1 /*
2 * Copyright (c) 1997, 2018, 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/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/connode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/matcher.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/regmask.hpp"
47 #include "utilities/align.hpp"
48 #include "utilities/copy.hpp"
49 #include "utilities/macros.hpp"
50 #include "utilities/vmError.hpp"
51 #if INCLUDE_ZGC
52 #include "gc/z/c2/zBarrierSetC2.hpp"
53 #endif
54
55 // Portions of code courtesy of Clifford Click
56
57 // Optimization - Graph Style
58
59 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
60
61 //=============================================================================
62 uint MemNode::size_of() const { return sizeof(*this); }
63
64 const TypePtr *MemNode::adr_type() const {
65 Node* adr = in(Address);
66 if (adr == NULL) return NULL; // node is dead
67 const TypePtr* cross_check = NULL;
68 DEBUG_ONLY(cross_check = _adr_type);
69 return calculate_adr_type(adr->bottom_type(), cross_check);
70 }
71
72 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
73 if (adr != NULL) {
74 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
75 return true;
76 }
77 }
78 return false;
79 }
80
81 #ifndef PRODUCT
82 void MemNode::dump_spec(outputStream *st) const {
83 if (in(Address) == NULL) return; // node is dead
84 #ifndef ASSERT
85 // fake the missing field
86 const TypePtr* _adr_type = NULL;
87 if (in(Address) != NULL)
88 _adr_type = in(Address)->bottom_type()->isa_ptr();
89 #endif
90 dump_adr_type(this, _adr_type, st);
91
92 Compile* C = Compile::current();
93 if (C->alias_type(_adr_type)->is_volatile()) {
94 st->print(" Volatile!");
95 }
96 if (_unaligned_access) {
97 st->print(" unaligned");
98 }
99 if (_mismatched_access) {
100 st->print(" mismatched");
101 }
102 }
103
104 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
105 st->print(" @");
106 if (adr_type == NULL) {
107 st->print("NULL");
108 } else {
109 adr_type->dump_on(st);
110 Compile* C = Compile::current();
111 Compile::AliasType* atp = NULL;
112 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
113 if (atp == NULL)
114 st->print(", idx=?\?;");
115 else if (atp->index() == Compile::AliasIdxBot)
116 st->print(", idx=Bot;");
117 else if (atp->index() == Compile::AliasIdxTop)
118 st->print(", idx=Top;");
119 else if (atp->index() == Compile::AliasIdxRaw)
120 st->print(", idx=Raw;");
121 else {
122 ciField* field = atp->field();
123 if (field) {
124 st->print(", name=");
125 field->print_name_on(st);
126 }
127 st->print(", idx=%d;", atp->index());
128 }
129 }
130 }
131
132 extern void print_alias_types();
133
134 #endif
135
136 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
137 assert((t_oop != NULL), "sanity");
138 bool is_instance = t_oop->is_known_instance_field();
139 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
140 (load != NULL) && load->is_Load() &&
141 (phase->is_IterGVN() != NULL);
142 if (!(is_instance || is_boxed_value_load))
143 return mchain; // don't try to optimize non-instance types
144 uint instance_id = t_oop->instance_id();
145 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
146 Node *prev = NULL;
147 Node *result = mchain;
148 while (prev != result) {
149 prev = result;
150 if (result == start_mem)
151 break; // hit one of our sentinels
152 // skip over a call which does not affect this memory slice
153 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
154 Node *proj_in = result->in(0);
155 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
156 break; // hit one of our sentinels
157 } else if (proj_in->is_Call()) {
158 // ArrayCopyNodes processed here as well
159 CallNode *call = proj_in->as_Call();
160 if (!call->may_modify(t_oop, phase)) { // returns false for instances
161 result = call->in(TypeFunc::Memory);
162 }
163 } else if (proj_in->is_Initialize()) {
164 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
165 // Stop if this is the initialization for the object instance which
166 // contains this memory slice, otherwise skip over it.
167 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
168 break;
169 }
170 if (is_instance) {
171 result = proj_in->in(TypeFunc::Memory);
172 } else if (is_boxed_value_load) {
173 Node* klass = alloc->in(AllocateNode::KlassNode);
174 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
175 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
176 result = proj_in->in(TypeFunc::Memory); // not related allocation
177 }
178 }
179 } else if (proj_in->is_MemBar()) {
180 ArrayCopyNode* ac = NULL;
181 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
182 break;
183 }
184 result = proj_in->in(TypeFunc::Memory);
185 } else {
186 assert(false, "unexpected projection");
187 }
188 } else if (result->is_ClearArray()) {
189 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
190 // Can not bypass initialization of the instance
191 // we are looking for.
192 break;
193 }
194 // Otherwise skip it (the call updated 'result' value).
195 } else if (result->is_MergeMem()) {
196 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
197 }
198 }
199 return result;
200 }
201
202 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
203 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
204 if (t_oop == NULL)
205 return mchain; // don't try to optimize non-oop types
206 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
207 bool is_instance = t_oop->is_known_instance_field();
208 PhaseIterGVN *igvn = phase->is_IterGVN();
209 if (is_instance && igvn != NULL && result->is_Phi()) {
210 PhiNode *mphi = result->as_Phi();
211 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
212 const TypePtr *t = mphi->adr_type();
213 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
214 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
215 t->is_oopptr()->cast_to_exactness(true)
216 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
217 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
218 // clone the Phi with our address type
219 result = mphi->split_out_instance(t_adr, igvn);
220 } else {
221 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
222 }
223 }
224 return result;
225 }
226
227 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
228 uint alias_idx = phase->C->get_alias_index(tp);
229 Node *mem = mmem;
230 #ifdef ASSERT
231 {
232 // Check that current type is consistent with the alias index used during graph construction
233 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
234 bool consistent = adr_check == NULL || adr_check->empty() ||
235 phase->C->must_alias(adr_check, alias_idx );
236 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
237 if( !consistent && adr_check != NULL && !adr_check->empty() &&
238 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
239 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
240 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
241 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
242 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
243 // don't assert if it is dead code.
244 consistent = true;
245 }
246 if( !consistent ) {
247 st->print("alias_idx==%d, adr_check==", alias_idx);
248 if( adr_check == NULL ) {
249 st->print("NULL");
250 } else {
251 adr_check->dump();
252 }
253 st->cr();
254 print_alias_types();
255 assert(consistent, "adr_check must match alias idx");
256 }
257 }
258 #endif
259 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
260 // means an array I have not precisely typed yet. Do not do any
261 // alias stuff with it any time soon.
262 const TypeOopPtr *toop = tp->isa_oopptr();
263 if( tp->base() != Type::AnyPtr &&
264 !(toop &&
265 toop->klass() != NULL &&
266 toop->klass()->is_java_lang_Object() &&
267 toop->offset() == Type::OffsetBot) ) {
268 // compress paths and change unreachable cycles to TOP
269 // If not, we can update the input infinitely along a MergeMem cycle
270 // Equivalent code in PhiNode::Ideal
271 Node* m = phase->transform(mmem);
272 // If transformed to a MergeMem, get the desired slice
273 // Otherwise the returned node represents memory for every slice
274 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
275 // Update input if it is progress over what we have now
276 }
277 return mem;
278 }
279
280 //--------------------------Ideal_common---------------------------------------
281 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
282 // Unhook non-raw memories from complete (macro-expanded) initializations.
283 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
284 // If our control input is a dead region, kill all below the region
285 Node *ctl = in(MemNode::Control);
286 if (ctl && remove_dead_region(phase, can_reshape))
287 return this;
288 ctl = in(MemNode::Control);
289 // Don't bother trying to transform a dead node
290 if (ctl && ctl->is_top()) return NodeSentinel;
291
292 PhaseIterGVN *igvn = phase->is_IterGVN();
293 // Wait if control on the worklist.
294 if (ctl && can_reshape && igvn != NULL) {
295 Node* bol = NULL;
296 Node* cmp = NULL;
297 if (ctl->in(0)->is_If()) {
298 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
299 bol = ctl->in(0)->in(1);
300 if (bol->is_Bool())
301 cmp = ctl->in(0)->in(1)->in(1);
302 }
303 if (igvn->_worklist.member(ctl) ||
304 (bol != NULL && igvn->_worklist.member(bol)) ||
305 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
306 // This control path may be dead.
307 // Delay this memory node transformation until the control is processed.
308 phase->is_IterGVN()->_worklist.push(this);
309 return NodeSentinel; // caller will return NULL
310 }
311 }
312 // Ignore if memory is dead, or self-loop
313 Node *mem = in(MemNode::Memory);
314 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
315 assert(mem != this, "dead loop in MemNode::Ideal");
316
317 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
318 // This memory slice may be dead.
319 // Delay this mem node transformation until the memory is processed.
320 phase->is_IterGVN()->_worklist.push(this);
321 return NodeSentinel; // caller will return NULL
322 }
323
324 Node *address = in(MemNode::Address);
325 const Type *t_adr = phase->type(address);
326 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
327
328 if (can_reshape && igvn != NULL &&
329 (igvn->_worklist.member(address) ||
330 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
331 // The address's base and type may change when the address is processed.
332 // Delay this mem node transformation until the address is processed.
333 phase->is_IterGVN()->_worklist.push(this);
334 return NodeSentinel; // caller will return NULL
335 }
336
337 // Do NOT remove or optimize the next lines: ensure a new alias index
338 // is allocated for an oop pointer type before Escape Analysis.
339 // Note: C++ will not remove it since the call has side effect.
340 if (t_adr->isa_oopptr()) {
341 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
342 }
343
344 Node* base = NULL;
345 if (address->is_AddP()) {
346 base = address->in(AddPNode::Base);
347 }
348 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
349 !t_adr->isa_rawptr()) {
350 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
351 // Skip this node optimization if its address has TOP base.
352 return NodeSentinel; // caller will return NULL
353 }
354
355 // Avoid independent memory operations
356 Node* old_mem = mem;
357
358 // The code which unhooks non-raw memories from complete (macro-expanded)
359 // initializations was removed. After macro-expansion all stores catched
360 // by Initialize node became raw stores and there is no information
361 // which memory slices they modify. So it is unsafe to move any memory
362 // operation above these stores. Also in most cases hooked non-raw memories
363 // were already unhooked by using information from detect_ptr_independence()
364 // and find_previous_store().
365
366 if (mem->is_MergeMem()) {
367 MergeMemNode* mmem = mem->as_MergeMem();
368 const TypePtr *tp = t_adr->is_ptr();
369
370 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
371 }
372
373 if (mem != old_mem) {
374 set_req(MemNode::Memory, mem);
375 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
376 igvn->_worklist.push(old_mem);
377 }
378 if (phase->type(mem) == Type::TOP) return NodeSentinel;
379 return this;
380 }
381
382 // let the subclass continue analyzing...
383 return NULL;
384 }
385
386 // Helper function for proving some simple control dominations.
387 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
388 // Already assumes that 'dom' is available at 'sub', and that 'sub'
389 // is not a constant (dominated by the method's StartNode).
390 // Used by MemNode::find_previous_store to prove that the
391 // control input of a memory operation predates (dominates)
392 // an allocation it wants to look past.
393 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
394 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
395 return false; // Conservative answer for dead code
396
397 // Check 'dom'. Skip Proj and CatchProj nodes.
398 dom = dom->find_exact_control(dom);
399 if (dom == NULL || dom->is_top())
400 return false; // Conservative answer for dead code
401
402 if (dom == sub) {
403 // For the case when, for example, 'sub' is Initialize and the original
404 // 'dom' is Proj node of the 'sub'.
405 return false;
406 }
407
408 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
409 return true;
410
411 // 'dom' dominates 'sub' if its control edge and control edges
412 // of all its inputs dominate or equal to sub's control edge.
413
414 // Currently 'sub' is either Allocate, Initialize or Start nodes.
415 // Or Region for the check in LoadNode::Ideal();
416 // 'sub' should have sub->in(0) != NULL.
417 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
418 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
419
420 // Get control edge of 'sub'.
421 Node* orig_sub = sub;
422 sub = sub->find_exact_control(sub->in(0));
423 if (sub == NULL || sub->is_top())
424 return false; // Conservative answer for dead code
425
426 assert(sub->is_CFG(), "expecting control");
427
428 if (sub == dom)
429 return true;
430
431 if (sub->is_Start() || sub->is_Root())
432 return false;
433
434 {
435 // Check all control edges of 'dom'.
436
437 ResourceMark rm;
438 Arena* arena = Thread::current()->resource_area();
439 Node_List nlist(arena);
440 Unique_Node_List dom_list(arena);
441
442 dom_list.push(dom);
443 bool only_dominating_controls = false;
444
445 for (uint next = 0; next < dom_list.size(); next++) {
446 Node* n = dom_list.at(next);
447 if (n == orig_sub)
448 return false; // One of dom's inputs dominated by sub.
449 if (!n->is_CFG() && n->pinned()) {
450 // Check only own control edge for pinned non-control nodes.
451 n = n->find_exact_control(n->in(0));
452 if (n == NULL || n->is_top())
453 return false; // Conservative answer for dead code
454 assert(n->is_CFG(), "expecting control");
455 dom_list.push(n);
456 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
457 only_dominating_controls = true;
458 } else if (n->is_CFG()) {
459 if (n->dominates(sub, nlist))
460 only_dominating_controls = true;
461 else
462 return false;
463 } else {
464 // First, own control edge.
465 Node* m = n->find_exact_control(n->in(0));
466 if (m != NULL) {
467 if (m->is_top())
468 return false; // Conservative answer for dead code
469 dom_list.push(m);
470 }
471 // Now, the rest of edges.
472 uint cnt = n->req();
473 for (uint i = 1; i < cnt; i++) {
474 m = n->find_exact_control(n->in(i));
475 if (m == NULL || m->is_top())
476 continue;
477 dom_list.push(m);
478 }
479 }
480 }
481 return only_dominating_controls;
482 }
483 }
484
485 //---------------------detect_ptr_independence---------------------------------
486 // Used by MemNode::find_previous_store to prove that two base
487 // pointers are never equal.
488 // The pointers are accompanied by their associated allocations,
489 // if any, which have been previously discovered by the caller.
490 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
491 Node* p2, AllocateNode* a2,
492 PhaseTransform* phase) {
493 // Attempt to prove that these two pointers cannot be aliased.
494 // They may both manifestly be allocations, and they should differ.
495 // Or, if they are not both allocations, they can be distinct constants.
496 // Otherwise, one is an allocation and the other a pre-existing value.
497 if (a1 == NULL && a2 == NULL) { // neither an allocation
498 return (p1 != p2) && p1->is_Con() && p2->is_Con();
499 } else if (a1 != NULL && a2 != NULL) { // both allocations
500 return (a1 != a2);
501 } else if (a1 != NULL) { // one allocation a1
502 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
503 return all_controls_dominate(p2, a1);
504 } else { //(a2 != NULL) // one allocation a2
505 return all_controls_dominate(p1, a2);
506 }
507 return false;
508 }
509
510
511 // Find an arraycopy that must have set (can_see_stored_value=true) or
512 // could have set (can_see_stored_value=false) the value for this load
513 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
514 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
515 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
516 Node* mb = mem->in(0);
517 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
518 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
519 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
520 if (ac->is_clonebasic()) {
521 intptr_t offset;
522 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
523 if (alloc != NULL && alloc == ld_alloc) {
524 return ac;
525 }
526 }
527 }
528 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
529 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
530
531 if (ac->is_arraycopy_validated() ||
532 ac->is_copyof_validated() ||
533 ac->is_copyofrange_validated()) {
534 Node* ld_addp = in(MemNode::Address);
535 if (ld_addp->is_AddP()) {
536 Node* ld_base = ld_addp->in(AddPNode::Address);
537 Node* ld_offs = ld_addp->in(AddPNode::Offset);
538
539 Node* dest = ac->in(ArrayCopyNode::Dest);
540
541 if (dest == ld_base) {
542 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
543 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
544 return ac;
545 }
546 if (!can_see_stored_value) {
547 mem = ac->in(TypeFunc::Memory);
548 }
549 }
550 }
551 }
552 }
553 return NULL;
554 }
555
556 // The logic for reordering loads and stores uses four steps:
557 // (a) Walk carefully past stores and initializations which we
558 // can prove are independent of this load.
559 // (b) Observe that the next memory state makes an exact match
560 // with self (load or store), and locate the relevant store.
561 // (c) Ensure that, if we were to wire self directly to the store,
562 // the optimizer would fold it up somehow.
563 // (d) Do the rewiring, and return, depending on some other part of
564 // the optimizer to fold up the load.
565 // This routine handles steps (a) and (b). Steps (c) and (d) are
566 // specific to loads and stores, so they are handled by the callers.
567 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
568 //
569 Node* MemNode::find_previous_store(PhaseTransform* phase) {
570 Node* ctrl = in(MemNode::Control);
571 Node* adr = in(MemNode::Address);
572 intptr_t offset = 0;
573 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
574 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
575
576 if (offset == Type::OffsetBot)
577 return NULL; // cannot unalias unless there are precise offsets
578
579 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
580 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
581
582 intptr_t size_in_bytes = memory_size();
583
584 Node* mem = in(MemNode::Memory); // start searching here...
585
586 int cnt = 50; // Cycle limiter
587 for (;;) { // While we can dance past unrelated stores...
588 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
589
590 Node* prev = mem;
591 if (mem->is_Store()) {
592 Node* st_adr = mem->in(MemNode::Address);
593 intptr_t st_offset = 0;
594 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
595 if (st_base == NULL)
596 break; // inscrutable pointer
597
598 // For raw accesses it's not enough to prove that constant offsets don't intersect.
599 // We need the bases to be the equal in order for the offset check to make sense.
600 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
601 break;
602 }
603
604 if (st_offset != offset && st_offset != Type::OffsetBot) {
605 const int MAX_STORE = BytesPerLong;
606 if (st_offset >= offset + size_in_bytes ||
607 st_offset <= offset - MAX_STORE ||
608 st_offset <= offset - mem->as_Store()->memory_size()) {
609 // Success: The offsets are provably independent.
610 // (You may ask, why not just test st_offset != offset and be done?
611 // The answer is that stores of different sizes can co-exist
612 // in the same sequence of RawMem effects. We sometimes initialize
613 // a whole 'tile' of array elements with a single jint or jlong.)
614 mem = mem->in(MemNode::Memory);
615 continue; // (a) advance through independent store memory
616 }
617 }
618 if (st_base != base &&
619 detect_ptr_independence(base, alloc,
620 st_base,
621 AllocateNode::Ideal_allocation(st_base, phase),
622 phase)) {
623 // Success: The bases are provably independent.
624 mem = mem->in(MemNode::Memory);
625 continue; // (a) advance through independent store memory
626 }
627
628 // (b) At this point, if the bases or offsets do not agree, we lose,
629 // since we have not managed to prove 'this' and 'mem' independent.
630 if (st_base == base && st_offset == offset) {
631 return mem; // let caller handle steps (c), (d)
632 }
633
634 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
635 InitializeNode* st_init = mem->in(0)->as_Initialize();
636 AllocateNode* st_alloc = st_init->allocation();
637 if (st_alloc == NULL)
638 break; // something degenerated
639 bool known_identical = false;
640 bool known_independent = false;
641 if (alloc == st_alloc)
642 known_identical = true;
643 else if (alloc != NULL)
644 known_independent = true;
645 else if (all_controls_dominate(this, st_alloc))
646 known_independent = true;
647
648 if (known_independent) {
649 // The bases are provably independent: Either they are
650 // manifestly distinct allocations, or else the control
651 // of this load dominates the store's allocation.
652 int alias_idx = phase->C->get_alias_index(adr_type());
653 if (alias_idx == Compile::AliasIdxRaw) {
654 mem = st_alloc->in(TypeFunc::Memory);
655 } else {
656 mem = st_init->memory(alias_idx);
657 }
658 continue; // (a) advance through independent store memory
659 }
660
661 // (b) at this point, if we are not looking at a store initializing
662 // the same allocation we are loading from, we lose.
