1 /*
2 * Copyright (c) 1997, 2019, 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 "opto/rootnode.hpp"
48 #include "opto/valuetypenode.hpp"
49 #include "utilities/align.hpp"
50 #include "utilities/copy.hpp"
51 #include "utilities/macros.hpp"
52 #include "utilities/vmError.hpp"
53 #if INCLUDE_ZGC
54 #include "gc/z/c2/zBarrierSetC2.hpp"
55 #endif
56
57 // Portions of code courtesy of Clifford Click
58
59 // Optimization - Graph Style
60
61 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
62
63 //=============================================================================
64 uint MemNode::size_of() const { return sizeof(*this); }
65
66 const TypePtr *MemNode::adr_type() const {
67 Node* adr = in(Address);
68 if (adr == NULL) return NULL; // node is dead
69 const TypePtr* cross_check = NULL;
70 DEBUG_ONLY(cross_check = _adr_type);
71 return calculate_adr_type(adr->bottom_type(), cross_check);
72 }
73
74 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
75 if (adr != NULL) {
76 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
77 return true;
78 }
79 }
80 return false;
81 }
82
83 #ifndef PRODUCT
84 void MemNode::dump_spec(outputStream *st) const {
85 if (in(Address) == NULL) return; // node is dead
86 #ifndef ASSERT
87 // fake the missing field
88 const TypePtr* _adr_type = NULL;
89 if (in(Address) != NULL)
90 _adr_type = in(Address)->bottom_type()->isa_ptr();
91 #endif
92 dump_adr_type(this, _adr_type, st);
93
94 Compile* C = Compile::current();
95 if (C->alias_type(_adr_type)->is_volatile()) {
96 st->print(" Volatile!");
97 }
98 if (_unaligned_access) {
99 st->print(" unaligned");
100 }
101 if (_mismatched_access) {
102 st->print(" mismatched");
103 }
104 if (_unsafe_access) {
105 st->print(" unsafe");
106 }
107 }
108
109 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
110 st->print(" @");
111 if (adr_type == NULL) {
112 st->print("NULL");
113 } else {
114 adr_type->dump_on(st);
115 Compile* C = Compile::current();
116 Compile::AliasType* atp = NULL;
117 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
118 if (atp == NULL)
119 st->print(", idx=?\?;");
120 else if (atp->index() == Compile::AliasIdxBot)
121 st->print(", idx=Bot;");
122 else if (atp->index() == Compile::AliasIdxTop)
123 st->print(", idx=Top;");
124 else if (atp->index() == Compile::AliasIdxRaw)
125 st->print(", idx=Raw;");
126 else {
127 ciField* field = atp->field();
128 if (field) {
129 st->print(", name=");
130 field->print_name_on(st);
131 }
132 st->print(", idx=%d;", atp->index());
133 }
134 }
135 }
136
137 extern void print_alias_types();
138
139 #endif
140
141 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
142 assert((t_oop != NULL), "sanity");
143 bool is_instance = t_oop->is_known_instance_field();
144 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
145 (load != NULL) && load->is_Load() &&
146 (phase->is_IterGVN() != NULL);
147 if (!(is_instance || is_boxed_value_load))
148 return mchain; // don't try to optimize non-instance types
149 uint instance_id = t_oop->instance_id();
150 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
151 Node *prev = NULL;
152 Node *result = mchain;
153 while (prev != result) {
154 prev = result;
155 if (result == start_mem)
156 break; // hit one of our sentinels
157 // skip over a call which does not affect this memory slice
158 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
159 Node *proj_in = result->in(0);
160 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
161 break; // hit one of our sentinels
162 } else if (proj_in->is_Call()) {
163 // ArrayCopyNodes processed here as well
164 CallNode *call = proj_in->as_Call();
165 if (!call->may_modify(t_oop, phase)) { // returns false for instances
166 result = call->in(TypeFunc::Memory);
167 }
168 } else if (proj_in->is_Initialize()) {
169 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
170 // Stop if this is the initialization for the object instance which
171 // contains this memory slice, otherwise skip over it.
172 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
173 break;
174 }
175 if (is_instance) {
176 result = proj_in->in(TypeFunc::Memory);
177 } else if (is_boxed_value_load) {
178 Node* klass = alloc->in(AllocateNode::KlassNode);
179 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
180 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
181 result = proj_in->in(TypeFunc::Memory); // not related allocation
182 }
183 }
184 } else if (proj_in->is_MemBar()) {
185 ArrayCopyNode* ac = NULL;
186 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
187 break;
188 }
189 result = proj_in->in(TypeFunc::Memory);
190 } else {
191 assert(false, "unexpected projection");
192 }
193 } else if (result->is_ClearArray()) {
194 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
195 // Can not bypass initialization of the instance
196 // we are looking for.
197 break;
198 }
199 // Otherwise skip it (the call updated 'result' value).
200 } else if (result->is_MergeMem()) {
201 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
202 }
203 }
204 return result;
205 }
206
207 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
208 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
209 if (t_oop == NULL)
210 return mchain; // don't try to optimize non-oop types
211 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
212 bool is_instance = t_oop->is_known_instance_field();
213 PhaseIterGVN *igvn = phase->is_IterGVN();
214 if (is_instance && igvn != NULL && result->is_Phi()) {
215 PhiNode *mphi = result->as_Phi();
216 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
217 const TypePtr *t = mphi->adr_type();
218 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
219 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
220 t->is_oopptr()->cast_to_exactness(true)
221 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
222 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
223 // clone the Phi with our address type
224 result = mphi->split_out_instance(t_adr, igvn);
225 } else {
226 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
227 }
228 }
229 return result;
230 }
231
232 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
233 uint alias_idx = phase->C->get_alias_index(tp);
234 Node *mem = mmem;
235 #ifdef ASSERT
236 {
237 // Check that current type is consistent with the alias index used during graph construction
238 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
239 bool consistent = adr_check == NULL || adr_check->empty() ||
240 phase->C->must_alias(adr_check, alias_idx );
241 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
242 if( !consistent && adr_check != NULL && !adr_check->empty() &&
243 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
244 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
245 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
246 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
247 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
248 // don't assert if it is dead code.
249 consistent = true;
250 }
251 if( !consistent ) {
252 st->print("alias_idx==%d, adr_check==", alias_idx);
253 if( adr_check == NULL ) {
254 st->print("NULL");
255 } else {
256 adr_check->dump();
257 }
258 st->cr();
259 print_alias_types();
260 assert(consistent, "adr_check must match alias idx");
261 }
262 }
263 #endif
264 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
265 // means an array I have not precisely typed yet. Do not do any
266 // alias stuff with it any time soon.
267 const TypeOopPtr *toop = tp->isa_oopptr();
268 if( tp->base() != Type::AnyPtr &&
269 !(toop &&
270 toop->klass() != NULL &&
271 toop->klass()->is_java_lang_Object() &&
272 toop->offset() == Type::OffsetBot) ) {
273 // compress paths and change unreachable cycles to TOP
274 // If not, we can update the input infinitely along a MergeMem cycle
275 // Equivalent code in PhiNode::Ideal
276 Node* m = phase->transform(mmem);
277 // If transformed to a MergeMem, get the desired slice
278 // Otherwise the returned node represents memory for every slice
279 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
280 // Update input if it is progress over what we have now
281 }
282 return mem;
283 }
284
285 //--------------------------Ideal_common---------------------------------------
286 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
287 // Unhook non-raw memories from complete (macro-expanded) initializations.
288 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
289 // If our control input is a dead region, kill all below the region
290 Node *ctl = in(MemNode::Control);
291 if (ctl && remove_dead_region(phase, can_reshape))
292 return this;
293 ctl = in(MemNode::Control);
294 // Don't bother trying to transform a dead node
295 if (ctl && ctl->is_top()) return NodeSentinel;
296
297 PhaseIterGVN *igvn = phase->is_IterGVN();
298 // Wait if control on the worklist.
299 if (ctl && can_reshape && igvn != NULL) {
300 Node* bol = NULL;
301 Node* cmp = NULL;
302 if (ctl->in(0)->is_If()) {
303 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
304 bol = ctl->in(0)->in(1);
305 if (bol->is_Bool())
306 cmp = ctl->in(0)->in(1)->in(1);
307 }
308 if (igvn->_worklist.member(ctl) ||
309 (bol != NULL && igvn->_worklist.member(bol)) ||
310 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
311 // This control path may be dead.
312 // Delay this memory node transformation until the control is processed.
313 phase->is_IterGVN()->_worklist.push(this);
314 return NodeSentinel; // caller will return NULL
315 }
316 }
317 // Ignore if memory is dead, or self-loop
318 Node *mem = in(MemNode::Memory);
319 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
320 assert(mem != this, "dead loop in MemNode::Ideal");
321
322 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
323 // This memory slice may be dead.
324 // Delay this mem node transformation until the memory is processed.
325 phase->is_IterGVN()->_worklist.push(this);
326 return NodeSentinel; // caller will return NULL
327 }
328
329 Node *address = in(MemNode::Address);
330 const Type *t_adr = phase->type(address);
331 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
332
333 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
334 // Unsafe off-heap access with zero address. Remove access and other control users
335 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
336 assert(ctl != NULL, "unsafe accesses should be control dependent");
337 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
338 Node* u = ctl->fast_out(i);
339 if (u != ctl) {
340 igvn->rehash_node_delayed(u);
341 int nb = u->replace_edge(ctl, phase->C->top());
342 --i, imax -= nb;
343 }
344 }
345 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
346 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
347 phase->C->root()->add_req(halt);
348 return this;
349 }
350
351 if (can_reshape && igvn != NULL &&
352 (igvn->_worklist.member(address) ||
353 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
354 // The address's base and type may change when the address is processed.
355 // Delay this mem node transformation until the address is processed.
356 phase->is_IterGVN()->_worklist.push(this);
357 return NodeSentinel; // caller will return NULL
358 }
359
360 // Do NOT remove or optimize the next lines: ensure a new alias index
361 // is allocated for an oop pointer type before Escape Analysis.
362 // Note: C++ will not remove it since the call has side effect.
363 if (t_adr->isa_oopptr()) {
364 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
365 }
366
367 Node* base = NULL;
368 if (address->is_AddP()) {
369 base = address->in(AddPNode::Base);
370 }
371 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
372 !t_adr->isa_rawptr()) {
373 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
374 // Skip this node optimization if its address has TOP base.
375 return NodeSentinel; // caller will return NULL
376 }
377
378 // Avoid independent memory operations
379 Node* old_mem = mem;
380
381 // The code which unhooks non-raw memories from complete (macro-expanded)
382 // initializations was removed. After macro-expansion all stores catched
383 // by Initialize node became raw stores and there is no information
384 // which memory slices they modify. So it is unsafe to move any memory
385 // operation above these stores. Also in most cases hooked non-raw memories
386 // were already unhooked by using information from detect_ptr_independence()
387 // and find_previous_store().
388
389 if (mem->is_MergeMem()) {
390 MergeMemNode* mmem = mem->as_MergeMem();
391 const TypePtr *tp = t_adr->is_ptr();
392
393 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
394 }
395
396 if (mem != old_mem) {
397 set_req(MemNode::Memory, mem);
398 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
399 igvn->_worklist.push(old_mem);
400 }
401 if (phase->type(mem) == Type::TOP) return NodeSentinel;
402 return this;
403 }
404
405 // let the subclass continue analyzing...
406 return NULL;
407 }
408
409 // Helper function for proving some simple control dominations.
410 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
411 // Already assumes that 'dom' is available at 'sub', and that 'sub'
412 // is not a constant (dominated by the method's StartNode).
413 // Used by MemNode::find_previous_store to prove that the
414 // control input of a memory operation predates (dominates)
415 // an allocation it wants to look past.
416 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
417 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
418 return false; // Conservative answer for dead code
419
420 // Check 'dom'. Skip Proj and CatchProj nodes.
421 dom = dom->find_exact_control(dom);
422 if (dom == NULL || dom->is_top())
423 return false; // Conservative answer for dead code
424
425 if (dom == sub) {
426 // For the case when, for example, 'sub' is Initialize and the original
427 // 'dom' is Proj node of the 'sub'.
428 return false;
429 }
430
431 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
432 return true;
433
434 // 'dom' dominates 'sub' if its control edge and control edges
435 // of all its inputs dominate or equal to sub's control edge.
436
437 // Currently 'sub' is either Allocate, Initialize or Start nodes.
438 // Or Region for the check in LoadNode::Ideal();
439 // 'sub' should have sub->in(0) != NULL.
440 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
441 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
442
443 // Get control edge of 'sub'.
444 Node* orig_sub = sub;
445 sub = sub->find_exact_control(sub->in(0));
446 if (sub == NULL || sub->is_top())
447 return false; // Conservative answer for dead code
448
449 assert(sub->is_CFG(), "expecting control");
450
451 if (sub == dom)
452 return true;
453
454 if (sub->is_Start() || sub->is_Root())
455 return false;
456
457 {
458 // Check all control edges of 'dom'.
459
460 ResourceMark rm;
461 Arena* arena = Thread::current()->resource_area();
462 Node_List nlist(arena);
463 Unique_Node_List dom_list(arena);
464
465 dom_list.push(dom);
466 bool only_dominating_controls = false;
467
468 for (uint next = 0; next < dom_list.size(); next++) {
469 Node* n = dom_list.at(next);
470 if (n == orig_sub)
471 return false; // One of dom's inputs dominated by sub.
472 if (!n->is_CFG() && n->pinned()) {
473 // Check only own control edge for pinned non-control nodes.
474 n = n->find_exact_control(n->in(0));
475 if (n == NULL || n->is_top())
476 return false; // Conservative answer for dead code
477 assert(n->is_CFG(), "expecting control");
478 dom_list.push(n);
479 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
480 only_dominating_controls = true;
481 } else if (n->is_CFG()) {
482 if (n->dominates(sub, nlist))
483 only_dominating_controls = true;
484 else
485 return false;
486 } else {
487 // First, own control edge.
488 Node* m = n->find_exact_control(n->in(0));
489 if (m != NULL) {
490 if (m->is_top())
491 return false; // Conservative answer for dead code
492 dom_list.push(m);
493 }
494 // Now, the rest of edges.
495 uint cnt = n->req();
496 for (uint i = 1; i < cnt; i++) {
497 m = n->find_exact_control(n->in(i));
498 if (m == NULL || m->is_top())
499 continue;
500 dom_list.push(m);
501 }
502 }
503 }
504 return only_dominating_controls;
505 }
506 }
507
508 //---------------------detect_ptr_independence---------------------------------
509 // Used by MemNode::find_previous_store to prove that two base
510 // pointers are never equal.
511 // The pointers are accompanied by their associated allocations,
512 // if any, which have been previously discovered by the caller.
513 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
514 Node* p2, AllocateNode* a2,
515 PhaseTransform* phase) {
516 // Attempt to prove that these two pointers cannot be aliased.
517 // They may both manifestly be allocations, and they should differ.
518 // Or, if they are not both allocations, they can be distinct constants.
519 // Otherwise, one is an allocation and the other a pre-existing value.
520 if (a1 == NULL && a2 == NULL) { // neither an allocation
521 return (p1 != p2) && p1->is_Con() && p2->is_Con();
522 } else if (a1 != NULL && a2 != NULL) { // both allocations
523 return (a1 != a2);
524 } else if (a1 != NULL) { // one allocation a1
525 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
526 return all_controls_dominate(p2, a1);
527 } else { //(a2 != NULL) // one allocation a2
528 return all_controls_dominate(p1, a2);
529 }
530 return false;
531 }
532
533
534 // Find an arraycopy that must have set (can_see_stored_value=true) or
535 // could have set (can_see_stored_value=false) the value for this load
536 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
537 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
538 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
539 Node* mb = mem->in(0);
540 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
541 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
542 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
543 if (ac->is_clonebasic()) {
544 intptr_t offset;
545 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
546 if (alloc != NULL && alloc == ld_alloc) {
547 return ac;
548 }
549 }
550 }
551 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
552 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
553
554 if (ac->is_arraycopy_validated() ||
555 ac->is_copyof_validated() ||
556 ac->is_copyofrange_validated()) {
557 Node* ld_addp = in(MemNode::Address);
558 if (ld_addp->is_AddP()) {
559 Node* ld_base = ld_addp->in(AddPNode::Address);
560 Node* ld_offs = ld_addp->in(AddPNode::Offset);
561
562 Node* dest = ac->in(ArrayCopyNode::Dest);
563
564 if (dest == ld_base) {
565 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
566 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
567 return ac;
568 }
569 if (!can_see_stored_value) {
570 mem = ac->in(TypeFunc::Memory);
571 }
572 }
573 }
574 }
575 }
576 return NULL;
577 }
578
579 // The logic for reordering loads and stores uses four steps:
580 // (a) Walk carefully past stores and initializations which we
581 // can prove are independent of this load.
582 // (b) Observe that the next memory state makes an exact match
583 // with self (load or store), and locate the relevant store.
584 // (c) Ensure that, if we were to wire self directly to the store,
585 // the optimizer would fold it up somehow.
586 // (d) Do the rewiring, and return, depending on some other part of
587 // the optimizer to fold up the load.
588 // This routine handles steps (a) and (b). Steps (c) and (d) are
589 // specific to loads and stores, so they are handled by the callers.
590 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
591 //
592 Node* MemNode::find_previous_store(PhaseTransform* phase) {
593 Node* ctrl = in(MemNode::Control);
594 Node* adr = in(MemNode::Address);
595 intptr_t offset = 0;
596 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
597 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
598
599 if (offset == Type::OffsetBot)
600 return NULL; // cannot unalias unless there are precise offsets
601
602 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
603 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
604
605 intptr_t size_in_bytes = memory_size();
606
607 Node* mem = in(MemNode::Memory); // start searching here...
608
609 int cnt = 50; // Cycle limiter
610 for (;;) { // While we can dance past unrelated stores...
611 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
612
613 Node* prev = mem;
614 if (mem->is_Store()) {
615 Node* st_adr = mem->in(MemNode::Address);
616 intptr_t st_offset = 0;
617 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
618 if (st_base == NULL)
619 break; // inscrutable pointer
620
621 // For raw accesses it's not enough to prove that constant offsets don't intersect.
622 // We need the bases to be the equal in order for the offset check to make sense.
623 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
624 break;
625 }
626
627 if (st_offset != offset && st_offset != Type::OffsetBot) {
628 const int MAX_STORE = BytesPerLong;
629 if (st_offset >= offset + size_in_bytes ||
630 st_offset <= offset - MAX_STORE ||
631 st_offset <= offset - mem->as_Store()->memory_size()) {
632 // Success: The offsets are provably independent.
633 // (You may ask, why not just test st_offset != offset and be done?
634 // The answer is that stores of different sizes can co-exist
635 // in the same sequence of RawMem effects. We sometimes initialize
636 // a whole 'tile' of array elements with a single jint or jlong.)
637 mem = mem->in(MemNode::Memory);
638 continue; // (a) advance through independent store memory
639 }
640 }
641 if (st_base != base &&
642 detect_ptr_independence(base, alloc,
643 st_base,
644 AllocateNode::Ideal_allocation(st_base, phase),
645 phase)) {
646 // Success: The bases are provably independent.
647 mem = mem->in(MemNode::Memory);
648 continue; // (a) advance through independent store memory
649 }
650
651 // (b) At this point, if the bases or offsets do not agree, we lose,
652 // since we have not managed to prove 'this' and 'mem' independent.
653 if (st_base == base && st_offset == offset) {
654 return mem; // let caller handle steps (c), (d)
655 }
656
657 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
658 InitializeNode* st_init = mem->in(0)->as_Initialize();
659 AllocateNode* st_alloc = st_init->allocation();
660 if (st_alloc == NULL)
661 break; // something degenerated
662 bool known_identical = false;
663 bool known_independent = false;
664 if (alloc == st_alloc)
665 known_identical = true;
666 else if (alloc != NULL)
667 known_independent = true;
668 else if (all_controls_dominate(this, st_alloc))
669 known_independent = true;
670
671 if (known_independent) {
672 // The bases are provably independent: Either they are
673 // manifestly distinct allocations, or else the control
674 // of this load dominates the store's allocation.
675 int alias_idx = phase->C->get_alias_index(adr_type());
676 if (alias_idx == Compile::AliasIdxRaw) {
677 mem = st_alloc->in(TypeFunc::Memory);
678 } else {
679 mem = st_init->memory(alias_idx);
680 }
681 continue; // (a) advance through independent store memory
682 }
683
684 // (b) at this point, if we are not looking at a store initializing
685 // the same allocation we are loading from, we lose.
686 if (known_identical) {
687 // From caller, can_see_stored_value will consult find_captured_store.
688 return mem; // let caller handle steps (c), (d)
689 }
690
691 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
692 if (prev != mem) {
693 // Found an arraycopy but it doesn't affect that load
694 continue;
695 }
696 // Found an arraycopy that may affect that load
697 return mem;
698 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
699 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
700 if (mem->is_Proj() && mem->in(0)->is_Call()) {
701 // ArrayCopyNodes processed here as well.
