned();
2997 }
2998 return nullptr;
2999 }
3000
3001 //------------------------------Value------------------------------------------
3002 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
3003 const Type *t = klass_value_common(phase);
3004 if (t == Type::TOP)
3005 return t;
3006
3007 return t->make_narrowklass();
3008 }
3009
3010 //------------------------------Identity---------------------------------------
3011 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
3012 // Also feed through the klass in Allocate(...klass...)._klass.
3013 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
3014 Node *x = klass_identity_common(phase);
3015
3016 const Type *t = phase->type( x );
3017 if( t == Type::TOP ) return x;
3018 if( t->isa_narrowklass()) return x;
3019 assert (!t->isa_narrowoop(), "no narrow oop here");
3020
3021 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
3022 }
3023
3024 //------------------------------Value-----------------------------------------
3025 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
3026 // Either input is TOP ==> the result is TOP
3027 const Type *t1 = phase->type( in(MemNode::Memory) );
3028 if( t1 == Type::TOP ) return Type::TOP;
3029 Node *adr = in(MemNode::Address);
3030 const Type *t2 = phase->type( adr );
3031 if( t2 == Type::TOP ) return Type::TOP;
3032 const TypePtr *tp = t2->is_ptr();
3033 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
3034 const TypeAryPtr *tap = tp->isa_aryptr();
3035 if( !tap ) return _type;
3036 return tap->size();
3037 }
3038
3039 //-------------------------------Ideal---------------------------------------
3040 // Feed through the length in AllocateArray(...length...)._length.
3041 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3042 Node* p = MemNode::Ideal_common(phase, can_reshape);
3043 if (p) return (p == NodeSentinel) ? nullptr : p;
3044
3045 // Take apart the address into an oop and offset.
3046 // Return 'this' if we cannot.
3047 Node* adr = in(MemNode::Address);
3048 intptr_t offset = 0;
3049 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3050 if (base == nullptr) return nullptr;
3051 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
3052 if (tary == nullptr) return nullptr;
3053
3054 // We can fetch the length directly through an AllocateArrayNode.
3055 // This works even if the length is not constant (clone or newArray).
3056 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3057 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
3058 if (alloc != nullptr) {
3059 Node* allocated_length = alloc->Ideal_length();
3060 Node* len = alloc->make_ideal_length(tary, phase);
3061 if (allocated_length != len) {
3062 // New CastII improves on this.
3063 return len;
3064 }
3065 }
3066 }
3067
3068 return nullptr;
3069 }
3070
3071 //------------------------------Identity---------------------------------------
3072 // Feed through the length in AllocateArray(...length...)._length.
3073 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
3074 Node* x = LoadINode::Identity(phase);
3075 if (x != this) return x;
3076
3077 // Take apart the address into an oop and offset.
3078 // Return 'this' if we cannot.
3079 Node* adr = in(MemNode::Address);
3080 intptr_t offset = 0;
3081 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3082 if (base == nullptr) return this;
3083 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
3084 if (tary == nullptr) return this;
3085
3086 // We can fetch the length directly through an AllocateArrayNode.
3087 // This works even if the length is not constant (clone or newArray).
3088 if (offset == arrayOopDesc::length_offset_in_bytes()) {
3089 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
3090 if (alloc != nullptr) {
3091 Node* allocated_length = alloc->Ideal_length();
3092 // Do not allow make_ideal_length to allocate a CastII node.
3093 Node* len = alloc->make_ideal_length(tary, phase, false);
3094 if (allocated_length == len) {
3095 // Return allocated_length only if it would not be improved by a CastII.
3096 return allocated_length;
3097 }
3098 }
3099 }
3100
3101 return this;
3102
3103 }
3104
3105 //=============================================================================
3106 //---------------------------StoreNode::make-----------------------------------
3107 // Polymorphic factory method:
3108 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo, bool require_atomic_access) {
3109 assert((mo == unordered || mo == release), "unexpected");
3110 Compile* C = gvn.C;
3111 assert(adr_type == nullptr || adr->is_top() || C->get_alias_index(gvn.type(adr)->is_ptr()) == C->get_alias_index(adr_type), "adr and adr_type must agree");
3112 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
3113 ctl != nullptr, "raw memory operations should have control edge");
3114
3115 switch (bt) {
3116 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
3117 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
3118 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
3119 case T_CHAR:
3120 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
3121 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
3122 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
3123 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
3124 case T_METADATA:
3125 case T_ADDRESS:
3126 case T_OBJECT:
3127 case T_ARRAY:
3128 #ifdef _LP64
3129 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3130 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
3131 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
3132 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
3133 (val->bottom_type()->isa_klassptr() && adr->bottom_type()->isa_rawptr())) {
3134 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
3135 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
3136 }
3137 #endif
3138 {
3139 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
3140 }
3141 default:
3142 guarantee(false, "unexpected basic type %s", type2name(bt));
3143 return (StoreNode*)nullptr;
3144 }
3145 }
3146
3147 //--------------------------bottom_type----------------------------------------
3148 const Type *StoreNode::bottom_type() const {
3149 return Type::MEMORY;
3150 }
3151
3152 //------------------------------hash-------------------------------------------
3153 uint StoreNode::hash() const {
3154 // unroll addition of interesting fields
3155 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
3156
3157 // Since they are not commoned, do not hash them:
3158 return NO_HASH;
3159 }
3160
3161 // Link together multiple stores (B/S/C/I) into a longer one.
3162 //
3163 // Example: _store = StoreB[i+3]
3164 //
3165 // RangeCheck[i+0] RangeCheck[i+0]
3166 // StoreB[i+0]
3167 // RangeCheck[i+3] RangeCheck[i+3]
3168 // StoreB[i+1] --> pass: fail:
3169 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
3170 // StoreB[i+3]
3171 //
3172 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
3173 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
3174 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
3175 // failing path.
3176 //
3177 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
3178 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
3179 // and possibly no RangeChecks remain between the stores.
3180 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
3181 // and those later RangeChecks can be removed. Example:
3182 //
3183 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
3184 // StoreB[i+0] StoreB[i+0] <- first store
3185 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
3186 // StoreB[i+1] StoreB[i+1] <- second store
3187 // RangeCheck[i+2] --> removed
3188 // StoreB[i+2] StoreB[i+2]
3189 // RangeCheck[i+3] --> removed
3190 // StoreB[i+3] StoreB[i+3] <- last store
3191 //
3192 // Thus, it is a common pattern that between the first and last store in a chain
3193 // of adjacent stores there remains exactly one RangeCheck, located between the
3194 // first and the second store (e.g. RangeCheck[i+3]).
3195 //
3196 class MergePrimitiveStores : public StackObj {
3197 private:
3198 PhaseGVN* const _phase;
3199 StoreNode* const _store;
3200 // State machine with initial state Unknown
3201 // Allowed transitions:
3202 // Unknown -> Const
3203 // Unknown -> Platform
3204 // Unknown -> Reverse
3205 // Unknown -> NotAdjacent
3206 // Const -> Const
3207 // Const -> NotAdjacent
3208 // Platform -> Platform
3209 // Platform -> NotAdjacent
3210 // Reverse -> Reverse
3211 // Reverse -> NotAdjacent
3212 // NotAdjacent -> NotAdjacent
3213 enum ValueOrder : uint8_t {
3214 Unknown, // Initial state
3215 Const, // Input values are const
3216 Platform, // Platform order
3217 Reverse, // Reverse platform order
3218 NotAdjacent // Not adjacent
3219 };
3220 ValueOrder _value_order;
3221
3222 NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
3223
3224 public:
3225 MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
3226 _phase(phase), _store(store), _value_order(ValueOrder::Unknown)
3227 NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
3228 {}
3229
3230 StoreNode* run();
3231
3232 private:
3233 bool is_compatible_store(const StoreNode* other_store) const;
3234 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
3235 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
3236 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase);
3237 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
3238 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
3239
3240 class Status {
3241 private:
3242 StoreNode* _found_store;
3243 bool _found_range_check;
3244
3245 Status(StoreNode* found_store, bool found_range_check)
3246 : _found_store(found_store), _found_range_check(found_range_check) {}
3247
3248 public:
3249 StoreNode* found_store() const { return _found_store; }
3250 bool found_range_check() const { return _found_range_check; }
3251 static Status make_failure() { return Status(nullptr, false); }
3252
3253 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
3254 if (cfg_status == CFGStatus::Failure) {
3255 return Status::make_failure();
3256 }
3257 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
3258 }
3259
3260 #ifndef PRODUCT
3261 void print_on(outputStream* st) const {
3262 if (_found_store == nullptr) {
3263 st->print_cr("None");
3264 } else {
3265 st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
3266 _found_range_check ? "RC" : "no-RC");
3267 }
3268 }
3269 #endif
3270 };
3271
3272 enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const;
3273 Status find_adjacent_use_store(const StoreNode* def_store) const;
3274 Status find_adjacent_def_store(const StoreNode* use_store) const;
3275 Status find_use_store(const StoreNode* def_store) const;
3276 Status find_def_store(const StoreNode* use_store) const;
3277 Status find_use_store_unidirectional(const StoreNode* def_store) const;
3278 Status find_def_store_unidirectional(const StoreNode* use_store) const;
3279
3280 void collect_merge_list(Node_List& merge_list) const;
3281 Node* make_merged_input_value(const Node_List& merge_list);
3282 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
3283
3284 #ifndef PRODUCT
3285 // Access to TraceMergeStores tags
3286 bool is_trace(TraceMergeStores::Tag tag) const {
3287 return _trace_tags.at(tag);
3288 }
3289
3290 bool is_trace_basic() const {
3291 return is_trace(TraceMergeStores::Tag::BASIC);
3292 }
3293
3294 bool is_trace_pointer_parsing() const {
3295 return is_trace(TraceMergeStores::Tag::POINTER_PARSING);
3296 }
3297
3298 bool is_trace_pointer_aliasing() const {
3299 return is_trace(TraceMergeStores::Tag::POINTER_ALIASING);
3300 }
3301
3302 bool is_trace_pointer_adjacency() const {
3303 return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY);
3304 }
3305
3306 bool is_trace_success() const {
3307 return is_trace(TraceMergeStores::Tag::SUCCESS);
3308 }
3309 #endif
3310
3311 NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
3312 };
3313
3314 StoreNode* MergePrimitiveStores::run() {
3315 // Check for B/S/C/I
3316 int opc = _store->Opcode();
3317 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
3318 return nullptr;
3319 }
3320
3321 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
3322
3323 // The _store must be the "last" store in a chain. If we find a use we could merge with
3324 // then that use or a store further down is the "last" store.
3325 Status status_use = find_adjacent_use_store(_store);
3326 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
3327 if (status_use.found_store() != nullptr) {
3328 return nullptr;
3329 }
3330
3331 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
3332 Status status_def = find_adjacent_def_store(_store);
3333 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
3334 Node* def_store = status_def.found_store();
3335 if (def_store == nullptr) {
3336 return nullptr;
3337 }
3338
3339 // Initialize value order
3340 _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn),
3341 _store->in(MemNode::ValueIn),
3342 _store->memory_size());
3343 assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked");
3344
3345 ResourceMark rm;
3346 Node_List merge_list;
3347 collect_merge_list(merge_list);
3348
3349 Node* merged_input_value = make_merged_input_value(merge_list);
3350 if (merged_input_value == nullptr) { return nullptr; }
3351
3352 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3353
3354 NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
3355
3356 return merged_store;
3357 }
3358
3359 // Check compatibility between _store and other_store.
3360 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
3361 int opc = _store->Opcode();
3362 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3363
3364 if (other_store == nullptr ||
3365 _store->Opcode() != other_store->Opcode()) {
3366 return false;
3367 }
3368
3369 return true;
3370 }
3371
3372 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3373 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3374 use_store->in(MemNode::ValueIn),
3375 def_store->memory_size())) {
3376 return false;
3377 }
3378
3379 ResourceMark rm;
3380 #ifndef PRODUCT
3381 const TraceMemPointer trace(is_trace_pointer_parsing(),
3382 is_trace_pointer_aliasing(),
3383 is_trace_pointer_adjacency(),
3384 true);
3385 #endif
3386 const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace));
3387 const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace));
3388 return pointer_def.is_adjacent_to_and_before(pointer_use);
3389 }
3390
3391 // Check input values n1 and n2 can be merged and return the value order
3392 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2,
3393 const int memory_size) const {
3394 // Pattern: [n1 = ConI, n2 = ConI]
3395 if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) {
3396 return ValueOrder::Const;
3397 }
3398
3399 Node const *base_n2;
3400 jint shift_n2;
3401 if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) {
3402 return ValueOrder::NotAdjacent;
3403 }
3404 Node const *base_n1;
3405 jint shift_n1;
3406 if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) {
3407 return ValueOrder::NotAdjacent;
3408 }
3409
3410 int bits_per_store = memory_size * 8;
3411 if (base_n1 != base_n2 ||
3412 abs(shift_n1 - shift_n2) != bits_per_store ||
3413 shift_n1 % bits_per_store != 0) {
3414 // Values are not adjacent
3415 return ValueOrder::NotAdjacent;
3416 }
3417
3418 // Detect value order
3419 #ifdef VM_LITTLE_ENDIAN
3420 return shift_n1 < shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3421 : ValueOrder::Reverse; // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3422 #else
3423 return shift_n1 > shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3424 : ValueOrder::Reverse; // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3425 #endif
3426 }
3427
3428 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3429 ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size);
3430
3431 switch (input_value_order) {
3432 case ValueOrder::NotAdjacent:
3433 return false;
3434 case ValueOrder::Reverse:
3435 if (memory_size != 1 ||
3436 !Matcher::match_rule_supported(Op_ReverseBytesS) ||
3437 !Matcher::match_rule_supported(Op_ReverseBytesI) ||
3438 !Matcher::match_rule_supported(Op_ReverseBytesL)) {
3439 // ReverseBytes are not supported by platform
3440 return false;
3441 }
3442 // fall-through.
3443 case ValueOrder::Const:
3444 case ValueOrder::Platform:
3445 if (_value_order == ValueOrder::Unknown) {
3446 // Initial state is Unknown, and we find a valid input value order
3447 return true;
3448 }
3449 // The value order can not be changed
3450 return _value_order == input_value_order;
3451 case ValueOrder::Unknown:
3452 default:
3453 ShouldNotReachHere();
3454 }
3455 return false;
3456 }
3457
3458 // Detect pattern: n = base_out >> shift_out
3459 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) {
3460 assert(n != nullptr, "precondition");
3461
3462 int opc = n->Opcode();
3463 if (opc == Op_ConvL2I) {
3464 n = n->in(1);
3465 opc = n->Opcode();
3466 }
3467
3468 if ((opc == Op_RShiftI ||
3469 opc == Op_RShiftL ||
3470 opc == Op_URShiftI ||
3471 opc == Op_URShiftL) &&
3472 n->in(2)->is_ConI()) {
3473 base_out = n->in(1);
3474 shift_out = n->in(2)->get_int();
3475 // The shift must be positive:
3476 return shift_out >= 0;
3477 }
3478
3479 if (phase->type(n)->isa_int() != nullptr ||
3480 phase->type(n)->isa_long() != nullptr) {
3481 // (base >> 0)
3482 base_out = n;
3483 shift_out = 0;
3484 return true;
3485 }
3486 return false;
3487 }
3488
3489 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3490 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3491 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3492
3493 Node* ctrl_use = use_store->in(MemNode::Control);
3494 Node* ctrl_def = def_store->in(MemNode::Control);
3495 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3496 return CFGStatus::Failure;
3497 }
3498
3499 if (ctrl_use == ctrl_def) {
3500 // Same ctrl -> no RangeCheck in between.