663 if (known_identical) {
664 // From caller, can_see_stored_value will consult find_captured_store.
665 return mem; // let caller handle steps (c), (d)
666 }
667
668 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
669 if (prev != mem) {
670 // Found an arraycopy but it doesn't affect that load
671 continue;
672 }
673 // Found an arraycopy that may affect that load
674 return mem;
675 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
676 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
677 if (mem->is_Proj() && mem->in(0)->is_Call()) {
678 // ArrayCopyNodes processed here as well.
679 CallNode *call = mem->in(0)->as_Call();
680 if (!call->may_modify(addr_t, phase)) {
681 mem = call->in(TypeFunc::Memory);
682 continue; // (a) advance through independent call memory
683 }
684 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
685 ArrayCopyNode* ac = NULL;
686 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
687 break;
688 }
689 mem = mem->in(0)->in(TypeFunc::Memory);
690 continue; // (a) advance through independent MemBar memory
691 } else if (mem->is_ClearArray()) {
692 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
693 // (the call updated 'mem' value)
694 continue; // (a) advance through independent allocation memory
695 } else {
696 // Can not bypass initialization of the instance
697 // we are looking for.
698 return mem;
699 }
700 } else if (mem->is_MergeMem()) {
701 int alias_idx = phase->C->get_alias_index(adr_type());
702 mem = mem->as_MergeMem()->memory_at(alias_idx);
703 continue; // (a) advance through independent MergeMem memory
704 }
705 }
706
707 // Unless there is an explicit 'continue', we must bail out here,
708 // because 'mem' is an inscrutable memory state (e.g., a call).
709 break;
710 }
711
712 return NULL; // bail out
713 }
714
715 //----------------------calculate_adr_type-------------------------------------
716 // Helper function. Notices when the given type of address hits top or bottom.
717 // Also, asserts a cross-check of the type against the expected address type.
718 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
719 if (t == Type::TOP) return NULL; // does not touch memory any more?
720 #ifdef PRODUCT
721 cross_check = NULL;
722 #else
723 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
724 #endif
725 const TypePtr* tp = t->isa_ptr();
726 if (tp == NULL) {
727 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
728 return TypePtr::BOTTOM; // touches lots of memory
729 } else {
730 #ifdef ASSERT
731 // %%%% [phh] We don't check the alias index if cross_check is
732 // TypeRawPtr::BOTTOM. Needs to be investigated.
733 if (cross_check != NULL &&
734 cross_check != TypePtr::BOTTOM &&
735 cross_check != TypeRawPtr::BOTTOM) {
736 // Recheck the alias index, to see if it has changed (due to a bug).
737 Compile* C = Compile::current();
738 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
739 "must stay in the original alias category");
740 // The type of the address must be contained in the adr_type,
741 // disregarding "null"-ness.
742 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
743 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
744 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
745 "real address must not escape from expected memory type");
746 }
747 #endif
748 return tp;
749 }
750 }
751
752 //=============================================================================
753 // Should LoadNode::Ideal() attempt to remove control edges?
754 bool LoadNode::can_remove_control() const {
755 return true;
756 }
757 uint LoadNode::size_of() const { return sizeof(*this); }
758 uint LoadNode::cmp( const Node &n ) const
759 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
760 const Type *LoadNode::bottom_type() const { return _type; }
761 uint LoadNode::ideal_reg() const {
762 return _type->ideal_reg();
763 }
764
765 #ifndef PRODUCT
766 void LoadNode::dump_spec(outputStream *st) const {
767 MemNode::dump_spec(st);
768 if( !Verbose && !WizardMode ) {
769 // standard dump does this in Verbose and WizardMode
770 st->print(" #"); _type->dump_on(st);
771 }
772 if (!depends_only_on_test()) {
773 st->print(" (does not depend only on test)");
774 }
775 }
776 #endif
777
778 #ifdef ASSERT
779 //----------------------------is_immutable_value-------------------------------
780 // Helper function to allow a raw load without control edge for some cases
781 bool LoadNode::is_immutable_value(Node* adr) {
782 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
783 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
784 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
785 in_bytes(JavaThread::osthread_offset())));
786 }
787 #endif
788
789 //----------------------------LoadNode::make-----------------------------------
790 // Polymorphic factory method:
791 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
792 ControlDependency control_dependency, bool unaligned, bool mismatched) {
793 Compile* C = gvn.C;
794
795 // sanity check the alias category against the created node type
796 assert(!(adr_type->isa_oopptr() &&
797 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
798 "use LoadKlassNode instead");
799 assert(!(adr_type->isa_aryptr() &&
800 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
801 "use LoadRangeNode instead");
802 // Check control edge of raw loads
803 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
804 // oop will be recorded in oop map if load crosses safepoint
805 rt->isa_oopptr() || is_immutable_value(adr),
806 "raw memory operations should have control edge");
807 LoadNode* load = NULL;
808 switch (bt) {
809 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
810 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
811 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
812 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
813 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
814 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
815 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
816 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
817 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
818 case T_OBJECT:
819 #ifdef _LP64
820 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
821 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
822 } else
823 #endif
824 {
825 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
826 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
827 }
828 break;
829 default:
830 ShouldNotReachHere();
831 break;
832 }
833 assert(load != NULL, "LoadNode should have been created");
834 if (unaligned) {
835 load->set_unaligned_access();
836 }
837 if (mismatched) {
838 load->set_mismatched_access();
839 }
840 if (load->Opcode() == Op_LoadN) {
841 Node* ld = gvn.transform(load);
842 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
843 }
844
845 return load;
846 }
847
848 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
849 ControlDependency control_dependency, bool unaligned, bool mismatched) {
850 bool require_atomic = true;
851 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
852 if (unaligned) {
853 load->set_unaligned_access();
854 }
855 if (mismatched) {
856 load->set_mismatched_access();
857 }
858 return load;
859 }
860
861 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
862 ControlDependency control_dependency, bool unaligned, bool mismatched) {
863 bool require_atomic = true;
864 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
865 if (unaligned) {
866 load->set_unaligned_access();
867 }
868 if (mismatched) {
869 load->set_mismatched_access();
870 }
871 return load;
872 }
873
874
875
876 //------------------------------hash-------------------------------------------
877 uint LoadNode::hash() const {
878 // unroll addition of interesting fields
879 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
880 }
881
882 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
883 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
884 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
885 bool is_stable_ary = FoldStableValues &&
886 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
887 tp->isa_aryptr()->is_stable();
888
889 return (eliminate_boxing && non_volatile) || is_stable_ary;
890 }
891
892 return false;
893 }
894
895 // Is the value loaded previously stored by an arraycopy? If so return
896 // a load node that reads from the source array so we may be able to
897 // optimize out the ArrayCopy node later.
898 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
899 #if INCLUDE_ZGC
900 if (UseZGC) {
901 if (bottom_type()->make_oopptr() != NULL) {
902 return NULL;
903 }
904 }
905 #endif
906
907 Node* ld_adr = in(MemNode::Address);
908 intptr_t ld_off = 0;
909 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
910 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
911 if (ac != NULL) {
912 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
913
914 Node* mem = ac->in(TypeFunc::Memory);
915 Node* ctl = ac->in(0);
916 Node* src = ac->in(ArrayCopyNode::Src);
917
918 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
919 return NULL;
920 }
921
922 LoadNode* ld = clone()->as_Load();
923 Node* addp = in(MemNode::Address)->clone();
924 if (ac->as_ArrayCopy()->is_clonebasic()) {
925 assert(ld_alloc != NULL, "need an alloc");
926 assert(addp->is_AddP(), "address must be addp");
927 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern");
928 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern");
929 addp->set_req(AddPNode::Base, src->in(AddPNode::Base));
930 addp->set_req(AddPNode::Address, src->in(AddPNode::Address));
931 } else {
932 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
933 ac->as_ArrayCopy()->is_copyof_validated() ||
934 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
935 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
936 addp->set_req(AddPNode::Base, src);
937 addp->set_req(AddPNode::Address, src);
938
939 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
940 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
941 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
942 uint shift = exact_log2(type2aelembytes(ary_elem));
943
944 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
945 #ifdef _LP64
946 diff = phase->transform(new ConvI2LNode(diff));
947 #endif
948 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
949
950 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
951 addp->set_req(AddPNode::Offset, offset);
952 }
953 addp = phase->transform(addp);
954 #ifdef ASSERT
955 const TypePtr* adr_type = phase->type(addp)->is_ptr();
956 ld->_adr_type = adr_type;
957 #endif
958 ld->set_req(MemNode::Address, addp);
959 ld->set_req(0, ctl);
960 ld->set_req(MemNode::Memory, mem);
961 // load depends on the tests that validate the arraycopy
962 ld->_control_dependency = Pinned;
963 return ld;
964 }
965 return NULL;
966 }
967
968
969 //---------------------------can_see_stored_value------------------------------
970 // This routine exists to make sure this set of tests is done the same
971 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
972 // will change the graph shape in a way which makes memory alive twice at the
973 // same time (uses the Oracle model of aliasing), then some
974 // LoadXNode::Identity will fold things back to the equivalence-class model
975 // of aliasing.
976 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
977 Node* ld_adr = in(MemNode::Address);
978 intptr_t ld_off = 0;
979 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
980 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
981 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
982 // This is more general than load from boxing objects.
983 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
984 uint alias_idx = atp->index();
985 bool final = !atp->is_rewritable();
986 Node* result = NULL;
987 Node* current = st;
988 // Skip through chains of MemBarNodes checking the MergeMems for
989 // new states for the slice of this load. Stop once any other
990 // kind of node is encountered. Loads from final memory can skip
991 // through any kind of MemBar but normal loads shouldn't skip
992 // through MemBarAcquire since the could allow them to move out of
993 // a synchronized region.
994 while (current->is_Proj()) {
995 int opc = current->in(0)->Opcode();
996 if ((final && (opc == Op_MemBarAcquire ||
997 opc == Op_MemBarAcquireLock ||
998 opc == Op_LoadFence)) ||
999 opc == Op_MemBarRelease ||
1000 opc == Op_StoreFence ||
1001 opc == Op_MemBarReleaseLock ||
1002 opc == Op_MemBarStoreStore ||
1003 opc == Op_MemBarCPUOrder) {
1004 Node* mem = current->in(0)->in(TypeFunc::Memory);
1005 if (mem->is_MergeMem()) {
1006 MergeMemNode* merge = mem->as_MergeMem();
1007 Node* new_st = merge->memory_at(alias_idx);
1008 if (new_st == merge->base_memory()) {
1009 // Keep searching
1010 current = new_st;
1011 continue;
1012 }
1013 // Save the new memory state for the slice and fall through
1014 // to exit.
1015 result = new_st;
1016 }
1017 }
1018 break;
1019 }
1020 if (result != NULL) {
1021 st = result;
1022 }
1023 }
1024
1025 // Loop around twice in the case Load -> Initialize -> Store.
1026 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1027 for (int trip = 0; trip <= 1; trip++) {
1028
1029 if (st->is_Store()) {
1030 Node* st_adr = st->in(MemNode::Address);
1031 if (!phase->eqv(st_adr, ld_adr)) {
1032 // Try harder before giving up... Match raw and non-raw pointers.
1033 intptr_t st_off = 0;
1034 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1035 if (alloc == NULL) return NULL;
1036 if (alloc != ld_alloc) return NULL;
1037 if (ld_off != st_off) return NULL;
1038 // At this point we have proven something like this setup:
1039 // A = Allocate(...)
1040 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
1041 // S = StoreQ(, AddP(, A.Parm , #Off), V)
1042 // (Actually, we haven't yet proven the Q's are the same.)
1043 // In other words, we are loading from a casted version of
1044 // the same pointer-and-offset that we stored to.
1045 // Thus, we are able to replace L by V.
1046 }
1047 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1048 if (store_Opcode() != st->Opcode())
1049 return NULL;
1050 return st->in(MemNode::ValueIn);
1051 }
1052
1053 // A load from a freshly-created object always returns zero.
1054 // (This can happen after LoadNode::Ideal resets the load's memory input
1055 // to find_captured_store, which returned InitializeNode::zero_memory.)
1056 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1057 (st->in(0) == ld_alloc) &&
1058 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1059 // return a zero value for the load's basic type
1060 // (This is one of the few places where a generic PhaseTransform
1061 // can create new nodes. Think of it as lazily manifesting
1062 // virtually pre-existing constants.)
1063 return phase->zerocon(memory_type());
1064 }
1065
1066 // A load from an initialization barrier can match a captured store.
1067 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1068 InitializeNode* init = st->in(0)->as_Initialize();
1069 AllocateNode* alloc = init->allocation();
1070 if ((alloc != NULL) && (alloc == ld_alloc)) {
1071 // examine a captured store value
1072 st = init->find_captured_store(ld_off, memory_size(), phase);
1073 if (st != NULL) {
1074 continue; // take one more trip around
1075 }
1076 }
1077 }
1078
1079 // Load boxed value from result of valueOf() call is input parameter.
1080 if (this->is_Load() && ld_adr->is_AddP() &&
1081 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1082 intptr_t ignore = 0;
1083 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1084 if (base != NULL && base->is_Proj() &&
1085 base->as_Proj()->_con == TypeFunc::Parms &&
1086 base->in(0)->is_CallStaticJava() &&
1087 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1088 return base->in(0)->in(TypeFunc::Parms);
1089 }
1090 }
1091
1092 break;
1093 }
1094
1095 return NULL;
1096 }
1097
1098 //----------------------is_instance_field_load_with_local_phi------------------
1099 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1100 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1101 in(Address)->is_AddP() ) {
1102 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1103 // Only instances and boxed values.
1104 if( t_oop != NULL &&
1105 (t_oop->is_ptr_to_boxed_value() ||
1106 t_oop->is_known_instance_field()) &&
1107 t_oop->offset() != Type::OffsetBot &&
1108 t_oop->offset() != Type::OffsetTop) {
1109 return true;
1110 }
1111 }
1112 return false;
1113 }
1114
1115 //------------------------------Identity---------------------------------------
1116 // Loads are identity if previous store is to same address
1117 Node* LoadNode::Identity(PhaseGVN* phase) {
1118 // If the previous store-maker is the right kind of Store, and the store is
1119 // to the same address, then we are equal to the value stored.
1120 Node* mem = in(Memory);
1121 Node* value = can_see_stored_value(mem, phase);
1122 if( value ) {
1123 // byte, short & char stores truncate naturally.
1124 // A load has to load the truncated value which requires
1125 // some sort of masking operation and that requires an
1126 // Ideal call instead of an Identity call.
1127 if (memory_size() < BytesPerInt) {
1128 // If the input to the store does not fit with the load's result type,
1129 // it must be truncated via an Ideal call.
1130 if (!phase->type(value)->higher_equal(phase->type(this)))
1131 return this;
1132 }
1133 // (This works even when value is a Con, but LoadNode::Value
1134 // usually runs first, producing the singleton type of the Con.)
1135 return value;
1136 }
1137
1138 // Search for an existing data phi which was generated before for the same
1139 // instance's field to avoid infinite generation of phis in a loop.
1140 Node *region = mem->in(0);
1141 if (is_instance_field_load_with_local_phi(region)) {
1142 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1143 int this_index = phase->C->get_alias_index(addr_t);
1144 int this_offset = addr_t->offset();
1145 int this_iid = addr_t->instance_id();
1146 if (!addr_t->is_known_instance() &&
1147 addr_t->is_ptr_to_boxed_value()) {
1148 // Use _idx of address base (could be Phi node) for boxed values.
1149 intptr_t ignore = 0;
1150 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1151 if (base == NULL) {
1152 return this;
1153 }
1154 this_iid = base->_idx;
1155 }
1156 const Type* this_type = bottom_type();
1157 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1158 Node* phi = region->fast_out(i);
1159 if (phi->is_Phi() && phi != mem &&
1160 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1161 return phi;
1162 }
1163 }
1164 }
1165
1166 return this;
1167 }
1168
1169 // Construct an equivalent unsigned load.
1170 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1171 BasicType bt = T_ILLEGAL;
1172 const Type* rt = NULL;
1173 switch (Opcode()) {
1174 case Op_LoadUB: return this;
1175 case Op_LoadUS: return this;
1176 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1177 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1178 default:
1179 assert(false, "no unsigned variant: %s", Name());
1180 return NULL;
1181 }
1182 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1183 raw_adr_type(), rt, bt, _mo, _control_dependency,
1184 is_unaligned_access(), is_mismatched_access());
1185 }
1186
1187 // Construct an equivalent signed load.
1188 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1189 BasicType bt = T_ILLEGAL;
1190 const Type* rt = NULL;
1191 switch (Opcode()) {
1192 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1193 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1194 case Op_LoadB: // fall through
1195 case Op_LoadS: // fall through
1196 case Op_LoadI: // fall through
1197 case Op_LoadL: return this;
1198 default:
1199 assert(false, "no signed variant: %s", Name());
1200 return NULL;
1201 }
1202 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1203 raw_adr_type(), rt, bt, _mo, _control_dependency,
1204 is_unaligned_access(), is_mismatched_access());
1205 }
1206
1207 // We're loading from an object which has autobox behaviour.
1208 // If this object is result of a valueOf call we'll have a phi
1209 // merging a newly allocated object and a load from the cache.
1210 // We want to replace this load with the original incoming
1211 // argument to the valueOf call.
1212 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1213 assert(phase->C->eliminate_boxing(), "sanity");
1214 intptr_t ignore = 0;
1215 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1216 if ((base == NULL) || base->is_Phi()) {
1217 // Push the loads from the phi that comes from valueOf up
1218 // through it to allow elimination of the loads and the recovery
1219 // of the original value. It is done in split_through_phi().
1220 return NULL;
1221 } else if (base->is_Load() ||
1222 (base->is_DecodeN() && base->in(1)->is_Load())) {
1223 // Eliminate the load of boxed value for integer types from the cache
1224 // array by deriving the value from the index into the array.
1225 // Capture the offset of the load and then reverse the computation.
1226
1227 // Get LoadN node which loads a boxing object from 'cache' array.
1228 if (base->is_DecodeN()) {
1229 base = base->in(1);
1230 }
1231 if (!base->in(Address)->is_AddP()) {
1232 return NULL; // Complex address
1233 }
1234 AddPNode* address = base->in(Address)->as_AddP();
1235 Node* cache_base = address->in(AddPNode::Base);
1236 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1237 // Get ConP node which is static 'cache' field.
1238 cache_base = cache_base->in(1);
1239 }
1240 if ((cache_base != NULL) && cache_base->is_Con()) {
1241 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1242 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1243 Node* elements[4];
1244 int shift = exact_log2(type2aelembytes(T_OBJECT));
1245 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1246 if (count > 0 && elements[0]->is_Con() &&
1247 (count == 1 ||
1248 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1249 elements[1]->in(2) == phase->intcon(shift)))) {
1250 ciObjArray* array = base_type->const_oop()->as_obj_array();
1251 // Fetch the box object cache[0] at the base of the array and get its value
1252 ciInstance* box = array->obj_at(0)->as_instance();
1253 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1254 assert(ik->is_box_klass(), "sanity");
1255 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1256 if (ik->nof_nonstatic_fields() == 1) {
1257 // This should be true nonstatic_field_at requires calling
1258 // nof_nonstatic_fields so check it anyway
1259 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1260 BasicType bt = c.basic_type();
1261 // Only integer types have boxing cache.
1262 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1263 bt == T_BYTE || bt == T_SHORT ||
1264 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1265 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1266 if (cache_low != (int)cache_low) {
1267 return NULL; // should not happen since cache is array indexed by value
1268 }
1269 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1270 if (offset != (int)offset) {
1271 return NULL; // should not happen since cache is array indexed by value
1272 }
1273 // Add up all the offsets making of the address of the load
1274 Node* result = elements[0];
1275 for (int i = 1; i < count; i++) {
1276 result = phase->transform(new AddXNode(result, elements[i]));
1277 }
1278 // Remove the constant offset from the address and then
1279 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1280 // remove the scaling of the offset to recover the original index.
