1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/javaClasses.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "gc/shared/tlab_globals.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "oops/objArrayKlass.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/regalloc.hpp"
38 #include "opto/compile.hpp"
39 #include "opto/connode.hpp"
40 #include "opto/convertnode.hpp"
41 #include "opto/loopnode.hpp"
42 #include "opto/machnode.hpp"
43 #include "opto/matcher.hpp"
44 #include "opto/memnode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/phaseX.hpp"
48 #include "opto/regmask.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/vectornode.hpp"
51 #include "utilities/align.hpp"
52 #include "utilities/copy.hpp"
53 #include "utilities/macros.hpp"
54 #include "utilities/powerOfTwo.hpp"
55 #include "utilities/vmError.hpp"
56
57 // Portions of code courtesy of Clifford Click
58
59 // Optimization - Graph Style
60
61 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
62
63 //=============================================================================
64 uint MemNode::size_of() const { return sizeof(*this); }
65
66 const TypePtr *MemNode::adr_type() const {
67 Node* adr = in(Address);
68 if (adr == nullptr) return nullptr; // node is dead
69 const TypePtr* cross_check = nullptr;
70 DEBUG_ONLY(cross_check = _adr_type);
71 return calculate_adr_type(adr->bottom_type(), cross_check);
72 }
73
74 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
75 if (adr != nullptr) {
76 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
77 return true;
78 }
79 }
80 return false;
81 }
82
83 #ifndef PRODUCT
84 void MemNode::dump_spec(outputStream *st) const {
85 if (in(Address) == nullptr) return; // node is dead
86 #ifndef ASSERT
87 // fake the missing field
88 const TypePtr* _adr_type = nullptr;
89 if (in(Address) != nullptr)
90 _adr_type = in(Address)->bottom_type()->isa_ptr();
91 #endif
92 dump_adr_type(this, _adr_type, st);
93
94 Compile* C = Compile::current();
95 if (C->alias_type(_adr_type)->is_volatile()) {
96 st->print(" Volatile!");
97 }
98 if (_unaligned_access) {
99 st->print(" unaligned");
100 }
101 if (_mismatched_access) {
102 st->print(" mismatched");
103 }
104 if (_unsafe_access) {
105 st->print(" unsafe");
106 }
107 }
108
109 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
110 st->print(" @");
111 if (adr_type == nullptr) {
112 st->print("null");
113 } else {
114 adr_type->dump_on(st);
115 Compile* C = Compile::current();
116 Compile::AliasType* atp = nullptr;
117 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
118 if (atp == nullptr)
119 st->print(", idx=?\?;");
120 else if (atp->index() == Compile::AliasIdxBot)
121 st->print(", idx=Bot;");
122 else if (atp->index() == Compile::AliasIdxTop)
123 st->print(", idx=Top;");
124 else if (atp->index() == Compile::AliasIdxRaw)
125 st->print(", idx=Raw;");
126 else {
127 ciField* field = atp->field();
128 if (field) {
129 st->print(", name=");
130 field->print_name_on(st);
131 }
132 st->print(", idx=%d;", atp->index());
133 }
134 }
135 }
136
137 extern void print_alias_types();
138
139 #endif
140
141 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
142 assert((t_oop != nullptr), "sanity");
143 bool is_instance = t_oop->is_known_instance_field();
144 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
145 (load != nullptr) && load->is_Load() &&
146 (phase->is_IterGVN() != nullptr);
147 if (!(is_instance || is_boxed_value_load))
148 return mchain; // don't try to optimize non-instance types
149 uint instance_id = t_oop->instance_id();
150 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
151 Node *prev = nullptr;
152 Node *result = mchain;
153 while (prev != result) {
154 prev = result;
155 if (result == start_mem)
156 break; // hit one of our sentinels
157 // skip over a call which does not affect this memory slice
158 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
159 Node *proj_in = result->in(0);
160 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
161 break; // hit one of our sentinels
162 } else if (proj_in->is_Call()) {
163 // ArrayCopyNodes processed here as well
164 CallNode *call = proj_in->as_Call();
165 if (!call->may_modify(t_oop, phase)) { // returns false for instances
166 result = call->in(TypeFunc::Memory);
167 }
168 } else if (proj_in->is_Initialize()) {
169 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
170 // Stop if this is the initialization for the object instance which
171 // contains this memory slice, otherwise skip over it.
172 if ((alloc == nullptr) || (alloc->_idx == instance_id)) {
173 break;
174 }
175 if (is_instance) {
176 result = proj_in->in(TypeFunc::Memory);
177 } else if (is_boxed_value_load) {
178 Node* klass = alloc->in(AllocateNode::KlassNode);
179 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
180 if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) {
181 result = proj_in->in(TypeFunc::Memory); // not related allocation
182 }
183 }
184 } else if (proj_in->is_MemBar()) {
185 ArrayCopyNode* ac = nullptr;
186 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
187 break;
188 }
189 result = proj_in->in(TypeFunc::Memory);
190 } else if (proj_in->is_top()) {
191 break; // dead code
192 } else {
193 assert(false, "unexpected projection");
194 }
195 } else if (result->is_ClearArray()) {
196 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
197 // Can not bypass initialization of the instance
198 // we are looking for.
199 break;
200 }
201 // Otherwise skip it (the call updated 'result' value).
202 } else if (result->is_MergeMem()) {
203 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty);
204 }
205 }
206 return result;
207 }
208
209 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
210 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
211 if (t_oop == nullptr)
212 return mchain; // don't try to optimize non-oop types
213 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
214 bool is_instance = t_oop->is_known_instance_field();
215 PhaseIterGVN *igvn = phase->is_IterGVN();
216 if (is_instance && igvn != nullptr && result->is_Phi()) {
217 PhiNode *mphi = result->as_Phi();
218 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
219 const TypePtr *t = mphi->adr_type();
220 bool do_split = false;
221 // In the following cases, Load memory input can be further optimized based on
222 // its precise address type
223 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
224 do_split = true;
225 } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
226 const TypeOopPtr* mem_t =
227 t->is_oopptr()->cast_to_exactness(true)
228 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
229 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
230 if (t_oop->isa_aryptr()) {
231 mem_t = mem_t->is_aryptr()
232 ->cast_to_stable(t_oop->is_aryptr()->is_stable())
233 ->cast_to_size(t_oop->is_aryptr()->size())
234 ->with_offset(t_oop->is_aryptr()->offset())
235 ->is_aryptr();
236 }
237 do_split = mem_t == t_oop;
238 }
239 if (do_split) {
240 // clone the Phi with our address type
241 result = mphi->split_out_instance(t_adr, igvn);
242 } else {
243 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
244 }
245 }
246 return result;
247 }
248
249 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
250 uint alias_idx = phase->C->get_alias_index(tp);
251 Node *mem = mmem;
252 #ifdef ASSERT
253 {
254 // Check that current type is consistent with the alias index used during graph construction
255 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
256 bool consistent = adr_check == nullptr || adr_check->empty() ||
257 phase->C->must_alias(adr_check, alias_idx );
258 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
259 if( !consistent && adr_check != nullptr && !adr_check->empty() &&
260 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
261 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
262 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
263 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
264 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
265 // don't assert if it is dead code.
266 consistent = true;
267 }
268 if( !consistent ) {
269 st->print("alias_idx==%d, adr_check==", alias_idx);
270 if( adr_check == nullptr ) {
271 st->print("null");
272 } else {
273 adr_check->dump();
274 }
275 st->cr();
276 print_alias_types();
277 assert(consistent, "adr_check must match alias idx");
278 }
279 }
280 #endif
281 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
282 // means an array I have not precisely typed yet. Do not do any
283 // alias stuff with it any time soon.
284 const TypeOopPtr *toop = tp->isa_oopptr();
285 if (tp->base() != Type::AnyPtr &&
286 !(toop &&
287 toop->isa_instptr() &&
288 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
289 toop->offset() == Type::OffsetBot)) {
290 // compress paths and change unreachable cycles to TOP
291 // If not, we can update the input infinitely along a MergeMem cycle
292 // Equivalent code in PhiNode::Ideal
293 Node* m = phase->transform(mmem);
294 // If transformed to a MergeMem, get the desired slice
295 // Otherwise the returned node represents memory for every slice
296 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
297 // Update input if it is progress over what we have now
298 }
299 return mem;
300 }
301
302 //--------------------------Ideal_common---------------------------------------
303 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
304 // Unhook non-raw memories from complete (macro-expanded) initializations.
305 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
306 // If our control input is a dead region, kill all below the region
307 Node *ctl = in(MemNode::Control);
308 if (ctl && remove_dead_region(phase, can_reshape))
309 return this;
310 ctl = in(MemNode::Control);
311 // Don't bother trying to transform a dead node
312 if (ctl && ctl->is_top()) return NodeSentinel;
313
314 PhaseIterGVN *igvn = phase->is_IterGVN();
315 // Wait if control on the worklist.
316 if (ctl && can_reshape && igvn != nullptr) {
317 Node* bol = nullptr;
318 Node* cmp = nullptr;
319 if (ctl->in(0)->is_If()) {
320 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
321 bol = ctl->in(0)->in(1);
322 if (bol->is_Bool())
323 cmp = ctl->in(0)->in(1)->in(1);
324 }
325 if (igvn->_worklist.member(ctl) ||
326 (bol != nullptr && igvn->_worklist.member(bol)) ||
327 (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
328 // This control path may be dead.
329 // Delay this memory node transformation until the control is processed.
330 igvn->_worklist.push(this);
331 return NodeSentinel; // caller will return null
332 }
333 }
334 // Ignore if memory is dead, or self-loop
335 Node *mem = in(MemNode::Memory);
336 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
337 assert(mem != this, "dead loop in MemNode::Ideal");
338
339 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
340 // This memory slice may be dead.
341 // Delay this mem node transformation until the memory is processed.
342 igvn->_worklist.push(this);
343 return NodeSentinel; // caller will return null
344 }
345
346 Node *address = in(MemNode::Address);
347 const Type *t_adr = phase->type(address);
348 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null
349
350 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
351 // Unsafe off-heap access with zero address. Remove access and other control users
352 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
353 assert(ctl != nullptr, "unsafe accesses should be control dependent");
354 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
355 Node* u = ctl->fast_out(i);
356 if (u != ctl) {
357 igvn->rehash_node_delayed(u);
358 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
359 --i, imax -= nb;
360 }
361 }
362 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
363 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
364 phase->C->root()->add_req(halt);
365 return this;
366 }
367
368 if (can_reshape && igvn != nullptr &&
369 (igvn->_worklist.member(address) ||
370 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
371 // The address's base and type may change when the address is processed.
372 // Delay this mem node transformation until the address is processed.
373 igvn->_worklist.push(this);
374 return NodeSentinel; // caller will return null
375 }
376
377 // Do NOT remove or optimize the next lines: ensure a new alias index
378 // is allocated for an oop pointer type before Escape Analysis.
379 // Note: C++ will not remove it since the call has side effect.
380 if (t_adr->isa_oopptr()) {
381 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
382 }
383
384 Node* base = nullptr;
385 if (address->is_AddP()) {
386 base = address->in(AddPNode::Base);
387 }
388 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
389 !t_adr->isa_rawptr()) {
390 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
391 // Skip this node optimization if its address has TOP base.
392 return NodeSentinel; // caller will return null
393 }
394
395 // Avoid independent memory operations
396 Node* old_mem = mem;
397
398 // The code which unhooks non-raw memories from complete (macro-expanded)
399 // initializations was removed. After macro-expansion all stores caught
400 // by Initialize node became raw stores and there is no information
401 // which memory slices they modify. So it is unsafe to move any memory
402 // operation above these stores. Also in most cases hooked non-raw memories
403 // were already unhooked by using information from detect_ptr_independence()
404 // and find_previous_store().
405
406 if (mem->is_MergeMem()) {
407 MergeMemNode* mmem = mem->as_MergeMem();
408 const TypePtr *tp = t_adr->is_ptr();
409
410 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
411 }
412
413 if (mem != old_mem) {
414 set_req_X(MemNode::Memory, mem, phase);
415 if (phase->type(mem) == Type::TOP) return NodeSentinel;
416 return this;
417 }
418
419 // let the subclass continue analyzing...
420 return nullptr;
421 }
422
423 // Helper function for proving some simple control dominations.
424 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
425 // Already assumes that 'dom' is available at 'sub', and that 'sub'
426 // is not a constant (dominated by the method's StartNode).
427 // Used by MemNode::find_previous_store to prove that the
428 // control input of a memory operation predates (dominates)
429 // an allocation it wants to look past.
430 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
431 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top())
432 return false; // Conservative answer for dead code
433
434 // Check 'dom'. Skip Proj and CatchProj nodes.
435 dom = dom->find_exact_control(dom);
436 if (dom == nullptr || dom->is_top())
437 return false; // Conservative answer for dead code
438
439 if (dom == sub) {
440 // For the case when, for example, 'sub' is Initialize and the original
441 // 'dom' is Proj node of the 'sub'.
442 return false;
443 }
444
445 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
446 return true;
447
448 // 'dom' dominates 'sub' if its control edge and control edges
449 // of all its inputs dominate or equal to sub's control edge.
450
451 // Currently 'sub' is either Allocate, Initialize or Start nodes.
452 // Or Region for the check in LoadNode::Ideal();
453 // 'sub' should have sub->in(0) != nullptr.
454 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
455 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
456
457 // Get control edge of 'sub'.
458 Node* orig_sub = sub;
459 sub = sub->find_exact_control(sub->in(0));
460 if (sub == nullptr || sub->is_top())
461 return false; // Conservative answer for dead code
462
463 assert(sub->is_CFG(), "expecting control");
464
465 if (sub == dom)
466 return true;
467
468 if (sub->is_Start() || sub->is_Root())
469 return false;
470
471 {
472 // Check all control edges of 'dom'.
473
474 ResourceMark rm;
475 Node_List nlist;
476 Unique_Node_List dom_list;
477
478 dom_list.push(dom);
479 bool only_dominating_controls = false;
480
481 for (uint next = 0; next < dom_list.size(); next++) {
482 Node* n = dom_list.at(next);
483 if (n == orig_sub)
484 return false; // One of dom's inputs dominated by sub.
485 if (!n->is_CFG() && n->pinned()) {
486 // Check only own control edge for pinned non-control nodes.
487 n = n->find_exact_control(n->in(0));
488 if (n == nullptr || n->is_top())
489 return false; // Conservative answer for dead code
490 assert(n->is_CFG(), "expecting control");
491 dom_list.push(n);
492 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
493 only_dominating_controls = true;
494 } else if (n->is_CFG()) {
495 if (n->dominates(sub, nlist))
496 only_dominating_controls = true;
497 else
498 return false;
499 } else {
500 // First, own control edge.
501 Node* m = n->find_exact_control(n->in(0));
502 if (m != nullptr) {
503 if (m->is_top())
504 return false; // Conservative answer for dead code
505 dom_list.push(m);
506 }
507 // Now, the rest of edges.
508 uint cnt = n->req();
509 for (uint i = 1; i < cnt; i++) {
510 m = n->find_exact_control(n->in(i));
511 if (m == nullptr || m->is_top())
512 continue;
513 dom_list.push(m);
514 }
515 }
516 }
517 return only_dominating_controls;
518 }
519 }
520
521 //---------------------detect_ptr_independence---------------------------------
522 // Used by MemNode::find_previous_store to prove that two base
523 // pointers are never equal.
524 // The pointers are accompanied by their associated allocations,
525 // if any, which have been previously discovered by the caller.
526 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
527 Node* p2, AllocateNode* a2,
528 PhaseTransform* phase) {
529 // Attempt to prove that these two pointers cannot be aliased.
530 // They may both manifestly be allocations, and they should differ.
531 // Or, if they are not both allocations, they can be distinct constants.
532 // Otherwise, one is an allocation and the other a pre-existing value.
533 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
534 return (p1 != p2) && p1->is_Con() && p2->is_Con();
535 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
536 return (a1 != a2);
537 } else if (a1 != nullptr) { // one allocation a1
538 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
539 return all_controls_dominate(p2, a1);
540 } else { //(a2 != null) // one allocation a2
541 return all_controls_dominate(p1, a2);
542 }
543 return false;
544 }
545
546
547 // Find an arraycopy ac that produces the memory state represented by parameter mem.
548 // Return ac if
549 // (a) can_see_stored_value=true and ac must have set the value for this load or if
550 // (b) can_see_stored_value=false and ac could have set the value for this load or if
551 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
552 // In case (c) change the parameter mem to the memory input of ac to skip it
553 // when searching stored value.
554 // Otherwise return null.
555 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
556 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
557 if (ac != nullptr) {
558 Node* ld_addp = in(MemNode::Address);
559 Node* src = ac->in(ArrayCopyNode::Src);
560 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
561
562 // This is a load from a cloned array. The corresponding arraycopy ac must
563 // have set the value for the load and we can return ac but only if the load
564 // is known to be within bounds. This is checked below.
565 if (ary_t != nullptr && ld_addp->is_AddP()) {
566 Node* ld_offs = ld_addp->in(AddPNode::Offset);
567 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
568 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
569 jlong elemsize = type2aelembytes(ary_elem);
570
571 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
572 const TypeInt* sizetype = ary_t->size();
573
574 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
575 // The load is known to be within bounds. It receives its value from ac.
576 return ac;
577 }
578 // The load is known to be out-of-bounds.
579 }
580 // The load could be out-of-bounds. It must not be hoisted but must remain
581 // dependent on the runtime range check. This is achieved by returning null.
582 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
583 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
584
585 if (ac->is_arraycopy_validated() ||
586 ac->is_copyof_validated() ||
587 ac->is_copyofrange_validated()) {
588 Node* ld_addp = in(MemNode::Address);
589 if (ld_addp->is_AddP()) {
590 Node* ld_base = ld_addp->in(AddPNode::Address);
591 Node* ld_offs = ld_addp->in(AddPNode::Offset);
592
593 Node* dest = ac->in(ArrayCopyNode::Dest);
594
595 if (dest == ld_base) {
596 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
597 assert(!ld_offs_t->empty(), "dead reference should be checked already");
598 // Take into account vector or unsafe access size
599 jlong ld_size_in_bytes = (jlong)memory_size();
600 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
601 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
602 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
603 return ac;
604 }
605 if (!can_see_stored_value) {
606 mem = ac->in(TypeFunc::Memory);
607 return ac;
608 }
609 }
610 }
611 }
612 }
613 return nullptr;
614 }
615
616 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
617 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
618 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
619 if (ld_alloc != nullptr) {
620 // Check if there is an array copy for a clone
621 Node* mb = mem->in(0);
622 ArrayCopyNode* ac = nullptr;
623 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
624 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
625 ac = mb->in(0)->in(0)->as_ArrayCopy();
626 } else {
627 // Step over GC barrier when ReduceInitialCardMarks is disabled
628 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
629 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
630
631 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
632 ac = control_proj_ac->in(0)->as_ArrayCopy();
633 }
634 }
635
636 if (ac != nullptr && ac->is_clonebasic()) {
637 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
638 if (alloc != nullptr && alloc == ld_alloc) {
639 return ac;
640 }
641 }
642 }
643 }
644 return nullptr;
645 }
646
647 // The logic for reordering loads and stores uses four steps:
648 // (a) Walk carefully past stores and initializations which we
649 // can prove are independent of this load.
650 // (b) Observe that the next memory state makes an exact match
651 // with self (load or store), and locate the relevant store.
652 // (c) Ensure that, if we were to wire self directly to the store,
653 // the optimizer would fold it up somehow.
654 // (d) Do the rewiring, and return, depending on some other part of
655 // the optimizer to fold up the load.
656 // This routine handles steps (a) and (b). Steps (c) and (d) are
657 // specific to loads and stores, so they are handled by the callers.
658 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
659 //
660 Node* MemNode::find_previous_store(PhaseValues* phase) {
661 Node* ctrl = in(MemNode::Control);
662 Node* adr = in(MemNode::Address);
663 intptr_t offset = 0;
664 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
665 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
666
667 if (offset == Type::OffsetBot)
668 return nullptr; // cannot unalias unless there are precise offsets
669
670 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
671 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
672
673 intptr_t size_in_bytes = memory_size();
674
675 Node* mem = in(MemNode::Memory); // start searching here...
676
677 int cnt = 50; // Cycle limiter
678 for (;;) { // While we can dance past unrelated stores...
679 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
680
681 Node* prev = mem;
682 if (mem->is_Store()) {
683 Node* st_adr = mem->in(MemNode::Address);
684 intptr_t st_offset = 0;
685 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
686 if (st_base == nullptr)
687 break; // inscrutable pointer
688
689 // For raw accesses it's not enough to prove that constant offsets don't intersect.
690 // We need the bases to be the equal in order for the offset check to make sense.
691 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
692 break;
693 }
694
695 if (st_offset != offset && st_offset != Type::OffsetBot) {
696 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
697 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
698 if (st_offset >= offset + size_in_bytes ||
699 st_offset <= offset - MAX_STORE ||
700 st_offset <= offset - mem->as_Store()->memory_size()) {
701 // Success: The offsets are provably independent.
702 // (You may ask, why not just test st_offset != offset and be done?
703 // The answer is that stores of different sizes can co-exist
704 // in the same sequence of RawMem effects. We sometimes initialize
705 // a whole 'tile' of array elements with a single jint or jlong.)
706 mem = mem->in(MemNode::Memory);
707 continue; // (a) advance through independent store memory
708 }
709 }
710 if (st_base != base &&
711 detect_ptr_independence(base, alloc,
712 st_base,
713 AllocateNode::Ideal_allocation(st_base),
714 phase)) {
715 // Success: The bases are provably independent.
716 mem = mem->in(MemNode::Memory);
717 continue; // (a) advance through independent store memory
718 }
719
720 // (b) At this point, if the bases or offsets do not agree, we lose,
721 // since we have not managed to prove 'this' and 'mem' independent.
722 if (st_base == base && st_offset == offset) {
723 return mem; // let caller handle steps (c), (d)
724 }
725
726 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
727 InitializeNode* st_init = mem->in(0)->as_Initialize();
728 AllocateNode* st_alloc = st_init->allocation();
729 if (st_alloc == nullptr)
730 break; // something degenerated
731 bool known_identical = false;
732 bool known_independent = false;
733 if (alloc == st_alloc)
734 known_identical = true;
735 else if (alloc != nullptr)
736 known_independent = true;
737 else if (all_controls_dominate(this, st_alloc))
738 known_independent = true;
739
740 if (known_independent) {
741 // The bases are provably independent: Either they are
742 // manifestly distinct allocations, or else the control
743 // of this load dominates the store's allocation.
744 int alias_idx = phase->C->get_alias_index(adr_type());
745 if (alias_idx == Compile::AliasIdxRaw) {
746 mem = st_alloc->in(TypeFunc::Memory);
747 } else {
748 mem = st_init->memory(alias_idx);
749 }
750 continue; // (a) advance through independent store memory
751 }
752
753 // (b) at this point, if we are not looking at a store initializing
754 // the same allocation we are loading from, we lose.
755 if (known_identical) {
756 // From caller, can_see_stored_value will consult find_captured_store.
757 return mem; // let caller handle steps (c), (d)
758 }
759
760 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) {
761 if (prev != mem) {
762 // Found an arraycopy but it doesn't affect that load
763 continue;
764 }
765 // Found an arraycopy that may affect that load
766 return mem;
767 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) {
768 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
769 if (mem->is_Proj() && mem->in(0)->is_Call()) {
770 // ArrayCopyNodes processed here as well.
771 CallNode *call = mem->in(0)->as_Call();
772 if (!call->may_modify(addr_t, phase)) {
773 mem = call->in(TypeFunc::Memory);
774 continue; // (a) advance through independent call memory
775 }
776 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
777 ArrayCopyNode* ac = nullptr;
778 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
779 break;
780 }
781 mem = mem->in(0)->in(TypeFunc::Memory);
782 continue; // (a) advance through independent MemBar memory
783 } else if (mem->is_ClearArray()) {
784 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
785 // (the call updated 'mem' value)
786 continue; // (a) advance through independent allocation memory
787 } else {
788 // Can not bypass initialization of the instance
789 // we are looking for.
790 return mem;
791 }
792 } else if (mem->is_MergeMem()) {
793 int alias_idx = phase->C->get_alias_index(adr_type());
794 mem = mem->as_MergeMem()->memory_at(alias_idx);
795 continue; // (a) advance through independent MergeMem memory
796 }
797 }
798
799 // Unless there is an explicit 'continue', we must bail out here,
800 // because 'mem' is an inscrutable memory state (e.g., a call).
