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));
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 = Pinned;
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 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1474
1475 // Split through Phi (see original code in loopopts.cpp).
1476 assert(C->have_alias_type(t_oop), "instance should have alias type");
1477
1478 // Do nothing here if Identity will find a value
1479 // (to avoid infinite chain of value phis generation).
1480 if (!phase->eqv(this, phase->apply_identity(this)))
1481 return NULL;
1482
1483 // Select Region to split through.
1484 Node* region;
1485 if (!base_is_phi) {
1486 assert(mem->is_Phi(), "sanity");
1487 region = mem->in(0);
1488 // Skip if the region dominates some control edge of the address.
1489 if (!MemNode::all_controls_dominate(address, region))
1490 return NULL;
1491 } else if (!mem->is_Phi()) {
1492 assert(base_is_phi, "sanity");
1493 region = base->in(0);
1494 // Skip if the region dominates some control edge of the memory.
1495 if (!MemNode::all_controls_dominate(mem, region))
1496 return NULL;
1497 } else if (base->in(0) != mem->in(0)) {
1498 assert(base_is_phi && mem->is_Phi(), "sanity");
1499 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1500 region = base->in(0);
1501 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1502 region = mem->in(0);
1503 } else {
1504 return NULL; // complex graph
1505 }
1506 } else {
1507 assert(base->in(0) == mem->in(0), "sanity");
1508 region = mem->in(0);
1509 }
1510
1511 const Type* this_type = this->bottom_type();
1512 int this_index = C->get_alias_index(t_oop);
1513 int this_offset = t_oop->offset();
1514 int this_iid = t_oop->instance_id();
1515 if (!t_oop->is_known_instance() && load_boxed_values) {
1516 // Use _idx of address base for boxed values.
1517 this_iid = base->_idx;
1518 }
1519 PhaseIterGVN* igvn = phase->is_IterGVN();
1520 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1521 for (uint i = 1; i < region->req(); i++) {
1522 Node* x;
1523 Node* the_clone = NULL;
1524 if (region->in(i) == C->top()) {
1525 x = C->top(); // Dead path? Use a dead data op
1526 } else {
1527 x = this->clone(); // Else clone up the data op
1528 the_clone = x; // Remember for possible deletion.
1529 // Alter data node to use pre-phi inputs
1530 if (this->in(0) == region) {
1531 x->set_req(0, region->in(i));
1532 } else {
1533 x->set_req(0, NULL);
1534 }
1535 if (mem->is_Phi() && (mem->in(0) == region)) {
1536 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1537 }
1538 if (address->is_Phi() && address->in(0) == region) {
1539 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1540 }
1541 if (base_is_phi && (base->in(0) == region)) {
1542 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1543 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1544 x->set_req(Address, adr_x);
1545 }
1546 }
1547 // Check for a 'win' on some paths
1548 const Type *t = x->Value(igvn);
1549
1550 bool singleton = t->singleton();
1551
1552 // See comments in PhaseIdealLoop::split_thru_phi().
1553 if (singleton && t == Type::TOP) {
1554 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1555 }
1556
1557 if (singleton) {
1558 x = igvn->makecon(t);
1559 } else {
1560 // We now call Identity to try to simplify the cloned node.
1561 // Note that some Identity methods call phase->type(this).
1562 // Make sure that the type array is big enough for
1563 // our new node, even though we may throw the node away.
1564 // (This tweaking with igvn only works because x is a new node.)
1565 igvn->set_type(x, t);
1566 // If x is a TypeNode, capture any more-precise type permanently into Node
1567 // otherwise it will be not updated during igvn->transform since
1568 // igvn->type(x) is set to x->Value() already.
1569 x->raise_bottom_type(t);
1570 Node *y = igvn->apply_identity(x);
1571 if (y != x) {
1572 x = y;
1573 } else {
1574 y = igvn->hash_find_insert(x);
1575 if (y) {
1576 x = y;
1577 } else {
1578 // Else x is a new node we are keeping
1579 // We do not need register_new_node_with_optimizer
1580 // because set_type has already been called.
1581 igvn->_worklist.push(x);
1582 }
1583 }
1584 }
1585 if (x != the_clone && the_clone != NULL) {
1586 igvn->remove_dead_node(the_clone);
1587 }
1588 phi->set_req(i, x);
1589 }
1590 // Record Phi
1591 igvn->register_new_node_with_optimizer(phi);
1592 return phi;
1593 }
1594
1595 //------------------------------Ideal------------------------------------------
1596 // If the load is from Field memory and the pointer is non-null, it might be possible to
1597 // zero out the control input.
1598 // If the offset is constant and the base is an object allocation,
1599 // try to hook me up to the exact initializing store.
1600 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1601 Node* p = MemNode::Ideal_common(phase, can_reshape);
1602 if (p) return (p == NodeSentinel) ? NULL : p;
1603
1604 Node* ctrl = in(MemNode::Control);
1605 Node* address = in(MemNode::Address);
1606 bool progress = false;
1607
1608 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1609 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1610
1611 // Skip up past a SafePoint control. Cannot do this for Stores because
1612 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1613 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1614 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1615 !addr_mark ) {
1616 ctrl = ctrl->in(0);
1617 set_req(MemNode::Control,ctrl);
1618 progress = true;
1619 }
1620
1621 intptr_t ignore = 0;
1622 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1623 if (base != NULL
1624 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1625 // Check for useless control edge in some common special cases
1626 if (in(MemNode::Control) != NULL
1627 && can_remove_control()
1628 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1629 && all_controls_dominate(base, phase->C->start())) {
1630 // A method-invariant, non-null address (constant or 'this' argument).
1631 set_req(MemNode::Control, NULL);
1632 progress = true;
1633 }
1634 }
1635
1636 Node* mem = in(MemNode::Memory);
1637 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1638
1639 if (can_reshape && (addr_t != NULL)) {
1640 // try to optimize our memory input
1641 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1642 if (opt_mem != mem) {
1643 set_req(MemNode::Memory, opt_mem);
1644 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1645 return this;
1646 }
1647 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1648 if ((t_oop != NULL) &&
1649 (t_oop->is_known_instance_field() ||
1650 t_oop->is_ptr_to_boxed_value())) {
1651 PhaseIterGVN *igvn = phase->is_IterGVN();
1652 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1653 // Delay this transformation until memory Phi is processed.
1654 phase->is_IterGVN()->_worklist.push(this);
1655 return NULL;
1656 }
1657 // Split instance field load through Phi.
1658 Node* result = split_through_phi(phase);
1659 if (result != NULL) return result;
1660
1661 if (t_oop->is_ptr_to_boxed_value()) {
1662 Node* result = eliminate_autobox(phase);
1663 if (result != NULL) return result;
1664 }
1665 }
1666 }
1667
1668 // Is there a dominating load that loads the same value? Leave
1669 // anything that is not a load of a field/array element (like
1670 // barriers etc.) alone
1671 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1672 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1673 Node *use = mem->fast_out(i);
1674 if (use != this &&
1675 use->Opcode() == Opcode() &&
1676 use->in(0) != NULL &&
1677 use->in(0) != in(0) &&
1678 use->in(Address) == in(Address)) {
1679 Node* ctl = in(0);
1680 for (int i = 0; i < 10 && ctl != NULL; i++) {
1681 ctl = IfNode::up_one_dom(ctl);
1682 if (ctl == use->in(0)) {
1683 set_req(0, use->in(0));
1684 return this;
1685 }
1686 }
1687 }
1688 }
1689 }
1690
1691 // Check for prior store with a different base or offset; make Load
1692 // independent. Skip through any number of them. Bail out if the stores
1693 // are in an endless dead cycle and report no progress. This is a key
1694 // transform for Reflection. However, if after skipping through the Stores
1695 // we can't then fold up against a prior store do NOT do the transform as
1696 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1697 // array memory alive twice: once for the hoisted Load and again after the
1698 // bypassed Store. This situation only works if EVERYBODY who does
1699 // anti-dependence work knows how to bypass. I.e. we need all
1700 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1701 // the alias index stuff. So instead, peek through Stores and IFF we can
1702 // fold up, do so.
1703 Node* prev_mem = find_previous_store(phase);
1704 if (prev_mem != NULL) {
1705 Node* value = can_see_arraycopy_value(prev_mem, phase);
1706 if (value != NULL) {
1707 return value;
1708 }
1709 }
1710 // Steps (a), (b): Walk past independent stores to find an exact match.
1711 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1712 // (c) See if we can fold up on the spot, but don't fold up here.
1713 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1714 // just return a prior value, which is done by Identity calls.
1715 if (can_see_stored_value(prev_mem, phase)) {
1716 // Make ready for step (d):
1717 set_req(MemNode::Memory, prev_mem);
1718 return this;
1719 }
1720 }
1721
1722 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1723 if (alloc != NULL && mem->is_Proj() &&
1724 mem->in(0) != NULL &&
1725 mem->in(0) == alloc->initialization() &&
1726 Opcode() == Op_LoadX &&
1727 alloc->initialization()->proj_out_or_null(0) != NULL) {
1728 InitializeNode* init = alloc->initialization();
1729 Node* control = init->proj_out(0);
1730 return alloc->make_ideal_mark(phase, address, control, mem, NULL);
1731 }
1732
1733 return progress ? this : NULL;
1734 }
1735
1736 // Helper to recognize certain Klass fields which are invariant across
1737 // some group of array types (e.g., int[] or all T[] where T < Object).
1738 const Type*
1739 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1740 ciKlass* klass) const {
1741 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1742 // The field is Klass::_modifier_flags. Return its (constant) value.
1743 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1744 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1745 return TypeInt::make(klass->modifier_flags());
1746 }
1747 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1748 // The field is Klass::_access_flags. Return its (constant) value.
1749 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1750 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1751 return TypeInt::make(klass->access_flags());
1752 }
1753 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1754 // The field is Klass::_layout_helper. Return its constant value if known.
1755 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1756 return TypeInt::make(klass->layout_helper());
1757 }
1758
1759 // No match.
1760 return NULL;
1761 }
1762
1763 //------------------------------Value-----------------------------------------
1764 const Type* LoadNode::Value(PhaseGVN* phase) const {
1765 // Either input is TOP ==> the result is TOP
1766 Node* mem = in(MemNode::Memory);
1767 const Type *t1 = phase->type(mem);
1768 if (t1 == Type::TOP) return Type::TOP;
1769 Node* adr = in(MemNode::Address);
1770 const TypePtr* tp = phase->type(adr)->isa_ptr();
1771 if (tp == NULL || tp->empty()) return Type::TOP;
1772 int off = tp->offset();
1773 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1774 Compile* C = phase->C;
1775
1776 // Try to guess loaded type from pointer type
1777 if (tp->isa_aryptr()) {
1778 const TypeAryPtr* ary = tp->is_aryptr();
1779 const Type* t = ary->elem();
1780
1781 // Determine whether the reference is beyond the header or not, by comparing
1782 // the offset against the offset of the start of the array's data.
1783 // Different array types begin at slightly different offsets (12 vs. 16).
1784 // We choose T_BYTE as an example base type that is least restrictive
1785 // as to alignment, which will therefore produce the smallest
1786 // possible base offset.
1787 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1788 const bool off_beyond_header = (off >= min_base_off);
1789
1790 // Try to constant-fold a stable array element.
1791 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1792 // Make sure the reference is not into the header and the offset is constant
1793 ciObject* aobj = ary->const_oop();
1794 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1795 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1796 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1797 stable_dimension,
1798 memory_type(), is_unsigned());
1799 if (con_type != NULL) {
1800 return con_type;
1801 }
1802 }
1803 }
1804
1805 // Don't do this for integer types. There is only potential profit if
1806 // the element type t is lower than _type; that is, for int types, if _type is
1807 // more restrictive than t. This only happens here if one is short and the other
1808 // char (both 16 bits), and in those cases we've made an intentional decision
1809 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1810 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1811 //
1812 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1813 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1814 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1815 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1816 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1817 // In fact, that could have been the original type of p1, and p1 could have
1818 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1819 // expression (LShiftL quux 3) independently optimized to the constant 8.
1820 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1821 && (_type->isa_vect() == NULL)
1822 && t->isa_valuetype() == NULL
1823 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1824 // t might actually be lower than _type, if _type is a unique
1825 // concrete subclass of abstract class t.
1826 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
1827 const Type* jt = t->join_speculative(_type);
1828 // In any case, do not allow the join, per se, to empty out the type.
1829 if (jt->empty() && !t->empty()) {
1830 // This can happen if a interface-typed array narrows to a class type.
1831 jt = _type;
1832 }
1833 #ifdef ASSERT
1834 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1835 // The pointers in the autobox arrays are always non-null
1836 Node* base = adr->in(AddPNode::Base);
1837 if ((base != NULL) && base->is_DecodeN()) {
1838 // Get LoadN node which loads IntegerCache.cache field
1839 base = base->in(1);
1840 }
1841 if ((base != NULL) && base->is_Con()) {
1842 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1843 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1844 // It could be narrow oop
1845 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1846 }
1847 }
1848 }
1849 #endif
1850 return jt;
1851 }
1852 }
1853 } else if (tp->base() == Type::InstPtr) {
1854 assert( off != Type::OffsetBot ||
1855 // arrays can be cast to Objects
1856 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1857 tp->is_oopptr()->klass() == ciEnv::current()->Class_klass() ||
1858 // unsafe field access may not have a constant offset
1859 C->has_unsafe_access(),
1860 "Field accesses must be precise" );
1861 // For oop loads, we expect the _type to be precise.