1281 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1282 // Peel the shift off directly but wrap it in a dummy node
1283 // since Ideal can't return existing nodes
1284 result = new RShiftXNode(result->in(1), phase->intcon(0));
1285 } else if (result->is_Add() && result->in(2)->is_Con() &&
1286 result->in(1)->Opcode() == Op_LShiftX &&
1287 result->in(1)->in(2) == phase->intcon(shift)) {
1288 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1289 // but for boxing cache access we know that X<<Z will not overflow
1290 // (there is range check) so we do this optimizatrion by hand here.
1291 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1292 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1293 } else {
1294 result = new RShiftXNode(result, phase->intcon(shift));
1295 }
1296 #ifdef _LP64
1297 if (bt != T_LONG) {
1298 result = new ConvL2INode(phase->transform(result));
1299 }
1300 #else
1301 if (bt == T_LONG) {
1302 result = new ConvI2LNode(phase->transform(result));
1303 }
1304 #endif
1305 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1306 // Need to preserve unboxing load type if it is unsigned.
1307 switch(this->Opcode()) {
1308 case Op_LoadUB:
1309 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1310 break;
1311 case Op_LoadUS:
1312 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1313 break;
1314 }
1315 return result;
1316 }
1317 }
1318 }
1319 }
1320 }
1321 return NULL;
1322 }
1323
1324 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1325 Node* region = phi->in(0);
1326 if (region == NULL) {
1327 return false; // Wait stable graph
1328 }
1329 uint cnt = phi->req();
1330 for (uint i = 1; i < cnt; i++) {
1331 Node* rc = region->in(i);
1332 if (rc == NULL || phase->type(rc) == Type::TOP)
1333 return false; // Wait stable graph
1334 Node* in = phi->in(i);
1335 if (in == NULL || phase->type(in) == Type::TOP)
1336 return false; // Wait stable graph
1337 }
1338 return true;
1339 }
1340 //------------------------------split_through_phi------------------------------
1341 // Split instance or boxed field load through Phi.
1342 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1343 Node* mem = in(Memory);
1344 Node* address = in(Address);
1345 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1346
1347 assert((t_oop != NULL) &&
1348 (t_oop->is_known_instance_field() ||
1349 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1350
1351 Compile* C = phase->C;
1352 intptr_t ignore = 0;
1353 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1354 bool base_is_phi = (base != NULL) && base->is_Phi();
1355 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1356 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1357 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1358
1359 if (!((mem->is_Phi() || base_is_phi) &&
1360 (load_boxed_values || t_oop->is_known_instance_field()))) {
1361 return NULL; // memory is not Phi
1362 }
1363
1364 if (mem->is_Phi()) {
1365 if (!stable_phi(mem->as_Phi(), phase)) {
1366 return NULL; // Wait stable graph
1367 }
1368 uint cnt = mem->req();
1369 // Check for loop invariant memory.
1370 if (cnt == 3) {
1371 for (uint i = 1; i < cnt; i++) {
1372 Node* in = mem->in(i);
1373 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1374 if (m == mem) {
1375 set_req(Memory, mem->in(cnt - i));
1376 return this; // made change
1377 }
1378 }
1379 }
1380 }
1381 if (base_is_phi) {
1382 if (!stable_phi(base->as_Phi(), phase)) {
1383 return NULL; // Wait stable graph
1384 }
1385 uint cnt = base->req();
1386 // Check for loop invariant memory.
1387 if (cnt == 3) {
1388 for (uint i = 1; i < cnt; i++) {
1389 if (base->in(i) == base) {
1390 return NULL; // Wait stable graph
1391 }
1392 }
1393 }
1394 }
1395
1396 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1397
1398 // Split through Phi (see original code in loopopts.cpp).
1399 assert(C->have_alias_type(t_oop), "instance should have alias type");
1400
1401 // Do nothing here if Identity will find a value
1402 // (to avoid infinite chain of value phis generation).
1403 if (!phase->eqv(this, this->Identity(phase)))
1404 return NULL;
1405
1406 // Select Region to split through.
1407 Node* region;
1408 if (!base_is_phi) {
1409 assert(mem->is_Phi(), "sanity");
1410 region = mem->in(0);
1411 // Skip if the region dominates some control edge of the address.
1412 if (!MemNode::all_controls_dominate(address, region))
1413 return NULL;
1414 } else if (!mem->is_Phi()) {
1415 assert(base_is_phi, "sanity");
1416 region = base->in(0);
1417 // Skip if the region dominates some control edge of the memory.
1418 if (!MemNode::all_controls_dominate(mem, region))
1419 return NULL;
1420 } else if (base->in(0) != mem->in(0)) {
1421 assert(base_is_phi && mem->is_Phi(), "sanity");
1422 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1423 region = base->in(0);
1424 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1425 region = mem->in(0);
1426 } else {
1427 return NULL; // complex graph
1428 }
1429 } else {
1430 assert(base->in(0) == mem->in(0), "sanity");
1431 region = mem->in(0);
1432 }
1433
1434 const Type* this_type = this->bottom_type();
1435 int this_index = C->get_alias_index(t_oop);
1436 int this_offset = t_oop->offset();
1437 int this_iid = t_oop->instance_id();
1438 if (!t_oop->is_known_instance() && load_boxed_values) {
1439 // Use _idx of address base for boxed values.
1440 this_iid = base->_idx;
1441 }
1442 PhaseIterGVN* igvn = phase->is_IterGVN();
1443 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1444 for (uint i = 1; i < region->req(); i++) {
1445 Node* x;
1446 Node* the_clone = NULL;
1447 if (region->in(i) == C->top()) {
1448 x = C->top(); // Dead path? Use a dead data op
1449 } else {
1450 x = this->clone(); // Else clone up the data op
1451 the_clone = x; // Remember for possible deletion.
1452 // Alter data node to use pre-phi inputs
1453 if (this->in(0) == region) {
1454 x->set_req(0, region->in(i));
1455 } else {
1456 x->set_req(0, NULL);
1457 }
1458 if (mem->is_Phi() && (mem->in(0) == region)) {
1459 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1460 }
1461 if (address->is_Phi() && address->in(0) == region) {
1462 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1463 }
1464 if (base_is_phi && (base->in(0) == region)) {
1465 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1466 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1467 x->set_req(Address, adr_x);
1468 }
1469 }
1470 // Check for a 'win' on some paths
1471 const Type *t = x->Value(igvn);
1472
1473 bool singleton = t->singleton();
1474
1475 // See comments in PhaseIdealLoop::split_thru_phi().
1476 if (singleton && t == Type::TOP) {
1477 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1478 }
1479
1480 if (singleton) {
1481 x = igvn->makecon(t);
1482 } else {
1483 // We now call Identity to try to simplify the cloned node.
1484 // Note that some Identity methods call phase->type(this).
1485 // Make sure that the type array is big enough for
1486 // our new node, even though we may throw the node away.
1487 // (This tweaking with igvn only works because x is a new node.)
1488 igvn->set_type(x, t);
1489 // If x is a TypeNode, capture any more-precise type permanently into Node
1490 // otherwise it will be not updated during igvn->transform since
1491 // igvn->type(x) is set to x->Value() already.
1492 x->raise_bottom_type(t);
1493 Node *y = x->Identity(igvn);
1494 if (y != x) {
1495 x = y;
1496 } else {
1497 y = igvn->hash_find_insert(x);
1498 if (y) {
1499 x = y;
1500 } else {
1501 // Else x is a new node we are keeping
1502 // We do not need register_new_node_with_optimizer
1503 // because set_type has already been called.
1504 igvn->_worklist.push(x);
1505 }
1506 }
1507 }
1508 if (x != the_clone && the_clone != NULL) {
1509 igvn->remove_dead_node(the_clone);
1510 }
1511 phi->set_req(i, x);
1512 }
1513 // Record Phi
1514 igvn->register_new_node_with_optimizer(phi);
1515 return phi;
1516 }
1517
1518 //------------------------------Ideal------------------------------------------
1519 // If the load is from Field memory and the pointer is non-null, it might be possible to
1520 // zero out the control input.
1521 // If the offset is constant and the base is an object allocation,
1522 // try to hook me up to the exact initializing store.
1523 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1524 Node* p = MemNode::Ideal_common(phase, can_reshape);
1525 if (p) return (p == NodeSentinel) ? NULL : p;
1526
1527 Node* ctrl = in(MemNode::Control);
1528 Node* address = in(MemNode::Address);
1529 bool progress = false;
1530
1531 // Skip up past a SafePoint control. Cannot do this for Stores because
1532 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1533 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1534 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1535 ctrl = ctrl->in(0);
1536 set_req(MemNode::Control,ctrl);
1537 progress = true;
1538 }
1539
1540 intptr_t ignore = 0;
1541 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1542 if (base != NULL
1543 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1544 // Check for useless control edge in some common special cases
1545 if (in(MemNode::Control) != NULL
1546 && can_remove_control()
1547 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1548 && all_controls_dominate(base, phase->C->start())) {
1549 // A method-invariant, non-null address (constant or 'this' argument).
1550 set_req(MemNode::Control, NULL);
1551 progress = true;
1552 }
1553 }
1554
1555 Node* mem = in(MemNode::Memory);
1556 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1557
1558 if (can_reshape && (addr_t != NULL)) {
1559 // try to optimize our memory input
1560 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1561 if (opt_mem != mem) {
1562 set_req(MemNode::Memory, opt_mem);
1563 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1564 return this;
1565 }
1566 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1567 if ((t_oop != NULL) &&
1568 (t_oop->is_known_instance_field() ||
1569 t_oop->is_ptr_to_boxed_value())) {
1570 PhaseIterGVN *igvn = phase->is_IterGVN();
1571 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1572 // Delay this transformation until memory Phi is processed.
1573 phase->is_IterGVN()->_worklist.push(this);
1574 return NULL;
1575 }
1576 // Split instance field load through Phi.
1577 Node* result = split_through_phi(phase);
1578 if (result != NULL) return result;
1579
1580 if (t_oop->is_ptr_to_boxed_value()) {
1581 Node* result = eliminate_autobox(phase);
1582 if (result != NULL) return result;
1583 }
1584 }
1585 }
1586
1587 // Is there a dominating load that loads the same value? Leave
1588 // anything that is not a load of a field/array element (like
1589 // barriers etc.) alone
1590 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1591 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1592 Node *use = mem->fast_out(i);
1593 if (use != this &&
1594 use->Opcode() == Opcode() &&
1595 use->in(0) != NULL &&
1596 use->in(0) != in(0) &&
1597 use->in(Address) == in(Address)) {
1598 Node* ctl = in(0);
1599 for (int i = 0; i < 10 && ctl != NULL; i++) {
1600 ctl = IfNode::up_one_dom(ctl);
1601 if (ctl == use->in(0)) {
1602 set_req(0, use->in(0));
1603 return this;
1604 }
1605 }
1606 }
1607 }
1608 }
1609
1610 // Check for prior store with a different base or offset; make Load
1611 // independent. Skip through any number of them. Bail out if the stores
1612 // are in an endless dead cycle and report no progress. This is a key
1613 // transform for Reflection. However, if after skipping through the Stores
1614 // we can't then fold up against a prior store do NOT do the transform as
1615 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1616 // array memory alive twice: once for the hoisted Load and again after the
1617 // bypassed Store. This situation only works if EVERYBODY who does
1618 // anti-dependence work knows how to bypass. I.e. we need all
1619 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1620 // the alias index stuff. So instead, peek through Stores and IFF we can
1621 // fold up, do so.
1622 Node* prev_mem = find_previous_store(phase);
1623 if (prev_mem != NULL) {
1624 Node* value = can_see_arraycopy_value(prev_mem, phase);
1625 if (value != NULL) {
1626 return value;
1627 }
1628 }
1629 // Steps (a), (b): Walk past independent stores to find an exact match.
1630 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1631 // (c) See if we can fold up on the spot, but don't fold up here.
1632 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1633 // just return a prior value, which is done by Identity calls.
1634 if (can_see_stored_value(prev_mem, phase)) {
1635 // Make ready for step (d):
1636 set_req(MemNode::Memory, prev_mem);
1637 return this;
1638 }
1639 }
1640
1641 return progress ? this : NULL;
1642 }
1643
1644 // Helper to recognize certain Klass fields which are invariant across
1645 // some group of array types (e.g., int[] or all T[] where T < Object).
1646 const Type*
1647 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1648 ciKlass* klass) const {
1649 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1650 // The field is Klass::_modifier_flags. Return its (constant) value.
1651 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1652 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1653 return TypeInt::make(klass->modifier_flags());
1654 }
1655 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1656 // The field is Klass::_access_flags. Return its (constant) value.
1657 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1658 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1659 return TypeInt::make(klass->access_flags());
1660 }
1661 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1662 // The field is Klass::_layout_helper. Return its constant value if known.
1663 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1664 return TypeInt::make(klass->layout_helper());
1665 }
1666
1667 // No match.
1668 return NULL;
1669 }
1670
1671 //------------------------------Value-----------------------------------------
1672 const Type* LoadNode::Value(PhaseGVN* phase) const {
1673 // Either input is TOP ==> the result is TOP
1674 Node* mem = in(MemNode::Memory);
1675 const Type *t1 = phase->type(mem);
1676 if (t1 == Type::TOP) return Type::TOP;
1677 Node* adr = in(MemNode::Address);
1678 const TypePtr* tp = phase->type(adr)->isa_ptr();
1679 if (tp == NULL || tp->empty()) return Type::TOP;
1680 int off = tp->offset();
1681 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1682 Compile* C = phase->C;
1683
1684 // Try to guess loaded type from pointer type
1685 if (tp->isa_aryptr()) {
1686 const TypeAryPtr* ary = tp->is_aryptr();
1687 const Type* t = ary->elem();
1688
1689 // Determine whether the reference is beyond the header or not, by comparing
1690 // the offset against the offset of the start of the array's data.
1691 // Different array types begin at slightly different offsets (12 vs. 16).
1692 // We choose T_BYTE as an example base type that is least restrictive
1693 // as to alignment, which will therefore produce the smallest
1694 // possible base offset.
1695 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1696 const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1697
1698 // Try to constant-fold a stable array element.
1699 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1700 // Make sure the reference is not into the header and the offset is constant
1701 ciObject* aobj = ary->const_oop();
1702 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1703 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1704 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1705 stable_dimension,
1706 memory_type(), is_unsigned());
1707 if (con_type != NULL) {
1708 return con_type;
1709 }
1710 }
1711 }
1712
1713 // Don't do this for integer types. There is only potential profit if
1714 // the element type t is lower than _type; that is, for int types, if _type is
1715 // more restrictive than t. This only happens here if one is short and the other
1716 // char (both 16 bits), and in those cases we've made an intentional decision
1717 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1718 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1719 //
1720 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1721 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1722 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1723 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1724 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1725 // In fact, that could have been the original type of p1, and p1 could have
1726 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1727 // expression (LShiftL quux 3) independently optimized to the constant 8.
1728 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1729 && (_type->isa_vect() == NULL)
1730 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1731 // t might actually be lower than _type, if _type is a unique
1732 // concrete subclass of abstract class t.
1733 if (off_beyond_header) { // is the offset beyond the header?
1734 const Type* jt = t->join_speculative(_type);
1735 // In any case, do not allow the join, per se, to empty out the type.
1736 if (jt->empty() && !t->empty()) {
1737 // This can happen if a interface-typed array narrows to a class type.
1738 jt = _type;
1739 }
1740 #ifdef ASSERT
1741 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1742 // The pointers in the autobox arrays are always non-null
1743 Node* base = adr->in(AddPNode::Base);
1744 if ((base != NULL) && base->is_DecodeN()) {
1745 // Get LoadN node which loads IntegerCache.cache field
1746 base = base->in(1);
1747 }
1748 if ((base != NULL) && base->is_Con()) {
1749 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1750 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1751 // It could be narrow oop
1752 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1753 }
1754 }
1755 }
1756 #endif
1757 return jt;
1758 }
1759 }
1760 } else if (tp->base() == Type::InstPtr) {
1761 assert( off != Type::OffsetBot ||
1762 // arrays can be cast to Objects
1763 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1764 // unsafe field access may not have a constant offset
1765 C->has_unsafe_access(),
1766 "Field accesses must be precise" );
1767 // For oop loads, we expect the _type to be precise.
1768
1769 // Optimize loads from constant fields.
1770 const TypeInstPtr* tinst = tp->is_instptr();
1771 ciObject* const_oop = tinst->const_oop();
1772 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1773 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1774 if (con_type != NULL) {
1775 return con_type;
1776 }
1777 }
1778 } else if (tp->base() == Type::KlassPtr) {
1779 assert( off != Type::OffsetBot ||
1780 // arrays can be cast to Objects
1781 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1782 // also allow array-loading from the primary supertype
1783 // array during subtype checks
1784 Opcode() == Op_LoadKlass,
1785 "Field accesses must be precise" );
1786 // For klass/static loads, we expect the _type to be precise
1787 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
1788 /* With mirrors being an indirect in the Klass*
1789 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1790 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1791 *
1792 * So check the type and klass of the node before the LoadP.
1793 */
1794 Node* adr2 = adr->in(MemNode::Address);
1795 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1796 if (tkls != NULL && !StressReflectiveCode) {
1797 ciKlass* klass = tkls->klass();
1798 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1799 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1800 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1801 return TypeInstPtr::make(klass->java_mirror());
1802 }
1803 }
1804 }
1805
1806 const TypeKlassPtr *tkls = tp->isa_klassptr();
1807 if (tkls != NULL && !StressReflectiveCode) {
1808 ciKlass* klass = tkls->klass();
1809 if (klass->is_loaded() && tkls->klass_is_exact()) {
1810 // We are loading a field from a Klass metaobject whose identity
1811 // is known at compile time (the type is "exact" or "precise").
1812 // Check for fields we know are maintained as constants by the VM.
1813 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1814 // The field is Klass::_super_check_offset. Return its (constant) value.
1815 // (Folds up type checking code.)
1816 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1817 return TypeInt::make(klass->super_check_offset());
1818 }
1819 // Compute index into primary_supers array
1820 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1821 // Check for overflowing; use unsigned compare to handle the negative case.
1822 if( depth < ciKlass::primary_super_limit() ) {
1823 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1824 // (Folds up type checking code.)
1825 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1826 ciKlass *ss = klass->super_of_depth(depth);
1827 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1828 }
1829 const Type* aift = load_array_final_field(tkls, klass);
1830 if (aift != NULL) return aift;
1831 }
1832
1833 // We can still check if we are loading from the primary_supers array at a
1834 // shallow enough depth. Even though the klass is not exact, entries less
1835 // than or equal to its super depth are correct.
1836 if (klass->is_loaded() ) {
1837 ciType *inner = klass;
1838 while( inner->is_obj_array_klass() )
1839 inner = inner->as_obj_array_klass()->base_element_type();
1840 if( inner->is_instance_klass() &&
1841 !inner->as_instance_klass()->flags().is_interface() ) {
1842 // Compute index into primary_supers array
1843 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1844 // Check for overflowing; use unsigned compare to handle the negative case.
1845 if( depth < ciKlass::primary_super_limit() &&
1846 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1847 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1848 // (Folds up type checking code.)
1849 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1850 ciKlass *ss = klass->super_of_depth(depth);
1851 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1852 }
1853 }
1854 }
1855
1856 // If the type is enough to determine that the thing is not an array,
1857 // we can give the layout_helper a positive interval type.
1858 // This will help short-circuit some reflective code.
1859 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1860 && !klass->is_array_klass() // not directly typed as an array
1861 && !klass->is_interface() // specifically not Serializable & Cloneable
1862 && !klass->is_java_lang_Object() // not the supertype of all T[]
1863 ) {
1864 // Note: When interfaces are reliable, we can narrow the interface
1865 // test to (klass != Serializable && klass != Cloneable).
1866 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1867 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1868 // The key property of this type is that it folds up tests
1869 // for array-ness, since it proves that the layout_helper is positive.