801 break;
802 }
803
804 return nullptr; // bail out
805 }
806
807 //----------------------calculate_adr_type-------------------------------------
808 // Helper function. Notices when the given type of address hits top or bottom.
809 // Also, asserts a cross-check of the type against the expected address type.
810 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
811 if (t == Type::TOP) return nullptr; // does not touch memory any more?
812 #ifdef ASSERT
813 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
814 #endif
815 const TypePtr* tp = t->isa_ptr();
816 if (tp == nullptr) {
817 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
818 return TypePtr::BOTTOM; // touches lots of memory
819 } else {
820 #ifdef ASSERT
821 // %%%% [phh] We don't check the alias index if cross_check is
822 // TypeRawPtr::BOTTOM. Needs to be investigated.
823 if (cross_check != nullptr &&
824 cross_check != TypePtr::BOTTOM &&
825 cross_check != TypeRawPtr::BOTTOM) {
826 // Recheck the alias index, to see if it has changed (due to a bug).
827 Compile* C = Compile::current();
828 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
829 "must stay in the original alias category");
830 // The type of the address must be contained in the adr_type,
831 // disregarding "null"-ness.
832 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
833 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
834 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
835 "real address must not escape from expected memory type");
836 }
837 #endif
838 return tp;
839 }
840 }
841
842 uint8_t MemNode::barrier_data(const Node* n) {
843 if (n->is_LoadStore()) {
844 return n->as_LoadStore()->barrier_data();
845 } else if (n->is_Mem()) {
846 return n->as_Mem()->barrier_data();
847 }
848 return 0;
849 }
850
851 //=============================================================================
852 // Should LoadNode::Ideal() attempt to remove control edges?
853 bool LoadNode::can_remove_control() const {
854 return !has_pinned_control_dependency();
855 }
856 uint LoadNode::size_of() const { return sizeof(*this); }
857 bool LoadNode::cmp(const Node &n) const {
858 LoadNode& load = (LoadNode &)n;
859 return Type::equals(_type, load._type) &&
860 _control_dependency == load._control_dependency &&
861 _mo == load._mo;
862 }
863 const Type *LoadNode::bottom_type() const { return _type; }
864 uint LoadNode::ideal_reg() const {
865 return _type->ideal_reg();
866 }
867
868 #ifndef PRODUCT
869 void LoadNode::dump_spec(outputStream *st) const {
870 MemNode::dump_spec(st);
871 if( !Verbose && !WizardMode ) {
872 // standard dump does this in Verbose and WizardMode
873 st->print(" #"); _type->dump_on(st);
874 }
875 if (!depends_only_on_test()) {
876 st->print(" (does not depend only on test, ");
877 if (control_dependency() == UnknownControl) {
878 st->print("unknown control");
879 } else if (control_dependency() == Pinned) {
880 st->print("pinned");
881 } else if (adr_type() == TypeRawPtr::BOTTOM) {
882 st->print("raw access");
883 } else {
884 st->print("unknown reason");
885 }
886 st->print(")");
887 }
888 }
889 #endif
890
891 #ifdef ASSERT
892 //----------------------------is_immutable_value-------------------------------
893 // Helper function to allow a raw load without control edge for some cases
894 bool LoadNode::is_immutable_value(Node* adr) {
895 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
896 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
897
898 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
899 int offsets[] = {
900 in_bytes(JavaThread::osthread_offset()),
901 in_bytes(JavaThread::threadObj_offset()),
902 in_bytes(JavaThread::vthread_offset()),
903 in_bytes(JavaThread::scopedValueCache_offset()),
904 };
905
906 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
907 if (offset == offsets[i]) {
908 return true;
909 }
910 }
911 }
912
913 return false;
914 }
915 #endif
916
917 //----------------------------LoadNode::make-----------------------------------
918 // Polymorphic factory method:
919 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
920 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
921 Compile* C = gvn.C;
922
923 // sanity check the alias category against the created node type
924 assert(!(adr_type->isa_oopptr() &&
925 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
926 "use LoadKlassNode instead");
927 assert(!(adr_type->isa_aryptr() &&
928 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
929 "use LoadRangeNode instead");
930 // Check control edge of raw loads
931 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
932 // oop will be recorded in oop map if load crosses safepoint
933 rt->isa_oopptr() || is_immutable_value(adr),
934 "raw memory operations should have control edge");
935 LoadNode* load = nullptr;
936 switch (bt) {
937 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
938 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
939 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
940 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
941 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
942 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
943 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
944 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
945 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
946 case T_OBJECT:
947 case T_NARROWOOP:
948 #ifdef _LP64
949 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
950 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
951 } else
952 #endif
953 {
954 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
955 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
956 }
957 break;
958 default:
959 ShouldNotReachHere();
960 break;
961 }
962 assert(load != nullptr, "LoadNode should have been created");
963 if (unaligned) {
964 load->set_unaligned_access();
965 }
966 if (mismatched) {
967 load->set_mismatched_access();
968 }
969 if (unsafe) {
970 load->set_unsafe_access();
971 }
972 load->set_barrier_data(barrier_data);
973 if (load->Opcode() == Op_LoadN) {
974 Node* ld = gvn.transform(load);
975 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
976 }
977
978 return load;
979 }
980
981 //------------------------------hash-------------------------------------------
982 uint LoadNode::hash() const {
983 // unroll addition of interesting fields
984 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
985 }
986
987 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
988 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
989 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
990 bool is_stable_ary = FoldStableValues &&
991 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
992 tp->isa_aryptr()->is_stable();
993
994 return (eliminate_boxing && non_volatile) || is_stable_ary;
995 }
996
997 return false;
998 }
999
1000 LoadNode* LoadNode::pin_array_access_node() const {
1001 const TypePtr* adr_type = this->adr_type();
1002 if (adr_type != nullptr && adr_type->isa_aryptr()) {
1003 return clone_pinned();
1004 }
1005 return nullptr;
1006 }
1007
1008 // Is the value loaded previously stored by an arraycopy? If so return
1009 // a load node that reads from the source array so we may be able to
1010 // optimize out the ArrayCopy node later.
1011 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1012 Node* ld_adr = in(MemNode::Address);
1013 intptr_t ld_off = 0;
1014 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1015 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1016 if (ac != nullptr) {
1017 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1018
1019 Node* mem = ac->in(TypeFunc::Memory);
1020 Node* ctl = ac->in(0);
1021 Node* src = ac->in(ArrayCopyNode::Src);
1022
1023 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1024 return nullptr;
1025 }
1026
1027 // load depends on the tests that validate the arraycopy
1028 LoadNode* ld = clone_pinned();
1029 Node* addp = in(MemNode::Address)->clone();
1030 if (ac->as_ArrayCopy()->is_clonebasic()) {
1031 assert(ld_alloc != nullptr, "need an alloc");
1032 assert(addp->is_AddP(), "address must be addp");
1033 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1034 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1035 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1036 addp->set_req(AddPNode::Base, src);
1037 addp->set_req(AddPNode::Address, src);
1038 } else {
1039 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1040 ac->as_ArrayCopy()->is_copyof_validated() ||
1041 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1042 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1043 addp->set_req(AddPNode::Base, src);
1044 addp->set_req(AddPNode::Address, src);
1045
1046 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1047 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1048 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1049
1050 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1051 uint shift = exact_log2(type2aelembytes(ary_elem));
1052
1053 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1054 #ifdef _LP64
1055 diff = phase->transform(new ConvI2LNode(diff));
1056 #endif
1057 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1058
1059 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1060 addp->set_req(AddPNode::Offset, offset);
1061 }
1062 addp = phase->transform(addp);
1063 #ifdef ASSERT
1064 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1065 ld->_adr_type = adr_type;
1066 #endif
1067 ld->set_req(MemNode::Address, addp);
1068 ld->set_req(0, ctl);
1069 ld->set_req(MemNode::Memory, mem);
1070 return ld;
1071 }
1072 return nullptr;
1073 }
1074
1075
1076 //---------------------------can_see_stored_value------------------------------
1077 // This routine exists to make sure this set of tests is done the same
1078 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1079 // will change the graph shape in a way which makes memory alive twice at the
1080 // same time (uses the Oracle model of aliasing), then some
1081 // LoadXNode::Identity will fold things back to the equivalence-class model
1082 // of aliasing.
1083 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1084 Node* ld_adr = in(MemNode::Address);
1085 intptr_t ld_off = 0;
1086 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1087 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1088 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1089 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1090 // This is more general than load from boxing objects.
1091 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1092 uint alias_idx = atp->index();
1093 Node* result = nullptr;
1094 Node* current = st;
1095 // Skip through chains of MemBarNodes checking the MergeMems for
1096 // new states for the slice of this load. Stop once any other
1097 // kind of node is encountered. Loads from final memory can skip
1098 // through any kind of MemBar but normal loads shouldn't skip
1099 // through MemBarAcquire since the could allow them to move out of
1100 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1101 // because alias info above them may be inaccurate (e.g., due to
1102 // mixed/mismatched unsafe accesses).
1103 bool is_final_mem = !atp->is_rewritable();
1104 while (current->is_Proj()) {
1105 int opc = current->in(0)->Opcode();
1106 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1107 opc == Op_MemBarAcquireLock ||
1108 opc == Op_LoadFence)) ||
1109 opc == Op_MemBarRelease ||
1110 opc == Op_StoreFence ||
1111 opc == Op_MemBarReleaseLock ||
1112 opc == Op_MemBarStoreStore ||
1113 opc == Op_StoreStoreFence) {
1114 Node* mem = current->in(0)->in(TypeFunc::Memory);
1115 if (mem->is_MergeMem()) {
1116 MergeMemNode* merge = mem->as_MergeMem();
1117 Node* new_st = merge->memory_at(alias_idx);
1118 if (new_st == merge->base_memory()) {
1119 // Keep searching
1120 current = new_st;
1121 continue;
1122 }
1123 // Save the new memory state for the slice and fall through
1124 // to exit.
1125 result = new_st;
1126 }
1127 }
1128 break;
1129 }
1130 if (result != nullptr) {
1131 st = result;
1132 }
1133 }
1134
1135 // Loop around twice in the case Load -> Initialize -> Store.
1136 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1137 for (int trip = 0; trip <= 1; trip++) {
1138
1139 if (st->is_Store()) {
1140 Node* st_adr = st->in(MemNode::Address);
1141 if (st_adr != ld_adr) {
1142 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1143 intptr_t st_off = 0;
1144 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1145 if (ld_base == nullptr) return nullptr;
1146 if (st_base == nullptr) return nullptr;
1147 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1148 if (ld_off != st_off) return nullptr;
1149 if (ld_off == Type::OffsetBot) return nullptr;
1150 // Same base, same offset.
1151 // Possible improvement for arrays: check index value instead of absolute offset.
1152
1153 // At this point we have proven something like this setup:
1154 // B = << base >>
1155 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1156 // S = StoreQ(AddP( B , #Off), V)
1157 // (Actually, we haven't yet proven the Q's are the same.)
1158 // In other words, we are loading from a casted version of
1159 // the same pointer-and-offset that we stored to.
1160 // Casted version may carry a dependency and it is respected.
1161 // Thus, we are able to replace L by V.
1162 }
1163 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1164 if (store_Opcode() != st->Opcode()) {
1165 return nullptr;
1166 }
1167 // LoadVector/StoreVector needs additional check to ensure the types match.
1168 if (st->is_StoreVector()) {
1169 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1170 const TypeVect* out_vt = as_LoadVector()->vect_type();
1171 if (in_vt != out_vt) {
1172 return nullptr;
1173 }
1174 }
1175 return st->in(MemNode::ValueIn);
1176 }
1177
1178 // A load from a freshly-created object always returns zero.
1179 // (This can happen after LoadNode::Ideal resets the load's memory input
1180 // to find_captured_store, which returned InitializeNode::zero_memory.)
1181 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1182 (st->in(0) == ld_alloc) &&
1183 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1184 // return a zero value for the load's basic type
1185 // (This is one of the few places where a generic PhaseTransform
1186 // can create new nodes. Think of it as lazily manifesting
1187 // virtually pre-existing constants.)
1188 if (memory_type() != T_VOID) {
1189 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1190 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1191 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1192 // by the ArrayCopyNode.
1193 return phase->zerocon(memory_type());
1194 }
1195 } else {
1196 // TODO: materialize all-zero vector constant
1197 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1198 }
1199 }
1200
1201 // A load from an initialization barrier can match a captured store.
1202 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1203 InitializeNode* init = st->in(0)->as_Initialize();
1204 AllocateNode* alloc = init->allocation();
1205 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1206 // examine a captured store value
1207 st = init->find_captured_store(ld_off, memory_size(), phase);
1208 if (st != nullptr) {
1209 continue; // take one more trip around
1210 }
1211 }
1212 }
1213
1214 // Load boxed value from result of valueOf() call is input parameter.
1215 if (this->is_Load() && ld_adr->is_AddP() &&
1216 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1217 intptr_t ignore = 0;
1218 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1219 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1220 base = bs->step_over_gc_barrier(base);
1221 if (base != nullptr && base->is_Proj() &&
1222 base->as_Proj()->_con == TypeFunc::Parms &&
1223 base->in(0)->is_CallStaticJava() &&
1224 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1225 return base->in(0)->in(TypeFunc::Parms);
1226 }
1227 }
1228
1229 break;
1230 }
1231
1232 return nullptr;
1233 }
1234
1235 //----------------------is_instance_field_load_with_local_phi------------------
1236 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1237 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1238 in(Address)->is_AddP() ) {
1239 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1240 // Only instances and boxed values.
1241 if( t_oop != nullptr &&
1242 (t_oop->is_ptr_to_boxed_value() ||
1243 t_oop->is_known_instance_field()) &&
1244 t_oop->offset() != Type::OffsetBot &&
1245 t_oop->offset() != Type::OffsetTop) {
1246 return true;
1247 }
1248 }
1249 return false;
1250 }
1251
1252 //------------------------------Identity---------------------------------------
1253 // Loads are identity if previous store is to same address
1254 Node* LoadNode::Identity(PhaseGVN* phase) {
1255 // If the previous store-maker is the right kind of Store, and the store is
1256 // to the same address, then we are equal to the value stored.
1257 Node* mem = in(Memory);
1258 Node* value = can_see_stored_value(mem, phase);
1259 if( value ) {
1260 // byte, short & char stores truncate naturally.
1261 // A load has to load the truncated value which requires
1262 // some sort of masking operation and that requires an
1263 // Ideal call instead of an Identity call.
1264 if (memory_size() < BytesPerInt) {
1265 // If the input to the store does not fit with the load's result type,
1266 // it must be truncated via an Ideal call.
1267 if (!phase->type(value)->higher_equal(phase->type(this)))
1268 return this;
1269 }
1270 // (This works even when value is a Con, but LoadNode::Value
1271 // usually runs first, producing the singleton type of the Con.)
1272 if (!has_pinned_control_dependency() || value->is_Con()) {
1273 return value;
1274 } else {
1275 return this;
1276 }
1277 }
1278
1279 if (has_pinned_control_dependency()) {
1280 return this;
1281 }
1282 // Search for an existing data phi which was generated before for the same
1283 // instance's field to avoid infinite generation of phis in a loop.
1284 Node *region = mem->in(0);
1285 if (is_instance_field_load_with_local_phi(region)) {
1286 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1287 int this_index = phase->C->get_alias_index(addr_t);
1288 int this_offset = addr_t->offset();
1289 int this_iid = addr_t->instance_id();
1290 if (!addr_t->is_known_instance() &&
1291 addr_t->is_ptr_to_boxed_value()) {
1292 // Use _idx of address base (could be Phi node) for boxed values.
1293 intptr_t ignore = 0;
1294 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1295 if (base == nullptr) {
1296 return this;
1297 }
1298 this_iid = base->_idx;
1299 }
1300 const Type* this_type = bottom_type();
1301 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1302 Node* phi = region->fast_out(i);
1303 if (phi->is_Phi() && phi != mem &&
1304 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1305 return phi;
1306 }
1307 }
1308 }
1309
1310 return this;
1311 }
1312
1313 // Construct an equivalent unsigned load.
1314 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1315 BasicType bt = T_ILLEGAL;
1316 const Type* rt = nullptr;
1317 switch (Opcode()) {
1318 case Op_LoadUB: return this;
1319 case Op_LoadUS: return this;
1320 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1321 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1322 default:
1323 assert(false, "no unsigned variant: %s", Name());
1324 return nullptr;
1325 }
1326 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1327 raw_adr_type(), rt, bt, _mo, _control_dependency,
1328 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1329 }
1330
1331 // Construct an equivalent signed load.
1332 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1333 BasicType bt = T_ILLEGAL;
1334 const Type* rt = nullptr;
1335 switch (Opcode()) {
1336 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1337 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1338 case Op_LoadB: // fall through
1339 case Op_LoadS: // fall through
1340 case Op_LoadI: // fall through
1341 case Op_LoadL: return this;
1342 default:
1343 assert(false, "no signed variant: %s", Name());
1344 return nullptr;
1345 }
1346 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1347 raw_adr_type(), rt, bt, _mo, _control_dependency,
1348 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1349 }
1350
1351 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1352 BasicType bt = rt->basic_type();
1353 switch (Opcode()) {
1354 case Op_LoadI: return (bt == T_FLOAT);
1355 case Op_LoadL: return (bt == T_DOUBLE);
1356 case Op_LoadF: return (bt == T_INT);
1357 case Op_LoadD: return (bt == T_LONG);
1358
1359 default: return false;
1360 }
1361 }
1362
1363 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1364 BasicType bt = rt->basic_type();
1365 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1366 bool is_mismatched = is_mismatched_access();
1367 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1368 if (raw_type == nullptr) {
1369 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1370 }
1371 const int op = Opcode();
1372 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1373 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1374 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1375 raw_adr_type(), rt, bt, _mo, _control_dependency,
1376 require_atomic_access, is_unaligned_access(), is_mismatched);
1377 }
1378
1379 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1380 BasicType bt = vt->basic_type();
1381 switch (Opcode()) {
1382 case Op_StoreI: return (bt == T_FLOAT);
1383 case Op_StoreL: return (bt == T_DOUBLE);
1384 case Op_StoreF: return (bt == T_INT);
1385 case Op_StoreD: return (bt == T_LONG);
1386
1387 default: return false;
1388 }
1389 }
1390
1391 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1392 BasicType bt = vt->basic_type();
1393 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1394 const int op = Opcode();
1395 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1396 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1397 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1398 raw_adr_type(), val, bt, _mo, require_atomic_access);
1399
1400 bool is_mismatched = is_mismatched_access();
1401 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1402 if (raw_type == nullptr) {
1403 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1404 }
1405 if (is_mismatched) {
1406 st->set_mismatched_access();
1407 }
1408 return st;
1409 }
1410
1411 // We're loading from an object which has autobox behaviour.
1412 // If this object is result of a valueOf call we'll have a phi
1413 // merging a newly allocated object and a load from the cache.
1414 // We want to replace this load with the original incoming
1415 // argument to the valueOf call.
1416 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1417 assert(igvn->C->eliminate_boxing(), "sanity");
1418 intptr_t ignore = 0;
1419 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1420 if ((base == nullptr) || base->is_Phi()) {
1421 // Push the loads from the phi that comes from valueOf up
1422 // through it to allow elimination of the loads and the recovery
1423 // of the original value. It is done in split_through_phi().
1424 return nullptr;
1425 } else if (base->is_Load() ||
1426 (base->is_DecodeN() && base->in(1)->is_Load())) {
1427 // Eliminate the load of boxed value for integer types from the cache
1428 // array by deriving the value from the index into the array.
1429 // Capture the offset of the load and then reverse the computation.
1430
1431 // Get LoadN node which loads a boxing object from 'cache' array.
1432 if (base->is_DecodeN()) {
1433 base = base->in(1);
1434 }
1435 if (!base->in(Address)->is_AddP()) {
1436 return nullptr; // Complex address
1437 }
1438 AddPNode* address = base->in(Address)->as_AddP();
1439 Node* cache_base = address->in(AddPNode::Base);
1440 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1441 // Get ConP node which is static 'cache' field.
1442 cache_base = cache_base->in(1);
1443 }
1444 if ((cache_base != nullptr) && cache_base->is_Con()) {
1445 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1446 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1447 Node* elements[4];
1448 int shift = exact_log2(type2aelembytes(T_OBJECT));
1449 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1450 if (count > 0 && elements[0]->is_Con() &&
1451 (count == 1 ||
1452 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1453 elements[1]->in(2) == igvn->intcon(shift)))) {
1454 ciObjArray* array = base_type->const_oop()->as_obj_array();
1455 // Fetch the box object cache[0] at the base of the array and get its value
1456 ciInstance* box = array->obj_at(0)->as_instance();
1457 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1458 assert(ik->is_box_klass(), "sanity");
1459 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1460 if (ik->nof_nonstatic_fields() == 1) {
1461 // This should be true nonstatic_field_at requires calling
1462 // nof_nonstatic_fields so check it anyway
1463 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1464 BasicType bt = c.basic_type();
1465 // Only integer types have boxing cache.
1466 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1467 bt == T_BYTE || bt == T_SHORT ||
1468 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1469 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1470 if (cache_low != (int)cache_low) {
1471 return nullptr; // should not happen since cache is array indexed by value
1472 }
1473 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1474 if (offset != (int)offset) {
1475 return nullptr; // should not happen since cache is array indexed by value
1476 }
1477 // Add up all the offsets making of the address of the load
1478 Node* result = elements[0];
1479 for (int i = 1; i < count; i++) {
1480 result = igvn->transform(new AddXNode(result, elements[i]));
1481 }
1482 // Remove the constant offset from the address and then
1483 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1484 // remove the scaling of the offset to recover the original index.
1485 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1486 // Peel the shift off directly but wrap it in a dummy node
1487 // since Ideal can't return existing nodes
1488 igvn->_worklist.push(result); // remove dead node later
1489 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1490 } else if (result->is_Add() && result->in(2)->is_Con() &&
1491 result->in(1)->Opcode() == Op_LShiftX &&
1492 result->in(1)->in(2) == igvn->intcon(shift)) {
1493 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1494 // but for boxing cache access we know that X<<Z will not overflow
1495 // (there is range check) so we do this optimizatrion by hand here.
1496 igvn->_worklist.push(result); // remove dead node later
1497 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1498 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1499 } else {
1500 result = new RShiftXNode(result, igvn->intcon(shift));
1501 }
1502 #ifdef _LP64
1503 if (bt != T_LONG) {
1504 result = new ConvL2INode(igvn->transform(result));
1505 }
1506 #else
1507 if (bt == T_LONG) {
1508 result = new ConvI2LNode(igvn->transform(result));
1509 }
1510 #endif
1511 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1512 // Need to preserve unboxing load type if it is unsigned.
1513 switch(this->Opcode()) {
1514 case Op_LoadUB:
1515 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1516 break;
1517 case Op_LoadUS:
1518 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1519 break;
1520 }
1521 return result;
1522 }
1523 }
1524 }
1525 }
1526 }
1527 return nullptr;
1528 }
1529
1530 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1531 Node* region = phi->in(0);
1532 if (region == nullptr) {
1533 return false; // Wait stable graph
1534 }
1535 uint cnt = phi->req();
1536 for (uint i = 1; i < cnt; i++) {
1537 Node* rc = region->in(i);
1538 if (rc == nullptr || phase->type(rc) == Type::TOP)
1539 return false; // Wait stable graph
1540 Node* in = phi->in(i);
1541 if (in == nullptr || phase->type(in) == Type::TOP)
1542 return false; // Wait stable graph
1543 }
1544 return true;
1545 }
1546
1547 //------------------------------split_through_phi------------------------------
1548 // Check whether a call to 'split_through_phi' would split this load through the
1549 // Phi *base*. This method is essentially a copy of the validations performed
1550 // by 'split_through_phi'. The first use of this method was in EA code as part
1551 // of simplification of allocation merges.