1862
1863 const TypeInstPtr* tinst = tp->is_instptr();
1864 BasicType bt = memory_type();
1865
1866 // Fold component and value mirror loads
1867 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1868 if (ik == phase->C->env()->Class_klass() && (off == java_lang_Class::component_mirror_offset_in_bytes() ||
1869 off == java_lang_Class::inline_mirror_offset_in_bytes())) {
1870 ciType* mirror_type = tinst->java_mirror_type();
1871 if (mirror_type != NULL) {
1872 const Type* const_oop = TypePtr::NULL_PTR;
1873 if (mirror_type->is_array_klass()) {
1874 const_oop = TypeInstPtr::make(mirror_type->as_array_klass()->component_mirror_instance());
1875 } else if (mirror_type->is_valuetype()) {
1876 const_oop = TypeInstPtr::make(mirror_type->as_value_klass()->inline_mirror_instance());
1877 }
1878 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop;
1879 }
1880 }
1881
1882 // Optimize loads from constant fields.
1883 ciObject* const_oop = tinst->const_oop();
1884 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1885 ciType* mirror_type = const_oop->as_instance()->java_mirror_type();
1886 if (mirror_type != NULL && mirror_type->is_valuetype()) {
1887 ciValueKlass* vk = mirror_type->as_value_klass();
1888 if (off == vk->default_value_offset()) {
1889 // Loading a special hidden field that contains the oop of the default value type
1890 const Type* const_oop = TypeInstPtr::make(vk->default_value_instance());
1891 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop;
1892 }
1893 }
1894 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
1895 if (con_type != NULL) {
1896 return con_type;
1897 }
1898 }
1899 } else if (tp->base() == Type::KlassPtr) {
1900 assert( off != Type::OffsetBot ||
1901 // arrays can be cast to Objects
1902 tp->is_klassptr()->klass() == NULL ||
1903 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1904 // also allow array-loading from the primary supertype
1905 // array during subtype checks
1906 Opcode() == Op_LoadKlass,
1907 "Field accesses must be precise" );
1908 // For klass/static loads, we expect the _type to be precise
1909 } else if (tp->base() == Type::RawPtr && !StressReflectiveCode) {
1910 if (adr->is_Load() && off == 0) {
1911 /* With mirrors being an indirect in the Klass*
1912 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1913 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1914 *
1915 * So check the type and klass of the node before the LoadP.
1916 */
1917 Node* adr2 = adr->in(MemNode::Address);
1918 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1919 if (tkls != NULL) {
1920 ciKlass* klass = tkls->klass();
1921 if (klass != NULL && klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1922 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1923 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1924 return TypeInstPtr::make(klass->java_mirror());
1925 }
1926 }
1927 } else {
1928 // Check for a load of the default value offset from the ValueKlassFixedBlock:
1929 // LoadI(LoadP(value_klass, adr_valueklass_fixed_block_offset), default_value_offset_offset)
1930 intptr_t offset = 0;
1931 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1932 if (base != NULL && base->is_Load() && offset == in_bytes(ValueKlass::default_value_offset_offset())) {
1933 const TypeKlassPtr* tkls = phase->type(base->in(MemNode::Address))->isa_klassptr();
1934 if (tkls != NULL && tkls->is_loaded() && tkls->klass_is_exact() && tkls->isa_valuetype() &&
1935 tkls->offset() == in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset())) {
1936 assert(base->Opcode() == Op_LoadP, "must load an oop from klass");
1937 assert(Opcode() == Op_LoadI, "must load an int from fixed block");
1938 return TypeInt::make(tkls->klass()->as_value_klass()->default_value_offset());
1939 }
1940 }
1941 }
1942 }
1943
1944 const TypeKlassPtr *tkls = tp->isa_klassptr();
1945 if (tkls != NULL && !StressReflectiveCode) {
1946 ciKlass* klass = tkls->klass();
1947 if (tkls->is_loaded() && tkls->klass_is_exact()) {
1948 // We are loading a field from a Klass metaobject whose identity
1949 // is known at compile time (the type is "exact" or "precise").
1950 // Check for fields we know are maintained as constants by the VM.
1951 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1952 // The field is Klass::_super_check_offset. Return its (constant) value.
1953 // (Folds up type checking code.)
1954 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1955 return TypeInt::make(klass->super_check_offset());
1956 }
1957 // Compute index into primary_supers array
1958 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1959 // Check for overflowing; use unsigned compare to handle the negative case.
1960 if( depth < ciKlass::primary_super_limit() ) {
1961 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1962 // (Folds up type checking code.)
1963 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1964 ciKlass *ss = klass->super_of_depth(depth);
1965 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1966 }
1967 const Type* aift = load_array_final_field(tkls, klass);
1968 if (aift != NULL) return aift;
1969 }
1970
1971 // We can still check if we are loading from the primary_supers array at a
1972 // shallow enough depth. Even though the klass is not exact, entries less
1973 // than or equal to its super depth are correct.
1974 if (tkls->is_loaded()) {
1975 ciType *inner = klass;
1976 while( inner->is_obj_array_klass() )
1977 inner = inner->as_obj_array_klass()->base_element_type();
1978 if( inner->is_instance_klass() &&
1979 !inner->as_instance_klass()->flags().is_interface() ) {
1980 // Compute index into primary_supers array
1981 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1982 // Check for overflowing; use unsigned compare to handle the negative case.
1983 if( depth < ciKlass::primary_super_limit() &&
1984 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1985 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1986 // (Folds up type checking code.)
1987 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1988 ciKlass *ss = klass->super_of_depth(depth);
1989 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1990 }
1991 }
1992 }
1993
1994 // If the type is enough to determine that the thing is not an array,
1995 // we can give the layout_helper a positive interval type.
1996 // This will help short-circuit some reflective code.
1997 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1998 && !klass->is_array_klass() // not directly typed as an array
1999 && !klass->is_interface() // specifically not Serializable & Cloneable
2000 && !klass->is_java_lang_Object() // not the supertype of all T[]
2001 ) {
2002 // Note: When interfaces are reliable, we can narrow the interface
2003 // test to (klass != Serializable && klass != Cloneable).
2004 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2005 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2006 // The key property of this type is that it folds up tests
2007 // for array-ness, since it proves that the layout_helper is positive.
2008 // Thus, a generic value like the basic object layout helper works fine.
2009 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2010 }
2011 }
2012
2013 // If we are loading from a freshly-allocated object, produce a zero,
2014 // if the load is provably beyond the header of the object.
2015 // (Also allow a variable load from a fresh array to produce zero.)
2016 const TypeOopPtr *tinst = tp->isa_oopptr();
2017 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
2018 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
2019 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2020 Node* value = can_see_stored_value(mem,phase);
2021 if (value != NULL && value->is_Con()) {
2022 assert(value->bottom_type()->higher_equal(_type),"sanity");
2023 return value->bottom_type();
2024 }
2025 }
2026
2027 if (is_instance) {
2028 // If we have an instance type and our memory input is the
2029 // programs's initial memory state, there is no matching store,
2030 // so just return a zero of the appropriate type
2031 Node *mem = in(MemNode::Memory);
2032 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2033 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2034 return Type::get_zero_type(_type->basic_type());
2035 }
2036 }
2037 return _type;
2038 }
2039
2040 //------------------------------match_edge-------------------------------------
2041 // Do we Match on this edge index or not? Match only the address.
2042 uint LoadNode::match_edge(uint idx) const {
2043 return idx == MemNode::Address;
2044 }
2045
2046 //--------------------------LoadBNode::Ideal--------------------------------------
2047 //
2048 // If the previous store is to the same address as this load,
2049 // and the value stored was larger than a byte, replace this load
2050 // with the value stored truncated to a byte. If no truncation is
2051 // needed, the replacement is done in LoadNode::Identity().
2052 //
2053 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2054 Node* mem = in(MemNode::Memory);
2055 Node* value = can_see_stored_value(mem,phase);
2056 if( value && !phase->type(value)->higher_equal( _type ) ) {
2057 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
2058 return new RShiftINode(result, phase->intcon(24));
2059 }
2060 // Identity call will handle the case where truncation is not needed.
2061 return LoadNode::Ideal(phase, can_reshape);
2062 }
2063
2064 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2065 Node* mem = in(MemNode::Memory);
2066 Node* value = can_see_stored_value(mem,phase);
2067 if (value != NULL && value->is_Con() &&
2068 !value->bottom_type()->higher_equal(_type)) {
2069 // If the input to the store does not fit with the load's result type,
2070 // it must be truncated. We can't delay until Ideal call since
2071 // a singleton Value is needed for split_thru_phi optimization.
2072 int con = value->get_int();
2073 return TypeInt::make((con << 24) >> 24);
2074 }
2075 return LoadNode::Value(phase);
2076 }
2077
2078 //--------------------------LoadUBNode::Ideal-------------------------------------
2079 //
2080 // If the previous store is to the same address as this load,
2081 // and the value stored was larger than a byte, replace this load
2082 // with the value stored truncated to a byte. If no truncation is
2083 // needed, the replacement is done in LoadNode::Identity().
2084 //
2085 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2086 Node* mem = in(MemNode::Memory);
2087 Node* value = can_see_stored_value(mem, phase);
2088 if (value && !phase->type(value)->higher_equal(_type))
2089 return new AndINode(value, phase->intcon(0xFF));
2090 // Identity call will handle the case where truncation is not needed.
2091 return LoadNode::Ideal(phase, can_reshape);
2092 }
2093
2094 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2095 Node* mem = in(MemNode::Memory);
2096 Node* value = can_see_stored_value(mem,phase);
2097 if (value != NULL && value->is_Con() &&
2098 !value->bottom_type()->higher_equal(_type)) {
2099 // If the input to the store does not fit with the load's result type,
2100 // it must be truncated. We can't delay until Ideal call since
2101 // a singleton Value is needed for split_thru_phi optimization.
2102 int con = value->get_int();
2103 return TypeInt::make(con & 0xFF);
2104 }
2105 return LoadNode::Value(phase);
2106 }
2107
2108 //--------------------------LoadUSNode::Ideal-------------------------------------
2109 //
2110 // If the previous store is to the same address as this load,
2111 // and the value stored was larger than a char, replace this load
2112 // with the value stored truncated to a char. If no truncation is
2113 // needed, the replacement is done in LoadNode::Identity().
2114 //
2115 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2116 Node* mem = in(MemNode::Memory);
2117 Node* value = can_see_stored_value(mem,phase);
2118 if( value && !phase->type(value)->higher_equal( _type ) )
2119 return new AndINode(value,phase->intcon(0xFFFF));
2120 // Identity call will handle the case where truncation is not needed.
2121 return LoadNode::Ideal(phase, can_reshape);
2122 }
2123
2124 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2125 Node* mem = in(MemNode::Memory);
2126 Node* value = can_see_stored_value(mem,phase);
2127 if (value != NULL && value->is_Con() &&
2128 !value->bottom_type()->higher_equal(_type)) {
2129 // If the input to the store does not fit with the load's result type,
2130 // it must be truncated. We can't delay until Ideal call since
2131 // a singleton Value is needed for split_thru_phi optimization.
2132 int con = value->get_int();
2133 return TypeInt::make(con & 0xFFFF);
2134 }
2135 return LoadNode::Value(phase);
2136 }
2137
2138 //--------------------------LoadSNode::Ideal--------------------------------------
2139 //
2140 // If the previous store is to the same address as this load,
2141 // and the value stored was larger than a short, replace this load
2142 // with the value stored truncated to a short. If no truncation is
2143 // needed, the replacement is done in LoadNode::Identity().
2144 //
2145 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2146 Node* mem = in(MemNode::Memory);
2147 Node* value = can_see_stored_value(mem,phase);
2148 if( value && !phase->type(value)->higher_equal( _type ) ) {
2149 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2150 return new RShiftINode(result, phase->intcon(16));
2151 }
2152 // Identity call will handle the case where truncation is not needed.
2153 return LoadNode::Ideal(phase, can_reshape);
2154 }
2155
2156 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2157 Node* mem = in(MemNode::Memory);
2158 Node* value = can_see_stored_value(mem,phase);
2159 if (value != NULL && value->is_Con() &&
2160 !value->bottom_type()->higher_equal(_type)) {
2161 // If the input to the store does not fit with the load's result type,
2162 // it must be truncated. We can't delay until Ideal call since
2163 // a singleton Value is needed for split_thru_phi optimization.
2164 int con = value->get_int();
2165 return TypeInt::make((con << 16) >> 16);
2166 }
2167 return LoadNode::Value(phase);
2168 }
2169
2170 //=============================================================================
2171 //----------------------------LoadKlassNode::make------------------------------
2172 // Polymorphic factory method:
2173 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2174 // sanity check the alias category against the created node type
2175 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2176 assert(adr_type != NULL, "expecting TypeKlassPtr");
2177 #ifdef _LP64
2178 if (adr_type->is_ptr_to_narrowklass()) {
2179 assert(UseCompressedClassPointers, "no compressed klasses");
2180 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2181 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2182 }
2183 #endif
2184 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2185 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2186 }
2187
2188 //------------------------------Value------------------------------------------
2189 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2190 return klass_value_common(phase);
2191 }
2192
2193 // In most cases, LoadKlassNode does not have the control input set. If the control
2194 // input is set, it must not be removed (by LoadNode::Ideal()).