3501 // Check: use_store must be the only use of def_store.
3502 if (def_store->outcnt() > 1) {
3503 return CFGStatus::Failure;
3504 }
3505 return CFGStatus::SuccessNoRangeCheck;
3506 }
3507
3508 // Different ctrl -> could have RangeCheck in between.
3509 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3510 // 2. ctrl separated by RangeCheck.
3511 if (def_store->outcnt() != 2) {
3512 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3513 }
3514 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3515 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3516 if (merge_mem == nullptr ||
3517 merge_mem->outcnt() != 1) {
3518 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3519 }
3520 if (!ctrl_use->is_IfProj() ||
3521 !ctrl_use->in(0)->is_RangeCheck() ||
3522 ctrl_use->in(0)->outcnt() != 2) {
3523 return CFGStatus::Failure; // Not RangeCheck.
3524 }
3525 IfProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3526 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3527 if (trap != merge_mem->unique_out() ||
3528 ctrl_use->in(0)->in(0) != ctrl_def) {
3529 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3530 }
3531
3532 return CFGStatus::SuccessWithRangeCheck;
3533 }
3534
3535 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3536 Status status_use = find_use_store(def_store);
3537 StoreNode* use_store = status_use.found_store();
3538 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3539 return Status::make_failure();
3540 }
3541 return status_use;
3542 }
3543
3544 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3545 Status status_def = find_def_store(use_store);
3546 StoreNode* def_store = status_def.found_store();
3547 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3548 return Status::make_failure();
3549 }
3550 return status_def;
3551 }
3552
3553 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3554 Status status_use = find_use_store_unidirectional(def_store);
3555
3556 #ifdef ASSERT
3557 StoreNode* use_store = status_use.found_store();
3558 if (use_store != nullptr) {
3559 Status status_def = find_def_store_unidirectional(use_store);
3560 assert(status_def.found_store() == def_store &&
3561 status_def.found_range_check() == status_use.found_range_check(),
3562 "find_use_store and find_def_store must be symmetric");
3563 }
3564 #endif
3565
3566 return status_use;
3567 }
3568
3569 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3570 Status status_def = find_def_store_unidirectional(use_store);
3571
3572 #ifdef ASSERT
3573 StoreNode* def_store = status_def.found_store();
3574 if (def_store != nullptr) {
3575 Status status_use = find_use_store_unidirectional(def_store);
3576 assert(status_use.found_store() == use_store &&
3577 status_use.found_range_check() == status_def.found_range_check(),
3578 "find_use_store and find_def_store must be symmetric");
3579 }
3580 #endif
3581
3582 return status_def;
3583 }
3584
3585 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3586 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3587
3588 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3589 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3590 if (is_compatible_store(use_store)) {
3591 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3592 }
3593 }
3594
3595 return Status::make_failure();
3596 }
3597
3598 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3599 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3600
3601 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3602 if (!is_compatible_store(def_store)) {
3603 return Status::make_failure();
3604 }
3605
3606 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3607 }
3608
3609 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3610 // The merged store can be at most 8 bytes.
3611 const uint merge_list_max_size = 8 / _store->memory_size();
3612 assert(merge_list_max_size >= 2 &&
3613 merge_list_max_size <= 8 &&
3614 is_power_of_2(merge_list_max_size),
3615 "must be 2, 4 or 8");
3616
3617 // Traverse up the chain of adjacent def stores.
3618 StoreNode* current = _store;
3619 merge_list.push(current);
3620 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3621 Status status = find_adjacent_def_store(current);
3622 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3623
3624 current = status.found_store();
3625 if (current != nullptr) {
3626 merge_list.push(current);
3627
3628 // We can have at most one RangeCheck.
3629 if (status.found_range_check()) {
3630 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3631 break;
3632 }
3633 }
3634 }
3635
3636 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3637
3638 // Truncate the merge_list to a power of 2.
3639 const uint pow2size = round_down_power_of_2(merge_list.size());
3640 assert(pow2size >= 2, "must be merging at least 2 stores");
3641 while (merge_list.size() > pow2size) { merge_list.pop(); }
3642
3643 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3644 }
3645
3646 // Merge the input values of the smaller stores to a single larger input value.
3647 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3648 int new_memory_size = _store->memory_size() * merge_list.size();
3649 Node* first = merge_list.at(merge_list.size()-1);
3650 Node* merged_input_value = nullptr;
3651 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3652 assert(_value_order == ValueOrder::Const, "must match");
3653 // Pattern: [ConI, ConI, ...] -> new constant
3654 jlong con = 0;
3655 jlong bits_per_store = _store->memory_size() * 8;
3656 jlong mask = (((jlong)1) << bits_per_store) - 1;
3657 for (uint i = 0; i < merge_list.size(); i++) {
3658 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3659 #ifdef VM_LITTLE_ENDIAN
3660 con = con << bits_per_store;
3661 con = con | (mask & con_i);
3662 #else // VM_LITTLE_ENDIAN
3663 con_i = (mask & con_i) << (i * bits_per_store);
3664 con = con | con_i;
3665 #endif // VM_LITTLE_ENDIAN
3666 }
3667 merged_input_value = _phase->longcon(con);
3668 } else {
3669 assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match");
3670 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3671 // | |
3672 // _store first
3673 //
3674 Node* hi = _store->in(MemNode::ValueIn);
3675 Node* lo = first->in(MemNode::ValueIn);
3676 #ifndef VM_LITTLE_ENDIAN
3677 // `_store` and `first` are swapped in the diagram above
3678 swap(hi, lo);
3679 #endif // !VM_LITTLE_ENDIAN
3680 if (_value_order == ValueOrder::Reverse) {
3681 swap(hi, lo);
3682 }
3683 Node const* hi_base;
3684 jint hi_shift;
3685 merged_input_value = lo;
3686 bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase);
3687 assert(is_true, "must detect con RShift");
3688 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3689 // look through
3690 merged_input_value = merged_input_value->in(1);
3691 }
3692 if (merged_input_value != hi_base) {
3693 // merged_input_value is not the base
3694 return nullptr;
3695 }
3696 }
3697
3698 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3699 // Example:
3700 //
3701 // long base = ...;
3702 // a[0] = (byte)(base >> 0);
3703 // a[1] = (byte)(base >> 8);
3704 //
3705 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3706 }
3707
3708 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3709 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3710 "merged_input_value is either int or long, and new_memory_size is small enough");
3711
3712 if (_value_order == ValueOrder::Reverse) {
3713 assert(_store->memory_size() == 1, "only implemented for bytes");
3714 if (new_memory_size == 8) {
3715 merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value));
3716 } else if (new_memory_size == 4) {
3717 merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value));
3718 } else {
3719 assert(new_memory_size == 2, "sanity check");
3720 merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value));
3721 }
3722 }
3723 return merged_input_value;
3724 }
3725
3726 // //
3727 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3728 // | | | | | | //
3729 // | | | | +---------------+ | //
3730 // | | | | | | | //
3731 // | | +---------+ | | +---------------+ //
3732 // | | | | | | | | //
3733 // +--------------+ | | v1 +------------------------------+ | | v1 //
3734 // | | | | | | | | | | | | //
3735 // RangeCheck first_store RangeCheck | | first_store //
3736 // | | | | | | | //
3737 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3738 // | | ===> | | | //
3739 // +--------------+ | a2 v2 | | | //
3740 // | | | | | | | | //
3741 // | second_store | | | //
3742 // | | | | | [v1 v2 ... vn] //
3743 // ... ... | | | | //
3744 // | | | | | v //
3745 // +--------------+ | an vn +--------------+ | | merged_input_value //
3746 // | | | | | | | | //
3747 // last_store (= _store) merged_store //
3748 // //
3749 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3750 Node* first_store = merge_list.at(merge_list.size()-1);
3751 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3752 Node* first_mem = first_store->in(MemNode::Memory);
3753 Node* first_adr = first_store->in(MemNode::Address);
3754
3755 const TypePtr* new_adr_type = _store->adr_type();
3756
3757 int new_memory_size = _store->memory_size() * merge_list.size();
3758 BasicType bt = T_ILLEGAL;
3759 switch (new_memory_size) {
3760 case 2: bt = T_SHORT; break;
3761 case 4: bt = T_INT; break;
3762 case 8: bt = T_LONG; break;
3763 }
3764
3765 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3766 new_adr_type, merged_input_value, bt, MemNode::unordered);
3767
3768 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3769 // are all on the same slice. Hence, we need no barriers.
3770 merged_store->set_mismatched_access();
3771
3772 // Constants above may now also be be packed -> put candidate on worklist
3773 _phase->is_IterGVN()->_worklist.push(first_mem);
3774
3775 return merged_store;
3776 }
3777
3778 #ifndef PRODUCT
3779 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3780 stringStream ss;
3781 ss.print_cr("[TraceMergeStores]: Replace");
3782 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3783 merge_list.at(i)->dump("\n", false, &ss);
3784 }
3785 ss.print_cr("[TraceMergeStores]: with");
3786 merged_input_value->dump("\n", false, &ss);
3787 merged_store->dump("\n", false, &ss);
3788 tty->print("%s", ss.as_string());
3789 }
3790 #endif
3791
3792 //------------------------------Ideal------------------------------------------
3793 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3794 // When a store immediately follows a relevant allocation/initialization,
3795 // try to capture it into the initialization, or hoist it above.
3796 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3797 Node* p = MemNode::Ideal_common(phase, can_reshape);
3798 if (p) return (p == NodeSentinel) ? nullptr : p;
3799
3800 Node* mem = in(MemNode::Memory);
3801 Node* address = in(MemNode::Address);
3802 Node* value = in(MemNode::ValueIn);
3803 // Back-to-back stores to same address? Fold em up. Generally
3804 // unsafe if I have intervening uses...
3805 if ((!this->is_StoreVector() || this->Opcode() == Op_StoreVector) &&
3806 phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::INLINES) {
3807 Node* st = mem;
3808 // If Store 'st' has more than one use, we cannot fold 'st' away.
3809 // For example, 'st' might be the final state at a conditional
3810 // return. Or, 'st' might be used by some node which is live at
3811 // the same time 'st' is live, which might be unschedulable. So,
3812 // require exactly ONE user until such time as we clone 'mem' for
3813 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3814 // true). Further, 'st' must be a contiguous store, otherwise
3815 // memory_size does not make sense for measuring overlap.
3816 while (st->is_Store() && st->outcnt() == 1 && (!st->is_StoreVector() || st->Opcode() == Op_StoreVector)) {
3817 // Looking at a dead closed cycle of memory?
3818 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3819 assert(Opcode() == st->Opcode() ||
3820 st->Opcode() == Op_StoreVector ||
3821 Opcode() == Op_StoreVector ||
3822 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3823 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3824 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3825 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
3826 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3827 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3828
3829 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3830 st->as_Store()->memory_size() <= this->memory_size()) {
3831 assert(!is_predicated_vector() && !is_StoreVectorMasked() &&
3832 !is_StoreVectorScatter() && !is_StoreVectorScatterMasked() &&
3833 !st->is_predicated_vector() && !st->is_StoreVectorMasked() &&
3834 !st->is_StoreVectorScatter() && !st->is_StoreVectorScatterMasked(),
3835 "optimization only correct for full-width stores without holes");
3836 Node* use = st->raw_out(0);
3837 if (phase->is_IterGVN()) {
3838 phase->is_IterGVN()->rehash_node_delayed(use);
3839 }
3840 // It's OK to do this in the parser, since DU info is always accurate,
3841 // and the parser always refers to nodes via SafePointNode maps.
3842 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3843 return this;
3844 }
3845 st = st->in(MemNode::Memory);
3846 }
3847 }
3848
3849
3850 // Capture an unaliased, unconditional, simple store into an initializer.
3851 // Or, if it is independent of the allocation, hoist it above the allocation.
3852 if (ReduceFieldZeroing && /*can_reshape &&*/
3853 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3854 InitializeNode* init = mem->in(0)->as_Initialize();
3855 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3856 if (offset > 0) {
3857 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3858 // If the InitializeNode captured me, it made a raw copy of me,
3859 // and I need to disappear.
3860 if (moved != nullptr) {
3861 // %%% hack to ensure that Ideal returns a new node:
3862 mem = MergeMemNode::make(mem);
3863 return mem; // fold me away
3864 }
3865 }
3866 }
3867
3868 // Fold reinterpret cast into memory operation:
3869 // StoreX mem (MoveY2X v) => StoreY mem v
3870 if (value->is_Move()) {
3871 const Type* vt = value->in(1)->bottom_type();
3872 if (has_reinterpret_variant(vt)) {
3873 if (phase->C->post_loop_opts_phase()) {
3874 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3875 } else {
3876 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3877 }
3878 }
3879 }
3880
3881 if (MergeStores && UseUnalignedAccesses) {
3882 if (phase->C->merge_stores_phase()) {
3883 MergePrimitiveStores merge(phase, this);
3884 Node* progress = merge.run();
3885 if (progress != nullptr) { return progress; }
3886 } else {
3887 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
3888 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
3889 // the stores, and we merge the wrong sequence of stores.
3890 // Example:
3891 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
3892 // Apply MergeStores:
3893 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
3894 // Remove more RangeChecks:
3895 // StoreI [ StoreL ] StoreI
3896 // But now it would have been better to do this instead:
3897 // [ StoreL ] [ StoreL ]
3898 phase->C->record_for_merge_stores_igvn(this);
3899 }
3900 }
3901
3902 return nullptr; // No further progress
3903 }
3904
3905 //------------------------------Value-----------------------------------------
3906 const Type* StoreNode::Value(PhaseGVN* phase) const {
3907 // Either input is TOP ==> the result is TOP
3908 const Type *t1 = phase->type( in(MemNode::Memory) );
3909 if( t1 == Type::TOP ) return Type::TOP;
3910 const Type *t2 = phase->type( in(MemNode::Address) );
3911 if( t2 == Type::TOP ) return Type::TOP;
3912 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3913 if( t3 == Type::TOP ) return Type::TOP;
3914 return Type::MEMORY;
3915 }
3916
3917 //------------------------------Identity---------------------------------------
3918 // Remove redundant stores:
3919 // Store(m, p, Load(m, p)) changes to m.
3920 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3921 Node* StoreNode::Identity(PhaseGVN* phase) {
3922 Node* mem = in(MemNode::Memory);
3923 Node* adr = in(MemNode::Address);
3924 Node* val = in(MemNode::ValueIn);
3925
3926 Node* result = this;
3927
3928 // Load then Store? Then the Store is useless
3929 if (val->is_Load() &&
3930 val->in(MemNode::Address)->eqv_uncast(adr) &&
3931 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3932 val->as_Load()->store_Opcode() == Opcode()) {
3933 if (!is_StoreVector()) {
3934 result = mem;
3935 } else if (Opcode() == Op_StoreVector && val->Opcode() == Op_LoadVector &&
3936 as_StoreVector()->vect_type() == val->as_LoadVector()->vect_type()) {
3937 // Ensure both are not masked accesses or gathers/scatters and vector types are the same
3938 result = mem;
3939 }
3940 }
3941
3942 // Two stores in a row of the same value?