702 CallNode *call = mem->in(0)->as_Call();
703 if (!call->may_modify(addr_t, phase)) {
704 mem = call->in(TypeFunc::Memory);
705 continue; // (a) advance through independent call memory
706 }
707 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
708 ArrayCopyNode* ac = NULL;
709 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
710 break;
711 }
712 mem = mem->in(0)->in(TypeFunc::Memory);
713 continue; // (a) advance through independent MemBar memory
714 } else if (mem->is_ClearArray()) {
715 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
716 // (the call updated 'mem' value)
717 continue; // (a) advance through independent allocation memory
718 } else {
719 // Can not bypass initialization of the instance
720 // we are looking for.
721 return mem;
722 }
723 } else if (mem->is_MergeMem()) {
724 int alias_idx = phase->C->get_alias_index(adr_type());
725 mem = mem->as_MergeMem()->memory_at(alias_idx);
726 continue; // (a) advance through independent MergeMem memory
727 }
728 }
729
730 // Unless there is an explicit 'continue', we must bail out here,
731 // because 'mem' is an inscrutable memory state (e.g., a call).
732 break;
733 }
734
735 return NULL; // bail out
736 }
737
738 //----------------------calculate_adr_type-------------------------------------
739 // Helper function. Notices when the given type of address hits top or bottom.
740 // Also, asserts a cross-check of the type against the expected address type.
741 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
742 if (t == Type::TOP) return NULL; // does not touch memory any more?
743 #ifdef PRODUCT
744 cross_check = NULL;
745 #else
746 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
747 #endif
748 const TypePtr* tp = t->isa_ptr();
749 if (tp == NULL) {
750 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
751 return TypePtr::BOTTOM; // touches lots of memory
752 } else {
753 #ifdef ASSERT
754 // %%%% [phh] We don't check the alias index if cross_check is
755 // TypeRawPtr::BOTTOM. Needs to be investigated.
756 if (cross_check != NULL &&
757 cross_check != TypePtr::BOTTOM &&
758 cross_check != TypeRawPtr::BOTTOM) {
759 // Recheck the alias index, to see if it has changed (due to a bug).
760 Compile* C = Compile::current();
761 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
762 "must stay in the original alias category");
763 // The type of the address must be contained in the adr_type,
764 // disregarding "null"-ness.
765 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
766 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
767 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
768 "real address must not escape from expected memory type");
769 }
770 #endif
771 return tp;
772 }
773 }
774
775 //=============================================================================
776 // Should LoadNode::Ideal() attempt to remove control edges?
777 bool LoadNode::can_remove_control() const {
778 return true;
779 }
780 uint LoadNode::size_of() const { return sizeof(*this); }
781 bool LoadNode::cmp( const Node &n ) const
782 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
783 const Type *LoadNode::bottom_type() const { return _type; }
784 uint LoadNode::ideal_reg() const {
785 return _type->ideal_reg();
786 }
787
788 #ifndef PRODUCT
789 void LoadNode::dump_spec(outputStream *st) const {
790 MemNode::dump_spec(st);
791 if( !Verbose && !WizardMode ) {
792 // standard dump does this in Verbose and WizardMode
793 st->print(" #"); _type->dump_on(st);
794 }
795 if (!depends_only_on_test()) {
796 st->print(" (does not depend only on test)");
797 }
798 }
799 #endif
800
801 #ifdef ASSERT
802 //----------------------------is_immutable_value-------------------------------
803 // Helper function to allow a raw load without control edge for some cases
804 bool LoadNode::is_immutable_value(Node* adr) {
805 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
806 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
807 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
808 in_bytes(JavaThread::osthread_offset())));
809 }
810 #endif
811
812 //----------------------------LoadNode::make-----------------------------------
813 // Polymorphic factory method:
814 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
815 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
816 Compile* C = gvn.C;
817
818 // sanity check the alias category against the created node type
819 assert(!(adr_type->isa_oopptr() &&
820 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
821 "use LoadKlassNode instead");
822 assert(!(adr_type->isa_aryptr() &&
823 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
824 "use LoadRangeNode instead");
825 // Check control edge of raw loads
826 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
827 // oop will be recorded in oop map if load crosses safepoint
828 rt->isa_oopptr() || is_immutable_value(adr),
829 "raw memory operations should have control edge");
830 LoadNode* load = NULL;
831 switch (bt) {
832 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
833 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
834 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
835 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
836 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
837 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
838 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
839 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
840 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
841 case T_VALUETYPE:
842 case T_OBJECT:
843 #ifdef _LP64
844 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
845 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
846 } else
847 #endif
848 {
849 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
850 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
851 }
852 break;
853 default:
854 ShouldNotReachHere();
855 break;
856 }
857 assert(load != NULL, "LoadNode should have been created");
858 if (unaligned) {
859 load->set_unaligned_access();
860 }
861 if (mismatched) {
862 load->set_mismatched_access();
863 }
864 if (unsafe) {
865 load->set_unsafe_access();
866 }
867 if (load->Opcode() == Op_LoadN) {
868 Node* ld = gvn.transform(load);
869 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
870 }
871
872 return load;
873 }
874
875 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
876 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
877 bool require_atomic = true;
878 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
879 if (unaligned) {
880 load->set_unaligned_access();
881 }
882 if (mismatched) {
883 load->set_mismatched_access();
884 }
885 if (unsafe) {
886 load->set_unsafe_access();
887 }
888 return load;
889 }
890
891 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
892 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
893 bool require_atomic = true;
894 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
895 if (unaligned) {
896 load->set_unaligned_access();
897 }
898 if (mismatched) {
899 load->set_mismatched_access();
900 }
901 if (unsafe) {
902 load->set_unsafe_access();
903 }
904 return load;
905 }
906
907
908
909 //------------------------------hash-------------------------------------------
910 uint LoadNode::hash() const {
911 // unroll addition of interesting fields
912 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
913 }
914
915 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
916 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
917 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
918 bool is_stable_ary = FoldStableValues &&
919 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
920 tp->isa_aryptr()->is_stable();
921
922 return (eliminate_boxing && non_volatile) || is_stable_ary;
923 }
924
925 return false;
926 }
927
928 // Is the value loaded previously stored by an arraycopy? If so return
929 // a load node that reads from the source array so we may be able to
930 // optimize out the ArrayCopy node later.
931 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
932 Node* ld_adr = in(MemNode::Address);
933 intptr_t ld_off = 0;
934 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
935 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
936 if (ac != NULL) {
937 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
938
939 Node* mem = ac->in(TypeFunc::Memory);
940 Node* ctl = ac->in(0);
941 Node* src = ac->in(ArrayCopyNode::Src);
942
943 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
944 return NULL;
945 }
946
947 LoadNode* ld = clone()->as_Load();
948 Node* addp = in(MemNode::Address)->clone();
949 if (ac->as_ArrayCopy()->is_clonebasic()) {
950 assert(ld_alloc != NULL, "need an alloc");
951 assert(addp->is_AddP(), "address must be addp");
952 assert(ac->in(ArrayCopyNode::Dest)->is_AddP(), "dest must be an address");
953 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
954 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base)), "strange pattern");
955 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address)), "strange pattern");
956 addp->set_req(AddPNode::Base, src->in(AddPNode::Base));
957 addp->set_req(AddPNode::Address, src->in(AddPNode::Address));
958 } else {
959 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
960 ac->as_ArrayCopy()->is_copyof_validated() ||
961 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
962 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
963 addp->set_req(AddPNode::Base, src);
964 addp->set_req(AddPNode::Address, src);
965
966 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
967 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
968 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
969 uint shift = exact_log2(type2aelembytes(ary_elem));
970
971 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
972 #ifdef _LP64
973 diff = phase->transform(new ConvI2LNode(diff));
974 #endif
975 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
976
977 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
978 addp->set_req(AddPNode::Offset, offset);
979 }
980 addp = phase->transform(addp);
981 #ifdef ASSERT
982 const TypePtr* adr_type = phase->type(addp)->is_ptr();
983 ld->_adr_type = adr_type;
984 #endif
985 ld->set_req(MemNode::Address, addp);
986 ld->set_req(0, ctl);
987 ld->set_req(MemNode::Memory, mem);
988 // load depends on the tests that validate the arraycopy
989 ld->_control_dependency = UnknownControl;
990 return ld;
991 }
992 return NULL;
993 }
994
995
996 //---------------------------can_see_stored_value------------------------------
997 // This routine exists to make sure this set of tests is done the same
998 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
999 // will change the graph shape in a way which makes memory alive twice at the
1000 // same time (uses the Oracle model of aliasing), then some
1001 // LoadXNode::Identity will fold things back to the equivalence-class model
1002 // of aliasing.
1003 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
1004 Node* ld_adr = in(MemNode::Address);
1005 intptr_t ld_off = 0;
1006 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1007 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase);
1008 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1009 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1010 // This is more general than load from boxing objects.
1011 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1012 uint alias_idx = atp->index();
1013 bool final = !atp->is_rewritable();
1014 Node* result = NULL;
1015 Node* current = st;
1016 // Skip through chains of MemBarNodes checking the MergeMems for
1017 // new states for the slice of this load. Stop once any other
1018 // kind of node is encountered. Loads from final memory can skip
1019 // through any kind of MemBar but normal loads shouldn't skip
1020 // through MemBarAcquire since the could allow them to move out of
1021 // a synchronized region.
1022 while (current->is_Proj()) {
1023 int opc = current->in(0)->Opcode();
1024 if ((final && (opc == Op_MemBarAcquire ||
1025 opc == Op_MemBarAcquireLock ||
1026 opc == Op_LoadFence)) ||
1027 opc == Op_MemBarRelease ||
1028 opc == Op_StoreFence ||
1029 opc == Op_MemBarReleaseLock ||
1030 opc == Op_MemBarStoreStore ||
1031 opc == Op_MemBarCPUOrder) {
1032 Node* mem = current->in(0)->in(TypeFunc::Memory);
1033 if (mem->is_MergeMem()) {
1034 MergeMemNode* merge = mem->as_MergeMem();
1035 Node* new_st = merge->memory_at(alias_idx);
1036 if (new_st == merge->base_memory()) {
1037 // Keep searching
1038 current = new_st;
1039 continue;
1040 }
1041 // Save the new memory state for the slice and fall through
1042 // to exit.
1043 result = new_st;
1044 }
1045 }
1046 break;
1047 }
1048 if (result != NULL) {
1049 st = result;
1050 }
1051 }
1052
1053 // Loop around twice in the case Load -> Initialize -> Store.
1054 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1055 for (int trip = 0; trip <= 1; trip++) {
1056
1057 if (st->is_Store()) {
1058 Node* st_adr = st->in(MemNode::Address);
1059 if (!phase->eqv(st_adr, ld_adr)) {
1060 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1061 intptr_t st_off = 0;
1062 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1063 if (ld_base == NULL) return NULL;
1064 if (st_base == NULL) return NULL;
1065 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL;
1066 if (ld_off != st_off) return NULL;
1067 if (ld_off == Type::OffsetBot) return NULL;
1068 // Same base, same offset.
1069 // Possible improvement for arrays: check index value instead of absolute offset.
1070
1071 // At this point we have proven something like this setup:
1072 // B = << base >>
1073 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1074 // S = StoreQ(AddP( B , #Off), V)
1075 // (Actually, we haven't yet proven the Q's are the same.)
1076 // In other words, we are loading from a casted version of
1077 // the same pointer-and-offset that we stored to.
1078 // Casted version may carry a dependency and it is respected.
1079 // Thus, we are able to replace L by V.
1080 }
1081 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1082 if (store_Opcode() != st->Opcode())
1083 return NULL;
1084 return st->in(MemNode::ValueIn);
1085 }
1086
1087 // A load from a freshly-created object always returns zero.
1088 // (This can happen after LoadNode::Ideal resets the load's memory input
1089 // to find_captured_store, which returned InitializeNode::zero_memory.)
1090 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1091 (st->in(0) == ld_alloc) &&
1092 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1093 // return a zero value for the load's basic type
1094 // (This is one of the few places where a generic PhaseTransform
1095 // can create new nodes. Think of it as lazily manifesting
1096 // virtually pre-existing constants.)
1097 assert(memory_type() != T_VALUETYPE, "should not be used for value types");
1098 Node* default_value = ld_alloc->in(AllocateNode::DefaultValue);
1099 if (default_value != NULL) {
1100 return default_value;
1101 }
1102 assert(ld_alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
1103 return phase->zerocon(memory_type());
1104 }
1105
1106 // A load from an initialization barrier can match a captured store.
1107 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1108 InitializeNode* init = st->in(0)->as_Initialize();
1109 AllocateNode* alloc = init->allocation();
1110 if ((alloc != NULL) && (alloc == ld_alloc)) {
1111 // examine a captured store value
1112 st = init->find_captured_store(ld_off, memory_size(), phase);
1113 if (st != NULL) {
1114 continue; // take one more trip around
1115 }
1116 }
1117 }
1118
1119 // Load boxed value from result of valueOf() call is input parameter.
1120 if (this->is_Load() && ld_adr->is_AddP() &&
1121 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1122 intptr_t ignore = 0;
1123 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1124 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1125 base = bs->step_over_gc_barrier(base);
1126 if (base != NULL && base->is_Proj() &&
1127 base->as_Proj()->_con == TypeFunc::Parms &&
1128 base->in(0)->is_CallStaticJava() &&
1129 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1130 return base->in(0)->in(TypeFunc::Parms);
1131 }
1132 }
1133
1134 break;
1135 }
1136
1137 return NULL;
1138 }
1139
1140 //----------------------is_instance_field_load_with_local_phi------------------
1141 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1142 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1143 in(Address)->is_AddP() ) {
1144 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1145 // Only instances and boxed values.
1146 if( t_oop != NULL &&
1147 (t_oop->is_ptr_to_boxed_value() ||
1148 t_oop->is_known_instance_field()) &&
1149 t_oop->offset() != Type::OffsetBot &&
1150 t_oop->offset() != Type::OffsetTop) {
1151 return true;
1152 }
1153 }
1154 return false;
1155 }
1156
1157 //------------------------------Identity---------------------------------------
1158 // Loads are identity if previous store is to same address
1159 Node* LoadNode::Identity(PhaseGVN* phase) {
1160 // Loading from a ValueTypePtr? The ValueTypePtr has the values of
1161 // all fields as input. Look for the field with matching offset.
1162 Node* addr = in(Address);
1163 intptr_t offset;
1164 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset);
1165 if (base != NULL && base->is_ValueTypePtr() && offset > oopDesc::klass_offset_in_bytes()) {
1166 Node* value = base->as_ValueTypePtr()->field_value_by_offset((int)offset, true);
1167 if (value->is_ValueType()) {
1168 // Non-flattened value type field
1169 ValueTypeNode* vt = value->as_ValueType();
1170 if (vt->is_allocated(phase)) {
1171 value = vt->get_oop();
1172 } else {
1173 // Not yet allocated, bail out
1174 value = NULL;
1175 }
1176 }
1177 if (value != NULL) {
1178 if (Opcode() == Op_LoadN) {
1179 // Encode oop value if we are loading a narrow oop
1180 assert(!phase->type(value)->isa_narrowoop(), "should already be decoded");
1181 value = phase->transform(new EncodePNode(value, bottom_type()));
1182 }
1183 return value;
1184 }
1185 }
1186
1187 // If the previous store-maker is the right kind of Store, and the store is
1188 // to the same address, then we are equal to the value stored.
1189 Node* mem = in(Memory);
1190 Node* value = can_see_stored_value(mem, phase);
1191 if( value ) {
1192 // byte, short & char stores truncate naturally.
1193 // A load has to load the truncated value which requires
1194 // some sort of masking operation and that requires an
1195 // Ideal call instead of an Identity call.
1196 if (memory_size() < BytesPerInt) {
1197 // If the input to the store does not fit with the load's result type,
1198 // it must be truncated via an Ideal call.
1199 if (!phase->type(value)->higher_equal(phase->type(this)))
1200 return this;
1201 }
1202 // (This works even when value is a Con, but LoadNode::Value
1203 // usually runs first, producing the singleton type of the Con.)
1204 return value;
1205 }
1206
1207 // Search for an existing data phi which was generated before for the same
1208 // instance's field to avoid infinite generation of phis in a loop.
1209 Node *region = mem->in(0);
1210 if (is_instance_field_load_with_local_phi(region)) {
1211 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1212 int this_index = phase->C->get_alias_index(addr_t);
1213 int this_offset = addr_t->offset();
1214 int this_iid = addr_t->instance_id();
1215 if (!addr_t->is_known_instance() &&
1216 addr_t->is_ptr_to_boxed_value()) {
1217 // Use _idx of address base (could be Phi node) for boxed values.
1218 intptr_t ignore = 0;
1219 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1220 if (base == NULL) {
1221 return this;
1222 }
1223 this_iid = base->_idx;
1224 }
1225 const Type* this_type = bottom_type();
1226 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1227 Node* phi = region->fast_out(i);
1228 if (phi->is_Phi() && phi != mem &&
1229 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1230 return phi;
1231 }
1232 }
1233 }
1234
1235 return this;
1236 }
1237
1238 // Construct an equivalent unsigned load.
1239 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1240 BasicType bt = T_ILLEGAL;
1241 const Type* rt = NULL;
1242 switch (Opcode()) {
1243 case Op_LoadUB: return this;
1244 case Op_LoadUS: return this;
1245 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1246 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1247 default:
1248 assert(false, "no unsigned variant: %s", Name());
1249 return NULL;
1250 }
1251 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1252 raw_adr_type(), rt, bt, _mo, _control_dependency,
1253 is_unaligned_access(), is_mismatched_access());
1254 }
1255
1256 // Construct an equivalent signed load.
1257 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1258 BasicType bt = T_ILLEGAL;
1259 const Type* rt = NULL;
1260 switch (Opcode()) {
1261 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1262 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1263 case Op_LoadB: // fall through
1264 case Op_LoadS: // fall through
1265 case Op_LoadI: // fall through
1266 case Op_LoadL: return this;
1267 default:
1268 assert(false, "no signed variant: %s", Name());
1269 return NULL;
1270 }
1271 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1272 raw_adr_type(), rt, bt, _mo, _control_dependency,
1273 is_unaligned_access(), is_mismatched_access());
1274 }
1275
1276 // We're loading from an object which has autobox behaviour.
1277 // If this object is result of a valueOf call we'll have a phi
1278 // merging a newly allocated object and a load from the cache.
1279 // We want to replace this load with the original incoming
1280 // argument to the valueOf call.
1281 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1282 assert(phase->C->eliminate_boxing(), "sanity");
1283 intptr_t ignore = 0;
1284 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1285 if ((base == NULL) || base->is_Phi()) {
1286 // Push the loads from the phi that comes from valueOf up
1287 // through it to allow elimination of the loads and the recovery
1288 // of the original value. It is done in split_through_phi().
1289 return NULL;
1290 } else if (base->is_Load() ||
1291 (base->is_DecodeN() && base->in(1)->is_Load())) {
1292 // Eliminate the load of boxed value for integer types from the cache
1293 // array by deriving the value from the index into the array.
1294 // Capture the offset of the load and then reverse the computation.
1295
1296 // Get LoadN node which loads a boxing object from 'cache' array.
1297 if (base->is_DecodeN()) {
1298 base = base->in(1);
1299 }
1300 if (!base->in(Address)->is_AddP()) {
1301 return NULL; // Complex address
1302 }
1303 AddPNode* address = base->in(Address)->as_AddP();
1304 Node* cache_base = address->in(AddPNode::Base);
1305 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1306 // Get ConP node which is static 'cache' field.
1307 cache_base = cache_base->in(1);
1308 }
1309 if ((cache_base != NULL) && cache_base->is_Con()) {
1310 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1311 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1312 Node* elements[4];
1313 int shift = exact_log2(type2aelembytes(T_OBJECT));
1314 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1315 if (count > 0 && elements[0]->is_Con() &&
1316 (count == 1 ||
1317 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1318 elements[1]->in(2) == phase->intcon(shift)))) {
1319 ciObjArray* array = base_type->const_oop()->as_obj_array();
1320 // Fetch the box object cache[0] at the base of the array and get its value
1321 ciInstance* box = array->obj_at(0)->as_instance();
1322 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1323 assert(ik->is_box_klass(), "sanity");
1324 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1325 if (ik->nof_nonstatic_fields() == 1) {
1326 // This should be true nonstatic_field_at requires calling
1327 // nof_nonstatic_fields so check it anyway
1328 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1329 BasicType bt = c.basic_type();
1330 // Only integer types have boxing cache.
1331 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1332 bt == T_BYTE || bt == T_SHORT ||
1333 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1334 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1335 if (cache_low != (int)cache_low) {
1336 return NULL; // should not happen since cache is array indexed by value
1337 }
1338 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1339 if (offset != (int)offset) {
1340 return NULL; // should not happen since cache is array indexed by value
1341 }
1342 // Add up all the offsets making of the address of the load
1343 Node* result = elements[0];
1344 for (int i = 1; i < count; i++) {
1345 result = phase->transform(new AddXNode(result, elements[i]));
1346 }
1347 // Remove the constant offset from the address and then
1348 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1349 // remove the scaling of the offset to recover the original index.
1350 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1351 // Peel the shift off directly but wrap it in a dummy node
1352 // since Ideal can't return existing nodes
1353 result = new RShiftXNode(result->in(1), phase->intcon(0));
1354 } else if (result->is_Add() && result->in(2)->is_Con() &&
1355 result->in(1)->Opcode() == Op_LShiftX &&
1356 result->in(1)->in(2) == phase->intcon(shift)) {
1357 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1358 // but for boxing cache access we know that X<<Z will not overflow
1359 // (there is range check) so we do this optimizatrion by hand here.