1870 // Thus, a generic value like the basic object layout helper works fine.
1871 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1872 }
1873 }
1874
1875 // If we are loading from a freshly-allocated object, produce a zero,
1876 // if the load is provably beyond the header of the object.
1877 // (Also allow a variable load from a fresh array to produce zero.)
1878 const TypeOopPtr *tinst = tp->isa_oopptr();
1879 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1880 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1881 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1882 Node* value = can_see_stored_value(mem,phase);
1883 if (value != NULL && value->is_Con()) {
1884 assert(value->bottom_type()->higher_equal(_type),"sanity");
1885 return value->bottom_type();
1886 }
1887 }
1888
1889 if (is_instance) {
1890 // If we have an instance type and our memory input is the
1891 // programs's initial memory state, there is no matching store,
1892 // so just return a zero of the appropriate type
1893 Node *mem = in(MemNode::Memory);
1894 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1895 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1896 return Type::get_zero_type(_type->basic_type());
1897 }
1898 }
1899 return _type;
1900 }
1901
1902 //------------------------------match_edge-------------------------------------
1903 // Do we Match on this edge index or not? Match only the address.
1904 uint LoadNode::match_edge(uint idx) const {
1905 return idx == MemNode::Address;
1906 }
1907
1908 //--------------------------LoadBNode::Ideal--------------------------------------
1909 //
1910 // If the previous store is to the same address as this load,
1911 // and the value stored was larger than a byte, replace this load
1912 // with the value stored truncated to a byte. If no truncation is
1913 // needed, the replacement is done in LoadNode::Identity().
1914 //
1915 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1916 Node* mem = in(MemNode::Memory);
1917 Node* value = can_see_stored_value(mem,phase);
1918 if( value && !phase->type(value)->higher_equal( _type ) ) {
1919 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1920 return new RShiftINode(result, phase->intcon(24));
1921 }
1922 // Identity call will handle the case where truncation is not needed.
1923 return LoadNode::Ideal(phase, can_reshape);
1924 }
1925
1926 const Type* LoadBNode::Value(PhaseGVN* phase) const {
1927 Node* mem = in(MemNode::Memory);
1928 Node* value = can_see_stored_value(mem,phase);
1929 if (value != NULL && value->is_Con() &&
1930 !value->bottom_type()->higher_equal(_type)) {
1931 // If the input to the store does not fit with the load's result type,
1932 // it must be truncated. We can't delay until Ideal call since
1933 // a singleton Value is needed for split_thru_phi optimization.
1934 int con = value->get_int();
1935 return TypeInt::make((con << 24) >> 24);
1936 }
1937 return LoadNode::Value(phase);
1938 }
1939
1940 //--------------------------LoadUBNode::Ideal-------------------------------------
1941 //
1942 // If the previous store is to the same address as this load,
1943 // and the value stored was larger than a byte, replace this load
1944 // with the value stored truncated to a byte. If no truncation is
1945 // needed, the replacement is done in LoadNode::Identity().
1946 //
1947 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1948 Node* mem = in(MemNode::Memory);
1949 Node* value = can_see_stored_value(mem, phase);
1950 if (value && !phase->type(value)->higher_equal(_type))
1951 return new AndINode(value, phase->intcon(0xFF));
1952 // Identity call will handle the case where truncation is not needed.
1953 return LoadNode::Ideal(phase, can_reshape);
1954 }
1955
1956 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
1957 Node* mem = in(MemNode::Memory);
1958 Node* value = can_see_stored_value(mem,phase);
1959 if (value != NULL && value->is_Con() &&
1960 !value->bottom_type()->higher_equal(_type)) {
1961 // If the input to the store does not fit with the load's result type,
1962 // it must be truncated. We can't delay until Ideal call since
1963 // a singleton Value is needed for split_thru_phi optimization.
1964 int con = value->get_int();
1965 return TypeInt::make(con & 0xFF);
1966 }
1967 return LoadNode::Value(phase);
1968 }
1969
1970 //--------------------------LoadUSNode::Ideal-------------------------------------
1971 //
1972 // If the previous store is to the same address as this load,
1973 // and the value stored was larger than a char, replace this load
1974 // with the value stored truncated to a char. If no truncation is
1975 // needed, the replacement is done in LoadNode::Identity().
1976 //
1977 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1978 Node* mem = in(MemNode::Memory);
1979 Node* value = can_see_stored_value(mem,phase);
1980 if( value && !phase->type(value)->higher_equal( _type ) )
1981 return new AndINode(value,phase->intcon(0xFFFF));
1982 // Identity call will handle the case where truncation is not needed.
1983 return LoadNode::Ideal(phase, can_reshape);
1984 }
1985
1986 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
1987 Node* mem = in(MemNode::Memory);
1988 Node* value = can_see_stored_value(mem,phase);
1989 if (value != NULL && value->is_Con() &&
1990 !value->bottom_type()->higher_equal(_type)) {
1991 // If the input to the store does not fit with the load's result type,
1992 // it must be truncated. We can't delay until Ideal call since
1993 // a singleton Value is needed for split_thru_phi optimization.
1994 int con = value->get_int();
1995 return TypeInt::make(con & 0xFFFF);
1996 }
1997 return LoadNode::Value(phase);
1998 }
1999
2000 //--------------------------LoadSNode::Ideal--------------------------------------
2001 //
2002 // If the previous store is to the same address as this load,
2003 // and the value stored was larger than a short, replace this load
2004 // with the value stored truncated to a short. If no truncation is
2005 // needed, the replacement is done in LoadNode::Identity().
2006 //
2007 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2008 Node* mem = in(MemNode::Memory);
2009 Node* value = can_see_stored_value(mem,phase);
2010 if( value && !phase->type(value)->higher_equal( _type ) ) {
2011 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2012 return new RShiftINode(result, phase->intcon(16));
2013 }
2014 // Identity call will handle the case where truncation is not needed.
2015 return LoadNode::Ideal(phase, can_reshape);
2016 }
2017
2018 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2019 Node* mem = in(MemNode::Memory);
2020 Node* value = can_see_stored_value(mem,phase);
2021 if (value != NULL && value->is_Con() &&
2022 !value->bottom_type()->higher_equal(_type)) {
2023 // If the input to the store does not fit with the load's result type,
2024 // it must be truncated. We can't delay until Ideal call since
2025 // a singleton Value is needed for split_thru_phi optimization.
2026 int con = value->get_int();
2027 return TypeInt::make((con << 16) >> 16);
2028 }
2029 return LoadNode::Value(phase);
2030 }
2031
2032 //=============================================================================
2033 //----------------------------LoadKlassNode::make------------------------------
2034 // Polymorphic factory method:
2035 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2036 // sanity check the alias category against the created node type
2037 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2038 assert(adr_type != NULL, "expecting TypeKlassPtr");
2039 #ifdef _LP64
2040 if (adr_type->is_ptr_to_narrowklass()) {
2041 assert(UseCompressedClassPointers, "no compressed klasses");
2042 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2043 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2044 }
2045 #endif
2046 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2047 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2048 }
2049
2050 //------------------------------Value------------------------------------------
2051 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2052 return klass_value_common(phase);
2053 }
2054
2055 // In most cases, LoadKlassNode does not have the control input set. If the control
2056 // input is set, it must not be removed (by LoadNode::Ideal()).
2057 bool LoadKlassNode::can_remove_control() const {
2058 return false;
2059 }
2060
2061 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2062 // Either input is TOP ==> the result is TOP
2063 const Type *t1 = phase->type( in(MemNode::Memory) );
2064 if (t1 == Type::TOP) return Type::TOP;
2065 Node *adr = in(MemNode::Address);
2066 const Type *t2 = phase->type( adr );
2067 if (t2 == Type::TOP) return Type::TOP;
2068 const TypePtr *tp = t2->is_ptr();
2069 if (TypePtr::above_centerline(tp->ptr()) ||
2070 tp->ptr() == TypePtr::Null) return Type::TOP;
2071
2072 // Return a more precise klass, if possible
2073 const TypeInstPtr *tinst = tp->isa_instptr();
2074 if (tinst != NULL) {
2075 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2076 int offset = tinst->offset();
2077 if (ik == phase->C->env()->Class_klass()
2078 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2079 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2080 // We are loading a special hidden field from a Class mirror object,
2081 // the field which points to the VM's Klass metaobject.
2082 ciType* t = tinst->java_mirror_type();
2083 // java_mirror_type returns non-null for compile-time Class constants.
2084 if (t != NULL) {
2085 // constant oop => constant klass
2086 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2087 if (t->is_void()) {
2088 // We cannot create a void array. Since void is a primitive type return null
2089 // klass. Users of this result need to do a null check on the returned klass.
2090 return TypePtr::NULL_PTR;
2091 }
2092 return TypeKlassPtr::make(ciArrayKlass::make(t));
2093 }
2094 if (!t->is_klass()) {
2095 // a primitive Class (e.g., int.class) has NULL for a klass field
2096 return TypePtr::NULL_PTR;
2097 }
2098 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2099 return TypeKlassPtr::make(t->as_klass());
2100 }
2101 // non-constant mirror, so we can't tell what's going on
2102 }
2103 if( !ik->is_loaded() )
2104 return _type; // Bail out if not loaded
2105 if (offset == oopDesc::klass_offset_in_bytes()) {
2106 if (tinst->klass_is_exact()) {
2107 return TypeKlassPtr::make(ik);
2108 }
2109 // See if we can become precise: no subklasses and no interface
2110 // (Note: We need to support verified interfaces.)
2111 if (!ik->is_interface() && !ik->has_subklass()) {
2112 //assert(!UseExactTypes, "this code should be useless with exact types");
2113 // Add a dependence; if any subclass added we need to recompile
2114 if (!ik->is_final()) {
2115 // %%% should use stronger assert_unique_concrete_subtype instead
2116 phase->C->dependencies()->assert_leaf_type(ik);
2117 }
2118 // Return precise klass
2119 return TypeKlassPtr::make(ik);
2120 }
2121
2122 // Return root of possible klass
2123 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2124 }
2125 }
2126
2127 // Check for loading klass from an array
2128 const TypeAryPtr *tary = tp->isa_aryptr();
2129 if( tary != NULL ) {
2130 ciKlass *tary_klass = tary->klass();
2131 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2132 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2133 if (tary->klass_is_exact()) {
2134 return TypeKlassPtr::make(tary_klass);
2135 }
2136 ciArrayKlass *ak = tary->klass()->as_array_klass();
2137 // If the klass is an object array, we defer the question to the
2138 // array component klass.
2139 if( ak->is_obj_array_klass() ) {
2140 assert( ak->is_loaded(), "" );
2141 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2142 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2143 ciInstanceKlass* ik = base_k->as_instance_klass();
2144 // See if we can become precise: no subklasses and no interface
2145 if (!ik->is_interface() && !ik->has_subklass()) {
2146 //assert(!UseExactTypes, "this code should be useless with exact types");
2147 // Add a dependence; if any subclass added we need to recompile
2148 if (!ik->is_final()) {
2149 phase->C->dependencies()->assert_leaf_type(ik);
2150 }
2151 // Return precise array klass
2152 return TypeKlassPtr::make(ak);
2153 }
2154 }
2155 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2156 } else { // Found a type-array?
2157 //assert(!UseExactTypes, "this code should be useless with exact types");
2158 assert( ak->is_type_array_klass(), "" );
2159 return TypeKlassPtr::make(ak); // These are always precise
2160 }
2161 }
2162 }
2163
2164 // Check for loading klass from an array klass
2165 const TypeKlassPtr *tkls = tp->isa_klassptr();
2166 if (tkls != NULL && !StressReflectiveCode) {
2167 ciKlass* klass = tkls->klass();
2168 if( !klass->is_loaded() )
2169 return _type; // Bail out if not loaded
2170 if( klass->is_obj_array_klass() &&
2171 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2172 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2173 // // Always returning precise element type is incorrect,
2174 // // e.g., element type could be object and array may contain strings
2175 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2176
2177 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2178 // according to the element type's subclassing.
2179 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2180 }
2181 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2182 tkls->offset() == in_bytes(Klass::super_offset())) {
2183 ciKlass* sup = klass->as_instance_klass()->super();
2184 // The field is Klass::_super. Return its (constant) value.
2185 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2186 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2187 }
2188 }
2189
2190 // Bailout case
2191 return LoadNode::Value(phase);
2192 }
2193
2194 //------------------------------Identity---------------------------------------
2195 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2196 // Also feed through the klass in Allocate(...klass...)._klass.
2197 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2198 return klass_identity_common(phase);
2199 }
2200
2201 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2202 Node* x = LoadNode::Identity(phase);
2203 if (x != this) return x;
2204
2205 // Take apart the address into an oop and and offset.
2206 // Return 'this' if we cannot.
2207 Node* adr = in(MemNode::Address);
2208 intptr_t offset = 0;
2209 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2210 if (base == NULL) return this;
2211 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2212 if (toop == NULL) return this;
2213
2214 // Step over potential GC barrier for OopHandle resolve
2215 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2216 if (bs->is_gc_barrier_node(base)) {
2217 base = bs->step_over_gc_barrier(base);
2218 }
2219
2220 // We can fetch the klass directly through an AllocateNode.
2221 // This works even if the klass is not constant (clone or newArray).
2222 if (offset == oopDesc::klass_offset_in_bytes()) {
2223 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2224 if (allocated_klass != NULL) {
2225 return allocated_klass;
2226 }
2227 }
2228
2229 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2230 // See inline_native_Class_query for occurrences of these patterns.
2231 // Java Example: x.getClass().isAssignableFrom(y)
2232 //
2233 // This improves reflective code, often making the Class
2234 // mirror go completely dead. (Current exception: Class
2235 // mirrors may appear in debug info, but we could clean them out by
2236 // introducing a new debug info operator for Klass.java_mirror).
2237
2238 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2239 && offset == java_lang_Class::klass_offset_in_bytes()) {
2240 if (base->is_Load()) {
2241 Node* base2 = base->in(MemNode::Address);
2242 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2243 Node* adr2 = base2->in(MemNode::Address);
2244 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2245 if (tkls != NULL && !tkls->empty()
2246 && (tkls->klass()->is_instance_klass() ||
2247 tkls->klass()->is_array_klass())
2248 && adr2->is_AddP()
2249 ) {
2250 int mirror_field = in_bytes(Klass::java_mirror_offset());
2251 if (tkls->offset() == mirror_field) {
2252 return adr2->in(AddPNode::Base);
2253 }
2254 }
2255 }
2256 }
2257 }
2258
2259 return this;
2260 }
2261
2262
2263 //------------------------------Value------------------------------------------
2264 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2265 const Type *t = klass_value_common(phase);
2266 if (t == Type::TOP)
2267 return t;
2268
2269 return t->make_narrowklass();
2270 }
2271
2272 //------------------------------Identity---------------------------------------
2273 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2274 // Also feed through the klass in Allocate(...klass...)._klass.
2275 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2276 Node *x = klass_identity_common(phase);
2277
2278 const Type *t = phase->type( x );
2279 if( t == Type::TOP ) return x;
2280 if( t->isa_narrowklass()) return x;
2281 assert (!t->isa_narrowoop(), "no narrow oop here");
2282
2283 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2284 }
2285
2286 //------------------------------Value-----------------------------------------
2287 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2288 // Either input is TOP ==> the result is TOP
2289 const Type *t1 = phase->type( in(MemNode::Memory) );
2290 if( t1 == Type::TOP ) return Type::TOP;
2291 Node *adr = in(MemNode::Address);
2292 const Type *t2 = phase->type( adr );
2293 if( t2 == Type::TOP ) return Type::TOP;
2294 const TypePtr *tp = t2->is_ptr();
2295 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2296 const TypeAryPtr *tap = tp->isa_aryptr();
2297 if( !tap ) return _type;
2298 return tap->size();
2299 }
2300
2301 //-------------------------------Ideal---------------------------------------
2302 // Feed through the length in AllocateArray(...length...)._length.
2303 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2304 Node* p = MemNode::Ideal_common(phase, can_reshape);
2305 if (p) return (p == NodeSentinel) ? NULL : p;
2306
2307 // Take apart the address into an oop and and offset.
2308 // Return 'this' if we cannot.
2309 Node* adr = in(MemNode::Address);
2310 intptr_t offset = 0;
2311 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2312 if (base == NULL) return NULL;
2313 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2314 if (tary == NULL) return NULL;
2315
2316 // We can fetch the length directly through an AllocateArrayNode.
2317 // This works even if the length is not constant (clone or newArray).
2318 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2319 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2320 if (alloc != NULL) {
2321 Node* allocated_length = alloc->Ideal_length();
2322 Node* len = alloc->make_ideal_length(tary, phase);
2323 if (allocated_length != len) {
2324 // New CastII improves on this.
2325 return len;
2326 }
2327 }
2328 }
2329
2330 return NULL;
2331 }
2332
2333 //------------------------------Identity---------------------------------------
2334 // Feed through the length in AllocateArray(...length...)._length.
2335 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2336 Node* x = LoadINode::Identity(phase);
2337 if (x != this) return x;
2338
2339 // Take apart the address into an oop and and offset.
2340 // Return 'this' if we cannot.
2341 Node* adr = in(MemNode::Address);
2342 intptr_t offset = 0;
2343 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2344 if (base == NULL) return this;
2345 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2346 if (tary == NULL) return this;
2347
2348 // We can fetch the length directly through an AllocateArrayNode.
2349 // This works even if the length is not constant (clone or newArray).
2350 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2351 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2352 if (alloc != NULL) {
2353 Node* allocated_length = alloc->Ideal_length();
2354 // Do not allow make_ideal_length to allocate a CastII node.
2355 Node* len = alloc->make_ideal_length(tary, phase, false);
2356 if (allocated_length == len) {
2357 // Return allocated_length only if it would not be improved by a CastII.
2358 return allocated_length;
2359 }
2360 }
2361 }
2362
2363 return this;
2364
2365 }
2366
2367 //=============================================================================
2368 //---------------------------StoreNode::make-----------------------------------
2369 // Polymorphic factory method:
2370 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2371 assert((mo == unordered || mo == release), "unexpected");
2372 Compile* C = gvn.C;
2373 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2374 ctl != NULL, "raw memory operations should have control edge");
2375
2376 switch (bt) {
2377 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2378 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2379 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2380 case T_CHAR:
2381 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2382 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2383 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2384 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2385 case T_METADATA:
2386 case T_ADDRESS:
2387 case T_OBJECT:
2388 #ifdef _LP64
2389 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2390 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2391 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2392 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2393 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2394 adr->bottom_type()->isa_rawptr())) {
2395 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2396 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2397 }
2398 #endif
2399 {
2400 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2401 }
2402 default:
2403 ShouldNotReachHere();
2404 return (StoreNode*)NULL;
2405 }
2406 }
2407
2408 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2409 bool require_atomic = true;
2410 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2411 }
2412
2413 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2414 bool require_atomic = true;
2415 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2416 }
2417
2418
2419 //--------------------------bottom_type----------------------------------------
2420 const Type *StoreNode::bottom_type() const {
2421 return Type::MEMORY;
2422 }
2423
2424 //------------------------------hash-------------------------------------------
2425 uint StoreNode::hash() const {
2426 // unroll addition of interesting fields
2427 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2428
2429 // Since they are not commoned, do not hash them:
2430 return NO_HASH;
2431 }
2432
2433 //------------------------------Ideal------------------------------------------
2434 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2435 // When a store immediately follows a relevant allocation/initialization,
2436 // try to capture it into the initialization, or hoist it above.
2437 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2438 Node* p = MemNode::Ideal_common(phase, can_reshape);
2439 if (p) return (p == NodeSentinel) ? NULL : p;
2440
2441 Node* mem = in(MemNode::Memory);
2442 Node* address = in(MemNode::Address);
2443 // Back-to-back stores to same address? Fold em up. Generally
2444 // unsafe if I have intervening uses... Also disallowed for StoreCM
2445 // since they must follow each StoreP operation. Redundant StoreCMs
2446 // are eliminated just before matching in final_graph_reshape.