3943 if (result == this &&
3944 mem->is_Store() &&
3945 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3946 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3947 mem->Opcode() == Opcode()) {
3948 if (!is_StoreVector()) {
3949 result = mem;
3950 } else {
3951 const StoreVectorNode* store_vector = as_StoreVector();
3952 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3953 const Node* store_indices = store_vector->indices();
3954 const Node* mem_indices = mem_vector->indices();
3955 const Node* store_mask = store_vector->mask();
3956 const Node* mem_mask = mem_vector->mask();
3957 // Ensure types, indices, and masks match
3958 if (store_vector->vect_type() == mem_vector->vect_type() &&
3959 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3960 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3961 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3962 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3963 result = mem;
3964 }
3965 }
3966 }
3967
3968 // Store of zero anywhere into a freshly-allocated object?
3969 // Then the store is useless.
3970 // (It must already have been captured by the InitializeNode.)
3971 if (result == this && ReduceFieldZeroing) {
3972 // a newly allocated object is already all-zeroes everywhere
3973 if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
3974 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::InitValue) == val)) {
3975 result = mem;
3976 }
3977
3978 if (result == this && phase->type(val)->is_zero_type()) {
3979 // the store may also apply to zero-bits in an earlier object
3980 Node* prev_mem = find_previous_store(phase);
3981 // Steps (a), (b): Walk past independent stores to find an exact match.
3982 if (prev_mem != nullptr) {
3983 if (prev_mem->is_top()) {
3984 // find_previous_store returns top when the access is dead
3985 return prev_mem;
3986 }
3987 Node* prev_val = can_see_stored_value(prev_mem, phase);
3988 if (prev_val != nullptr && prev_val == val) {
3989 // prev_val and val might differ by a cast; it would be good
3990 // to keep the more informative of the two.
3991 result = mem;
3992 }
3993 }
3994 }
3995 }
3996
3997 PhaseIterGVN* igvn = phase->is_IterGVN();
3998 if (result != this && igvn != nullptr) {
3999 MemBarNode* trailing = trailing_membar();
4000 if (trailing != nullptr) {
4001 #ifdef ASSERT
4002 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
4003 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
4004 #endif
4005 trailing->remove(igvn);
4006 }
4007 }
4008
4009 return result;
4010 }
4011
4012 //------------------------------match_edge-------------------------------------
4013 // Do we Match on this edge index or not? Match only memory & value
4014 uint StoreNode::match_edge(uint idx) const {
4015 return idx == MemNode::Address || idx == MemNode::ValueIn;
4016 }
4017
4018 //------------------------------cmp--------------------------------------------
4019 // Do not common stores up together. They generally have to be split
4020 // back up anyways, so do not bother.
4021 bool StoreNode::cmp( const Node &n ) const {
4022 return (&n == this); // Always fail except on self
4023 }
4024
4025 //------------------------------Ideal_masked_input-----------------------------
4026 // Check for a useless mask before a partial-word store
4027 // (StoreB ... (AndI valIn conIa) )
4028 // If (conIa & mask == mask) this simplifies to
4029 // (StoreB ... (valIn) )
4030 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
4031 Node *val = in(MemNode::ValueIn);
4032 if( val->Opcode() == Op_AndI ) {
4033 const TypeInt *t = phase->type( val->in(2) )->isa_int();
4034 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
4035 set_req_X(MemNode::ValueIn, val->in(1), phase);
4036 return this;
4037 }
4038 }
4039 return nullptr;
4040 }
4041
4042
4043 //------------------------------Ideal_sign_extended_input----------------------
4044 // Check for useless sign-extension before a partial-word store
4045 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
4046 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to
4047 // (StoreB ... (v))
4048 // If (conIL > conIR) under some conditions, it can be simplified into
4049 // (StoreB ... (LShiftI _ v (conIL - conIR)))
4050 // This case happens when the value of the store was itself a left shift, that
4051 // gets merged into the inner left shift of the sign-extension. For instance,
4052 // if we have
4053 // array_of_shorts[0] = (short)(v << 2)
4054 // We get a structure such as:
4055 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16))
4056 // that is simplified into
4057 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)).
4058 // It is thus useful to handle cases where conIL > conIR. But this simplification
4059 // does not always hold. Let's see in which cases it's valid.
4060 //
4061 // Let's assume we have the following 32 bits integer v:
4062 // +----------------------------------+
4063 // | v[0..31] |
4064 // +----------------------------------+
4065 // 31 0
4066 // that will be stuffed in 8 bits byte after a shift left and a shift right of
4067 // potentially different magnitudes.
4068 // We denote num_rejected_bits the number of bits of the discarded part. In this
4069 // case, num_rejected_bits == 24.
4070 //
4071 // Statement (proved further below in case analysis):
4072 // Given:
4073 // - 0 <= conIL < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
4074 // - 0 <= conIR < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
4075 // - conIL >= conIR
4076 // - num_rejected_bits >= conIR
4077 // Then this form:
4078 // (RShiftI _ (LShiftI _ v conIL) conIR)
4079 // can be replaced with this form:
4080 // (LShiftI _ v (conIL-conIR))
4081 //
4082 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte)
4083 // have to be correct, as the higher bits are rejected / truncated by the store.
4084 //
4085 // The hypotheses
4086 // 0 <= conIL < BitsPerJavaInteger
4087 // 0 <= conIR < BitsPerJavaInteger
4088 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed,
4089 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent
4090 // to v << 1).
4091 //
4092 //
4093 // If you don't like case analysis, jump after the conclusion.
4094 // ### Case 1 : conIL == conIR
4095 // ###### Case 1.1: conIL == conIR == num_rejected_bits
4096 // If we do the shift left then right by 24 bits, we get:
4097 // after: v << 24
4098 // +---------+------------------------+
4099 // | v[0..7] | 0 |
4100 // +---------+------------------------+
4101 // 31 24 23 0
4102 // after: (v << 24) >> 24
4103 // +------------------------+---------+
4104 // | sign bit | v[0..7] |
4105 // +------------------------+---------+
4106 // 31 8 7 0
4107 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the
4108 // result) are the same before and after, so, indeed, simplifying is correct.
4109
4110 // ###### Case 1.2: conIL == conIR < num_rejected_bits
4111 // If we do the shift left then right by 22 bits, we get:
4112 // after: v << 22
4113 // +---------+------------------------+
4114 // | v[0..9] | 0 |
4115 // +---------+------------------------+
4116 // 31 22 21 0
4117 // after: (v << 22) >> 22
4118 // +------------------------+---------+
4119 // | sign bit | v[0..9] |
4120 // +------------------------+---------+
4121 // 31 10 9 0
4122 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still
4123 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is
4124 // still correct.
4125
4126 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits
4127 // If we do the shift left then right by 26 bits, we get:
4128 // after: v << 26
4129 // +---------+------------------------+
4130 // | v[0..5] | 0 |
4131 // +---------+------------------------+
4132 // 31 26 25 0
4133 // after: (v << 26) >> 26
4134 // +------------------------+---------+
4135 // | sign bit | v[0..5] |
4136 // +------------------------+---------+
4137 // 31 6 5 0
4138 // The non-rejected bits are made of
4139 // - 0-5 => the bits 0 to 5 of v
4140 // - 6-7 => the sign bit of v[0..5] (that is v[5])
4141 // Simplifying this as v is not correct.
4142 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1
4143 //
4144 // ### Case 2: conIL > conIR
4145 // ###### Case 2.1: num_rejected_bits == conIR
4146 // We take conIL == 26 for this example.
4147 // after: v << 26
4148 // +---------+------------------------+
4149 // | v[0..5] | 0 |
4150 // +---------+------------------------+
4151 // 31 26 25 0
4152 // after: (v << 26) >> 24
4153 // +------------------+---------+-----+
4154 // | sign bit | v[0..5] | 0 |
4155 // +------------------+---------+-----+
4156 // 31 8 7 2 1 0
4157 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR).
4158 // The bits 6 and 7 of v have been thrown away after the shift left.
4159 // The simplification is still correct.
4160 //
4161 // ###### Case 2.2: num_rejected_bits > conIR.
4162 // Let's say conIL == 26 and conIR == 22.
4163 // after: v << 26
4164 // +---------+------------------------+
4165 // | v[0..5] | 0 |
4166 // +---------+------------------------+
4167 // 31 26 25 0
4168 // after: (v << 26) >> 22
4169 // +------------------+---------+-----+
4170 // | sign bit | v[0..5] | 0 |
4171 // +------------------+---------+-----+
4172 // 31 10 9 4 3 0
4173 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)):
4174 // - 0-3 => 0
4175 // - 4-7 => bits 0 to 3 of v
4176 // The simplification is still correct.
4177 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't
4178 // matter as they are not in the 8 lower bits: they will be cut out by the store.
4179 //
4180 // ###### But! Case 2.3: num_rejected_bits < conIR.
4181 // Let's see that this case is not as easy to simplify.
4182 // Let's say conIL == 28 and conIR == 26.
4183 // after: v << 28
4184 // +---------+------------------------+
4185 // | v[0..3] | 0 |
4186 // +---------+------------------------+
4187 // 31 28 27 0
4188 // after: (v << 28) >> 26
4189 // +------------------+---------+-----+
4190 // | sign bit | v[0..3] | 0 |
4191 // +------------------+---------+-----+
4192 // 31 6 5 2 1 0
4193 // The non-rejected bits are made of
4194 // - 0-1 => 0
4195 // - 2-5 => the bits 0 to 3 of v
4196 // - 6-7 => the sign bit of v[0..3] (that is v[3])
4197 // Simplifying this as (v << 2) is not correct.
4198 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2.
4199 //
4200 // ### Conclusion:
4201 // Our hypotheses are indeed sufficient:
4202 // - 0 <= conIL < BitsPerJavaInteger
4203 // - 0 <= conIR < BitsPerJavaInteger
4204 // - conIL >= conIR
4205 // - num_rejected_bits >= conIR
4206 //
4207 // ### A rationale without case analysis:
4208 // After the shift left, conIL upper bits of v are discarded and conIL lower bit
4209 // zeroes are added. After the shift right, conIR lower bits of the previous result
4210 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during
4211 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But
4212 // the point of the simplification is to get an expression of the form
4213 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the
4214 // simplification is not correct if conIL < conIR.
4215 //
4216 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the
4217 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If
4218 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will
4219 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to
4220 // (v << (conIL-conIR)) if conIR > num_rejected_bits.
4221 //
4222 // We do not treat the case conIL < conIR here since the point of this function is
4223 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need
4224 // of treating conIL > conIR comes from the cases where the sign-extended value is
4225 // also left-shift expression. Computing the sign-extension of a right-shift expression
4226 // doesn't yield a situation such as
4227 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
4228 // where conIL < conIR.
4229 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) {
4230 Node* shr = in(MemNode::ValueIn);
4231 if (shr->Opcode() == Op_RShiftI) {
4232 const TypeInt* conIR = phase->type(shr->in(2))->isa_int();
4233 if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) {
4234 Node* shl = shr->in(1);
4235 if (shl->Opcode() == Op_LShiftI) {
4236 const TypeInt* conIL = phase->type(shl->in(2))->isa_int();
4237 if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) {
4238 if (conIL->get_con() == conIR->get_con()) {
4239 set_req_X(MemNode::ValueIn, shl->in(1), phase);
4240 return this;
4241 }
4242 if (conIL->get_con() > conIR->get_con()) {
4243 Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con())));
4244 set_req_X(MemNode::ValueIn, new_shl, phase);
4245 return this;
4246 }
4247 }
4248 }
4249 }
4250 }
4251 return nullptr;
4252 }
4253
4254 //------------------------------value_never_loaded-----------------------------------
4255 // Determine whether there are any possible loads of the value stored.
4256 // For simplicity, we actually check if there are any loads from the
4257 // address stored to, not just for loads of the value stored by this node.
4258 //
4259 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
4260 Node *adr = in(Address);
4261 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
4262 if (adr_oop == nullptr)
4263 return false;
4264 if (!adr_oop->is_known_instance_field())
4265 return false; // if not a distinct instance, there may be aliases of the address
4266 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
4267 Node *use = adr->fast_out(i);
4268 if (use->is_Load() || use->is_LoadStore()) {
4269 return false;
4270 }
4271 }
4272 return true;
4273 }
4274
4275 MemBarNode* StoreNode::trailing_membar() const {
4276 if (is_release()) {
4277 MemBarNode* trailing_mb = nullptr;
4278 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4279 Node* u = fast_out(i);
4280 if (u->is_MemBar()) {
4281 if (u->as_MemBar()->trailing_store()) {
4282 assert(u->Opcode() == Op_MemBarVolatile, "");
4283 assert(trailing_mb == nullptr, "only one");
4284 trailing_mb = u->as_MemBar();
4285 #ifdef ASSERT
4286 Node* leading = u->as_MemBar()->leading_membar();
4287 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4288 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
4289 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
4290 #endif
4291 } else {
4292 assert(u->as_MemBar()->standalone(), "");
4293 }
4294 }
4295 }
4296 return trailing_mb;
4297 }
4298 return nullptr;
4299 }
4300
4301
4302 //=============================================================================
4303 //------------------------------Ideal------------------------------------------
4304 // If the store is from an AND mask that leaves the low bits untouched, then
4305 // we can skip the AND operation. If the store is from a sign-extension
4306 // (a left shift, then right shift) we can skip both.
4307 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
4308 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
4309 if( progress != nullptr ) return progress;
4310
4311 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
4312 if( progress != nullptr ) return progress;
4313
4314 // Finally check the default case
4315 return StoreNode::Ideal(phase, can_reshape);
4316 }
4317
4318 //=============================================================================
4319 //------------------------------Ideal------------------------------------------
4320 // If the store is from an AND mask that leaves the low bits untouched, then
4321 // we can skip the AND operation
4322 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
4323 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
4324 if( progress != nullptr ) return progress;
4325
4326 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
4327 if( progress != nullptr ) return progress;
4328
4329 // Finally check the default case
4330 return StoreNode::Ideal(phase, can_reshape);
4331 }
4332
4333 //=============================================================================
4334 //----------------------------------SCMemProjNode------------------------------
4335 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
4336 {
4337 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
4338 return Type::TOP;
4339 }
4340 return bottom_type();
4341 }
4342
4343 //=============================================================================
4344 //----------------------------------LoadStoreNode------------------------------
4345 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
4346 : Node(required),
4347 _type(rt),
4348 _barrier_data(0)
4349 {
4350 init_req(MemNode::Control, c );
4351 init_req(MemNode::Memory , mem);
4352 init_req(MemNode::Address, adr);
4353 init_req(MemNode::ValueIn, val);
4354 init_class_id(Class_LoadStore);
4355 DEBUG_ONLY(_adr_type = at; adr_type();)
4356 }
4357
4358 //------------------------------Value-----------------------------------------
4359 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
4360 // Either input is TOP ==> the result is TOP
4361 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
4362 return Type::TOP;
4363 }
4364 const Type* t = phase->type(in(MemNode::Memory));
4365 if (t == Type::TOP) {
4366 return Type::TOP;
4367 }
4368 t = phase->type(in(MemNode::Address));
4369 if (t == Type::TOP) {
4370 return Type::TOP;
4371 }
4372 t = phase->type(in(MemNode::ValueIn));
4373 if (t == Type::TOP) {
4374 return Type::TOP;
4375 }
4376 return bottom_type();
4377 }
4378
4379 const TypePtr* LoadStoreNode::adr_type() const {
4380 const TypePtr* cross_check = DEBUG_ONLY(_adr_type) NOT_DEBUG(nullptr);
4381 return MemNode::calculate_adr_type(in(MemNode::Address)->bottom_type(), cross_check);
4382 }
4383
4384 uint LoadStoreNode::ideal_reg() const {
4385 return _type->ideal_reg();
4386 }
4387
4388 // This method conservatively checks if the result of a LoadStoreNode is
4389 // used, that is, if it returns true, then it is definitely the case that
4390 // the result of the node is not needed.