1552 // Some differences from original method (split_through_phi):
1553 // - If base->is_CastPP(): base = base->in(1)
1554 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1555 Node* mem = in(Memory);
1556 Node* address = in(Address);
1557 intptr_t ignore = 0;
1558 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1559
1560 if (base->is_CastPP()) {
1561 base = base->in(1);
1562 }
1563
1564 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1565 return false;
1566 }
1567
1568 if (!mem->is_Phi()) {
1569 if (!MemNode::all_controls_dominate(mem, base->in(0)))
1570 return false;
1571 } else if (base->in(0) != mem->in(0)) {
1572 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1573 return false;
1574 }
1575 }
1576
1577 return true;
1578 }
1579
1580 //------------------------------split_through_phi------------------------------
1581 // Split instance or boxed field load through Phi.
1582 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1583 if (req() > 3) {
1584 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1585 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1586 return nullptr;
1587 }
1588 Node* mem = in(Memory);
1589 Node* address = in(Address);
1590 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1591
1592 assert((t_oop != nullptr) &&
1593 (ignore_missing_instance_id ||
1594 t_oop->is_known_instance_field() ||
1595 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1596
1597 Compile* C = phase->C;
1598 intptr_t ignore = 0;
1599 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1600 bool base_is_phi = (base != nullptr) && base->is_Phi();
1601 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1602 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1603 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1604
1605 if (!((mem->is_Phi() || base_is_phi) &&
1606 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1607 return nullptr; // Neither memory or base are Phi
1608 }
1609
1610 if (mem->is_Phi()) {
1611 if (!stable_phi(mem->as_Phi(), phase)) {
1612 return nullptr; // Wait stable graph
1613 }
1614 uint cnt = mem->req();
1615 // Check for loop invariant memory.
1616 if (cnt == 3) {
1617 for (uint i = 1; i < cnt; i++) {
1618 Node* in = mem->in(i);
1619 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1620 if (m == mem) {
1621 if (i == 1) {
1622 // if the first edge was a loop, check second edge too.
1623 // If both are replaceable - we are in an infinite loop
1624 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1625 if (n == mem) {
1626 break;
1627 }
1628 }
1629 set_req(Memory, mem->in(cnt - i));
1630 return this; // made change
1631 }
1632 }
1633 }
1634 }
1635 if (base_is_phi) {
1636 if (!stable_phi(base->as_Phi(), phase)) {
1637 return nullptr; // Wait stable graph
1638 }
1639 uint cnt = base->req();
1640 // Check for loop invariant memory.
1641 if (cnt == 3) {
1642 for (uint i = 1; i < cnt; i++) {
1643 if (base->in(i) == base) {
1644 return nullptr; // Wait stable graph
1645 }
1646 }
1647 }
1648 }
1649
1650 // Split through Phi (see original code in loopopts.cpp).
1651 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1652
1653 // Do nothing here if Identity will find a value
1654 // (to avoid infinite chain of value phis generation).
1655 if (this != Identity(phase)) {
1656 return nullptr;
1657 }
1658
1659 // Select Region to split through.
1660 Node* region;
1661 if (!base_is_phi) {
1662 assert(mem->is_Phi(), "sanity");
1663 region = mem->in(0);
1664 // Skip if the region dominates some control edge of the address.
1665 if (!MemNode::all_controls_dominate(address, region))
1666 return nullptr;
1667 } else if (!mem->is_Phi()) {
1668 assert(base_is_phi, "sanity");
1669 region = base->in(0);
1670 // Skip if the region dominates some control edge of the memory.
1671 if (!MemNode::all_controls_dominate(mem, region))
1672 return nullptr;
1673 } else if (base->in(0) != mem->in(0)) {
1674 assert(base_is_phi && mem->is_Phi(), "sanity");
1675 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1676 region = base->in(0);
1677 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1678 region = mem->in(0);
1679 } else {
1680 return nullptr; // complex graph
1681 }
1682 } else {
1683 assert(base->in(0) == mem->in(0), "sanity");
1684 region = mem->in(0);
1685 }
1686
1687 Node* phi = nullptr;
1688 const Type* this_type = this->bottom_type();
1689 PhaseIterGVN* igvn = phase->is_IterGVN();
1690 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1691 int this_index = C->get_alias_index(t_oop);
1692 int this_offset = t_oop->offset();
1693 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1694 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1695 } else if (ignore_missing_instance_id) {
1696 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1697 } else {
1698 return nullptr;
1699 }
1700
1701 for (uint i = 1; i < region->req(); i++) {
1702 Node* x;
1703 Node* the_clone = nullptr;
1704 Node* in = region->in(i);
1705 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1706 in != nullptr && in->is_OuterStripMinedLoop()) {
1707 // No node should go in the outer strip mined loop
1708 in = in->in(LoopNode::EntryControl);
1709 }
1710 if (in == nullptr || in == C->top()) {
1711 x = C->top(); // Dead path? Use a dead data op
1712 } else {
1713 x = this->clone(); // Else clone up the data op
1714 the_clone = x; // Remember for possible deletion.
1715 // Alter data node to use pre-phi inputs
1716 if (this->in(0) == region) {
1717 x->set_req(0, in);
1718 } else {
1719 x->set_req(0, nullptr);
1720 }
1721 if (mem->is_Phi() && (mem->in(0) == region)) {
1722 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1723 }
1724 if (address->is_Phi() && address->in(0) == region) {
1725 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1726 }
1727 if (base_is_phi && (base->in(0) == region)) {
1728 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1729 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1730 x->set_req(Address, adr_x);
1731 }
1732 }
1733 // Check for a 'win' on some paths
1734 const Type *t = x->Value(igvn);
1735
1736 bool singleton = t->singleton();
1737
1738 // See comments in PhaseIdealLoop::split_thru_phi().
1739 if (singleton && t == Type::TOP) {
1740 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1741 }
1742
1743 if (singleton) {
1744 x = igvn->makecon(t);
1745 } else {
1746 // We now call Identity to try to simplify the cloned node.
1747 // Note that some Identity methods call phase->type(this).
1748 // Make sure that the type array is big enough for
1749 // our new node, even though we may throw the node away.
1750 // (This tweaking with igvn only works because x is a new node.)
1751 igvn->set_type(x, t);
1752 // If x is a TypeNode, capture any more-precise type permanently into Node
1753 // otherwise it will be not updated during igvn->transform since
1754 // igvn->type(x) is set to x->Value() already.
1755 x->raise_bottom_type(t);
1756 Node* y = x->Identity(igvn);
1757 if (y != x) {
1758 x = y;
1759 } else {
1760 y = igvn->hash_find_insert(x);
1761 if (y) {
1762 x = y;
1763 } else {
1764 // Else x is a new node we are keeping
1765 // We do not need register_new_node_with_optimizer
1766 // because set_type has already been called.
1767 igvn->_worklist.push(x);
1768 }
1769 }
1770 }
1771 if (x != the_clone && the_clone != nullptr) {
1772 igvn->remove_dead_node(the_clone);
1773 }
1774 phi->set_req(i, x);
1775 }
1776 // Record Phi
1777 igvn->register_new_node_with_optimizer(phi);
1778 return phi;
1779 }
1780
1781 AllocateNode* LoadNode::is_new_object_mark_load() const {
1782 if (Opcode() == Op_LoadX) {
1783 Node* address = in(MemNode::Address);
1784 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1785 Node* mem = in(MemNode::Memory);
1786 if (alloc != nullptr && mem->is_Proj() &&
1787 mem->in(0) != nullptr &&
1788 mem->in(0) == alloc->initialization() &&
1789 alloc->initialization()->proj_out_or_null(0) != nullptr) {
1790 return alloc;
1791 }
1792 }
1793 return nullptr;
1794 }
1795
1796
1797 //------------------------------Ideal------------------------------------------
1798 // If the load is from Field memory and the pointer is non-null, it might be possible to
1799 // zero out the control input.
1800 // If the offset is constant and the base is an object allocation,
1801 // try to hook me up to the exact initializing store.
1802 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1803 if (has_pinned_control_dependency()) {
1804 return nullptr;
1805 }
1806 Node* p = MemNode::Ideal_common(phase, can_reshape);
1807 if (p) return (p == NodeSentinel) ? nullptr : p;
1808
1809 Node* ctrl = in(MemNode::Control);
1810 Node* address = in(MemNode::Address);
1811 bool progress = false;
1812
1813 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1814 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1815
1816 // Skip up past a SafePoint control. Cannot do this for Stores because
1817 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1818 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
1819 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1820 !addr_mark &&
1821 (depends_only_on_test() || has_unknown_control_dependency())) {
1822 ctrl = ctrl->in(0);
1823 set_req(MemNode::Control,ctrl);
1824 progress = true;
1825 }
1826
1827 intptr_t ignore = 0;
1828 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1829 if (base != nullptr
1830 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1831 // Check for useless control edge in some common special cases
1832 if (in(MemNode::Control) != nullptr
1833 && can_remove_control()
1834 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1835 && all_controls_dominate(base, phase->C->start())) {
1836 // A method-invariant, non-null address (constant or 'this' argument).
1837 set_req(MemNode::Control, nullptr);
1838 progress = true;
1839 }
1840 }
1841
1842 Node* mem = in(MemNode::Memory);
1843 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1844
1845 if (can_reshape && (addr_t != nullptr)) {
1846 // try to optimize our memory input
1847 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1848 if (opt_mem != mem) {
1849 set_req_X(MemNode::Memory, opt_mem, phase);
1850 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
1851 return this;
1852 }
1853 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1854 if ((t_oop != nullptr) &&
1855 (t_oop->is_known_instance_field() ||
1856 t_oop->is_ptr_to_boxed_value())) {
1857 PhaseIterGVN *igvn = phase->is_IterGVN();
1858 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
1859 if (igvn->_worklist.member(opt_mem)) {
1860 // Delay this transformation until memory Phi is processed.
1861 igvn->_worklist.push(this);
1862 return nullptr;
1863 }
1864 // Split instance field load through Phi.
1865 Node* result = split_through_phi(phase);
1866 if (result != nullptr) return result;
1867
1868 if (t_oop->is_ptr_to_boxed_value()) {
1869 Node* result = eliminate_autobox(igvn);
1870 if (result != nullptr) return result;
1871 }
1872 }
1873 }
1874
1875 // Is there a dominating load that loads the same value? Leave
1876 // anything that is not a load of a field/array element (like
1877 // barriers etc.) alone
1878 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
1879 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1880 Node *use = mem->fast_out(i);
1881 if (use != this &&
1882 use->Opcode() == Opcode() &&
1883 use->in(0) != nullptr &&
1884 use->in(0) != in(0) &&
1885 use->in(Address) == in(Address)) {
1886 Node* ctl = in(0);
1887 for (int i = 0; i < 10 && ctl != nullptr; i++) {
1888 ctl = IfNode::up_one_dom(ctl);
1889 if (ctl == use->in(0)) {
1890 set_req(0, use->in(0));
1891 return this;
1892 }
1893 }
1894 }
1895 }
1896 }
1897
1898 // Check for prior store with a different base or offset; make Load
1899 // independent. Skip through any number of them. Bail out if the stores
1900 // are in an endless dead cycle and report no progress. This is a key
1901 // transform for Reflection. However, if after skipping through the Stores
1902 // we can't then fold up against a prior store do NOT do the transform as
1903 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1904 // array memory alive twice: once for the hoisted Load and again after the
1905 // bypassed Store. This situation only works if EVERYBODY who does
1906 // anti-dependence work knows how to bypass. I.e. we need all
1907 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1908 // the alias index stuff. So instead, peek through Stores and IFF we can
1909 // fold up, do so.
1910 Node* prev_mem = find_previous_store(phase);
1911 if (prev_mem != nullptr) {
1912 Node* value = can_see_arraycopy_value(prev_mem, phase);
1913 if (value != nullptr) {
1914 return value;
1915 }
1916 }
1917 // Steps (a), (b): Walk past independent stores to find an exact match.
1918 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
1919 // (c) See if we can fold up on the spot, but don't fold up here.
1920 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1921 // just return a prior value, which is done by Identity calls.
1922 if (can_see_stored_value(prev_mem, phase)) {
1923 // Make ready for step (d):
1924 set_req_X(MemNode::Memory, prev_mem, phase);
1925 return this;
1926 }
1927 }
1928
1929 return progress ? this : nullptr;
1930 }
1931
1932 // Helper to recognize certain Klass fields which are invariant across
1933 // some group of array types (e.g., int[] or all T[] where T < Object).
1934 const Type*
1935 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1936 ciKlass* klass) const {
1937 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1938 // The field is Klass::_modifier_flags. Return its (constant) value.
1939 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1940 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1941 return TypeInt::make(klass->modifier_flags());
1942 }
1943 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1944 // The field is Klass::_access_flags. Return its (constant) value.
1945 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1946 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1947 return TypeInt::make(klass->access_flags());
1948 }
1949 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1950 // The field is Klass::_layout_helper. Return its constant value if known.
1951 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1952 return TypeInt::make(klass->layout_helper());
1953 }
1954
1955 // No match.
1956 return nullptr;
1957 }
1958
1959 //------------------------------Value-----------------------------------------
1960 const Type* LoadNode::Value(PhaseGVN* phase) const {
1961 // Either input is TOP ==> the result is TOP
1962 Node* mem = in(MemNode::Memory);
1963 const Type *t1 = phase->type(mem);
1964 if (t1 == Type::TOP) return Type::TOP;
1965 Node* adr = in(MemNode::Address);
1966 const TypePtr* tp = phase->type(adr)->isa_ptr();
1967 if (tp == nullptr || tp->empty()) return Type::TOP;
1968 int off = tp->offset();
1969 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1970 Compile* C = phase->C;
1971
1972 // Try to guess loaded type from pointer type
1973 if (tp->isa_aryptr()) {
1974 const TypeAryPtr* ary = tp->is_aryptr();
1975 const Type* t = ary->elem();
1976
1977 // Determine whether the reference is beyond the header or not, by comparing
1978 // the offset against the offset of the start of the array's data.
1979 // Different array types begin at slightly different offsets (12 vs. 16).
1980 // We choose T_BYTE as an example base type that is least restrictive
1981 // as to alignment, which will therefore produce the smallest
1982 // possible base offset.
1983 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1984 const bool off_beyond_header = (off >= min_base_off);
1985
1986 // Try to constant-fold a stable array element.
1987 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1988 // Make sure the reference is not into the header and the offset is constant
1989 ciObject* aobj = ary->const_oop();
1990 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1991 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1992 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1993 stable_dimension,
1994 memory_type(), is_unsigned());
1995 if (con_type != nullptr) {
1996 return con_type;
1997 }
1998 }
1999 }
2000
2001 // Don't do this for integer types. There is only potential profit if
2002 // the element type t is lower than _type; that is, for int types, if _type is
2003 // more restrictive than t. This only happens here if one is short and the other
2004 // char (both 16 bits), and in those cases we've made an intentional decision
2005 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2006 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2007 //
2008 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2009 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2010 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2011 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2012 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2013 // In fact, that could have been the original type of p1, and p1 could have
2014 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2015 // expression (LShiftL quux 3) independently optimized to the constant 8.
2016 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2017 && (_type->isa_vect() == nullptr)
2018 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2019 // t might actually be lower than _type, if _type is a unique
2020 // concrete subclass of abstract class t.
2021 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2022 const Type* jt = t->join_speculative(_type);
2023 // In any case, do not allow the join, per se, to empty out the type.
2024 if (jt->empty() && !t->empty()) {
2025 // This can happen if a interface-typed array narrows to a class type.
2026 jt = _type;
2027 }
2028 #ifdef ASSERT
2029 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2030 // The pointers in the autobox arrays are always non-null
2031 Node* base = adr->in(AddPNode::Base);
2032 if ((base != nullptr) && base->is_DecodeN()) {
2033 // Get LoadN node which loads IntegerCache.cache field
2034 base = base->in(1);
2035 }
2036 if ((base != nullptr) && base->is_Con()) {
2037 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2038 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2039 // It could be narrow oop
2040 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2041 }
2042 }
2043 }
2044 #endif
2045 return jt;
2046 }
2047 }
2048 } else if (tp->base() == Type::InstPtr) {
2049 assert( off != Type::OffsetBot ||
2050 // arrays can be cast to Objects
2051 !tp->isa_instptr() ||
2052 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2053 // unsafe field access may not have a constant offset
2054 C->has_unsafe_access(),
2055 "Field accesses must be precise" );
2056 // For oop loads, we expect the _type to be precise.
2057
2058 // Optimize loads from constant fields.
2059 const TypeInstPtr* tinst = tp->is_instptr();
2060 ciObject* const_oop = tinst->const_oop();
2061 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2062 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
2063 if (con_type != nullptr) {
2064 return con_type;
2065 }
2066 }
2067 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2068 assert(off != Type::OffsetBot ||
2069 !tp->isa_instklassptr() ||
2070 // arrays can be cast to Objects
2071 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2072 // also allow array-loading from the primary supertype
2073 // array during subtype checks
2074 Opcode() == Op_LoadKlass,
2075 "Field accesses must be precise");
2076 // For klass/static loads, we expect the _type to be precise
2077 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2078 /* With mirrors being an indirect in the Klass*
2079 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2080 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2081 *
2082 * So check the type and klass of the node before the LoadP.
2083 */
2084 Node* adr2 = adr->in(MemNode::Address);
2085 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2086 if (tkls != nullptr && !StressReflectiveCode) {
2087 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2088 ciKlass* klass = tkls->exact_klass();
2089 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2090 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2091 return TypeInstPtr::make(klass->java_mirror());
2092 }
2093 }
2094 }
2095
2096 const TypeKlassPtr *tkls = tp->isa_klassptr();
2097 if (tkls != nullptr) {
2098 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2099 ciKlass* klass = tkls->exact_klass();
2100 // We are loading a field from a Klass metaobject whose identity
2101 // is known at compile time (the type is "exact" or "precise").
2102 // Check for fields we know are maintained as constants by the VM.
2103 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2104 // The field is Klass::_super_check_offset. Return its (constant) value.
2105 // (Folds up type checking code.)
2106 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2107 return TypeInt::make(klass->super_check_offset());
2108 }
2109 // Compute index into primary_supers array
2110 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2111 // Check for overflowing; use unsigned compare to handle the negative case.
2112 if( depth < ciKlass::primary_super_limit() ) {
2113 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2114 // (Folds up type checking code.)
2115 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2116 ciKlass *ss = klass->super_of_depth(depth);
2117 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2118 }
2119 const Type* aift = load_array_final_field(tkls, klass);
2120 if (aift != nullptr) return aift;
2121 }
2122
2123 // We can still check if we are loading from the primary_supers array at a
2124 // shallow enough depth. Even though the klass is not exact, entries less
2125 // than or equal to its super depth are correct.
2126 if (tkls->is_loaded()) {
2127 ciKlass* klass = nullptr;
2128 if (tkls->isa_instklassptr()) {
2129 klass = tkls->is_instklassptr()->instance_klass();
2130 } else {
2131 int dims;
2132 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2133 if (inner->isa_instklassptr()) {
2134 klass = inner->is_instklassptr()->instance_klass();
2135 klass = ciObjArrayKlass::make(klass, dims);
2136 }
2137 }
2138 if (klass != nullptr) {
2139 // Compute index into primary_supers array
2140 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2141 // Check for overflowing; use unsigned compare to handle the negative case.
2142 if (depth < ciKlass::primary_super_limit() &&
2143 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2144 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2145 // (Folds up type checking code.)
2146 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2147 ciKlass *ss = klass->super_of_depth(depth);
2148 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2149 }
2150 }
2151 }
2152
2153 // If the type is enough to determine that the thing is not an array,
2154 // we can give the layout_helper a positive interval type.
2155 // This will help short-circuit some reflective code.
2156 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2157 tkls->isa_instklassptr() && // not directly typed as an array
2158 !tkls->is_instklassptr()->instance_klass()->is_java_lang_Object() // not the supertype of all T[] and specifically not Serializable & Cloneable
2159 ) {
2160 // Note: When interfaces are reliable, we can narrow the interface
2161 // test to (klass != Serializable && klass != Cloneable).
2162 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2163 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2164 // The key property of this type is that it folds up tests
2165 // for array-ness, since it proves that the layout_helper is positive.
2166 // Thus, a generic value like the basic object layout helper works fine.
2167 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2168 }
2169 }
2170
2171 // If we are loading from a freshly-allocated object, produce a zero,
2172 // if the load is provably beyond the header of the object.
2173 // (Also allow a variable load from a fresh array to produce zero.)
2174 const TypeOopPtr *tinst = tp->isa_oopptr();
2175 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field();
2176 bool is_boxed_value = (tinst != nullptr) && tinst->is_ptr_to_boxed_value();
2177 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2178 Node* value = can_see_stored_value(mem,phase);
2179 if (value != nullptr && value->is_Con()) {
2180 assert(value->bottom_type()->higher_equal(_type),"sanity");
2181 return value->bottom_type();
2182 }
2183 }
2184
2185 bool is_vect = (_type->isa_vect() != nullptr);
2186 if (is_instance && !is_vect) {
2187 // If we have an instance type and our memory input is the
2188 // programs's initial memory state, there is no matching store,
2189 // so just return a zero of the appropriate type -
2190 // except if it is vectorized - then we have no zero constant.
2191 Node *mem = in(MemNode::Memory);
2192 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2193 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2194 return Type::get_zero_type(_type->basic_type());
2195 }
2196 }
2197
2198 Node* alloc = is_new_object_mark_load();
2199 if (alloc != nullptr) {
2200 return TypeX::make(markWord::prototype().value());
2201 }
2202
2203 return _type;
2204 }
2205
2206 //------------------------------match_edge-------------------------------------
2207 // Do we Match on this edge index or not? Match only the address.
2208 uint LoadNode::match_edge(uint idx) const {
2209 return idx == MemNode::Address;
2210 }
2211
2212 //--------------------------LoadBNode::Ideal--------------------------------------
2213 //
2214 // If the previous store is to the same address as this load,
2215 // and the value stored was larger than a byte, replace this load
2216 // with the value stored truncated to a byte. If no truncation is
2217 // needed, the replacement is done in LoadNode::Identity().
2218 //
2219 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2220 Node* mem = in(MemNode::Memory);
2221 Node* value = can_see_stored_value(mem,phase);
2222 if (value != nullptr) {
2223 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2224 if (narrow != value) {
2225 return narrow;
2226 }
2227 }
2228 // Identity call will handle the case where truncation is not needed.
2229 return LoadNode::Ideal(phase, can_reshape);
2230 }
2231
2232 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2233 Node* mem = in(MemNode::Memory);
2234 Node* value = can_see_stored_value(mem,phase);
2235 if (value != nullptr && value->is_Con() &&
2236 !value->bottom_type()->higher_equal(_type)) {
2237 // If the input to the store does not fit with the load's result type,
2238 // it must be truncated. We can't delay until Ideal call since
2239 // a singleton Value is needed for split_thru_phi optimization.
2240 int con = value->get_int();
2241 return TypeInt::make((con << 24) >> 24);
2242 }
2243 return LoadNode::Value(phase);
2244 }
2245
2246 //--------------------------LoadUBNode::Ideal-------------------------------------
2247 //
2248 // If the previous store is to the same address as this load,
2249 // and the value stored was larger than a byte, replace this load
2250 // with the value stored truncated to a byte. If no truncation is
2251 // needed, the replacement is done in LoadNode::Identity().
2252 //
2253 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2254 Node* mem = in(MemNode::Memory);
2255 Node* value = can_see_stored_value(mem, phase);
2256 if (value != nullptr) {
2257 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2258 if (narrow != value) {
2259 return narrow;
2260 }
2261 }
2262 // Identity call will handle the case where truncation is not needed.
2263 return LoadNode::Ideal(phase, can_reshape);
2264 }
2265
2266 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2267 Node* mem = in(MemNode::Memory);
2268 Node* value = can_see_stored_value(mem,phase);
2269 if (value != nullptr && value->is_Con() &&
2270 !value->bottom_type()->higher_equal(_type)) {
2271 // If the input to the store does not fit with the load's result type,
2272 // it must be truncated. We can't delay until Ideal call since
2273 // a singleton Value is needed for split_thru_phi optimization.
2274 int con = value->get_int();
2275 return TypeInt::make(con & 0xFF);
2276 }
2277 return LoadNode::Value(phase);
2278 }
2279
2280 //--------------------------LoadUSNode::Ideal-------------------------------------
2281 //
2282 // If the previous store is to the same address as this load,
2283 // and the value stored was larger than a char, replace this load
2284 // with the value stored truncated to a char. If no truncation is
2285 // needed, the replacement is done in LoadNode::Identity().