1360 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1361 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1362 } else {
1363 result = new RShiftXNode(result, phase->intcon(shift));
1364 }
1365 #ifdef _LP64
1366 if (bt != T_LONG) {
1367 result = new ConvL2INode(phase->transform(result));
1368 }
1369 #else
1370 if (bt == T_LONG) {
1371 result = new ConvI2LNode(phase->transform(result));
1372 }
1373 #endif
1374 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1375 // Need to preserve unboxing load type if it is unsigned.
1376 switch(this->Opcode()) {
1377 case Op_LoadUB:
1378 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1379 break;
1380 case Op_LoadUS:
1381 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1382 break;
1383 }
1384 return result;
1385 }
1386 }
1387 }
1388 }
1389 }
1390 return NULL;
1391 }
1392
1393 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1394 Node* region = phi->in(0);
1395 if (region == NULL) {
1396 return false; // Wait stable graph
1397 }
1398 uint cnt = phi->req();
1399 for (uint i = 1; i < cnt; i++) {
1400 Node* rc = region->in(i);
1401 if (rc == NULL || phase->type(rc) == Type::TOP)
1402 return false; // Wait stable graph
1403 Node* in = phi->in(i);
1404 if (in == NULL || phase->type(in) == Type::TOP)
1405 return false; // Wait stable graph
1406 }
1407 return true;
1408 }
1409 //------------------------------split_through_phi------------------------------
1410 // Split instance or boxed field load through Phi.
1411 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1412 Node* mem = in(Memory);
1413 Node* address = in(Address);
1414 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1415
1416 assert((t_oop != NULL) &&
1417 (t_oop->is_known_instance_field() ||
1418 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1419
1420 Compile* C = phase->C;
1421 intptr_t ignore = 0;
1422 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1423 bool base_is_phi = (base != NULL) && base->is_Phi();
1424 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1425 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1426 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1427
1428 if (!((mem->is_Phi() || base_is_phi) &&
1429 (load_boxed_values || t_oop->is_known_instance_field()))) {
1430 return NULL; // memory is not Phi
1431 }
1432
1433 if (mem->is_Phi()) {
1434 if (!stable_phi(mem->as_Phi(), phase)) {
1435 return NULL; // Wait stable graph
1436 }
1437 uint cnt = mem->req();
1438 // Check for loop invariant memory.
1439 if (cnt == 3) {
1440 for (uint i = 1; i < cnt; i++) {
1441 Node* in = mem->in(i);
1442 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1443 if (m == mem) {
1444 if (i == 1) {
1445 // if the first edge was a loop, check second edge too.
1446 // If both are replaceable - we are in an infinite loop
1447 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1448 if (n == mem) {
1449 break;
1450 }
1451 }
1452 set_req(Memory, mem->in(cnt - i));
1453 return this; // made change
1454 }
1455 }
1456 }
1457 }
1458 if (base_is_phi) {
1459 if (!stable_phi(base->as_Phi(), phase)) {
1460 return NULL; // Wait stable graph
1461 }
1462 uint cnt = base->req();
1463 // Check for loop invariant memory.
1464 if (cnt == 3) {
1465 for (uint i = 1; i < cnt; i++) {
1466 if (base->in(i) == base) {
1467 return NULL; // Wait stable graph
1468 }
1469 }
1470 }
1471 }
1472
1473 // Split through Phi (see original code in loopopts.cpp).
1474 assert(C->have_alias_type(t_oop), "instance should have alias type");
1475
1476 // Do nothing here if Identity will find a value
1477 // (to avoid infinite chain of value phis generation).
1478 if (!phase->eqv(this, phase->apply_identity(this)))
1479 return NULL;
1480
1481 // Select Region to split through.
1482 Node* region;
1483 if (!base_is_phi) {
1484 assert(mem->is_Phi(), "sanity");
1485 region = mem->in(0);
1486 // Skip if the region dominates some control edge of the address.
1487 if (!MemNode::all_controls_dominate(address, region))
1488 return NULL;
1489 } else if (!mem->is_Phi()) {
1490 assert(base_is_phi, "sanity");
1491 region = base->in(0);
1492 // Skip if the region dominates some control edge of the memory.
1493 if (!MemNode::all_controls_dominate(mem, region))
1494 return NULL;
1495 } else if (base->in(0) != mem->in(0)) {
1496 assert(base_is_phi && mem->is_Phi(), "sanity");
1497 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1498 region = base->in(0);
1499 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1500 region = mem->in(0);
1501 } else {
1502 return NULL; // complex graph
1503 }
1504 } else {
1505 assert(base->in(0) == mem->in(0), "sanity");
1506 region = mem->in(0);
1507 }
1508
1509 const Type* this_type = this->bottom_type();
1510 int this_index = C->get_alias_index(t_oop);
1511 int this_offset = t_oop->offset();
1512 int this_iid = t_oop->instance_id();
1513 if (!t_oop->is_known_instance() && load_boxed_values) {
1514 // Use _idx of address base for boxed values.
1515 this_iid = base->_idx;
1516 }
1517 PhaseIterGVN* igvn = phase->is_IterGVN();
1518 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1519 for (uint i = 1; i < region->req(); i++) {
1520 Node* x;
1521 Node* the_clone = NULL;
1522 if (region->in(i) == C->top()) {
1523 x = C->top(); // Dead path? Use a dead data op
1524 } else {
1525 x = this->clone(); // Else clone up the data op
1526 the_clone = x; // Remember for possible deletion.
1527 // Alter data node to use pre-phi inputs
1528 if (this->in(0) == region) {
1529 x->set_req(0, region->in(i));
1530 } else {
1531 x->set_req(0, NULL);
1532 }
1533 if (mem->is_Phi() && (mem->in(0) == region)) {
1534 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1535 }
1536 if (address->is_Phi() && address->in(0) == region) {
1537 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1538 }
1539 if (base_is_phi && (base->in(0) == region)) {
1540 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1541 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1542 x->set_req(Address, adr_x);
1543 }
1544 }
1545 // Check for a 'win' on some paths
1546 const Type *t = x->Value(igvn);
1547
1548 bool singleton = t->singleton();
1549
1550 // See comments in PhaseIdealLoop::split_thru_phi().
1551 if (singleton && t == Type::TOP) {
1552 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1553 }
1554
1555 if (singleton) {
1556 x = igvn->makecon(t);
1557 } else {
1558 // We now call Identity to try to simplify the cloned node.
1559 // Note that some Identity methods call phase->type(this).
1560 // Make sure that the type array is big enough for
1561 // our new node, even though we may throw the node away.
1562 // (This tweaking with igvn only works because x is a new node.)
1563 igvn->set_type(x, t);
1564 // If x is a TypeNode, capture any more-precise type permanently into Node
1565 // otherwise it will be not updated during igvn->transform since
1566 // igvn->type(x) is set to x->Value() already.
1567 x->raise_bottom_type(t);
1568 Node *y = igvn->apply_identity(x);
1569 if (y != x) {
1570 x = y;
1571 } else {
1572 y = igvn->hash_find_insert(x);
1573 if (y) {
1574 x = y;
1575 } else {
1576 // Else x is a new node we are keeping
1577 // We do not need register_new_node_with_optimizer
1578 // because set_type has already been called.
1579 igvn->_worklist.push(x);
1580 }
1581 }
1582 }
1583 if (x != the_clone && the_clone != NULL) {
1584 igvn->remove_dead_node(the_clone);
1585 }
1586 phi->set_req(i, x);
1587 }
1588 // Record Phi
1589 igvn->register_new_node_with_optimizer(phi);
1590 return phi;
1591 }
1592
1593 AllocateNode* LoadNode::is_new_object_mark_load(PhaseGVN *phase) const {
1594 if (Opcode() == Op_LoadX) {
1595 Node* address = in(MemNode::Address);
1596 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1597 Node* mem = in(MemNode::Memory);
1598 if (alloc != NULL && mem->is_Proj() &&
1599 mem->in(0) != NULL &&
1600 mem->in(0) == alloc->initialization() &&
1601 alloc->initialization()->proj_out_or_null(0) != NULL) {
1602 return alloc;
1603 }
1604 }
1605 return NULL;
1606 }
1607
1608
1609 //------------------------------Ideal------------------------------------------
1610 // If the load is from Field memory and the pointer is non-null, it might be possible to
1611 // zero out the control input.
1612 // If the offset is constant and the base is an object allocation,
1613 // try to hook me up to the exact initializing store.
1614 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1615 Node* p = MemNode::Ideal_common(phase, can_reshape);
1616 if (p) return (p == NodeSentinel) ? NULL : p;
1617
1618 Node* ctrl = in(MemNode::Control);
1619 Node* address = in(MemNode::Address);
1620 bool progress = false;
1621
1622 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1623 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1624
1625 // Skip up past a SafePoint control. Cannot do this for Stores because
1626 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1627 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1628 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1629 !addr_mark &&
1630 (depends_only_on_test() || has_unknown_control_dependency())) {
1631 ctrl = ctrl->in(0);
1632 set_req(MemNode::Control,ctrl);
1633 progress = true;
1634 }
1635
1636 intptr_t ignore = 0;
1637 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1638 if (base != NULL
1639 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1640 // Check for useless control edge in some common special cases
1641 if (in(MemNode::Control) != NULL
1642 && can_remove_control()
1643 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1644 && all_controls_dominate(base, phase->C->start())) {
1645 // A method-invariant, non-null address (constant or 'this' argument).
1646 set_req(MemNode::Control, NULL);
1647 progress = true;
1648 }
1649 }
1650
1651 Node* mem = in(MemNode::Memory);
1652 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1653
1654 if (can_reshape && (addr_t != NULL)) {
1655 // try to optimize our memory input
1656 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1657 if (opt_mem != mem) {
1658 set_req(MemNode::Memory, opt_mem);
1659 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1660 return this;
1661 }
1662 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1663 if ((t_oop != NULL) &&
1664 (t_oop->is_known_instance_field() ||
1665 t_oop->is_ptr_to_boxed_value())) {
1666 PhaseIterGVN *igvn = phase->is_IterGVN();
1667 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1668 // Delay this transformation until memory Phi is processed.
1669 phase->is_IterGVN()->_worklist.push(this);
1670 return NULL;
1671 }
1672 // Split instance field load through Phi.
1673 Node* result = split_through_phi(phase);
1674 if (result != NULL) return result;
1675
1676 if (t_oop->is_ptr_to_boxed_value()) {
1677 Node* result = eliminate_autobox(phase);
1678 if (result != NULL) return result;
1679 }
1680 }
1681 }
1682
1683 // Is there a dominating load that loads the same value? Leave
1684 // anything that is not a load of a field/array element (like
1685 // barriers etc.) alone
1686 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1687 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1688 Node *use = mem->fast_out(i);
1689 if (use != this &&
1690 use->Opcode() == Opcode() &&
1691 use->in(0) != NULL &&
1692 use->in(0) != in(0) &&
1693 use->in(Address) == in(Address)) {
1694 Node* ctl = in(0);
1695 for (int i = 0; i < 10 && ctl != NULL; i++) {
1696 ctl = IfNode::up_one_dom(ctl);
1697 if (ctl == use->in(0)) {
1698 set_req(0, use->in(0));
1699 return this;
1700 }
1701 }
1702 }
1703 }
1704 }
1705
1706 // Check for prior store with a different base or offset; make Load
1707 // independent. Skip through any number of them. Bail out if the stores
1708 // are in an endless dead cycle and report no progress. This is a key
1709 // transform for Reflection. However, if after skipping through the Stores
1710 // we can't then fold up against a prior store do NOT do the transform as
1711 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1712 // array memory alive twice: once for the hoisted Load and again after the
1713 // bypassed Store. This situation only works if EVERYBODY who does
1714 // anti-dependence work knows how to bypass. I.e. we need all
1715 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1716 // the alias index stuff. So instead, peek through Stores and IFF we can
1717 // fold up, do so.
1718 Node* prev_mem = find_previous_store(phase);
1719 if (prev_mem != NULL) {
1720 Node* value = can_see_arraycopy_value(prev_mem, phase);
1721 if (value != NULL) {
1722 return value;
1723 }
1724 }
1725 // Steps (a), (b): Walk past independent stores to find an exact match.
1726 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1727 // (c) See if we can fold up on the spot, but don't fold up here.
1728 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1729 // just return a prior value, which is done by Identity calls.
1730 if (can_see_stored_value(prev_mem, phase)) {
1731 // Make ready for step (d):
1732 set_req(MemNode::Memory, prev_mem);
1733 return this;
1734 }
1735 }
1736
1737 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1738 if (alloc != NULL && mem->is_Proj() &&
1739 mem->in(0) != NULL &&
1740 mem->in(0) == alloc->initialization() &&
1741 Opcode() == Op_LoadX &&
1742 alloc->initialization()->proj_out_or_null(0) != NULL) {
1743 InitializeNode* init = alloc->initialization();
1744 Node* control = init->proj_out(0);
1745 return alloc->make_ideal_mark(phase, address, control, mem);
1746 }
1747
1748 return progress ? this : NULL;
1749 }
1750
1751 // Helper to recognize certain Klass fields which are invariant across
1752 // some group of array types (e.g., int[] or all T[] where T < Object).
1753 const Type*
1754 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1755 ciKlass* klass) const {
1756 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1757 // The field is Klass::_modifier_flags. Return its (constant) value.
1758 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1759 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1760 return TypeInt::make(klass->modifier_flags());
1761 }
1762 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1763 // The field is Klass::_access_flags. Return its (constant) value.
1764 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1765 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1766 return TypeInt::make(klass->access_flags());
1767 }
1768 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1769 // The field is Klass::_layout_helper. Return its constant value if known.
1770 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1771 return TypeInt::make(klass->layout_helper());
1772 }
1773
1774 // No match.
1775 return NULL;
1776 }
1777
1778 //------------------------------Value-----------------------------------------
1779 const Type* LoadNode::Value(PhaseGVN* phase) const {
1780 // Either input is TOP ==> the result is TOP
1781 Node* mem = in(MemNode::Memory);
1782 const Type *t1 = phase->type(mem);
1783 if (t1 == Type::TOP) return Type::TOP;
1784 Node* adr = in(MemNode::Address);
1785 const TypePtr* tp = phase->type(adr)->isa_ptr();
1786 if (tp == NULL || tp->empty()) return Type::TOP;
1787 int off = tp->offset();
1788 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1789 Compile* C = phase->C;
1790
1791 // Try to guess loaded type from pointer type
1792 if (tp->isa_aryptr()) {
1793 const TypeAryPtr* ary = tp->is_aryptr();
1794 const Type* t = ary->elem();
1795
1796 // Determine whether the reference is beyond the header or not, by comparing
1797 // the offset against the offset of the start of the array's data.
1798 // Different array types begin at slightly different offsets (12 vs. 16).
1799 // We choose T_BYTE as an example base type that is least restrictive
1800 // as to alignment, which will therefore produce the smallest
1801 // possible base offset.
1802 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1803 const bool off_beyond_header = (off >= min_base_off);
1804
1805 // Try to constant-fold a stable array element.
1806 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1807 // Make sure the reference is not into the header and the offset is constant
1808 ciObject* aobj = ary->const_oop();
1809 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1810 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1811 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1812 stable_dimension,
1813 memory_type(), is_unsigned());
1814 if (con_type != NULL) {
1815 return con_type;
1816 }
1817 }
1818 }
1819
1820 // Don't do this for integer types. There is only potential profit if
1821 // the element type t is lower than _type; that is, for int types, if _type is
1822 // more restrictive than t. This only happens here if one is short and the other
1823 // char (both 16 bits), and in those cases we've made an intentional decision
1824 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1825 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1826 //
1827 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1828 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1829 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1830 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1831 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1832 // In fact, that could have been the original type of p1, and p1 could have
1833 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1834 // expression (LShiftL quux 3) independently optimized to the constant 8.
1835 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1836 && (_type->isa_vect() == NULL)
1837 && t->isa_valuetype() == NULL
1838 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1839 // t might actually be lower than _type, if _type is a unique
1840 // concrete subclass of abstract class t.
1841 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
1842 const Type* jt = t->join_speculative(_type);
1843 // In any case, do not allow the join, per se, to empty out the type.
1844 if (jt->empty() && !t->empty()) {
1845 // This can happen if a interface-typed array narrows to a class type.
1846 jt = _type;
1847 }
1848 #ifdef ASSERT
1849 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1850 // The pointers in the autobox arrays are always non-null
1851 Node* base = adr->in(AddPNode::Base);
1852 if ((base != NULL) && base->is_DecodeN()) {
1853 // Get LoadN node which loads IntegerCache.cache field
1854 base = base->in(1);
1855 }
1856 if ((base != NULL) && base->is_Con()) {
1857 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1858 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1859 // It could be narrow oop
1860 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1861 }
1862 }
1863 }
1864 #endif
1865 return jt;
1866 }
1867 }
1868 } else if (tp->base() == Type::InstPtr) {
1869 assert( off != Type::OffsetBot ||
1870 // arrays can be cast to Objects
1871 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1872 tp->is_oopptr()->klass() == ciEnv::current()->Class_klass() ||
1873 // unsafe field access may not have a constant offset
1874 C->has_unsafe_access(),
1875 "Field accesses must be precise" );
1876 // For oop loads, we expect the _type to be precise.
1877
1878 const TypeInstPtr* tinst = tp->is_instptr();
1879 BasicType bt = memory_type();
1880
1881 // Fold component and value mirror loads
1882 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1883 if (ik == phase->C->env()->Class_klass() && (off == java_lang_Class::component_mirror_offset_in_bytes() ||
1884 off == java_lang_Class::inline_mirror_offset_in_bytes())) {
1885 ciType* mirror_type = tinst->java_mirror_type();
1886 if (mirror_type != NULL) {
1887 const Type* const_oop = TypePtr::NULL_PTR;
1888 if (mirror_type->is_array_klass()) {
1889 const_oop = TypeInstPtr::make(mirror_type->as_array_klass()->component_mirror_instance());
1890 } else if (mirror_type->is_valuetype()) {
1891 const_oop = TypeInstPtr::make(mirror_type->as_value_klass()->inline_mirror_instance());
1892 }
1893 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop;
1894 }
1895 }
1896
1897 // Optimize loads from constant fields.
1898 ciObject* const_oop = tinst->const_oop();
1899 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1900 ciType* mirror_type = const_oop->as_instance()->java_mirror_type();
1901 if (mirror_type != NULL && mirror_type->is_valuetype()) {
1902 ciValueKlass* vk = mirror_type->as_value_klass();
1903 if (off == vk->default_value_offset()) {
1904 // Loading a special hidden field that contains the oop of the default value type
1905 const Type* const_oop = TypeInstPtr::make(vk->default_value_instance());
1906 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop;
1907 }
1908 }
1909 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
1910 if (con_type != NULL) {
1911 return con_type;
1912 }
1913 }
1914 } else if (tp->base() == Type::KlassPtr) {
1915 assert( off != Type::OffsetBot ||
1916 // arrays can be cast to Objects
1917 tp->is_klassptr()->klass() == NULL ||
1918 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1919 // also allow array-loading from the primary supertype
1920 // array during subtype checks
1921 Opcode() == Op_LoadKlass,
1922 "Field accesses must be precise" );
1923 // For klass/static loads, we expect the _type to be precise
1924 } else if (tp->base() == Type::RawPtr && !StressReflectiveCode) {
1925 if (adr->is_Load() && off == 0) {
1926 /* With mirrors being an indirect in the Klass*
1927 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1928 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1929 *
1930 * So check the type and klass of the node before the LoadP.
1931 */
1932 Node* adr2 = adr->in(MemNode::Address);
1933 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1934 if (tkls != NULL) {
1935 ciKlass* klass = tkls->klass();
1936 if (klass != NULL && klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1937 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1938 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1939 return TypeInstPtr::make(klass->java_mirror());
1940 }
1941 }
1942 } else {
1943 // Check for a load of the default value offset from the ValueKlassFixedBlock:
1944 // LoadI(LoadP(value_klass, adr_valueklass_fixed_block_offset), default_value_offset_offset)
1945 intptr_t offset = 0;
1946 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1947 if (base != NULL && base->is_Load() && offset == in_bytes(ValueKlass::default_value_offset_offset())) {
1948 const TypeKlassPtr* tkls = phase->type(base->in(MemNode::Address))->isa_klassptr();
1949 if (tkls != NULL && tkls->is_loaded() && tkls->klass_is_exact() && tkls->isa_valuetype() &&
1950 tkls->offset() == in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset())) {
1951 assert(base->Opcode() == Op_LoadP, "must load an oop from klass");
1952 assert(Opcode() == Op_LoadI, "must load an int from fixed block");
1953 return TypeInt::make(tkls->klass()->as_value_klass()->default_value_offset());
1954 }
1955 }
1956 }
1957 }
1958
1959 const TypeKlassPtr *tkls = tp->isa_klassptr();
1960 if (tkls != NULL && !StressReflectiveCode) {
1961 ciKlass* klass = tkls->klass();
1962 if (tkls->is_loaded() && tkls->klass_is_exact()) {
1963 // We are loading a field from a Klass metaobject whose identity
1964 // is known at compile time (the type is "exact" or "precise").