2447 {
2448 Node* st = mem;
2449 // If Store 'st' has more than one use, we cannot fold 'st' away.
2450 // For example, 'st' might be the final state at a conditional
2451 // return. Or, 'st' might be used by some node which is live at
2452 // the same time 'st' is live, which might be unschedulable. So,
2453 // require exactly ONE user until such time as we clone 'mem' for
2454 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2455 // true).
2456 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2457 // Looking at a dead closed cycle of memory?
2458 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2459 assert(Opcode() == st->Opcode() ||
2460 st->Opcode() == Op_StoreVector ||
2461 Opcode() == Op_StoreVector ||
2462 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2463 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2464 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2465 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2466 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2467
2468 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2469 st->as_Store()->memory_size() <= this->memory_size()) {
2470 Node* use = st->raw_out(0);
2471 phase->igvn_rehash_node_delayed(use);
2472 if (can_reshape) {
2473 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2474 } else {
2475 // It's OK to do this in the parser, since DU info is always accurate,
2476 // and the parser always refers to nodes via SafePointNode maps.
2477 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2478 }
2479 return this;
2480 }
2481 st = st->in(MemNode::Memory);
2482 }
2483 }
2484
2485
2486 // Capture an unaliased, unconditional, simple store into an initializer.
2487 // Or, if it is independent of the allocation, hoist it above the allocation.
2488 if (ReduceFieldZeroing && /*can_reshape &&*/
2489 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2490 InitializeNode* init = mem->in(0)->as_Initialize();
2491 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2492 if (offset > 0) {
2493 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2494 // If the InitializeNode captured me, it made a raw copy of me,
2495 // and I need to disappear.
2496 if (moved != NULL) {
2497 // %%% hack to ensure that Ideal returns a new node:
2498 mem = MergeMemNode::make(mem);
2499 return mem; // fold me away
2500 }
2501 }
2502 }
2503
2504 return NULL; // No further progress
2505 }
2506
2507 //------------------------------Value-----------------------------------------
2508 const Type* StoreNode::Value(PhaseGVN* phase) const {
2509 // Either input is TOP ==> the result is TOP
2510 const Type *t1 = phase->type( in(MemNode::Memory) );
2511 if( t1 == Type::TOP ) return Type::TOP;
2512 const Type *t2 = phase->type( in(MemNode::Address) );
2513 if( t2 == Type::TOP ) return Type::TOP;
2514 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2515 if( t3 == Type::TOP ) return Type::TOP;
2516 return Type::MEMORY;
2517 }
2518
2519 //------------------------------Identity---------------------------------------
2520 // Remove redundant stores:
2521 // Store(m, p, Load(m, p)) changes to m.
2522 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2523 Node* StoreNode::Identity(PhaseGVN* phase) {
2524 Node* mem = in(MemNode::Memory);
2525 Node* adr = in(MemNode::Address);
2526 Node* val = in(MemNode::ValueIn);
2527
2528 Node* result = this;
2529
2530 // Load then Store? Then the Store is useless
2531 if (val->is_Load() &&
2532 val->in(MemNode::Address)->eqv_uncast(adr) &&
2533 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2534 val->as_Load()->store_Opcode() == Opcode()) {
2535 result = mem;
2536 }
2537
2538 // Two stores in a row of the same value?
2539 if (result == this &&
2540 mem->is_Store() &&
2541 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2542 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2543 mem->Opcode() == Opcode()) {
2544 result = mem;
2545 }
2546
2547 // Store of zero anywhere into a freshly-allocated object?
2548 // Then the store is useless.
2549 // (It must already have been captured by the InitializeNode.)
2550 if (result == this &&
2551 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2552 // a newly allocated object is already all-zeroes everywhere
2553 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2554 result = mem;
2555 }
2556
2557 if (result == this) {
2558 // the store may also apply to zero-bits in an earlier object
2559 Node* prev_mem = find_previous_store(phase);
2560 // Steps (a), (b): Walk past independent stores to find an exact match.
2561 if (prev_mem != NULL) {
2562 Node* prev_val = can_see_stored_value(prev_mem, phase);
2563 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2564 // prev_val and val might differ by a cast; it would be good
2565 // to keep the more informative of the two.
2566 result = mem;
2567 }
2568 }
2569 }
2570 }
2571
2572 if (result != this && phase->is_IterGVN() != NULL) {
2573 MemBarNode* trailing = trailing_membar();
2574 if (trailing != NULL) {
2575 #ifdef ASSERT
2576 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2577 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2578 #endif
2579 PhaseIterGVN* igvn = phase->is_IterGVN();
2580 trailing->remove(igvn);
2581 }
2582 }
2583
2584 return result;
2585 }
2586
2587 //------------------------------match_edge-------------------------------------
2588 // Do we Match on this edge index or not? Match only memory & value
2589 uint StoreNode::match_edge(uint idx) const {
2590 return idx == MemNode::Address || idx == MemNode::ValueIn;
2591 }
2592
2593 //------------------------------cmp--------------------------------------------
2594 // Do not common stores up together. They generally have to be split
2595 // back up anyways, so do not bother.
2596 uint StoreNode::cmp( const Node &n ) const {
2597 return (&n == this); // Always fail except on self
2598 }
2599
2600 //------------------------------Ideal_masked_input-----------------------------
2601 // Check for a useless mask before a partial-word store
2602 // (StoreB ... (AndI valIn conIa) )
2603 // If (conIa & mask == mask) this simplifies to
2604 // (StoreB ... (valIn) )
2605 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2606 Node *val = in(MemNode::ValueIn);
2607 if( val->Opcode() == Op_AndI ) {
2608 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2609 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2610 set_req(MemNode::ValueIn, val->in(1));
2611 return this;
2612 }
2613 }
2614 return NULL;
2615 }
2616
2617
2618 //------------------------------Ideal_sign_extended_input----------------------
2619 // Check for useless sign-extension before a partial-word store
2620 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2621 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2622 // (StoreB ... (valIn) )
2623 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2624 Node *val = in(MemNode::ValueIn);
2625 if( val->Opcode() == Op_RShiftI ) {
2626 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2627 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2628 Node *shl = val->in(1);
2629 if( shl->Opcode() == Op_LShiftI ) {
2630 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2631 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2632 set_req(MemNode::ValueIn, shl->in(1));
2633 return this;
2634 }
2635 }
2636 }
2637 }
2638 return NULL;
2639 }
2640
2641 //------------------------------value_never_loaded-----------------------------------
2642 // Determine whether there are any possible loads of the value stored.
2643 // For simplicity, we actually check if there are any loads from the
2644 // address stored to, not just for loads of the value stored by this node.
2645 //
2646 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2647 Node *adr = in(Address);
2648 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2649 if (adr_oop == NULL)
2650 return false;
2651 if (!adr_oop->is_known_instance_field())
2652 return false; // if not a distinct instance, there may be aliases of the address
2653 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2654 Node *use = adr->fast_out(i);
2655 if (use->is_Load() || use->is_LoadStore()) {
2656 return false;
2657 }
2658 }
2659 return true;
2660 }
2661
2662 MemBarNode* StoreNode::trailing_membar() const {
2663 if (is_release()) {
2664 MemBarNode* trailing_mb = NULL;
2665 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2666 Node* u = fast_out(i);
2667 if (u->is_MemBar()) {
2668 if (u->as_MemBar()->trailing_store()) {
2669 assert(u->Opcode() == Op_MemBarVolatile, "");
2670 assert(trailing_mb == NULL, "only one");
2671 trailing_mb = u->as_MemBar();
2672 #ifdef ASSERT
2673 Node* leading = u->as_MemBar()->leading_membar();
2674 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2675 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2676 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2677 #endif
2678 } else {
2679 assert(u->as_MemBar()->standalone(), "");
2680 }
2681 }
2682 }
2683 return trailing_mb;
2684 }
2685 return NULL;
2686 }
2687
2688
2689 //=============================================================================
2690 //------------------------------Ideal------------------------------------------
2691 // If the store is from an AND mask that leaves the low bits untouched, then
2692 // we can skip the AND operation. If the store is from a sign-extension
2693 // (a left shift, then right shift) we can skip both.
2694 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2695 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2696 if( progress != NULL ) return progress;
2697
2698 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2699 if( progress != NULL ) return progress;
2700
2701 // Finally check the default case
2702 return StoreNode::Ideal(phase, can_reshape);
2703 }
2704
2705 //=============================================================================
2706 //------------------------------Ideal------------------------------------------
2707 // If the store is from an AND mask that leaves the low bits untouched, then
2708 // we can skip the AND operation
2709 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2710 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2711 if( progress != NULL ) return progress;
2712
2713 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2714 if( progress != NULL ) return progress;
2715
2716 // Finally check the default case
2717 return StoreNode::Ideal(phase, can_reshape);
2718 }
2719
2720 //=============================================================================
2721 //------------------------------Identity---------------------------------------
2722 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2723 // No need to card mark when storing a null ptr
2724 Node* my_store = in(MemNode::OopStore);
2725 if (my_store->is_Store()) {
2726 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2727 if( t1 == TypePtr::NULL_PTR ) {
2728 return in(MemNode::Memory);
2729 }
2730 }
2731 return this;
2732 }
2733
2734 //=============================================================================
2735 //------------------------------Ideal---------------------------------------
2736 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2737 Node* progress = StoreNode::Ideal(phase, can_reshape);
2738 if (progress != NULL) return progress;
2739
2740 Node* my_store = in(MemNode::OopStore);
2741 if (my_store->is_MergeMem()) {
2742 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2743 set_req(MemNode::OopStore, mem);
2744 return this;
2745 }
2746
2747 return NULL;
2748 }
2749
2750 //------------------------------Value-----------------------------------------
2751 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2752 // Either input is TOP ==> the result is TOP
2753 const Type *t = phase->type( in(MemNode::Memory) );
2754 if( t == Type::TOP ) return Type::TOP;
2755 t = phase->type( in(MemNode::Address) );
2756 if( t == Type::TOP ) return Type::TOP;
2757 t = phase->type( in(MemNode::ValueIn) );
2758 if( t == Type::TOP ) return Type::TOP;
2759 // If extra input is TOP ==> the result is TOP
2760 t = phase->type( in(MemNode::OopStore) );
2761 if( t == Type::TOP ) return Type::TOP;
2762
2763 return StoreNode::Value( phase );
2764 }
2765
2766
2767 //=============================================================================
2768 //----------------------------------SCMemProjNode------------------------------
2769 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2770 {
2771 return bottom_type();
2772 }
2773
2774 //=============================================================================
2775 //----------------------------------LoadStoreNode------------------------------
2776 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2777 : Node(required),
2778 _type(rt),
2779 _adr_type(at)
2780 {
2781 init_req(MemNode::Control, c );
2782 init_req(MemNode::Memory , mem);
2783 init_req(MemNode::Address, adr);
2784 init_req(MemNode::ValueIn, val);
2785 init_class_id(Class_LoadStore);
2786 }
2787
2788 uint LoadStoreNode::ideal_reg() const {
2789 return _type->ideal_reg();
2790 }
2791
2792 bool LoadStoreNode::result_not_used() const {
2793 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2794 Node *x = fast_out(i);
2795 if (x->Opcode() == Op_SCMemProj) continue;
2796 return false;
2797 }
2798 return true;
2799 }
2800
2801 MemBarNode* LoadStoreNode::trailing_membar() const {
2802 MemBarNode* trailing = NULL;
2803 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2804 Node* u = fast_out(i);
2805 if (u->is_MemBar()) {
2806 if (u->as_MemBar()->trailing_load_store()) {
2807 assert(u->Opcode() == Op_MemBarAcquire, "");
2808 assert(trailing == NULL, "only one");
2809 trailing = u->as_MemBar();
2810 #ifdef ASSERT
2811 Node* leading = trailing->leading_membar();
2812 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2813 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
2814 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
2815 #endif
2816 } else {
2817 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
2818 }
2819 }
2820 }
2821
2822 return trailing;
2823 }
2824
2825 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2826
2827 //=============================================================================
2828 //----------------------------------LoadStoreConditionalNode--------------------
2829 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2830 init_req(ExpectedIn, ex );
2831 }
2832
2833 //=============================================================================
2834 //-------------------------------adr_type--------------------------------------
2835 const TypePtr* ClearArrayNode::adr_type() const {
2836 Node *adr = in(3);
2837 if (adr == NULL) return NULL; // node is dead
2838 return MemNode::calculate_adr_type(adr->bottom_type());
2839 }
2840
2841 //------------------------------match_edge-------------------------------------
2842 // Do we Match on this edge index or not? Do not match memory
2843 uint ClearArrayNode::match_edge(uint idx) const {
2844 return idx > 1;
2845 }
2846
2847 //------------------------------Identity---------------------------------------
2848 // Clearing a zero length array does nothing
2849 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2850 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2851 }
2852
2853 //------------------------------Idealize---------------------------------------
2854 // Clearing a short array is faster with stores
2855 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2856 // Already know this is a large node, do not try to ideal it
2857 if (!IdealizeClearArrayNode || _is_large) return NULL;
2858
2859 const int unit = BytesPerLong;
2860 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2861 if (!t) return NULL;
2862 if (!t->is_con()) return NULL;
2863 intptr_t raw_count = t->get_con();
2864 intptr_t size = raw_count;
2865 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2866 // Clearing nothing uses the Identity call.
2867 // Negative clears are possible on dead ClearArrays
2868 // (see jck test stmt114.stmt11402.val).
2869 if (size <= 0 || size % unit != 0) return NULL;
2870 intptr_t count = size / unit;
2871 // Length too long; communicate this to matchers and assemblers.
2872 // Assemblers are responsible to produce fast hardware clears for it.
2873 if (size > InitArrayShortSize) {
2874 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
2875 }
2876 Node *mem = in(1);
2877 if( phase->type(mem)==Type::TOP ) return NULL;
2878 Node *adr = in(3);
2879 const Type* at = phase->type(adr);
2880 if( at==Type::TOP ) return NULL;
2881 const TypePtr* atp = at->isa_ptr();
2882 // adjust atp to be the correct array element address type
2883 if (atp == NULL) atp = TypePtr::BOTTOM;
2884 else atp = atp->add_offset(Type::OffsetBot);
2885 // Get base for derived pointer purposes
2886 if( adr->Opcode() != Op_AddP ) Unimplemented();
2887 Node *base = adr->in(1);
2888
2889 Node *zero = phase->makecon(TypeLong::ZERO);
2890 Node *off = phase->MakeConX(BytesPerLong);
2891 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2892 count--;
2893 while( count-- ) {
2894 mem = phase->transform(mem);
2895 adr = phase->transform(new AddPNode(base,adr,off));
2896 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2897 }
2898 return mem;
2899 }
2900
2901 //----------------------------step_through----------------------------------
2902 // Return allocation input memory edge if it is different instance
2903 // or itself if it is the one we are looking for.
2904 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2905 Node* n = *np;
2906 assert(n->is_ClearArray(), "sanity");
2907 intptr_t offset;
2908 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2909 // This method is called only before Allocate nodes are expanded
2910 // during macro nodes expansion. Before that ClearArray nodes are
2911 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2912 // Allocate nodes are expanded) which follows allocations.
2913 assert(alloc != NULL, "should have allocation");
2914 if (alloc->_idx == instance_id) {
2915 // Can not bypass initialization of the instance we are looking for.
2916 return false;
2917 }
2918 // Otherwise skip it.
2919 InitializeNode* init = alloc->initialization();
2920 if (init != NULL)
2921 *np = init->in(TypeFunc::Memory);
2922 else
2923 *np = alloc->in(TypeFunc::Memory);
2924 return true;
2925 }
2926
2927 //----------------------------clear_memory-------------------------------------
2928 // Generate code to initialize object storage to zero.
2929 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2930 intptr_t start_offset,
2931 Node* end_offset,
2932 PhaseGVN* phase) {
2933 intptr_t offset = start_offset;
2934
2935 int unit = BytesPerLong;
2936 if ((offset % unit) != 0) {
2937 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
2938 adr = phase->transform(adr);
2939 const TypePtr* atp = TypeRawPtr::BOTTOM;
2940 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2941 mem = phase->transform(mem);
2942 offset += BytesPerInt;
2943 }
2944 assert((offset % unit) == 0, "");
2945
2946 // Initialize the remaining stuff, if any, with a ClearArray.
2947 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2948 }
2949
2950 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2951 Node* start_offset,
2952 Node* end_offset,
2953 PhaseGVN* phase) {
2954 if (start_offset == end_offset) {
2955 // nothing to do
2956 return mem;
2957 }
2958
2959 int unit = BytesPerLong;
2960 Node* zbase = start_offset;
2961 Node* zend = end_offset;
2962
2963 // Scale to the unit required by the CPU:
2964 if (!Matcher::init_array_count_is_in_bytes) {
2965 Node* shift = phase->intcon(exact_log2(unit));
2966 zbase = phase->transform(new URShiftXNode(zbase, shift) );
2967 zend = phase->transform(new URShiftXNode(zend, shift) );
2968 }
2969
2970 // Bulk clear double-words
2971 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
2972 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
2973 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
2974 return phase->transform(mem);
2975 }
2976
2977 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2978 intptr_t start_offset,
2979 intptr_t end_offset,
2980 PhaseGVN* phase) {
2981 if (start_offset == end_offset) {
2982 // nothing to do
2983 return mem;
2984 }
2985
2986 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2987 intptr_t done_offset = end_offset;
2988 if ((done_offset % BytesPerLong) != 0) {
2989 done_offset -= BytesPerInt;
2990 }
2991 if (done_offset > start_offset) {
2992 mem = clear_memory(ctl, mem, dest,
2993 start_offset, phase->MakeConX(done_offset), phase);
2994 }
2995 if (done_offset < end_offset) { // emit the final 32-bit store
2996 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
2997 adr = phase->transform(adr);
2998 const TypePtr* atp = TypeRawPtr::BOTTOM;
2999 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3000 mem = phase->transform(mem);
3001 done_offset += BytesPerInt;
3002 }
3003 assert(done_offset == end_offset, "");
3004 return mem;
3005 }
3006
3007 //=============================================================================
3008 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3009 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3010 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3011 #ifdef ASSERT
3012 , _pair_idx(0)
3013 #endif
3014 {
3015 init_class_id(Class_MemBar);
3016 Node* top = C->top();
3017 init_req(TypeFunc::I_O,top);
3018 init_req(TypeFunc::FramePtr,top);
3019 init_req(TypeFunc::ReturnAdr,top);
3020 if (precedent != NULL)
3021 init_req(TypeFunc::Parms, precedent);
3022 }
3023
3024 //------------------------------cmp--------------------------------------------
3025 uint MemBarNode::hash() const { return NO_HASH; }
3026 uint MemBarNode::cmp( const Node &n ) const {
3027 return (&n == this); // Always fail except on self
3028 }
3029
3030 //------------------------------make-------------------------------------------
3031 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3032 switch (opcode) {
3033 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3034 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3035 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3036 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3037 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3038 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3039 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3040 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3041 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3042 case Op_Initialize: return new InitializeNode(C, atp, pn);
3043 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3044 default: ShouldNotReachHere(); return NULL;
3045 }
3046 }
3047
3048 void MemBarNode::remove(PhaseIterGVN *igvn) {
3049 if (outcnt() != 2) {
3050 return;
3051 }
3052 if (trailing_store() || trailing_load_store()) {
3053 MemBarNode* leading = leading_membar();
3054 if (leading != NULL) {
3055 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3056 leading->remove(igvn);
3057 }
3058 }
3059 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3060 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3061 }
3062
3063 //------------------------------Ideal------------------------------------------
3064 // Return a node which is more "ideal" than the current node. Strip out
3065 // control copies
3066 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3067 if (remove_dead_region(phase, can_reshape)) return this;
3068 // Don't bother trying to transform a dead node
3069 if (in(0) && in(0)->is_top()) {
3070 return NULL;
3071 }
3072
3073 #if INCLUDE_ZGC
3074 if (UseZGC) {
3075 if (req() == (Precedent+1) && in(MemBarNode::Precedent)->in(0) != NULL && in(MemBarNode::Precedent)->in(0)->is_LoadBarrier()) {
3076 Node* load_node = in(MemBarNode::Precedent)->in(0)->in(LoadBarrierNode::Oop);
3077 set_req(MemBarNode::Precedent, load_node);
3078 return this;
3079 }
3080 }
3081 #endif
3082
3083 bool progress = false;
3084 // Eliminate volatile MemBars for scalar replaced objects.