2195 bool LoadKlassNode::can_remove_control() const {
2196 return false;
2197 }
2198
2199 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2200 // Either input is TOP ==> the result is TOP
2201 const Type *t1 = phase->type( in(MemNode::Memory) );
2202 if (t1 == Type::TOP) return Type::TOP;
2203 Node *adr = in(MemNode::Address);
2204 const Type *t2 = phase->type( adr );
2205 if (t2 == Type::TOP) return Type::TOP;
2206 const TypePtr *tp = t2->is_ptr();
2207 if (TypePtr::above_centerline(tp->ptr()) ||
2208 tp->ptr() == TypePtr::Null) return Type::TOP;
2209
2210 // Return a more precise klass, if possible
2211 const TypeInstPtr *tinst = tp->isa_instptr();
2212 if (tinst != NULL) {
2213 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2214 int offset = tinst->offset();
2215 if (ik == phase->C->env()->Class_klass()
2216 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2217 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2218 // We are loading a special hidden field from a Class mirror object,
2219 // the field which points to the VM's Klass metaobject.
2220 bool is_indirect_type = true;
2221 ciType* t = tinst->java_mirror_type(&is_indirect_type);
2222 // java_mirror_type returns non-null for compile-time Class constants.
2223 if (t != NULL) {
2224 // constant oop => constant klass
2225 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2226 if (t->is_void()) {
2227 // We cannot create a void array. Since void is a primitive type return null
2228 // klass. Users of this result need to do a null check on the returned klass.
2229 return TypePtr::NULL_PTR;
2230 }
2231 return TypeKlassPtr::make(ciArrayKlass::make(t, /* never_null */ !is_indirect_type));
2232 }
2233 if (!t->is_klass()) {
2234 // a primitive Class (e.g., int.class) has NULL for a klass field
2235 return TypePtr::NULL_PTR;
2236 }
2237 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2238 return TypeKlassPtr::make(t->as_klass());
2239 }
2240 // non-constant mirror, so we can't tell what's going on
2241 }
2242 if( !ik->is_loaded() )
2243 return _type; // Bail out if not loaded
2244 if (offset == oopDesc::klass_offset_in_bytes()) {
2245 if (tinst->klass_is_exact()) {
2246 return TypeKlassPtr::make(ik);
2247 }
2248 // See if we can become precise: no subklasses and no interface
2249 // (Note: We need to support verified interfaces.)
2250 if (!ik->is_interface() && !ik->has_subklass()) {
2251 //assert(!UseExactTypes, "this code should be useless with exact types");
2252 // Add a dependence; if any subclass added we need to recompile
2253 if (!ik->is_final()) {
2254 // %%% should use stronger assert_unique_concrete_subtype instead
2255 phase->C->dependencies()->assert_leaf_type(ik);
2256 }
2257 // Return precise klass
2258 return TypeKlassPtr::make(ik);
2259 }
2260
2261 // Return root of possible klass
2262 return TypeKlassPtr::make(TypePtr::NotNull, ik, Type::Offset(0), tinst->flatten_array());
2263 }
2264 }
2265
2266 // Check for loading klass from an array
2267 const TypeAryPtr *tary = tp->isa_aryptr();
2268 if (tary != NULL) {
2269 ciKlass *tary_klass = tary->klass();
2270 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2271 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2272 ciArrayKlass* ak = tary_klass->as_array_klass();
2273 // Do not fold klass loads from [V?. The runtime type might be [V due to [V <: [V?
2274 // and the klass for [V is not equal to the klass for [V?.
2275 if (tary->klass_is_exact()) {
2276 return TypeKlassPtr::make(tary_klass);
2277 }
2278
2279 // If the klass is an object array, we defer the question to the
2280 // array component klass.
2281 if (ak->is_obj_array_klass()) {
2282 assert(ak->is_loaded(), "");
2283 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2284 if (base_k->is_loaded() && base_k->is_instance_klass()) {
2285 ciInstanceKlass *ik = base_k->as_instance_klass();
2286 // See if we can become precise: no subklasses and no interface
2287 if (!ik->is_interface() && !ik->has_subklass() && (!ik->is_valuetype() || ak->storage_properties().is_null_free())) {
2288 //assert(!UseExactTypes, "this code should be useless with exact types");
2289 // Add a dependence; if any subclass added we need to recompile
2290 if (!ik->is_final()) {
2291 phase->C->dependencies()->assert_leaf_type(ik);
2292 }
2293 // Return precise array klass
2294 return TypeKlassPtr::make(ak);
2295 }
2296 }
2297 return TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0), false);
2298 } else if (ak->is_type_array_klass()) {
2299 //assert(!UseExactTypes, "this code should be useless with exact types");
2300 return TypeKlassPtr::make(ak); // These are always precise
2301 }
2302 }
2303 }
2304
2305 // Check for loading klass from an array klass
2306 const TypeKlassPtr *tkls = tp->isa_klassptr();
2307 if (tkls != NULL && !StressReflectiveCode) {
2308 if (!tkls->is_loaded()) {
2309 return _type; // Bail out if not loaded
2310 }
2311 ciKlass* klass = tkls->klass();
2312 if( klass->is_obj_array_klass() &&
2313 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2314 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2315 // // Always returning precise element type is incorrect,
2316 // // e.g., element type could be object and array may contain strings
2317 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2318
2319 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2320 // according to the element type's subclassing.
2321 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0), elem->flatten_array());
2322 } else if (klass->is_value_array_klass() &&
2323 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2324 ciKlass* elem = klass->as_value_array_klass()->element_klass();
2325 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0), true);
2326 }
2327 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2328 tkls->offset() == in_bytes(Klass::super_offset())) {
2329 ciKlass* sup = klass->as_instance_klass()->super();
2330 // The field is Klass::_super. Return its (constant) value.
2331 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2332 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2333 }
2334 }
2335
2336 // Bailout case
2337 return LoadNode::Value(phase);
2338 }
2339
2340 //------------------------------Identity---------------------------------------
2341 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2342 // Also feed through the klass in Allocate(...klass...)._klass.
2343 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2344 return klass_identity_common(phase);
2345 }
2346
2347 const Type* GetStoragePropertyNode::Value(PhaseGVN* phase) const {
2348 if (in(1) != NULL) {
2349 const Type* in1_t = phase->type(in(1));
2350 if (in1_t == Type::TOP) {
2351 return Type::TOP;
2352 }
2353 const TypeKlassPtr* tk = in1_t->make_ptr()->is_klassptr();
2354 ciArrayKlass* ak = tk->klass()->as_array_klass();
2355 ciKlass* elem = ak->element_klass();
2356 if (tk->klass_is_exact() || (!elem->is_java_lang_Object() && !elem->is_interface() && !elem->is_valuetype())) {
2357 int props_shift = in1_t->isa_narrowklass() ? oopDesc::narrow_storage_props_shift : oopDesc::wide_storage_props_shift;
2358 ArrayStorageProperties props = ak->storage_properties();
2359 intptr_t storage_properties = 0;
2360 if ((Opcode() == Op_GetFlattenedProperty && props.is_flattened()) ||
2361 (Opcode() == Op_GetNullFreeProperty && props.is_null_free())){
2362 storage_properties = 1;
2363 }
2364 if (in1_t->isa_narrowklass()) {
2365 return TypeInt::make((int)storage_properties);
2366 }
2367 return TypeX::make(storage_properties);
2368 }
2369 }
2370 return bottom_type();
2371 }
2372
2373 Node* GetStoragePropertyNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2374 if (!can_reshape) {
2375 return NULL;
2376 }
2377 if (in(1) != NULL && in(1)->is_Phi()) {
2378 Node* phi = in(1);
2379 Node* r = phi->in(0);
2380 Node* new_phi = new PhiNode(r, bottom_type());
2381 for (uint i = 1; i < r->req(); i++) {
2382 Node* in = phi->in(i);
2383 if (in == NULL) continue;
2384 Node* n = clone();
2385 n->set_req(1, in);
2386 new_phi->init_req(i, phase->transform(n));
2387 }
2388 return new_phi;
2389 }
2390 return NULL;
2391 }
2392
2393 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2394 Node* x = LoadNode::Identity(phase);
2395 if (x != this) return x;
2396
2397 // Take apart the address into an oop and and offset.
2398 // Return 'this' if we cannot.
2399 Node* adr = in(MemNode::Address);
2400 intptr_t offset = 0;
2401 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2402 if (base == NULL) return this;
2403 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2404 if (toop == NULL) return this;
2405
2406 // Step over potential GC barrier for OopHandle resolve
2407 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2408 if (bs->is_gc_barrier_node(base)) {
2409 base = bs->step_over_gc_barrier(base);
2410 }
2411
2412 // We can fetch the klass directly through an AllocateNode.
2413 // This works even if the klass is not constant (clone or newArray).
2414 if (offset == oopDesc::klass_offset_in_bytes()) {
2415 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2416 if (allocated_klass != NULL) {
2417 return allocated_klass;
2418 }
2419 }
2420
2421 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2422 // See inline_native_Class_query for occurrences of these patterns.
2423 // Java Example: x.getClass().isAssignableFrom(y)
2424 //
2425 // This improves reflective code, often making the Class
2426 // mirror go completely dead. (Current exception: Class
2427 // mirrors may appear in debug info, but we could clean them out by
2428 // introducing a new debug info operator for Klass.java_mirror).
2429
2430 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2431 && offset == java_lang_Class::klass_offset_in_bytes()) {
2432 if (base->is_Load()) {
2433 Node* base2 = base->in(MemNode::Address);
2434 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2435 Node* adr2 = base2->in(MemNode::Address);
2436 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2437 if (tkls != NULL && !tkls->empty()
2438 && (tkls->klass()->is_instance_klass() ||
2439 tkls->klass()->is_array_klass())
2440 && adr2->is_AddP()
2441 ) {
2442 int mirror_field = in_bytes(Klass::java_mirror_offset());
2443 if (tkls->offset() == mirror_field) {
2444 return adr2->in(AddPNode::Base);
2445 }
2446 }
2447 }
2448 }
2449 }
2450
2451 return this;
2452 }
2453
2454
2455 //------------------------------Value------------------------------------------
2456 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2457 const Type *t = klass_value_common(phase);
2458 if (t == Type::TOP)
2459 return t;
2460
2461 return t->make_narrowklass();
2462 }
2463
2464 //------------------------------Identity---------------------------------------
2465 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2466 // Also feed through the klass in Allocate(...klass...)._klass.
2467 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2468 Node *x = klass_identity_common(phase);
2469
2470 const Type *t = phase->type( x );
2471 if( t == Type::TOP ) return x;
2472 if( t->isa_narrowklass()) return x;
2473 assert (!t->isa_narrowoop(), "no narrow oop here");
2474
2475 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2476 }
2477
2478 //------------------------------Value-----------------------------------------
2479 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2480 // Either input is TOP ==> the result is TOP
2481 const Type *t1 = phase->type( in(MemNode::Memory) );
2482 if( t1 == Type::TOP ) return Type::TOP;
2483 Node *adr = in(MemNode::Address);
2484 const Type *t2 = phase->type( adr );
2485 if( t2 == Type::TOP ) return Type::TOP;
2486 const TypePtr *tp = t2->is_ptr();
2487 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2488 const TypeAryPtr *tap = tp->isa_aryptr();
2489 if( !tap ) return _type;
2490 return tap->size();
2491 }
2492
2493 //-------------------------------Ideal---------------------------------------
2494 // Feed through the length in AllocateArray(...length...)._length.
2495 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2496 Node* p = MemNode::Ideal_common(phase, can_reshape);
2497 if (p) return (p == NodeSentinel) ? NULL : p;
2498
2499 // Take apart the address into an oop and and offset.
2500 // Return 'this' if we cannot.
2501 Node* adr = in(MemNode::Address);
2502 intptr_t offset = 0;
2503 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2504 if (base == NULL) return NULL;
2505 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2506 if (tary == NULL) return NULL;
2507
2508 // We can fetch the length directly through an AllocateArrayNode.
2509 // This works even if the length is not constant (clone or newArray).
2510 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2511 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2512 if (alloc != NULL) {
2513 Node* allocated_length = alloc->Ideal_length();
2514 Node* len = alloc->make_ideal_length(tary, phase);
2515 if (allocated_length != len) {
2516 // New CastII improves on this.
2517 return len;
2518 }
2519 }
2520 }
2521
2522 return NULL;
2523 }
2524
2525 //------------------------------Identity---------------------------------------
2526 // Feed through the length in AllocateArray(...length...)._length.
2527 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2528 Node* x = LoadINode::Identity(phase);
2529 if (x != this) return x;
2530
2531 // Take apart the address into an oop and and offset.
2532 // Return 'this' if we cannot.
2533 Node* adr = in(MemNode::Address);
2534 intptr_t offset = 0;
2535 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2536 if (base == NULL) return this;
2537 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2538 if (tary == NULL) return this;
2539
2540 // We can fetch the length directly through an AllocateArrayNode.
2541 // This works even if the length is not constant (clone or newArray).
2542 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2543 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2544 if (alloc != NULL) {
2545 Node* allocated_length = alloc->Ideal_length();
2546 // Do not allow make_ideal_length to allocate a CastII node.
2547 Node* len = alloc->make_ideal_length(tary, phase, false);
2548 if (allocated_length == len) {
2549 // Return allocated_length only if it would not be improved by a CastII.