1965 // Check for fields we know are maintained as constants by the VM.
1966 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1967 // The field is Klass::_super_check_offset. Return its (constant) value.
1968 // (Folds up type checking code.)
1969 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1970 return TypeInt::make(klass->super_check_offset());
1971 }
1972 // Compute index into primary_supers array
1973 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1974 // Check for overflowing; use unsigned compare to handle the negative case.
1975 if( depth < ciKlass::primary_super_limit() ) {
1976 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1977 // (Folds up type checking code.)
1978 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1979 ciKlass *ss = klass->super_of_depth(depth);
1980 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1981 }
1982 const Type* aift = load_array_final_field(tkls, klass);
1983 if (aift != NULL) return aift;
1984 }
1985
1986 // We can still check if we are loading from the primary_supers array at a
1987 // shallow enough depth. Even though the klass is not exact, entries less
1988 // than or equal to its super depth are correct.
1989 if (tkls->is_loaded()) {
1990 ciType *inner = klass;
1991 while( inner->is_obj_array_klass() )
1992 inner = inner->as_obj_array_klass()->base_element_type();
1993 if( inner->is_instance_klass() &&
1994 !inner->as_instance_klass()->flags().is_interface() ) {
1995 // Compute index into primary_supers array
1996 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1997 // Check for overflowing; use unsigned compare to handle the negative case.
1998 if( depth < ciKlass::primary_super_limit() &&
1999 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
2000 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2001 // (Folds up type checking code.)
2002 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2003 ciKlass *ss = klass->super_of_depth(depth);
2004 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2005 }
2006 }
2007 }
2008
2009 // If the type is enough to determine that the thing is not an array,
2010 // we can give the layout_helper a positive interval type.
2011 // This will help short-circuit some reflective code.
2012 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
2013 && !klass->is_array_klass() // not directly typed as an array
2014 && !klass->is_interface() // specifically not Serializable & Cloneable
2015 && !klass->is_java_lang_Object() // not the supertype of all T[]
2016 ) {
2017 // Note: When interfaces are reliable, we can narrow the interface
2018 // test to (klass != Serializable && klass != Cloneable).
2019 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2020 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2021 // The key property of this type is that it folds up tests
2022 // for array-ness, since it proves that the layout_helper is positive.
2023 // Thus, a generic value like the basic object layout helper works fine.
2024 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2025 }
2026 }
2027
2028 // If we are loading from a freshly-allocated object, produce a zero,
2029 // if the load is provably beyond the header of the object.
2030 // (Also allow a variable load from a fresh array to produce zero.)
2031 const TypeOopPtr *tinst = tp->isa_oopptr();
2032 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
2033 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
2034 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2035 Node* value = can_see_stored_value(mem,phase);
2036 if (value != NULL && value->is_Con()) {
2037 assert(value->bottom_type()->higher_equal(_type),"sanity");
2038 return value->bottom_type();
2039 }
2040 }
2041
2042 if (is_instance) {
2043 // If we have an instance type and our memory input is the
2044 // programs's initial memory state, there is no matching store,
2045 // so just return a zero of the appropriate type
2046 Node *mem = in(MemNode::Memory);
2047 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2048 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2049 return Type::get_zero_type(_type->basic_type());
2050 }
2051 }
2052
2053 Node* alloc = is_new_object_mark_load(phase);
2054 if (alloc != NULL && !(alloc->Opcode() == Op_Allocate && UseBiasedLocking)) {
2055 return TypeX::make(markWord::prototype().value());
2056 }
2057
2058 return _type;
2059 }
2060
2061 //------------------------------match_edge-------------------------------------
2062 // Do we Match on this edge index or not? Match only the address.
2063 uint LoadNode::match_edge(uint idx) const {
2064 return idx == MemNode::Address;
2065 }
2066
2067 //--------------------------LoadBNode::Ideal--------------------------------------
2068 //
2069 // If the previous store is to the same address as this load,
2070 // and the value stored was larger than a byte, replace this load
2071 // with the value stored truncated to a byte. If no truncation is
2072 // needed, the replacement is done in LoadNode::Identity().
2073 //
2074 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2075 Node* mem = in(MemNode::Memory);
2076 Node* value = can_see_stored_value(mem,phase);
2077 if( value && !phase->type(value)->higher_equal( _type ) ) {
2078 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
2079 return new RShiftINode(result, phase->intcon(24));
2080 }
2081 // Identity call will handle the case where truncation is not needed.
2082 return LoadNode::Ideal(phase, can_reshape);
2083 }
2084
2085 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2086 Node* mem = in(MemNode::Memory);
2087 Node* value = can_see_stored_value(mem,phase);
2088 if (value != NULL && value->is_Con() &&
2089 !value->bottom_type()->higher_equal(_type)) {
2090 // If the input to the store does not fit with the load's result type,
2091 // it must be truncated. We can't delay until Ideal call since
2092 // a singleton Value is needed for split_thru_phi optimization.
2093 int con = value->get_int();
2094 return TypeInt::make((con << 24) >> 24);
2095 }
2096 return LoadNode::Value(phase);
2097 }
2098
2099 //--------------------------LoadUBNode::Ideal-------------------------------------
2100 //
2101 // If the previous store is to the same address as this load,
2102 // and the value stored was larger than a byte, replace this load
2103 // with the value stored truncated to a byte. If no truncation is
2104 // needed, the replacement is done in LoadNode::Identity().
2105 //
2106 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2107 Node* mem = in(MemNode::Memory);
2108 Node* value = can_see_stored_value(mem, phase);
2109 if (value && !phase->type(value)->higher_equal(_type))
2110 return new AndINode(value, phase->intcon(0xFF));
2111 // Identity call will handle the case where truncation is not needed.
2112 return LoadNode::Ideal(phase, can_reshape);
2113 }
2114
2115 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2116 Node* mem = in(MemNode::Memory);
2117 Node* value = can_see_stored_value(mem,phase);
2118 if (value != NULL && value->is_Con() &&
2119 !value->bottom_type()->higher_equal(_type)) {
2120 // If the input to the store does not fit with the load's result type,
2121 // it must be truncated. We can't delay until Ideal call since
2122 // a singleton Value is needed for split_thru_phi optimization.
2123 int con = value->get_int();
2124 return TypeInt::make(con & 0xFF);
2125 }
2126 return LoadNode::Value(phase);
2127 }
2128
2129 //--------------------------LoadUSNode::Ideal-------------------------------------
2130 //
2131 // If the previous store is to the same address as this load,
2132 // and the value stored was larger than a char, replace this load
2133 // with the value stored truncated to a char. If no truncation is
2134 // needed, the replacement is done in LoadNode::Identity().
2135 //
2136 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2137 Node* mem = in(MemNode::Memory);
2138 Node* value = can_see_stored_value(mem,phase);
2139 if( value && !phase->type(value)->higher_equal( _type ) )
2140 return new AndINode(value,phase->intcon(0xFFFF));
2141 // Identity call will handle the case where truncation is not needed.
2142 return LoadNode::Ideal(phase, can_reshape);
2143 }
2144
2145 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2146 Node* mem = in(MemNode::Memory);
2147 Node* value = can_see_stored_value(mem,phase);
2148 if (value != NULL && value->is_Con() &&
2149 !value->bottom_type()->higher_equal(_type)) {
2150 // If the input to the store does not fit with the load's result type,
2151 // it must be truncated. We can't delay until Ideal call since
2152 // a singleton Value is needed for split_thru_phi optimization.
2153 int con = value->get_int();
2154 return TypeInt::make(con & 0xFFFF);
2155 }
2156 return LoadNode::Value(phase);
2157 }
2158
2159 //--------------------------LoadSNode::Ideal--------------------------------------
2160 //
2161 // If the previous store is to the same address as this load,
2162 // and the value stored was larger than a short, replace this load
2163 // with the value stored truncated to a short. If no truncation is
2164 // needed, the replacement is done in LoadNode::Identity().
2165 //
2166 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2167 Node* mem = in(MemNode::Memory);
2168 Node* value = can_see_stored_value(mem,phase);
2169 if( value && !phase->type(value)->higher_equal( _type ) ) {
2170 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2171 return new RShiftINode(result, phase->intcon(16));
2172 }
2173 // Identity call will handle the case where truncation is not needed.
2174 return LoadNode::Ideal(phase, can_reshape);
2175 }
2176
2177 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2178 Node* mem = in(MemNode::Memory);
2179 Node* value = can_see_stored_value(mem,phase);
2180 if (value != NULL && value->is_Con() &&
2181 !value->bottom_type()->higher_equal(_type)) {
2182 // If the input to the store does not fit with the load's result type,
2183 // it must be truncated. We can't delay until Ideal call since
2184 // a singleton Value is needed for split_thru_phi optimization.
2185 int con = value->get_int();
2186 return TypeInt::make((con << 16) >> 16);
2187 }
2188 return LoadNode::Value(phase);
2189 }
2190
2191 //=============================================================================
2192 //----------------------------LoadKlassNode::make------------------------------
2193 // Polymorphic factory method:
2194 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at,
2195 const TypeKlassPtr* tk, bool clear_prop_bits) {
2196 // sanity check the alias category against the created node type
2197 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2198 assert(adr_type != NULL, "expecting TypeKlassPtr");
2199 #ifdef _LP64
2200 if (adr_type->is_ptr_to_narrowklass()) {
2201 assert(UseCompressedClassPointers, "no compressed klasses");
2202 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered, clear_prop_bits));
2203 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2204 }
2205 #endif
2206 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2207 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered, clear_prop_bits);
2208 }
2209
2210 //------------------------------Value------------------------------------------
2211 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2212 return klass_value_common(phase);
2213 }
2214
2215 // In most cases, LoadKlassNode does not have the control input set. If the control
2216 // input is set, it must not be removed (by LoadNode::Ideal()).
2217 bool LoadKlassNode::can_remove_control() const {
2218 return false;
2219 }
2220
2221 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2222 // Either input is TOP ==> the result is TOP
2223 const Type *t1 = phase->type( in(MemNode::Memory) );
2224 if (t1 == Type::TOP) return Type::TOP;
2225 Node *adr = in(MemNode::Address);
2226 const Type *t2 = phase->type( adr );
2227 if (t2 == Type::TOP) return Type::TOP;
2228 const TypePtr *tp = t2->is_ptr();
2229 if (TypePtr::above_centerline(tp->ptr()) ||
2230 tp->ptr() == TypePtr::Null) return Type::TOP;
2231
2232 // Return a more precise klass, if possible
2233 const TypeInstPtr *tinst = tp->isa_instptr();
2234 if (tinst != NULL) {
2235 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2236 int offset = tinst->offset();
2237 if (ik == phase->C->env()->Class_klass()
2238 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2239 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2240 // We are loading a special hidden field from a Class mirror object,
2241 // the field which points to the VM's Klass metaobject.
2242 bool is_indirect_type = true;
2243 ciType* t = tinst->java_mirror_type(&is_indirect_type);
2244 // java_mirror_type returns non-null for compile-time Class constants.
2245 if (t != NULL) {
2246 // constant oop => constant klass
2247 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2248 if (t->is_void()) {
2249 // We cannot create a void array. Since void is a primitive type return null
2250 // klass. Users of this result need to do a null check on the returned klass.
2251 return TypePtr::NULL_PTR;
2252 }
2253 return TypeKlassPtr::make(ciArrayKlass::make(t, /* never_null */ !is_indirect_type));
2254 }
2255 if (!t->is_klass()) {
2256 // a primitive Class (e.g., int.class) has NULL for a klass field
2257 return TypePtr::NULL_PTR;
2258 }
2259 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2260 return TypeKlassPtr::make(t->as_klass());
2261 }
2262 // non-constant mirror, so we can't tell what's going on
2263 }
2264 if( !ik->is_loaded() )
2265 return _type; // Bail out if not loaded
2266 if (offset == oopDesc::klass_offset_in_bytes()) {
2267 if (tinst->klass_is_exact()) {
2268 return TypeKlassPtr::make(ik);
2269 }
2270 // See if we can become precise: no subklasses and no interface
2271 // (Note: We need to support verified interfaces.)
2272 if (!ik->is_interface() && !ik->has_subklass()) {
2273 //assert(!UseExactTypes, "this code should be useless with exact types");
2274 // Add a dependence; if any subclass added we need to recompile
2275 if (!ik->is_final()) {
2276 // %%% should use stronger assert_unique_concrete_subtype instead
2277 phase->C->dependencies()->assert_leaf_type(ik);
2278 }
2279 // Return precise klass
2280 return TypeKlassPtr::make(ik);
2281 }
2282
2283 // Return root of possible klass
2284 return TypeKlassPtr::make(TypePtr::NotNull, ik, Type::Offset(0), tinst->flat_array());
2285 }
2286 }
2287
2288 // Check for loading klass from an array
2289 const TypeAryPtr *tary = tp->isa_aryptr();
2290 if (tary != NULL) {
2291 ciKlass *tary_klass = tary->klass();
2292 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2293 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2294 ciArrayKlass* ak = tary_klass->as_array_klass();
2295 // Do not fold klass loads from [V?. The runtime type might be [V due to [V <: [V?
2296 // and the klass for [V is not equal to the klass for [V?.
2297 if (tary->klass_is_exact()) {
2298 return TypeKlassPtr::make(tary_klass);
2299 }
2300
2301 // If the klass is an object array, we defer the question to the
2302 // array component klass.
2303 if (ak->is_obj_array_klass()) {
2304 assert(ak->is_loaded(), "");
2305 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2306 if (base_k->is_loaded() && base_k->is_instance_klass()) {
2307 ciInstanceKlass *ik = base_k->as_instance_klass();
2308 // See if we can become precise: no subklasses and no interface
2309 if (!ik->is_interface() && !ik->has_subklass() && (!ik->is_valuetype() || ak->storage_properties().is_null_free())) {
2310 //assert(!UseExactTypes, "this code should be useless with exact types");
2311 // Add a dependence; if any subclass added we need to recompile
2312 if (!ik->is_final()) {
2313 phase->C->dependencies()->assert_leaf_type(ik);
2314 }
2315 // Return precise array klass
2316 return TypeKlassPtr::make(ak);
2317 }
2318 }
2319 return TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0), false);
2320 } else if (ak->is_type_array_klass()) {
2321 //assert(!UseExactTypes, "this code should be useless with exact types");
2322 return TypeKlassPtr::make(ak); // These are always precise
2323 }
2324 }
2325 }
2326
2327 // Check for loading klass from an array klass
2328 const TypeKlassPtr *tkls = tp->isa_klassptr();
2329 if (tkls != NULL && !StressReflectiveCode) {
2330 if (!tkls->is_loaded()) {
2331 return _type; // Bail out if not loaded
2332 }
2333 ciKlass* klass = tkls->klass();
2334 if( klass->is_obj_array_klass() &&
2335 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2336 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2337 // // Always returning precise element type is incorrect,
2338 // // e.g., element type could be object and array may contain strings
2339 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2340
2341 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2342 // according to the element type's subclassing.
2343 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0), elem->flatten_array());
2344 } else if (klass->is_value_array_klass() &&
2345 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2346 ciKlass* elem = klass->as_value_array_klass()->element_klass();
2347 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0), true);
2348 }
2349 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2350 tkls->offset() == in_bytes(Klass::super_offset())) {
2351 ciKlass* sup = klass->as_instance_klass()->super();
2352 // The field is Klass::_super. Return its (constant) value.
2353 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2354 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2355 }
2356 }
2357
2358 // Bailout case
2359 return LoadNode::Value(phase);
2360 }
2361
2362 //------------------------------Identity---------------------------------------
2363 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2364 // Also feed through the klass in Allocate(...klass...)._klass.
2365 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2366 return klass_identity_common(phase);
2367 }
2368
2369 const Type* GetStoragePropertyNode::Value(PhaseGVN* phase) const {
2370 if (in(1) != NULL) {
2371 const Type* in1_t = phase->type(in(1));
2372 if (in1_t == Type::TOP) {
2373 return Type::TOP;
2374 }
2375 const TypeKlassPtr* tk = in1_t->make_ptr()->is_klassptr();
2376 ciArrayKlass* ak = tk->klass()->as_array_klass();
2377 ciKlass* elem = ak->element_klass();
2378 if (tk->klass_is_exact() || (!elem->is_java_lang_Object() && !elem->is_interface() && !elem->is_valuetype())) {
2379 int props_shift = in1_t->isa_narrowklass() ? oopDesc::narrow_storage_props_shift : oopDesc::wide_storage_props_shift;
2380 ArrayStorageProperties props = ak->storage_properties();
2381 intptr_t storage_properties = 0;
2382 if ((Opcode() == Op_GetFlattenedProperty && props.is_flattened()) ||
2383 (Opcode() == Op_GetNullFreeProperty && props.is_null_free())) {
2384 storage_properties = 1;
2385 }
2386 if (in1_t->isa_narrowklass()) {
2387 return TypeInt::make((int)storage_properties);
2388 }
2389 return TypeX::make(storage_properties);
2390 }
2391 }
2392 return bottom_type();
2393 }
2394
2395 Node* GetStoragePropertyNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2396 if (!can_reshape) {
2397 return NULL;
2398 }
2399 if (in(1) != NULL && in(1)->is_Phi()) {
2400 Node* phi = in(1);
2401 Node* r = phi->in(0);
2402 Node* new_phi = new PhiNode(r, bottom_type());
2403 for (uint i = 1; i < r->req(); i++) {
2404 Node* in = phi->in(i);
2405 if (in == NULL) continue;
2406 Node* n = clone();
2407 n->set_req(1, in);
2408 new_phi->init_req(i, phase->transform(n));
2409 }
2410 return new_phi;
2411 }
2412 return NULL;
2413 }
2414
2415 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2416 Node* x = LoadNode::Identity(phase);
2417 if (x != this) return x;
2418
2419 // Take apart the address into an oop and and offset.
2420 // Return 'this' if we cannot.
2421 Node* adr = in(MemNode::Address);
2422 intptr_t offset = 0;
2423 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2424 if (base == NULL) return this;
2425 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2426 if (toop == NULL) return this;
2427
2428 // Step over potential GC barrier for OopHandle resolve
2429 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2430 if (bs->is_gc_barrier_node(base)) {
2431 base = bs->step_over_gc_barrier(base);
2432 }
2433
2434 // We can fetch the klass directly through an AllocateNode.
2435 // This works even if the klass is not constant (clone or newArray).
2436 if (offset == oopDesc::klass_offset_in_bytes()) {
2437 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2438 if (allocated_klass != NULL) {
2439 return allocated_klass;
2440 }
2441 }
2442
2443 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2444 // See inline_native_Class_query for occurrences of these patterns.
2445 // Java Example: x.getClass().isAssignableFrom(y)
2446 //
2447 // This improves reflective code, often making the Class
2448 // mirror go completely dead. (Current exception: Class
2449 // mirrors may appear in debug info, but we could clean them out by
2450 // introducing a new debug info operator for Klass.java_mirror).
2451
2452 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2453 && offset == java_lang_Class::klass_offset_in_bytes()) {
2454 if (base->is_Load()) {
2455 Node* base2 = base->in(MemNode::Address);
2456 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2457 Node* adr2 = base2->in(MemNode::Address);
2458 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2459 if (tkls != NULL && !tkls->empty()
2460 && (tkls->klass()->is_instance_klass() ||
2461 tkls->klass()->is_array_klass())
2462 && adr2->is_AddP()
2463 ) {
2464 int mirror_field = in_bytes(Klass::java_mirror_offset());
2465 if (tkls->offset() == mirror_field) {
2466 return adr2->in(AddPNode::Base);
2467 }
2468 }
2469 }
2470 }
2471 }
2472
2473 return this;
2474 }
2475
2476
2477 //------------------------------Value------------------------------------------
2478 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2479 const Type *t = klass_value_common(phase);
2480 if (t == Type::TOP)
2481 return t;
2482
2483 return t->make_narrowklass();
2484 }
2485
2486 //------------------------------Identity---------------------------------------
2487 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2488 // Also feed through the klass in Allocate(...klass...)._klass.
2489 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2490 Node *x = klass_identity_common(phase);
2491
2492 const Type *t = phase->type( x );
2493 if( t == Type::TOP ) return x;
2494 if( t->isa_narrowklass()) return x;
2495 assert (!t->isa_narrowoop(), "no narrow oop here");
2496
2497 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2498 }
2499
2500 //------------------------------Value-----------------------------------------
2501 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2502 // Either input is TOP ==> the result is TOP
2503 const Type *t1 = phase->type( in(MemNode::Memory) );
2504 if( t1 == Type::TOP ) return Type::TOP;
2505 Node *adr = in(MemNode::Address);
2506 const Type *t2 = phase->type( adr );
2507 if( t2 == Type::TOP ) return Type::TOP;
2508 const TypePtr *tp = t2->is_ptr();
2509 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2510 const TypeAryPtr *tap = tp->isa_aryptr();
2511 if( !tap ) return _type;
2512 return tap->size();
2513 }
2514
2515 //-------------------------------Ideal---------------------------------------
2516 // Feed through the length in AllocateArray(...length...)._length.
2517 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2518 Node* p = MemNode::Ideal_common(phase, can_reshape);
2519 if (p) return (p == NodeSentinel) ? NULL : p;
2520
2521 // Take apart the address into an oop and and offset.
