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