2286 //
2287 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2288 Node* mem = in(MemNode::Memory);
2289 Node* value = can_see_stored_value(mem,phase);
2290 if (value != nullptr) {
2291 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2292 if (narrow != value) {
2293 return narrow;
2294 }
2295 }
2296 // Identity call will handle the case where truncation is not needed.
2297 return LoadNode::Ideal(phase, can_reshape);
2298 }
2299
2300 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2301 Node* mem = in(MemNode::Memory);
2302 Node* value = can_see_stored_value(mem,phase);
2303 if (value != nullptr && value->is_Con() &&
2304 !value->bottom_type()->higher_equal(_type)) {
2305 // If the input to the store does not fit with the load's result type,
2306 // it must be truncated. We can't delay until Ideal call since
2307 // a singleton Value is needed for split_thru_phi optimization.
2308 int con = value->get_int();
2309 return TypeInt::make(con & 0xFFFF);
2310 }
2311 return LoadNode::Value(phase);
2312 }
2313
2314 //--------------------------LoadSNode::Ideal--------------------------------------
2315 //
2316 // If the previous store is to the same address as this load,
2317 // and the value stored was larger than a short, replace this load
2318 // with the value stored truncated to a short. If no truncation is
2319 // needed, the replacement is done in LoadNode::Identity().
2320 //
2321 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2322 Node* mem = in(MemNode::Memory);
2323 Node* value = can_see_stored_value(mem,phase);
2324 if (value != nullptr) {
2325 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2326 if (narrow != value) {
2327 return narrow;
2328 }
2329 }
2330 // Identity call will handle the case where truncation is not needed.
2331 return LoadNode::Ideal(phase, can_reshape);
2332 }
2333
2334 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2335 Node* mem = in(MemNode::Memory);
2336 Node* value = can_see_stored_value(mem,phase);
2337 if (value != nullptr && value->is_Con() &&
2338 !value->bottom_type()->higher_equal(_type)) {
2339 // If the input to the store does not fit with the load's result type,
2340 // it must be truncated. We can't delay until Ideal call since
2341 // a singleton Value is needed for split_thru_phi optimization.
2342 int con = value->get_int();
2343 return TypeInt::make((con << 16) >> 16);
2344 }
2345 return LoadNode::Value(phase);
2346 }
2347
2348 //=============================================================================
2349 //----------------------------LoadKlassNode::make------------------------------
2350 // Polymorphic factory method:
2351 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2352 // sanity check the alias category against the created node type
2353 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2354 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2355 #ifdef _LP64
2356 if (adr_type->is_ptr_to_narrowklass()) {
2357 assert(UseCompressedClassPointers, "no compressed klasses");
2358 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2359 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2360 }
2361 #endif
2362 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2363 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2364 }
2365
2366 //------------------------------Value------------------------------------------
2367 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2368 return klass_value_common(phase);
2369 }
2370
2371 // In most cases, LoadKlassNode does not have the control input set. If the control
2372 // input is set, it must not be removed (by LoadNode::Ideal()).
2373 bool LoadKlassNode::can_remove_control() const {
2374 return false;
2375 }
2376
2377 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2378 // Either input is TOP ==> the result is TOP
2379 const Type *t1 = phase->type( in(MemNode::Memory) );
2380 if (t1 == Type::TOP) return Type::TOP;
2381 Node *adr = in(MemNode::Address);
2382 const Type *t2 = phase->type( adr );
2383 if (t2 == Type::TOP) return Type::TOP;
2384 const TypePtr *tp = t2->is_ptr();
2385 if (TypePtr::above_centerline(tp->ptr()) ||
2386 tp->ptr() == TypePtr::Null) return Type::TOP;
2387
2388 // Return a more precise klass, if possible
2389 const TypeInstPtr *tinst = tp->isa_instptr();
2390 if (tinst != nullptr) {
2391 ciInstanceKlass* ik = tinst->instance_klass();
2392 int offset = tinst->offset();
2393 if (ik == phase->C->env()->Class_klass()
2394 && (offset == java_lang_Class::klass_offset() ||
2395 offset == java_lang_Class::array_klass_offset())) {
2396 // We are loading a special hidden field from a Class mirror object,
2397 // the field which points to the VM's Klass metaobject.
2398 ciType* t = tinst->java_mirror_type();
2399 // java_mirror_type returns non-null for compile-time Class constants.
2400 if (t != nullptr) {
2401 // constant oop => constant klass
2402 if (offset == java_lang_Class::array_klass_offset()) {
2403 if (t->is_void()) {
2404 // We cannot create a void array. Since void is a primitive type return null
2405 // klass. Users of this result need to do a null check on the returned klass.
2406 return TypePtr::NULL_PTR;
2407 }
2408 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2409 }
2410 if (!t->is_klass()) {
2411 // a primitive Class (e.g., int.class) has null for a klass field
2412 return TypePtr::NULL_PTR;
2413 }
2414 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2415 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2416 }
2417 // non-constant mirror, so we can't tell what's going on
2418 }
2419 if (!tinst->is_loaded())
2420 return _type; // Bail out if not loaded
2421 if (offset == oopDesc::klass_offset_in_bytes()) {
2422 return tinst->as_klass_type(true);
2423 }
2424 }
2425
2426 // Check for loading klass from an array
2427 const TypeAryPtr *tary = tp->isa_aryptr();
2428 if (tary != nullptr &&
2429 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2430 return tary->as_klass_type(true);
2431 }
2432
2433 // Check for loading klass from an array klass
2434 const TypeKlassPtr *tkls = tp->isa_klassptr();
2435 if (tkls != nullptr && !StressReflectiveCode) {
2436 if (!tkls->is_loaded())
2437 return _type; // Bail out if not loaded
2438 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2439 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2440 // // Always returning precise element type is incorrect,
2441 // // e.g., element type could be object and array may contain strings
2442 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2443
2444 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2445 // according to the element type's subclassing.
2446 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2447 }
2448 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2449 tkls->offset() == in_bytes(Klass::super_offset())) {
2450 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2451 // The field is Klass::_super. Return its (constant) value.
2452 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2453 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2454 }
2455 }
2456
2457 if (tkls != nullptr && !UseSecondarySupersCache
2458 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2459 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2460 return TypePtr::NULL_PTR;
2461 }
2462
2463 // Bailout case
2464 return LoadNode::Value(phase);
2465 }
2466
2467 //------------------------------Identity---------------------------------------
2468 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2469 // Also feed through the klass in Allocate(...klass...)._klass.
2470 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2471 return klass_identity_common(phase);
2472 }
2473
2474 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2475 Node* x = LoadNode::Identity(phase);
2476 if (x != this) return x;
2477
2478 // Take apart the address into an oop and offset.
2479 // Return 'this' if we cannot.
2480 Node* adr = in(MemNode::Address);
2481 intptr_t offset = 0;
2482 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2483 if (base == nullptr) return this;
2484 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2485 if (toop == nullptr) return this;
2486
2487 // Step over potential GC barrier for OopHandle resolve
2488 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2489 if (bs->is_gc_barrier_node(base)) {
2490 base = bs->step_over_gc_barrier(base);
2491 }
2492
2493 // We can fetch the klass directly through an AllocateNode.
2494 // This works even if the klass is not constant (clone or newArray).
2495 if (offset == oopDesc::klass_offset_in_bytes()) {
2496 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2497 if (allocated_klass != nullptr) {
2498 return allocated_klass;
2499 }
2500 }
2501
2502 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2503 // See inline_native_Class_query for occurrences of these patterns.
2504 // Java Example: x.getClass().isAssignableFrom(y)
2505 //
2506 // This improves reflective code, often making the Class
2507 // mirror go completely dead. (Current exception: Class
2508 // mirrors may appear in debug info, but we could clean them out by
2509 // introducing a new debug info operator for Klass.java_mirror).
2510
2511 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2512 && offset == java_lang_Class::klass_offset()) {
2513 if (base->is_Load()) {
2514 Node* base2 = base->in(MemNode::Address);
2515 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2516 Node* adr2 = base2->in(MemNode::Address);
2517 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2518 if (tkls != nullptr && !tkls->empty()
2519 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2520 && adr2->is_AddP()
2521 ) {
2522 int mirror_field = in_bytes(Klass::java_mirror_offset());
2523 if (tkls->offset() == mirror_field) {
2524 return adr2->in(AddPNode::Base);
2525 }
2526 }
2527 }
2528 }
2529 }
2530
2531 return this;
2532 }
2533
2534 LoadNode* LoadNode::clone_pinned() const {
2535 LoadNode* ld = clone()->as_Load();
2536 ld->_control_dependency = UnknownControl;
2537 return ld;
2538 }
2539
2540
2541 //------------------------------Value------------------------------------------
2542 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2543 const Type *t = klass_value_common(phase);
2544 if (t == Type::TOP)
2545 return t;
2546
2547 return t->make_narrowklass();
2548 }
2549
2550 //------------------------------Identity---------------------------------------
2551 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2552 // Also feed through the klass in Allocate(...klass...)._klass.
2553 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2554 Node *x = klass_identity_common(phase);
2555
2556 const Type *t = phase->type( x );
2557 if( t == Type::TOP ) return x;
2558 if( t->isa_narrowklass()) return x;
2559 assert (!t->isa_narrowoop(), "no narrow oop here");
2560
2561 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2562 }
2563
2564 //------------------------------Value-----------------------------------------
2565 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2566 // Either input is TOP ==> the result is TOP
2567 const Type *t1 = phase->type( in(MemNode::Memory) );
2568 if( t1 == Type::TOP ) return Type::TOP;
2569 Node *adr = in(MemNode::Address);
2570 const Type *t2 = phase->type( adr );
2571 if( t2 == Type::TOP ) return Type::TOP;
2572 const TypePtr *tp = t2->is_ptr();
2573 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2574 const TypeAryPtr *tap = tp->isa_aryptr();
2575 if( !tap ) return _type;
2576 return tap->size();
2577 }
2578
2579 //-------------------------------Ideal---------------------------------------
2580 // Feed through the length in AllocateArray(...length...)._length.
2581 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2582 Node* p = MemNode::Ideal_common(phase, can_reshape);
2583 if (p) return (p == NodeSentinel) ? nullptr : p;
2584
2585 // Take apart the address into an oop and offset.
2586 // Return 'this' if we cannot.
2587 Node* adr = in(MemNode::Address);
2588 intptr_t offset = 0;
2589 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2590 if (base == nullptr) return nullptr;
2591 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2592 if (tary == nullptr) return nullptr;
2593
2594 // We can fetch the length directly through an AllocateArrayNode.
2595 // This works even if the length is not constant (clone or newArray).
2596 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2597 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2598 if (alloc != nullptr) {
2599 Node* allocated_length = alloc->Ideal_length();
2600 Node* len = alloc->make_ideal_length(tary, phase);
2601 if (allocated_length != len) {
2602 // New CastII improves on this.
2603 return len;
2604 }
2605 }
2606 }
2607
2608 return nullptr;
2609 }
2610
2611 //------------------------------Identity---------------------------------------
2612 // Feed through the length in AllocateArray(...length...)._length.
2613 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2614 Node* x = LoadINode::Identity(phase);
2615 if (x != this) return x;
2616
2617 // Take apart the address into an oop and offset.
2618 // Return 'this' if we cannot.
2619 Node* adr = in(MemNode::Address);
2620 intptr_t offset = 0;
2621 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2622 if (base == nullptr) return this;
2623 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2624 if (tary == nullptr) return this;
2625
2626 // We can fetch the length directly through an AllocateArrayNode.
2627 // This works even if the length is not constant (clone or newArray).
2628 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2629 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2630 if (alloc != nullptr) {
2631 Node* allocated_length = alloc->Ideal_length();
2632 // Do not allow make_ideal_length to allocate a CastII node.
2633 Node* len = alloc->make_ideal_length(tary, phase, false);
2634 if (allocated_length == len) {
2635 // Return allocated_length only if it would not be improved by a CastII.
2636 return allocated_length;
2637 }
2638 }
2639 }
2640
2641 return this;
2642
2643 }
2644
2645 //=============================================================================
2646 //---------------------------StoreNode::make-----------------------------------
2647 // Polymorphic factory method:
2648 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) {
2649 assert((mo == unordered || mo == release), "unexpected");
2650 Compile* C = gvn.C;
2651 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2652 ctl != nullptr, "raw memory operations should have control edge");
2653
2654 switch (bt) {
2655 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2656 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2657 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2658 case T_CHAR:
2659 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2660 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2661 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2662 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2663 case T_METADATA:
2664 case T_ADDRESS:
2665 case T_OBJECT:
2666 #ifdef _LP64
2667 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2668 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2669 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2670 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2671 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2672 adr->bottom_type()->isa_rawptr())) {
2673 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2674 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2675 }
2676 #endif
2677 {
2678 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2679 }
2680 default:
2681 ShouldNotReachHere();
2682 return (StoreNode*)nullptr;
2683 }
2684 }
2685
2686 //--------------------------bottom_type----------------------------------------
2687 const Type *StoreNode::bottom_type() const {
2688 return Type::MEMORY;
2689 }
2690
2691 //------------------------------hash-------------------------------------------
2692 uint StoreNode::hash() const {
2693 // unroll addition of interesting fields
2694 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2695
2696 // Since they are not commoned, do not hash them:
2697 return NO_HASH;
2698 }
2699
2700 // Class to parse array pointers, and determine if they are adjacent. We parse the form:
2701 //
2702 // pointer = base
2703 // + constant_offset
2704 // + LShiftL( ConvI2L(int_offset + int_con), int_offset_shift)
2705 // + sum(other_offsets)
2706 //
2707 //
2708 // Note: we accumulate all constant offsets into constant_offset, even the int constant behind
2709 // the "LShiftL(ConvI2L(...))" pattern. We convert "ConvI2L(int_offset + int_con)" to
2710 // "ConvI2L(int_offset) + int_con", which is only safe if we can assume that either all
2711 // compared addresses have an overflow for "int_offset + int_con" or none.
2712 // For loads and stores on arrays, we know that if one overflows and the other not, then
2713 // the two addresses lay almost max_int indices apart, but the maximal array size is
2714 // only about half of that. Therefore, the RangeCheck on at least one of them must have
2715 // failed.
2716 //
2717 // constant_offset += LShiftL( ConvI2L(int_con), int_offset_shift)
2718 //
2719 // pointer = base
2720 // + constant_offset
2721 // + LShiftL( ConvI2L(int_offset), int_offset_shift)
2722 // + sum(other_offsets)
2723 //
2724 class ArrayPointer {
2725 private:
2726 const bool _is_valid; // The parsing succeeded
2727 const Node* _pointer; // The final pointer to the position in the array
2728 const Node* _base; // Base address of the array
2729 const jlong _constant_offset; // Sum of collected constant offsets
2730 const Node* _int_offset; // (optional) Offset behind LShiftL and ConvI2L
2731 const jint _int_offset_shift; // (optional) Shift value for int_offset
2732 const GrowableArray<Node*>* _other_offsets; // List of other AddP offsets
2733
2734 ArrayPointer(const bool is_valid,
2735 const Node* pointer,
2736 const Node* base,
2737 const jlong constant_offset,
2738 const Node* int_offset,
2739 const jint int_offset_shift,
2740 const GrowableArray<Node*>* other_offsets) :
2741 _is_valid(is_valid),
2742 _pointer(pointer),
2743 _base(base),
2744 _constant_offset(constant_offset),
2745 _int_offset(int_offset),
2746 _int_offset_shift(int_offset_shift),
2747 _other_offsets(other_offsets)
2748 {
2749 assert(_pointer != nullptr, "must always have pointer");
2750 assert(is_valid == (_base != nullptr), "have base exactly if valid");
2751 assert(is_valid == (_other_offsets != nullptr), "have other_offsets exactly if valid");
2752 }
2753
2754 static ArrayPointer make_invalid(const Node* pointer) {
2755 return ArrayPointer(false, pointer, nullptr, 0, nullptr, 0, nullptr);
2756 }
2757
2758 static bool parse_int_offset(Node* offset, Node*& int_offset, jint& int_offset_shift) {
2759 // offset = LShiftL( ConvI2L(int_offset), int_offset_shift)
2760 if (offset->Opcode() == Op_LShiftL &&
2761 offset->in(1)->Opcode() == Op_ConvI2L &&
2762 offset->in(2)->Opcode() == Op_ConI) {
2763 int_offset = offset->in(1)->in(1); // LShiftL -> ConvI2L -> int_offset
2764 int_offset_shift = offset->in(2)->get_int(); // LShiftL -> int_offset_shift
2765 return true;
2766 }
2767
2768 // offset = ConvI2L(int_offset) = LShiftL( ConvI2L(int_offset), 0)
2769 if (offset->Opcode() == Op_ConvI2L) {
2770 int_offset = offset->in(1);
2771 int_offset_shift = 0;
2772 return true;
2773 }
2774
2775 // parse failed
2776 return false;
2777 }
2778
2779 public:
2780 // Parse the structure above the pointer
2781 static ArrayPointer make(PhaseGVN* phase, const Node* pointer) {
2782 assert(phase->type(pointer)->isa_aryptr() != nullptr, "must be array pointer");
2783 if (!pointer->is_AddP()) { return ArrayPointer::make_invalid(pointer); }
2784
2785 const Node* base = pointer->in(AddPNode::Base);
2786 if (base == nullptr) { return ArrayPointer::make_invalid(pointer); }
2787
2788 const int search_depth = 5;
2789 Node* offsets[search_depth];
2790 int count = pointer->as_AddP()->unpack_offsets(offsets, search_depth);
2791
2792 // We expect at least a constant each
2793 if (count <= 0) { return ArrayPointer::make_invalid(pointer); }
2794
2795 // We extract the form:
2796 //
2797 // pointer = base
2798 // + constant_offset
2799 // + LShiftL( ConvI2L(int_offset + int_con), int_offset_shift)
2800 // + sum(other_offsets)
2801 //
2802 jlong constant_offset = 0;
2803 Node* int_offset = nullptr;
2804 jint int_offset_shift = 0;
2805 GrowableArray<Node*>* other_offsets = new GrowableArray<Node*>(count);
2806
2807 for (int i = 0; i < count; i++) {
2808 Node* offset = offsets[i];
2809 if (offset->Opcode() == Op_ConI) {
2810 // Constant int offset
2811 constant_offset += offset->get_int();
2812 } else if (offset->Opcode() == Op_ConL) {
2813 // Constant long offset
2814 constant_offset += offset->get_long();
2815 } else if(int_offset == nullptr && parse_int_offset(offset, int_offset, int_offset_shift)) {
2816 // LShiftL( ConvI2L(int_offset), int_offset_shift)
2817 int_offset = int_offset->uncast();
2818 if (int_offset->Opcode() == Op_AddI && int_offset->in(2)->Opcode() == Op_ConI) {
2819 // LShiftL( ConvI2L(int_offset + int_con), int_offset_shift)
2820 constant_offset += ((jlong)int_offset->in(2)->get_int()) << int_offset_shift;
2821 int_offset = int_offset->in(1);
2822 }
2823 } else {
2824 // All others
2825 other_offsets->append(offset);
2826 }
2827 }
2828
2829 return ArrayPointer(true, pointer, base, constant_offset, int_offset, int_offset_shift, other_offsets);
2830 }
2831
2832 bool is_adjacent_to_and_before(const ArrayPointer& other, const jlong data_size) const {
2833 if (!_is_valid || !other._is_valid) { return false; }
2834
2835 // Offset adjacent?
2836 if (this->_constant_offset + data_size != other._constant_offset) { return false; }
2837
2838 // All other components identical?
2839 if (this->_base != other._base ||
2840 this->_int_offset != other._int_offset ||
2841 this->_int_offset_shift != other._int_offset_shift ||
2842 this->_other_offsets->length() != other._other_offsets->length()) {
2843 return false;
2844 }
2845
2846 for (int i = 0; i < this->_other_offsets->length(); i++) {
2847 Node* o1 = this->_other_offsets->at(i);
2848 Node* o2 = other._other_offsets->at(i);
2849 if (o1 != o2) { return false; }
2850 }
2851
2852 return true;
2853 }
2854
2855 #ifndef PRODUCT
2856 void dump() {
2857 if (!_is_valid) {
2858 tty->print("ArrayPointer[%d %s, invalid]", _pointer->_idx, _pointer->Name());
2859 return;
2860 }
2861 tty->print("ArrayPointer[%d %s, base[%d %s] + %lld",
2862 _pointer->_idx, _pointer->Name(),
2863 _base->_idx, _base->Name(),
2864 (long long)_constant_offset);
2865 if (_int_offset != 0) {
2866 tty->print(" + I2L[%d %s] << %d",
2867 _int_offset->_idx, _int_offset->Name(), _int_offset_shift);
2868 }
2869 for (int i = 0; i < _other_offsets->length(); i++) {
2870 Node* n = _other_offsets->at(i);
2871 tty->print(" + [%d %s]", n->_idx, n->Name());
2872 }
2873 tty->print_cr("]");
2874 }
2875 #endif
2876 };
2877
2878 // Link together multiple stores (B/S/C/I) into a longer one.
2879 //
2880 // Example: _store = StoreB[i+3]
2881 //
2882 // RangeCheck[i+0] RangeCheck[i+0]
2883 // StoreB[i+0]
2884 // RangeCheck[i+3] RangeCheck[i+3]
2885 // StoreB[i+1] --> pass: fail:
2886 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2887 // StoreB[i+3]
2888 //
2889 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2890 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2891 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2892 // failing path.
2893 //
2894 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2895 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2896 // and possibly no RangeChecks remain between the stores.
2897 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2898 // and those later RangeChecks can be removed. Example:
2899 //
2900 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2901 // StoreB[i+0] StoreB[i+0] <- first store
2902 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2903 // StoreB[i+1] StoreB[i+1] <- second store
2904 // RangeCheck[i+2] --> removed
2905 // StoreB[i+2] StoreB[i+2]
2906 // RangeCheck[i+3] --> removed
2907 // StoreB[i+3] StoreB[i+3] <- last store
2908 //
2909 // Thus, it is a common pattern that between the first and last store in a chain
2910 // of adjacent stores there remains exactly one RangeCheck, located between the
2911 // first and the second store (e.g. RangeCheck[i+3]).
2912 //
2913 class MergePrimitiveArrayStores : public StackObj {
2914 private:
2915 PhaseGVN* _phase;
2916 StoreNode* _store;
2917
2918 public:
2919 MergePrimitiveArrayStores(PhaseGVN* phase, StoreNode* store) : _phase(phase), _store(store) {}
2920
2921 StoreNode* run();
2922
2923 private:
2924 bool is_compatible_store(const StoreNode* other_store) const;
2925 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
2926 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
2927 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out);
2928 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
2929 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
2930
2931 class Status {
2932 private:
2933 StoreNode* _found_store;
2934 bool _found_range_check;
2935
2936 Status(StoreNode* found_store, bool found_range_check)
2937 : _found_store(found_store), _found_range_check(found_range_check) {}
2938
2939 public:
2940 StoreNode* found_store() const { return _found_store; }
2941 bool found_range_check() const { return _found_range_check; }
2942 static Status make_failure() { return Status(nullptr, false); }
2943
2944 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
2945 if (cfg_status == CFGStatus::Failure) {
2946 return Status::make_failure();
2947 }
2948 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
2949 }
2950 };
2951
2952 Status find_adjacent_use_store(const StoreNode* def_store) const;
2953 Status find_adjacent_def_store(const StoreNode* use_store) const;
2954 Status find_use_store(const StoreNode* def_store) const;
2955 Status find_def_store(const StoreNode* use_store) const;
2956 Status find_use_store_unidirectional(const StoreNode* def_store) const;
2957 Status find_def_store_unidirectional(const StoreNode* use_store) const;
2958
2959 void collect_merge_list(Node_List& merge_list) const;
2960 Node* make_merged_input_value(const Node_List& merge_list);
2961 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
2962
2963 DEBUG_ONLY( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
2964 };
2965
2966 StoreNode* MergePrimitiveArrayStores::run() {
2967 // Check for B/S/C/I
2968 int opc = _store->Opcode();
2969 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
2970 return nullptr;
2971 }
2972
2973 // Only merge stores on arrays, and the stores must have the same size as the elements.