2522 // Return 'this' if we cannot.
2523 Node* adr = in(MemNode::Address);
2524 intptr_t offset = 0;
2525 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2526 if (base == NULL) return NULL;
2527 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2528 if (tary == NULL) return NULL;
2529
2530 // We can fetch the length directly through an AllocateArrayNode.
2531 // This works even if the length is not constant (clone or newArray).
2532 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2533 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2534 if (alloc != NULL) {
2535 Node* allocated_length = alloc->Ideal_length();
2536 Node* len = alloc->make_ideal_length(tary, phase);
2537 if (allocated_length != len) {
2538 // New CastII improves on this.
2539 return len;
2540 }
2541 }
2542 }
2543
2544 return NULL;
2545 }
2546
2547 //------------------------------Identity---------------------------------------
2548 // Feed through the length in AllocateArray(...length...)._length.
2549 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2550 Node* x = LoadINode::Identity(phase);
2551 if (x != this) return x;
2552
2553 // Take apart the address into an oop and and offset.
2554 // Return 'this' if we cannot.
2555 Node* adr = in(MemNode::Address);
2556 intptr_t offset = 0;
2557 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2558 if (base == NULL) return this;
2559 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2560 if (tary == NULL) return this;
2561
2562 // We can fetch the length directly through an AllocateArrayNode.
2563 // This works even if the length is not constant (clone or newArray).
2564 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2565 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2566 if (alloc != NULL) {
2567 Node* allocated_length = alloc->Ideal_length();
2568 // Do not allow make_ideal_length to allocate a CastII node.
2569 Node* len = alloc->make_ideal_length(tary, phase, false);
2570 if (allocated_length == len) {
2571 // Return allocated_length only if it would not be improved by a CastII.
2572 return allocated_length;
2573 }
2574 }
2575 }
2576
2577 return this;
2578
2579 }
2580
2581 //=============================================================================
2582 //---------------------------StoreNode::make-----------------------------------
2583 // Polymorphic factory method:
2584 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2585 assert((mo == unordered || mo == release), "unexpected");
2586 Compile* C = gvn.C;
2587 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2588 ctl != NULL, "raw memory operations should have control edge");
2589
2590 switch (bt) {
2591 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2592 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2593 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2594 case T_CHAR:
2595 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2596 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2597 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2598 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2599 case T_METADATA:
2600 case T_ADDRESS:
2601 case T_VALUETYPE:
2602 case T_OBJECT:
2603 #ifdef _LP64
2604 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2605 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2606 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2607 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2608 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2609 adr->bottom_type()->isa_rawptr())) {
2610 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2611 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2612 }
2613 #endif
2614 {
2615 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2616 }
2617 default:
2618 ShouldNotReachHere();
2619 return (StoreNode*)NULL;
2620 }
2621 }
2622
2623 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2624 bool require_atomic = true;
2625 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2626 }
2627
2628 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2629 bool require_atomic = true;
2630 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2631 }
2632
2633
2634 //--------------------------bottom_type----------------------------------------
2635 const Type *StoreNode::bottom_type() const {
2636 return Type::MEMORY;
2637 }
2638
2639 //------------------------------hash-------------------------------------------
2640 uint StoreNode::hash() const {
2641 // unroll addition of interesting fields
2642 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2643
2644 // Since they are not commoned, do not hash them:
2645 return NO_HASH;
2646 }
2647
2648 //------------------------------Ideal------------------------------------------
2649 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2650 // When a store immediately follows a relevant allocation/initialization,
2651 // try to capture it into the initialization, or hoist it above.
2652 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2653 Node* p = MemNode::Ideal_common(phase, can_reshape);
2654 if (p) return (p == NodeSentinel) ? NULL : p;
2655
2656 Node* mem = in(MemNode::Memory);
2657 Node* address = in(MemNode::Address);
2658 // Back-to-back stores to same address? Fold em up. Generally
2659 // unsafe if I have intervening uses... Also disallowed for StoreCM
2660 // since they must follow each StoreP operation. Redundant StoreCMs
2661 // are eliminated just before matching in final_graph_reshape.
2662 if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::VALUES) {
2663 Node* st = mem;
2664 // If Store 'st' has more than one use, we cannot fold 'st' away.
2665 // For example, 'st' might be the final state at a conditional
2666 // return. Or, 'st' might be used by some node which is live at
2667 // the same time 'st' is live, which might be unschedulable. So,
2668 // require exactly ONE user until such time as we clone 'mem' for
2669 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2670 // true).
2671 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2672 // Looking at a dead closed cycle of memory?
2673 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2674 assert(Opcode() == st->Opcode() ||
2675 st->Opcode() == Op_StoreVector ||
2676 Opcode() == Op_StoreVector ||
2677 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2678 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2679 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2680 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
2681 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2682 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2683
2684 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2685 st->as_Store()->memory_size() <= this->memory_size()) {
2686 Node* use = st->raw_out(0);
2687 phase->igvn_rehash_node_delayed(use);
2688 if (can_reshape) {
2689 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2690 } else {
2691 // It's OK to do this in the parser, since DU info is always accurate,
2692 // and the parser always refers to nodes via SafePointNode maps.
2693 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2694 }
2695 return this;
2696 }
2697 st = st->in(MemNode::Memory);
2698 }
2699 }
2700
2701
2702 // Capture an unaliased, unconditional, simple store into an initializer.
2703 // Or, if it is independent of the allocation, hoist it above the allocation.
2704 if (ReduceFieldZeroing && /*can_reshape &&*/
2705 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2706 InitializeNode* init = mem->in(0)->as_Initialize();
2707 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2708 if (offset > 0) {
2709 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2710 // If the InitializeNode captured me, it made a raw copy of me,
2711 // and I need to disappear.
2712 if (moved != NULL) {
2713 // %%% hack to ensure that Ideal returns a new node:
2714 mem = MergeMemNode::make(mem);
2715 return mem; // fold me away
2716 }
2717 }
2718 }
2719
2720 return NULL; // No further progress
2721 }
2722
2723 //------------------------------Value-----------------------------------------
2724 const Type* StoreNode::Value(PhaseGVN* phase) const {
2725 // Either input is TOP ==> the result is TOP
2726 const Type *t1 = phase->type( in(MemNode::Memory) );
2727 if( t1 == Type::TOP ) return Type::TOP;
2728 const Type *t2 = phase->type( in(MemNode::Address) );
2729 if( t2 == Type::TOP ) return Type::TOP;
2730 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2731 if( t3 == Type::TOP ) return Type::TOP;
2732 return Type::MEMORY;
2733 }
2734
2735 //------------------------------Identity---------------------------------------
2736 // Remove redundant stores:
2737 // Store(m, p, Load(m, p)) changes to m.
2738 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2739 Node* StoreNode::Identity(PhaseGVN* phase) {
2740 Node* mem = in(MemNode::Memory);
2741 Node* adr = in(MemNode::Address);
2742 Node* val = in(MemNode::ValueIn);
2743
2744 Node* result = this;
2745
2746 // Load then Store? Then the Store is useless
2747 if (val->is_Load() &&
2748 val->in(MemNode::Address)->eqv_uncast(adr) &&
2749 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2750 val->as_Load()->store_Opcode() == Opcode()) {
2751 result = mem;
2752 }
2753
2754 // Two stores in a row of the same value?
2755 if (result == this &&
2756 mem->is_Store() &&
2757 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2758 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2759 mem->Opcode() == Opcode()) {
2760 result = mem;
2761 }
2762
2763 // Store of zero anywhere into a freshly-allocated object?
2764 // Then the store is useless.
2765 // (It must already have been captured by the InitializeNode.)
2766 if (result == this && ReduceFieldZeroing) {
2767 // a newly allocated object is already all-zeroes everywhere
2768 if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
2769 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) {
2770 assert(!phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == NULL, "storing null to value array is forbidden");
2771 result = mem;
2772 }
2773
2774 if (result == this) {
2775 // the store may also apply to zero-bits in an earlier object
2776 Node* prev_mem = find_previous_store(phase);
2777 // Steps (a), (b): Walk past independent stores to find an exact match.
2778 if (prev_mem != NULL) {
2779 Node* prev_val = can_see_stored_value(prev_mem, phase);
2780 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2781 // prev_val and val might differ by a cast; it would be good
2782 // to keep the more informative of the two.
2783 if (phase->type(val)->is_zero_type()) {
2784 result = mem;
2785 } else if (prev_mem->is_Proj() && prev_mem->in(0)->is_Initialize()) {
2786 InitializeNode* init = prev_mem->in(0)->as_Initialize();
2787 AllocateNode* alloc = init->allocation();
2788 if (alloc != NULL && alloc->in(AllocateNode::DefaultValue) == val) {
2789 result = mem;
2790 }
2791 }
2792 }
2793 }
2794 }
2795 }
2796
2797 if (result != this && phase->is_IterGVN() != NULL) {
2798 MemBarNode* trailing = trailing_membar();
2799 if (trailing != NULL) {
2800 #ifdef ASSERT
2801 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2802 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2803 #endif
2804 PhaseIterGVN* igvn = phase->is_IterGVN();
2805 trailing->remove(igvn);
2806 }
2807 }
2808
2809 return result;
2810 }
2811
2812 //------------------------------match_edge-------------------------------------
2813 // Do we Match on this edge index or not? Match only memory & value
2814 uint StoreNode::match_edge(uint idx) const {
2815 return idx == MemNode::Address || idx == MemNode::ValueIn;
2816 }
2817
2818 //------------------------------cmp--------------------------------------------
2819 // Do not common stores up together. They generally have to be split
2820 // back up anyways, so do not bother.
2821 bool StoreNode::cmp( const Node &n ) const {
2822 return (&n == this); // Always fail except on self
2823 }
2824
2825 //------------------------------Ideal_masked_input-----------------------------
2826 // Check for a useless mask before a partial-word store
2827 // (StoreB ... (AndI valIn conIa) )
2828 // If (conIa & mask == mask) this simplifies to
2829 // (StoreB ... (valIn) )
2830 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2831 Node *val = in(MemNode::ValueIn);
2832 if( val->Opcode() == Op_AndI ) {
2833 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2834 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2835 set_req(MemNode::ValueIn, val->in(1));
2836 return this;
2837 }
2838 }
2839 return NULL;
2840 }
2841
2842
2843 //------------------------------Ideal_sign_extended_input----------------------
2844 // Check for useless sign-extension before a partial-word store
2845 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2846 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2847 // (StoreB ... (valIn) )
2848 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2849 Node *val = in(MemNode::ValueIn);
2850 if( val->Opcode() == Op_RShiftI ) {
2851 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2852 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2853 Node *shl = val->in(1);
2854 if( shl->Opcode() == Op_LShiftI ) {
2855 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2856 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2857 set_req(MemNode::ValueIn, shl->in(1));
2858 return this;
2859 }
2860 }
2861 }
2862 }
2863 return NULL;
2864 }
2865
2866 //------------------------------value_never_loaded-----------------------------------
2867 // Determine whether there are any possible loads of the value stored.
2868 // For simplicity, we actually check if there are any loads from the
2869 // address stored to, not just for loads of the value stored by this node.
2870 //
2871 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2872 Node *adr = in(Address);
2873 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2874 if (adr_oop == NULL)
2875 return false;
2876 if (!adr_oop->is_known_instance_field())
2877 return false; // if not a distinct instance, there may be aliases of the address
2878 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2879 Node *use = adr->fast_out(i);
2880 if (use->is_Load() || use->is_LoadStore()) {
2881 return false;
2882 }
2883 }
2884 return true;
2885 }
2886
2887 MemBarNode* StoreNode::trailing_membar() const {
2888 if (is_release()) {
2889 MemBarNode* trailing_mb = NULL;
2890 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2891 Node* u = fast_out(i);
2892 if (u->is_MemBar()) {
2893 if (u->as_MemBar()->trailing_store()) {
2894 assert(u->Opcode() == Op_MemBarVolatile, "");
2895 assert(trailing_mb == NULL, "only one");
2896 trailing_mb = u->as_MemBar();
2897 #ifdef ASSERT
2898 Node* leading = u->as_MemBar()->leading_membar();
2899 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2900 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2901 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2902 #endif
2903 } else {
2904 assert(u->as_MemBar()->standalone(), "");
2905 }
2906 }
2907 }
2908 return trailing_mb;
2909 }
2910 return NULL;
2911 }
2912
2913
2914 //=============================================================================
2915 //------------------------------Ideal------------------------------------------
2916 // If the store is from an AND mask that leaves the low bits untouched, then
2917 // we can skip the AND operation. If the store is from a sign-extension
2918 // (a left shift, then right shift) we can skip both.
2919 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2920 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2921 if( progress != NULL ) return progress;
2922
2923 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2924 if( progress != NULL ) return progress;
2925
2926 // Finally check the default case
2927 return StoreNode::Ideal(phase, can_reshape);
2928 }
2929
2930 //=============================================================================
2931 //------------------------------Ideal------------------------------------------
2932 // If the store is from an AND mask that leaves the low bits untouched, then
2933 // we can skip the AND operation
2934 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2935 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2936 if( progress != NULL ) return progress;
2937
2938 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2939 if( progress != NULL ) return progress;
2940
2941 // Finally check the default case
2942 return StoreNode::Ideal(phase, can_reshape);
2943 }
2944
2945 //=============================================================================
2946 //------------------------------Identity---------------------------------------
2947 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2948 // No need to card mark when storing a null ptr
2949 Node* my_store = in(MemNode::OopStore);
2950 if (my_store->is_Store()) {
2951 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2952 if( t1 == TypePtr::NULL_PTR ) {
2953 return in(MemNode::Memory);
2954 }
2955 }
2956 return this;
2957 }
2958
2959 //=============================================================================
2960 //------------------------------Ideal---------------------------------------
2961 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2962 Node* progress = StoreNode::Ideal(phase, can_reshape);
2963 if (progress != NULL) return progress;
2964
2965 Node* my_store = in(MemNode::OopStore);
2966 if (my_store->is_MergeMem()) {
2967 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2968 set_req(MemNode::OopStore, mem);
2969 return this;
2970 }
2971
2972 return NULL;
2973 }
2974
2975 //------------------------------Value-----------------------------------------
2976 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2977 // Either input is TOP ==> the result is TOP
2978 const Type *t = phase->type( in(MemNode::Memory) );
2979 if( t == Type::TOP ) return Type::TOP;
2980 t = phase->type( in(MemNode::Address) );
2981 if( t == Type::TOP ) return Type::TOP;
2982 t = phase->type( in(MemNode::ValueIn) );
2983 if( t == Type::TOP ) return Type::TOP;
2984 // If extra input is TOP ==> the result is TOP
2985 t = phase->type( in(MemNode::OopStore) );
2986 if( t == Type::TOP ) return Type::TOP;
2987
2988 return StoreNode::Value( phase );
2989 }
2990
2991
2992 //=============================================================================
2993 //----------------------------------SCMemProjNode------------------------------
2994 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2995 {
2996 return bottom_type();
2997 }
2998
2999 //=============================================================================
3000 //----------------------------------LoadStoreNode------------------------------
3001 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3002 : Node(required),
3003 _type(rt),
3004 _adr_type(at),
3005 _has_barrier(false)
3006 {
3007 init_req(MemNode::Control, c );
3008 init_req(MemNode::Memory , mem);
3009 init_req(MemNode::Address, adr);
3010 init_req(MemNode::ValueIn, val);
3011 init_class_id(Class_LoadStore);
3012 }
3013
3014 uint LoadStoreNode::ideal_reg() const {
3015 return _type->ideal_reg();
3016 }
3017
3018 bool LoadStoreNode::result_not_used() const {
3019 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
3020 Node *x = fast_out(i);
3021 if (x->Opcode() == Op_SCMemProj) continue;
3022 return false;
3023 }
3024 return true;
3025 }
3026
3027 MemBarNode* LoadStoreNode::trailing_membar() const {
3028 MemBarNode* trailing = NULL;
3029 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3030 Node* u = fast_out(i);
3031 if (u->is_MemBar()) {
3032 if (u->as_MemBar()->trailing_load_store()) {
3033 assert(u->Opcode() == Op_MemBarAcquire, "");
3034 assert(trailing == NULL, "only one");
3035 trailing = u->as_MemBar();
3036 #ifdef ASSERT
3037 Node* leading = trailing->leading_membar();
3038 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3039 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3040 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3041 #endif
3042 } else {
3043 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3044 }
3045 }
3046 }
3047
3048 return trailing;
3049 }
3050
3051 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3052
3053 //=============================================================================
3054 //----------------------------------LoadStoreConditionalNode--------------------
3055 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
3056 init_req(ExpectedIn, ex );
3057 }
3058
3059 //=============================================================================
3060 //-------------------------------adr_type--------------------------------------
3061 const TypePtr* ClearArrayNode::adr_type() const {
3062 Node *adr = in(3);
3063 if (adr == NULL) return NULL; // node is dead
3064 return MemNode::calculate_adr_type(adr->bottom_type());
3065 }
3066
3067 //------------------------------match_edge-------------------------------------
3068 // Do we Match on this edge index or not? Do not match memory
3069 uint ClearArrayNode::match_edge(uint idx) const {
3070 return idx > 1;
3071 }
3072
3073 //------------------------------Identity---------------------------------------
3074 // Clearing a zero length array does nothing
3075 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3076 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
3077 }
3078
3079 //------------------------------Idealize---------------------------------------
3080 // Clearing a short array is faster with stores
3081 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3082 // Already know this is a large node, do not try to ideal it
3083 if (!IdealizeClearArrayNode || _is_large) return NULL;
3084
3085 const int unit = BytesPerLong;
3086 const TypeX* t = phase->type(in(2))->isa_intptr_t();
3087 if (!t) return NULL;
3088 if (!t->is_con()) return NULL;
3089 intptr_t raw_count = t->get_con();
3090 intptr_t size = raw_count;
3091 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3092 // Clearing nothing uses the Identity call.
3093 // Negative clears are possible on dead ClearArrays
3094 // (see jck test stmt114.stmt11402.val).
3095 if (size <= 0 || size % unit != 0) return NULL;
3096 intptr_t count = size / unit;
3097 // Length too long; communicate this to matchers and assemblers.
3098 // Assemblers are responsible to produce fast hardware clears for it.
3099 if (size > InitArrayShortSize) {
3100 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
3101 }
3102 Node *mem = in(1);
3103 if( phase->type(mem)==Type::TOP ) return NULL;
3104 Node *adr = in(3);
3105 const Type* at = phase->type(adr);
3106 if( at==Type::TOP ) return NULL;
3107 const TypePtr* atp = at->isa_ptr();
3108 // adjust atp to be the correct array element address type
3109 if (atp == NULL) atp = TypePtr::BOTTOM;
3110 else atp = atp->add_offset(Type::OffsetBot);
3111 // Get base for derived pointer purposes
3112 if( adr->Opcode() != Op_AddP ) Unimplemented();
3113 Node *base = adr->in(1);
3114
3115 Node *val = in(4);
3116 Node *off = phase->MakeConX(BytesPerLong);
3117 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
3118 count--;
3119 while( count-- ) {
3120 mem = phase->transform(mem);
3121 adr = phase->transform(new AddPNode(base,adr,off));
3122 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
3123 }
3124 return mem;
3125 }
3126
3127 //----------------------------step_through----------------------------------
3128 // Return allocation input memory edge if it is different instance
3129 // or itself if it is the one we are looking for.
3130 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
3131 Node* n = *np;
3132 assert(n->is_ClearArray(), "sanity");
3133 intptr_t offset;
3134 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3135 // This method is called only before Allocate nodes are expanded
3136 // during macro nodes expansion. Before that ClearArray nodes are
3137 // only generated in PhaseMacroExpand::generate_arraycopy() (before
3138 // Allocate nodes are expanded) which follows allocations.
3139 assert(alloc != NULL, "should have allocation");
3140 if (alloc->_idx == instance_id) {
3141 // Can not bypass initialization of the instance we are looking for.
3142 return false;
3143 }
3144 // Otherwise skip it.
3145 InitializeNode* init = alloc->initialization();
3146 if (init != NULL)
3147 *np = init->in(TypeFunc::Memory);
3148 else
3149 *np = alloc->in(TypeFunc::Memory);
3150 return true;
3151 }
3152
3153 //----------------------------clear_memory-------------------------------------
3154 // Generate code to initialize object storage to zero.
3155 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3156 Node* val,
3157 Node* raw_val,
3158 intptr_t start_offset,
3159 Node* end_offset,
3160 PhaseGVN* phase) {
3161 intptr_t offset = start_offset;
3162
3163 int unit = BytesPerLong;
3164 if ((offset % unit) != 0) {
3165 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3166 adr = phase->transform(adr);
3167 const TypePtr* atp = TypeRawPtr::BOTTOM;
3168 if (val != NULL) {
3169 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3170 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3171 } else {
3172 assert(raw_val == NULL, "val may not be null");
3173 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3174 }
3175 mem = phase->transform(mem);
3176 offset += BytesPerInt;
3177 }
3178 assert((offset % unit) == 0, "");
3179
3180 // Initialize the remaining stuff, if any, with a ClearArray.
