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