2550 return allocated_length;
2551 }
2552 }
2553 }
2554
2555 return this;
2556
2557 }
2558
2559 //=============================================================================
2560 //---------------------------StoreNode::make-----------------------------------
2561 // Polymorphic factory method:
2562 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2563 assert((mo == unordered || mo == release), "unexpected");
2564 Compile* C = gvn.C;
2565 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2566 ctl != NULL, "raw memory operations should have control edge");
2567
2568 switch (bt) {
2569 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2570 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2571 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2572 case T_CHAR:
2573 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2574 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2575 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2576 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2577 case T_METADATA:
2578 case T_ADDRESS:
2579 case T_VALUETYPE:
2580 case T_OBJECT:
2581 #ifdef _LP64
2582 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2583 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2584 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2585 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2586 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2587 adr->bottom_type()->isa_rawptr())) {
2588 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2589 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2590 }
2591 #endif
2592 {
2593 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2594 }
2595 default:
2596 ShouldNotReachHere();
2597 return (StoreNode*)NULL;
2598 }
2599 }
2600
2601 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2602 bool require_atomic = true;
2603 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2604 }
2605
2606 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2607 bool require_atomic = true;
2608 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2609 }
2610
2611
2612 //--------------------------bottom_type----------------------------------------
2613 const Type *StoreNode::bottom_type() const {
2614 return Type::MEMORY;
2615 }
2616
2617 //------------------------------hash-------------------------------------------
2618 uint StoreNode::hash() const {
2619 // unroll addition of interesting fields
2620 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2621
2622 // Since they are not commoned, do not hash them:
2623 return NO_HASH;
2624 }
2625
2626 //------------------------------Ideal------------------------------------------
2627 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2628 // When a store immediately follows a relevant allocation/initialization,
2629 // try to capture it into the initialization, or hoist it above.
2630 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2631 Node* p = MemNode::Ideal_common(phase, can_reshape);
2632 if (p) return (p == NodeSentinel) ? NULL : p;
2633
2634 Node* mem = in(MemNode::Memory);
2635 Node* address = in(MemNode::Address);
2636 // Back-to-back stores to same address? Fold em up. Generally
2637 // unsafe if I have intervening uses... Also disallowed for StoreCM
2638 // since they must follow each StoreP operation. Redundant StoreCMs
2639 // are eliminated just before matching in final_graph_reshape.
2640 if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::VALUES) {
2641 Node* st = mem;
2642 // If Store 'st' has more than one use, we cannot fold 'st' away.
2643 // For example, 'st' might be the final state at a conditional
2644 // return. Or, 'st' might be used by some node which is live at
2645 // the same time 'st' is live, which might be unschedulable. So,
2646 // require exactly ONE user until such time as we clone 'mem' for
2647 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2648 // true).
2649 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2650 // Looking at a dead closed cycle of memory?
2651 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2652 assert(Opcode() == st->Opcode() ||
2653 st->Opcode() == Op_StoreVector ||
2654 Opcode() == Op_StoreVector ||
2655 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2656 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2657 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2658 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
2659 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2660 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2661
2662 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2663 st->as_Store()->memory_size() <= this->memory_size()) {
2664 Node* use = st->raw_out(0);
2665 phase->igvn_rehash_node_delayed(use);
2666 if (can_reshape) {
2667 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2668 } else {
2669 // It's OK to do this in the parser, since DU info is always accurate,
2670 // and the parser always refers to nodes via SafePointNode maps.
2671 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2672 }
2673 return this;
2674 }
2675 st = st->in(MemNode::Memory);
2676 }
2677 }
2678
2679
2680 // Capture an unaliased, unconditional, simple store into an initializer.
2681 // Or, if it is independent of the allocation, hoist it above the allocation.
2682 if (ReduceFieldZeroing && /*can_reshape &&*/
2683 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2684 InitializeNode* init = mem->in(0)->as_Initialize();
2685 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2686 if (offset > 0) {
2687 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2688 // If the InitializeNode captured me, it made a raw copy of me,
2689 // and I need to disappear.
2690 if (moved != NULL) {
2691 // %%% hack to ensure that Ideal returns a new node:
2692 mem = MergeMemNode::make(mem);
2693 return mem; // fold me away
2694 }
2695 }
2696 }
2697
2698 return NULL; // No further progress
2699 }
2700
2701 //------------------------------Value-----------------------------------------
2702 const Type* StoreNode::Value(PhaseGVN* phase) const {
2703 // Either input is TOP ==> the result is TOP
2704 const Type *t1 = phase->type( in(MemNode::Memory) );
2705 if( t1 == Type::TOP ) return Type::TOP;
2706 const Type *t2 = phase->type( in(MemNode::Address) );
2707 if( t2 == Type::TOP ) return Type::TOP;
2708 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2709 if( t3 == Type::TOP ) return Type::TOP;
2710 return Type::MEMORY;
2711 }
2712
2713 //------------------------------Identity---------------------------------------
2714 // Remove redundant stores:
2715 // Store(m, p, Load(m, p)) changes to m.
2716 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2717 Node* StoreNode::Identity(PhaseGVN* phase) {
2718 Node* mem = in(MemNode::Memory);
2719 Node* adr = in(MemNode::Address);
2720 Node* val = in(MemNode::ValueIn);
2721
2722 Node* result = this;
2723
2724 // Load then Store? Then the Store is useless
2725 if (val->is_Load() &&
2726 val->in(MemNode::Address)->eqv_uncast(adr) &&
2727 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2728 val->as_Load()->store_Opcode() == Opcode()) {
2729 result = mem;
2730 }
2731
2732 // Two stores in a row of the same value?
2733 if (result == this &&
2734 mem->is_Store() &&
2735 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2736 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2737 mem->Opcode() == Opcode()) {
2738 result = mem;
2739 }
2740
2741 // Store of zero anywhere into a freshly-allocated object?
2742 // Then the store is useless.
2743 // (It must already have been captured by the InitializeNode.)
2744 if (result == this && ReduceFieldZeroing) {
2745 // a newly allocated object is already all-zeroes everywhere
2746 if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
2747 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) {
2748 assert(!phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == NULL, "storing null to value array is forbidden");
2749 result = mem;
2750 }
2751
2752 if (result == this) {
2753 // the store may also apply to zero-bits in an earlier object
2754 Node* prev_mem = find_previous_store(phase);
2755 // Steps (a), (b): Walk past independent stores to find an exact match.
2756 if (prev_mem != NULL) {
2757 Node* prev_val = can_see_stored_value(prev_mem, phase);
2758 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2759 // prev_val and val might differ by a cast; it would be good
2760 // to keep the more informative of the two.
2761 if (phase->type(val)->is_zero_type()) {
2762 result = mem;
2763 } else if (prev_mem->is_Proj() && prev_mem->in(0)->is_Initialize()) {
2764 InitializeNode* init = prev_mem->in(0)->as_Initialize();
2765 AllocateNode* alloc = init->allocation();
2766 if (alloc != NULL && alloc->in(AllocateNode::DefaultValue) == val) {
2767 result = mem;
2768 }
2769 }
2770 }
2771 }
2772 }
2773 }
2774
2775 if (result != this && phase->is_IterGVN() != NULL) {
2776 MemBarNode* trailing = trailing_membar();
2777 if (trailing != NULL) {
2778 #ifdef ASSERT
2779 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2780 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2781 #endif
2782 PhaseIterGVN* igvn = phase->is_IterGVN();
2783 trailing->remove(igvn);
2784 }
2785 }
2786
2787 return result;
2788 }
2789
2790 //------------------------------match_edge-------------------------------------
2791 // Do we Match on this edge index or not? Match only memory & value
2792 uint StoreNode::match_edge(uint idx) const {
2793 return idx == MemNode::Address || idx == MemNode::ValueIn;
2794 }
2795
2796 //------------------------------cmp--------------------------------------------
2797 // Do not common stores up together. They generally have to be split
2798 // back up anyways, so do not bother.
2799 bool StoreNode::cmp( const Node &n ) const {
2800 return (&n == this); // Always fail except on self
2801 }
2802
2803 //------------------------------Ideal_masked_input-----------------------------
2804 // Check for a useless mask before a partial-word store
2805 // (StoreB ... (AndI valIn conIa) )
2806 // If (conIa & mask == mask) this simplifies to
2807 // (StoreB ... (valIn) )
2808 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2809 Node *val = in(MemNode::ValueIn);
2810 if( val->Opcode() == Op_AndI ) {
2811 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2812 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2813 set_req(MemNode::ValueIn, val->in(1));
2814 return this;
2815 }
2816 }
2817 return NULL;
2818 }
2819
2820
2821 //------------------------------Ideal_sign_extended_input----------------------
2822 // Check for useless sign-extension before a partial-word store
2823 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2824 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2825 // (StoreB ... (valIn) )
2826 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2827 Node *val = in(MemNode::ValueIn);
2828 if( val->Opcode() == Op_RShiftI ) {
2829 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2830 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2831 Node *shl = val->in(1);
2832 if( shl->Opcode() == Op_LShiftI ) {
2833 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2834 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2835 set_req(MemNode::ValueIn, shl->in(1));
2836 return this;
2837 }
2838 }
2839 }
2840 }
2841 return NULL;
2842 }
2843
2844 //------------------------------value_never_loaded-----------------------------------
2845 // Determine whether there are any possible loads of the value stored.
2846 // For simplicity, we actually check if there are any loads from the
2847 // address stored to, not just for loads of the value stored by this node.
2848 //
2849 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2850 Node *adr = in(Address);
2851 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2852 if (adr_oop == NULL)
2853 return false;
2854 if (!adr_oop->is_known_instance_field())
2855 return false; // if not a distinct instance, there may be aliases of the address
2856 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2857 Node *use = adr->fast_out(i);
2858 if (use->is_Load() || use->is_LoadStore()) {
2859 return false;
2860 }
2861 }
2862 return true;
2863 }
2864
2865 MemBarNode* StoreNode::trailing_membar() const {
2866 if (is_release()) {
2867 MemBarNode* trailing_mb = NULL;
2868 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2869 Node* u = fast_out(i);
2870 if (u->is_MemBar()) {
2871 if (u->as_MemBar()->trailing_store()) {
2872 assert(u->Opcode() == Op_MemBarVolatile, "");
2873 assert(trailing_mb == NULL, "only one");
2874 trailing_mb = u->as_MemBar();
2875 #ifdef ASSERT
2876 Node* leading = u->as_MemBar()->leading_membar();
2877 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2878 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2879 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2880 #endif
2881 } else {
2882 assert(u->as_MemBar()->standalone(), "");
2883 }
2884 }
2885 }
2886 return trailing_mb;
2887 }
2888 return NULL;
2889 }
2890
2891
2892 //=============================================================================
2893 //------------------------------Ideal------------------------------------------
2894 // If the store is from an AND mask that leaves the low bits untouched, then
2895 // we can skip the AND operation. If the store is from a sign-extension
2896 // (a left shift, then right shift) we can skip both.
2897 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2898 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2899 if( progress != NULL ) return progress;
2900
2901 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2902 if( progress != NULL ) return progress;
2903
2904 // Finally check the default case
2905 return StoreNode::Ideal(phase, can_reshape);
2906 }
2907
2908 //=============================================================================
2909 //------------------------------Ideal------------------------------------------
2910 // If the store is from an AND mask that leaves the low bits untouched, then
2911 // we can skip the AND operation
2912 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2913 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2914 if( progress != NULL ) return progress;
2915
2916 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2917 if( progress != NULL ) return progress;
2918
2919 // Finally check the default case
2920 return StoreNode::Ideal(phase, can_reshape);
2921 }
2922
2923 //=============================================================================
2924 //------------------------------Identity---------------------------------------
2925 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2926 // No need to card mark when storing a null ptr
2927 Node* my_store = in(MemNode::OopStore);
2928 if (my_store->is_Store()) {
2929 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2930 if( t1 == TypePtr::NULL_PTR ) {
2931 return in(MemNode::Memory);
2932 }
2933 }
2934 return this;
2935 }
2936
2937 //=============================================================================
2938 //------------------------------Ideal---------------------------------------
2939 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2940 Node* progress = StoreNode::Ideal(phase, can_reshape);
2941 if (progress != NULL) return progress;
2942
2943 Node* my_store = in(MemNode::OopStore);
2944 if (my_store->is_MergeMem()) {
2945 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2946 set_req(MemNode::OopStore, mem);
2947 return this;
2948 }
2949
2950 return NULL;
2951 }
2952
2953 //------------------------------Value-----------------------------------------
2954 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2955 // Either input is TOP ==> the result is TOP
2956 const Type *t = phase->type( in(MemNode::Memory) );
2957 if( t == Type::TOP ) return Type::TOP;
2958 t = phase->type( in(MemNode::Address) );
2959 if( t == Type::TOP ) return Type::TOP;
2960 t = phase->type( in(MemNode::ValueIn) );
2961 if( t == Type::TOP ) return Type::TOP;
2962 // If extra input is TOP ==> the result is TOP
2963 t = phase->type( in(MemNode::OopStore) );
2964 if( t == Type::TOP ) return Type::TOP;
2965
2966 return StoreNode::Value( phase );
2967 }
2968
2969
2970 //=============================================================================
2971 //----------------------------------SCMemProjNode------------------------------
2972 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2973 {
2974 return bottom_type();
2975 }
2976
2977 //=============================================================================
2978 //----------------------------------LoadStoreNode------------------------------
2979 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2980 : Node(required),
2981 _type(rt),
2982 _adr_type(at),
2983 _has_barrier(false)
2984 {
2985 init_req(MemNode::Control, c );
2986 init_req(MemNode::Memory , mem);
2987 init_req(MemNode::Address, adr);
2988 init_req(MemNode::ValueIn, val);
2989 init_class_id(Class_LoadStore);
2990 }
2991
2992 uint LoadStoreNode::ideal_reg() const {
2993 return _type->ideal_reg();
2994 }
2995
2996 bool LoadStoreNode::result_not_used() const {
2997 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2998 Node *x = fast_out(i);
2999 if (x->Opcode() == Op_SCMemProj) continue;
3000 return false;
3001 }
3002 return true;
3003 }
3004
3005 MemBarNode* LoadStoreNode::trailing_membar() const {
3006 MemBarNode* trailing = NULL;
3007 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3008 Node* u = fast_out(i);
3009 if (u->is_MemBar()) {
3010 if (u->as_MemBar()->trailing_load_store()) {
3011 assert(u->Opcode() == Op_MemBarAcquire, "");
3012 assert(trailing == NULL, "only one");
3013 trailing = u->as_MemBar();
3014 #ifdef ASSERT
3015 Node* leading = trailing->leading_membar();
3016 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3017 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3018 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3019 #endif
3020 } else {
3021 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3022 }
3023 }
3024 }
3025
3026 return trailing;
3027 }
3028
3029 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3030
3031 //=============================================================================
3032 //----------------------------------LoadStoreConditionalNode--------------------
3033 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
3034 init_req(ExpectedIn, ex );
3035 }
3036
3037 //=============================================================================
3038 //-------------------------------adr_type--------------------------------------
3039 const TypePtr* ClearArrayNode::adr_type() const {
3040 Node *adr = in(3);
3041 if (adr == NULL) return NULL; // node is dead
3042 return MemNode::calculate_adr_type(adr->bottom_type());
3043 }
3044
3045 //------------------------------match_edge-------------------------------------
3046 // Do we Match on this edge index or not? Do not match memory
3047 uint ClearArrayNode::match_edge(uint idx) const {
3048 return idx > 1;
3049 }
3050
3051 //------------------------------Identity---------------------------------------
3052 // Clearing a zero length array does nothing
3053 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3054 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
3055 }
3056
3057 //------------------------------Idealize---------------------------------------
3058 // Clearing a short array is faster with stores
3059 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3060 // Already know this is a large node, do not try to ideal it
3061 if (!IdealizeClearArrayNode || _is_large) return NULL;
3062
3063 const int unit = BytesPerLong;
3064 const TypeX* t = phase->type(in(2))->isa_intptr_t();
3065 if (!t) return NULL;
3066 if (!t->is_con()) return NULL;
3067 intptr_t raw_count = t->get_con();
3068 intptr_t size = raw_count;
3069 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3070 // Clearing nothing uses the Identity call.