3181 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
3182 }
3183
3184 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3185 Node* raw_val,
3186 Node* start_offset,
3187 Node* end_offset,
3188 PhaseGVN* phase) {
3189 if (start_offset == end_offset) {
3190 // nothing to do
3191 return mem;
3192 }
3193
3194 int unit = BytesPerLong;
3195 Node* zbase = start_offset;
3196 Node* zend = end_offset;
3197
3198 // Scale to the unit required by the CPU:
3199 if (!Matcher::init_array_count_is_in_bytes) {
3200 Node* shift = phase->intcon(exact_log2(unit));
3201 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3202 zend = phase->transform(new URShiftXNode(zend, shift) );
3203 }
3204
3205 // Bulk clear double-words
3206 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3207 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3208 if (raw_val == NULL) {
3209 raw_val = phase->MakeConX(0);
3210 }
3211 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
3212 return phase->transform(mem);
3213 }
3214
3215 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3216 Node* val,
3217 Node* raw_val,
3218 intptr_t start_offset,
3219 intptr_t end_offset,
3220 PhaseGVN* phase) {
3221 if (start_offset == end_offset) {
3222 // nothing to do
3223 return mem;
3224 }
3225
3226 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3227 intptr_t done_offset = end_offset;
3228 if ((done_offset % BytesPerLong) != 0) {
3229 done_offset -= BytesPerInt;
3230 }
3231 if (done_offset > start_offset) {
3232 mem = clear_memory(ctl, mem, dest, val, raw_val,
3233 start_offset, phase->MakeConX(done_offset), phase);
3234 }
3235 if (done_offset < end_offset) { // emit the final 32-bit store
3236 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3237 adr = phase->transform(adr);
3238 const TypePtr* atp = TypeRawPtr::BOTTOM;
3239 if (val != NULL) {
3240 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3241 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3242 } else {
3243 assert(raw_val == NULL, "val may not be null");
3244 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3245 }
3246 mem = phase->transform(mem);
3247 done_offset += BytesPerInt;
3248 }
3249 assert(done_offset == end_offset, "");
3250 return mem;
3251 }
3252
3253 //=============================================================================
3254 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3255 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3256 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3257 #ifdef ASSERT
3258 , _pair_idx(0)
3259 #endif
3260 {
3261 init_class_id(Class_MemBar);
3262 Node* top = C->top();
3263 init_req(TypeFunc::I_O,top);
3264 init_req(TypeFunc::FramePtr,top);
3265 init_req(TypeFunc::ReturnAdr,top);
3266 if (precedent != NULL)
3267 init_req(TypeFunc::Parms, precedent);
3268 }
3269
3270 //------------------------------cmp--------------------------------------------
3271 uint MemBarNode::hash() const { return NO_HASH; }
3272 bool MemBarNode::cmp( const Node &n ) const {
3273 return (&n == this); // Always fail except on self
3274 }
3275
3276 //------------------------------make-------------------------------------------
3277 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3278 switch (opcode) {
3279 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3280 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3281 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3282 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3283 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3284 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3285 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3286 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3287 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3288 case Op_Initialize: return new InitializeNode(C, atp, pn);
3289 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3290 default: ShouldNotReachHere(); return NULL;
3291 }
3292 }
3293
3294 void MemBarNode::remove(PhaseIterGVN *igvn) {
3295 if (outcnt() != 2) {
3296 return;
3297 }
3298 if (trailing_store() || trailing_load_store()) {
3299 MemBarNode* leading = leading_membar();
3300 if (leading != NULL) {
3301 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3302 leading->remove(igvn);
3303 }
3304 }
3305 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3306 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3307 }
3308
3309 //------------------------------Ideal------------------------------------------
3310 // Return a node which is more "ideal" than the current node. Strip out
3311 // control copies
3312 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3313 if (remove_dead_region(phase, can_reshape)) return this;
3314 // Don't bother trying to transform a dead node
3315 if (in(0) && in(0)->is_top()) {
3316 return NULL;
3317 }
3318
3319 bool progress = false;
3320 // Eliminate volatile MemBars for scalar replaced objects.
3321 if (can_reshape && req() == (Precedent+1)) {
3322 bool eliminate = false;
3323 int opc = Opcode();
3324 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3325 // Volatile field loads and stores.
3326 Node* my_mem = in(MemBarNode::Precedent);
3327 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3328 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3329 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3330 // replace this Precedent (decodeN) with the Load instead.
3331 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3332 Node* load_node = my_mem->in(1);
3333 set_req(MemBarNode::Precedent, load_node);
3334 phase->is_IterGVN()->_worklist.push(my_mem);
3335 my_mem = load_node;
3336 } else {
3337 assert(my_mem->unique_out() == this, "sanity");
3338 del_req(Precedent);
3339 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3340 my_mem = NULL;
3341 }
3342 progress = true;
3343 }
3344 if (my_mem != NULL && my_mem->is_Mem()) {
3345 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3346 // Check for scalar replaced object reference.
3347 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3348 t_oop->offset() != Type::OffsetBot &&
3349 t_oop->offset() != Type::OffsetTop) {
3350 eliminate = true;
3351 }
3352 }
3353 } else if (opc == Op_MemBarRelease) {
3354 // Final field stores.
3355 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3356 if ((alloc != NULL) && alloc->is_Allocate() &&
3357 alloc->as_Allocate()->does_not_escape_thread()) {
3358 // The allocated object does not escape.
3359 eliminate = true;
3360 }
3361 }
3362 if (eliminate) {
3363 // Replace MemBar projections by its inputs.
3364 PhaseIterGVN* igvn = phase->is_IterGVN();
3365 remove(igvn);
3366 // Must return either the original node (now dead) or a new node
3367 // (Do not return a top here, since that would break the uniqueness of top.)
3368 return new ConINode(TypeInt::ZERO);
3369 }
3370 }
3371 return progress ? this : NULL;
3372 }
3373
3374 //------------------------------Value------------------------------------------
3375 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3376 if( !in(0) ) return Type::TOP;
3377 if( phase->type(in(0)) == Type::TOP )
3378 return Type::TOP;
3379 return TypeTuple::MEMBAR;
3380 }
3381
3382 //------------------------------match------------------------------------------
3383 // Construct projections for memory.
3384 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
3385 switch (proj->_con) {
3386 case TypeFunc::Control:
3387 case TypeFunc::Memory:
3388 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3389 }
3390 ShouldNotReachHere();
3391 return NULL;
3392 }
3393
3394 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3395 trailing->_kind = TrailingStore;
3396 leading->_kind = LeadingStore;
3397 #ifdef ASSERT
3398 trailing->_pair_idx = leading->_idx;
3399 leading->_pair_idx = leading->_idx;
3400 #endif
3401 }
3402
3403 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3404 trailing->_kind = TrailingLoadStore;
3405 leading->_kind = LeadingLoadStore;
3406 #ifdef ASSERT
3407 trailing->_pair_idx = leading->_idx;
3408 leading->_pair_idx = leading->_idx;
3409 #endif
3410 }
3411
3412 MemBarNode* MemBarNode::trailing_membar() const {
3413 ResourceMark rm;
3414 Node* trailing = (Node*)this;
3415 VectorSet seen(Thread::current()->resource_area());
3416 Node_Stack multis(0);
3417 do {
3418 Node* c = trailing;
3419 uint i = 0;
3420 do {
3421 trailing = NULL;
3422 for (; i < c->outcnt(); i++) {
3423 Node* next = c->raw_out(i);
3424 if (next != c && next->is_CFG()) {
3425 if (c->is_MultiBranch()) {
3426 if (multis.node() == c) {
3427 multis.set_index(i+1);
3428 } else {
3429 multis.push(c, i+1);
3430 }
3431 }
3432 trailing = next;
3433 break;
3434 }
3435 }
3436 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3437 break;
3438 }
3439 while (multis.size() > 0) {
3440 c = multis.node();
3441 i = multis.index();
3442 if (i < c->req()) {
3443 break;
3444 }
3445 multis.pop();
3446 }
3447 } while (multis.size() > 0);
3448 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3449
3450 MemBarNode* mb = trailing->as_MemBar();
3451 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3452 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3453 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3454 return mb;
3455 }
3456
3457 MemBarNode* MemBarNode::leading_membar() const {
3458 ResourceMark rm;
3459 VectorSet seen(Thread::current()->resource_area());
3460 Node_Stack regions(0);
3461 Node* leading = in(0);
3462 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3463 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3464 leading = NULL;
3465 while (regions.size() > 0 && leading == NULL) {
3466 Node* r = regions.node();
3467 uint i = regions.index();
3468 if (i < r->req()) {
3469 leading = r->in(i);
3470 regions.set_index(i+1);
3471 } else {
3472 regions.pop();
3473 }
3474 }
3475 if (leading == NULL) {
3476 assert(regions.size() == 0, "all paths should have been tried");
3477 return NULL;
3478 }
3479 }
3480 if (leading->is_Region()) {
3481 regions.push(leading, 2);
3482 leading = leading->in(1);
3483 } else {
3484 leading = leading->in(0);
3485 }
3486 }
3487 #ifdef ASSERT
3488 Unique_Node_List wq;
3489 wq.push((Node*)this);
3490 uint found = 0;
3491 for (uint i = 0; i < wq.size(); i++) {
3492 Node* n = wq.at(i);
3493 if (n->is_Region()) {
3494 for (uint j = 1; j < n->req(); j++) {
3495 Node* in = n->in(j);
3496 if (in != NULL && !in->is_top()) {
3497 wq.push(in);
3498 }
3499 }
3500 } else {
3501 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3502 assert(n == leading, "consistency check failed");
3503 found++;
3504 } else {
3505 Node* in = n->in(0);
3506 if (in != NULL && !in->is_top()) {
3507 wq.push(in);
3508 }
3509 }
3510 }
3511 }
3512 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3513 #endif
3514 if (leading == NULL) {
3515 return NULL;
3516 }
3517 MemBarNode* mb = leading->as_MemBar();
3518 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3519 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3520 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3521 return mb;
3522 }
3523
3524 //===========================InitializeNode====================================
3525 // SUMMARY:
3526 // This node acts as a memory barrier on raw memory, after some raw stores.
3527 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3528 // The Initialize can 'capture' suitably constrained stores as raw inits.
3529 // It can coalesce related raw stores into larger units (called 'tiles').
3530 // It can avoid zeroing new storage for memory units which have raw inits.
3531 // At macro-expansion, it is marked 'complete', and does not optimize further.
3532 //
3533 // EXAMPLE:
3534 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3535 // ctl = incoming control; mem* = incoming memory
3536 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3537 // First allocate uninitialized memory and fill in the header:
3538 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3539 // ctl := alloc.Control; mem* := alloc.Memory*
3540 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3541 // Then initialize to zero the non-header parts of the raw memory block:
3542 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3543 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3544 // After the initialize node executes, the object is ready for service:
3545 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3546 // Suppose its body is immediately initialized as {1,2}:
3547 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3548 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3549 // mem.SLICE(#short[*]) := store2
3550 //
3551 // DETAILS:
3552 // An InitializeNode collects and isolates object initialization after
3553 // an AllocateNode and before the next possible safepoint. As a
3554 // memory barrier (MemBarNode), it keeps critical stores from drifting
3555 // down past any safepoint or any publication of the allocation.
3556 // Before this barrier, a newly-allocated object may have uninitialized bits.
3557 // After this barrier, it may be treated as a real oop, and GC is allowed.
3558 //
3559 // The semantics of the InitializeNode include an implicit zeroing of
3560 // the new object from object header to the end of the object.
3561 // (The object header and end are determined by the AllocateNode.)
3562 //
3563 // Certain stores may be added as direct inputs to the InitializeNode.
3564 // These stores must update raw memory, and they must be to addresses
3565 // derived from the raw address produced by AllocateNode, and with
3566 // a constant offset. They must be ordered by increasing offset.
3567 // The first one is at in(RawStores), the last at in(req()-1).
3568 // Unlike most memory operations, they are not linked in a chain,
3569 // but are displayed in parallel as users of the rawmem output of
3570 // the allocation.
3571 //
3572 // (See comments in InitializeNode::capture_store, which continue
3573 // the example given above.)
3574 //
3575 // When the associated Allocate is macro-expanded, the InitializeNode
3576 // may be rewritten to optimize collected stores. A ClearArrayNode
3577 // may also be created at that point to represent any required zeroing.
3578 // The InitializeNode is then marked 'complete', prohibiting further
3579 // capturing of nearby memory operations.
3580 //
3581 // During macro-expansion, all captured initializations which store
3582 // constant values of 32 bits or smaller are coalesced (if advantageous)
3583 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3584 // initialized in fewer memory operations. Memory words which are
3585 // covered by neither tiles nor non-constant stores are pre-zeroed
3586 // by explicit stores of zero. (The code shape happens to do all
3587 // zeroing first, then all other stores, with both sequences occurring
3588 // in order of ascending offsets.)
3589 //
3590 // Alternatively, code may be inserted between an AllocateNode and its
3591 // InitializeNode, to perform arbitrary initialization of the new object.
3592 // E.g., the object copying intrinsics insert complex data transfers here.
3593 // The initialization must then be marked as 'complete' disable the
3594 // built-in zeroing semantics and the collection of initializing stores.
3595 //
3596 // While an InitializeNode is incomplete, reads from the memory state
3597 // produced by it are optimizable if they match the control edge and
3598 // new oop address associated with the allocation/initialization.
3599 // They return a stored value (if the offset matches) or else zero.
3600 // A write to the memory state, if it matches control and address,
3601 // and if it is to a constant offset, may be 'captured' by the
3602 // InitializeNode. It is cloned as a raw memory operation and rewired
3603 // inside the initialization, to the raw oop produced by the allocation.
3604 // Operations on addresses which are provably distinct (e.g., to
3605 // other AllocateNodes) are allowed to bypass the initialization.
3606 //
3607 // The effect of all this is to consolidate object initialization
3608 // (both arrays and non-arrays, both piecewise and bulk) into a
3609 // single location, where it can be optimized as a unit.
3610 //
3611 // Only stores with an offset less than TrackedInitializationLimit words
3612 // will be considered for capture by an InitializeNode. This puts a
3613 // reasonable limit on the complexity of optimized initializations.
3614
3615 //---------------------------InitializeNode------------------------------------
3616 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3617 : MemBarNode(C, adr_type, rawoop),
3618 _is_complete(Incomplete), _does_not_escape(false)
3619 {
3620 init_class_id(Class_Initialize);
3621
3622 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3623 assert(in(RawAddress) == rawoop, "proper init");
3624 // Note: allocation() can be NULL, for secondary initialization barriers
3625 }
3626
3627 // Since this node is not matched, it will be processed by the
3628 // register allocator. Declare that there are no constraints
3629 // on the allocation of the RawAddress edge.
3630 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3631 // This edge should be set to top, by the set_complete. But be conservative.
3632 if (idx == InitializeNode::RawAddress)
3633 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3634 return RegMask::Empty;
3635 }
3636
3637 Node* InitializeNode::memory(uint alias_idx) {
3638 Node* mem = in(Memory);
3639 if (mem->is_MergeMem()) {
3640 return mem->as_MergeMem()->memory_at(alias_idx);
3641 } else {
3642 // incoming raw memory is not split
3643 return mem;
3644 }
3645 }
3646
3647 bool InitializeNode::is_non_zero() {
3648 if (is_complete()) return false;
3649 remove_extra_zeroes();
3650 return (req() > RawStores);
3651 }
3652
3653 void InitializeNode::set_complete(PhaseGVN* phase) {
3654 assert(!is_complete(), "caller responsibility");
3655 _is_complete = Complete;
3656
3657 // After this node is complete, it contains a bunch of
3658 // raw-memory initializations. There is no need for
3659 // it to have anything to do with non-raw memory effects.
3660 // Therefore, tell all non-raw users to re-optimize themselves,
3661 // after skipping the memory effects of this initialization.
3662 PhaseIterGVN* igvn = phase->is_IterGVN();
3663 if (igvn) igvn->add_users_to_worklist(this);
3664 }
3665
3666 // convenience function
3667 // return false if the init contains any stores already
3668 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3669 InitializeNode* init = initialization();
3670 if (init == NULL || init->is_complete()) {
3671 return false;
3672 }
3673 init->remove_extra_zeroes();
3674 // for now, if this allocation has already collected any inits, bail:
3675 if (init->is_non_zero()) return false;
3676 init->set_complete(phase);
3677 return true;
3678 }
3679
3680 void InitializeNode::remove_extra_zeroes() {
3681 if (req() == RawStores) return;
3682 Node* zmem = zero_memory();
3683 uint fill = RawStores;
3684 for (uint i = fill; i < req(); i++) {
3685 Node* n = in(i);
3686 if (n->is_top() || n == zmem) continue; // skip
3687 if (fill < i) set_req(fill, n); // compact
3688 ++fill;
3689 }
3690 // delete any empty spaces created:
3691 while (fill < req()) {
3692 del_req(fill);
3693 }
3694 }
3695
3696 // Helper for remembering which stores go with which offsets.
3697 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3698 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3699 intptr_t offset = -1;
3700 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3701 phase, offset);
3702 if (base == NULL) return -1; // something is dead,
3703 if (offset < 0) return -1; // dead, dead
3704 return offset;
3705 }
3706
3707 // Helper for proving that an initialization expression is
3708 // "simple enough" to be folded into an object initialization.
3709 // Attempts to prove that a store's initial value 'n' can be captured
3710 // within the initialization without creating a vicious cycle, such as:
3711 // { Foo p = new Foo(); p.next = p; }
3712 // True for constants and parameters and small combinations thereof.
3713 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3714 if (n == NULL) return true; // (can this really happen?)
3715 if (n->is_Proj()) n = n->in(0);
3716 if (n == this) return false; // found a cycle
3717 if (n->is_Con()) return true;
3718 if (n->is_Start()) return true; // params, etc., are OK
3719 if (n->is_Root()) return true; // even better
3720
3721 Node* ctl = n->in(0);
3722 if (ctl != NULL && !ctl->is_top()) {
3723 if (ctl->is_Proj()) ctl = ctl->in(0);
3724 if (ctl == this) return false;
3725
3726 // If we already know that the enclosing memory op is pinned right after
3727 // the init, then any control flow that the store has picked up
3728 // must have preceded the init, or else be equal to the init.
3729 // Even after loop optimizations (which might change control edges)
3730 // a store is never pinned *before* the availability of its inputs.
3731 if (!MemNode::all_controls_dominate(n, this))
3732 return false; // failed to prove a good control
3733 }
3734
3735 // Check data edges for possible dependencies on 'this'.
3736 if ((count += 1) > 20) return false; // complexity limit
3737 for (uint i = 1; i < n->req(); i++) {
3738 Node* m = n->in(i);
3739 if (m == NULL || m == n || m->is_top()) continue;
3740 uint first_i = n->find_edge(m);
3741 if (i != first_i) continue; // process duplicate edge just once
3742 if (!detect_init_independence(m, count)) {
3743 return false;
3744 }
3745 }
3746
3747 return true;
3748 }
3749
3750 // Here are all the checks a Store must pass before it can be moved into
3751 // an initialization. Returns zero if a check fails.
3752 // On success, returns the (constant) offset to which the store applies,
3753 // within the initialized memory.
3754 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3755 const int FAIL = 0;
3756 if (st->req() != MemNode::ValueIn + 1)
3757 return FAIL; // an inscrutable StoreNode (card mark?)
3758 Node* ctl = st->in(MemNode::Control);
3759 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3760 return FAIL; // must be unconditional after the initialization
3761 Node* mem = st->in(MemNode::Memory);
3762 if (!(mem->is_Proj() && mem->in(0) == this))
3763 return FAIL; // must not be preceded by other stores
3764 Node* adr = st->in(MemNode::Address);
3765 intptr_t offset;
3766 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3767 if (alloc == NULL)
3768 return FAIL; // inscrutable address
3769 if (alloc != allocation())
3770 return FAIL; // wrong allocation! (store needs to float up)
3771 int size_in_bytes = st->memory_size();
3772 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3773 return FAIL; // mismatched access
3774 }
3775 Node* val = st->in(MemNode::ValueIn);
3776 int complexity_count = 0;
3777 if (!detect_init_independence(val, complexity_count))
3778 return FAIL; // stored value must be 'simple enough'
3779
3780 // The Store can be captured only if nothing after the allocation
3781 // and before the Store is using the memory location that the store
3782 // overwrites.
3783 bool failed = false;
3784 // If is_complete_with_arraycopy() is true the shape of the graph is
3785 // well defined and is safe so no need for extra checks.
3786 if (!is_complete_with_arraycopy()) {
3787 // We are going to look at each use of the memory state following
3788 // the allocation to make sure nothing reads the memory that the
3789 // Store writes.
3790 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3791 int alias_idx = phase->C->get_alias_index(t_adr);
3792 ResourceMark rm;
3793 Unique_Node_List mems;
3794 mems.push(mem);
3795 Node* unique_merge = NULL;
3796 for (uint next = 0; next < mems.size(); ++next) {
3797 Node *m = mems.at(next);
3798 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3799 Node *n = m->fast_out(j);
3800 if (n->outcnt() == 0) {
3801 continue;
3802 }
3803 if (n == st) {
3804 continue;
3805 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3806 // If the control of this use is different from the control
3807 // of the Store which is right after the InitializeNode then
3808 // this node cannot be between the InitializeNode and the
3809 // Store.