4391 // For example, GetAndAdd can be matched into a lock_add instead of a
4392 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
4393 bool LoadStoreNode::result_not_used() const {
4394 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4395 Node *x = fast_out(i);
4396 if (x->Opcode() == Op_SCMemProj || x->is_ReachabilityFence()) {
4397 continue;
4398 }
4399 if (x->bottom_type() == TypeTuple::MEMBAR &&
4400 !x->is_Call() &&
4401 x->Opcode() != Op_Blackhole) {
4402 continue;
4403 }
4404 return false;
4405 }
4406 return true;
4407 }
4408
4409 MemBarNode* LoadStoreNode::trailing_membar() const {
4410 MemBarNode* trailing = nullptr;
4411 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4412 Node* u = fast_out(i);
4413 if (u->is_MemBar()) {
4414 if (u->as_MemBar()->trailing_load_store()) {
4415 assert(u->Opcode() == Op_MemBarAcquire, "");
4416 assert(trailing == nullptr, "only one");
4417 trailing = u->as_MemBar();
4418 #ifdef ASSERT
4419 Node* leading = trailing->leading_membar();
4420 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4421 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
4422 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
4423 #endif
4424 } else {
4425 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
4426 }
4427 }
4428 }
4429
4430 return trailing;
4431 }
4432
4433 uint LoadStoreNode::size_of() const { return sizeof(*this); }
4434
4435 #ifndef PRODUCT
4436 void LoadStoreNode::dump_spec(outputStream* st) const {
4437 if (in(MemNode::Address) == nullptr) {
4438 // node is dead
4439 return;
4440 }
4441 #ifndef ASSERT
4442 // fake the missing field
4443 const TypePtr* _adr_type = in(MemNode::Address)->bottom_type()->isa_ptr();
4444 #endif
4445 MemNode::dump_adr_type(_adr_type, st);
4446
4447 Compile* C = Compile::current();
4448 if (C->alias_type(_adr_type)->is_volatile()) {
4449 st->print(" Volatile!");
4450 }
4451 st->print(" barrier(0x%x)", _barrier_data);
4452 }
4453 #endif
4454
4455 //=============================================================================
4456 //----------------------------------LoadStoreConditionalNode--------------------
4457 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
4458 init_req(ExpectedIn, ex );
4459 }
4460
4461 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
4462 // Either input is TOP ==> the result is TOP
4463 const Type* t = phase->type(in(ExpectedIn));
4464 if (t == Type::TOP) {
4465 return Type::TOP;
4466 }
4467 return LoadStoreNode::Value(phase);
4468 }
4469
4470 //=============================================================================
4471 //-------------------------------adr_type--------------------------------------
4472 const TypePtr* ClearArrayNode::adr_type() const {
4473 Node *adr = in(3);
4474 if (adr == nullptr) return nullptr; // node is dead
4475 return MemNode::calculate_adr_type(adr->bottom_type());
4476 }
4477
4478 //------------------------------match_edge-------------------------------------
4479 // Do we Match on this edge index or not? Do not match memory
4480 uint ClearArrayNode::match_edge(uint idx) const {
4481 return idx > 1;
4482 }
4483
4484 //------------------------------Identity---------------------------------------
4485 // Clearing a zero length array does nothing
4486 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
4487 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
4488 }
4489
4490 //------------------------------Idealize---------------------------------------
4491 // Clearing a short array is faster with stores
4492 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4493 // Already know this is a large node, do not try to ideal it
4494 if (_is_large) return nullptr;
4495
4496 const int unit = BytesPerLong;
4497 const TypeX* t = phase->type(in(2))->isa_intptr_t();
4498 if (!t) return nullptr;
4499 if (!t->is_con()) return nullptr;
4500 intptr_t raw_count = t->get_con();
4501 intptr_t size = raw_count;
4502 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
4503 // Clearing nothing uses the Identity call.
4504 // Negative clears are possible on dead ClearArrays
4505 // (see jck test stmt114.stmt11402.val).
4506 if (size <= 0 || size % unit != 0) return nullptr;
4507 intptr_t count = size / unit;
4508 // Length too long; communicate this to matchers and assemblers.
4509 // Assemblers are responsible to produce fast hardware clears for it.
4510 if (size > InitArrayShortSize) {
4511 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
4512 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
4513 return nullptr;
4514 }
4515 if (!IdealizeClearArrayNode) return nullptr;
4516 Node *mem = in(1);
4517 if( phase->type(mem)==Type::TOP ) return nullptr;
4518 Node *adr = in(3);
4519 const Type* at = phase->type(adr);
4520 if( at==Type::TOP ) return nullptr;
4521 const TypePtr* atp = at->isa_ptr();
4522 // adjust atp to be the correct array element address type
4523 if (atp == nullptr) atp = TypePtr::BOTTOM;
4524 else atp = atp->add_offset(Type::OffsetBot);
4525 // Get base for derived pointer purposes
4526 if( adr->Opcode() != Op_AddP ) Unimplemented();
4527 Node *base = adr->in(1);
4528
4529 Node *val = in(4);
4530 Node *off = phase->MakeConX(BytesPerLong);
4531 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4532 count--;
4533 while (count--) {
4534 mem = phase->transform(mem);
4535 adr = phase->transform(AddPNode::make_with_base(base,adr,off));
4536 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4537 }
4538 return mem;
4539 }
4540
4541 //----------------------------step_through----------------------------------
4542 // Return allocation input memory edge if it is different instance
4543 // or itself if it is the one we are looking for.
4544 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4545 Node* n = *np;
4546 assert(n->is_ClearArray(), "sanity");
4547 intptr_t offset;
4548 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4549 // This method is called only before Allocate nodes are expanded
4550 // during macro nodes expansion. Before that ClearArray nodes are
4551 // only generated in PhaseMacroExpand::generate_arraycopy() (before
4552 // Allocate nodes are expanded) which follows allocations.
4553 assert(alloc != nullptr, "should have allocation");
4554 if (alloc->_idx == instance_id) {
4555 // Can not bypass initialization of the instance we are looking for.
4556 return false;
4557 }
4558 // Otherwise skip it.
4559 InitializeNode* init = alloc->initialization();
4560 if (init != nullptr)
4561 *np = init->in(TypeFunc::Memory);
4562 else
4563 *np = alloc->in(TypeFunc::Memory);
4564 return true;
4565 }
4566
4567 Node* ClearArrayNode::make_address(Node* dest, Node* offset, bool raw_base, PhaseGVN* phase) {
4568 Node* base = dest;
4569 if (raw_base) {
4570 // May be called as part of the initialization of a just allocated object
4571 base = phase->C->top();
4572 }
4573 return phase->transform(AddPNode::make_with_base(base, dest, offset));
4574 }
4575
4576 //----------------------------clear_memory-------------------------------------
4577 // Generate code to initialize object storage to zero.
4578 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4579 Node* val,
4580 Node* raw_val,
4581 intptr_t start_offset,
4582 Node* end_offset,
4583 bool raw_base,
4584 PhaseGVN* phase) {
4585 intptr_t offset = start_offset;
4586
4587 int unit = BytesPerLong;
4588 if ((offset % unit) != 0) {
4589 Node* adr = make_address(dest, phase->MakeConX(offset), raw_base, phase);
4590 const TypePtr* atp = TypeRawPtr::BOTTOM;
4591 if (val != nullptr) {
4592 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4593 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4594 } else {
4595 assert(raw_val == nullptr, "val may not be null");
4596 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4597 }
4598 mem = phase->transform(mem);
4599 offset += BytesPerInt;
4600 }
4601 assert((offset % unit) == 0, "");
4602
4603 // Initialize the remaining stuff, if any, with a ClearArray.
4604 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, raw_base, phase);
4605 }
4606
4607 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4608 Node* raw_val,
4609 Node* start_offset,
4610 Node* end_offset,
4611 bool raw_base,
4612 PhaseGVN* phase) {
4613 if (start_offset == end_offset) {
4614 // nothing to do
4615 return mem;
4616 }
4617
4618 int unit = BytesPerLong;
4619 Node* zbase = start_offset;
4620 Node* zend = end_offset;
4621
4622 // Scale to the unit required by the CPU:
4623 if (!Matcher::init_array_count_is_in_bytes) {
4624 Node* shift = phase->intcon(exact_log2(unit));
4625 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4626 zend = phase->transform(new URShiftXNode(zend, shift) );
4627 }
4628
4629 // Bulk clear double-words
4630 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4631 Node* adr = make_address(dest, start_offset, raw_base, phase);
4632 if (raw_val == nullptr) {
4633 raw_val = phase->MakeConX(0);
4634 }
4635 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
4636 return phase->transform(mem);
4637 }
4638
4639 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4640 Node* val,
4641 Node* raw_val,
4642 intptr_t start_offset,
4643 intptr_t end_offset,
4644 bool raw_base,
4645 PhaseGVN* phase) {
4646 if (start_offset == end_offset) {
4647 // nothing to do
4648 return mem;
4649 }
4650
4651 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4652 intptr_t done_offset = end_offset;
4653 if ((done_offset % BytesPerLong) != 0) {
4654 done_offset -= BytesPerInt;
4655 }
4656 if (done_offset > start_offset) {
4657 mem = clear_memory(ctl, mem, dest, val, raw_val,
4658 start_offset, phase->MakeConX(done_offset), raw_base, phase);
4659 }
4660 if (done_offset < end_offset) { // emit the final 32-bit store
4661 Node* adr = make_address(dest, phase->MakeConX(done_offset), raw_base, phase);
4662 const TypePtr* atp = TypeRawPtr::BOTTOM;
4663 if (val != nullptr) {
4664 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4665 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4666 } else {
4667 assert(raw_val == nullptr, "val may not be null");
4668 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4669 }
4670 mem = phase->transform(mem);
4671 done_offset += BytesPerInt;
4672 }
4673 assert(done_offset == end_offset, "");
4674 return mem;
4675 }
4676
4677 //=============================================================================
4678 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4679 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4680 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4681 #ifdef ASSERT
4682 , _pair_idx(0)
4683 #endif
4684 {
4685 init_class_id(Class_MemBar);
4686 Node* top = C->top();
4687 init_req(TypeFunc::I_O,top);
4688 init_req(TypeFunc::FramePtr,top);
4689 init_req(TypeFunc::ReturnAdr,top);
4690 if (precedent != nullptr)
4691 init_req(TypeFunc::Parms, precedent);
4692 }
4693
4694 //------------------------------cmp--------------------------------------------
4695 uint MemBarNode::hash() const { return NO_HASH; }
4696 bool MemBarNode::cmp( const Node &n ) const {
4697 return (&n == this); // Always fail except on self
4698 }
4699
4700 //------------------------------make-------------------------------------------
4701 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4702 switch (opcode) {
4703 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4704 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4705 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4706 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4707 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4708 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4709 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4710 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4711 case Op_MemBarStoreLoad: return new MemBarStoreLoadNode(C, atp, pn);
4712 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4713 case Op_MemBarFull: return new MemBarFullNode(C, atp, pn);
4714 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4715 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4716 case Op_Initialize: return new InitializeNode(C, atp, pn);
4717 default: ShouldNotReachHere(); return nullptr;
4718 }
4719 }
4720
4721 void MemBarNode::remove(PhaseIterGVN *igvn) {
4722 assert(outcnt() > 0 && (outcnt() <= 2 || Opcode() == Op_Initialize), "Only one or two out edges allowed");
4723 if (trailing_store() || trailing_load_store()) {
4724 MemBarNode* leading = leading_membar();
4725 if (leading != nullptr) {
4726 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4727 leading->remove(igvn);
4728 }
4729 }
4730 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4731 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4732 }
4733 if (is_Initialize()) {
4734 as_Initialize()->replace_mem_projs_by(in(TypeFunc::Memory), igvn);
4735 } else {
4736 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4737 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4738 }
4739 }
4740 }
4741
4742 //------------------------------Ideal------------------------------------------
4743 // Return a node which is more "ideal" than the current node. Strip out
4744 // control copies
4745 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4746 if (remove_dead_region(phase, can_reshape)) return this;
4747 // Don't bother trying to transform a dead node
4748 if (in(0) && in(0)->is_top()) {
4749 return nullptr;
4750 }
4751
4752 bool progress = false;
4753 // Eliminate volatile MemBars for scalar replaced objects.
4754 if (can_reshape && req() == (Precedent+1)) {
4755 bool eliminate = false;
4756 int opc = Opcode();
4757 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4758 // Volatile field loads and stores.
4759 Node* my_mem = in(MemBarNode::Precedent);
4760 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4761 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4762 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4763 // replace this Precedent (decodeN) with the Load instead.
4764 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4765 Node* load_node = my_mem->in(1);
4766 set_req(MemBarNode::Precedent, load_node);
4767 phase->is_IterGVN()->_worklist.push(my_mem);
4768 my_mem = load_node;
4769 } else {
4770 assert(my_mem->unique_out() == this, "sanity");
4771 assert(!trailing_load_store(), "load store node can't be eliminated");
4772 del_req(Precedent);
4773 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4774 my_mem = nullptr;
4775 }
4776 progress = true;
4777 }
4778 if (my_mem != nullptr && my_mem->is_Mem()) {
4779 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4780 // Check for scalar replaced object reference.
4781 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4782 t_oop->offset() != Type::OffsetBot &&
4783 t_oop->offset() != Type::OffsetTop) {
4784 eliminate = true;
4785 }
4786 }
4787 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4788 // Final field stores.
4789 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4790 if ((alloc != nullptr) && alloc->is_Allocate() &&
4791 alloc->as_Allocate()->does_not_escape_thread()) {
4792 // The allocated object does not escape.
4793 eliminate = true;
4794 }
4795 }
4796 if (eliminate) {
4797 // Replace MemBar projections by its inputs.
4798 PhaseIterGVN* igvn = phase->is_IterGVN();
4799 remove(igvn);
4800 // Must return either the original node (now dead) or a new node
4801 // (Do not return a top here, since that would break the uniqueness of top.)