3071 // Negative clears are possible on dead ClearArrays
3072 // (see jck test stmt114.stmt11402.val).
3073 if (size <= 0 || size % unit != 0) return NULL;
3074 intptr_t count = size / unit;
3075 // Length too long; communicate this to matchers and assemblers.
3076 // Assemblers are responsible to produce fast hardware clears for it.
3077 if (size > InitArrayShortSize) {
3078 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
3079 }
3080 Node *mem = in(1);
3081 if( phase->type(mem)==Type::TOP ) return NULL;
3082 Node *adr = in(3);
3083 const Type* at = phase->type(adr);
3084 if( at==Type::TOP ) return NULL;
3085 const TypePtr* atp = at->isa_ptr();
3086 // adjust atp to be the correct array element address type
3087 if (atp == NULL) atp = TypePtr::BOTTOM;
3088 else atp = atp->add_offset(Type::OffsetBot);
3089 // Get base for derived pointer purposes
3090 if( adr->Opcode() != Op_AddP ) Unimplemented();
3091 Node *base = adr->in(1);
3092
3093 Node *val = in(4);
3094 Node *off = phase->MakeConX(BytesPerLong);
3095 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
3096 count--;
3097 while( count-- ) {
3098 mem = phase->transform(mem);
3099 adr = phase->transform(new AddPNode(base,adr,off));
3100 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
3101 }
3102 return mem;
3103 }
3104
3105 //----------------------------step_through----------------------------------
3106 // Return allocation input memory edge if it is different instance
3107 // or itself if it is the one we are looking for.
3108 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
3109 Node* n = *np;
3110 assert(n->is_ClearArray(), "sanity");
3111 intptr_t offset;
3112 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3113 // This method is called only before Allocate nodes are expanded
3114 // during macro nodes expansion. Before that ClearArray nodes are
3115 // only generated in PhaseMacroExpand::generate_arraycopy() (before
3116 // Allocate nodes are expanded) which follows allocations.
3117 assert(alloc != NULL, "should have allocation");
3118 if (alloc->_idx == instance_id) {
3119 // Can not bypass initialization of the instance we are looking for.
3120 return false;
3121 }
3122 // Otherwise skip it.
3123 InitializeNode* init = alloc->initialization();
3124 if (init != NULL)
3125 *np = init->in(TypeFunc::Memory);
3126 else
3127 *np = alloc->in(TypeFunc::Memory);
3128 return true;
3129 }
3130
3131 //----------------------------clear_memory-------------------------------------
3132 // Generate code to initialize object storage to zero.
3133 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3134 Node* val,
3135 Node* raw_val,
3136 intptr_t start_offset,
3137 Node* end_offset,
3138 PhaseGVN* phase) {
3139 intptr_t offset = start_offset;
3140
3141 int unit = BytesPerLong;
3142 if ((offset % unit) != 0) {
3143 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3144 adr = phase->transform(adr);
3145 const TypePtr* atp = TypeRawPtr::BOTTOM;
3146 if (val != NULL) {
3147 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3148 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3149 } else {
3150 assert(raw_val == NULL, "val may not be null");
3151 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3152 }
3153 mem = phase->transform(mem);
3154 offset += BytesPerInt;
3155 }
3156 assert((offset % unit) == 0, "");
3157
3158 // Initialize the remaining stuff, if any, with a ClearArray.
3159 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
3160 }
3161
3162 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3163 Node* raw_val,
3164 Node* start_offset,
3165 Node* end_offset,
3166 PhaseGVN* phase) {
3167 if (start_offset == end_offset) {
3168 // nothing to do
3169 return mem;
3170 }
3171
3172 int unit = BytesPerLong;
3173 Node* zbase = start_offset;
3174 Node* zend = end_offset;
3175
3176 // Scale to the unit required by the CPU:
3177 if (!Matcher::init_array_count_is_in_bytes) {
3178 Node* shift = phase->intcon(exact_log2(unit));
3179 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3180 zend = phase->transform(new URShiftXNode(zend, shift) );
3181 }
3182
3183 // Bulk clear double-words
3184 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3185 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3186 if (raw_val == NULL) {
3187 raw_val = phase->MakeConX(0);
3188 }
3189 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
3190 return phase->transform(mem);
3191 }
3192
3193 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3194 Node* val,
3195 Node* raw_val,
3196 intptr_t start_offset,
3197 intptr_t end_offset,
3198 PhaseGVN* phase) {
3199 if (start_offset == end_offset) {
3200 // nothing to do
3201 return mem;
3202 }
3203
3204 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3205 intptr_t done_offset = end_offset;
3206 if ((done_offset % BytesPerLong) != 0) {
3207 done_offset -= BytesPerInt;
3208 }
3209 if (done_offset > start_offset) {
3210 mem = clear_memory(ctl, mem, dest, val, raw_val,
3211 start_offset, phase->MakeConX(done_offset), phase);
3212 }
3213 if (done_offset < end_offset) { // emit the final 32-bit store
3214 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3215 adr = phase->transform(adr);
3216 const TypePtr* atp = TypeRawPtr::BOTTOM;
3217 if (val != NULL) {
3218 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3219 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3220 } else {
3221 assert(raw_val == NULL, "val may not be null");
3222 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3223 }
3224 mem = phase->transform(mem);
3225 done_offset += BytesPerInt;
3226 }
3227 assert(done_offset == end_offset, "");
3228 return mem;
3229 }
3230
3231 //=============================================================================
3232 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3233 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3234 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3235 #ifdef ASSERT
3236 , _pair_idx(0)
3237 #endif
3238 {
3239 init_class_id(Class_MemBar);
3240 Node* top = C->top();
3241 init_req(TypeFunc::I_O,top);
3242 init_req(TypeFunc::FramePtr,top);
3243 init_req(TypeFunc::ReturnAdr,top);
3244 if (precedent != NULL)
3245 init_req(TypeFunc::Parms, precedent);
3246 }
3247
3248 //------------------------------cmp--------------------------------------------
3249 uint MemBarNode::hash() const { return NO_HASH; }
3250 bool MemBarNode::cmp( const Node &n ) const {
3251 return (&n == this); // Always fail except on self
3252 }
3253
3254 //------------------------------make-------------------------------------------
3255 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3256 switch (opcode) {
3257 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3258 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3259 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3260 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3261 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3262 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3263 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3264 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3265 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3266 case Op_Initialize: return new InitializeNode(C, atp, pn);
3267 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3268 default: ShouldNotReachHere(); return NULL;
3269 }
3270 }
3271
3272 void MemBarNode::remove(PhaseIterGVN *igvn) {
3273 if (outcnt() != 2) {
3274 return;
3275 }
3276 if (trailing_store() || trailing_load_store()) {
3277 MemBarNode* leading = leading_membar();
3278 if (leading != NULL) {
3279 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3280 leading->remove(igvn);
3281 }
3282 }
3283 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3284 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3285 }
3286
3287 //------------------------------Ideal------------------------------------------
3288 // Return a node which is more "ideal" than the current node. Strip out
3289 // control copies
3290 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3291 if (remove_dead_region(phase, can_reshape)) return this;
3292 // Don't bother trying to transform a dead node
3293 if (in(0) && in(0)->is_top()) {
3294 return NULL;
3295 }
3296
3297 bool progress = false;
3298 // Eliminate volatile MemBars for scalar replaced objects.
3299 if (can_reshape && req() == (Precedent+1)) {
3300 bool eliminate = false;
3301 int opc = Opcode();
3302 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3303 // Volatile field loads and stores.
3304 Node* my_mem = in(MemBarNode::Precedent);
3305 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3306 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3307 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3308 // replace this Precedent (decodeN) with the Load instead.
3309 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3310 Node* load_node = my_mem->in(1);
3311 set_req(MemBarNode::Precedent, load_node);
3312 phase->is_IterGVN()->_worklist.push(my_mem);
3313 my_mem = load_node;
3314 } else {
3315 assert(my_mem->unique_out() == this, "sanity");
3316 del_req(Precedent);
3317 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3318 my_mem = NULL;
3319 }
3320 progress = true;
3321 }
3322 if (my_mem != NULL && my_mem->is_Mem()) {
3323 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3324 // Check for scalar replaced object reference.
3325 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3326 t_oop->offset() != Type::OffsetBot &&
3327 t_oop->offset() != Type::OffsetTop) {
3328 eliminate = true;
3329 }
3330 }
3331 } else if (opc == Op_MemBarRelease) {
3332 // Final field stores.
3333 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3334 if ((alloc != NULL) && alloc->is_Allocate() &&
3335 alloc->as_Allocate()->does_not_escape_thread()) {
3336 // The allocated object does not escape.
3337 eliminate = true;
3338 }
3339 }
3340 if (eliminate) {
3341 // Replace MemBar projections by its inputs.
3342 PhaseIterGVN* igvn = phase->is_IterGVN();
3343 remove(igvn);
3344 // Must return either the original node (now dead) or a new node
3345 // (Do not return a top here, since that would break the uniqueness of top.)
3346 return new ConINode(TypeInt::ZERO);
3347 }
3348 }
3349 return progress ? this : NULL;
3350 }
3351
3352 //------------------------------Value------------------------------------------
3353 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3354 if( !in(0) ) return Type::TOP;
3355 if( phase->type(in(0)) == Type::TOP )
3356 return Type::TOP;
3357 return TypeTuple::MEMBAR;
3358 }
3359
3360 //------------------------------match------------------------------------------
3361 // Construct projections for memory.
3362 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
3363 switch (proj->_con) {
3364 case TypeFunc::Control:
3365 case TypeFunc::Memory:
3366 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3367 }
3368 ShouldNotReachHere();
3369 return NULL;
3370 }
3371
3372 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3373 trailing->_kind = TrailingStore;
3374 leading->_kind = LeadingStore;
3375 #ifdef ASSERT
3376 trailing->_pair_idx = leading->_idx;
3377 leading->_pair_idx = leading->_idx;
3378 #endif
3379 }
3380
3381 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3382 trailing->_kind = TrailingLoadStore;
3383 leading->_kind = LeadingLoadStore;
3384 #ifdef ASSERT
3385 trailing->_pair_idx = leading->_idx;
3386 leading->_pair_idx = leading->_idx;
3387 #endif
3388 }
3389
3390 MemBarNode* MemBarNode::trailing_membar() const {
3391 ResourceMark rm;
3392 Node* trailing = (Node*)this;
3393 VectorSet seen(Thread::current()->resource_area());
3394 Node_Stack multis(0);
3395 do {
3396 Node* c = trailing;
3397 uint i = 0;
3398 do {
3399 trailing = NULL;
3400 for (; i < c->outcnt(); i++) {
3401 Node* next = c->raw_out(i);
3402 if (next != c && next->is_CFG()) {
3403 if (c->is_MultiBranch()) {
3404 if (multis.node() == c) {
3405 multis.set_index(i+1);
3406 } else {
3407 multis.push(c, i+1);
3408 }
3409 }
3410 trailing = next;
3411 break;
3412 }
3413 }
3414 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3415 break;
3416 }
3417 while (multis.size() > 0) {
3418 c = multis.node();
3419 i = multis.index();
3420 if (i < c->req()) {
3421 break;
3422 }
3423 multis.pop();
3424 }
3425 } while (multis.size() > 0);
3426 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3427
3428 MemBarNode* mb = trailing->as_MemBar();
3429 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3430 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3431 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3432 return mb;
3433 }
3434
3435 MemBarNode* MemBarNode::leading_membar() const {
3436 ResourceMark rm;
3437 VectorSet seen(Thread::current()->resource_area());
3438 Node_Stack regions(0);
3439 Node* leading = in(0);
3440 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3441 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3442 leading = NULL;
3443 while (regions.size() > 0 && leading == NULL) {
3444 Node* r = regions.node();
3445 uint i = regions.index();
3446 if (i < r->req()) {
3447 leading = r->in(i);
3448 regions.set_index(i+1);
3449 } else {
3450 regions.pop();
3451 }
3452 }
3453 if (leading == NULL) {
3454 assert(regions.size() == 0, "all paths should have been tried");
3455 return NULL;
3456 }
3457 }
3458 if (leading->is_Region()) {
3459 regions.push(leading, 2);
3460 leading = leading->in(1);
3461 } else {
3462 leading = leading->in(0);
3463 }
3464 }
3465 #ifdef ASSERT
3466 Unique_Node_List wq;
3467 wq.push((Node*)this);
3468 uint found = 0;
3469 for (uint i = 0; i < wq.size(); i++) {
3470 Node* n = wq.at(i);
3471 if (n->is_Region()) {
3472 for (uint j = 1; j < n->req(); j++) {
3473 Node* in = n->in(j);
3474 if (in != NULL && !in->is_top()) {
3475 wq.push(in);
3476 }
3477 }
3478 } else {
3479 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3480 assert(n == leading, "consistency check failed");
3481 found++;
3482 } else {
3483 Node* in = n->in(0);
3484 if (in != NULL && !in->is_top()) {
3485 wq.push(in);
3486 }
3487 }
3488 }
3489 }
3490 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3491 #endif
3492 if (leading == NULL) {
3493 return NULL;
3494 }
3495 MemBarNode* mb = leading->as_MemBar();
3496 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3497 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3498 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3499 return mb;
3500 }
3501
3502 //===========================InitializeNode====================================
3503 // SUMMARY:
3504 // This node acts as a memory barrier on raw memory, after some raw stores.