3810 continue;
3811 } else if (n->is_MergeMem()) {
3812 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3813 // We can hit a MergeMemNode (that will likely go away
3814 // later) that is a direct use of the memory state
3815 // following the InitializeNode on the same slice as the
3816 // store node that we'd like to capture. We need to check
3817 // the uses of the MergeMemNode.
3818 mems.push(n);
3819 }
3820 } else if (n->is_Mem()) {
3821 Node* other_adr = n->in(MemNode::Address);
3822 if (other_adr == adr) {
3823 failed = true;
3824 break;
3825 } else {
3826 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3827 if (other_t_adr != NULL) {
3828 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3829 if (other_alias_idx == alias_idx) {
3830 // A load from the same memory slice as the store right
3831 // after the InitializeNode. We check the control of the
3832 // object/array that is loaded from. If it's the same as
3833 // the store control then we cannot capture the store.
3834 assert(!n->is_Store(), "2 stores to same slice on same control?");
3835 Node* base = other_adr;
3836 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3837 base = base->in(AddPNode::Base);
3838 if (base != NULL) {
3839 base = base->uncast();
3840 if (base->is_Proj() && base->in(0) == alloc) {
3841 failed = true;
3842 break;
3843 }
3844 }
3845 }
3846 }
3847 }
3848 } else {
3849 failed = true;
3850 break;
3851 }
3852 }
3853 }
3854 }
3855 if (failed) {
3856 if (!can_reshape) {
3857 // We decided we couldn't capture the store during parsing. We
3858 // should try again during the next IGVN once the graph is
3859 // cleaner.
3860 phase->C->record_for_igvn(st);
3861 }
3862 return FAIL;
3863 }
3864
3865 return offset; // success
3866 }
3867
3868 // Find the captured store in(i) which corresponds to the range
3869 // [start..start+size) in the initialized object.
3870 // If there is one, return its index i. If there isn't, return the
3871 // negative of the index where it should be inserted.
3872 // Return 0 if the queried range overlaps an initialization boundary
3873 // or if dead code is encountered.
3874 // If size_in_bytes is zero, do not bother with overlap checks.
3875 int InitializeNode::captured_store_insertion_point(intptr_t start,
3876 int size_in_bytes,
3877 PhaseTransform* phase) {
3878 const int FAIL = 0, MAX_STORE = BytesPerLong;
3879
3880 if (is_complete())
3881 return FAIL; // arraycopy got here first; punt
3882
3883 assert(allocation() != NULL, "must be present");
3884
3885 // no negatives, no header fields:
3886 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3887
3888 // after a certain size, we bail out on tracking all the stores:
3889 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3890 if (start >= ti_limit) return FAIL;
3891
3892 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3893 if (i >= limit) return -(int)i; // not found; here is where to put it
3894
3895 Node* st = in(i);
3896 intptr_t st_off = get_store_offset(st, phase);
3897 if (st_off < 0) {
3898 if (st != zero_memory()) {
3899 return FAIL; // bail out if there is dead garbage
3900 }
3901 } else if (st_off > start) {
3902 // ...we are done, since stores are ordered
3903 if (st_off < start + size_in_bytes) {
3904 return FAIL; // the next store overlaps
3905 }
3906 return -(int)i; // not found; here is where to put it
3907 } else if (st_off < start) {
3908 if (size_in_bytes != 0 &&
3909 start < st_off + MAX_STORE &&
3910 start < st_off + st->as_Store()->memory_size()) {
3911 return FAIL; // the previous store overlaps
3912 }
3913 } else {
3914 if (size_in_bytes != 0 &&
3915 st->as_Store()->memory_size() != size_in_bytes) {
3916 return FAIL; // mismatched store size
3917 }
3918 return i;
3919 }
3920
3921 ++i;
3922 }
3923 }
3924
3925 // Look for a captured store which initializes at the offset 'start'
3926 // with the given size. If there is no such store, and no other
3927 // initialization interferes, then return zero_memory (the memory
3928 // projection of the AllocateNode).
3929 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3930 PhaseTransform* phase) {
3931 assert(stores_are_sane(phase), "");
3932 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3933 if (i == 0) {
3934 return NULL; // something is dead
3935 } else if (i < 0) {
3936 return zero_memory(); // just primordial zero bits here
3937 } else {
3938 Node* st = in(i); // here is the store at this position
3939 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3940 return st;
3941 }
3942 }
3943
3944 // Create, as a raw pointer, an address within my new object at 'offset'.
3945 Node* InitializeNode::make_raw_address(intptr_t offset,
3946 PhaseTransform* phase) {
3947 Node* addr = in(RawAddress);
3948 if (offset != 0) {
3949 Compile* C = phase->C;
3950 addr = phase->transform( new AddPNode(C->top(), addr,
3951 phase->MakeConX(offset)) );
3952 }
3953 return addr;
3954 }
3955
3956 // Clone the given store, converting it into a raw store
3957 // initializing a field or element of my new object.
3958 // Caller is responsible for retiring the original store,
3959 // with subsume_node or the like.
3960 //
3961 // From the example above InitializeNode::InitializeNode,
3962 // here are the old stores to be captured:
3963 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3964 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3965 //
3966 // Here is the changed code; note the extra edges on init:
3967 // alloc = (Allocate ...)
3968 // rawoop = alloc.RawAddress
3969 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3970 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3971 // init = (Initialize alloc.Control alloc.Memory rawoop
3972 // rawstore1 rawstore2)
3973 //
3974 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3975 PhaseTransform* phase, bool can_reshape) {
3976 assert(stores_are_sane(phase), "");
3977
3978 if (start < 0) return NULL;
3979 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3980
3981 Compile* C = phase->C;
3982 int size_in_bytes = st->memory_size();
3983 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3984 if (i == 0) return NULL; // bail out
3985 Node* prev_mem = NULL; // raw memory for the captured store
3986 if (i > 0) {
3987 prev_mem = in(i); // there is a pre-existing store under this one
3988 set_req(i, C->top()); // temporarily disconnect it
3989 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3990 } else {
3991 i = -i; // no pre-existing store
3992 prev_mem = zero_memory(); // a slice of the newly allocated object
3993 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3994 set_req(--i, C->top()); // reuse this edge; it has been folded away
3995 else
3996 ins_req(i, C->top()); // build a new edge
3997 }
3998 Node* new_st = st->clone();
3999 new_st->set_req(MemNode::Control, in(Control));
4000 new_st->set_req(MemNode::Memory, prev_mem);
4001 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
4002 new_st = phase->transform(new_st);
4003
4004 // At this point, new_st might have swallowed a pre-existing store
4005 // at the same offset, or perhaps new_st might have disappeared,
4006 // if it redundantly stored the same value (or zero to fresh memory).
4007
4008 // In any case, wire it in:
4009 phase->igvn_rehash_node_delayed(this);
4010 set_req(i, new_st);
4011
4012 // The caller may now kill the old guy.
4013 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
4014 assert(check_st == new_st || check_st == NULL, "must be findable");
4015 assert(!is_complete(), "");
4016 return new_st;
4017 }
4018
4019 static bool store_constant(jlong* tiles, int num_tiles,
4020 intptr_t st_off, int st_size,
4021 jlong con) {
4022 if ((st_off & (st_size-1)) != 0)
4023 return false; // strange store offset (assume size==2**N)
4024 address addr = (address)tiles + st_off;
4025 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4026 switch (st_size) {
4027 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
4028 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
4029 case sizeof(jint): *(jint*) addr = (jint) con; break;
4030 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
4031 default: return false; // strange store size (detect size!=2**N here)
4032 }
4033 return true; // return success to caller
4034 }
4035
4036 // Coalesce subword constants into int constants and possibly
4037 // into long constants. The goal, if the CPU permits,
4038 // is to initialize the object with a small number of 64-bit tiles.
4039 // Also, convert floating-point constants to bit patterns.
4040 // Non-constants are not relevant to this pass.
4041 //
4042 // In terms of the running example on InitializeNode::InitializeNode
4043 // and InitializeNode::capture_store, here is the transformation
4044 // of rawstore1 and rawstore2 into rawstore12:
4045 // alloc = (Allocate ...)
4046 // rawoop = alloc.RawAddress
4047 // tile12 = 0x00010002
4048 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4049 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4050 //
4051 void
4052 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4053 Node* size_in_bytes,
4054 PhaseGVN* phase) {
4055 Compile* C = phase->C;
4056
4057 assert(stores_are_sane(phase), "");
4058 // Note: After this pass, they are not completely sane,
4059 // since there may be some overlaps.
4060
4061 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4062
4063 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4064 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4065 size_limit = MIN2(size_limit, ti_limit);
4066 size_limit = align_up(size_limit, BytesPerLong);
4067 int num_tiles = size_limit / BytesPerLong;
4068
4069 // allocate space for the tile map:
4070 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4071 jlong tiles_buf[small_len];
4072 Node* nodes_buf[small_len];
4073 jlong inits_buf[small_len];
4074 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4075 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4076 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4077 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4078 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4079 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4080 // tiles: exact bitwise model of all primitive constants
4081 // nodes: last constant-storing node subsumed into the tiles model
4082 // inits: which bytes (in each tile) are touched by any initializations
4083
4084 //// Pass A: Fill in the tile model with any relevant stores.
4085
4086 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4087 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4088 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4089 Node* zmem = zero_memory(); // initially zero memory state
4090 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4091 Node* st = in(i);
4092 intptr_t st_off = get_store_offset(st, phase);
4093
4094 // Figure out the store's offset and constant value:
4095 if (st_off < header_size) continue; //skip (ignore header)
4096 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
4097 int st_size = st->as_Store()->memory_size();
4098 if (st_off + st_size > size_limit) break;
4099
4100 // Record which bytes are touched, whether by constant or not.
4101 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4102 continue; // skip (strange store size)
4103
4104 const Type* val = phase->type(st->in(MemNode::ValueIn));
4105 if (!val->singleton()) continue; //skip (non-con store)
4106 BasicType type = val->basic_type();
4107
4108 jlong con = 0;
4109 switch (type) {
4110 case T_INT: con = val->is_int()->get_con(); break;
4111 case T_LONG: con = val->is_long()->get_con(); break;
4112 case T_FLOAT: con = jint_cast(val->getf()); break;
4113 case T_DOUBLE: con = jlong_cast(val->getd()); break;
4114 default: continue; //skip (odd store type)
4115 }
4116
4117 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4118 st->Opcode() == Op_StoreL) {
4119 continue; // This StoreL is already optimal.
4120 }
4121
4122 // Store down the constant.
4123 store_constant(tiles, num_tiles, st_off, st_size, con);
4124
4125 intptr_t j = st_off >> LogBytesPerLong;
4126
4127 if (type == T_INT && st_size == BytesPerInt
4128 && (st_off & BytesPerInt) == BytesPerInt) {
4129 jlong lcon = tiles[j];
4130 if (!Matcher::isSimpleConstant64(lcon) &&
4131 st->Opcode() == Op_StoreI) {
4132 // This StoreI is already optimal by itself.
4133 jint* intcon = (jint*) &tiles[j];
4134 intcon[1] = 0; // undo the store_constant()
4135
4136 // If the previous store is also optimal by itself, back up and
4137 // undo the action of the previous loop iteration... if we can.
4138 // But if we can't, just let the previous half take care of itself.
4139 st = nodes[j];
4140 st_off -= BytesPerInt;
4141 con = intcon[0];
4142 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
4143 assert(st_off >= header_size, "still ignoring header");
4144 assert(get_store_offset(st, phase) == st_off, "must be");
4145 assert(in(i-1) == zmem, "must be");
4146 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4147 assert(con == tcon->is_int()->get_con(), "must be");
4148 // Undo the effects of the previous loop trip, which swallowed st:
4149 intcon[0] = 0; // undo store_constant()
4150 set_req(i-1, st); // undo set_req(i, zmem)
4151 nodes[j] = NULL; // undo nodes[j] = st
4152 --old_subword; // undo ++old_subword
4153 }
4154 continue; // This StoreI is already optimal.
4155 }
4156 }
4157
4158 // This store is not needed.
4159 set_req(i, zmem);
4160 nodes[j] = st; // record for the moment
4161 if (st_size < BytesPerLong) // something has changed
4162 ++old_subword; // includes int/float, but who's counting...
4163 else ++old_long;
4164 }
4165
4166 if ((old_subword + old_long) == 0)
4167 return; // nothing more to do
4168
4169 //// Pass B: Convert any non-zero tiles into optimal constant stores.
4170 // Be sure to insert them before overlapping non-constant stores.
4171 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
4172 for (int j = 0; j < num_tiles; j++) {
4173 jlong con = tiles[j];
4174 jlong init = inits[j];
4175 if (con == 0) continue;
4176 jint con0, con1; // split the constant, address-wise
4177 jint init0, init1; // split the init map, address-wise
4178 { union { jlong con; jint intcon[2]; } u;
4179 u.con = con;
4180 con0 = u.intcon[0];
4181 con1 = u.intcon[1];
4182 u.con = init;
4183 init0 = u.intcon[0];
4184 init1 = u.intcon[1];
4185 }
4186
4187 Node* old = nodes[j];
4188 assert(old != NULL, "need the prior store");
4189 intptr_t offset = (j * BytesPerLong);
4190
4191 bool split = !Matcher::isSimpleConstant64(con);
4192
4193 if (offset < header_size) {
4194 assert(offset + BytesPerInt >= header_size, "second int counts");
4195 assert(*(jint*)&tiles[j] == 0, "junk in header");
4196 split = true; // only the second word counts
4197 // Example: int a[] = { 42 ... }
4198 } else if (con0 == 0 && init0 == -1) {
4199 split = true; // first word is covered by full inits
4200 // Example: int a[] = { ... foo(), 42 ... }
4201 } else if (con1 == 0 && init1 == -1) {
4202 split = true; // second word is covered by full inits
4203 // Example: int a[] = { ... 42, foo() ... }
4204 }
4205
4206 // Here's a case where init0 is neither 0 nor -1:
4207 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4208 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4209 // In this case the tile is not split; it is (jlong)42.
4210 // The big tile is stored down, and then the foo() value is inserted.
4211 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4212
4213 Node* ctl = old->in(MemNode::Control);
4214 Node* adr = make_raw_address(offset, phase);
4215 const TypePtr* atp = TypeRawPtr::BOTTOM;
4216
4217 // One or two coalesced stores to plop down.
4218 Node* st[2];
4219 intptr_t off[2];
4220 int nst = 0;
4221 if (!split) {
4222 ++new_long;
4223 off[nst] = offset;
4224 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4225 phase->longcon(con), T_LONG, MemNode::unordered);
4226 } else {
4227 // Omit either if it is a zero.
4228 if (con0 != 0) {
4229 ++new_int;
4230 off[nst] = offset;
4231 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4232 phase->intcon(con0), T_INT, MemNode::unordered);
4233 }
4234 if (con1 != 0) {
4235 ++new_int;
4236 offset += BytesPerInt;
4237 adr = make_raw_address(offset, phase);
4238 off[nst] = offset;
4239 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4240 phase->intcon(con1), T_INT, MemNode::unordered);
4241 }
4242 }
4243
4244 // Insert second store first, then the first before the second.
4245 // Insert each one just before any overlapping non-constant stores.
4246 while (nst > 0) {
4247 Node* st1 = st[--nst];
4248 C->copy_node_notes_to(st1, old);
4249 st1 = phase->transform(st1);
4250 offset = off[nst];
4251 assert(offset >= header_size, "do not smash header");
4252 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4253 guarantee(ins_idx != 0, "must re-insert constant store");
4254 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4255 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4256 set_req(--ins_idx, st1);
4257 else
4258 ins_req(ins_idx, st1);
4259 }
4260 }
4261
4262 if (PrintCompilation && WizardMode)
4263 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4264 old_subword, old_long, new_int, new_long);
4265 if (C->log() != NULL)
4266 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4267 old_subword, old_long, new_int, new_long);
4268
4269 // Clean up any remaining occurrences of zmem:
4270 remove_extra_zeroes();
4271 }
4272
4273 // Explore forward from in(start) to find the first fully initialized
4274 // word, and return its offset. Skip groups of subword stores which
4275 // together initialize full words. If in(start) is itself part of a
4276 // fully initialized word, return the offset of in(start). If there
4277 // are no following full-word stores, or if something is fishy, return
4278 // a negative value.
4279 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4280 int int_map = 0;
4281 intptr_t int_map_off = 0;
4282 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4283
4284 for (uint i = start, limit = req(); i < limit; i++) {
4285 Node* st = in(i);
4286
4287 intptr_t st_off = get_store_offset(st, phase);
4288 if (st_off < 0) break; // return conservative answer
4289
4290 int st_size = st->as_Store()->memory_size();
4291 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4292 return st_off; // we found a complete word init
4293 }
4294
4295 // update the map:
4296
4297 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4298 if (this_int_off != int_map_off) {
4299 // reset the map:
4300 int_map = 0;
4301 int_map_off = this_int_off;
4302 }
4303
4304 int subword_off = st_off - this_int_off;
4305 int_map |= right_n_bits(st_size) << subword_off;
4306 if ((int_map & FULL_MAP) == FULL_MAP) {
4307 return this_int_off; // we found a complete word init
4308 }
4309
4310 // Did this store hit or cross the word boundary?
4311 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4312 if (next_int_off == this_int_off + BytesPerInt) {
4313 // We passed the current int, without fully initializing it.
4314 int_map_off = next_int_off;
4315 int_map >>= BytesPerInt;
4316 } else if (next_int_off > this_int_off + BytesPerInt) {
4317 // We passed the current and next int.
4318 return this_int_off + BytesPerInt;
4319 }
4320 }
4321
4322 return -1;
4323 }
4324
4325
4326 // Called when the associated AllocateNode is expanded into CFG.
4327 // At this point, we may perform additional optimizations.
4328 // Linearize the stores by ascending offset, to make memory
4329 // activity as coherent as possible.
4330 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4331 intptr_t header_size,
4332 Node* size_in_bytes,
4333 PhaseGVN* phase) {
4334 assert(!is_complete(), "not already complete");
4335 assert(stores_are_sane(phase), "");
4336 assert(allocation() != NULL, "must be present");
4337
4338 remove_extra_zeroes();
4339
4340 if (ReduceFieldZeroing || ReduceBulkZeroing)
4341 // reduce instruction count for common initialization patterns
4342 coalesce_subword_stores(header_size, size_in_bytes, phase);
4343
4344 Node* zmem = zero_memory(); // initially zero memory state
4345 Node* inits = zmem; // accumulating a linearized chain of inits
4346 #ifdef ASSERT
4347 intptr_t first_offset = allocation()->minimum_header_size();
4348 intptr_t last_init_off = first_offset; // previous init offset
4349 intptr_t last_init_end = first_offset; // previous init offset+size
4350 intptr_t last_tile_end = first_offset; // previous tile offset+size
4351 #endif
4352 intptr_t zeroes_done = header_size;
4353
4354 bool do_zeroing = true; // we might give up if inits are very sparse
4355 int big_init_gaps = 0; // how many large gaps have we seen?
4356
4357 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4358 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4359
4360 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4361 Node* st = in(i);
4362 intptr_t st_off = get_store_offset(st, phase);
4363 if (st_off < 0)
4364 break; // unknown junk in the inits
4365 if (st->in(MemNode::Memory) != zmem)
4366 break; // complicated store chains somehow in list
4367
4368 int st_size = st->as_Store()->memory_size();
4369 intptr_t next_init_off = st_off + st_size;
4370
4371 if (do_zeroing && zeroes_done < next_init_off) {
4372 // See if this store needs a zero before it or under it.
4373 intptr_t zeroes_needed = st_off;
4374
4375 if (st_size < BytesPerInt) {
4376 // Look for subword stores which only partially initialize words.
4377 // If we find some, we must lay down some word-level zeroes first,
4378 // underneath the subword stores.
4379 //
4380 // Examples:
4381 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4382 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4383 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4384 //
4385 // Note: coalesce_subword_stores may have already done this,
4386 // if it was prompted by constant non-zero subword initializers.
4387 // But this case can still arise with non-constant stores.
4388
4389 intptr_t next_full_store = find_next_fullword_store(i, phase);
4390
4391 // In the examples above:
4392 // in(i) p q r s x y z
4393 // st_off 12 13 14 15 12 13 14
4394 // st_size 1 1 1 1 1 1 1
4395 // next_full_s. 12 16 16 16 16 16 16
4396 // z's_done 12 16 16 16 12 16 12
4397 // z's_needed 12 16 16 16 16 16 16
4398 // zsize 0 0 0 0 4 0 4
4399 if (next_full_store < 0) {
4400 // Conservative tack: Zero to end of current word.
4401 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4402 } else {
4403 // Zero to beginning of next fully initialized word.
4404 // Or, don't zero at all, if we are already in that word.
4405 assert(next_full_store >= zeroes_needed, "must go forward");
4406 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4407 zeroes_needed = next_full_store;
4408 }
4409 }
4410
4411 if (zeroes_needed > zeroes_done) {
4412 intptr_t zsize = zeroes_needed - zeroes_done;
4413 // Do some incremental zeroing on rawmem, in parallel with inits.