4802 return new ConINode(TypeInt::ZERO);
4803 }
4804 }
4805 return progress ? this : nullptr;
4806 }
4807
4808 //------------------------------Value------------------------------------------
4809 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4810 if( !in(0) ) return Type::TOP;
4811 if( phase->type(in(0)) == Type::TOP )
4812 return Type::TOP;
4813 return TypeTuple::MEMBAR;
4814 }
4815
4816 //------------------------------match------------------------------------------
4817 // Construct projections for memory.
4818 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
4819 switch (proj->_con) {
4820 case TypeFunc::Control:
4821 case TypeFunc::Memory:
4822 return new MachProjNode(this, proj->_con, RegMask::EMPTY, MachProjNode::unmatched_proj);
4823 }
4824 ShouldNotReachHere();
4825 return nullptr;
4826 }
4827
4828 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4829 trailing->_kind = TrailingStore;
4830 leading->_kind = LeadingStore;
4831 #ifdef ASSERT
4832 trailing->_pair_idx = leading->_idx;
4833 leading->_pair_idx = leading->_idx;
4834 #endif
4835 }
4836
4837 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4838 trailing->_kind = TrailingLoadStore;
4839 leading->_kind = LeadingLoadStore;
4840 #ifdef ASSERT
4841 trailing->_pair_idx = leading->_idx;
4842 leading->_pair_idx = leading->_idx;
4843 #endif
4844 }
4845
4846 MemBarNode* MemBarNode::trailing_membar() const {
4847 ResourceMark rm;
4848 Node* trailing = (Node*)this;
4849 VectorSet seen;
4850 Node_Stack multis(0);
4851 do {
4852 Node* c = trailing;
4853 uint i = 0;
4854 do {
4855 trailing = nullptr;
4856 for (; i < c->outcnt(); i++) {
4857 Node* next = c->raw_out(i);
4858 if (next != c && next->is_CFG()) {
4859 if (c->is_MultiBranch()) {
4860 if (multis.node() == c) {
4861 multis.set_index(i+1);
4862 } else {
4863 multis.push(c, i+1);
4864 }
4865 }
4866 trailing = next;
4867 break;
4868 }
4869 }
4870 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4871 break;
4872 }
4873 while (multis.size() > 0) {
4874 c = multis.node();
4875 i = multis.index();
4876 if (i < c->req()) {
4877 break;
4878 }
4879 multis.pop();
4880 }
4881 } while (multis.size() > 0);
4882 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4883
4884 MemBarNode* mb = trailing->as_MemBar();
4885 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4886 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4887 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4888 return mb;
4889 }
4890
4891 MemBarNode* MemBarNode::leading_membar() const {
4892 ResourceMark rm;
4893 VectorSet seen;
4894 Node_Stack regions(0);
4895 Node* leading = in(0);
4896 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4897 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4898 leading = nullptr;
4899 while (regions.size() > 0 && leading == nullptr) {
4900 Node* r = regions.node();
4901 uint i = regions.index();
4902 if (i < r->req()) {
4903 leading = r->in(i);
4904 regions.set_index(i+1);
4905 } else {
4906 regions.pop();
4907 }
4908 }
4909 if (leading == nullptr) {
4910 assert(regions.size() == 0, "all paths should have been tried");
4911 return nullptr;
4912 }
4913 }
4914 if (leading->is_Region()) {
4915 regions.push(leading, 2);
4916 leading = leading->in(1);
4917 } else {
4918 leading = leading->in(0);
4919 }
4920 }
4921 #ifdef ASSERT
4922 Unique_Node_List wq;
4923 wq.push((Node*)this);
4924 uint found = 0;
4925 for (uint i = 0; i < wq.size(); i++) {
4926 Node* n = wq.at(i);
4927 if (n->is_Region()) {
4928 for (uint j = 1; j < n->req(); j++) {
4929 Node* in = n->in(j);
4930 if (in != nullptr && !in->is_top()) {
4931 wq.push(in);
4932 }
4933 }
4934 } else {
4935 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4936 assert(n == leading, "consistency check failed");
4937 found++;
4938 } else {
4939 Node* in = n->in(0);
4940 if (in != nullptr && !in->is_top()) {
4941 wq.push(in);
4942 }
4943 }
4944 }
4945 }
4946 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4947 #endif
4948 if (leading == nullptr) {
4949 return nullptr;
4950 }
4951 MemBarNode* mb = leading->as_MemBar();
4952 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4953 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4954 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4955 return mb;
4956 }
4957
4958
4959 //===========================InitializeNode====================================
4960 // SUMMARY:
4961 // This node acts as a memory barrier on raw memory, after some raw stores.
4962 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4963 // The Initialize can 'capture' suitably constrained stores as raw inits.
4964 // It can coalesce related raw stores into larger units (called 'tiles').
4965 // It can avoid zeroing new storage for memory units which have raw inits.
4966 // At macro-expansion, it is marked 'complete', and does not optimize further.
4967 //
4968 // EXAMPLE:
4969 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4970 // ctl = incoming control; mem* = incoming memory
4971 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4972 // First allocate uninitialized memory and fill in the header:
4973 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4974 // ctl := alloc.Control; mem* := alloc.Memory*
4975 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4976 // Then initialize to zero the non-header parts of the raw memory block:
4977 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4978 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4979 // After the initialize node executes, the object is ready for service:
4980 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4981 // Suppose its body is immediately initialized as {1,2}:
4982 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4983 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4984 // mem.SLICE(#short[*]) := store2
4985 //
4986 // DETAILS:
4987 // An InitializeNode collects and isolates object initialization after
4988 // an AllocateNode and before the next possible safepoint. As a
4989 // memory barrier (MemBarNode), it keeps critical stores from drifting
4990 // down past any safepoint or any publication of the allocation.
4991 // Before this barrier, a newly-allocated object may have uninitialized bits.
4992 // After this barrier, it may be treated as a real oop, and GC is allowed.
4993 //
4994 // The semantics of the InitializeNode include an implicit zeroing of
4995 // the new object from object header to the end of the object.
4996 // (The object header and end are determined by the AllocateNode.)
4997 //
4998 // Certain stores may be added as direct inputs to the InitializeNode.
4999 // These stores must update raw memory, and they must be to addresses
5000 // derived from the raw address produced by AllocateNode, and with
5001 // a constant offset. They must be ordered by increasing offset.
5002 // The first one is at in(RawStores), the last at in(req()-1).
5003 // Unlike most memory operations, they are not linked in a chain,
5004 // but are displayed in parallel as users of the rawmem output of
5005 // the allocation.
5006 //
5007 // (See comments in InitializeNode::capture_store, which continue
5008 // the example given above.)
5009 //
5010 // When the associated Allocate is macro-expanded, the InitializeNode
5011 // may be rewritten to optimize collected stores. A ClearArrayNode
5012 // may also be created at that point to represent any required zeroing.
5013 // The InitializeNode is then marked 'complete', prohibiting further
5014 // capturing of nearby memory operations.
5015 //
5016 // During macro-expansion, all captured initializations which store
5017 // constant values of 32 bits or smaller are coalesced (if advantageous)
5018 // into larger 'tiles' 32 or 64 bits. This allows an object to be
5019 // initialized in fewer memory operations. Memory words which are
5020 // covered by neither tiles nor non-constant stores are pre-zeroed
5021 // by explicit stores of zero. (The code shape happens to do all
5022 // zeroing first, then all other stores, with both sequences occurring
5023 // in order of ascending offsets.)
5024 //
5025 // Alternatively, code may be inserted between an AllocateNode and its
5026 // InitializeNode, to perform arbitrary initialization of the new object.
5027 // E.g., the object copying intrinsics insert complex data transfers here.
5028 // The initialization must then be marked as 'complete' disable the
5029 // built-in zeroing semantics and the collection of initializing stores.
5030 //
5031 // While an InitializeNode is incomplete, reads from the memory state
5032 // produced by it are optimizable if they match the control edge and
5033 // new oop address associated with the allocation/initialization.
5034 // They return a stored value (if the offset matches) or else zero.
5035 // A write to the memory state, if it matches control and address,
5036 // and if it is to a constant offset, may be 'captured' by the
5037 // InitializeNode. It is cloned as a raw memory operation and rewired
5038 // inside the initialization, to the raw oop produced by the allocation.
5039 // Operations on addresses which are provably distinct (e.g., to
5040 // other AllocateNodes) are allowed to bypass the initialization.
5041 //
5042 // The effect of all this is to consolidate object initialization
5043 // (both arrays and non-arrays, both piecewise and bulk) into a
5044 // single location, where it can be optimized as a unit.
5045 //
5046 // Only stores with an offset less than TrackedInitializationLimit words
5047 // will be considered for capture by an InitializeNode. This puts a
5048 // reasonable limit on the complexity of optimized initializations.
5049
5050 //---------------------------InitializeNode------------------------------------
5051 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
5052 : MemBarNode(C, adr_type, rawoop),
5053 _is_complete(Incomplete), _does_not_escape(false)
5054 {
5055 init_class_id(Class_Initialize);
5056
5057 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
5058 assert(in(RawAddress) == rawoop, "proper init");
5059 // Note: allocation() can be null, for secondary initialization barriers
5060 }
5061
5062 // Since this node is not matched, it will be processed by the
5063 // register allocator. Declare that there are no constraints
5064 // on the allocation of the RawAddress edge.
5065 const RegMask &InitializeNode::in_RegMask(uint idx) const {
5066 // This edge should be set to top, by the set_complete. But be conservative.
5067 if (idx == InitializeNode::RawAddress)
5068 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
5069 return RegMask::EMPTY;
5070 }
5071
5072 Node* InitializeNode::memory(uint alias_idx) {
5073 Node* mem = in(Memory);
5074 if (mem->is_MergeMem()) {
5075 return mem->as_MergeMem()->memory_at(alias_idx);
5076 } else {
5077 // incoming raw memory is not split
5078 return mem;
5079 }
5080 }
5081
5082 bool InitializeNode::is_non_zero() {
5083 if (is_complete()) return false;
5084 remove_extra_zeroes();
5085 return (req() > RawStores);
5086 }
5087
5088 void InitializeNode::set_complete(PhaseGVN* phase) {
5089 assert(!is_complete(), "caller responsibility");
5090 _is_complete = Complete;
5091
5092 // After this node is complete, it contains a bunch of
5093 // raw-memory initializations. There is no need for
5094 // it to have anything to do with non-raw memory effects.
5095 // Therefore, tell all non-raw users to re-optimize themselves,
5096 // after skipping the memory effects of this initialization.
5097 PhaseIterGVN* igvn = phase->is_IterGVN();
5098 if (igvn) igvn->add_users_to_worklist(this);
5099 }
5100
5101 // convenience function
5102 // return false if the init contains any stores already
5103 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
5104 InitializeNode* init = initialization();
5105 if (init == nullptr || init->is_complete()) {
5106 return false;
5107 }
5108 init->remove_extra_zeroes();
5109 // for now, if this allocation has already collected any inits, bail:
5110 if (init->is_non_zero()) return false;
5111 init->set_complete(phase);
5112 return true;
5113 }
5114
5115 void InitializeNode::remove_extra_zeroes() {
5116 if (req() == RawStores) return;
5117 Node* zmem = zero_memory();
5118 uint fill = RawStores;
5119 for (uint i = fill; i < req(); i++) {
5120 Node* n = in(i);
5121 if (n->is_top() || n == zmem) continue; // skip
5122 if (fill < i) set_req(fill, n); // compact
5123 ++fill;
5124 }
5125 // delete any empty spaces created:
5126 while (fill < req()) {
5127 del_req(fill);
5128 }
5129 }
5130
5131 // Helper for remembering which stores go with which offsets.
5132 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
5133 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
5134 intptr_t offset = -1;
5135 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
5136 phase, offset);
5137 if (base == nullptr) return -1; // something is dead,
5138 if (offset < 0) return -1; // dead, dead
5139 return offset;
5140 }
5141
5142 // Helper for proving that an initialization expression is
5143 // "simple enough" to be folded into an object initialization.
5144 // Attempts to prove that a store's initial value 'n' can be captured
5145 // within the initialization without creating a vicious cycle, such as:
5146 // { Foo p = new Foo(); p.next = p; }
5147 // True for constants and parameters and small combinations thereof.
5148 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
5149 ResourceMark rm;
5150 Unique_Node_List worklist;
5151 worklist.push(value);
5152
5153 uint complexity_limit = 20;
5154 for (uint j = 0; j < worklist.size(); j++) {
5155 if (j >= complexity_limit) {
5156 return false; // Bail out if processed too many nodes
5157 }
5158
5159 Node* n = worklist.at(j);
5160 if (n == nullptr) continue; // (can this really happen?)
5161 if (n->is_Proj()) n = n->in(0);
5162 if (n == this) return false; // found a cycle
5163 if (n->is_Con()) continue;
5164 if (n->is_Start()) continue; // params, etc., are OK
5165 if (n->is_Root()) continue; // even better
5166
5167 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
5168 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
5169 continue;
5170 }
5171
5172 Node* ctl = n->in(0);
5173 if (ctl != nullptr && !ctl->is_top()) {
5174 if (ctl->is_Proj()) ctl = ctl->in(0);
5175 if (ctl == this) return false;
5176
5177 // If we already know that the enclosing memory op is pinned right after
5178 // the init, then any control flow that the store has picked up
5179 // must have preceded the init, or else be equal to the init.
5180 // Even after loop optimizations (which might change control edges)
5181 // a store is never pinned *before* the availability of its inputs.
5182 if (!MemNode::all_controls_dominate(n, this, phase)) {
5183 return false; // failed to prove a good control
5184 }
5185 }
5186
5187 // Check data edges for possible dependencies on 'this'.
5188 for (uint i = 1; i < n->req(); i++) {
5189 Node* m = n->in(i);
5190 if (m == nullptr || m == n || m->is_top()) continue;
5191
5192 // Only process data inputs once
5193 worklist.push(m);
5194 }
5195 }
5196
5197 return true;
5198 }
5199
5200 // Here are all the checks a Store must pass before it can be moved into
5201 // an initialization. Returns zero if a check fails.
5202 // On success, returns the (constant) offset to which the store applies,
5203 // within the initialized memory.
5204 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
5205 const int FAIL = 0;
5206 if (st->req() != MemNode::ValueIn + 1)
5207 return FAIL; // an inscrutable StoreNode (card mark?)
5208 Node* ctl = st->in(MemNode::Control);
5209 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
5210 return FAIL; // must be unconditional after the initialization
5211 Node* mem = st->in(MemNode::Memory);
5212 if (!(mem->is_Proj() && mem->in(0) == this))
5213 return FAIL; // must not be preceded by other stores
5214 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5215 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
5216 !bs->can_initialize_object(st)) {
5217 return FAIL;
5218 }
5219 Node* adr = st->in(MemNode::Address);
5220 intptr_t offset;
5221 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
5222 if (alloc == nullptr)
5223 return FAIL; // inscrutable address
5224 if (alloc != allocation())
5225 return FAIL; // wrong allocation! (store needs to float up)
5226 int size_in_bytes = st->memory_size();
5227 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
5228 return FAIL; // mismatched access
5229 }
5230 Node* val = st->in(MemNode::ValueIn);
5231
5232 if (!detect_init_independence(val, phase))
5233 return FAIL; // stored value must be 'simple enough'
5234
5235 // The Store can be captured only if nothing after the allocation
5236 // and before the Store is using the memory location that the store
5237 // overwrites.