2974 const TypePtr* ptr_t = _store->adr_type();
2975 if (ptr_t == nullptr) {
2976 return nullptr;
2977 }
2978 const TypeAryPtr* aryptr_t = ptr_t->isa_aryptr();
2979 if (aryptr_t == nullptr) {
2980 return nullptr;
2981 }
2982 BasicType bt = aryptr_t->elem()->array_element_basic_type();
2983 if (!is_java_primitive(bt) ||
2984 type2aelembytes(bt) != _store->memory_size()) {
2985 return nullptr;
2986 }
2987
2988 // The _store must be the "last" store in a chain. If we find a use we could merge with
2989 // then that use or a store further down is the "last" store.
2990 Status status_use = find_adjacent_use_store(_store);
2991 if (status_use.found_store() != nullptr) {
2992 return nullptr;
2993 }
2994
2995 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
2996 Status status_def = find_adjacent_def_store(_store);
2997 if (status_def.found_store() == nullptr) {
2998 return nullptr;
2999 }
3000
3001 ResourceMark rm;
3002 Node_List merge_list;
3003 collect_merge_list(merge_list);
3004
3005 Node* merged_input_value = make_merged_input_value(merge_list);
3006 if (merged_input_value == nullptr) { return nullptr; }
3007
3008 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3009
3010 DEBUG_ONLY( if(TraceMergeStores) { trace(merge_list, merged_input_value, merged_store); } )
3011
3012 return merged_store;
3013 }
3014
3015 // Check compatibility between _store and other_store.
3016 bool MergePrimitiveArrayStores::is_compatible_store(const StoreNode* other_store) const {
3017 int opc = _store->Opcode();
3018 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3019 assert(_store->adr_type()->isa_aryptr() != nullptr, "must be array store");
3020
3021 if (other_store == nullptr ||
3022 _store->Opcode() != other_store->Opcode() ||
3023 other_store->adr_type() == nullptr ||
3024 other_store->adr_type()->isa_aryptr() == nullptr) {
3025 return false;
3026 }
3027
3028 // Check that the size of the stores, and the array elements are all the same.
3029 const TypeAryPtr* aryptr_t1 = _store->adr_type()->is_aryptr();
3030 const TypeAryPtr* aryptr_t2 = other_store->adr_type()->is_aryptr();
3031 BasicType aryptr_bt1 = aryptr_t1->elem()->array_element_basic_type();
3032 BasicType aryptr_bt2 = aryptr_t2->elem()->array_element_basic_type();
3033 if (!is_java_primitive(aryptr_bt1) || !is_java_primitive(aryptr_bt2)) {
3034 return false;
3035 }
3036 int size1 = type2aelembytes(aryptr_bt1);
3037 int size2 = type2aelembytes(aryptr_bt2);
3038 if (size1 != size2 ||
3039 size1 != _store->memory_size() ||
3040 _store->memory_size() != other_store->memory_size()) {
3041 return false;
3042 }
3043 return true;
3044 }
3045
3046 bool MergePrimitiveArrayStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3047 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3048 use_store->in(MemNode::ValueIn),
3049 def_store->memory_size())) {
3050 return false;
3051 }
3052
3053 ResourceMark rm;
3054 ArrayPointer array_pointer_use = ArrayPointer::make(_phase, use_store->in(MemNode::Address));
3055 ArrayPointer array_pointer_def = ArrayPointer::make(_phase, def_store->in(MemNode::Address));
3056 if (!array_pointer_def.is_adjacent_to_and_before(array_pointer_use, use_store->memory_size())) {
3057 return false;
3058 }
3059
3060 return true;
3061 }
3062
3063 bool MergePrimitiveArrayStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3064 // Pattern: [n1 = ConI, n2 = ConI]
3065 if (n1->Opcode() == Op_ConI) {
3066 return n2->Opcode() == Op_ConI;
3067 }
3068
3069 // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3070 #ifndef VM_LITTLE_ENDIAN
3071 // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3072 // Swapping n1 with n2 gives same pattern as on little endian platforms.
3073 swap(n1, n2);
3074 #endif // !VM_LITTLE_ENDIAN
3075 Node const* base_n2;
3076 jint shift_n2;
3077 if (!is_con_RShift(n2, base_n2, shift_n2)) {
3078 return false;
3079 }
3080 if (n1->Opcode() == Op_ConvL2I) {
3081 // look through
3082 n1 = n1->in(1);
3083 }
3084 Node const* base_n1;
3085 jint shift_n1;
3086 if (n1 == base_n2) {
3087 // n1 = base = base >> 0
3088 base_n1 = n1;
3089 shift_n1 = 0;
3090 } else if (!is_con_RShift(n1, base_n1, shift_n1)) {
3091 return false;
3092 }
3093 int bits_per_store = memory_size * 8;
3094 if (base_n1 != base_n2 ||
3095 shift_n1 + bits_per_store != shift_n2 ||
3096 shift_n1 % bits_per_store != 0) {
3097 return false;
3098 }
3099
3100 // both load from same value with correct shift
3101 return true;
3102 }
3103
3104 // Detect pattern: n = base_out >> shift_out
3105 bool MergePrimitiveArrayStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out) {
3106 assert(n != nullptr, "precondition");
3107
3108 int opc = n->Opcode();
3109 if (opc == Op_ConvL2I) {
3110 n = n->in(1);
3111 opc = n->Opcode();
3112 }
3113
3114 if ((opc == Op_RShiftI ||
3115 opc == Op_RShiftL ||
3116 opc == Op_URShiftI ||
3117 opc == Op_URShiftL) &&
3118 n->in(2)->is_ConI()) {
3119 base_out = n->in(1);
3120 shift_out = n->in(2)->get_int();
3121 // The shift must be positive:
3122 return shift_out >= 0;
3123 }
3124 return false;
3125 }
3126
3127 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3128 MergePrimitiveArrayStores::CFGStatus MergePrimitiveArrayStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3129 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3130
3131 Node* ctrl_use = use_store->in(MemNode::Control);
3132 Node* ctrl_def = def_store->in(MemNode::Control);
3133 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3134 return CFGStatus::Failure;
3135 }
3136
3137 if (ctrl_use == ctrl_def) {
3138 // Same ctrl -> no RangeCheck in between.
3139 // Check: use_store must be the only use of def_store.
3140 if (def_store->outcnt() > 1) {
3141 return CFGStatus::Failure;
3142 }
3143 return CFGStatus::SuccessNoRangeCheck;
3144 }
3145
3146 // Different ctrl -> could have RangeCheck in between.
3147 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3148 // 2. ctrl separated by RangeCheck.
3149 if (def_store->outcnt() != 2) {
3150 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3151 }
3152 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3153 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3154 if (merge_mem == nullptr ||
3155 merge_mem->outcnt() != 1) {
3156 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3157 }
3158 if (!ctrl_use->is_IfProj() ||
3159 !ctrl_use->in(0)->is_RangeCheck() ||
3160 ctrl_use->in(0)->outcnt() != 2) {
3161 return CFGStatus::Failure; // Not RangeCheck.
3162 }
3163 ProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3164 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3165 if (trap != merge_mem->unique_out() ||
3166 ctrl_use->in(0)->in(0) != ctrl_def) {
3167 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3168 }
3169
3170 return CFGStatus::SuccessWithRangeCheck;
3171 }
3172
3173 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_adjacent_use_store(const StoreNode* def_store) const {
3174 Status status_use = find_use_store(def_store);
3175 StoreNode* use_store = status_use.found_store();
3176 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3177 return Status::make_failure();
3178 }
3179 return status_use;
3180 }
3181
3182 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_adjacent_def_store(const StoreNode* use_store) const {
3183 Status status_def = find_def_store(use_store);
3184 StoreNode* def_store = status_def.found_store();
3185 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3186 return Status::make_failure();
3187 }
3188 return status_def;
3189 }
3190
3191 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_use_store(const StoreNode* def_store) const {
3192 Status status_use = find_use_store_unidirectional(def_store);
3193
3194 #ifdef ASSERT
3195 StoreNode* use_store = status_use.found_store();
3196 if (use_store != nullptr) {
3197 Status status_def = find_def_store_unidirectional(use_store);
3198 assert(status_def.found_store() == def_store &&
3199 status_def.found_range_check() == status_use.found_range_check(),
3200 "find_use_store and find_def_store must be symmetric");
3201 }
3202 #endif
3203
3204 return status_use;
3205 }
3206
3207 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_def_store(const StoreNode* use_store) const {
3208 Status status_def = find_def_store_unidirectional(use_store);
3209
3210 #ifdef ASSERT
3211 StoreNode* def_store = status_def.found_store();
3212 if (def_store != nullptr) {
3213 Status status_use = find_use_store_unidirectional(def_store);
3214 assert(status_use.found_store() == use_store &&
3215 status_use.found_range_check() == status_def.found_range_check(),
3216 "find_use_store and find_def_store must be symmetric");
3217 }
3218 #endif
3219
3220 return status_def;
3221 }
3222
3223 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3224 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3225
3226 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3227 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3228 if (is_compatible_store(use_store)) {
3229 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3230 }
3231 }
3232
3233 return Status::make_failure();
3234 }
3235
3236 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3237 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3238
3239 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3240 if (!is_compatible_store(def_store)) {
3241 return Status::make_failure();
3242 }
3243
3244 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3245 }
3246
3247 void MergePrimitiveArrayStores::collect_merge_list(Node_List& merge_list) const {
3248 // The merged store can be at most 8 bytes.
3249 const uint merge_list_max_size = 8 / _store->memory_size();
3250 assert(merge_list_max_size >= 2 &&
3251 merge_list_max_size <= 8 &&
3252 is_power_of_2(merge_list_max_size),
3253 "must be 2, 4 or 8");
3254
3255 // Traverse up the chain of adjacent def stores.
3256 StoreNode* current = _store;
3257 merge_list.push(current);
3258 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3259 Status status = find_adjacent_def_store(current);
3260 current = status.found_store();
3261 if (current != nullptr) {
3262 merge_list.push(current);
3263
3264 // We can have at most one RangeCheck.
3265 if (status.found_range_check()) {
3266 break;
3267 }
3268 }
3269 }
3270
3271 // Truncate the merge_list to a power of 2.
3272 const uint pow2size = round_down_power_of_2(merge_list.size());
3273 assert(pow2size >= 2, "must be merging at least 2 stores");
3274 while (merge_list.size() > pow2size) { merge_list.pop(); }
3275 }
3276
3277 // Merge the input values of the smaller stores to a single larger input value.
3278 Node* MergePrimitiveArrayStores::make_merged_input_value(const Node_List& merge_list) {
3279 int new_memory_size = _store->memory_size() * merge_list.size();
3280 Node* first = merge_list.at(merge_list.size()-1);
3281 Node* merged_input_value = nullptr;
3282 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3283 // Pattern: [ConI, ConI, ...] -> new constant
3284 jlong con = 0;
3285 jlong bits_per_store = _store->memory_size() * 8;
3286 jlong mask = (((jlong)1) << bits_per_store) - 1;
3287 for (uint i = 0; i < merge_list.size(); i++) {
3288 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3289 #ifdef VM_LITTLE_ENDIAN
3290 con = con << bits_per_store;
3291 con = con | (mask & con_i);
3292 #else // VM_LITTLE_ENDIAN
3293 con_i = (mask & con_i) << (i * bits_per_store);
3294 con = con | con_i;
3295 #endif // VM_LITTLE_ENDIAN
3296 }
3297 merged_input_value = _phase->longcon(con);
3298 } else {
3299 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3300 // | |
3301 // _store first
3302 //
3303 Node* hi = _store->in(MemNode::ValueIn);
3304 Node* lo = first->in(MemNode::ValueIn);
3305 #ifndef VM_LITTLE_ENDIAN
3306 // `_store` and `first` are swapped in the diagram above
3307 swap(hi, lo);
3308 #endif // !VM_LITTLE_ENDIAN
3309 Node const* hi_base;
3310 jint hi_shift;
3311 merged_input_value = lo;
3312 bool is_true = is_con_RShift(hi, hi_base, hi_shift);
3313 assert(is_true, "must detect con RShift");
3314 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3315 // look through
3316 merged_input_value = merged_input_value->in(1);
3317 }
3318 if (merged_input_value != hi_base) {
3319 // merged_input_value is not the base
3320 return nullptr;
3321 }
3322 }
3323
3324 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3325 // Example:
3326 //
3327 // long base = ...;
3328 // a[0] = (byte)(base >> 0);
3329 // a[1] = (byte)(base >> 8);
3330 //
3331 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3332 }
3333
3334 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3335 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3336 "merged_input_value is either int or long, and new_memory_size is small enough");
3337
3338 return merged_input_value;
3339 }
3340
3341 // //
3342 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3343 // | | | | | | //
3344 // | | | | +---------------+ | //
3345 // | | | | | | | //
3346 // | | +---------+ | | +---------------+ //
3347 // | | | | | | | | //
3348 // +--------------+ | | v1 +------------------------------+ | | v1 //
3349 // | | | | | | | | | | | | //
3350 // RangeCheck first_store RangeCheck | | first_store //
3351 // | | | | | | | //
3352 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3353 // | | ===> | | | //
3354 // +--------------+ | a2 v2 | | | //
3355 // | | | | | | | | //
3356 // | second_store | | | //
3357 // | | | | | [v1 v2 ... vn] //
3358 // ... ... | | | | //
3359 // | | | | | v //
3360 // +--------------+ | an vn +--------------+ | | merged_input_value //
3361 // | | | | | | | | //
3362 // last_store (= _store) merged_store //
3363 // //
3364 StoreNode* MergePrimitiveArrayStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3365 Node* first_store = merge_list.at(merge_list.size()-1);
3366 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3367 Node* first_mem = first_store->in(MemNode::Memory);
3368 Node* first_adr = first_store->in(MemNode::Address);
3369
3370 const TypePtr* new_adr_type = _store->adr_type();
3371
3372 int new_memory_size = _store->memory_size() * merge_list.size();
3373 BasicType bt = T_ILLEGAL;
3374 switch (new_memory_size) {
3375 case 2: bt = T_SHORT; break;
3376 case 4: bt = T_INT; break;
3377 case 8: bt = T_LONG; break;
3378 }
3379
3380 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3381 new_adr_type, merged_input_value, bt, MemNode::unordered);
3382
3383 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3384 // are all on the same slice. Hence, we need no barriers.
3385 merged_store->set_mismatched_access();
3386
3387 // Constants above may now also be be packed -> put candidate on worklist
3388 _phase->is_IterGVN()->_worklist.push(first_mem);
3389
3390 return merged_store;
3391 }
3392
3393 #ifdef ASSERT
3394 void MergePrimitiveArrayStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3395 stringStream ss;
3396 ss.print_cr("[TraceMergeStores]: Replace");
3397 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3398 merge_list.at(i)->dump("\n", false, &ss);
3399 }
3400 ss.print_cr("[TraceMergeStores]: with");
3401 merged_input_value->dump("\n", false, &ss);
3402 merged_store->dump("\n", false, &ss);
3403 tty->print("%s", ss.as_string());
3404 }
3405 #endif
3406
3407 //------------------------------Ideal------------------------------------------
3408 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3409 // When a store immediately follows a relevant allocation/initialization,
3410 // try to capture it into the initialization, or hoist it above.
3411 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3412 Node* p = MemNode::Ideal_common(phase, can_reshape);
3413 if (p) return (p == NodeSentinel) ? nullptr : p;
3414
3415 Node* mem = in(MemNode::Memory);
3416 Node* address = in(MemNode::Address);
3417 Node* value = in(MemNode::ValueIn);
3418 // Back-to-back stores to same address? Fold em up. Generally
3419 // unsafe if I have intervening uses... Also disallowed for StoreCM
3420 // since they must follow each StoreP operation. Redundant StoreCMs
3421 // are eliminated just before matching in final_graph_reshape.
3422 {
3423 Node* st = mem;
3424 // If Store 'st' has more than one use, we cannot fold 'st' away.
3425 // For example, 'st' might be the final state at a conditional
3426 // return. Or, 'st' might be used by some node which is live at
3427 // the same time 'st' is live, which might be unschedulable. So,
3428 // require exactly ONE user until such time as we clone 'mem' for
3429 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3430 // true).
3431 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
3432 // Looking at a dead closed cycle of memory?
3433 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3434 assert(Opcode() == st->Opcode() ||
3435 st->Opcode() == Op_StoreVector ||
3436 Opcode() == Op_StoreVector ||
3437 st->Opcode() == Op_StoreVectorScatter ||
3438 Opcode() == Op_StoreVectorScatter ||
3439 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3440 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3441 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3442 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3443 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3444
3445 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3446 st->as_Store()->memory_size() <= this->memory_size()) {
3447 Node* use = st->raw_out(0);
3448 if (phase->is_IterGVN()) {
3449 phase->is_IterGVN()->rehash_node_delayed(use);
3450 }
3451 // It's OK to do this in the parser, since DU info is always accurate,
3452 // and the parser always refers to nodes via SafePointNode maps.
3453 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3454 return this;
3455 }
3456 st = st->in(MemNode::Memory);
3457 }
3458 }
3459
3460
3461 // Capture an unaliased, unconditional, simple store into an initializer.
3462 // Or, if it is independent of the allocation, hoist it above the allocation.
3463 if (ReduceFieldZeroing && /*can_reshape &&*/
3464 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3465 InitializeNode* init = mem->in(0)->as_Initialize();
3466 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3467 if (offset > 0) {
3468 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3469 // If the InitializeNode captured me, it made a raw copy of me,
3470 // and I need to disappear.
3471 if (moved != nullptr) {
3472 // %%% hack to ensure that Ideal returns a new node:
3473 mem = MergeMemNode::make(mem);
3474 return mem; // fold me away
3475 }
3476 }
3477 }
3478
3479 // Fold reinterpret cast into memory operation:
3480 // StoreX mem (MoveY2X v) => StoreY mem v
3481 if (value->is_Move()) {
3482 const Type* vt = value->in(1)->bottom_type();
3483 if (has_reinterpret_variant(vt)) {
3484 if (phase->C->post_loop_opts_phase()) {
3485 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3486 } else {
3487 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3488 }
3489 }
3490 }
3491
3492 if (MergeStores && UseUnalignedAccesses) {
3493 if (phase->C->post_loop_opts_phase()) {
3494 MergePrimitiveArrayStores merge(phase, this);
3495 Node* progress = merge.run();
3496 if (progress != nullptr) { return progress; }
3497 } else {
3498 phase->C->record_for_post_loop_opts_igvn(this);
3499 }
3500 }
3501
3502 return nullptr; // No further progress
3503 }
3504
3505 //------------------------------Value-----------------------------------------
3506 const Type* StoreNode::Value(PhaseGVN* phase) const {
3507 // Either input is TOP ==> the result is TOP
3508 const Type *t1 = phase->type( in(MemNode::Memory) );
3509 if( t1 == Type::TOP ) return Type::TOP;
3510 const Type *t2 = phase->type( in(MemNode::Address) );
3511 if( t2 == Type::TOP ) return Type::TOP;
3512 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3513 if( t3 == Type::TOP ) return Type::TOP;
3514 return Type::MEMORY;
3515 }
3516
3517 //------------------------------Identity---------------------------------------
3518 // Remove redundant stores:
3519 // Store(m, p, Load(m, p)) changes to m.
3520 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3521 Node* StoreNode::Identity(PhaseGVN* phase) {
3522 Node* mem = in(MemNode::Memory);
3523 Node* adr = in(MemNode::Address);
3524 Node* val = in(MemNode::ValueIn);
3525
3526 Node* result = this;
3527
3528 // Load then Store? Then the Store is useless
3529 if (val->is_Load() &&
3530 val->in(MemNode::Address)->eqv_uncast(adr) &&
3531 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3532 val->as_Load()->store_Opcode() == Opcode()) {
3533 // Ensure vector type is the same
3534 if (!is_StoreVector() || (mem->is_LoadVector() && as_StoreVector()->vect_type() == mem->as_LoadVector()->vect_type())) {
3535 result = mem;
3536 }
3537 }
3538
3539 // Two stores in a row of the same value?
3540 if (result == this &&
3541 mem->is_Store() &&
3542 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3543 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3544 mem->Opcode() == Opcode()) {
3545 if (!is_StoreVector()) {
3546 result = mem;
3547 } else {
3548 const StoreVectorNode* store_vector = as_StoreVector();
3549 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3550 const Node* store_indices = store_vector->indices();
3551 const Node* mem_indices = mem_vector->indices();
3552 const Node* store_mask = store_vector->mask();
3553 const Node* mem_mask = mem_vector->mask();
3554 // Ensure types, indices, and masks match
3555 if (store_vector->vect_type() == mem_vector->vect_type() &&
3556 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3557 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3558 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3559 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3560 result = mem;
3561 }
3562 }
3563 }
3564
3565 // Store of zero anywhere into a freshly-allocated object?
3566 // Then the store is useless.
3567 // (It must already have been captured by the InitializeNode.)
3568 if (result == this &&
3569 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
3570 // a newly allocated object is already all-zeroes everywhere
3571 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
3572 result = mem;
3573 }
3574
3575 if (result == this) {
3576 // the store may also apply to zero-bits in an earlier object
3577 Node* prev_mem = find_previous_store(phase);
3578 // Steps (a), (b): Walk past independent stores to find an exact match.
3579 if (prev_mem != nullptr) {
3580 Node* prev_val = can_see_stored_value(prev_mem, phase);
3581 if (prev_val != nullptr && prev_val == val) {
3582 // prev_val and val might differ by a cast; it would be good
3583 // to keep the more informative of the two.
3584 result = mem;
3585 }
3586 }
3587 }
3588 }
3589
3590 PhaseIterGVN* igvn = phase->is_IterGVN();
3591 if (result != this && igvn != nullptr) {
3592 MemBarNode* trailing = trailing_membar();
3593 if (trailing != nullptr) {
3594 #ifdef ASSERT
3595 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3596 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3597 #endif
3598 trailing->remove(igvn);
3599 }
3600 }
3601
3602 return result;
3603 }
3604
3605 //------------------------------match_edge-------------------------------------
3606 // Do we Match on this edge index or not? Match only memory & value
3607 uint StoreNode::match_edge(uint idx) const {
3608 return idx == MemNode::Address || idx == MemNode::ValueIn;
3609 }
3610
3611 //------------------------------cmp--------------------------------------------
3612 // Do not common stores up together. They generally have to be split
3613 // back up anyways, so do not bother.
3614 bool StoreNode::cmp( const Node &n ) const {
3615 return (&n == this); // Always fail except on self
3616 }
3617
3618 //------------------------------Ideal_masked_input-----------------------------
3619 // Check for a useless mask before a partial-word store
3620 // (StoreB ... (AndI valIn conIa) )
3621 // If (conIa & mask == mask) this simplifies to
3622 // (StoreB ... (valIn) )
3623 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3624 Node *val = in(MemNode::ValueIn);
3625 if( val->Opcode() == Op_AndI ) {
3626 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3627 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3628 set_req_X(MemNode::ValueIn, val->in(1), phase);
3629 return this;
3630 }
3631 }
3632 return nullptr;
3633 }
3634
3635
3636 //------------------------------Ideal_sign_extended_input----------------------
3637 // Check for useless sign-extension before a partial-word store
3638 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
3639 // If (conIL == conIR && conIR <= num_bits) this simplifies to
3640 // (StoreB ... (valIn) )
3641 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
3642 Node *val = in(MemNode::ValueIn);
3643 if( val->Opcode() == Op_RShiftI ) {
3644 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3645 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
3646 Node *shl = val->in(1);
3647 if( shl->Opcode() == Op_LShiftI ) {
3648 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
3649 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
3650 set_req_X(MemNode::ValueIn, shl->in(1), phase);
3651 return this;
3652 }
3653 }
3654 }
3655 }
3656 return nullptr;
3657 }
3658
3659 //------------------------------value_never_loaded-----------------------------------
3660 // Determine whether there are any possible loads of the value stored.
3661 // For simplicity, we actually check if there are any loads from the
3662 // address stored to, not just for loads of the value stored by this node.