4414 zeroes_done = align_down(zeroes_done, BytesPerInt);
4415 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4416 allocation()->in(AllocateNode::DefaultValue),
4417 allocation()->in(AllocateNode::RawDefaultValue),
4418 zeroes_done, zeroes_needed,
4419 phase);
4420 zeroes_done = zeroes_needed;
4421 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4422 do_zeroing = false; // leave the hole, next time
4423 }
4424 }
4425
4426 // Collect the store and move on:
4427 st->set_req(MemNode::Memory, inits);
4428 inits = st; // put it on the linearized chain
4429 set_req(i, zmem); // unhook from previous position
4430
4431 if (zeroes_done == st_off)
4432 zeroes_done = next_init_off;
4433
4434 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4435
4436 #ifdef ASSERT
4437 // Various order invariants. Weaker than stores_are_sane because
4438 // a large constant tile can be filled in by smaller non-constant stores.
4439 assert(st_off >= last_init_off, "inits do not reverse");
4440 last_init_off = st_off;
4441 const Type* val = NULL;
4442 if (st_size >= BytesPerInt &&
4443 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4444 (int)val->basic_type() < (int)T_OBJECT) {
4445 assert(st_off >= last_tile_end, "tiles do not overlap");
4446 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4447 last_tile_end = MAX2(last_tile_end, next_init_off);
4448 } else {
4449 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4450 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4451 assert(st_off >= last_init_end, "inits do not overlap");
4452 last_init_end = next_init_off; // it's a non-tile
4453 }
4454 #endif //ASSERT
4455 }
4456
4457 remove_extra_zeroes(); // clear out all the zmems left over
4458 add_req(inits);
4459
4460 if (!(UseTLAB && ZeroTLAB)) {
4461 // If anything remains to be zeroed, zero it all now.
4462 zeroes_done = align_down(zeroes_done, BytesPerInt);
4463 // if it is the last unused 4 bytes of an instance, forget about it
4464 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4465 if (zeroes_done + BytesPerLong >= size_limit) {
4466 AllocateNode* alloc = allocation();
4467 assert(alloc != NULL, "must be present");
4468 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4469 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4470 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4471 if (zeroes_done == k->layout_helper())
4472 zeroes_done = size_limit;
4473 }
4474 }
4475 if (zeroes_done < size_limit) {
4476 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4477 allocation()->in(AllocateNode::DefaultValue),
4478 allocation()->in(AllocateNode::RawDefaultValue),
4479 zeroes_done, size_in_bytes, phase);
4480 }
4481 }
4482
4483 set_complete(phase);
4484 return rawmem;
4485 }
4486
4487
4488 #ifdef ASSERT
4489 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4490 if (is_complete())
4491 return true; // stores could be anything at this point
4492 assert(allocation() != NULL, "must be present");
4493 intptr_t last_off = allocation()->minimum_header_size();
4494 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4495 Node* st = in(i);
4496 intptr_t st_off = get_store_offset(st, phase);
4497 if (st_off < 0) continue; // ignore dead garbage
4498 if (last_off > st_off) {
4499 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4500 this->dump(2);
4501 assert(false, "ascending store offsets");
4502 return false;
4503 }
4504 last_off = st_off + st->as_Store()->memory_size();
4505 }
4506 return true;
4507 }
4508 #endif //ASSERT
4509
4510
4511
4512
4513 //============================MergeMemNode=====================================
4514 //
4515 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4516 // contributing store or call operations. Each contributor provides the memory
4517 // state for a particular "alias type" (see Compile::alias_type). For example,
4518 // if a MergeMem has an input X for alias category #6, then any memory reference
4519 // to alias category #6 may use X as its memory state input, as an exact equivalent
4520 // to using the MergeMem as a whole.
4521 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4522 //
4523 // (Here, the <N> notation gives the index of the relevant adr_type.)
4524 //
4525 // In one special case (and more cases in the future), alias categories overlap.
4526 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4527 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4528 // it is exactly equivalent to that state W:
4529 // MergeMem(<Bot>: W) <==> W
4530 //
4531 // Usually, the merge has more than one input. In that case, where inputs
4532 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4533 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4534 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4535 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4536 //
4537 // A merge can take a "wide" memory state as one of its narrow inputs.
4538 // This simply means that the merge observes out only the relevant parts of
4539 // the wide input. That is, wide memory states arriving at narrow merge inputs
4540 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4541 //
4542 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4543 // and that memory slices "leak through":
4544 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4545 //
4546 // But, in such a cascade, repeated memory slices can "block the leak":
4547 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4548 //
4549 // In the last example, Y is not part of the combined memory state of the
4550 // outermost MergeMem. The system must, of course, prevent unschedulable
4551 // memory states from arising, so you can be sure that the state Y is somehow
4552 // a precursor to state Y'.
4553 //
4554 //
4555 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4556 // of each MergeMemNode array are exactly the numerical alias indexes, including
4557 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4558 // Compile::alias_type (and kin) produce and manage these indexes.
4559 //
4560 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4561 // (Note that this provides quick access to the top node inside MergeMem methods,
4562 // without the need to reach out via TLS to Compile::current.)
4563 //
4564 // As a consequence of what was just described, a MergeMem that represents a full
4565 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4566 // containing all alias categories.
4567 //
4568 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4569 //
4570 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4571 // a memory state for the alias type <N>, or else the top node, meaning that
4572 // there is no particular input for that alias type. Note that the length of
4573 // a MergeMem is variable, and may be extended at any time to accommodate new
4574 // memory states at larger alias indexes. When merges grow, they are of course
4575 // filled with "top" in the unused in() positions.
4576 //
4577 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4578 // (Top was chosen because it works smoothly with passes like GCM.)
4579 //
4580 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4581 // the type of random VM bits like TLS references.) Since it is always the
4582 // first non-Bot memory slice, some low-level loops use it to initialize an
4583 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4584 //
4585 //
4586 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4587 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4588 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4589 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4590 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4591 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4592 //
4593 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4594 // really that different from the other memory inputs. An abbreviation called
4595 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4596 //
4597 //
4598 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4599 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4600 // that "emerges though" the base memory will be marked as excluding the alias types
4601 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4602 //
4603 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4604 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4605 //
4606 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4607 // (It is currently unimplemented.) As you can see, the resulting merge is
4608 // actually a disjoint union of memory states, rather than an overlay.
4609 //
4610
4611 //------------------------------MergeMemNode-----------------------------------
4612 Node* MergeMemNode::make_empty_memory() {
4613 Node* empty_memory = (Node*) Compile::current()->top();
4614 assert(empty_memory->is_top(), "correct sentinel identity");
4615 return empty_memory;
4616 }
4617
4618 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4619 init_class_id(Class_MergeMem);
4620 // all inputs are nullified in Node::Node(int)
4621 // set_input(0, NULL); // no control input
4622
4623 // Initialize the edges uniformly to top, for starters.
4624 Node* empty_mem = make_empty_memory();
4625 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4626 init_req(i,empty_mem);
4627 }
4628 assert(empty_memory() == empty_mem, "");
4629
4630 if( new_base != NULL && new_base->is_MergeMem() ) {
4631 MergeMemNode* mdef = new_base->as_MergeMem();
4632 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4633 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4634 mms.set_memory(mms.memory2());
4635 }
4636 assert(base_memory() == mdef->base_memory(), "");
4637 } else {
4638 set_base_memory(new_base);
4639 }
4640 }
4641
4642 // Make a new, untransformed MergeMem with the same base as 'mem'.
4643 // If mem is itself a MergeMem, populate the result with the same edges.
4644 MergeMemNode* MergeMemNode::make(Node* mem) {
4645 return new MergeMemNode(mem);
4646 }
4647
4648 //------------------------------cmp--------------------------------------------
4649 uint MergeMemNode::hash() const { return NO_HASH; }
4650 bool MergeMemNode::cmp( const Node &n ) const {
4651 return (&n == this); // Always fail except on self
4652 }
4653
4654 //------------------------------Identity---------------------------------------
4655 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4656 // Identity if this merge point does not record any interesting memory
4657 // disambiguations.
4658 Node* base_mem = base_memory();
4659 Node* empty_mem = empty_memory();
4660 if (base_mem != empty_mem) { // Memory path is not dead?
4661 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4662 Node* mem = in(i);
4663 if (mem != empty_mem && mem != base_mem) {
4664 return this; // Many memory splits; no change
4665 }
4666 }
4667 }
4668 return base_mem; // No memory splits; ID on the one true input
4669 }
4670
4671 //------------------------------Ideal------------------------------------------
4672 // This method is invoked recursively on chains of MergeMem nodes
4673 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4674 // Remove chain'd MergeMems
4675 //
4676 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4677 // relative to the "in(Bot)". Since we are patching both at the same time,
4678 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4679 // but rewrite each "in(i)" relative to the new "in(Bot)".
4680 Node *progress = NULL;
4681
4682
4683 Node* old_base = base_memory();
4684 Node* empty_mem = empty_memory();
4685 if (old_base == empty_mem)
4686 return NULL; // Dead memory path.
4687
4688 MergeMemNode* old_mbase;
4689 if (old_base != NULL && old_base->is_MergeMem())
4690 old_mbase = old_base->as_MergeMem();
4691 else
4692 old_mbase = NULL;
4693 Node* new_base = old_base;
4694
4695 // simplify stacked MergeMems in base memory
4696 if (old_mbase) new_base = old_mbase->base_memory();
4697
4698 // the base memory might contribute new slices beyond my req()
4699 if (old_mbase) grow_to_match(old_mbase);
4700
4701 // Look carefully at the base node if it is a phi.
4702 PhiNode* phi_base;
4703 if (new_base != NULL && new_base->is_Phi())
4704 phi_base = new_base->as_Phi();
4705 else
4706 phi_base = NULL;
4707
4708 Node* phi_reg = NULL;
4709 uint phi_len = (uint)-1;
4710 if (phi_base != NULL && !phi_base->is_copy()) {
4711 // do not examine phi if degraded to a copy
4712 phi_reg = phi_base->region();
4713 phi_len = phi_base->req();
4714 // see if the phi is unfinished
4715 for (uint i = 1; i < phi_len; i++) {
4716 if (phi_base->in(i) == NULL) {
4717 // incomplete phi; do not look at it yet!
4718 phi_reg = NULL;
4719 phi_len = (uint)-1;
4720 break;
4721 }
4722 }
4723 }
4724
4725 // Note: We do not call verify_sparse on entry, because inputs
4726 // can normalize to the base_memory via subsume_node or similar
4727 // mechanisms. This method repairs that damage.
4728
4729 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4730
4731 // Look at each slice.
4732 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4733 Node* old_in = in(i);
4734 // calculate the old memory value
4735 Node* old_mem = old_in;
4736 if (old_mem == empty_mem) old_mem = old_base;
4737 assert(old_mem == memory_at(i), "");
4738
4739 // maybe update (reslice) the old memory value
4740
4741 // simplify stacked MergeMems
4742 Node* new_mem = old_mem;
4743 MergeMemNode* old_mmem;
4744 if (old_mem != NULL && old_mem->is_MergeMem())
4745 old_mmem = old_mem->as_MergeMem();
4746 else
4747 old_mmem = NULL;
4748 if (old_mmem == this) {
4749 // This can happen if loops break up and safepoints disappear.
4750 // A merge of BotPtr (default) with a RawPtr memory derived from a
4751 // safepoint can be rewritten to a merge of the same BotPtr with
4752 // the BotPtr phi coming into the loop. If that phi disappears
4753 // also, we can end up with a self-loop of the mergemem.
4754 // In general, if loops degenerate and memory effects disappear,
4755 // a mergemem can be left looking at itself. This simply means
4756 // that the mergemem's default should be used, since there is
4757 // no longer any apparent effect on this slice.
4758 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4759 // from start. Update the input to TOP.
4760 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4761 }
4762 else if (old_mmem != NULL) {
4763 new_mem = old_mmem->memory_at(i);
4764 }
4765 // else preceding memory was not a MergeMem
4766
4767 // replace equivalent phis (unfortunately, they do not GVN together)
4768 if (new_mem != NULL && new_mem != new_base &&
4769 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4770 if (new_mem->is_Phi()) {
4771 PhiNode* phi_mem = new_mem->as_Phi();
4772 for (uint i = 1; i < phi_len; i++) {
4773 if (phi_base->in(i) != phi_mem->in(i)) {
4774 phi_mem = NULL;
4775 break;
4776 }
4777 }
4778 if (phi_mem != NULL) {
4779 // equivalent phi nodes; revert to the def
4780 new_mem = new_base;
4781 }
4782 }
4783 }
4784
4785 // maybe store down a new value
4786 Node* new_in = new_mem;
4787 if (new_in == new_base) new_in = empty_mem;
4788
4789 if (new_in != old_in) {
4790 // Warning: Do not combine this "if" with the previous "if"
4791 // A memory slice might have be be rewritten even if it is semantically
4792 // unchanged, if the base_memory value has changed.
4793 set_req(i, new_in);
4794 progress = this; // Report progress
4795 }
4796 }
4797
4798 if (new_base != old_base) {
4799 set_req(Compile::AliasIdxBot, new_base);
4800 // Don't use set_base_memory(new_base), because we need to update du.
4801 assert(base_memory() == new_base, "");
4802 progress = this;
4803 }
4804
4805 if( base_memory() == this ) {
4806 // a self cycle indicates this memory path is dead
4807 set_req(Compile::AliasIdxBot, empty_mem);
4808 }
4809
4810 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4811 // Recursion must occur after the self cycle check above
4812 if( base_memory()->is_MergeMem() ) {
4813 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4814 Node *m = phase->transform(new_mbase); // Rollup any cycles
4815 if( m != NULL &&
4816 (m->is_top() ||
4817 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4818 // propagate rollup of dead cycle to self
4819 set_req(Compile::AliasIdxBot, empty_mem);
4820 }
4821 }
4822
4823 if( base_memory() == empty_mem ) {
4824 progress = this;
4825 // Cut inputs during Parse phase only.
4826 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4827 if( !can_reshape ) {
4828 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4829 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4830 }
4831 }
4832 }
4833
4834 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4835 // Check if PhiNode::Ideal's "Split phis through memory merges"
4836 // transform should be attempted. Look for this->phi->this cycle.
4837 uint merge_width = req();
4838 if (merge_width > Compile::AliasIdxRaw) {
4839 PhiNode* phi = base_memory()->as_Phi();
4840 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4841 if (phi->in(i) == this) {
4842 phase->is_IterGVN()->_worklist.push(phi);
4843 break;
4844 }
4845 }
4846 }
4847 }
4848
4849 assert(progress || verify_sparse(), "please, no dups of base");
4850 return progress;
4851 }
4852
4853 //-------------------------set_base_memory-------------------------------------
4854 void MergeMemNode::set_base_memory(Node *new_base) {
4855 Node* empty_mem = empty_memory();
4856 set_req(Compile::AliasIdxBot, new_base);
4857 assert(memory_at(req()) == new_base, "must set default memory");
4858 // Clear out other occurrences of new_base:
4859 if (new_base != empty_mem) {
4860 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4861 if (in(i) == new_base) set_req(i, empty_mem);
4862 }
4863 }
4864 }
4865
4866 //------------------------------out_RegMask------------------------------------
4867 const RegMask &MergeMemNode::out_RegMask() const {
4868 return RegMask::Empty;
4869 }
4870
4871 //------------------------------dump_spec--------------------------------------
4872 #ifndef PRODUCT
4873 void MergeMemNode::dump_spec(outputStream *st) const {
4874 st->print(" {");
4875 Node* base_mem = base_memory();
4876 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4877 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4878 if (mem == base_mem) { st->print(" -"); continue; }
4879 st->print( " N%d:", mem->_idx );
4880 Compile::current()->get_adr_type(i)->dump_on(st);
4881 }
4882 st->print(" }");
4883 }
4884 #endif // !PRODUCT
4885
4886
4887 #ifdef ASSERT
4888 static bool might_be_same(Node* a, Node* b) {
4889 if (a == b) return true;
4890 if (!(a->is_Phi() || b->is_Phi())) return false;
4891 // phis shift around during optimization
4892 return true; // pretty stupid...
4893 }
4894
4895 // verify a narrow slice (either incoming or outgoing)
4896 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4897 if (!VerifyAliases) return; // don't bother to verify unless requested
4898 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4899 if (Node::in_dump()) return; // muzzle asserts when printing
4900 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4901 assert(n != NULL, "");
4902 // Elide intervening MergeMem's
4903 while (n->is_MergeMem()) {
4904 n = n->as_MergeMem()->memory_at(alias_idx);
4905 }
4906 Compile* C = Compile::current();
4907 const TypePtr* n_adr_type = n->adr_type();
4908 if (n == m->empty_memory()) {
4909 // Implicit copy of base_memory()
4910 } else if (n_adr_type != TypePtr::BOTTOM) {
4911 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4912 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4913 } else {
4914 // A few places like make_runtime_call "know" that VM calls are narrow,
4915 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4916 bool expected_wide_mem = false;
4917 if (n == m->base_memory()) {
4918 expected_wide_mem = true;
4919 } else if (alias_idx == Compile::AliasIdxRaw ||
4920 n == m->memory_at(Compile::AliasIdxRaw)) {
4921 expected_wide_mem = true;
4922 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4923 // memory can "leak through" calls on channels that
4924 // are write-once. Allow this also.
4925 expected_wide_mem = true;
4926 }
4927 assert(expected_wide_mem, "expected narrow slice replacement");
4928 }
4929 }
4930 #else // !ASSERT
4931 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4932 #endif
4933
4934
4935 //-----------------------------memory_at---------------------------------------
4936 Node* MergeMemNode::memory_at(uint alias_idx) const {
4937 assert(alias_idx >= Compile::AliasIdxRaw ||
4938 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4939 "must avoid base_memory and AliasIdxTop");
4940
4941 // Otherwise, it is a narrow slice.
4942 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4943 Compile *C = Compile::current();
4944 if (is_empty_memory(n)) {
4945 // the array is sparse; empty slots are the "top" node
4946 n = base_memory();
4947 assert(Node::in_dump()
4948 || n == NULL || n->bottom_type() == Type::TOP
4949 || n->adr_type() == NULL // address is TOP
4950 || n->adr_type() == TypePtr::BOTTOM
4951 || n->adr_type() == TypeRawPtr::BOTTOM
4952 || Compile::current()->AliasLevel() == 0,
4953 "must be a wide memory");
4954 // AliasLevel == 0 if we are organizing the memory states manually.
4955 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4956 } else {
4957 // make sure the stored slice is sane
4958 #ifdef ASSERT
4959 if (VMError::is_error_reported() || Node::in_dump()) {
4960 } else if (might_be_same(n, base_memory())) {
4961 // Give it a pass: It is a mostly harmless repetition of the base.
4962 // This can arise normally from node subsumption during optimization.
4963 } else {
4964 verify_memory_slice(this, alias_idx, n);
4965 }
4966 #endif
4967 }
4968 return n;
4969 }
4970
4971 //---------------------------set_memory_at-------------------------------------
4972 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4973 verify_memory_slice(this, alias_idx, n);
4974 Node* empty_mem = empty_memory();
4975 if (n == base_memory()) n = empty_mem; // collapse default
4976 uint need_req = alias_idx+1;
4977 if (req() < need_req) {
4978 if (n == empty_mem) return; // already the default, so do not grow me
4979 // grow the sparse array
4980 do {
4981 add_req(empty_mem);
4982 } while (req() < need_req);
4983 }
4984 set_req( alias_idx, n );
4985 }
4986
4987
4988
4989 //--------------------------iteration_setup------------------------------------
4990 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4991 if (other != NULL) {
4992 grow_to_match(other);
4993 // invariant: the finite support of mm2 is within mm->req()
4994 #ifdef ASSERT
4995 for (uint i = req(); i < other->req(); i++) {
4996 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4997 }
4998 #endif
4999 }
5000 // Replace spurious copies of base_memory by top.
5001 Node* base_mem = base_memory();
5002 if (base_mem != NULL && !base_mem->is_top()) {
5003 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
5004 if (in(i) == base_mem)
5005 set_req(i, empty_memory());
5006 }
5007 }
5008 }
5009
5010 //---------------------------grow_to_match-------------------------------------
5011 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
5012 Node* empty_mem = empty_memory();
5013 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
5014 // look for the finite support of the other memory
5015 for (uint i = other->req(); --i >= req(); ) {
5016 if (other->in(i) != empty_mem) {
5017 uint new_len = i+1;
5018 while (req() < new_len) add_req(empty_mem);
5019 break;
5020 }
5021 }
5022 }
5023
5024 //---------------------------verify_sparse-------------------------------------
5025 #ifndef PRODUCT
5026 bool MergeMemNode::verify_sparse() const {
5027 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
5028 Node* base_mem = base_memory();
5029 // The following can happen in degenerate cases, since empty==top.
5030 if (is_empty_memory(base_mem)) return true;
5031 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5032 assert(in(i) != NULL, "sane slice");
5033 if (in(i) == base_mem) return false; // should have been the sentinel value!
5034 }
5035 return true;
5036 }
5037
5038 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5039 Node* n;
5040 n = mm->in(idx);
5041 if (mem == n) return true; // might be empty_memory()
5042 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5043 if (mem == n) return true;
5044 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
5045 if (mem == n) return true;
5046 if (n == NULL) break;
5047 }
5048 return false;
5049 }
5050 #endif // !PRODUCT