5238 bool failed = false;
5239 // If is_complete_with_arraycopy() is true the shape of the graph is
5240 // well defined and is safe so no need for extra checks.
5241 if (!is_complete_with_arraycopy()) {
5242 // We are going to look at each use of the memory state following
5243 // the allocation to make sure nothing reads the memory that the
5244 // Store writes.
5245 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
5246 int alias_idx = phase->C->get_alias_index(t_adr);
5247 ResourceMark rm;
5248 Unique_Node_List mems;
5249 mems.push(mem);
5250 Node* unique_merge = nullptr;
5251 for (uint next = 0; next < mems.size(); ++next) {
5252 Node *m = mems.at(next);
5253 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5254 Node *n = m->fast_out(j);
5255 if (n->outcnt() == 0) {
5256 continue;
5257 }
5258 if (n == st) {
5259 continue;
5260 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
5261 // If the control of this use is different from the control
5262 // of the Store which is right after the InitializeNode then
5263 // this node cannot be between the InitializeNode and the
5264 // Store.
5265 continue;
5266 } else if (n->is_MergeMem()) {
5267 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
5268 // We can hit a MergeMemNode (that will likely go away
5269 // later) that is a direct use of the memory state
5270 // following the InitializeNode on the same slice as the
5271 // store node that we'd like to capture. We need to check
5272 // the uses of the MergeMemNode.
5273 mems.push(n);
5274 }
5275 } else if (n->is_Mem()) {
5276 Node* other_adr = n->in(MemNode::Address);
5277 if (other_adr == adr) {
5278 failed = true;
5279 break;
5280 } else {
5281 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
5282 if (other_t_adr != nullptr) {
5283 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
5284 if (other_alias_idx == alias_idx) {
5285 // A load from the same memory slice as the store right
5286 // after the InitializeNode. We check the control of the
5287 // object/array that is loaded from. If it's the same as
5288 // the store control then we cannot capture the store.
5289 assert(!n->is_Store(), "2 stores to same slice on same control?");
5290 Node* base = other_adr;
5291 if (base->is_Phi()) {
5292 // In rare case, base may be a PhiNode and it may read
5293 // the same memory slice between InitializeNode and store.
5294 failed = true;
5295 break;
5296 }
5297 assert(base->is_AddP(), "should be addp but is %s", base->Name());
5298 base = base->in(AddPNode::Base);
5299 if (base != nullptr) {
5300 base = base->uncast();
5301 if (base->is_Proj() && base->in(0) == alloc) {
5302 failed = true;
5303 break;
5304 }
5305 }
5306 }
5307 }
5308 }
5309 } else {
5310 failed = true;
5311 break;
5312 }
5313 }
5314 }
5315 }
5316 if (failed) {
5317 if (!can_reshape) {
5318 // We decided we couldn't capture the store during parsing. We
5319 // should try again during the next IGVN once the graph is
5320 // cleaner.
5321 phase->C->record_for_igvn(st);
5322 }
5323 return FAIL;
5324 }
5325
5326 return offset; // success
5327 }
5328
5329 // Find the captured store in(i) which corresponds to the range
5330 // [start..start+size) in the initialized object.
5331 // If there is one, return its index i. If there isn't, return the
5332 // negative of the index where it should be inserted.
5333 // Return 0 if the queried range overlaps an initialization boundary
5334 // or if dead code is encountered.
5335 // If size_in_bytes is zero, do not bother with overlap checks.
5336 int InitializeNode::captured_store_insertion_point(intptr_t start,
5337 int size_in_bytes,
5338 PhaseValues* phase) {
5339 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
5340
5341 if (is_complete())
5342 return FAIL; // arraycopy got here first; punt
5343
5344 assert(allocation() != nullptr, "must be present");
5345
5346 // no negatives, no header fields:
5347 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
5348
5349 // after a certain size, we bail out on tracking all the stores:
5350 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5351 if (start >= ti_limit) return FAIL;
5352
5353 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
5354 if (i >= limit) return -(int)i; // not found; here is where to put it
5355
5356 Node* st = in(i);
5357 intptr_t st_off = get_store_offset(st, phase);
5358 if (st_off < 0) {
5359 if (st != zero_memory()) {
5360 return FAIL; // bail out if there is dead garbage
5361 }
5362 } else if (st_off > start) {
5363 // ...we are done, since stores are ordered
5364 if (st_off < start + size_in_bytes) {
5365 return FAIL; // the next store overlaps
5366 }
5367 return -(int)i; // not found; here is where to put it
5368 } else if (st_off < start) {
5369 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
5370 if (size_in_bytes != 0 &&
5371 start < st_off + MAX_STORE &&
5372 start < st_off + st->as_Store()->memory_size()) {
5373 return FAIL; // the previous store overlaps
5374 }
5375 } else {
5376 if (size_in_bytes != 0 &&
5377 st->as_Store()->memory_size() != size_in_bytes) {
5378 return FAIL; // mismatched store size
5379 }
5380 return i;
5381 }
5382
5383 ++i;
5384 }
5385 }
5386
5387 // Look for a captured store which initializes at the offset 'start'
5388 // with the given size. If there is no such store, and no other
5389 // initialization interferes, then return zero_memory (the memory
5390 // projection of the AllocateNode).
5391 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
5392 PhaseValues* phase) {
5393 assert(stores_are_sane(phase), "");
5394 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5395 if (i == 0) {
5396 return nullptr; // something is dead
5397 } else if (i < 0) {
5398 return zero_memory(); // just primordial zero bits here
5399 } else {
5400 Node* st = in(i); // here is the store at this position
5401 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
5402 return st;
5403 }
5404 }
5405
5406 // Create, as a raw pointer, an address within my new object at 'offset'.
5407 Node* InitializeNode::make_raw_address(intptr_t offset,
5408 PhaseGVN* phase) {
5409 Node* addr = in(RawAddress);
5410 if (offset != 0) {
5411 Compile* C = phase->C;
5412 addr = phase->transform(AddPNode::make_off_heap(addr, phase->MakeConX(offset)));
5413 }
5414 return addr;
5415 }
5416
5417 // Clone the given store, converting it into a raw store
5418 // initializing a field or element of my new object.
5419 // Caller is responsible for retiring the original store,
5420 // with subsume_node or the like.
5421 //
5422 // From the example above InitializeNode::InitializeNode,
5423 // here are the old stores to be captured:
5424 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
5425 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
5426 //
5427 // Here is the changed code; note the extra edges on init:
5428 // alloc = (Allocate ...)
5429 // rawoop = alloc.RawAddress
5430 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
5431 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
5432 // init = (Initialize alloc.Control alloc.Memory rawoop
5433 // rawstore1 rawstore2)
5434 //
5435 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
5436 PhaseGVN* phase, bool can_reshape) {
5437 assert(stores_are_sane(phase), "");
5438
5439 if (start < 0) return nullptr;
5440 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
5441
5442 Compile* C = phase->C;
5443 int size_in_bytes = st->memory_size();
5444 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5445 if (i == 0) return nullptr; // bail out
5446 Node* prev_mem = nullptr; // raw memory for the captured store
5447 if (i > 0) {
5448 prev_mem = in(i); // there is a pre-existing store under this one
5449 set_req(i, C->top()); // temporarily disconnect it
5450 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
5451 } else {
5452 i = -i; // no pre-existing store
5453 prev_mem = zero_memory(); // a slice of the newly allocated object
5454 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
5455 set_req(--i, C->top()); // reuse this edge; it has been folded away
5456 else
5457 ins_req(i, C->top()); // build a new edge
5458 }
5459 Node* new_st = st->clone_with_adr_type(TypeRawPtr::BOTTOM);
5460 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5461 new_st->set_req(MemNode::Control, in(Control));
5462 new_st->set_req(MemNode::Memory, prev_mem);
5463 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
5464 bs->eliminate_gc_barrier_data(new_st);
5465 new_st = phase->transform(new_st);
5466
5467 // At this point, new_st might have swallowed a pre-existing store
5468 // at the same offset, or perhaps new_st might have disappeared,
5469 // if it redundantly stored the same value (or zero to fresh memory).
5470
5471 // In any case, wire it in:
5472 PhaseIterGVN* igvn = phase->is_IterGVN();
5473 if (igvn) {
5474 igvn->rehash_node_delayed(this);
5475 }
5476 set_req(i, new_st);
5477
5478 // The caller may now kill the old guy.
5479 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
5480 assert(check_st == new_st || check_st == nullptr, "must be findable");
5481 assert(!is_complete(), "");
5482 return new_st;
5483 }
5484
5485 static bool store_constant(jlong* tiles, int num_tiles,
5486 intptr_t st_off, int st_size,
5487 jlong con) {
5488 if ((st_off & (st_size-1)) != 0)
5489 return false; // strange store offset (assume size==2**N)
5490 address addr = (address)tiles + st_off;
5491 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
5492 switch (st_size) {
5493 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
5494 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
5495 case sizeof(jint): *(jint*) addr = (jint) con; break;
5496 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
5497 default: return false; // strange store size (detect size!=2**N here)
5498 }
5499 return true; // return success to caller
5500 }
5501
5502 // Coalesce subword constants into int constants and possibly
5503 // into long constants. The goal, if the CPU permits,
5504 // is to initialize the object with a small number of 64-bit tiles.
5505 // Also, convert floating-point constants to bit patterns.
5506 // Non-constants are not relevant to this pass.
5507 //
5508 // In terms of the running example on InitializeNode::InitializeNode
5509 // and InitializeNode::capture_store, here is the transformation
5510 // of rawstore1 and rawstore2 into rawstore12:
5511 // alloc = (Allocate ...)
5512 // rawoop = alloc.RawAddress
5513 // tile12 = 0x00010002
5514 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
5515 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
5516 //
5517 void
5518 InitializeNode::coalesce_subword_stores(intptr_t header_size,
5519 Node* size_in_bytes,
5520 PhaseGVN* phase) {
5521 Compile* C = phase->C;
5522
5523 assert(stores_are_sane(phase), "");
5524 // Note: After this pass, they are not completely sane,
5525 // since there may be some overlaps.
5526
5527 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
5528
5529 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5530 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
5531 size_limit = MIN2(size_limit, ti_limit);
5532 size_limit = align_up(size_limit, BytesPerLong);
5533 int num_tiles = size_limit / BytesPerLong;
5534
5535 // allocate space for the tile map:
5536 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5537 jlong tiles_buf[small_len];
5538 Node* nodes_buf[small_len];
5539 jlong inits_buf[small_len];
5540 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5541 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5542 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5543 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5544 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5545 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5546 // tiles: exact bitwise model of all primitive constants
5547 // nodes: last constant-storing node subsumed into the tiles model
5548 // inits: which bytes (in each tile) are touched by any initializations
5549
5550 //// Pass A: Fill in the tile model with any relevant stores.
5551
5552 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5553 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5554 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5555 Node* zmem = zero_memory(); // initially zero memory state
5556 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5557 Node* st = in(i);
5558 intptr_t st_off = get_store_offset(st, phase);
5559
5560 // Figure out the store's offset and constant value:
5561 if (st_off < header_size) continue; //skip (ignore header)
5562 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
5563 int st_size = st->as_Store()->memory_size();
5564 if (st_off + st_size > size_limit) break;
5565
5566 // Record which bytes are touched, whether by constant or not.
5567 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5568 continue; // skip (strange store size)
5569
5570 const Type* val = phase->type(st->in(MemNode::ValueIn));
5571 if (!val->singleton()) continue; //skip (non-con store)
5572 BasicType type = val->basic_type();
5573
5574 jlong con = 0;
5575 switch (type) {
5576 case T_INT: con = val->is_int()->get_con(); break;
5577 case T_LONG: con = val->is_long()->get_con(); break;
5578 case T_FLOAT: con = jint_cast(val->getf()); break;
5579 case T_DOUBLE: con = jlong_cast(val->getd()); break;
5580 default: continue; //skip (odd store type)
5581 }
5582
5583 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5584 st->Opcode() == Op_StoreL) {
5585 continue; // This StoreL is already optimal.
5586 }
5587
5588 // Store down the constant.
5589 store_constant(tiles, num_tiles, st_off, st_size, con);
5590
5591 intptr_t j = st_off >> LogBytesPerLong;
5592
5593 if (type == T_INT && st_size == BytesPerInt
5594 && (st_off & BytesPerInt) == BytesPerInt) {
5595 jlong lcon = tiles[j];
5596 if (!Matcher::isSimpleConstant64(lcon) &&
5597 st->Opcode() == Op_StoreI) {
5598 // This StoreI is already optimal by itself.
5599 jint* intcon = (jint*) &tiles[j];
5600 intcon[1] = 0; // undo the store_constant()
5601
5602 // If the previous store is also optimal by itself, back up and
5603 // undo the action of the previous loop iteration... if we can.
5604 // But if we can't, just let the previous half take care of itself.
5605 st = nodes[j];
5606 st_off -= BytesPerInt;
5607 con = intcon[0];
5608 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5609 assert(st_off >= header_size, "still ignoring header");
5610 assert(get_store_offset(st, phase) == st_off, "must be");
5611 assert(in(i-1) == zmem, "must be");
5612 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5613 assert(con == tcon->is_int()->get_con(), "must be");
5614 // Undo the effects of the previous loop trip, which swallowed st:
5615 intcon[0] = 0; // undo store_constant()
5616 set_req(i-1, st); // undo set_req(i, zmem)
5617 nodes[j] = nullptr; // undo nodes[j] = st
5618 --old_subword; // undo ++old_subword
5619 }
5620 continue; // This StoreI is already optimal.
5621 }
5622 }
5623
5624 // This store is not needed.
5625 set_req(i, zmem);
5626 nodes[j] = st; // record for the moment
5627 if (st_size < BytesPerLong) // something has changed
5628 ++old_subword; // includes int/float, but who's counting...
5629 else ++old_long;
5630 }
5631
5632 if ((old_subword + old_long) == 0)
5633 return; // nothing more to do
5634
5635 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5636 // Be sure to insert them before overlapping non-constant stores.
5637 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5638 for (int j = 0; j < num_tiles; j++) {
5639 jlong con = tiles[j];
5640 jlong init = inits[j];
5641 if (con == 0) continue;
5642 jint con0, con1; // split the constant, address-wise
5643 jint init0, init1; // split the init map, address-wise
5644 { union { jlong con; jint intcon[2]; } u;
5645 u.con = con;
5646 con0 = u.intcon[0];
5647 con1 = u.intcon[1];
5648 u.con = init;
5649 init0 = u.intcon[0];
5650 init1 = u.intcon[1];
5651 }
5652
5653 Node* old = nodes[j];
5654 assert(old != nullptr, "need the prior store");
5655 intptr_t offset = (j * BytesPerLong);
5656
5657 bool split = !Matcher::isSimpleConstant64(con);
5658
5659 if (offset < header_size) {
5660 assert(offset + BytesPerInt >= header_size, "second int counts");
5661 assert(*(jint*)&tiles[j] == 0, "junk in header");
5662 split = true; // only the second word counts
5663 // Example: int a[] = { 42 ... }
5664 } else if (con0 == 0 && init0 == -1) {
5665 split = true; // first word is covered by full inits
5666 // Example: int a[] = { ... foo(), 42 ... }
5667 } else if (con1 == 0 && init1 == -1) {
5668 split = true; // second word is covered by full inits
5669 // Example: int a[] = { ... 42, foo() ... }
5670 }
5671
5672 // Here's a case where init0 is neither 0 nor -1:
5673 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5674 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5675 // In this case the tile is not split; it is (jlong)42.