3505 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3506 // The Initialize can 'capture' suitably constrained stores as raw inits.
3507 // It can coalesce related raw stores into larger units (called 'tiles').
3508 // It can avoid zeroing new storage for memory units which have raw inits.
3509 // At macro-expansion, it is marked 'complete', and does not optimize further.
3510 //
3511 // EXAMPLE:
3512 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3513 // ctl = incoming control; mem* = incoming memory
3514 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3515 // First allocate uninitialized memory and fill in the header:
3516 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3517 // ctl := alloc.Control; mem* := alloc.Memory*
3518 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3519 // Then initialize to zero the non-header parts of the raw memory block:
3520 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3521 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3522 // After the initialize node executes, the object is ready for service:
3523 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3524 // Suppose its body is immediately initialized as {1,2}:
3525 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3526 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3527 // mem.SLICE(#short[*]) := store2
3528 //
3529 // DETAILS:
3530 // An InitializeNode collects and isolates object initialization after
3531 // an AllocateNode and before the next possible safepoint. As a
3532 // memory barrier (MemBarNode), it keeps critical stores from drifting
3533 // down past any safepoint or any publication of the allocation.
3534 // Before this barrier, a newly-allocated object may have uninitialized bits.
3535 // After this barrier, it may be treated as a real oop, and GC is allowed.
3536 //
3537 // The semantics of the InitializeNode include an implicit zeroing of
3538 // the new object from object header to the end of the object.
3539 // (The object header and end are determined by the AllocateNode.)
3540 //
3541 // Certain stores may be added as direct inputs to the InitializeNode.
3542 // These stores must update raw memory, and they must be to addresses
3543 // derived from the raw address produced by AllocateNode, and with
3544 // a constant offset. They must be ordered by increasing offset.
3545 // The first one is at in(RawStores), the last at in(req()-1).
3546 // Unlike most memory operations, they are not linked in a chain,
3547 // but are displayed in parallel as users of the rawmem output of
3548 // the allocation.
3549 //
3550 // (See comments in InitializeNode::capture_store, which continue
3551 // the example given above.)
3552 //
3553 // When the associated Allocate is macro-expanded, the InitializeNode
3554 // may be rewritten to optimize collected stores. A ClearArrayNode
3555 // may also be created at that point to represent any required zeroing.
3556 // The InitializeNode is then marked 'complete', prohibiting further
3557 // capturing of nearby memory operations.
3558 //
3559 // During macro-expansion, all captured initializations which store
3560 // constant values of 32 bits or smaller are coalesced (if advantageous)
3561 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3562 // initialized in fewer memory operations. Memory words which are
3563 // covered by neither tiles nor non-constant stores are pre-zeroed
3564 // by explicit stores of zero. (The code shape happens to do all
3565 // zeroing first, then all other stores, with both sequences occurring
3566 // in order of ascending offsets.)
3567 //
3568 // Alternatively, code may be inserted between an AllocateNode and its
3569 // InitializeNode, to perform arbitrary initialization of the new object.
3570 // E.g., the object copying intrinsics insert complex data transfers here.
3571 // The initialization must then be marked as 'complete' disable the
3572 // built-in zeroing semantics and the collection of initializing stores.
3573 //
3574 // While an InitializeNode is incomplete, reads from the memory state
3575 // produced by it are optimizable if they match the control edge and
3576 // new oop address associated with the allocation/initialization.
3577 // They return a stored value (if the offset matches) or else zero.
3578 // A write to the memory state, if it matches control and address,
3579 // and if it is to a constant offset, may be 'captured' by the
3580 // InitializeNode. It is cloned as a raw memory operation and rewired
3581 // inside the initialization, to the raw oop produced by the allocation.
3582 // Operations on addresses which are provably distinct (e.g., to
3583 // other AllocateNodes) are allowed to bypass the initialization.
3584 //
3585 // The effect of all this is to consolidate object initialization
3586 // (both arrays and non-arrays, both piecewise and bulk) into a
3587 // single location, where it can be optimized as a unit.
3588 //
3589 // Only stores with an offset less than TrackedInitializationLimit words
3590 // will be considered for capture by an InitializeNode. This puts a
3591 // reasonable limit on the complexity of optimized initializations.
3592
3593 //---------------------------InitializeNode------------------------------------
3594 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3595 : MemBarNode(C, adr_type, rawoop),
3596 _is_complete(Incomplete), _does_not_escape(false)
3597 {
3598 init_class_id(Class_Initialize);
3599
3600 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3601 assert(in(RawAddress) == rawoop, "proper init");
3602 // Note: allocation() can be NULL, for secondary initialization barriers
3603 }
3604
3605 // Since this node is not matched, it will be processed by the
3606 // register allocator. Declare that there are no constraints
3607 // on the allocation of the RawAddress edge.
3608 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3609 // This edge should be set to top, by the set_complete. But be conservative.
3610 if (idx == InitializeNode::RawAddress)
3611 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3612 return RegMask::Empty;
3613 }
3614
3615 Node* InitializeNode::memory(uint alias_idx) {
3616 Node* mem = in(Memory);
3617 if (mem->is_MergeMem()) {
3618 return mem->as_MergeMem()->memory_at(alias_idx);
3619 } else {
3620 // incoming raw memory is not split
3621 return mem;
3622 }
3623 }
3624
3625 bool InitializeNode::is_non_zero() {
3626 if (is_complete()) return false;
3627 remove_extra_zeroes();
3628 return (req() > RawStores);
3629 }
3630
3631 void InitializeNode::set_complete(PhaseGVN* phase) {
3632 assert(!is_complete(), "caller responsibility");
3633 _is_complete = Complete;
3634
3635 // After this node is complete, it contains a bunch of
3636 // raw-memory initializations. There is no need for
3637 // it to have anything to do with non-raw memory effects.
3638 // Therefore, tell all non-raw users to re-optimize themselves,
3639 // after skipping the memory effects of this initialization.
3640 PhaseIterGVN* igvn = phase->is_IterGVN();
3641 if (igvn) igvn->add_users_to_worklist(this);
3642 }
3643
3644 // convenience function
3645 // return false if the init contains any stores already
3646 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3647 InitializeNode* init = initialization();
3648 if (init == NULL || init->is_complete()) {
3649 return false;
3650 }
3651 init->remove_extra_zeroes();
3652 // for now, if this allocation has already collected any inits, bail:
3653 if (init->is_non_zero()) return false;
3654 init->set_complete(phase);
3655 return true;
3656 }
3657
3658 void InitializeNode::remove_extra_zeroes() {
3659 if (req() == RawStores) return;
3660 Node* zmem = zero_memory();
3661 uint fill = RawStores;
3662 for (uint i = fill; i < req(); i++) {
3663 Node* n = in(i);
3664 if (n->is_top() || n == zmem) continue; // skip
3665 if (fill < i) set_req(fill, n); // compact
3666 ++fill;
3667 }
3668 // delete any empty spaces created:
3669 while (fill < req()) {
3670 del_req(fill);
3671 }
3672 }
3673
3674 // Helper for remembering which stores go with which offsets.
3675 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3676 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3677 intptr_t offset = -1;
3678 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3679 phase, offset);
3680 if (base == NULL) return -1; // something is dead,
3681 if (offset < 0) return -1; // dead, dead
3682 return offset;
3683 }
3684
3685 // Helper for proving that an initialization expression is
3686 // "simple enough" to be folded into an object initialization.
3687 // Attempts to prove that a store's initial value 'n' can be captured
3688 // within the initialization without creating a vicious cycle, such as:
3689 // { Foo p = new Foo(); p.next = p; }
3690 // True for constants and parameters and small combinations thereof.
3691 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3692 if (n == NULL) return true; // (can this really happen?)
3693 if (n->is_Proj()) n = n->in(0);
3694 if (n == this) return false; // found a cycle
3695 if (n->is_Con()) return true;
3696 if (n->is_Start()) return true; // params, etc., are OK
3697 if (n->is_Root()) return true; // even better
3698
3699 Node* ctl = n->in(0);
3700 if (ctl != NULL && !ctl->is_top()) {
3701 if (ctl->is_Proj()) ctl = ctl->in(0);
3702 if (ctl == this) return false;
3703
3704 // If we already know that the enclosing memory op is pinned right after
3705 // the init, then any control flow that the store has picked up
3706 // must have preceded the init, or else be equal to the init.
3707 // Even after loop optimizations (which might change control edges)
3708 // a store is never pinned *before* the availability of its inputs.
3709 if (!MemNode::all_controls_dominate(n, this))
3710 return false; // failed to prove a good control
3711 }
3712
3713 // Check data edges for possible dependencies on 'this'.
3714 if ((count += 1) > 20) return false; // complexity limit
3715 for (uint i = 1; i < n->req(); i++) {
3716 Node* m = n->in(i);
3717 if (m == NULL || m == n || m->is_top()) continue;
3718 uint first_i = n->find_edge(m);
3719 if (i != first_i) continue; // process duplicate edge just once
3720 if (!detect_init_independence(m, count)) {
3721 return false;
3722 }
3723 }
3724
3725 return true;
3726 }
3727
3728 // Here are all the checks a Store must pass before it can be moved into
3729 // an initialization. Returns zero if a check fails.
3730 // On success, returns the (constant) offset to which the store applies,
3731 // within the initialized memory.
3732 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3733 const int FAIL = 0;
3734 if (st->req() != MemNode::ValueIn + 1)
3735 return FAIL; // an inscrutable StoreNode (card mark?)
3736 Node* ctl = st->in(MemNode::Control);
3737 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3738 return FAIL; // must be unconditional after the initialization
3739 Node* mem = st->in(MemNode::Memory);
3740 if (!(mem->is_Proj() && mem->in(0) == this))
3741 return FAIL; // must not be preceded by other stores
3742 Node* adr = st->in(MemNode::Address);
3743 intptr_t offset;
3744 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3745 if (alloc == NULL)
3746 return FAIL; // inscrutable address
3747 if (alloc != allocation())
3748 return FAIL; // wrong allocation! (store needs to float up)
3749 int size_in_bytes = st->memory_size();
3750 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3751 return FAIL; // mismatched access
3752 }
3753 Node* val = st->in(MemNode::ValueIn);
3754 int complexity_count = 0;
3755 if (!detect_init_independence(val, complexity_count))
3756 return FAIL; // stored value must be 'simple enough'
3757
3758 // The Store can be captured only if nothing after the allocation
3759 // and before the Store is using the memory location that the store
3760 // overwrites.
3761 bool failed = false;
3762 // If is_complete_with_arraycopy() is true the shape of the graph is
3763 // well defined and is safe so no need for extra checks.
3764 if (!is_complete_with_arraycopy()) {
3765 // We are going to look at each use of the memory state following
3766 // the allocation to make sure nothing reads the memory that the
3767 // Store writes.
3768 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3769 int alias_idx = phase->C->get_alias_index(t_adr);
3770 ResourceMark rm;
3771 Unique_Node_List mems;
3772 mems.push(mem);
3773 Node* unique_merge = NULL;
3774 for (uint next = 0; next < mems.size(); ++next) {
3775 Node *m = mems.at(next);
3776 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3777 Node *n = m->fast_out(j);
3778 if (n->outcnt() == 0) {
3779 continue;
3780 }
3781 if (n == st) {
3782 continue;
3783 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3784 // If the control of this use is different from the control
3785 // of the Store which is right after the InitializeNode then
3786 // this node cannot be between the InitializeNode and the
3787 // Store.
3788 continue;
3789 } else if (n->is_MergeMem()) {
3790 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3791 // We can hit a MergeMemNode (that will likely go away
3792 // later) that is a direct use of the memory state
3793 // following the InitializeNode on the same slice as the
3794 // store node that we'd like to capture. We need to check
3795 // the uses of the MergeMemNode.