3663 //
3664 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
3665 Node *adr = in(Address);
3666 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
3667 if (adr_oop == nullptr)
3668 return false;
3669 if (!adr_oop->is_known_instance_field())
3670 return false; // if not a distinct instance, there may be aliases of the address
3671 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
3672 Node *use = adr->fast_out(i);
3673 if (use->is_Load() || use->is_LoadStore()) {
3674 return false;
3675 }
3676 }
3677 return true;
3678 }
3679
3680 MemBarNode* StoreNode::trailing_membar() const {
3681 if (is_release()) {
3682 MemBarNode* trailing_mb = nullptr;
3683 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3684 Node* u = fast_out(i);
3685 if (u->is_MemBar()) {
3686 if (u->as_MemBar()->trailing_store()) {
3687 assert(u->Opcode() == Op_MemBarVolatile, "");
3688 assert(trailing_mb == nullptr, "only one");
3689 trailing_mb = u->as_MemBar();
3690 #ifdef ASSERT
3691 Node* leading = u->as_MemBar()->leading_membar();
3692 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3693 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
3694 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
3695 #endif
3696 } else {
3697 assert(u->as_MemBar()->standalone(), "");
3698 }
3699 }
3700 }
3701 return trailing_mb;
3702 }
3703 return nullptr;
3704 }
3705
3706
3707 //=============================================================================
3708 //------------------------------Ideal------------------------------------------
3709 // If the store is from an AND mask that leaves the low bits untouched, then
3710 // we can skip the AND operation. If the store is from a sign-extension
3711 // (a left shift, then right shift) we can skip both.
3712 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
3713 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
3714 if( progress != nullptr ) return progress;
3715
3716 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
3717 if( progress != nullptr ) return progress;
3718
3719 // Finally check the default case
3720 return StoreNode::Ideal(phase, can_reshape);
3721 }
3722
3723 //=============================================================================
3724 //------------------------------Ideal------------------------------------------
3725 // If the store is from an AND mask that leaves the low bits untouched, then
3726 // we can skip the AND operation
3727 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
3728 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
3729 if( progress != nullptr ) return progress;
3730
3731 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
3732 if( progress != nullptr ) return progress;
3733
3734 // Finally check the default case
3735 return StoreNode::Ideal(phase, can_reshape);
3736 }
3737
3738 //=============================================================================
3739 //------------------------------Identity---------------------------------------
3740 Node* StoreCMNode::Identity(PhaseGVN* phase) {
3741 // No need to card mark when storing a null ptr
3742 Node* my_store = in(MemNode::OopStore);
3743 if (my_store->is_Store()) {
3744 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
3745 if( t1 == TypePtr::NULL_PTR ) {
3746 return in(MemNode::Memory);
3747 }
3748 }
3749 return this;
3750 }
3751
3752 //=============================================================================
3753 //------------------------------Ideal---------------------------------------
3754 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
3755 Node* progress = StoreNode::Ideal(phase, can_reshape);
3756 if (progress != nullptr) return progress;
3757
3758 Node* my_store = in(MemNode::OopStore);
3759 if (my_store->is_MergeMem()) {
3760 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
3761 set_req_X(MemNode::OopStore, mem, phase);
3762 return this;
3763 }
3764
3765 return nullptr;
3766 }
3767
3768 //------------------------------Value-----------------------------------------
3769 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
3770 // Either input is TOP ==> the result is TOP (checked in StoreNode::Value).
3771 // If extra input is TOP ==> the result is TOP
3772 const Type* t = phase->type(in(MemNode::OopStore));
3773 if (t == Type::TOP) {
3774 return Type::TOP;
3775 }
3776 return StoreNode::Value(phase);
3777 }
3778
3779
3780 //=============================================================================
3781 //----------------------------------SCMemProjNode------------------------------
3782 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3783 {
3784 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
3785 return Type::TOP;
3786 }
3787 return bottom_type();
3788 }
3789
3790 //=============================================================================
3791 //----------------------------------LoadStoreNode------------------------------
3792 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3793 : Node(required),
3794 _type(rt),
3795 _adr_type(at),
3796 _barrier_data(0)
3797 {
3798 init_req(MemNode::Control, c );
3799 init_req(MemNode::Memory , mem);
3800 init_req(MemNode::Address, adr);
3801 init_req(MemNode::ValueIn, val);
3802 init_class_id(Class_LoadStore);
3803 }
3804
3805 //------------------------------Value-----------------------------------------
3806 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
3807 // Either input is TOP ==> the result is TOP
3808 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
3809 return Type::TOP;
3810 }
3811 const Type* t = phase->type(in(MemNode::Memory));
3812 if (t == Type::TOP) {
3813 return Type::TOP;
3814 }
3815 t = phase->type(in(MemNode::Address));
3816 if (t == Type::TOP) {
3817 return Type::TOP;
3818 }
3819 t = phase->type(in(MemNode::ValueIn));
3820 if (t == Type::TOP) {
3821 return Type::TOP;
3822 }
3823 return bottom_type();
3824 }
3825
3826 uint LoadStoreNode::ideal_reg() const {
3827 return _type->ideal_reg();
3828 }
3829
3830 // This method conservatively checks if the result of a LoadStoreNode is
3831 // used, that is, if it returns true, then it is definitely the case that
3832 // the result of the node is not needed.
3833 // For example, GetAndAdd can be matched into a lock_add instead of a
3834 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
3835 bool LoadStoreNode::result_not_used() const {
3836 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3837 Node *x = fast_out(i);
3838 if (x->Opcode() == Op_SCMemProj) {
3839 continue;
3840 }
3841 if (x->bottom_type() == TypeTuple::MEMBAR &&
3842 !x->is_Call() &&
3843 x->Opcode() != Op_Blackhole) {
3844 continue;
3845 }
3846 return false;
3847 }
3848 return true;
3849 }
3850
3851 MemBarNode* LoadStoreNode::trailing_membar() const {
3852 MemBarNode* trailing = nullptr;
3853 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3854 Node* u = fast_out(i);
3855 if (u->is_MemBar()) {
3856 if (u->as_MemBar()->trailing_load_store()) {
3857 assert(u->Opcode() == Op_MemBarAcquire, "");
3858 assert(trailing == nullptr, "only one");
3859 trailing = u->as_MemBar();
3860 #ifdef ASSERT
3861 Node* leading = trailing->leading_membar();
3862 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3863 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3864 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3865 #endif
3866 } else {
3867 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3868 }
3869 }
3870 }
3871
3872 return trailing;
3873 }
3874
3875 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3876
3877 //=============================================================================
3878 //----------------------------------LoadStoreConditionalNode--------------------
3879 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
3880 init_req(ExpectedIn, ex );
3881 }
3882
3883 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
3884 // Either input is TOP ==> the result is TOP
3885 const Type* t = phase->type(in(ExpectedIn));
3886 if (t == Type::TOP) {
3887 return Type::TOP;
3888 }
3889 return LoadStoreNode::Value(phase);
3890 }
3891
3892 //=============================================================================
3893 //-------------------------------adr_type--------------------------------------
3894 const TypePtr* ClearArrayNode::adr_type() const {
3895 Node *adr = in(3);
3896 if (adr == nullptr) return nullptr; // node is dead
3897 return MemNode::calculate_adr_type(adr->bottom_type());
3898 }
3899
3900 //------------------------------match_edge-------------------------------------
3901 // Do we Match on this edge index or not? Do not match memory
3902 uint ClearArrayNode::match_edge(uint idx) const {
3903 return idx > 1;
3904 }
3905
3906 //------------------------------Identity---------------------------------------
3907 // Clearing a zero length array does nothing
3908 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3909 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
3910 }
3911
3912 //------------------------------Idealize---------------------------------------
3913 // Clearing a short array is faster with stores
3914 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3915 // Already know this is a large node, do not try to ideal it
3916 if (_is_large) return nullptr;
3917
3918 const int unit = BytesPerLong;
3919 const TypeX* t = phase->type(in(2))->isa_intptr_t();
3920 if (!t) return nullptr;
3921 if (!t->is_con()) return nullptr;
3922 intptr_t raw_count = t->get_con();
3923 intptr_t size = raw_count;
3924 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3925 // Clearing nothing uses the Identity call.
3926 // Negative clears are possible on dead ClearArrays
3927 // (see jck test stmt114.stmt11402.val).
3928 if (size <= 0 || size % unit != 0) return nullptr;
3929 intptr_t count = size / unit;
3930 // Length too long; communicate this to matchers and assemblers.
3931 // Assemblers are responsible to produce fast hardware clears for it.
3932 if (size > InitArrayShortSize) {
3933 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
3934 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
3935 return nullptr;
3936 }
3937 if (!IdealizeClearArrayNode) return nullptr;
3938 Node *mem = in(1);
3939 if( phase->type(mem)==Type::TOP ) return nullptr;
3940 Node *adr = in(3);
3941 const Type* at = phase->type(adr);
3942 if( at==Type::TOP ) return nullptr;
3943 const TypePtr* atp = at->isa_ptr();
3944 // adjust atp to be the correct array element address type
3945 if (atp == nullptr) atp = TypePtr::BOTTOM;
3946 else atp = atp->add_offset(Type::OffsetBot);
3947 // Get base for derived pointer purposes
3948 if( adr->Opcode() != Op_AddP ) Unimplemented();
3949 Node *base = adr->in(1);
3950
3951 Node *zero = phase->makecon(TypeLong::ZERO);
3952 Node *off = phase->MakeConX(BytesPerLong);
3953 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3954 count--;
3955 while( count-- ) {
3956 mem = phase->transform(mem);
3957 adr = phase->transform(new AddPNode(base,adr,off));
3958 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3959 }
3960 return mem;
3961 }
3962
3963 //----------------------------step_through----------------------------------
3964 // Return allocation input memory edge if it is different instance
3965 // or itself if it is the one we are looking for.
3966 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
3967 Node* n = *np;
3968 assert(n->is_ClearArray(), "sanity");
3969 intptr_t offset;
3970 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3971 // This method is called only before Allocate nodes are expanded
3972 // during macro nodes expansion. Before that ClearArray nodes are
3973 // only generated in PhaseMacroExpand::generate_arraycopy() (before
3974 // Allocate nodes are expanded) which follows allocations.
3975 assert(alloc != nullptr, "should have allocation");
3976 if (alloc->_idx == instance_id) {
3977 // Can not bypass initialization of the instance we are looking for.
3978 return false;
3979 }
3980 // Otherwise skip it.
3981 InitializeNode* init = alloc->initialization();
3982 if (init != nullptr)
3983 *np = init->in(TypeFunc::Memory);
3984 else
3985 *np = alloc->in(TypeFunc::Memory);
3986 return true;
3987 }
3988
3989 //----------------------------clear_memory-------------------------------------
3990 // Generate code to initialize object storage to zero.
3991 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3992 intptr_t start_offset,
3993 Node* end_offset,
3994 PhaseGVN* phase) {
3995 intptr_t offset = start_offset;
3996
3997 int unit = BytesPerLong;
3998 if ((offset % unit) != 0) {
3999 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
4000 adr = phase->transform(adr);
4001 const TypePtr* atp = TypeRawPtr::BOTTOM;
4002 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4003 mem = phase->transform(mem);
4004 offset += BytesPerInt;
4005 }
4006 assert((offset % unit) == 0, "");
4007
4008 // Initialize the remaining stuff, if any, with a ClearArray.
4009 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
4010 }
4011
4012 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4013 Node* start_offset,
4014 Node* end_offset,
4015 PhaseGVN* phase) {
4016 if (start_offset == end_offset) {
4017 // nothing to do
4018 return mem;
4019 }
4020
4021 int unit = BytesPerLong;
4022 Node* zbase = start_offset;
4023 Node* zend = end_offset;
4024
4025 // Scale to the unit required by the CPU:
4026 if (!Matcher::init_array_count_is_in_bytes) {
4027 Node* shift = phase->intcon(exact_log2(unit));
4028 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4029 zend = phase->transform(new URShiftXNode(zend, shift) );
4030 }
4031
4032 // Bulk clear double-words
4033 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4034 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
4035 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
4036 return phase->transform(mem);
4037 }
4038
4039 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4040 intptr_t start_offset,
4041 intptr_t end_offset,
4042 PhaseGVN* phase) {
4043 if (start_offset == end_offset) {
4044 // nothing to do
4045 return mem;
4046 }
4047
4048 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4049 intptr_t done_offset = end_offset;
4050 if ((done_offset % BytesPerLong) != 0) {
4051 done_offset -= BytesPerInt;
4052 }
4053 if (done_offset > start_offset) {
4054 mem = clear_memory(ctl, mem, dest,
4055 start_offset, phase->MakeConX(done_offset), phase);
4056 }
4057 if (done_offset < end_offset) { // emit the final 32-bit store
4058 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
4059 adr = phase->transform(adr);
4060 const TypePtr* atp = TypeRawPtr::BOTTOM;
4061 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4062 mem = phase->transform(mem);
4063 done_offset += BytesPerInt;
4064 }
4065 assert(done_offset == end_offset, "");
4066 return mem;
4067 }
4068
4069 //=============================================================================
4070 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4071 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4072 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4073 #ifdef ASSERT
4074 , _pair_idx(0)
4075 #endif
4076 {
4077 init_class_id(Class_MemBar);
4078 Node* top = C->top();
4079 init_req(TypeFunc::I_O,top);
4080 init_req(TypeFunc::FramePtr,top);
4081 init_req(TypeFunc::ReturnAdr,top);
4082 if (precedent != nullptr)
4083 init_req(TypeFunc::Parms, precedent);
4084 }
4085
4086 //------------------------------cmp--------------------------------------------
4087 uint MemBarNode::hash() const { return NO_HASH; }
4088 bool MemBarNode::cmp( const Node &n ) const {
4089 return (&n == this); // Always fail except on self
4090 }
4091
4092 //------------------------------make-------------------------------------------
4093 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4094 switch (opcode) {
4095 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4096 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4097 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4098 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4099 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4100 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4101 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4102 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4103 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4104 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4105 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4106 case Op_Initialize: return new InitializeNode(C, atp, pn);
4107 default: ShouldNotReachHere(); return nullptr;
4108 }
4109 }
4110
4111 void MemBarNode::remove(PhaseIterGVN *igvn) {
4112 if (outcnt() != 2) {
4113 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
4114 assert(outcnt() == 1, "Only control then");
4115 }
4116 if (trailing_store() || trailing_load_store()) {
4117 MemBarNode* leading = leading_membar();
4118 if (leading != nullptr) {
4119 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4120 leading->remove(igvn);
4121 }
4122 }
4123 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4124 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4125 }
4126 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4127 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4128 }
4129 }
4130
4131 //------------------------------Ideal------------------------------------------
4132 // Return a node which is more "ideal" than the current node. Strip out
4133 // control copies
4134 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4135 if (remove_dead_region(phase, can_reshape)) return this;
4136 // Don't bother trying to transform a dead node
4137 if (in(0) && in(0)->is_top()) {
4138 return nullptr;
4139 }
4140
4141 bool progress = false;
4142 // Eliminate volatile MemBars for scalar replaced objects.
4143 if (can_reshape && req() == (Precedent+1)) {
4144 bool eliminate = false;
4145 int opc = Opcode();
4146 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4147 // Volatile field loads and stores.
4148 Node* my_mem = in(MemBarNode::Precedent);
4149 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4150 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4151 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4152 // replace this Precedent (decodeN) with the Load instead.
4153 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4154 Node* load_node = my_mem->in(1);
4155 set_req(MemBarNode::Precedent, load_node);
4156 phase->is_IterGVN()->_worklist.push(my_mem);
4157 my_mem = load_node;
4158 } else {
4159 assert(my_mem->unique_out() == this, "sanity");
4160 assert(!trailing_load_store(), "load store node can't be eliminated");
4161 del_req(Precedent);
4162 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4163 my_mem = nullptr;
4164 }
4165 progress = true;
4166 }
4167 if (my_mem != nullptr && my_mem->is_Mem()) {
4168 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4169 // Check for scalar replaced object reference.
4170 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4171 t_oop->offset() != Type::OffsetBot &&
4172 t_oop->offset() != Type::OffsetTop) {
4173 eliminate = true;
4174 }
4175 }
4176 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4177 // Final field stores.
4178 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4179 if ((alloc != nullptr) && alloc->is_Allocate() &&
4180 alloc->as_Allocate()->does_not_escape_thread()) {
4181 // The allocated object does not escape.
4182 eliminate = true;
4183 }
4184 }
4185 if (eliminate) {
4186 // Replace MemBar projections by its inputs.
4187 PhaseIterGVN* igvn = phase->is_IterGVN();
4188 remove(igvn);
4189 // Must return either the original node (now dead) or a new node
4190 // (Do not return a top here, since that would break the uniqueness of top.)
4191 return new ConINode(TypeInt::ZERO);
4192 }
4193 }
4194 return progress ? this : nullptr;
4195 }
4196
4197 //------------------------------Value------------------------------------------
4198 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4199 if( !in(0) ) return Type::TOP;
4200 if( phase->type(in(0)) == Type::TOP )
4201 return Type::TOP;
4202 return TypeTuple::MEMBAR;
4203 }
4204
4205 //------------------------------match------------------------------------------
4206 // Construct projections for memory.
4207 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
4208 switch (proj->_con) {
4209 case TypeFunc::Control:
4210 case TypeFunc::Memory:
4211 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
4212 }
4213 ShouldNotReachHere();
4214 return nullptr;
4215 }
4216
4217 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4218 trailing->_kind = TrailingStore;
4219 leading->_kind = LeadingStore;
4220 #ifdef ASSERT
4221 trailing->_pair_idx = leading->_idx;
4222 leading->_pair_idx = leading->_idx;
4223 #endif
4224 }
4225
4226 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4227 trailing->_kind = TrailingLoadStore;
4228 leading->_kind = LeadingLoadStore;
4229 #ifdef ASSERT
4230 trailing->_pair_idx = leading->_idx;
4231 leading->_pair_idx = leading->_idx;
4232 #endif
4233 }
4234
4235 MemBarNode* MemBarNode::trailing_membar() const {
4236 ResourceMark rm;
4237 Node* trailing = (Node*)this;
4238 VectorSet seen;
4239 Node_Stack multis(0);
4240 do {
4241 Node* c = trailing;
4242 uint i = 0;
4243 do {
4244 trailing = nullptr;
4245 for (; i < c->outcnt(); i++) {
4246 Node* next = c->raw_out(i);
4247 if (next != c && next->is_CFG()) {
4248 if (c->is_MultiBranch()) {
4249 if (multis.node() == c) {
4250 multis.set_index(i+1);
4251 } else {
4252 multis.push(c, i+1);
4253 }
4254 }
4255 trailing = next;
4256 break;
4257 }
4258 }
4259 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4260 break;
4261 }
4262 while (multis.size() > 0) {
4263 c = multis.node();
4264 i = multis.index();
4265 if (i < c->req()) {
4266 break;
4267 }
4268 multis.pop();
4269 }
4270 } while (multis.size() > 0);
4271 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4272
4273 MemBarNode* mb = trailing->as_MemBar();
4274 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4275 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4276 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4277 return mb;
4278 }
4279
4280 MemBarNode* MemBarNode::leading_membar() const {
4281 ResourceMark rm;
4282 VectorSet seen;
4283 Node_Stack regions(0);
4284 Node* leading = in(0);
4285 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4286 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4287 leading = nullptr;
4288 while (regions.size() > 0 && leading == nullptr) {
4289 Node* r = regions.node();
4290 uint i = regions.index();
4291 if (i < r->req()) {
4292 leading = r->in(i);
4293 regions.set_index(i+1);
4294 } else {
4295 regions.pop();
4296 }
4297 }
4298 if (leading == nullptr) {
4299 assert(regions.size() == 0, "all paths should have been tried");
4300 return nullptr;
4301 }
4302 }
4303 if (leading->is_Region()) {
4304 regions.push(leading, 2);
4305 leading = leading->in(1);
4306 } else {
4307 leading = leading->in(0);
4308 }
4309 }
4310 #ifdef ASSERT
4311 Unique_Node_List wq;
4312 wq.push((Node*)this);
4313 uint found = 0;
4314 for (uint i = 0; i < wq.size(); i++) {
4315 Node* n = wq.at(i);
4316 if (n->is_Region()) {
4317 for (uint j = 1; j < n->req(); j++) {
4318 Node* in = n->in(j);
4319 if (in != nullptr && !in->is_top()) {
4320 wq.push(in);
4321 }
4322 }
4323 } else {
4324 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4325 assert(n == leading, "consistency check failed");
4326 found++;
4327 } else {
4328 Node* in = n->in(0);
4329 if (in != nullptr && !in->is_top()) {
4330 wq.push(in);
4331 }
4332 }
4333 }
4334 }
4335 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4336 #endif
4337 if (leading == nullptr) {
4338 return nullptr;
4339 }
4340 MemBarNode* mb = leading->as_MemBar();
4341 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4342 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4343 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4344 return mb;
4345 }
4346
4347
4348 //===========================InitializeNode====================================
4349 // SUMMARY:
4350 // This node acts as a memory barrier on raw memory, after some raw stores.
4351 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4352 // The Initialize can 'capture' suitably constrained stores as raw inits.
4353 // It can coalesce related raw stores into larger units (called 'tiles').
4354 // It can avoid zeroing new storage for memory units which have raw inits.
4355 // At macro-expansion, it is marked 'complete', and does not optimize further.
4356 //
4357 // EXAMPLE:
4358 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4359 // ctl = incoming control; mem* = incoming memory
4360 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4361 // First allocate uninitialized memory and fill in the header:
4362 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4363 // ctl := alloc.Control; mem* := alloc.Memory*
4364 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4365 // Then initialize to zero the non-header parts of the raw memory block:
4366 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4367 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4368 // After the initialize node executes, the object is ready for service:
4369 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4370 // Suppose its body is immediately initialized as {1,2}:
4371 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4372 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4373 // mem.SLICE(#short[*]) := store2
4374 //
4375 // DETAILS:
4376 // An InitializeNode collects and isolates object initialization after
4377 // an AllocateNode and before the next possible safepoint. As a
4378 // memory barrier (MemBarNode), it keeps critical stores from drifting
4379 // down past any safepoint or any publication of the allocation.
4380 // Before this barrier, a newly-allocated object may have uninitialized bits.
4381 // After this barrier, it may be treated as a real oop, and GC is allowed.
4382 //
4383 // The semantics of the InitializeNode include an implicit zeroing of
4384 // the new object from object header to the end of the object.
4385 // (The object header and end are determined by the AllocateNode.)
4386 //
4387 // Certain stores may be added as direct inputs to the InitializeNode.
4388 // These stores must update raw memory, and they must be to addresses
4389 // derived from the raw address produced by AllocateNode, and with
4390 // a constant offset. They must be ordered by increasing offset.
4391 // The first one is at in(RawStores), the last at in(req()-1).
4392 // Unlike most memory operations, they are not linked in a chain,
4393 // but are displayed in parallel as users of the rawmem output of
4394 // the allocation.
4395 //
4396 // (See comments in InitializeNode::capture_store, which continue
4397 // the example given above.)
4398 //
4399 // When the associated Allocate is macro-expanded, the InitializeNode
4400 // may be rewritten to optimize collected stores. A ClearArrayNode
4401 // may also be created at that point to represent any required zeroing.
4402 // The InitializeNode is then marked 'complete', prohibiting further
4403 // capturing of nearby memory operations.
4404 //
4405 // During macro-expansion, all captured initializations which store
4406 // constant values of 32 bits or smaller are coalesced (if advantageous)
4407 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4408 // initialized in fewer memory operations. Memory words which are
4409 // covered by neither tiles nor non-constant stores are pre-zeroed
4410 // by explicit stores of zero. (The code shape happens to do all
4411 // zeroing first, then all other stores, with both sequences occurring
4412 // in order of ascending offsets.)
4413 //
4414 // Alternatively, code may be inserted between an AllocateNode and its
4415 // InitializeNode, to perform arbitrary initialization of the new object.
4416 // E.g., the object copying intrinsics insert complex data transfers here.
4417 // The initialization must then be marked as 'complete' disable the
4418 // built-in zeroing semantics and the collection of initializing stores.
4419 //
4420 // While an InitializeNode is incomplete, reads from the memory state
4421 // produced by it are optimizable if they match the control edge and
4422 // new oop address associated with the allocation/initialization.
4423 // They return a stored value (if the offset matches) or else zero.
4424 // A write to the memory state, if it matches control and address,
4425 // and if it is to a constant offset, may be 'captured' by the
4426 // InitializeNode. It is cloned as a raw memory operation and rewired
4427 // inside the initialization, to the raw oop produced by the allocation.
4428 // Operations on addresses which are provably distinct (e.g., to
4429 // other AllocateNodes) are allowed to bypass the initialization.