5676 // The big tile is stored down, and then the foo() value is inserted.
5677 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5678
5679 Node* ctl = old->in(MemNode::Control);
5680 Node* adr = make_raw_address(offset, phase);
5681 const TypePtr* atp = TypeRawPtr::BOTTOM;
5682
5683 // One or two coalesced stores to plop down.
5684 Node* st[2];
5685 intptr_t off[2];
5686 int nst = 0;
5687 if (!split) {
5688 ++new_long;
5689 off[nst] = offset;
5690 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5691 phase->longcon(con), T_LONG, MemNode::unordered);
5692 } else {
5693 // Omit either if it is a zero.
5694 if (con0 != 0) {
5695 ++new_int;
5696 off[nst] = offset;
5697 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5698 phase->intcon(con0), T_INT, MemNode::unordered);
5699 }
5700 if (con1 != 0) {
5701 ++new_int;
5702 offset += BytesPerInt;
5703 adr = make_raw_address(offset, phase);
5704 off[nst] = offset;
5705 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5706 phase->intcon(con1), T_INT, MemNode::unordered);
5707 }
5708 }
5709
5710 // Insert second store first, then the first before the second.
5711 // Insert each one just before any overlapping non-constant stores.
5712 while (nst > 0) {
5713 Node* st1 = st[--nst];
5714 C->copy_node_notes_to(st1, old);
5715 st1 = phase->transform(st1);
5716 offset = off[nst];
5717 assert(offset >= header_size, "do not smash header");
5718 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5719 guarantee(ins_idx != 0, "must re-insert constant store");
5720 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5721 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5722 set_req(--ins_idx, st1);
5723 else
5724 ins_req(ins_idx, st1);
5725 }
5726 }
5727
5728 if (PrintCompilation && WizardMode)
5729 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5730 old_subword, old_long, new_int, new_long);
5731 if (C->log() != nullptr)
5732 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5733 old_subword, old_long, new_int, new_long);
5734
5735 // Clean up any remaining occurrences of zmem:
5736 remove_extra_zeroes();
5737 }
5738
5739 // Explore forward from in(start) to find the first fully initialized
5740 // word, and return its offset. Skip groups of subword stores which
5741 // together initialize full words. If in(start) is itself part of a
5742 // fully initialized word, return the offset of in(start). If there
5743 // are no following full-word stores, or if something is fishy, return
5744 // a negative value.
5745 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5746 int int_map = 0;
5747 intptr_t int_map_off = 0;
5748 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5749
5750 for (uint i = start, limit = req(); i < limit; i++) {
5751 Node* st = in(i);
5752
5753 intptr_t st_off = get_store_offset(st, phase);
5754 if (st_off < 0) break; // return conservative answer
5755
5756 int st_size = st->as_Store()->memory_size();
5757 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5758 return st_off; // we found a complete word init
5759 }
5760
5761 // update the map:
5762
5763 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5764 if (this_int_off != int_map_off) {
5765 // reset the map:
5766 int_map = 0;
5767 int_map_off = this_int_off;
5768 }
5769
5770 int subword_off = st_off - this_int_off;
5771 int_map |= right_n_bits(st_size) << subword_off;
5772 if ((int_map & FULL_MAP) == FULL_MAP) {
5773 return this_int_off; // we found a complete word init
5774 }
5775
5776 // Did this store hit or cross the word boundary?
5777 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5778 if (next_int_off == this_int_off + BytesPerInt) {
5779 // We passed the current int, without fully initializing it.
5780 int_map_off = next_int_off;
5781 int_map >>= BytesPerInt;
5782 } else if (next_int_off > this_int_off + BytesPerInt) {
5783 // We passed the current and next int.
5784 return this_int_off + BytesPerInt;
5785 }
5786 }
5787
5788 return -1;
5789 }
5790
5791
5792 // Called when the associated AllocateNode is expanded into CFG.
5793 // At this point, we may perform additional optimizations.
5794 // Linearize the stores by ascending offset, to make memory
5795 // activity as coherent as possible.
5796 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5797 intptr_t header_size,
5798 Node* size_in_bytes,
5799 PhaseIterGVN* phase) {
5800 assert(!is_complete(), "not already complete");
5801 assert(stores_are_sane(phase), "");
5802 assert(allocation() != nullptr, "must be present");
5803
5804 remove_extra_zeroes();
5805
5806 if (ReduceFieldZeroing || ReduceBulkZeroing)
5807 // reduce instruction count for common initialization patterns
5808 coalesce_subword_stores(header_size, size_in_bytes, phase);
5809
5810 Node* zmem = zero_memory(); // initially zero memory state
5811 Node* inits = zmem; // accumulating a linearized chain of inits
5812 #ifdef ASSERT
5813 intptr_t first_offset = allocation()->minimum_header_size();
5814 intptr_t last_init_off = first_offset; // previous init offset
5815 intptr_t last_init_end = first_offset; // previous init offset+size
5816 intptr_t last_tile_end = first_offset; // previous tile offset+size
5817 #endif
5818 intptr_t zeroes_done = header_size;
5819
5820 bool do_zeroing = true; // we might give up if inits are very sparse
5821 int big_init_gaps = 0; // how many large gaps have we seen?
5822
5823 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5824 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5825
5826 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5827 Node* st = in(i);
5828 intptr_t st_off = get_store_offset(st, phase);
5829 if (st_off < 0)
5830 break; // unknown junk in the inits
5831 if (st->in(MemNode::Memory) != zmem)
5832 break; // complicated store chains somehow in list
5833
5834 int st_size = st->as_Store()->memory_size();
5835 intptr_t next_init_off = st_off + st_size;
5836
5837 if (do_zeroing && zeroes_done < next_init_off) {
5838 // See if this store needs a zero before it or under it.
5839 intptr_t zeroes_needed = st_off;
5840
5841 if (st_size < BytesPerInt) {
5842 // Look for subword stores which only partially initialize words.
5843 // If we find some, we must lay down some word-level zeroes first,
5844 // underneath the subword stores.
5845 //
5846 // Examples:
5847 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5848 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5849 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5850 //
5851 // Note: coalesce_subword_stores may have already done this,
5852 // if it was prompted by constant non-zero subword initializers.
5853 // But this case can still arise with non-constant stores.
5854
5855 intptr_t next_full_store = find_next_fullword_store(i, phase);
5856
5857 // In the examples above:
5858 // in(i) p q r s x y z
5859 // st_off 12 13 14 15 12 13 14
5860 // st_size 1 1 1 1 1 1 1
5861 // next_full_s. 12 16 16 16 16 16 16
5862 // z's_done 12 16 16 16 12 16 12
5863 // z's_needed 12 16 16 16 16 16 16
5864 // zsize 0 0 0 0 4 0 4
5865 if (next_full_store < 0) {
5866 // Conservative tack: Zero to end of current word.
5867 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5868 } else {
5869 // Zero to beginning of next fully initialized word.
5870 // Or, don't zero at all, if we are already in that word.
5871 assert(next_full_store >= zeroes_needed, "must go forward");
5872 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5873 zeroes_needed = next_full_store;
5874 }
5875 }
5876
5877 if (zeroes_needed > zeroes_done) {
5878 intptr_t zsize = zeroes_needed - zeroes_done;
5879 // Do some incremental zeroing on rawmem, in parallel with inits.
5880 zeroes_done = align_down(zeroes_done, BytesPerInt);
5881 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5882 allocation()->in(AllocateNode::InitValue),
5883 allocation()->in(AllocateNode::RawInitValue),
5884 zeroes_done, zeroes_needed,
5885 true,
5886 phase);
5887 zeroes_done = zeroes_needed;
5888 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5889 do_zeroing = false; // leave the hole, next time
5890 }
5891 }
5892
5893 // Collect the store and move on:
5894 phase->replace_input_of(st, MemNode::Memory, inits);
5895 inits = st; // put it on the linearized chain
5896 set_req(i, zmem); // unhook from previous position
5897
5898 if (zeroes_done == st_off)
5899 zeroes_done = next_init_off;
5900
5901 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5902
5903 #ifdef ASSERT
5904 // Various order invariants. Weaker than stores_are_sane because
5905 // a large constant tile can be filled in by smaller non-constant stores.
5906 assert(st_off >= last_init_off, "inits do not reverse");
5907 last_init_off = st_off;
5908 const Type* val = nullptr;
5909 if (st_size >= BytesPerInt &&
5910 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5911 (int)val->basic_type() < (int)T_OBJECT) {
5912 assert(st_off >= last_tile_end, "tiles do not overlap");
5913 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5914 last_tile_end = MAX2(last_tile_end, next_init_off);
5915 } else {
5916 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5917 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5918 assert(st_off >= last_init_end, "inits do not overlap");
5919 last_init_end = next_init_off; // it's a non-tile
5920 }
5921 #endif //ASSERT
5922 }
5923
5924 remove_extra_zeroes(); // clear out all the zmems left over
5925 add_req(inits);
5926
5927 if (!(UseTLAB && ZeroTLAB)) {
5928 // If anything remains to be zeroed, zero it all now.
5929 zeroes_done = align_down(zeroes_done, BytesPerInt);
5930 // if it is the last unused 4 bytes of an instance, forget about it
5931 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5932 if (zeroes_done + BytesPerLong >= size_limit) {
5933 AllocateNode* alloc = allocation();
5934 assert(alloc != nullptr, "must be present");
5935 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5936 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5937 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5938 if (zeroes_done == k->layout_helper())
5939 zeroes_done = size_limit;
5940 }
5941 }
5942 if (zeroes_done < size_limit) {
5943 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5944 allocation()->in(AllocateNode::InitValue),
5945 allocation()->in(AllocateNode::RawInitValue),
5946 zeroes_done, size_in_bytes, true, phase);
5947 }
5948 }
5949
5950 set_complete(phase);
5951 return rawmem;
5952 }
5953
5954 void InitializeNode::replace_mem_projs_by(Node* mem, Compile* C) {
5955 auto replace_proj = [&](ProjNode* proj) {
5956 C->gvn_replace_by(proj, mem);
5957 return CONTINUE;
5958 };
5959 apply_to_projs(replace_proj, TypeFunc::Memory);
5960 }
5961
5962 void InitializeNode::replace_mem_projs_by(Node* mem, PhaseIterGVN* igvn) {
5963 DUIterator_Fast imax, i = fast_outs(imax);
5964 auto replace_proj = [&](ProjNode* proj) {
5965 igvn->replace_node(proj, mem);
5966 --i; --imax;
5967 return CONTINUE;
5968 };
5969 apply_to_projs(imax, i, replace_proj, TypeFunc::Memory);
5970 }
5971
5972 bool InitializeNode::already_has_narrow_mem_proj_with_adr_type(const TypePtr* adr_type) const {
5973 auto find_proj = [&](ProjNode* proj) {
5974 if (proj->adr_type() == adr_type) {
5975 return BREAK_AND_RETURN_CURRENT_PROJ;
5976 }
5977 return CONTINUE;
5978 };
5979 DUIterator_Fast imax, i = fast_outs(imax);
5980 return apply_to_narrow_mem_projs_any_iterator(UsesIteratorFast(imax, i, this), find_proj) != nullptr;
5981 }
5982
5983 MachProjNode* InitializeNode::mem_mach_proj() const {
5984 auto find_proj = [](ProjNode* proj) {
5985 if (proj->is_MachProj()) {
5986 return BREAK_AND_RETURN_CURRENT_PROJ;
5987 }
5988 return CONTINUE;
5989 };
5990 ProjNode* proj = apply_to_projs(find_proj, TypeFunc::Memory);
5991 if (proj == nullptr) {
5992 return nullptr;
5993 }
5994 return proj->as_MachProj();
5995 }
5996
5997 #ifdef ASSERT
5998 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5999 if (is_complete())
6000 return true; // stores could be anything at this point
6001 assert(allocation() != nullptr, "must be present");
6002 intptr_t last_off = allocation()->minimum_header_size();
6003 for (uint i = InitializeNode::RawStores; i < req(); i++) {
6004 Node* st = in(i);
6005 intptr_t st_off = get_store_offset(st, phase);
6006 if (st_off < 0) continue; // ignore dead garbage
6007 if (last_off > st_off) {
6008 tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off);
6009 this->dump(2);
6010 assert(false, "ascending store offsets");
6011 return false;
6012 }
6013 last_off = st_off + st->as_Store()->memory_size();
6014 }
6015 return true;
6016 }
6017 #endif //ASSERT
6018
6019
6020
6021
6022 //============================MergeMemNode=====================================
6023 //
6024 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
6025 // contributing store or call operations. Each contributor provides the memory
6026 // state for a particular "alias type" (see Compile::alias_type). For example,
6027 // if a MergeMem has an input X for alias category #6, then any memory reference
6028 // to alias category #6 may use X as its memory state input, as an exact equivalent
6029 // to using the MergeMem as a whole.
6030 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
6031 //
6032 // (Here, the <N> notation gives the index of the relevant adr_type.)
6033 //
6034 // In one special case (and more cases in the future), alias categories overlap.
6035 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
6036 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
6037 // it is exactly equivalent to that state W:
6038 // MergeMem(<Bot>: W) <==> W
6039 //
6040 // Usually, the merge has more than one input. In that case, where inputs
6041 // overlap (i.e., one is Bot), the narrower alias type determines the memory
6042 // state for that type, and the wider alias type (Bot) fills in everywhere else:
6043 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
6044 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
6045 //
6046 // A merge can take a "wide" memory state as one of its narrow inputs.
6047 // This simply means that the merge observes out only the relevant parts of
6048 // the wide input. That is, wide memory states arriving at narrow merge inputs
6049 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
6050 //
6051 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
6052 // and that memory slices "leak through":
6053 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
6054 //
6055 // But, in such a cascade, repeated memory slices can "block the leak":
6056 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
6057 //
6058 // In the last example, Y is not part of the combined memory state of the
6059 // outermost MergeMem. The system must, of course, prevent unschedulable
6060 // memory states from arising, so you can be sure that the state Y is somehow
6061 // a precursor to state Y'.
6062 //
6063 //
6064 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
6065 // of each MergeMemNode array are exactly the numerical alias indexes, including
6066 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
6067 // Compile::alias_type (and kin) produce and manage these indexes.
6068 //
6069 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
6070 // (Note that this provides quick access to the top node inside MergeMem methods,
6071 // without the need to reach out via TLS to Compile::current.)
6072 //
6073 // As a consequence of what was just described, a MergeMem that represents a full
6074 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
6075 // containing all alias categories.
6076 //
6077 // MergeMem nodes never (?) have control inputs, so in(0) is null.
6078 //
6079 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
6080 // a memory state for the alias type <N>, or else the top node, meaning that
6081 // there is no particular input for that alias type. Note that the length of
6082 // a MergeMem is variable, and may be extended at any time to accommodate new
6083 // memory states at larger alias indexes. When merges grow, they are of course
6084 // filled with "top" in the unused in() positions.