3796 mems.push(n);
3797 }
3798 } else if (n->is_Mem()) {
3799 Node* other_adr = n->in(MemNode::Address);
3800 if (other_adr == adr) {
3801 failed = true;
3802 break;
3803 } else {
3804 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3805 if (other_t_adr != NULL) {
3806 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3807 if (other_alias_idx == alias_idx) {
3808 // A load from the same memory slice as the store right
3809 // after the InitializeNode. We check the control of the
3810 // object/array that is loaded from. If it's the same as
3811 // the store control then we cannot capture the store.
3812 assert(!n->is_Store(), "2 stores to same slice on same control?");
3813 Node* base = other_adr;
3814 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3815 base = base->in(AddPNode::Base);
3816 if (base != NULL) {
3817 base = base->uncast();
3818 if (base->is_Proj() && base->in(0) == alloc) {
3819 failed = true;
3820 break;
3821 }
3822 }
3823 }
3824 }
3825 }
3826 } else {
3827 failed = true;
3828 break;
3829 }
3830 }
3831 }
3832 }
3833 if (failed) {
3834 if (!can_reshape) {
3835 // We decided we couldn't capture the store during parsing. We
3836 // should try again during the next IGVN once the graph is
3837 // cleaner.
3838 phase->C->record_for_igvn(st);
3839 }
3840 return FAIL;
3841 }
3842
3843 return offset; // success
3844 }
3845
3846 // Find the captured store in(i) which corresponds to the range
3847 // [start..start+size) in the initialized object.
3848 // If there is one, return its index i. If there isn't, return the
3849 // negative of the index where it should be inserted.
3850 // Return 0 if the queried range overlaps an initialization boundary
3851 // or if dead code is encountered.
3852 // If size_in_bytes is zero, do not bother with overlap checks.
3853 int InitializeNode::captured_store_insertion_point(intptr_t start,
3854 int size_in_bytes,
3855 PhaseTransform* phase) {
3856 const int FAIL = 0, MAX_STORE = BytesPerLong;
3857
3858 if (is_complete())
3859 return FAIL; // arraycopy got here first; punt
3860
3861 assert(allocation() != NULL, "must be present");
3862
3863 // no negatives, no header fields:
3864 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3865
3866 // after a certain size, we bail out on tracking all the stores:
3867 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3868 if (start >= ti_limit) return FAIL;
3869
3870 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3871 if (i >= limit) return -(int)i; // not found; here is where to put it
3872
3873 Node* st = in(i);
3874 intptr_t st_off = get_store_offset(st, phase);
3875 if (st_off < 0) {
3876 if (st != zero_memory()) {
3877 return FAIL; // bail out if there is dead garbage
3878 }
3879 } else if (st_off > start) {
3880 // ...we are done, since stores are ordered
3881 if (st_off < start + size_in_bytes) {
3882 return FAIL; // the next store overlaps
3883 }
3884 return -(int)i; // not found; here is where to put it
3885 } else if (st_off < start) {
3886 if (size_in_bytes != 0 &&
3887 start < st_off + MAX_STORE &&
3888 start < st_off + st->as_Store()->memory_size()) {
3889 return FAIL; // the previous store overlaps
3890 }
3891 } else {
3892 if (size_in_bytes != 0 &&
3893 st->as_Store()->memory_size() != size_in_bytes) {
3894 return FAIL; // mismatched store size
3895 }
3896 return i;
3897 }
3898
3899 ++i;
3900 }
3901 }
3902
3903 // Look for a captured store which initializes at the offset 'start'
3904 // with the given size. If there is no such store, and no other
3905 // initialization interferes, then return zero_memory (the memory
3906 // projection of the AllocateNode).
3907 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3908 PhaseTransform* phase) {
3909 assert(stores_are_sane(phase), "");
3910 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3911 if (i == 0) {
3912 return NULL; // something is dead
3913 } else if (i < 0) {
3914 return zero_memory(); // just primordial zero bits here
3915 } else {
3916 Node* st = in(i); // here is the store at this position
3917 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3918 return st;
3919 }
3920 }
3921
3922 // Create, as a raw pointer, an address within my new object at 'offset'.
3923 Node* InitializeNode::make_raw_address(intptr_t offset,
3924 PhaseTransform* phase) {
3925 Node* addr = in(RawAddress);
3926 if (offset != 0) {
3927 Compile* C = phase->C;
3928 addr = phase->transform( new AddPNode(C->top(), addr,
3929 phase->MakeConX(offset)) );
3930 }
3931 return addr;
3932 }
3933
3934 // Clone the given store, converting it into a raw store
3935 // initializing a field or element of my new object.
3936 // Caller is responsible for retiring the original store,
3937 // with subsume_node or the like.
3938 //
3939 // From the example above InitializeNode::InitializeNode,
3940 // here are the old stores to be captured:
3941 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3942 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3943 //
3944 // Here is the changed code; note the extra edges on init:
3945 // alloc = (Allocate ...)
3946 // rawoop = alloc.RawAddress
3947 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3948 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3949 // init = (Initialize alloc.Control alloc.Memory rawoop
3950 // rawstore1 rawstore2)
3951 //
3952 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3953 PhaseTransform* phase, bool can_reshape) {
3954 assert(stores_are_sane(phase), "");
3955
3956 if (start < 0) return NULL;
3957 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3958
3959 Compile* C = phase->C;
3960 int size_in_bytes = st->memory_size();
3961 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3962 if (i == 0) return NULL; // bail out
3963 Node* prev_mem = NULL; // raw memory for the captured store
3964 if (i > 0) {
3965 prev_mem = in(i); // there is a pre-existing store under this one
3966 set_req(i, C->top()); // temporarily disconnect it
3967 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3968 } else {
3969 i = -i; // no pre-existing store
3970 prev_mem = zero_memory(); // a slice of the newly allocated object
3971 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3972 set_req(--i, C->top()); // reuse this edge; it has been folded away
3973 else
3974 ins_req(i, C->top()); // build a new edge
3975 }
3976 Node* new_st = st->clone();
3977 new_st->set_req(MemNode::Control, in(Control));
3978 new_st->set_req(MemNode::Memory, prev_mem);
3979 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3980 new_st = phase->transform(new_st);
3981
3982 // At this point, new_st might have swallowed a pre-existing store
3983 // at the same offset, or perhaps new_st might have disappeared,
3984 // if it redundantly stored the same value (or zero to fresh memory).
3985
3986 // In any case, wire it in:
3987 phase->igvn_rehash_node_delayed(this);
3988 set_req(i, new_st);
3989
3990 // The caller may now kill the old guy.
3991 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3992 assert(check_st == new_st || check_st == NULL, "must be findable");
3993 assert(!is_complete(), "");
3994 return new_st;
3995 }
3996
3997 static bool store_constant(jlong* tiles, int num_tiles,
3998 intptr_t st_off, int st_size,
3999 jlong con) {
4000 if ((st_off & (st_size-1)) != 0)
4001 return false; // strange store offset (assume size==2**N)
4002 address addr = (address)tiles + st_off;
4003 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4004 switch (st_size) {
4005 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
4006 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
4007 case sizeof(jint): *(jint*) addr = (jint) con; break;
4008 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
4009 default: return false; // strange store size (detect size!=2**N here)
4010 }
4011 return true; // return success to caller
4012 }
4013
4014 // Coalesce subword constants into int constants and possibly
4015 // into long constants. The goal, if the CPU permits,
4016 // is to initialize the object with a small number of 64-bit tiles.
4017 // Also, convert floating-point constants to bit patterns.
4018 // Non-constants are not relevant to this pass.
4019 //
4020 // In terms of the running example on InitializeNode::InitializeNode
4021 // and InitializeNode::capture_store, here is the transformation
4022 // of rawstore1 and rawstore2 into rawstore12:
4023 // alloc = (Allocate ...)
4024 // rawoop = alloc.RawAddress
4025 // tile12 = 0x00010002
4026 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4027 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4028 //
4029 void
4030 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4031 Node* size_in_bytes,
4032 PhaseGVN* phase) {
4033 Compile* C = phase->C;
4034
4035 assert(stores_are_sane(phase), "");
4036 // Note: After this pass, they are not completely sane,
4037 // since there may be some overlaps.
4038
4039 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4040
4041 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4042 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4043 size_limit = MIN2(size_limit, ti_limit);
4044 size_limit = align_up(size_limit, BytesPerLong);
4045 int num_tiles = size_limit / BytesPerLong;
4046
4047 // allocate space for the tile map:
4048 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4049 jlong tiles_buf[small_len];
4050 Node* nodes_buf[small_len];
4051 jlong inits_buf[small_len];
4052 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4053 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4054 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4055 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4056 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4057 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4058 // tiles: exact bitwise model of all primitive constants
4059 // nodes: last constant-storing node subsumed into the tiles model
4060 // inits: which bytes (in each tile) are touched by any initializations
4061
4062 //// Pass A: Fill in the tile model with any relevant stores.
4063
4064 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4065 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4066 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4067 Node* zmem = zero_memory(); // initially zero memory state
4068 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4069 Node* st = in(i);
4070 intptr_t st_off = get_store_offset(st, phase);
4071
4072 // Figure out the store's offset and constant value:
4073 if (st_off < header_size) continue; //skip (ignore header)
4074 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
4075 int st_size = st->as_Store()->memory_size();
4076 if (st_off + st_size > size_limit) break;
4077
4078 // Record which bytes are touched, whether by constant or not.
4079 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4080 continue; // skip (strange store size)
4081
4082 const Type* val = phase->type(st->in(MemNode::ValueIn));
4083 if (!val->singleton()) continue; //skip (non-con store)
4084 BasicType type = val->basic_type();
4085
4086 jlong con = 0;
4087 switch (type) {
4088 case T_INT: con = val->is_int()->get_con(); break;
4089 case T_LONG: con = val->is_long()->get_con(); break;
4090 case T_FLOAT: con = jint_cast(val->getf()); break;
4091 case T_DOUBLE: con = jlong_cast(val->getd()); break;
4092 default: continue; //skip (odd store type)
4093 }
4094
4095 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4096 st->Opcode() == Op_StoreL) {
4097 continue; // This StoreL is already optimal.
4098 }
4099
4100 // Store down the constant.
4101 store_constant(tiles, num_tiles, st_off, st_size, con);
4102
4103 intptr_t j = st_off >> LogBytesPerLong;
4104
4105 if (type == T_INT && st_size == BytesPerInt
4106 && (st_off & BytesPerInt) == BytesPerInt) {
4107 jlong lcon = tiles[j];
4108 if (!Matcher::isSimpleConstant64(lcon) &&
4109 st->Opcode() == Op_StoreI) {
4110 // This StoreI is already optimal by itself.
4111 jint* intcon = (jint*) &tiles[j];
4112 intcon[1] = 0; // undo the store_constant()
4113
4114 // If the previous store is also optimal by itself, back up and
4115 // undo the action of the previous loop iteration... if we can.
4116 // But if we can't, just let the previous half take care of itself.
4117 st = nodes[j];
4118 st_off -= BytesPerInt;
4119 con = intcon[0];
4120 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
4121 assert(st_off >= header_size, "still ignoring header");
4122 assert(get_store_offset(st, phase) == st_off, "must be");
4123 assert(in(i-1) == zmem, "must be");
4124 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4125 assert(con == tcon->is_int()->get_con(), "must be");
4126 // Undo the effects of the previous loop trip, which swallowed st:
4127 intcon[0] = 0; // undo store_constant()
4128 set_req(i-1, st); // undo set_req(i, zmem)
4129 nodes[j] = NULL; // undo nodes[j] = st
4130 --old_subword; // undo ++old_subword
4131 }
4132 continue; // This StoreI is already optimal.
4133 }
4134 }
4135
4136 // This store is not needed.
4137 set_req(i, zmem);
4138 nodes[j] = st; // record for the moment
4139 if (st_size < BytesPerLong) // something has changed
4140 ++old_subword; // includes int/float, but who's counting...
4141 else ++old_long;
4142 }
4143
4144 if ((old_subword + old_long) == 0)
4145 return; // nothing more to do
4146
4147 //// Pass B: Convert any non-zero tiles into optimal constant stores.
4148 // Be sure to insert them before overlapping non-constant stores.
4149 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
4150 for (int j = 0; j < num_tiles; j++) {
4151 jlong con = tiles[j];
4152 jlong init = inits[j];
4153 if (con == 0) continue;
4154 jint con0, con1; // split the constant, address-wise
4155 jint init0, init1; // split the init map, address-wise
4156 { union { jlong con; jint intcon[2]; } u;
4157 u.con = con;
4158 con0 = u.intcon[0];
4159 con1 = u.intcon[1];
4160 u.con = init;
4161 init0 = u.intcon[0];
4162 init1 = u.intcon[1];
4163 }
4164
4165 Node* old = nodes[j];
4166 assert(old != NULL, "need the prior store");
4167 intptr_t offset = (j * BytesPerLong);
4168
4169 bool split = !Matcher::isSimpleConstant64(con);
4170
4171 if (offset < header_size) {
4172 assert(offset + BytesPerInt >= header_size, "second int counts");
4173 assert(*(jint*)&tiles[j] == 0, "junk in header");
4174 split = true; // only the second word counts
4175 // Example: int a[] = { 42 ... }
4176 } else if (con0 == 0 && init0 == -1) {
4177 split = true; // first word is covered by full inits
4178 // Example: int a[] = { ... foo(), 42 ... }
4179 } else if (con1 == 0 && init1 == -1) {
4180 split = true; // second word is covered by full inits
4181 // Example: int a[] = { ... 42, foo() ... }
4182 }
4183
4184 // Here's a case where init0 is neither 0 nor -1:
4185 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4186 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4187 // In this case the tile is not split; it is (jlong)42.