3085 if (can_reshape && req() == (Precedent+1)) {
3086 bool eliminate = false;
3087 int opc = Opcode();
3088 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3089 // Volatile field loads and stores.
3090 Node* my_mem = in(MemBarNode::Precedent);
3091 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3092 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3093 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3094 // replace this Precedent (decodeN) with the Load instead.
3095 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3096 Node* load_node = my_mem->in(1);
3097 set_req(MemBarNode::Precedent, load_node);
3098 phase->is_IterGVN()->_worklist.push(my_mem);
3099 my_mem = load_node;
3100 } else {
3101 assert(my_mem->unique_out() == this, "sanity");
3102 del_req(Precedent);
3103 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3104 my_mem = NULL;
3105 }
3106 progress = true;
3107 }
3108 if (my_mem != NULL && my_mem->is_Mem()) {
3109 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3110 // Check for scalar replaced object reference.
3111 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3112 t_oop->offset() != Type::OffsetBot &&
3113 t_oop->offset() != Type::OffsetTop) {
3114 eliminate = true;
3115 }
3116 }
3117 } else if (opc == Op_MemBarRelease) {
3118 // Final field stores.
3119 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3120 if ((alloc != NULL) && alloc->is_Allocate() &&
3121 alloc->as_Allocate()->does_not_escape_thread()) {
3122 // The allocated object does not escape.
3123 eliminate = true;
3124 }
3125 }
3126 if (eliminate) {
3127 // Replace MemBar projections by its inputs.
3128 PhaseIterGVN* igvn = phase->is_IterGVN();
3129 remove(igvn);
3130 // Must return either the original node (now dead) or a new node
3131 // (Do not return a top here, since that would break the uniqueness of top.)
3132 return new ConINode(TypeInt::ZERO);
3133 }
3134 }
3135 return progress ? this : NULL;
3136 }
3137
3138 //------------------------------Value------------------------------------------
3139 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3140 if( !in(0) ) return Type::TOP;
3141 if( phase->type(in(0)) == Type::TOP )
3142 return Type::TOP;
3143 return TypeTuple::MEMBAR;
3144 }
3145
3146 //------------------------------match------------------------------------------
3147 // Construct projections for memory.
3148 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3149 switch (proj->_con) {
3150 case TypeFunc::Control:
3151 case TypeFunc::Memory:
3152 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3153 }
3154 ShouldNotReachHere();
3155 return NULL;
3156 }
3157
3158 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3159 trailing->_kind = TrailingStore;
3160 leading->_kind = LeadingStore;
3161 #ifdef ASSERT
3162 trailing->_pair_idx = leading->_idx;
3163 leading->_pair_idx = leading->_idx;
3164 #endif
3165 }
3166
3167 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3168 trailing->_kind = TrailingLoadStore;
3169 leading->_kind = LeadingLoadStore;
3170 #ifdef ASSERT
3171 trailing->_pair_idx = leading->_idx;
3172 leading->_pair_idx = leading->_idx;
3173 #endif
3174 }
3175
3176 MemBarNode* MemBarNode::trailing_membar() const {
3177 ResourceMark rm;
3178 Node* trailing = (Node*)this;
3179 VectorSet seen(Thread::current()->resource_area());
3180 Node_Stack multis(0);
3181 do {
3182 Node* c = trailing;
3183 uint i = 0;
3184 do {
3185 trailing = NULL;
3186 for (; i < c->outcnt(); i++) {
3187 Node* next = c->raw_out(i);
3188 if (next != c && next->is_CFG()) {
3189 if (c->is_MultiBranch()) {
3190 if (multis.node() == c) {
3191 multis.set_index(i+1);
3192 } else {
3193 multis.push(c, i+1);
3194 }
3195 }
3196 trailing = next;
3197 break;
3198 }
3199 }
3200 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3201 break;
3202 }
3203 while (multis.size() > 0) {
3204 c = multis.node();
3205 i = multis.index();
3206 if (i < c->req()) {
3207 break;
3208 }
3209 multis.pop();
3210 }
3211 } while (multis.size() > 0);
3212 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3213
3214 MemBarNode* mb = trailing->as_MemBar();
3215 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3216 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3217 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3218 return mb;
3219 }
3220
3221 MemBarNode* MemBarNode::leading_membar() const {
3222 ResourceMark rm;
3223 VectorSet seen(Thread::current()->resource_area());
3224 Node_Stack regions(0);
3225 Node* leading = in(0);
3226 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3227 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3228 leading = NULL;
3229 while (regions.size() > 0) {
3230 Node* r = regions.node();
3231 uint i = regions.index();
3232 if (i < r->req()) {
3233 leading = r->in(i);
3234 regions.set_index(i+1);
3235 } else {
3236 regions.pop();
3237 }
3238 }
3239 if (leading == NULL) {
3240 assert(regions.size() == 0, "all paths should have been tried");
3241 return NULL;
3242 }
3243 }
3244 if (leading->is_Region()) {
3245 regions.push(leading, 2);
3246 leading = leading->in(1);
3247 } else {
3248 leading = leading->in(0);
3249 }
3250 }
3251 #ifdef ASSERT
3252 Unique_Node_List wq;
3253 wq.push((Node*)this);
3254 uint found = 0;
3255 for (uint i = 0; i < wq.size(); i++) {
3256 Node* n = wq.at(i);
3257 if (n->is_Region()) {
3258 for (uint j = 1; j < n->req(); j++) {
3259 Node* in = n->in(j);
3260 if (in != NULL && !in->is_top()) {
3261 wq.push(in);
3262 }
3263 }
3264 } else {
3265 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3266 assert(n == leading, "consistency check failed");
3267 found++;
3268 } else {
3269 Node* in = n->in(0);
3270 if (in != NULL && !in->is_top()) {
3271 wq.push(in);
3272 }
3273 }
3274 }
3275 }
3276 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3277 #endif
3278 if (leading == NULL) {
3279 return NULL;
3280 }
3281 MemBarNode* mb = leading->as_MemBar();
3282 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3283 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3284 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3285 return mb;
3286 }
3287
3288 //===========================InitializeNode====================================
3289 // SUMMARY:
3290 // This node acts as a memory barrier on raw memory, after some raw stores.
3291 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3292 // The Initialize can 'capture' suitably constrained stores as raw inits.
3293 // It can coalesce related raw stores into larger units (called 'tiles').
3294 // It can avoid zeroing new storage for memory units which have raw inits.
3295 // At macro-expansion, it is marked 'complete', and does not optimize further.
3296 //
3297 // EXAMPLE:
3298 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3299 // ctl = incoming control; mem* = incoming memory
3300 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3301 // First allocate uninitialized memory and fill in the header:
3302 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3303 // ctl := alloc.Control; mem* := alloc.Memory*
3304 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3305 // Then initialize to zero the non-header parts of the raw memory block:
3306 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3307 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3308 // After the initialize node executes, the object is ready for service:
3309 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3310 // Suppose its body is immediately initialized as {1,2}:
3311 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3312 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3313 // mem.SLICE(#short[*]) := store2
3314 //
3315 // DETAILS:
3316 // An InitializeNode collects and isolates object initialization after
3317 // an AllocateNode and before the next possible safepoint. As a
3318 // memory barrier (MemBarNode), it keeps critical stores from drifting
3319 // down past any safepoint or any publication of the allocation.
3320 // Before this barrier, a newly-allocated object may have uninitialized bits.
3321 // After this barrier, it may be treated as a real oop, and GC is allowed.
3322 //
3323 // The semantics of the InitializeNode include an implicit zeroing of
3324 // the new object from object header to the end of the object.
3325 // (The object header and end are determined by the AllocateNode.)
3326 //
3327 // Certain stores may be added as direct inputs to the InitializeNode.
3328 // These stores must update raw memory, and they must be to addresses
3329 // derived from the raw address produced by AllocateNode, and with
3330 // a constant offset. They must be ordered by increasing offset.
3331 // The first one is at in(RawStores), the last at in(req()-1).
3332 // Unlike most memory operations, they are not linked in a chain,
3333 // but are displayed in parallel as users of the rawmem output of
3334 // the allocation.
3335 //
3336 // (See comments in InitializeNode::capture_store, which continue
3337 // the example given above.)
3338 //
3339 // When the associated Allocate is macro-expanded, the InitializeNode
3340 // may be rewritten to optimize collected stores. A ClearArrayNode
3341 // may also be created at that point to represent any required zeroing.
3342 // The InitializeNode is then marked 'complete', prohibiting further
3343 // capturing of nearby memory operations.
3344 //
3345 // During macro-expansion, all captured initializations which store
3346 // constant values of 32 bits or smaller are coalesced (if advantageous)
3347 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3348 // initialized in fewer memory operations. Memory words which are
3349 // covered by neither tiles nor non-constant stores are pre-zeroed
3350 // by explicit stores of zero. (The code shape happens to do all
3351 // zeroing first, then all other stores, with both sequences occurring
3352 // in order of ascending offsets.)
3353 //
3354 // Alternatively, code may be inserted between an AllocateNode and its
3355 // InitializeNode, to perform arbitrary initialization of the new object.
3356 // E.g., the object copying intrinsics insert complex data transfers here.
3357 // The initialization must then be marked as 'complete' disable the
3358 // built-in zeroing semantics and the collection of initializing stores.
3359 //
3360 // While an InitializeNode is incomplete, reads from the memory state
3361 // produced by it are optimizable if they match the control edge and
3362 // new oop address associated with the allocation/initialization.
3363 // They return a stored value (if the offset matches) or else zero.
3364 // A write to the memory state, if it matches control and address,
3365 // and if it is to a constant offset, may be 'captured' by the
3366 // InitializeNode. It is cloned as a raw memory operation and rewired
3367 // inside the initialization, to the raw oop produced by the allocation.
3368 // Operations on addresses which are provably distinct (e.g., to
3369 // other AllocateNodes) are allowed to bypass the initialization.
3370 //
3371 // The effect of all this is to consolidate object initialization
3372 // (both arrays and non-arrays, both piecewise and bulk) into a
3373 // single location, where it can be optimized as a unit.
3374 //
3375 // Only stores with an offset less than TrackedInitializationLimit words
3376 // will be considered for capture by an InitializeNode. This puts a
3377 // reasonable limit on the complexity of optimized initializations.
3378
3379 //---------------------------InitializeNode------------------------------------
3380 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3381 : MemBarNode(C, adr_type, rawoop),
3382 _is_complete(Incomplete), _does_not_escape(false)
3383 {
3384 init_class_id(Class_Initialize);
3385
3386 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3387 assert(in(RawAddress) == rawoop, "proper init");
3388 // Note: allocation() can be NULL, for secondary initialization barriers
3389 }
3390
3391 // Since this node is not matched, it will be processed by the
3392 // register allocator. Declare that there are no constraints
3393 // on the allocation of the RawAddress edge.
3394 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3395 // This edge should be set to top, by the set_complete. But be conservative.
3396 if (idx == InitializeNode::RawAddress)
3397 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3398 return RegMask::Empty;
3399 }
3400
3401 Node* InitializeNode::memory(uint alias_idx) {
3402 Node* mem = in(Memory);
3403 if (mem->is_MergeMem()) {
3404 return mem->as_MergeMem()->memory_at(alias_idx);
3405 } else {
3406 // incoming raw memory is not split
3407 return mem;
3408 }
3409 }
3410
3411 bool InitializeNode::is_non_zero() {
3412 if (is_complete()) return false;
3413 remove_extra_zeroes();
3414 return (req() > RawStores);
3415 }
3416
3417 void InitializeNode::set_complete(PhaseGVN* phase) {
3418 assert(!is_complete(), "caller responsibility");
3419 _is_complete = Complete;
3420
3421 // After this node is complete, it contains a bunch of
3422 // raw-memory initializations. There is no need for
3423 // it to have anything to do with non-raw memory effects.
3424 // Therefore, tell all non-raw users to re-optimize themselves,
3425 // after skipping the memory effects of this initialization.
3426 PhaseIterGVN* igvn = phase->is_IterGVN();
3427 if (igvn) igvn->add_users_to_worklist(this);
3428 }
3429
3430 // convenience function
3431 // return false if the init contains any stores already
3432 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3433 InitializeNode* init = initialization();
3434 if (init == NULL || init->is_complete()) return false;
3435 init->remove_extra_zeroes();
3436 // for now, if this allocation has already collected any inits, bail:
3437 if (init->is_non_zero()) return false;
3438 init->set_complete(phase);
3439 return true;
3440 }
3441
3442 void InitializeNode::remove_extra_zeroes() {
3443 if (req() == RawStores) return;
3444 Node* zmem = zero_memory();
3445 uint fill = RawStores;
3446 for (uint i = fill; i < req(); i++) {
3447 Node* n = in(i);
3448 if (n->is_top() || n == zmem) continue; // skip
3449 if (fill < i) set_req(fill, n); // compact
3450 ++fill;
3451 }
3452 // delete any empty spaces created:
3453 while (fill < req()) {
3454 del_req(fill);
3455 }
3456 }
3457
3458 // Helper for remembering which stores go with which offsets.
3459 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3460 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3461 intptr_t offset = -1;
3462 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3463 phase, offset);
3464 if (base == NULL) return -1; // something is dead,
3465 if (offset < 0) return -1; // dead, dead
3466 return offset;
3467 }
3468
3469 // Helper for proving that an initialization expression is
3470 // "simple enough" to be folded into an object initialization.
3471 // Attempts to prove that a store's initial value 'n' can be captured
3472 // within the initialization without creating a vicious cycle, such as:
3473 // { Foo p = new Foo(); p.next = p; }
3474 // True for constants and parameters and small combinations thereof.
3475 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3476 if (n == NULL) return true; // (can this really happen?)
3477 if (n->is_Proj()) n = n->in(0);
3478 if (n == this) return false; // found a cycle
3479 if (n->is_Con()) return true;
3480 if (n->is_Start()) return true; // params, etc., are OK
3481 if (n->is_Root()) return true; // even better
3482
3483 Node* ctl = n->in(0);
3484 if (ctl != NULL && !ctl->is_top()) {
3485 if (ctl->is_Proj()) ctl = ctl->in(0);
3486 if (ctl == this) return false;
3487
3488 // If we already know that the enclosing memory op is pinned right after
3489 // the init, then any control flow that the store has picked up
3490 // must have preceded the init, or else be equal to the init.
3491 // Even after loop optimizations (which might change control edges)
3492 // a store is never pinned *before* the availability of its inputs.
3493 if (!MemNode::all_controls_dominate(n, this))
3494 return false; // failed to prove a good control
3495 }
3496
3497 // Check data edges for possible dependencies on 'this'.
3498 if ((count += 1) > 20) return false; // complexity limit
3499 for (uint i = 1; i < n->req(); i++) {
3500 Node* m = n->in(i);
3501 if (m == NULL || m == n || m->is_top()) continue;
3502 uint first_i = n->find_edge(m);
3503 if (i != first_i) continue; // process duplicate edge just once
3504 if (!detect_init_independence(m, count)) {
3505 return false;
3506 }
3507 }
3508
3509 return true;
3510 }
3511
3512 // Here are all the checks a Store must pass before it can be moved into
3513 // an initialization. Returns zero if a check fails.
3514 // On success, returns the (constant) offset to which the store applies,
3515 // within the initialized memory.
3516 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3517 const int FAIL = 0;
3518 if (st->is_unaligned_access()) {
3519 return FAIL;
3520 }
3521 if (st->req() != MemNode::ValueIn + 1)
3522 return FAIL; // an inscrutable StoreNode (card mark?)
3523 Node* ctl = st->in(MemNode::Control);
3524 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3525 return FAIL; // must be unconditional after the initialization
3526 Node* mem = st->in(MemNode::Memory);
3527 if (!(mem->is_Proj() && mem->in(0) == this))
3528 return FAIL; // must not be preceded by other stores
3529 Node* adr = st->in(MemNode::Address);
3530 intptr_t offset;
3531 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3532 if (alloc == NULL)
3533 return FAIL; // inscrutable address
3534 if (alloc != allocation())
3535 return FAIL; // wrong allocation! (store needs to float up)
3536 Node* val = st->in(MemNode::ValueIn);
3537 int complexity_count = 0;
3538 if (!detect_init_independence(val, complexity_count))
3539 return FAIL; // stored value must be 'simple enough'
3540
3541 // The Store can be captured only if nothing after the allocation
3542 // and before the Store is using the memory location that the store
3543 // overwrites.
3544 bool failed = false;
3545 // If is_complete_with_arraycopy() is true the shape of the graph is
3546 // well defined and is safe so no need for extra checks.
3547 if (!is_complete_with_arraycopy()) {
3548 // We are going to look at each use of the memory state following
3549 // the allocation to make sure nothing reads the memory that the
3550 // Store writes.
3551 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3552 int alias_idx = phase->C->get_alias_index(t_adr);
3553 ResourceMark rm;
3554 Unique_Node_List mems;
3555 mems.push(mem);
3556 Node* unique_merge = NULL;
3557 for (uint next = 0; next < mems.size(); ++next) {
3558 Node *m = mems.at(next);
3559 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3560 Node *n = m->fast_out(j);
3561 if (n->outcnt() == 0) {
3562 continue;
3563 }
3564 if (n == st) {
3565 continue;
3566 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3567 // If the control of this use is different from the control
3568 // of the Store which is right after the InitializeNode then
3569 // this node cannot be between the InitializeNode and the
3570 // Store.
3571 continue;
3572 } else if (n->is_MergeMem()) {
3573 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3574 // We can hit a MergeMemNode (that will likely go away
3575 // later) that is a direct use of the memory state
3576 // following the InitializeNode on the same slice as the
3577 // store node that we'd like to capture. We need to check
3578 // the uses of the MergeMemNode.
3579 mems.push(n);
3580 }
3581 } else if (n->is_Mem()) {
3582 Node* other_adr = n->in(MemNode::Address);
3583 if (other_adr == adr) {
3584 failed = true;
3585 break;
3586 } else {
3587 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3588 if (other_t_adr != NULL) {
3589 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3590 if (other_alias_idx == alias_idx) {
3591 // A load from the same memory slice as the store right
3592 // after the InitializeNode. We check the control of the
3593 // object/array that is loaded from. If it's the same as
3594 // the store control then we cannot capture the store.
3595 assert(!n->is_Store(), "2 stores to same slice on same control?");
3596 Node* base = other_adr;
3597 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3598 base = base->in(AddPNode::Base);
3599 if (base != NULL) {
3600 base = base->uncast();
3601 if (base->is_Proj() && base->in(0) == alloc) {
3602 failed = true;
3603 break;
3604 }
3605 }
3606 }
3607 }
3608 }
3609 } else {
3610 failed = true;
3611 break;
3612 }
3613 }
3614 }
3615 }
3616 if (failed) {
3617 if (!can_reshape) {
3618 // We decided we couldn't capture the store during parsing. We
3619 // should try again during the next IGVN once the graph is
3620 // cleaner.
3621 phase->C->record_for_igvn(st);
3622 }
3623 return FAIL;
3624 }
3625
3626 return offset; // success
3627 }
3628
3629 // Find the captured store in(i) which corresponds to the range
3630 // [start..start+size) in the initialized object.
3631 // If there is one, return its index i. If there isn't, return the
3632 // negative of the index where it should be inserted.
3633 // Return 0 if the queried range overlaps an initialization boundary
3634 // or if dead code is encountered.
3635 // If size_in_bytes is zero, do not bother with overlap checks.