6085 //
6086 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
6087 // (Top was chosen because it works smoothly with passes like GCM.)
6088 //
6089 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
6090 // the type of random VM bits like TLS references.) Since it is always the
6091 // first non-Bot memory slice, some low-level loops use it to initialize an
6092 // index variable: for (i = AliasIdxRaw; i < req(); i++).
6093 //
6094 //
6095 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
6096 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
6097 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
6098 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
6099 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
6100 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
6101 //
6102 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
6103 // really that different from the other memory inputs. An abbreviation called
6104 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
6105 //
6106 //
6107 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
6108 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
6109 // that "emerges though" the base memory will be marked as excluding the alias types
6110 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
6111 //
6112 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
6113 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
6114 //
6115 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
6116 // (It is currently unimplemented.) As you can see, the resulting merge is
6117 // actually a disjoint union of memory states, rather than an overlay.
6118 //
6119
6120 //------------------------------MergeMemNode-----------------------------------
6121 Node* MergeMemNode::make_empty_memory() {
6122 Node* empty_memory = (Node*) Compile::current()->top();
6123 assert(empty_memory->is_top(), "correct sentinel identity");
6124 return empty_memory;
6125 }
6126
6127 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
6128 init_class_id(Class_MergeMem);
6129 // all inputs are nullified in Node::Node(int)
6130 // set_input(0, nullptr); // no control input
6131
6132 // Initialize the edges uniformly to top, for starters.
6133 Node* empty_mem = make_empty_memory();
6134 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
6135 init_req(i,empty_mem);
6136 }
6137 assert(empty_memory() == empty_mem, "");
6138
6139 if( new_base != nullptr && new_base->is_MergeMem() ) {
6140 MergeMemNode* mdef = new_base->as_MergeMem();
6141 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
6142 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
6143 mms.set_memory(mms.memory2());
6144 }
6145 assert(base_memory() == mdef->base_memory(), "");
6146 } else {
6147 set_base_memory(new_base);
6148 }
6149 }
6150
6151 // Make a new, untransformed MergeMem with the same base as 'mem'.
6152 // If mem is itself a MergeMem, populate the result with the same edges.
6153 MergeMemNode* MergeMemNode::make(Node* mem) {
6154 return new MergeMemNode(mem);
6155 }
6156
6157 //------------------------------cmp--------------------------------------------
6158 uint MergeMemNode::hash() const { return NO_HASH; }
6159 bool MergeMemNode::cmp( const Node &n ) const {
6160 return (&n == this); // Always fail except on self
6161 }
6162
6163 //------------------------------Identity---------------------------------------
6164 Node* MergeMemNode::Identity(PhaseGVN* phase) {
6165 // Identity if this merge point does not record any interesting memory
6166 // disambiguations.
6167 Node* base_mem = base_memory();
6168 Node* empty_mem = empty_memory();
6169 if (base_mem != empty_mem) { // Memory path is not dead?
6170 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6171 Node* mem = in(i);
6172 if (mem != empty_mem && mem != base_mem) {
6173 return this; // Many memory splits; no change
6174 }
6175 }
6176 }
6177 return base_mem; // No memory splits; ID on the one true input
6178 }
6179
6180 //------------------------------Ideal------------------------------------------
6181 // This method is invoked recursively on chains of MergeMem nodes
6182 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
6183 if (Identity(phase) != this) {
6184 // Let Identity handle this case
6185 return nullptr;
6186 }
6187
6188 // Remove chain'd MergeMems
6189 //
6190 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
6191 // relative to the "in(Bot)". Since we are patching both at the same time,
6192 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
6193 // but rewrite each "in(i)" relative to the new "in(Bot)".
6194 Node *progress = nullptr;
6195
6196
6197 Node* old_base = base_memory();
6198 Node* empty_mem = empty_memory();
6199 if (old_base == empty_mem)
6200 return nullptr; // Dead memory path.
6201
6202 MergeMemNode* old_mbase;
6203 if (old_base != nullptr && old_base->is_MergeMem())
6204 old_mbase = old_base->as_MergeMem();
6205 else
6206 old_mbase = nullptr;
6207 Node* new_base = old_base;
6208
6209 // simplify stacked MergeMems in base memory
6210 if (old_mbase) new_base = old_mbase->base_memory();
6211
6212 // the base memory might contribute new slices beyond my req()
6213 if (old_mbase) grow_to_match(old_mbase);
6214
6215 // Note: We do not call verify_sparse on entry, because inputs
6216 // can normalize to the base_memory via subsume_node or similar
6217 // mechanisms. This method repairs that damage.
6218
6219 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
6220
6221 // Look at each slice.
6222 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6223 Node* old_in = in(i);
6224 // calculate the old memory value
6225 Node* old_mem = old_in;
6226 if (old_mem == empty_mem) old_mem = old_base;
6227 assert(old_mem == memory_at(i), "");
6228
6229 // maybe update (reslice) the old memory value
6230
6231 // simplify stacked MergeMems
6232 Node* new_mem = old_mem;
6233 MergeMemNode* old_mmem;
6234 if (old_mem != nullptr && old_mem->is_MergeMem())
6235 old_mmem = old_mem->as_MergeMem();
6236 else
6237 old_mmem = nullptr;
6238 if (old_mmem == this) {
6239 // This can happen if loops break up and safepoints disappear.
6240 // A merge of BotPtr (default) with a RawPtr memory derived from a
6241 // safepoint can be rewritten to a merge of the same BotPtr with
6242 // the BotPtr phi coming into the loop. If that phi disappears
6243 // also, we can end up with a self-loop of the mergemem.
6244 // In general, if loops degenerate and memory effects disappear,
6245 // a mergemem can be left looking at itself. This simply means
6246 // that the mergemem's default should be used, since there is
6247 // no longer any apparent effect on this slice.
6248 // Note: If a memory slice is a MergeMem cycle, it is unreachable
6249 // from start. Update the input to TOP.
6250 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
6251 }
6252 else if (old_mmem != nullptr) {
6253 new_mem = old_mmem->memory_at(i);
6254 }
6255 // else preceding memory was not a MergeMem
6256
6257 // maybe store down a new value
6258 Node* new_in = new_mem;
6259 if (new_in == new_base) new_in = empty_mem;
6260
6261 if (new_in != old_in) {
6262 // Warning: Do not combine this "if" with the previous "if"
6263 // A memory slice might have be be rewritten even if it is semantically
6264 // unchanged, if the base_memory value has changed.
6265 set_req_X(i, new_in, phase);
6266 progress = this; // Report progress
6267 }
6268 }
6269
6270 if (new_base != old_base) {
6271 set_req_X(Compile::AliasIdxBot, new_base, phase);
6272 // Don't use set_base_memory(new_base), because we need to update du.
6273 assert(base_memory() == new_base, "");
6274 progress = this;
6275 }
6276
6277 if( base_memory() == this ) {
6278 // a self cycle indicates this memory path is dead
6279 set_req(Compile::AliasIdxBot, empty_mem);
6280 }
6281
6282 // Resolve external cycles by calling Ideal on a MergeMem base_memory
6283 // Recursion must occur after the self cycle check above
6284 if( base_memory()->is_MergeMem() ) {
6285 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
6286 Node *m = phase->transform(new_mbase); // Rollup any cycles
6287 if( m != nullptr &&
6288 (m->is_top() ||
6289 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
6290 // propagate rollup of dead cycle to self
6291 set_req(Compile::AliasIdxBot, empty_mem);
6292 }
6293 }
6294
6295 if( base_memory() == empty_mem ) {
6296 progress = this;
6297 // Cut inputs during Parse phase only.
6298 // During Optimize phase a dead MergeMem node will be subsumed by Top.
6299 if( !can_reshape ) {
6300 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6301 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
6302 }
6303 }
6304 }
6305
6306 if( !progress && base_memory()->is_Phi() && can_reshape ) {
6307 // Check if PhiNode::Ideal's "Split phis through memory merges"
6308 // transform should be attempted. Look for this->phi->this cycle.
6309 uint merge_width = req();
6310 if (merge_width > Compile::AliasIdxRaw) {
6311 PhiNode* phi = base_memory()->as_Phi();
6312 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
6313 if (phi->in(i) == this) {
6314 phase->is_IterGVN()->_worklist.push(phi);
6315 break;
6316 }
6317 }
6318 }
6319 }
6320
6321 assert(progress || verify_sparse(), "please, no dups of base");
6322 return progress;
6323 }
6324
6325 //-------------------------set_base_memory-------------------------------------
6326 void MergeMemNode::set_base_memory(Node *new_base) {
6327 Node* empty_mem = empty_memory();
6328 set_req(Compile::AliasIdxBot, new_base);
6329 assert(memory_at(req()) == new_base, "must set default memory");
6330 // Clear out other occurrences of new_base:
6331 if (new_base != empty_mem) {
6332 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6333 if (in(i) == new_base) set_req(i, empty_mem);
6334 }
6335 }
6336 }
6337
6338 //------------------------------out_RegMask------------------------------------
6339 const RegMask &MergeMemNode::out_RegMask() const {
6340 return RegMask::EMPTY;
6341 }
6342
6343 //------------------------------dump_spec--------------------------------------
6344 #ifndef PRODUCT
6345 void MergeMemNode::dump_spec(outputStream *st) const {
6346 st->print(" {");
6347 Node* base_mem = base_memory();
6348 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
6349 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
6350 if (mem == base_mem) { st->print(" -"); continue; }
6351 st->print( " N%d:", mem->_idx );
6352 Compile::current()->get_adr_type(i)->dump_on(st);
6353 }
6354 st->print(" }");
6355 }
6356 #endif // !PRODUCT
6357
6358
6359 #ifdef ASSERT
6360 static bool might_be_same(Node* a, Node* b) {
6361 if (a == b) return true;
6362 if (!(a->is_Phi() || b->is_Phi())) return false;
6363 // phis shift around during optimization
6364 return true; // pretty stupid...
6365 }
6366
6367 // verify a narrow slice (either incoming or outgoing)
6368 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
6369 if (!VerifyAliases) return; // don't bother to verify unless requested
6370 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
6371 if (Node::in_dump()) return; // muzzle asserts when printing
6372 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
6373 assert(n != nullptr, "");
6374 // Elide intervening MergeMem's
6375 while (n->is_MergeMem()) {
6376 n = n->as_MergeMem()->memory_at(alias_idx);
6377 }
6378 Compile* C = Compile::current();
6379 const TypePtr* n_adr_type = n->adr_type();
6380 if (n == m->empty_memory()) {
6381 // Implicit copy of base_memory()
6382 } else if (n_adr_type != TypePtr::BOTTOM) {
6383 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
6384 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
6385 } else {
6386 // A few places like make_runtime_call "know" that VM calls are narrow,
6387 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
6388 bool expected_wide_mem = false;
6389 if (n == m->base_memory()) {
6390 expected_wide_mem = true;
6391 } else if (alias_idx == Compile::AliasIdxRaw ||
6392 n == m->memory_at(Compile::AliasIdxRaw)) {
6393 expected_wide_mem = true;
6394 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
6395 // memory can "leak through" calls on channels that
6396 // are write-once. Allow this also.
6397 expected_wide_mem = true;
6398 }
6399 assert(expected_wide_mem, "expected narrow slice replacement");
6400 }
6401 }
6402 #else // !ASSERT
6403 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
6404 #endif
6405
6406
6407 //-----------------------------memory_at---------------------------------------
6408 Node* MergeMemNode::memory_at(uint alias_idx) const {
6409 assert(alias_idx >= Compile::AliasIdxRaw ||
6410 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
6411 "must avoid base_memory and AliasIdxTop");
6412
6413 // Otherwise, it is a narrow slice.
6414 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
6415 if (is_empty_memory(n)) {
6416 // the array is sparse; empty slots are the "top" node
6417 n = base_memory();
6418 assert(Node::in_dump()
6419 || n == nullptr || n->bottom_type() == Type::TOP
6420 || n->adr_type() == nullptr // address is TOP
6421 || n->adr_type() == TypePtr::BOTTOM
6422 || n->adr_type() == TypeRawPtr::BOTTOM
6423 || n->is_NarrowMemProj()
6424 || !Compile::current()->do_aliasing(),
6425 "must be a wide memory");
6426 // do_aliasing == false if we are organizing the memory states manually.
6427 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
6428 } else {
6429 // make sure the stored slice is sane
6430 #ifdef ASSERT
6431 if (VMError::is_error_reported() || Node::in_dump()) {
6432 } else if (might_be_same(n, base_memory())) {
6433 // Give it a pass: It is a mostly harmless repetition of the base.
6434 // This can arise normally from node subsumption during optimization.
6435 } else {
6436 verify_memory_slice(this, alias_idx, n);
6437 }
6438 #endif
6439 }
6440 return n;
6441 }
6442
6443 //---------------------------set_memory_at-------------------------------------
6444 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
6445 verify_memory_slice(this, alias_idx, n);
6446 Node* empty_mem = empty_memory();
6447 if (n == base_memory()) n = empty_mem; // collapse default
6448 uint need_req = alias_idx+1;
6449 if (req() < need_req) {
6450 if (n == empty_mem) return; // already the default, so do not grow me
6451 // grow the sparse array
6452 do {
6453 add_req(empty_mem);
6454 } while (req() < need_req);
6455 }
6456 set_req( alias_idx, n );
6457 }
6458
6459
6460
6461 //--------------------------iteration_setup------------------------------------
6462 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
6463 if (other != nullptr) {
6464 grow_to_match(other);
6465 // invariant: the finite support of mm2 is within mm->req()
6466 #ifdef ASSERT
6467 for (uint i = req(); i < other->req(); i++) {
6468 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
6469 }
6470 #endif
6471 }
6472 // Replace spurious copies of base_memory by top.
6473 Node* base_mem = base_memory();
6474 if (base_mem != nullptr && !base_mem->is_top()) {
6475 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
6476 if (in(i) == base_mem)
6477 set_req(i, empty_memory());
6478 }
6479 }
6480 }
6481
6482 //---------------------------grow_to_match-------------------------------------
6483 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
6484 Node* empty_mem = empty_memory();
6485 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
6486 // look for the finite support of the other memory
6487 for (uint i = other->req(); --i >= req(); ) {
6488 if (other->in(i) != empty_mem) {
6489 uint new_len = i+1;
6490 while (req() < new_len) add_req(empty_mem);
6491 break;
6492 }
6493 }
6494 }
6495
6496 //---------------------------verify_sparse-------------------------------------
6497 #ifndef PRODUCT
6498 bool MergeMemNode::verify_sparse() const {
6499 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
6500 Node* base_mem = base_memory();
6501 // The following can happen in degenerate cases, since empty==top.
6502 if (is_empty_memory(base_mem)) return true;
6503 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6504 assert(in(i) != nullptr, "sane slice");
6505 if (in(i) == base_mem) return false; // should have been the sentinel value!
6506 }
6507 return true;
6508 }
6509
6510 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
6511 Node* n;
6512 n = mm->in(idx);
6513 if (mem == n) return true; // might be empty_memory()
6514 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
6515 if (mem == n) return true;
6516 return false;
6517 }
6518 #endif // !PRODUCT