4430 //
4431 // The effect of all this is to consolidate object initialization
4432 // (both arrays and non-arrays, both piecewise and bulk) into a
4433 // single location, where it can be optimized as a unit.
4434 //
4435 // Only stores with an offset less than TrackedInitializationLimit words
4436 // will be considered for capture by an InitializeNode. This puts a
4437 // reasonable limit on the complexity of optimized initializations.
4438
4439 //---------------------------InitializeNode------------------------------------
4440 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4441 : MemBarNode(C, adr_type, rawoop),
4442 _is_complete(Incomplete), _does_not_escape(false)
4443 {
4444 init_class_id(Class_Initialize);
4445
4446 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4447 assert(in(RawAddress) == rawoop, "proper init");
4448 // Note: allocation() can be null, for secondary initialization barriers
4449 }
4450
4451 // Since this node is not matched, it will be processed by the
4452 // register allocator. Declare that there are no constraints
4453 // on the allocation of the RawAddress edge.
4454 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4455 // This edge should be set to top, by the set_complete. But be conservative.
4456 if (idx == InitializeNode::RawAddress)
4457 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4458 return RegMask::Empty;
4459 }
4460
4461 Node* InitializeNode::memory(uint alias_idx) {
4462 Node* mem = in(Memory);
4463 if (mem->is_MergeMem()) {
4464 return mem->as_MergeMem()->memory_at(alias_idx);
4465 } else {
4466 // incoming raw memory is not split
4467 return mem;
4468 }
4469 }
4470
4471 bool InitializeNode::is_non_zero() {
4472 if (is_complete()) return false;
4473 remove_extra_zeroes();
4474 return (req() > RawStores);
4475 }
4476
4477 void InitializeNode::set_complete(PhaseGVN* phase) {
4478 assert(!is_complete(), "caller responsibility");
4479 _is_complete = Complete;
4480
4481 // After this node is complete, it contains a bunch of
4482 // raw-memory initializations. There is no need for
4483 // it to have anything to do with non-raw memory effects.
4484 // Therefore, tell all non-raw users to re-optimize themselves,
4485 // after skipping the memory effects of this initialization.
4486 PhaseIterGVN* igvn = phase->is_IterGVN();
4487 if (igvn) igvn->add_users_to_worklist(this);
4488 }
4489
4490 // convenience function
4491 // return false if the init contains any stores already
4492 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4493 InitializeNode* init = initialization();
4494 if (init == nullptr || init->is_complete()) return false;
4495 init->remove_extra_zeroes();
4496 // for now, if this allocation has already collected any inits, bail:
4497 if (init->is_non_zero()) return false;
4498 init->set_complete(phase);
4499 return true;
4500 }
4501
4502 void InitializeNode::remove_extra_zeroes() {
4503 if (req() == RawStores) return;
4504 Node* zmem = zero_memory();
4505 uint fill = RawStores;
4506 for (uint i = fill; i < req(); i++) {
4507 Node* n = in(i);
4508 if (n->is_top() || n == zmem) continue; // skip
4509 if (fill < i) set_req(fill, n); // compact
4510 ++fill;
4511 }
4512 // delete any empty spaces created:
4513 while (fill < req()) {
4514 del_req(fill);
4515 }
4516 }
4517
4518 // Helper for remembering which stores go with which offsets.
4519 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4520 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4521 intptr_t offset = -1;
4522 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4523 phase, offset);
4524 if (base == nullptr) return -1; // something is dead,
4525 if (offset < 0) return -1; // dead, dead
4526 return offset;
4527 }
4528
4529 // Helper for proving that an initialization expression is
4530 // "simple enough" to be folded into an object initialization.
4531 // Attempts to prove that a store's initial value 'n' can be captured
4532 // within the initialization without creating a vicious cycle, such as:
4533 // { Foo p = new Foo(); p.next = p; }
4534 // True for constants and parameters and small combinations thereof.
4535 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4536 ResourceMark rm;
4537 Unique_Node_List worklist;
4538 worklist.push(value);
4539
4540 uint complexity_limit = 20;
4541 for (uint j = 0; j < worklist.size(); j++) {
4542 if (j >= complexity_limit) {
4543 return false; // Bail out if processed too many nodes
4544 }
4545
4546 Node* n = worklist.at(j);
4547 if (n == nullptr) continue; // (can this really happen?)
4548 if (n->is_Proj()) n = n->in(0);
4549 if (n == this) return false; // found a cycle
4550 if (n->is_Con()) continue;
4551 if (n->is_Start()) continue; // params, etc., are OK
4552 if (n->is_Root()) continue; // even better
4553
4554 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4555 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4556 continue;
4557 }
4558
4559 Node* ctl = n->in(0);
4560 if (ctl != nullptr && !ctl->is_top()) {
4561 if (ctl->is_Proj()) ctl = ctl->in(0);
4562 if (ctl == this) return false;
4563
4564 // If we already know that the enclosing memory op is pinned right after
4565 // the init, then any control flow that the store has picked up
4566 // must have preceded the init, or else be equal to the init.
4567 // Even after loop optimizations (which might change control edges)
4568 // a store is never pinned *before* the availability of its inputs.
4569 if (!MemNode::all_controls_dominate(n, this))
4570 return false; // failed to prove a good control
4571 }
4572
4573 // Check data edges for possible dependencies on 'this'.
4574 for (uint i = 1; i < n->req(); i++) {
4575 Node* m = n->in(i);
4576 if (m == nullptr || m == n || m->is_top()) continue;
4577
4578 // Only process data inputs once
4579 worklist.push(m);
4580 }
4581 }
4582
4583 return true;
4584 }
4585
4586 // Here are all the checks a Store must pass before it can be moved into
4587 // an initialization. Returns zero if a check fails.
4588 // On success, returns the (constant) offset to which the store applies,
4589 // within the initialized memory.
4590 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4591 const int FAIL = 0;
4592 if (st->req() != MemNode::ValueIn + 1)
4593 return FAIL; // an inscrutable StoreNode (card mark?)
4594 Node* ctl = st->in(MemNode::Control);
4595 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4596 return FAIL; // must be unconditional after the initialization
4597 Node* mem = st->in(MemNode::Memory);
4598 if (!(mem->is_Proj() && mem->in(0) == this))
4599 return FAIL; // must not be preceded by other stores
4600 Node* adr = st->in(MemNode::Address);
4601 intptr_t offset;
4602 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4603 if (alloc == nullptr)
4604 return FAIL; // inscrutable address
4605 if (alloc != allocation())
4606 return FAIL; // wrong allocation! (store needs to float up)
4607 int size_in_bytes = st->memory_size();
4608 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4609 return FAIL; // mismatched access
4610 }
4611 Node* val = st->in(MemNode::ValueIn);
4612
4613 if (!detect_init_independence(val, phase))
4614 return FAIL; // stored value must be 'simple enough'
4615
4616 // The Store can be captured only if nothing after the allocation
4617 // and before the Store is using the memory location that the store
4618 // overwrites.
4619 bool failed = false;
4620 // If is_complete_with_arraycopy() is true the shape of the graph is
4621 // well defined and is safe so no need for extra checks.
4622 if (!is_complete_with_arraycopy()) {
4623 // We are going to look at each use of the memory state following
4624 // the allocation to make sure nothing reads the memory that the
4625 // Store writes.
4626 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4627 int alias_idx = phase->C->get_alias_index(t_adr);
4628 ResourceMark rm;
4629 Unique_Node_List mems;
4630 mems.push(mem);
4631 Node* unique_merge = nullptr;
4632 for (uint next = 0; next < mems.size(); ++next) {
4633 Node *m = mems.at(next);
4634 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4635 Node *n = m->fast_out(j);
4636 if (n->outcnt() == 0) {
4637 continue;
4638 }
4639 if (n == st) {
4640 continue;
4641 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4642 // If the control of this use is different from the control
4643 // of the Store which is right after the InitializeNode then
4644 // this node cannot be between the InitializeNode and the
4645 // Store.
4646 continue;
4647 } else if (n->is_MergeMem()) {
4648 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4649 // We can hit a MergeMemNode (that will likely go away
4650 // later) that is a direct use of the memory state
4651 // following the InitializeNode on the same slice as the
4652 // store node that we'd like to capture. We need to check
4653 // the uses of the MergeMemNode.
4654 mems.push(n);
4655 }
4656 } else if (n->is_Mem()) {
4657 Node* other_adr = n->in(MemNode::Address);
4658 if (other_adr == adr) {
4659 failed = true;
4660 break;
4661 } else {
4662 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4663 if (other_t_adr != nullptr) {
4664 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4665 if (other_alias_idx == alias_idx) {
4666 // A load from the same memory slice as the store right
4667 // after the InitializeNode. We check the control of the
4668 // object/array that is loaded from. If it's the same as
4669 // the store control then we cannot capture the store.
4670 assert(!n->is_Store(), "2 stores to same slice on same control?");
4671 Node* base = other_adr;
4672 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4673 base = base->in(AddPNode::Base);
4674 if (base != nullptr) {
4675 base = base->uncast();
4676 if (base->is_Proj() && base->in(0) == alloc) {
4677 failed = true;
4678 break;
4679 }
4680 }
4681 }
4682 }
4683 }
4684 } else {
4685 failed = true;
4686 break;
4687 }
4688 }
4689 }
4690 }
4691 if (failed) {
4692 if (!can_reshape) {
4693 // We decided we couldn't capture the store during parsing. We
4694 // should try again during the next IGVN once the graph is
4695 // cleaner.
4696 phase->C->record_for_igvn(st);
4697 }
4698 return FAIL;
4699 }
4700
4701 return offset; // success
4702 }
4703
4704 // Find the captured store in(i) which corresponds to the range
4705 // [start..start+size) in the initialized object.
4706 // If there is one, return its index i. If there isn't, return the
4707 // negative of the index where it should be inserted.
4708 // Return 0 if the queried range overlaps an initialization boundary
4709 // or if dead code is encountered.
4710 // If size_in_bytes is zero, do not bother with overlap checks.
4711 int InitializeNode::captured_store_insertion_point(intptr_t start,
4712 int size_in_bytes,
4713 PhaseValues* phase) {
4714 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
4715
4716 if (is_complete())
4717 return FAIL; // arraycopy got here first; punt
4718
4719 assert(allocation() != nullptr, "must be present");
4720
4721 // no negatives, no header fields:
4722 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
4723
4724 // after a certain size, we bail out on tracking all the stores:
4725 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4726 if (start >= ti_limit) return FAIL;
4727
4728 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4729 if (i >= limit) return -(int)i; // not found; here is where to put it
4730
4731 Node* st = in(i);
4732 intptr_t st_off = get_store_offset(st, phase);
4733 if (st_off < 0) {
4734 if (st != zero_memory()) {
4735 return FAIL; // bail out if there is dead garbage
4736 }
4737 } else if (st_off > start) {
4738 // ...we are done, since stores are ordered
4739 if (st_off < start + size_in_bytes) {
4740 return FAIL; // the next store overlaps
4741 }
4742 return -(int)i; // not found; here is where to put it
4743 } else if (st_off < start) {
4744 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
4745 if (size_in_bytes != 0 &&
4746 start < st_off + MAX_STORE &&
4747 start < st_off + st->as_Store()->memory_size()) {
4748 return FAIL; // the previous store overlaps
4749 }
4750 } else {
4751 if (size_in_bytes != 0 &&
4752 st->as_Store()->memory_size() != size_in_bytes) {
4753 return FAIL; // mismatched store size
4754 }
4755 return i;
4756 }
4757
4758 ++i;
4759 }
4760 }
4761
4762 // Look for a captured store which initializes at the offset 'start'
4763 // with the given size. If there is no such store, and no other
4764 // initialization interferes, then return zero_memory (the memory
4765 // projection of the AllocateNode).
4766 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
4767 PhaseValues* phase) {
4768 assert(stores_are_sane(phase), "");
4769 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4770 if (i == 0) {
4771 return nullptr; // something is dead
4772 } else if (i < 0) {
4773 return zero_memory(); // just primordial zero bits here
4774 } else {
4775 Node* st = in(i); // here is the store at this position
4776 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
4777 return st;
4778 }
4779 }
4780
4781 // Create, as a raw pointer, an address within my new object at 'offset'.
4782 Node* InitializeNode::make_raw_address(intptr_t offset,
4783 PhaseGVN* phase) {
4784 Node* addr = in(RawAddress);
4785 if (offset != 0) {
4786 Compile* C = phase->C;
4787 addr = phase->transform( new AddPNode(C->top(), addr,
4788 phase->MakeConX(offset)) );
4789 }
4790 return addr;
4791 }
4792
4793 // Clone the given store, converting it into a raw store
4794 // initializing a field or element of my new object.
4795 // Caller is responsible for retiring the original store,
4796 // with subsume_node or the like.
4797 //
4798 // From the example above InitializeNode::InitializeNode,
4799 // here are the old stores to be captured:
4800 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4801 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4802 //
4803 // Here is the changed code; note the extra edges on init:
4804 // alloc = (Allocate ...)
4805 // rawoop = alloc.RawAddress
4806 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
4807 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
4808 // init = (Initialize alloc.Control alloc.Memory rawoop
4809 // rawstore1 rawstore2)
4810 //
4811 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
4812 PhaseGVN* phase, bool can_reshape) {
4813 assert(stores_are_sane(phase), "");
4814
4815 if (start < 0) return nullptr;
4816 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
4817
4818 Compile* C = phase->C;
4819 int size_in_bytes = st->memory_size();
4820 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4821 if (i == 0) return nullptr; // bail out
4822 Node* prev_mem = nullptr; // raw memory for the captured store
4823 if (i > 0) {
4824 prev_mem = in(i); // there is a pre-existing store under this one
4825 set_req(i, C->top()); // temporarily disconnect it
4826 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
4827 } else {
4828 i = -i; // no pre-existing store
4829 prev_mem = zero_memory(); // a slice of the newly allocated object
4830 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
4831 set_req(--i, C->top()); // reuse this edge; it has been folded away
4832 else
4833 ins_req(i, C->top()); // build a new edge
4834 }
4835 Node* new_st = st->clone();
4836 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4837 new_st->set_req(MemNode::Control, in(Control));
4838 new_st->set_req(MemNode::Memory, prev_mem);
4839 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
4840 bs->eliminate_gc_barrier_data(new_st);
4841 new_st = phase->transform(new_st);
4842
4843 // At this point, new_st might have swallowed a pre-existing store
4844 // at the same offset, or perhaps new_st might have disappeared,
4845 // if it redundantly stored the same value (or zero to fresh memory).
4846
4847 // In any case, wire it in:
4848 PhaseIterGVN* igvn = phase->is_IterGVN();
4849 if (igvn) {
4850 igvn->rehash_node_delayed(this);
4851 }
4852 set_req(i, new_st);
4853
4854 // The caller may now kill the old guy.
4855 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
4856 assert(check_st == new_st || check_st == nullptr, "must be findable");
4857 assert(!is_complete(), "");
4858 return new_st;
4859 }
4860
4861 static bool store_constant(jlong* tiles, int num_tiles,
4862 intptr_t st_off, int st_size,
4863 jlong con) {
4864 if ((st_off & (st_size-1)) != 0)
4865 return false; // strange store offset (assume size==2**N)
4866 address addr = (address)tiles + st_off;
4867 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4868 switch (st_size) {
4869 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
4870 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
4871 case sizeof(jint): *(jint*) addr = (jint) con; break;
4872 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
4873 default: return false; // strange store size (detect size!=2**N here)
4874 }
4875 return true; // return success to caller
4876 }
4877
4878 // Coalesce subword constants into int constants and possibly
4879 // into long constants. The goal, if the CPU permits,
4880 // is to initialize the object with a small number of 64-bit tiles.
4881 // Also, convert floating-point constants to bit patterns.
4882 // Non-constants are not relevant to this pass.
4883 //
4884 // In terms of the running example on InitializeNode::InitializeNode
4885 // and InitializeNode::capture_store, here is the transformation
4886 // of rawstore1 and rawstore2 into rawstore12:
4887 // alloc = (Allocate ...)
4888 // rawoop = alloc.RawAddress
4889 // tile12 = 0x00010002
4890 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4891 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4892 //
4893 void
4894 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4895 Node* size_in_bytes,
4896 PhaseGVN* phase) {
4897 Compile* C = phase->C;
4898
4899 assert(stores_are_sane(phase), "");
4900 // Note: After this pass, they are not completely sane,
4901 // since there may be some overlaps.
4902
4903 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4904
4905 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4906 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4907 size_limit = MIN2(size_limit, ti_limit);
4908 size_limit = align_up(size_limit, BytesPerLong);
4909 int num_tiles = size_limit / BytesPerLong;
4910
4911 // allocate space for the tile map:
4912 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4913 jlong tiles_buf[small_len];
4914 Node* nodes_buf[small_len];
4915 jlong inits_buf[small_len];
4916 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4917 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4918 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4919 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4920 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4921 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4922 // tiles: exact bitwise model of all primitive constants
4923 // nodes: last constant-storing node subsumed into the tiles model
4924 // inits: which bytes (in each tile) are touched by any initializations
4925
4926 //// Pass A: Fill in the tile model with any relevant stores.
4927
4928 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4929 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4930 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4931 Node* zmem = zero_memory(); // initially zero memory state
4932 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4933 Node* st = in(i);
4934 intptr_t st_off = get_store_offset(st, phase);
4935
4936 // Figure out the store's offset and constant value:
4937 if (st_off < header_size) continue; //skip (ignore header)
4938 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
4939 int st_size = st->as_Store()->memory_size();
4940 if (st_off + st_size > size_limit) break;
4941
4942 // Record which bytes are touched, whether by constant or not.
4943 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4944 continue; // skip (strange store size)
4945
4946 const Type* val = phase->type(st->in(MemNode::ValueIn));
4947 if (!val->singleton()) continue; //skip (non-con store)
4948 BasicType type = val->basic_type();
4949
4950 jlong con = 0;
4951 switch (type) {
4952 case T_INT: con = val->is_int()->get_con(); break;
4953 case T_LONG: con = val->is_long()->get_con(); break;
4954 case T_FLOAT: con = jint_cast(val->getf()); break;
4955 case T_DOUBLE: con = jlong_cast(val->getd()); break;
4956 default: continue; //skip (odd store type)
4957 }
4958
4959 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4960 st->Opcode() == Op_StoreL) {
4961 continue; // This StoreL is already optimal.
4962 }
4963
4964 // Store down the constant.
4965 store_constant(tiles, num_tiles, st_off, st_size, con);
4966
4967 intptr_t j = st_off >> LogBytesPerLong;
4968
4969 if (type == T_INT && st_size == BytesPerInt
4970 && (st_off & BytesPerInt) == BytesPerInt) {
4971 jlong lcon = tiles[j];
4972 if (!Matcher::isSimpleConstant64(lcon) &&
4973 st->Opcode() == Op_StoreI) {
4974 // This StoreI is already optimal by itself.
4975 jint* intcon = (jint*) &tiles[j];
4976 intcon[1] = 0; // undo the store_constant()
4977
4978 // If the previous store is also optimal by itself, back up and
4979 // undo the action of the previous loop iteration... if we can.
4980 // But if we can't, just let the previous half take care of itself.
4981 st = nodes[j];
4982 st_off -= BytesPerInt;
4983 con = intcon[0];
4984 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
4985 assert(st_off >= header_size, "still ignoring header");
4986 assert(get_store_offset(st, phase) == st_off, "must be");
4987 assert(in(i-1) == zmem, "must be");
4988 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4989 assert(con == tcon->is_int()->get_con(), "must be");
4990 // Undo the effects of the previous loop trip, which swallowed st:
4991 intcon[0] = 0; // undo store_constant()
4992 set_req(i-1, st); // undo set_req(i, zmem)
4993 nodes[j] = nullptr; // undo nodes[j] = st
4994 --old_subword; // undo ++old_subword
4995 }
4996 continue; // This StoreI is already optimal.
4997 }
4998 }
4999
5000 // This store is not needed.
5001 set_req(i, zmem);
5002 nodes[j] = st; // record for the moment
5003 if (st_size < BytesPerLong) // something has changed
5004 ++old_subword; // includes int/float, but who's counting...
5005 else ++old_long;
5006 }
5007
5008 if ((old_subword + old_long) == 0)
5009 return; // nothing more to do
5010
5011 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5012 // Be sure to insert them before overlapping non-constant stores.
5013 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5014 for (int j = 0; j < num_tiles; j++) {
5015 jlong con = tiles[j];
5016 jlong init = inits[j];
5017 if (con == 0) continue;
5018 jint con0, con1; // split the constant, address-wise
5019 jint init0, init1; // split the init map, address-wise
5020 { union { jlong con; jint intcon[2]; } u;
5021 u.con = con;
5022 con0 = u.intcon[0];
5023 con1 = u.intcon[1];
5024 u.con = init;
5025 init0 = u.intcon[0];
5026 init1 = u.intcon[1];
5027 }
5028
5029 Node* old = nodes[j];
5030 assert(old != nullptr, "need the prior store");
5031 intptr_t offset = (j * BytesPerLong);
5032
5033 bool split = !Matcher::isSimpleConstant64(con);
5034
5035 if (offset < header_size) {
5036 assert(offset + BytesPerInt >= header_size, "second int counts");
5037 assert(*(jint*)&tiles[j] == 0, "junk in header");
5038 split = true; // only the second word counts
5039 // Example: int a[] = { 42 ... }
5040 } else if (con0 == 0 && init0 == -1) {
5041 split = true; // first word is covered by full inits
5042 // Example: int a[] = { ... foo(), 42 ... }
5043 } else if (con1 == 0 && init1 == -1) {
5044 split = true; // second word is covered by full inits
5045 // Example: int a[] = { ... 42, foo() ... }
5046 }
5047
5048 // Here's a case where init0 is neither 0 nor -1:
5049 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5050 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5051 // In this case the tile is not split; it is (jlong)42.
5052 // The big tile is stored down, and then the foo() value is inserted.
5053 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5054
5055 Node* ctl = old->in(MemNode::Control);
5056 Node* adr = make_raw_address(offset, phase);
5057 const TypePtr* atp = TypeRawPtr::BOTTOM;
5058
5059 // One or two coalesced stores to plop down.
5060 Node* st[2];
5061 intptr_t off[2];
5062 int nst = 0;
5063 if (!split) {
5064 ++new_long;
5065 off[nst] = offset;
5066 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5067 phase->longcon(con), T_LONG, MemNode::unordered);
5068 } else {
5069 // Omit either if it is a zero.
5070 if (con0 != 0) {
5071 ++new_int;
5072 off[nst] = offset;
5073 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5074 phase->intcon(con0), T_INT, MemNode::unordered);
5075 }
5076 if (con1 != 0) {
5077 ++new_int;
5078 offset += BytesPerInt;
5079 adr = make_raw_address(offset, phase);
5080 off[nst] = offset;
5081 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5082 phase->intcon(con1), T_INT, MemNode::unordered);
5083 }
5084 }
5085
5086 // Insert second store first, then the first before the second.
5087 // Insert each one just before any overlapping non-constant stores.
5088 while (nst > 0) {
5089 Node* st1 = st[--nst];
5090 C->copy_node_notes_to(st1, old);
5091 st1 = phase->transform(st1);
5092 offset = off[nst];
5093 assert(offset >= header_size, "do not smash header");
5094 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5095 guarantee(ins_idx != 0, "must re-insert constant store");
5096 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5097 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5098 set_req(--ins_idx, st1);
5099 else
5100 ins_req(ins_idx, st1);
5101 }
5102 }
5103
5104 if (PrintCompilation && WizardMode)
5105 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5106 old_subword, old_long, new_int, new_long);
5107 if (C->log() != nullptr)
5108 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5109 old_subword, old_long, new_int, new_long);
5110
5111 // Clean up any remaining occurrences of zmem:
5112 remove_extra_zeroes();
5113 }
5114
5115 // Explore forward from in(start) to find the first fully initialized
5116 // word, and return its offset. Skip groups of subword stores which
5117 // together initialize full words. If in(start) is itself part of a
5118 // fully initialized word, return the offset of in(start). If there
5119 // are no following full-word stores, or if something is fishy, return
5120 // a negative value.