3636 int InitializeNode::captured_store_insertion_point(intptr_t start,
3637 int size_in_bytes,
3638 PhaseTransform* phase) {
3639 const int FAIL = 0, MAX_STORE = BytesPerLong;
3640
3641 if (is_complete())
3642 return FAIL; // arraycopy got here first; punt
3643
3644 assert(allocation() != NULL, "must be present");
3645
3646 // no negatives, no header fields:
3647 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3648
3649 // after a certain size, we bail out on tracking all the stores:
3650 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3651 if (start >= ti_limit) return FAIL;
3652
3653 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3654 if (i >= limit) return -(int)i; // not found; here is where to put it
3655
3656 Node* st = in(i);
3657 intptr_t st_off = get_store_offset(st, phase);
3658 if (st_off < 0) {
3659 if (st != zero_memory()) {
3660 return FAIL; // bail out if there is dead garbage
3661 }
3662 } else if (st_off > start) {
3663 // ...we are done, since stores are ordered
3664 if (st_off < start + size_in_bytes) {
3665 return FAIL; // the next store overlaps
3666 }
3667 return -(int)i; // not found; here is where to put it
3668 } else if (st_off < start) {
3669 if (size_in_bytes != 0 &&
3670 start < st_off + MAX_STORE &&
3671 start < st_off + st->as_Store()->memory_size()) {
3672 return FAIL; // the previous store overlaps
3673 }
3674 } else {
3675 if (size_in_bytes != 0 &&
3676 st->as_Store()->memory_size() != size_in_bytes) {
3677 return FAIL; // mismatched store size
3678 }
3679 return i;
3680 }
3681
3682 ++i;
3683 }
3684 }
3685
3686 // Look for a captured store which initializes at the offset 'start'
3687 // with the given size. If there is no such store, and no other
3688 // initialization interferes, then return zero_memory (the memory
3689 // projection of the AllocateNode).
3690 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3691 PhaseTransform* phase) {
3692 assert(stores_are_sane(phase), "");
3693 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3694 if (i == 0) {
3695 return NULL; // something is dead
3696 } else if (i < 0) {
3697 return zero_memory(); // just primordial zero bits here
3698 } else {
3699 Node* st = in(i); // here is the store at this position
3700 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3701 return st;
3702 }
3703 }
3704
3705 // Create, as a raw pointer, an address within my new object at 'offset'.
3706 Node* InitializeNode::make_raw_address(intptr_t offset,
3707 PhaseTransform* phase) {
3708 Node* addr = in(RawAddress);
3709 if (offset != 0) {
3710 Compile* C = phase->C;
3711 addr = phase->transform( new AddPNode(C->top(), addr,
3712 phase->MakeConX(offset)) );
3713 }
3714 return addr;
3715 }
3716
3717 // Clone the given store, converting it into a raw store
3718 // initializing a field or element of my new object.
3719 // Caller is responsible for retiring the original store,
3720 // with subsume_node or the like.
3721 //
3722 // From the example above InitializeNode::InitializeNode,
3723 // here are the old stores to be captured:
3724 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3725 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3726 //
3727 // Here is the changed code; note the extra edges on init:
3728 // alloc = (Allocate ...)
3729 // rawoop = alloc.RawAddress
3730 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3731 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3732 // init = (Initialize alloc.Control alloc.Memory rawoop
3733 // rawstore1 rawstore2)
3734 //
3735 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3736 PhaseTransform* phase, bool can_reshape) {
3737 assert(stores_are_sane(phase), "");
3738
3739 if (start < 0) return NULL;
3740 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3741
3742 Compile* C = phase->C;
3743 int size_in_bytes = st->memory_size();
3744 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3745 if (i == 0) return NULL; // bail out
3746 Node* prev_mem = NULL; // raw memory for the captured store
3747 if (i > 0) {
3748 prev_mem = in(i); // there is a pre-existing store under this one
3749 set_req(i, C->top()); // temporarily disconnect it
3750 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3751 } else {
3752 i = -i; // no pre-existing store
3753 prev_mem = zero_memory(); // a slice of the newly allocated object
3754 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3755 set_req(--i, C->top()); // reuse this edge; it has been folded away
3756 else
3757 ins_req(i, C->top()); // build a new edge
3758 }
3759 Node* new_st = st->clone();
3760 new_st->set_req(MemNode::Control, in(Control));
3761 new_st->set_req(MemNode::Memory, prev_mem);
3762 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3763 new_st = phase->transform(new_st);
3764
3765 // At this point, new_st might have swallowed a pre-existing store
3766 // at the same offset, or perhaps new_st might have disappeared,
3767 // if it redundantly stored the same value (or zero to fresh memory).
3768
3769 // In any case, wire it in:
3770 phase->igvn_rehash_node_delayed(this);
3771 set_req(i, new_st);
3772
3773 // The caller may now kill the old guy.
3774 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3775 assert(check_st == new_st || check_st == NULL, "must be findable");
3776 assert(!is_complete(), "");
3777 return new_st;
3778 }
3779
3780 static bool store_constant(jlong* tiles, int num_tiles,
3781 intptr_t st_off, int st_size,
3782 jlong con) {
3783 if ((st_off & (st_size-1)) != 0)
3784 return false; // strange store offset (assume size==2**N)
3785 address addr = (address)tiles + st_off;
3786 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3787 switch (st_size) {
3788 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3789 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3790 case sizeof(jint): *(jint*) addr = (jint) con; break;
3791 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3792 default: return false; // strange store size (detect size!=2**N here)
3793 }
3794 return true; // return success to caller
3795 }
3796
3797 // Coalesce subword constants into int constants and possibly
3798 // into long constants. The goal, if the CPU permits,
3799 // is to initialize the object with a small number of 64-bit tiles.
3800 // Also, convert floating-point constants to bit patterns.
3801 // Non-constants are not relevant to this pass.
3802 //
3803 // In terms of the running example on InitializeNode::InitializeNode
3804 // and InitializeNode::capture_store, here is the transformation
3805 // of rawstore1 and rawstore2 into rawstore12:
3806 // alloc = (Allocate ...)
3807 // rawoop = alloc.RawAddress
3808 // tile12 = 0x00010002
3809 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3810 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3811 //
3812 void
3813 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3814 Node* size_in_bytes,
3815 PhaseGVN* phase) {
3816 Compile* C = phase->C;
3817
3818 assert(stores_are_sane(phase), "");
3819 // Note: After this pass, they are not completely sane,
3820 // since there may be some overlaps.
3821
3822 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3823
3824 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3825 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3826 size_limit = MIN2(size_limit, ti_limit);
3827 size_limit = align_up(size_limit, BytesPerLong);
3828 int num_tiles = size_limit / BytesPerLong;
3829
3830 // allocate space for the tile map:
3831 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3832 jlong tiles_buf[small_len];
3833 Node* nodes_buf[small_len];
3834 jlong inits_buf[small_len];
3835 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3836 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3837 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3838 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3839 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3840 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3841 // tiles: exact bitwise model of all primitive constants
3842 // nodes: last constant-storing node subsumed into the tiles model
3843 // inits: which bytes (in each tile) are touched by any initializations
3844
3845 //// Pass A: Fill in the tile model with any relevant stores.
3846
3847 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3848 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3849 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3850 Node* zmem = zero_memory(); // initially zero memory state
3851 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3852 Node* st = in(i);
3853 intptr_t st_off = get_store_offset(st, phase);
3854
3855 // Figure out the store's offset and constant value:
3856 if (st_off < header_size) continue; //skip (ignore header)
3857 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3858 int st_size = st->as_Store()->memory_size();
3859 if (st_off + st_size > size_limit) break;
3860
3861 // Record which bytes are touched, whether by constant or not.
3862 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3863 continue; // skip (strange store size)
3864
3865 const Type* val = phase->type(st->in(MemNode::ValueIn));
3866 if (!val->singleton()) continue; //skip (non-con store)
3867 BasicType type = val->basic_type();
3868
3869 jlong con = 0;
3870 switch (type) {
3871 case T_INT: con = val->is_int()->get_con(); break;
3872 case T_LONG: con = val->is_long()->get_con(); break;
3873 case T_FLOAT: con = jint_cast(val->getf()); break;
3874 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3875 default: continue; //skip (odd store type)
3876 }
3877
3878 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3879 st->Opcode() == Op_StoreL) {
3880 continue; // This StoreL is already optimal.
3881 }
3882
3883 // Store down the constant.
3884 store_constant(tiles, num_tiles, st_off, st_size, con);
3885
3886 intptr_t j = st_off >> LogBytesPerLong;
3887
3888 if (type == T_INT && st_size == BytesPerInt
3889 && (st_off & BytesPerInt) == BytesPerInt) {
3890 jlong lcon = tiles[j];
3891 if (!Matcher::isSimpleConstant64(lcon) &&
3892 st->Opcode() == Op_StoreI) {
3893 // This StoreI is already optimal by itself.
3894 jint* intcon = (jint*) &tiles[j];
3895 intcon[1] = 0; // undo the store_constant()
3896
3897 // If the previous store is also optimal by itself, back up and
3898 // undo the action of the previous loop iteration... if we can.
3899 // But if we can't, just let the previous half take care of itself.
3900 st = nodes[j];
3901 st_off -= BytesPerInt;
3902 con = intcon[0];
3903 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3904 assert(st_off >= header_size, "still ignoring header");
3905 assert(get_store_offset(st, phase) == st_off, "must be");
3906 assert(in(i-1) == zmem, "must be");
3907 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3908 assert(con == tcon->is_int()->get_con(), "must be");
3909 // Undo the effects of the previous loop trip, which swallowed st:
3910 intcon[0] = 0; // undo store_constant()
3911 set_req(i-1, st); // undo set_req(i, zmem)
3912 nodes[j] = NULL; // undo nodes[j] = st
3913 --old_subword; // undo ++old_subword
3914 }
3915 continue; // This StoreI is already optimal.
3916 }
3917 }
3918
3919 // This store is not needed.
3920 set_req(i, zmem);
3921 nodes[j] = st; // record for the moment
3922 if (st_size < BytesPerLong) // something has changed
3923 ++old_subword; // includes int/float, but who's counting...
3924 else ++old_long;
3925 }
3926
3927 if ((old_subword + old_long) == 0)
3928 return; // nothing more to do
3929
3930 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3931 // Be sure to insert them before overlapping non-constant stores.
3932 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3933 for (int j = 0; j < num_tiles; j++) {
3934 jlong con = tiles[j];
3935 jlong init = inits[j];
3936 if (con == 0) continue;
3937 jint con0, con1; // split the constant, address-wise
3938 jint init0, init1; // split the init map, address-wise
3939 { union { jlong con; jint intcon[2]; } u;
3940 u.con = con;
3941 con0 = u.intcon[0];
3942 con1 = u.intcon[1];
3943 u.con = init;
3944 init0 = u.intcon[0];
3945 init1 = u.intcon[1];
3946 }
3947
3948 Node* old = nodes[j];
3949 assert(old != NULL, "need the prior store");
3950 intptr_t offset = (j * BytesPerLong);
3951
3952 bool split = !Matcher::isSimpleConstant64(con);
3953
3954 if (offset < header_size) {
3955 assert(offset + BytesPerInt >= header_size, "second int counts");
3956 assert(*(jint*)&tiles[j] == 0, "junk in header");
3957 split = true; // only the second word counts
3958 // Example: int a[] = { 42 ... }
3959 } else if (con0 == 0 && init0 == -1) {
3960 split = true; // first word is covered by full inits
3961 // Example: int a[] = { ... foo(), 42 ... }
3962 } else if (con1 == 0 && init1 == -1) {
3963 split = true; // second word is covered by full inits
3964 // Example: int a[] = { ... 42, foo() ... }
3965 }
3966
3967 // Here's a case where init0 is neither 0 nor -1:
3968 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3969 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3970 // In this case the tile is not split; it is (jlong)42.
3971 // The big tile is stored down, and then the foo() value is inserted.
3972 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3973
3974 Node* ctl = old->in(MemNode::Control);
3975 Node* adr = make_raw_address(offset, phase);
3976 const TypePtr* atp = TypeRawPtr::BOTTOM;
3977
3978 // One or two coalesced stores to plop down.
3979 Node* st[2];
3980 intptr_t off[2];
3981 int nst = 0;
3982 if (!split) {
3983 ++new_long;
3984 off[nst] = offset;
3985 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3986 phase->longcon(con), T_LONG, MemNode::unordered);
3987 } else {
3988 // Omit either if it is a zero.
3989 if (con0 != 0) {
3990 ++new_int;
3991 off[nst] = offset;
3992 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3993 phase->intcon(con0), T_INT, MemNode::unordered);
3994 }
3995 if (con1 != 0) {
3996 ++new_int;
3997 offset += BytesPerInt;
3998 adr = make_raw_address(offset, phase);
3999 off[nst] = offset;
4000 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4001 phase->intcon(con1), T_INT, MemNode::unordered);
4002 }
4003 }
4004
4005 // Insert second store first, then the first before the second.
4006 // Insert each one just before any overlapping non-constant stores.
4007 while (nst > 0) {
4008 Node* st1 = st[--nst];
4009 C->copy_node_notes_to(st1, old);
4010 st1 = phase->transform(st1);
4011 offset = off[nst];
4012 assert(offset >= header_size, "do not smash header");
4013 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4014 guarantee(ins_idx != 0, "must re-insert constant store");
4015 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4016 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4017 set_req(--ins_idx, st1);
4018 else
4019 ins_req(ins_idx, st1);
4020 }
4021 }
4022
4023 if (PrintCompilation && WizardMode)
4024 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4025 old_subword, old_long, new_int, new_long);
4026 if (C->log() != NULL)
4027 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4028 old_subword, old_long, new_int, new_long);
4029
4030 // Clean up any remaining occurrences of zmem:
4031 remove_extra_zeroes();
4032 }
4033
4034 // Explore forward from in(start) to find the first fully initialized
4035 // word, and return its offset. Skip groups of subword stores which
4036 // together initialize full words. If in(start) is itself part of a
4037 // fully initialized word, return the offset of in(start). If there
4038 // are no following full-word stores, or if something is fishy, return
4039 // a negative value.
4040 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4041 int int_map = 0;
4042 intptr_t int_map_off = 0;
4043 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4044
4045 for (uint i = start, limit = req(); i < limit; i++) {
4046 Node* st = in(i);
4047
4048 intptr_t st_off = get_store_offset(st, phase);
4049 if (st_off < 0) break; // return conservative answer
4050
4051 int st_size = st->as_Store()->memory_size();
4052 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4053 return st_off; // we found a complete word init
4054 }
4055
4056 // update the map:
4057
4058 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4059 if (this_int_off != int_map_off) {
4060 // reset the map:
4061 int_map = 0;
4062 int_map_off = this_int_off;
4063 }
4064
4065 int subword_off = st_off - this_int_off;
4066 int_map |= right_n_bits(st_size) << subword_off;
4067 if ((int_map & FULL_MAP) == FULL_MAP) {
4068 return this_int_off; // we found a complete word init
4069 }
4070
4071 // Did this store hit or cross the word boundary?
4072 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4073 if (next_int_off == this_int_off + BytesPerInt) {
4074 // We passed the current int, without fully initializing it.
4075 int_map_off = next_int_off;
4076 int_map >>= BytesPerInt;
4077 } else if (next_int_off > this_int_off + BytesPerInt) {
4078 // We passed the current and next int.
4079 return this_int_off + BytesPerInt;
4080 }
4081 }
4082
4083 return -1;
4084 }
4085
4086
4087 // Called when the associated AllocateNode is expanded into CFG.
4088 // At this point, we may perform additional optimizations.
4089 // Linearize the stores by ascending offset, to make memory
4090 // activity as coherent as possible.
4091 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4092 intptr_t header_size,
4093 Node* size_in_bytes,
4094 PhaseGVN* phase) {
4095 assert(!is_complete(), "not already complete");
4096 assert(stores_are_sane(phase), "");
4097 assert(allocation() != NULL, "must be present");
4098
4099 remove_extra_zeroes();
4100
4101 if (ReduceFieldZeroing || ReduceBulkZeroing)
4102 // reduce instruction count for common initialization patterns
4103 coalesce_subword_stores(header_size, size_in_bytes, phase);
4104
4105 Node* zmem = zero_memory(); // initially zero memory state
4106 Node* inits = zmem; // accumulating a linearized chain of inits
4107 #ifdef ASSERT
4108 intptr_t first_offset = allocation()->minimum_header_size();
4109 intptr_t last_init_off = first_offset; // previous init offset
4110 intptr_t last_init_end = first_offset; // previous init offset+size
4111 intptr_t last_tile_end = first_offset; // previous tile offset+size
4112 #endif
4113 intptr_t zeroes_done = header_size;
4114
4115 bool do_zeroing = true; // we might give up if inits are very sparse
4116 int big_init_gaps = 0; // how many large gaps have we seen?
4117
4118 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4119 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4120
4121 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4122 Node* st = in(i);
4123 intptr_t st_off = get_store_offset(st, phase);
4124 if (st_off < 0)
4125 break; // unknown junk in the inits
4126 if (st->in(MemNode::Memory) != zmem)
4127 break; // complicated store chains somehow in list
4128
4129 int st_size = st->as_Store()->memory_size();
4130 intptr_t next_init_off = st_off + st_size;
4131
4132 if (do_zeroing && zeroes_done < next_init_off) {
4133 // See if this store needs a zero before it or under it.
4134 intptr_t zeroes_needed = st_off;
4135
4136 if (st_size < BytesPerInt) {
4137 // Look for subword stores which only partially initialize words.
4138 // If we find some, we must lay down some word-level zeroes first,
4139 // underneath the subword stores.
4140 //
4141 // Examples:
4142 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4143 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4144 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4145 //
4146 // Note: coalesce_subword_stores may have already done this,
4147 // if it was prompted by constant non-zero subword initializers.
4148 // But this case can still arise with non-constant stores.
4149
4150 intptr_t next_full_store = find_next_fullword_store(i, phase);
4151
4152 // In the examples above:
4153 // in(i) p q r s x y z
4154 // st_off 12 13 14 15 12 13 14
4155 // st_size 1 1 1 1 1 1 1
4156 // next_full_s. 12 16 16 16 16 16 16
4157 // z's_done 12 16 16 16 12 16 12
4158 // z's_needed 12 16 16 16 16 16 16
4159 // zsize 0 0 0 0 4 0 4
4160 if (next_full_store < 0) {
4161 // Conservative tack: Zero to end of current word.
4162 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4163 } else {
4164 // Zero to beginning of next fully initialized word.
4165 // Or, don't zero at all, if we are already in that word.
4166 assert(next_full_store >= zeroes_needed, "must go forward");
4167 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4168 zeroes_needed = next_full_store;
4169 }
4170 }
4171
4172 if (zeroes_needed > zeroes_done) {
4173 intptr_t zsize = zeroes_needed - zeroes_done;
4174 // Do some incremental zeroing on rawmem, in parallel with inits.
4175 zeroes_done = align_down(zeroes_done, BytesPerInt);
4176 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4177 zeroes_done, zeroes_needed,
4178 phase);
4179 zeroes_done = zeroes_needed;
4180 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4181 do_zeroing = false; // leave the hole, next time
4182 }
4183 }
4184
4185 // Collect the store and move on:
4186 st->set_req(MemNode::Memory, inits);
4187 inits = st; // put it on the linearized chain
4188 set_req(i, zmem); // unhook from previous position
4189
4190 if (zeroes_done == st_off)
4191 zeroes_done = next_init_off;
4192
4193 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4194
4195 #ifdef ASSERT
4196 // Various order invariants. Weaker than stores_are_sane because
4197 // a large constant tile can be filled in by smaller non-constant stores.
4198 assert(st_off >= last_init_off, "inits do not reverse");
4199 last_init_off = st_off;
4200 const Type* val = NULL;
4201 if (st_size >= BytesPerInt &&
4202 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4203 (int)val->basic_type() < (int)T_OBJECT) {
4204 assert(st_off >= last_tile_end, "tiles do not overlap");
4205 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4206 last_tile_end = MAX2(last_tile_end, next_init_off);
4207 } else {
4208 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4209 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4210 assert(st_off >= last_init_end, "inits do not overlap");
4211 last_init_end = next_init_off; // it's a non-tile
4212 }
4213 #endif //ASSERT
4214 }
4215
4216 remove_extra_zeroes(); // clear out all the zmems left over
4217 add_req(inits);
4218
4219 if (!(UseTLAB && ZeroTLAB)) {
4220 // If anything remains to be zeroed, zero it all now.
4221 zeroes_done = align_down(zeroes_done, BytesPerInt);
4222 // if it is the last unused 4 bytes of an instance, forget about it
4223 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4224 if (zeroes_done + BytesPerLong >= size_limit) {
4225 AllocateNode* alloc = allocation();
4226 assert(alloc != NULL, "must be present");
4227 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4228 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4229 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4230 if (zeroes_done == k->layout_helper())
4231 zeroes_done = size_limit;
4232 }
4233 }
4234 if (zeroes_done < size_limit) {
4235 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4236 zeroes_done, size_in_bytes, phase);
4237 }
4238 }
4239
4240 set_complete(phase);
4241 return rawmem;
4242 }
4243
4244
4245 #ifdef ASSERT
4246 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4247 if (is_complete())
4248 return true; // stores could be anything at this point
4249 assert(allocation() != NULL, "must be present");
4250 intptr_t last_off = allocation()->minimum_header_size();
4251 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4252 Node* st = in(i);
4253 intptr_t st_off = get_store_offset(st, phase);
4254 if (st_off < 0) continue; // ignore dead garbage
4255 if (last_off > st_off) {
4256 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4257 this->dump(2);
4258 assert(false, "ascending store offsets");
4259 return false;
4260 }
4261 last_off = st_off + st->as_Store()->memory_size();
4262 }
4263 return true;
4264 }
4265 #endif //ASSERT
4266
4267
4268
4269
4270 //============================MergeMemNode=====================================
4271 //
4272 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4273 // contributing store or call operations. Each contributor provides the memory
4274 // state for a particular "alias type" (see Compile::alias_type). For example,
4275 // if a MergeMem has an input X for alias category #6, then any memory reference
4276 // to alias category #6 may use X as its memory state input, as an exact equivalent
4277 // to using the MergeMem as a whole.
4278 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4279 //
4280 // (Here, the <N> notation gives the index of the relevant adr_type.)
4281 //
4282 // In one special case (and more cases in the future), alias categories overlap.
4283 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4284 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4285 // it is exactly equivalent to that state W:
4286 // MergeMem(<Bot>: W) <==> W
4287 //
4288 // Usually, the merge has more than one input. In that case, where inputs
4289 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4290 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4291 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4292 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4293 //
4294 // A merge can take a "wide" memory state as one of its narrow inputs.
4295 // This simply means that the merge observes out only the relevant parts of
4296 // the wide input. That is, wide memory states arriving at narrow merge inputs
4297 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4298 //
4299 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4300 // and that memory slices "leak through":
4301 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4302 //
4303 // But, in such a cascade, repeated memory slices can "block the leak":
4304 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4305 //
4306 // In the last example, Y is not part of the combined memory state of the
4307 // outermost MergeMem. The system must, of course, prevent unschedulable
4308 // memory states from arising, so you can be sure that the state Y is somehow
4309 // a precursor to state Y'.
4310 //
4311 //
4312 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4313 // of each MergeMemNode array are exactly the numerical alias indexes, including
4314 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4315 // Compile::alias_type (and kin) produce and manage these indexes.
4316 //
4317 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4318 // (Note that this provides quick access to the top node inside MergeMem methods,
4319 // without the need to reach out via TLS to Compile::current.)
4320 //
4321 // As a consequence of what was just described, a MergeMem that represents a full
4322 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4323 // containing all alias categories.
4324 //
4325 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4326 //
4327 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4328 // a memory state for the alias type <N>, or else the top node, meaning that
4329 // there is no particular input for that alias type. Note that the length of
4330 // a MergeMem is variable, and may be extended at any time to accommodate new
4331 // memory states at larger alias indexes. When merges grow, they are of course
4332 // filled with "top" in the unused in() positions.
4333 //
4334 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4335 // (Top was chosen because it works smoothly with passes like GCM.)
4336 //
4337 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4338 // the type of random VM bits like TLS references.) Since it is always the
4339 // first non-Bot memory slice, some low-level loops use it to initialize an
4340 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4341 //
4342 //
4343 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4344 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4345 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4346 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4347 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4348 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4349 //
4350 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4351 // really that different from the other memory inputs. An abbreviation called
4352 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4353 //
4354 //
4355 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4356 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4357 // that "emerges though" the base memory will be marked as excluding the alias types
4358 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4359 //
4360 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4361 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4362 //
4363 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4364 // (It is currently unimplemented.) As you can see, the resulting merge is
4365 // actually a disjoint union of memory states, rather than an overlay.
4366 //
4367
4368 //------------------------------MergeMemNode-----------------------------------
4369 Node* MergeMemNode::make_empty_memory() {
4370 Node* empty_memory = (Node*) Compile::current()->top();
4371 assert(empty_memory->is_top(), "correct sentinel identity");
4372 return empty_memory;
4373 }
4374
4375 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4376 init_class_id(Class_MergeMem);
4377 // all inputs are nullified in Node::Node(int)
4378 // set_input(0, NULL); // no control input
4379
4380 // Initialize the edges uniformly to top, for starters.
4381 Node* empty_mem = make_empty_memory();
4382 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4383 init_req(i,empty_mem);
4384 }
4385 assert(empty_memory() == empty_mem, "");
4386
4387 if( new_base != NULL && new_base->is_MergeMem() ) {
4388 MergeMemNode* mdef = new_base->as_MergeMem();
4389 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4390 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4391 mms.set_memory(mms.memory2());
4392 }
4393 assert(base_memory() == mdef->base_memory(), "");
4394 } else {
4395 set_base_memory(new_base);
4396 }
4397 }
4398
4399 // Make a new, untransformed MergeMem with the same base as 'mem'.
4400 // If mem is itself a MergeMem, populate the result with the same edges.
4401 MergeMemNode* MergeMemNode::make(Node* mem) {
4402 return new MergeMemNode(mem);
4403 }
4404
4405 //------------------------------cmp--------------------------------------------
4406 uint MergeMemNode::hash() const { return NO_HASH; }
4407 uint MergeMemNode::cmp( const Node &n ) const {
4408 return (&n == this); // Always fail except on self
4409 }
4410
4411 //------------------------------Identity---------------------------------------
4412 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4413 // Identity if this merge point does not record any interesting memory
4414 // disambiguations.
4415 Node* base_mem = base_memory();
4416 Node* empty_mem = empty_memory();
4417 if (base_mem != empty_mem) { // Memory path is not dead?
4418 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4419 Node* mem = in(i);
4420 if (mem != empty_mem && mem != base_mem) {
4421 return this; // Many memory splits; no change
4422 }
4423 }
4424 }
4425 return base_mem; // No memory splits; ID on the one true input
4426 }
4427
4428 //------------------------------Ideal------------------------------------------
4429 // This method is invoked recursively on chains of MergeMem nodes
4430 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4431 // Remove chain'd MergeMems
4432 //
4433 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4434 // relative to the "in(Bot)". Since we are patching both at the same time,
4435 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4436 // but rewrite each "in(i)" relative to the new "in(Bot)".
4437 Node *progress = NULL;
4438
4439
4440 Node* old_base = base_memory();
4441 Node* empty_mem = empty_memory();
4442 if (old_base == empty_mem)
4443 return NULL; // Dead memory path.
4444
4445 MergeMemNode* old_mbase;
4446 if (old_base != NULL && old_base->is_MergeMem())
4447 old_mbase = old_base->as_MergeMem();
4448 else
4449 old_mbase = NULL;
4450 Node* new_base = old_base;
4451
4452 // simplify stacked MergeMems in base memory
4453 if (old_mbase) new_base = old_mbase->base_memory();
4454
4455 // the base memory might contribute new slices beyond my req()
4456 if (old_mbase) grow_to_match(old_mbase);
4457
4458 // Look carefully at the base node if it is a phi.
4459 PhiNode* phi_base;
4460 if (new_base != NULL && new_base->is_Phi())
4461 phi_base = new_base->as_Phi();
4462 else
4463 phi_base = NULL;
4464
4465 Node* phi_reg = NULL;
4466 uint phi_len = (uint)-1;
4467 if (phi_base != NULL && !phi_base->is_copy()) {
4468 // do not examine phi if degraded to a copy
4469 phi_reg = phi_base->region();
4470 phi_len = phi_base->req();
4471 // see if the phi is unfinished
4472 for (uint i = 1; i < phi_len; i++) {
4473 if (phi_base->in(i) == NULL) {
4474 // incomplete phi; do not look at it yet!
4475 phi_reg = NULL;
4476 phi_len = (uint)-1;
4477 break;
4478 }
4479 }
4480 }
4481
4482 // Note: We do not call verify_sparse on entry, because inputs
4483 // can normalize to the base_memory via subsume_node or similar
4484 // mechanisms. This method repairs that damage.
4485
4486 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4487
4488 // Look at each slice.
4489 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4490 Node* old_in = in(i);
4491 // calculate the old memory value
4492 Node* old_mem = old_in;
4493 if (old_mem == empty_mem) old_mem = old_base;
4494 assert(old_mem == memory_at(i), "");
4495
4496 // maybe update (reslice) the old memory value
4497
4498 // simplify stacked MergeMems
4499 Node* new_mem = old_mem;
4500 MergeMemNode* old_mmem;
4501 if (old_mem != NULL && old_mem->is_MergeMem())
4502 old_mmem = old_mem->as_MergeMem();
4503 else
4504 old_mmem = NULL;
4505 if (old_mmem == this) {
4506 // This can happen if loops break up and safepoints disappear.
4507 // A merge of BotPtr (default) with a RawPtr memory derived from a
4508 // safepoint can be rewritten to a merge of the same BotPtr with
4509 // the BotPtr phi coming into the loop. If that phi disappears
4510 // also, we can end up with a self-loop of the mergemem.
4511 // In general, if loops degenerate and memory effects disappear,
4512 // a mergemem can be left looking at itself. This simply means
4513 // that the mergemem's default should be used, since there is
4514 // no longer any apparent effect on this slice.
4515 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4516 // from start. Update the input to TOP.
4517 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4518 }
4519 else if (old_mmem != NULL) {
4520 new_mem = old_mmem->memory_at(i);
4521 }
4522 // else preceding memory was not a MergeMem
4523
4524 // replace equivalent phis (unfortunately, they do not GVN together)
4525 if (new_mem != NULL && new_mem != new_base &&
4526 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4527 if (new_mem->is_Phi()) {
4528 PhiNode* phi_mem = new_mem->as_Phi();
4529 for (uint i = 1; i < phi_len; i++) {
4530 if (phi_base->in(i) != phi_mem->in(i)) {
4531 phi_mem = NULL;
4532 break;
4533 }
4534 }
4535 if (phi_mem != NULL) {
4536 // equivalent phi nodes; revert to the def
4537 new_mem = new_base;
4538 }
4539 }
4540 }
4541
4542 // maybe store down a new value
4543 Node* new_in = new_mem;
4544 if (new_in == new_base) new_in = empty_mem;
4545
4546 if (new_in != old_in) {
4547 // Warning: Do not combine this "if" with the previous "if"
4548 // A memory slice might have be be rewritten even if it is semantically
4549 // unchanged, if the base_memory value has changed.
4550 set_req(i, new_in);
4551 progress = this; // Report progress
4552 }
4553 }
4554
4555 if (new_base != old_base) {
4556 set_req(Compile::AliasIdxBot, new_base);
4557 // Don't use set_base_memory(new_base), because we need to update du.
4558 assert(base_memory() == new_base, "");
4559 progress = this;
4560 }
4561
4562 if( base_memory() == this ) {
4563 // a self cycle indicates this memory path is dead
4564 set_req(Compile::AliasIdxBot, empty_mem);
4565 }
4566
4567 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4568 // Recursion must occur after the self cycle check above
4569 if( base_memory()->is_MergeMem() ) {
4570 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4571 Node *m = phase->transform(new_mbase); // Rollup any cycles
4572 if( m != NULL &&
4573 (m->is_top() ||
4574 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4575 // propagate rollup of dead cycle to self
4576 set_req(Compile::AliasIdxBot, empty_mem);
4577 }
4578 }
4579
4580 if( base_memory() == empty_mem ) {
4581 progress = this;
4582 // Cut inputs during Parse phase only.
4583 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4584 if( !can_reshape ) {
4585 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4586 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4587 }
4588 }
4589 }
4590
4591 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4592 // Check if PhiNode::Ideal's "Split phis through memory merges"
4593 // transform should be attempted. Look for this->phi->this cycle.
4594 uint merge_width = req();
4595 if (merge_width > Compile::AliasIdxRaw) {
4596 PhiNode* phi = base_memory()->as_Phi();
4597 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4598 if (phi->in(i) == this) {
4599 phase->is_IterGVN()->_worklist.push(phi);
4600 break;
4601 }
4602 }
4603 }
4604 }
4605
4606 assert(progress || verify_sparse(), "please, no dups of base");
4607 return progress;
4608 }
4609
4610 //-------------------------set_base_memory-------------------------------------
4611 void MergeMemNode::set_base_memory(Node *new_base) {
4612 Node* empty_mem = empty_memory();
4613 set_req(Compile::AliasIdxBot, new_base);
4614 assert(memory_at(req()) == new_base, "must set default memory");
4615 // Clear out other occurrences of new_base:
4616 if (new_base != empty_mem) {
4617 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4618 if (in(i) == new_base) set_req(i, empty_mem);
4619 }
4620 }
4621 }
4622
4623 //------------------------------out_RegMask------------------------------------
4624 const RegMask &MergeMemNode::out_RegMask() const {
4625 return RegMask::Empty;
4626 }
4627
4628 //------------------------------dump_spec--------------------------------------
4629 #ifndef PRODUCT
4630 void MergeMemNode::dump_spec(outputStream *st) const {
4631 st->print(" {");
4632 Node* base_mem = base_memory();
4633 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4634 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4635 if (mem == base_mem) { st->print(" -"); continue; }
4636 st->print( " N%d:", mem->_idx );
4637 Compile::current()->get_adr_type(i)->dump_on(st);
4638 }
4639 st->print(" }");
4640 }
4641 #endif // !PRODUCT
4642
4643
4644 #ifdef ASSERT
4645 static bool might_be_same(Node* a, Node* b) {
4646 if (a == b) return true;
4647 if (!(a->is_Phi() || b->is_Phi())) return false;
4648 // phis shift around during optimization
4649 return true; // pretty stupid...
4650 }
4651
4652 // verify a narrow slice (either incoming or outgoing)
4653 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4654 if (!VerifyAliases) return; // don't bother to verify unless requested
4655 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4656 if (Node::in_dump()) return; // muzzle asserts when printing
4657 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4658 assert(n != NULL, "");
4659 // Elide intervening MergeMem's
4660 while (n->is_MergeMem()) {
4661 n = n->as_MergeMem()->memory_at(alias_idx);
4662 }
4663 Compile* C = Compile::current();
4664 const TypePtr* n_adr_type = n->adr_type();
4665 if (n == m->empty_memory()) {
4666 // Implicit copy of base_memory()
4667 } else if (n_adr_type != TypePtr::BOTTOM) {
4668 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4669 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4670 } else {
4671 // A few places like make_runtime_call "know" that VM calls are narrow,
4672 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4673 bool expected_wide_mem = false;
4674 if (n == m->base_memory()) {
4675 expected_wide_mem = true;
4676 } else if (alias_idx == Compile::AliasIdxRaw ||
4677 n == m->memory_at(Compile::AliasIdxRaw)) {
4678 expected_wide_mem = true;
4679 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4680 // memory can "leak through" calls on channels that
4681 // are write-once. Allow this also.
4682 expected_wide_mem = true;
4683 }
4684 assert(expected_wide_mem, "expected narrow slice replacement");
4685 }
4686 }
4687 #else // !ASSERT
4688 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4689 #endif
4690
4691
4692 //-----------------------------memory_at---------------------------------------
4693 Node* MergeMemNode::memory_at(uint alias_idx) const {
4694 assert(alias_idx >= Compile::AliasIdxRaw ||
4695 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4696 "must avoid base_memory and AliasIdxTop");
4697
4698 // Otherwise, it is a narrow slice.
4699 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4700 Compile *C = Compile::current();
4701 if (is_empty_memory(n)) {
4702 // the array is sparse; empty slots are the "top" node
4703 n = base_memory();
4704 assert(Node::in_dump()
4705 || n == NULL || n->bottom_type() == Type::TOP
4706 || n->adr_type() == NULL // address is TOP
4707 || n->adr_type() == TypePtr::BOTTOM
4708 || n->adr_type() == TypeRawPtr::BOTTOM
4709 || Compile::current()->AliasLevel() == 0,
4710 "must be a wide memory");
4711 // AliasLevel == 0 if we are organizing the memory states manually.
4712 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4713 } else {
4714 // make sure the stored slice is sane
4715 #ifdef ASSERT
4716 if (VMError::is_error_reported() || Node::in_dump()) {
4717 } else if (might_be_same(n, base_memory())) {
4718 // Give it a pass: It is a mostly harmless repetition of the base.
4719 // This can arise normally from node subsumption during optimization.
4720 } else {
4721 verify_memory_slice(this, alias_idx, n);
4722 }
4723 #endif
4724 }
4725 return n;
4726 }
4727
4728 //---------------------------set_memory_at-------------------------------------
4729 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4730 verify_memory_slice(this, alias_idx, n);
4731 Node* empty_mem = empty_memory();
4732 if (n == base_memory()) n = empty_mem; // collapse default
4733 uint need_req = alias_idx+1;
4734 if (req() < need_req) {
4735 if (n == empty_mem) return; // already the default, so do not grow me
4736 // grow the sparse array
4737 do {
4738 add_req(empty_mem);
4739 } while (req() < need_req);
4740 }
4741 set_req( alias_idx, n );
4742 }
4743
4744
4745
4746 //--------------------------iteration_setup------------------------------------
4747 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4748 if (other != NULL) {
4749 grow_to_match(other);
4750 // invariant: the finite support of mm2 is within mm->req()
4751 #ifdef ASSERT
4752 for (uint i = req(); i < other->req(); i++) {
4753 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4754 }
4755 #endif
4756 }
4757 // Replace spurious copies of base_memory by top.
4758 Node* base_mem = base_memory();
4759 if (base_mem != NULL && !base_mem->is_top()) {
4760 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4761 if (in(i) == base_mem)
4762 set_req(i, empty_memory());
4763 }
4764 }
4765 }
4766
4767 //---------------------------grow_to_match-------------------------------------
4768 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4769 Node* empty_mem = empty_memory();
4770 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4771 // look for the finite support of the other memory
4772 for (uint i = other->req(); --i >= req(); ) {
4773 if (other->in(i) != empty_mem) {
4774 uint new_len = i+1;
4775 while (req() < new_len) add_req(empty_mem);
4776 break;
4777 }
4778 }
4779 }
4780
4781 //---------------------------verify_sparse-------------------------------------
4782 #ifndef PRODUCT
4783 bool MergeMemNode::verify_sparse() const {
4784 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4785 Node* base_mem = base_memory();
4786 // The following can happen in degenerate cases, since empty==top.
4787 if (is_empty_memory(base_mem)) return true;
4788 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4789 assert(in(i) != NULL, "sane slice");
4790 if (in(i) == base_mem) return false; // should have been the sentinel value!
4791 }
4792 return true;
4793 }
4794
4795 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4796 Node* n;
4797 n = mm->in(idx);
4798 if (mem == n) return true; // might be empty_memory()
4799 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4800 if (mem == n) return true;
4801 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4802 if (mem == n) return true;
4803 if (n == NULL) break;
4804 }
4805 return false;
4806 }
4807 #endif // !PRODUCT