5121 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5122 int int_map = 0;
5123 intptr_t int_map_off = 0;
5124 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5125
5126 for (uint i = start, limit = req(); i < limit; i++) {
5127 Node* st = in(i);
5128
5129 intptr_t st_off = get_store_offset(st, phase);
5130 if (st_off < 0) break; // return conservative answer
5131
5132 int st_size = st->as_Store()->memory_size();
5133 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5134 return st_off; // we found a complete word init
5135 }
5136
5137 // update the map:
5138
5139 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5140 if (this_int_off != int_map_off) {
5141 // reset the map:
5142 int_map = 0;
5143 int_map_off = this_int_off;
5144 }
5145
5146 int subword_off = st_off - this_int_off;
5147 int_map |= right_n_bits(st_size) << subword_off;
5148 if ((int_map & FULL_MAP) == FULL_MAP) {
5149 return this_int_off; // we found a complete word init
5150 }
5151
5152 // Did this store hit or cross the word boundary?
5153 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5154 if (next_int_off == this_int_off + BytesPerInt) {
5155 // We passed the current int, without fully initializing it.
5156 int_map_off = next_int_off;
5157 int_map >>= BytesPerInt;
5158 } else if (next_int_off > this_int_off + BytesPerInt) {
5159 // We passed the current and next int.
5160 return this_int_off + BytesPerInt;
5161 }
5162 }
5163
5164 return -1;
5165 }
5166
5167
5168 // Called when the associated AllocateNode is expanded into CFG.
5169 // At this point, we may perform additional optimizations.
5170 // Linearize the stores by ascending offset, to make memory
5171 // activity as coherent as possible.
5172 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5173 intptr_t header_size,
5174 Node* size_in_bytes,
5175 PhaseIterGVN* phase) {
5176 assert(!is_complete(), "not already complete");
5177 assert(stores_are_sane(phase), "");
5178 assert(allocation() != nullptr, "must be present");
5179
5180 remove_extra_zeroes();
5181
5182 if (ReduceFieldZeroing || ReduceBulkZeroing)
5183 // reduce instruction count for common initialization patterns
5184 coalesce_subword_stores(header_size, size_in_bytes, phase);
5185
5186 Node* zmem = zero_memory(); // initially zero memory state
5187 Node* inits = zmem; // accumulating a linearized chain of inits
5188 #ifdef ASSERT
5189 intptr_t first_offset = allocation()->minimum_header_size();
5190 intptr_t last_init_off = first_offset; // previous init offset
5191 intptr_t last_init_end = first_offset; // previous init offset+size
5192 intptr_t last_tile_end = first_offset; // previous tile offset+size
5193 #endif
5194 intptr_t zeroes_done = header_size;
5195
5196 bool do_zeroing = true; // we might give up if inits are very sparse
5197 int big_init_gaps = 0; // how many large gaps have we seen?
5198
5199 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5200 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5201
5202 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5203 Node* st = in(i);
5204 intptr_t st_off = get_store_offset(st, phase);
5205 if (st_off < 0)
5206 break; // unknown junk in the inits
5207 if (st->in(MemNode::Memory) != zmem)
5208 break; // complicated store chains somehow in list
5209
5210 int st_size = st->as_Store()->memory_size();
5211 intptr_t next_init_off = st_off + st_size;
5212
5213 if (do_zeroing && zeroes_done < next_init_off) {
5214 // See if this store needs a zero before it or under it.
5215 intptr_t zeroes_needed = st_off;
5216
5217 if (st_size < BytesPerInt) {
5218 // Look for subword stores which only partially initialize words.
5219 // If we find some, we must lay down some word-level zeroes first,
5220 // underneath the subword stores.
5221 //
5222 // Examples:
5223 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5224 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5225 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5226 //
5227 // Note: coalesce_subword_stores may have already done this,
5228 // if it was prompted by constant non-zero subword initializers.
5229 // But this case can still arise with non-constant stores.
5230
5231 intptr_t next_full_store = find_next_fullword_store(i, phase);
5232
5233 // In the examples above:
5234 // in(i) p q r s x y z
5235 // st_off 12 13 14 15 12 13 14
5236 // st_size 1 1 1 1 1 1 1
5237 // next_full_s. 12 16 16 16 16 16 16
5238 // z's_done 12 16 16 16 12 16 12
5239 // z's_needed 12 16 16 16 16 16 16
5240 // zsize 0 0 0 0 4 0 4
5241 if (next_full_store < 0) {
5242 // Conservative tack: Zero to end of current word.
5243 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5244 } else {
5245 // Zero to beginning of next fully initialized word.
5246 // Or, don't zero at all, if we are already in that word.
5247 assert(next_full_store >= zeroes_needed, "must go forward");
5248 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5249 zeroes_needed = next_full_store;
5250 }
5251 }
5252
5253 if (zeroes_needed > zeroes_done) {
5254 intptr_t zsize = zeroes_needed - zeroes_done;
5255 // Do some incremental zeroing on rawmem, in parallel with inits.
5256 zeroes_done = align_down(zeroes_done, BytesPerInt);
5257 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5258 zeroes_done, zeroes_needed,
5259 phase);
5260 zeroes_done = zeroes_needed;
5261 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5262 do_zeroing = false; // leave the hole, next time
5263 }
5264 }
5265
5266 // Collect the store and move on:
5267 phase->replace_input_of(st, MemNode::Memory, inits);
5268 inits = st; // put it on the linearized chain
5269 set_req(i, zmem); // unhook from previous position
5270
5271 if (zeroes_done == st_off)
5272 zeroes_done = next_init_off;
5273
5274 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5275
5276 #ifdef ASSERT
5277 // Various order invariants. Weaker than stores_are_sane because
5278 // a large constant tile can be filled in by smaller non-constant stores.
5279 assert(st_off >= last_init_off, "inits do not reverse");
5280 last_init_off = st_off;
5281 const Type* val = nullptr;
5282 if (st_size >= BytesPerInt &&
5283 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5284 (int)val->basic_type() < (int)T_OBJECT) {
5285 assert(st_off >= last_tile_end, "tiles do not overlap");
5286 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5287 last_tile_end = MAX2(last_tile_end, next_init_off);
5288 } else {
5289 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5290 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5291 assert(st_off >= last_init_end, "inits do not overlap");
5292 last_init_end = next_init_off; // it's a non-tile
5293 }
5294 #endif //ASSERT
5295 }
5296
5297 remove_extra_zeroes(); // clear out all the zmems left over
5298 add_req(inits);
5299
5300 if (!(UseTLAB && ZeroTLAB)) {
5301 // If anything remains to be zeroed, zero it all now.
5302 zeroes_done = align_down(zeroes_done, BytesPerInt);
5303 // if it is the last unused 4 bytes of an instance, forget about it
5304 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5305 if (zeroes_done + BytesPerLong >= size_limit) {
5306 AllocateNode* alloc = allocation();
5307 assert(alloc != nullptr, "must be present");
5308 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5309 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5310 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5311 if (zeroes_done == k->layout_helper())
5312 zeroes_done = size_limit;
5313 }
5314 }
5315 if (zeroes_done < size_limit) {
5316 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5317 zeroes_done, size_in_bytes, phase);
5318 }
5319 }
5320
5321 set_complete(phase);
5322 return rawmem;
5323 }
5324
5325
5326 #ifdef ASSERT
5327 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5328 if (is_complete())
5329 return true; // stores could be anything at this point
5330 assert(allocation() != nullptr, "must be present");
5331 intptr_t last_off = allocation()->minimum_header_size();
5332 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5333 Node* st = in(i);
5334 intptr_t st_off = get_store_offset(st, phase);
5335 if (st_off < 0) continue; // ignore dead garbage
5336 if (last_off > st_off) {
5337 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
5338 this->dump(2);
5339 assert(false, "ascending store offsets");
5340 return false;
5341 }
5342 last_off = st_off + st->as_Store()->memory_size();
5343 }
5344 return true;
5345 }
5346 #endif //ASSERT
5347
5348
5349
5350
5351 //============================MergeMemNode=====================================
5352 //
5353 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5354 // contributing store or call operations. Each contributor provides the memory
5355 // state for a particular "alias type" (see Compile::alias_type). For example,
5356 // if a MergeMem has an input X for alias category #6, then any memory reference
5357 // to alias category #6 may use X as its memory state input, as an exact equivalent
5358 // to using the MergeMem as a whole.
5359 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5360 //
5361 // (Here, the <N> notation gives the index of the relevant adr_type.)
5362 //
5363 // In one special case (and more cases in the future), alias categories overlap.
5364 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5365 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5366 // it is exactly equivalent to that state W:
5367 // MergeMem(<Bot>: W) <==> W
5368 //
5369 // Usually, the merge has more than one input. In that case, where inputs
5370 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5371 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5372 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5373 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5374 //
5375 // A merge can take a "wide" memory state as one of its narrow inputs.
5376 // This simply means that the merge observes out only the relevant parts of
5377 // the wide input. That is, wide memory states arriving at narrow merge inputs
5378 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5379 //
5380 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5381 // and that memory slices "leak through":
5382 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5383 //
5384 // But, in such a cascade, repeated memory slices can "block the leak":
5385 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5386 //
5387 // In the last example, Y is not part of the combined memory state of the
5388 // outermost MergeMem. The system must, of course, prevent unschedulable
5389 // memory states from arising, so you can be sure that the state Y is somehow
5390 // a precursor to state Y'.
5391 //
5392 //
5393 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5394 // of each MergeMemNode array are exactly the numerical alias indexes, including
5395 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5396 // Compile::alias_type (and kin) produce and manage these indexes.
5397 //
5398 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5399 // (Note that this provides quick access to the top node inside MergeMem methods,
5400 // without the need to reach out via TLS to Compile::current.)
5401 //
5402 // As a consequence of what was just described, a MergeMem that represents a full
5403 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5404 // containing all alias categories.
5405 //
5406 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5407 //
5408 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5409 // a memory state for the alias type <N>, or else the top node, meaning that
5410 // there is no particular input for that alias type. Note that the length of
5411 // a MergeMem is variable, and may be extended at any time to accommodate new
5412 // memory states at larger alias indexes. When merges grow, they are of course
5413 // filled with "top" in the unused in() positions.
5414 //
5415 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5416 // (Top was chosen because it works smoothly with passes like GCM.)
5417 //
5418 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5419 // the type of random VM bits like TLS references.) Since it is always the
5420 // first non-Bot memory slice, some low-level loops use it to initialize an
5421 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5422 //
5423 //
5424 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5425 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5426 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5427 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5428 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5429 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5430 //
5431 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5432 // really that different from the other memory inputs. An abbreviation called
5433 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5434 //
5435 //
5436 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5437 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5438 // that "emerges though" the base memory will be marked as excluding the alias types
5439 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5440 //
5441 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5442 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5443 //
5444 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5445 // (It is currently unimplemented.) As you can see, the resulting merge is
5446 // actually a disjoint union of memory states, rather than an overlay.
5447 //
5448
5449 //------------------------------MergeMemNode-----------------------------------
5450 Node* MergeMemNode::make_empty_memory() {
5451 Node* empty_memory = (Node*) Compile::current()->top();
5452 assert(empty_memory->is_top(), "correct sentinel identity");
5453 return empty_memory;
5454 }
5455
5456 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5457 init_class_id(Class_MergeMem);
5458 // all inputs are nullified in Node::Node(int)
5459 // set_input(0, nullptr); // no control input
5460
5461 // Initialize the edges uniformly to top, for starters.
5462 Node* empty_mem = make_empty_memory();
5463 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5464 init_req(i,empty_mem);
5465 }
5466 assert(empty_memory() == empty_mem, "");
5467
5468 if( new_base != nullptr && new_base->is_MergeMem() ) {
5469 MergeMemNode* mdef = new_base->as_MergeMem();
5470 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5471 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5472 mms.set_memory(mms.memory2());
5473 }
5474 assert(base_memory() == mdef->base_memory(), "");
5475 } else {
5476 set_base_memory(new_base);
5477 }
5478 }
5479
5480 // Make a new, untransformed MergeMem with the same base as 'mem'.
5481 // If mem is itself a MergeMem, populate the result with the same edges.
5482 MergeMemNode* MergeMemNode::make(Node* mem) {
5483 return new MergeMemNode(mem);
5484 }
5485
5486 //------------------------------cmp--------------------------------------------
5487 uint MergeMemNode::hash() const { return NO_HASH; }
5488 bool MergeMemNode::cmp( const Node &n ) const {
5489 return (&n == this); // Always fail except on self
5490 }
5491
5492 //------------------------------Identity---------------------------------------
5493 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5494 // Identity if this merge point does not record any interesting memory
5495 // disambiguations.
5496 Node* base_mem = base_memory();
5497 Node* empty_mem = empty_memory();
5498 if (base_mem != empty_mem) { // Memory path is not dead?
5499 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5500 Node* mem = in(i);
5501 if (mem != empty_mem && mem != base_mem) {
5502 return this; // Many memory splits; no change
5503 }
5504 }
5505 }
5506 return base_mem; // No memory splits; ID on the one true input
5507 }
5508
5509 //------------------------------Ideal------------------------------------------
5510 // This method is invoked recursively on chains of MergeMem nodes
5511 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5512 // Remove chain'd MergeMems
5513 //
5514 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5515 // relative to the "in(Bot)". Since we are patching both at the same time,
5516 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5517 // but rewrite each "in(i)" relative to the new "in(Bot)".
5518 Node *progress = nullptr;
5519
5520
5521 Node* old_base = base_memory();
5522 Node* empty_mem = empty_memory();
5523 if (old_base == empty_mem)
5524 return nullptr; // Dead memory path.
5525
5526 MergeMemNode* old_mbase;
5527 if (old_base != nullptr && old_base->is_MergeMem())
5528 old_mbase = old_base->as_MergeMem();
5529 else
5530 old_mbase = nullptr;
5531 Node* new_base = old_base;
5532
5533 // simplify stacked MergeMems in base memory
5534 if (old_mbase) new_base = old_mbase->base_memory();
5535
5536 // the base memory might contribute new slices beyond my req()
5537 if (old_mbase) grow_to_match(old_mbase);
5538
5539 // Note: We do not call verify_sparse on entry, because inputs
5540 // can normalize to the base_memory via subsume_node or similar
5541 // mechanisms. This method repairs that damage.
5542
5543 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5544
5545 // Look at each slice.
5546 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5547 Node* old_in = in(i);
5548 // calculate the old memory value
5549 Node* old_mem = old_in;
5550 if (old_mem == empty_mem) old_mem = old_base;
5551 assert(old_mem == memory_at(i), "");
5552
5553 // maybe update (reslice) the old memory value
5554
5555 // simplify stacked MergeMems
5556 Node* new_mem = old_mem;
5557 MergeMemNode* old_mmem;
5558 if (old_mem != nullptr && old_mem->is_MergeMem())
5559 old_mmem = old_mem->as_MergeMem();
5560 else
5561 old_mmem = nullptr;
5562 if (old_mmem == this) {
5563 // This can happen if loops break up and safepoints disappear.
5564 // A merge of BotPtr (default) with a RawPtr memory derived from a
5565 // safepoint can be rewritten to a merge of the same BotPtr with
5566 // the BotPtr phi coming into the loop. If that phi disappears
5567 // also, we can end up with a self-loop of the mergemem.
5568 // In general, if loops degenerate and memory effects disappear,
5569 // a mergemem can be left looking at itself. This simply means
5570 // that the mergemem's default should be used, since there is
5571 // no longer any apparent effect on this slice.
5572 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5573 // from start. Update the input to TOP.
5574 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5575 }
5576 else if (old_mmem != nullptr) {
5577 new_mem = old_mmem->memory_at(i);
5578 }
5579 // else preceding memory was not a MergeMem
5580
5581 // maybe store down a new value
5582 Node* new_in = new_mem;
5583 if (new_in == new_base) new_in = empty_mem;
5584
5585 if (new_in != old_in) {
5586 // Warning: Do not combine this "if" with the previous "if"
5587 // A memory slice might have be be rewritten even if it is semantically
5588 // unchanged, if the base_memory value has changed.
5589 set_req_X(i, new_in, phase);
5590 progress = this; // Report progress
5591 }
5592 }
5593
5594 if (new_base != old_base) {
5595 set_req_X(Compile::AliasIdxBot, new_base, phase);
5596 // Don't use set_base_memory(new_base), because we need to update du.
5597 assert(base_memory() == new_base, "");
5598 progress = this;
5599 }
5600
5601 if( base_memory() == this ) {
5602 // a self cycle indicates this memory path is dead
5603 set_req(Compile::AliasIdxBot, empty_mem);
5604 }
5605
5606 // Resolve external cycles by calling Ideal on a MergeMem base_memory
5607 // Recursion must occur after the self cycle check above
5608 if( base_memory()->is_MergeMem() ) {
5609 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5610 Node *m = phase->transform(new_mbase); // Rollup any cycles
5611 if( m != nullptr &&
5612 (m->is_top() ||
5613 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5614 // propagate rollup of dead cycle to self
5615 set_req(Compile::AliasIdxBot, empty_mem);
5616 }
5617 }
5618
5619 if( base_memory() == empty_mem ) {
5620 progress = this;
5621 // Cut inputs during Parse phase only.
5622 // During Optimize phase a dead MergeMem node will be subsumed by Top.
5623 if( !can_reshape ) {
5624 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5625 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5626 }
5627 }
5628 }
5629
5630 if( !progress && base_memory()->is_Phi() && can_reshape ) {
5631 // Check if PhiNode::Ideal's "Split phis through memory merges"
5632 // transform should be attempted. Look for this->phi->this cycle.
5633 uint merge_width = req();
5634 if (merge_width > Compile::AliasIdxRaw) {
5635 PhiNode* phi = base_memory()->as_Phi();
5636 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5637 if (phi->in(i) == this) {
5638 phase->is_IterGVN()->_worklist.push(phi);
5639 break;
5640 }
5641 }
5642 }
5643 }
5644
5645 assert(progress || verify_sparse(), "please, no dups of base");
5646 return progress;
5647 }
5648
5649 //-------------------------set_base_memory-------------------------------------
5650 void MergeMemNode::set_base_memory(Node *new_base) {
5651 Node* empty_mem = empty_memory();
5652 set_req(Compile::AliasIdxBot, new_base);
5653 assert(memory_at(req()) == new_base, "must set default memory");
5654 // Clear out other occurrences of new_base:
5655 if (new_base != empty_mem) {
5656 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5657 if (in(i) == new_base) set_req(i, empty_mem);
5658 }
5659 }
5660 }
5661
5662 //------------------------------out_RegMask------------------------------------
5663 const RegMask &MergeMemNode::out_RegMask() const {
5664 return RegMask::Empty;
5665 }
5666
5667 //------------------------------dump_spec--------------------------------------
5668 #ifndef PRODUCT
5669 void MergeMemNode::dump_spec(outputStream *st) const {
5670 st->print(" {");
5671 Node* base_mem = base_memory();
5672 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
5673 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
5674 if (mem == base_mem) { st->print(" -"); continue; }
5675 st->print( " N%d:", mem->_idx );
5676 Compile::current()->get_adr_type(i)->dump_on(st);
5677 }
5678 st->print(" }");
5679 }
5680 #endif // !PRODUCT
5681
5682
5683 #ifdef ASSERT
5684 static bool might_be_same(Node* a, Node* b) {
5685 if (a == b) return true;
5686 if (!(a->is_Phi() || b->is_Phi())) return false;
5687 // phis shift around during optimization
5688 return true; // pretty stupid...
5689 }
5690
5691 // verify a narrow slice (either incoming or outgoing)
5692 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
5693 if (!VerifyAliases) return; // don't bother to verify unless requested
5694 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
5695 if (Node::in_dump()) return; // muzzle asserts when printing
5696 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
5697 assert(n != nullptr, "");
5698 // Elide intervening MergeMem's
5699 while (n->is_MergeMem()) {
5700 n = n->as_MergeMem()->memory_at(alias_idx);
5701 }
5702 Compile* C = Compile::current();
5703 const TypePtr* n_adr_type = n->adr_type();
5704 if (n == m->empty_memory()) {
5705 // Implicit copy of base_memory()
5706 } else if (n_adr_type != TypePtr::BOTTOM) {
5707 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
5708 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
5709 } else {
5710 // A few places like make_runtime_call "know" that VM calls are narrow,
5711 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
5712 bool expected_wide_mem = false;
5713 if (n == m->base_memory()) {
5714 expected_wide_mem = true;
5715 } else if (alias_idx == Compile::AliasIdxRaw ||
5716 n == m->memory_at(Compile::AliasIdxRaw)) {
5717 expected_wide_mem = true;
5718 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
5719 // memory can "leak through" calls on channels that
5720 // are write-once. Allow this also.
5721 expected_wide_mem = true;
5722 }
5723 assert(expected_wide_mem, "expected narrow slice replacement");
5724 }
5725 }
5726 #else // !ASSERT
5727 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
5728 #endif
5729
5730
5731 //-----------------------------memory_at---------------------------------------
5732 Node* MergeMemNode::memory_at(uint alias_idx) const {
5733 assert(alias_idx >= Compile::AliasIdxRaw ||
5734 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
5735 "must avoid base_memory and AliasIdxTop");
5736
5737 // Otherwise, it is a narrow slice.
5738 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
5739 if (is_empty_memory(n)) {
5740 // the array is sparse; empty slots are the "top" node
5741 n = base_memory();
5742 assert(Node::in_dump()
5743 || n == nullptr || n->bottom_type() == Type::TOP
5744 || n->adr_type() == nullptr // address is TOP
5745 || n->adr_type() == TypePtr::BOTTOM
5746 || n->adr_type() == TypeRawPtr::BOTTOM
5747 || !Compile::current()->do_aliasing(),
5748 "must be a wide memory");
5749 // do_aliasing == false if we are organizing the memory states manually.
5750 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
5751 } else {
5752 // make sure the stored slice is sane
5753 #ifdef ASSERT
5754 if (VMError::is_error_reported() || Node::in_dump()) {
5755 } else if (might_be_same(n, base_memory())) {
5756 // Give it a pass: It is a mostly harmless repetition of the base.
5757 // This can arise normally from node subsumption during optimization.
5758 } else {
5759 verify_memory_slice(this, alias_idx, n);
5760 }
5761 #endif
5762 }
5763 return n;
5764 }
5765
5766 //---------------------------set_memory_at-------------------------------------
5767 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
5768 verify_memory_slice(this, alias_idx, n);
5769 Node* empty_mem = empty_memory();
5770 if (n == base_memory()) n = empty_mem; // collapse default
5771 uint need_req = alias_idx+1;
5772 if (req() < need_req) {
5773 if (n == empty_mem) return; // already the default, so do not grow me
5774 // grow the sparse array
5775 do {
5776 add_req(empty_mem);
5777 } while (req() < need_req);
5778 }
5779 set_req( alias_idx, n );
5780 }
5781
5782
5783
5784 //--------------------------iteration_setup------------------------------------
5785 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
5786 if (other != nullptr) {
5787 grow_to_match(other);
5788 // invariant: the finite support of mm2 is within mm->req()
5789 #ifdef ASSERT
5790 for (uint i = req(); i < other->req(); i++) {
5791 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
5792 }
5793 #endif
5794 }
5795 // Replace spurious copies of base_memory by top.
5796 Node* base_mem = base_memory();
5797 if (base_mem != nullptr && !base_mem->is_top()) {
5798 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
5799 if (in(i) == base_mem)
5800 set_req(i, empty_memory());
5801 }
5802 }
5803 }
5804
5805 //---------------------------grow_to_match-------------------------------------
5806 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
5807 Node* empty_mem = empty_memory();
5808 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
5809 // look for the finite support of the other memory
5810 for (uint i = other->req(); --i >= req(); ) {
5811 if (other->in(i) != empty_mem) {
5812 uint new_len = i+1;
5813 while (req() < new_len) add_req(empty_mem);
5814 break;
5815 }
5816 }
5817 }
5818
5819 //---------------------------verify_sparse-------------------------------------
5820 #ifndef PRODUCT
5821 bool MergeMemNode::verify_sparse() const {
5822 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
5823 Node* base_mem = base_memory();
5824 // The following can happen in degenerate cases, since empty==top.
5825 if (is_empty_memory(base_mem)) return true;
5826 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5827 assert(in(i) != nullptr, "sane slice");
5828 if (in(i) == base_mem) return false; // should have been the sentinel value!
5829 }
5830 return true;
5831 }
5832
5833 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5834 Node* n;
5835 n = mm->in(idx);
5836 if (mem == n) return true; // might be empty_memory()
5837 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5838 if (mem == n) return true;
5839 return false;
5840 }
5841 #endif // !PRODUCT