4188 // The big tile is stored down, and then the foo() value is inserted.
4189 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4190
4191 Node* ctl = old->in(MemNode::Control);
4192 Node* adr = make_raw_address(offset, phase);
4193 const TypePtr* atp = TypeRawPtr::BOTTOM;
4194
4195 // One or two coalesced stores to plop down.
4196 Node* st[2];
4197 intptr_t off[2];
4198 int nst = 0;
4199 if (!split) {
4200 ++new_long;
4201 off[nst] = offset;
4202 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4203 phase->longcon(con), T_LONG, MemNode::unordered);
4204 } else {
4205 // Omit either if it is a zero.
4206 if (con0 != 0) {
4207 ++new_int;
4208 off[nst] = offset;
4209 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4210 phase->intcon(con0), T_INT, MemNode::unordered);
4211 }
4212 if (con1 != 0) {
4213 ++new_int;
4214 offset += BytesPerInt;
4215 adr = make_raw_address(offset, phase);
4216 off[nst] = offset;
4217 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4218 phase->intcon(con1), T_INT, MemNode::unordered);
4219 }
4220 }
4221
4222 // Insert second store first, then the first before the second.
4223 // Insert each one just before any overlapping non-constant stores.
4224 while (nst > 0) {
4225 Node* st1 = st[--nst];
4226 C->copy_node_notes_to(st1, old);
4227 st1 = phase->transform(st1);
4228 offset = off[nst];
4229 assert(offset >= header_size, "do not smash header");
4230 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4231 guarantee(ins_idx != 0, "must re-insert constant store");
4232 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4233 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4234 set_req(--ins_idx, st1);
4235 else
4236 ins_req(ins_idx, st1);
4237 }
4238 }
4239
4240 if (PrintCompilation && WizardMode)
4241 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4242 old_subword, old_long, new_int, new_long);
4243 if (C->log() != NULL)
4244 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4245 old_subword, old_long, new_int, new_long);
4246
4247 // Clean up any remaining occurrences of zmem:
4248 remove_extra_zeroes();
4249 }
4250
4251 // Explore forward from in(start) to find the first fully initialized
4252 // word, and return its offset. Skip groups of subword stores which
4253 // together initialize full words. If in(start) is itself part of a
4254 // fully initialized word, return the offset of in(start). If there
4255 // are no following full-word stores, or if something is fishy, return
4256 // a negative value.
4257 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4258 int int_map = 0;
4259 intptr_t int_map_off = 0;
4260 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4261
4262 for (uint i = start, limit = req(); i < limit; i++) {
4263 Node* st = in(i);
4264
4265 intptr_t st_off = get_store_offset(st, phase);
4266 if (st_off < 0) break; // return conservative answer
4267
4268 int st_size = st->as_Store()->memory_size();
4269 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4270 return st_off; // we found a complete word init
4271 }
4272
4273 // update the map:
4274
4275 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4276 if (this_int_off != int_map_off) {
4277 // reset the map:
4278 int_map = 0;
4279 int_map_off = this_int_off;
4280 }
4281
4282 int subword_off = st_off - this_int_off;
4283 int_map |= right_n_bits(st_size) << subword_off;
4284 if ((int_map & FULL_MAP) == FULL_MAP) {
4285 return this_int_off; // we found a complete word init
4286 }
4287
4288 // Did this store hit or cross the word boundary?
4289 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4290 if (next_int_off == this_int_off + BytesPerInt) {
4291 // We passed the current int, without fully initializing it.
4292 int_map_off = next_int_off;
4293 int_map >>= BytesPerInt;
4294 } else if (next_int_off > this_int_off + BytesPerInt) {
4295 // We passed the current and next int.
4296 return this_int_off + BytesPerInt;
4297 }
4298 }
4299
4300 return -1;
4301 }
4302
4303
4304 // Called when the associated AllocateNode is expanded into CFG.
4305 // At this point, we may perform additional optimizations.
4306 // Linearize the stores by ascending offset, to make memory
4307 // activity as coherent as possible.
4308 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4309 intptr_t header_size,
4310 Node* size_in_bytes,
4311 PhaseGVN* phase) {
4312 assert(!is_complete(), "not already complete");
4313 assert(stores_are_sane(phase), "");
4314 assert(allocation() != NULL, "must be present");
4315
4316 remove_extra_zeroes();
4317
4318 if (ReduceFieldZeroing || ReduceBulkZeroing)
4319 // reduce instruction count for common initialization patterns
4320 coalesce_subword_stores(header_size, size_in_bytes, phase);
4321
4322 Node* zmem = zero_memory(); // initially zero memory state
4323 Node* inits = zmem; // accumulating a linearized chain of inits
4324 #ifdef ASSERT
4325 intptr_t first_offset = allocation()->minimum_header_size();
4326 intptr_t last_init_off = first_offset; // previous init offset
4327 intptr_t last_init_end = first_offset; // previous init offset+size
4328 intptr_t last_tile_end = first_offset; // previous tile offset+size
4329 #endif
4330 intptr_t zeroes_done = header_size;
4331
4332 bool do_zeroing = true; // we might give up if inits are very sparse
4333 int big_init_gaps = 0; // how many large gaps have we seen?
4334
4335 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4336 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4337
4338 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4339 Node* st = in(i);
4340 intptr_t st_off = get_store_offset(st, phase);
4341 if (st_off < 0)
4342 break; // unknown junk in the inits
4343 if (st->in(MemNode::Memory) != zmem)
4344 break; // complicated store chains somehow in list
4345
4346 int st_size = st->as_Store()->memory_size();
4347 intptr_t next_init_off = st_off + st_size;
4348
4349 if (do_zeroing && zeroes_done < next_init_off) {
4350 // See if this store needs a zero before it or under it.
4351 intptr_t zeroes_needed = st_off;
4352
4353 if (st_size < BytesPerInt) {
4354 // Look for subword stores which only partially initialize words.
4355 // If we find some, we must lay down some word-level zeroes first,
4356 // underneath the subword stores.
4357 //
4358 // Examples:
4359 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4360 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4361 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4362 //
4363 // Note: coalesce_subword_stores may have already done this,
4364 // if it was prompted by constant non-zero subword initializers.
4365 // But this case can still arise with non-constant stores.
4366
4367 intptr_t next_full_store = find_next_fullword_store(i, phase);
4368
4369 // In the examples above:
4370 // in(i) p q r s x y z
4371 // st_off 12 13 14 15 12 13 14
4372 // st_size 1 1 1 1 1 1 1
4373 // next_full_s. 12 16 16 16 16 16 16
4374 // z's_done 12 16 16 16 12 16 12
4375 // z's_needed 12 16 16 16 16 16 16
4376 // zsize 0 0 0 0 4 0 4
4377 if (next_full_store < 0) {
4378 // Conservative tack: Zero to end of current word.
4379 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4380 } else {
4381 // Zero to beginning of next fully initialized word.
4382 // Or, don't zero at all, if we are already in that word.
4383 assert(next_full_store >= zeroes_needed, "must go forward");
4384 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4385 zeroes_needed = next_full_store;
4386 }
4387 }
4388
4389 if (zeroes_needed > zeroes_done) {
4390 intptr_t zsize = zeroes_needed - zeroes_done;
4391 // Do some incremental zeroing on rawmem, in parallel with inits.
4392 zeroes_done = align_down(zeroes_done, BytesPerInt);
4393 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4394 allocation()->in(AllocateNode::DefaultValue),
4395 allocation()->in(AllocateNode::RawDefaultValue),
4396 zeroes_done, zeroes_needed,
4397 phase);
4398 zeroes_done = zeroes_needed;
4399 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4400 do_zeroing = false; // leave the hole, next time
4401 }
4402 }
4403
4404 // Collect the store and move on:
4405 st->set_req(MemNode::Memory, inits);
4406 inits = st; // put it on the linearized chain
4407 set_req(i, zmem); // unhook from previous position
4408
4409 if (zeroes_done == st_off)
4410 zeroes_done = next_init_off;
4411
4412 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4413
4414 #ifdef ASSERT
4415 // Various order invariants. Weaker than stores_are_sane because
4416 // a large constant tile can be filled in by smaller non-constant stores.
4417 assert(st_off >= last_init_off, "inits do not reverse");
4418 last_init_off = st_off;
4419 const Type* val = NULL;
4420 if (st_size >= BytesPerInt &&
4421 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4422 (int)val->basic_type() < (int)T_OBJECT) {
4423 assert(st_off >= last_tile_end, "tiles do not overlap");
4424 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4425 last_tile_end = MAX2(last_tile_end, next_init_off);
4426 } else {
4427 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4428 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4429 assert(st_off >= last_init_end, "inits do not overlap");
4430 last_init_end = next_init_off; // it's a non-tile
4431 }
4432 #endif //ASSERT
4433 }
4434
4435 remove_extra_zeroes(); // clear out all the zmems left over
4436 add_req(inits);
4437
4438 if (!(UseTLAB && ZeroTLAB)) {
4439 // If anything remains to be zeroed, zero it all now.
4440 zeroes_done = align_down(zeroes_done, BytesPerInt);
4441 // if it is the last unused 4 bytes of an instance, forget about it
4442 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4443 if (zeroes_done + BytesPerLong >= size_limit) {
4444 AllocateNode* alloc = allocation();
4445 assert(alloc != NULL, "must be present");
4446 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4447 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4448 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4449 if (zeroes_done == k->layout_helper())
4450 zeroes_done = size_limit;
4451 }
4452 }
4453 if (zeroes_done < size_limit) {
4454 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4455 allocation()->in(AllocateNode::DefaultValue),
4456 allocation()->in(AllocateNode::RawDefaultValue),
4457 zeroes_done, size_in_bytes, phase);
4458 }
4459 }
4460
4461 set_complete(phase);
4462 return rawmem;
4463 }
4464
4465
4466 #ifdef ASSERT
4467 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4468 if (is_complete())
4469 return true; // stores could be anything at this point
4470 assert(allocation() != NULL, "must be present");
4471 intptr_t last_off = allocation()->minimum_header_size();
4472 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4473 Node* st = in(i);
4474 intptr_t st_off = get_store_offset(st, phase);
4475 if (st_off < 0) continue; // ignore dead garbage
4476 if (last_off > st_off) {
4477 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4478 this->dump(2);
4479 assert(false, "ascending store offsets");
4480 return false;
4481 }
4482 last_off = st_off + st->as_Store()->memory_size();
4483 }
4484 return true;
4485 }
4486 #endif //ASSERT
4487
4488
4489
4490
4491 //============================MergeMemNode=====================================
4492 //
4493 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4494 // contributing store or call operations. Each contributor provides the memory
4495 // state for a particular "alias type" (see Compile::alias_type). For example,
4496 // if a MergeMem has an input X for alias category #6, then any memory reference
4497 // to alias category #6 may use X as its memory state input, as an exact equivalent
4498 // to using the MergeMem as a whole.
4499 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4500 //
4501 // (Here, the <N> notation gives the index of the relevant adr_type.)
4502 //
4503 // In one special case (and more cases in the future), alias categories overlap.
4504 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4505 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4506 // it is exactly equivalent to that state W:
4507 // MergeMem(<Bot>: W) <==> W
4508 //
4509 // Usually, the merge has more than one input. In that case, where inputs
4510 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4511 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4512 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4513 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4514 //
4515 // A merge can take a "wide" memory state as one of its narrow inputs.
4516 // This simply means that the merge observes out only the relevant parts of
4517 // the wide input. That is, wide memory states arriving at narrow merge inputs
4518 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4519 //
4520 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4521 // and that memory slices "leak through":
4522 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4523 //
4524 // But, in such a cascade, repeated memory slices can "block the leak":
4525 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4526 //
4527 // In the last example, Y is not part of the combined memory state of the
4528 // outermost MergeMem. The system must, of course, prevent unschedulable
4529 // memory states from arising, so you can be sure that the state Y is somehow
4530 // a precursor to state Y'.
4531 //
4532 //
4533 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4534 // of each MergeMemNode array are exactly the numerical alias indexes, including
4535 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4536 // Compile::alias_type (and kin) produce and manage these indexes.
4537 //
4538 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4539 // (Note that this provides quick access to the top node inside MergeMem methods,
4540 // without the need to reach out via TLS to Compile::current.)
4541 //
4542 // As a consequence of what was just described, a MergeMem that represents a full
4543 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4544 // containing all alias categories.
4545 //
4546 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4547 //
4548 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4549 // a memory state for the alias type <N>, or else the top node, meaning that
4550 // there is no particular input for that alias type. Note that the length of
4551 // a MergeMem is variable, and may be extended at any time to accommodate new
4552 // memory states at larger alias indexes. When merges grow, they are of course
4553 // filled with "top" in the unused in() positions.
4554 //
4555 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4556 // (Top was chosen because it works smoothly with passes like GCM.)
4557 //
4558 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4559 // the type of random VM bits like TLS references.) Since it is always the
4560 // first non-Bot memory slice, some low-level loops use it to initialize an
4561 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4562 //
4563 //
4564 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4565 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4566 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4567 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4568 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4569 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4570 //
4571 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4572 // really that different from the other memory inputs. An abbreviation called
4573 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4574 //
4575 //
4576 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4577 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4578 // that "emerges though" the base memory will be marked as excluding the alias types
4579 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4580 //
4581 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4582 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4583 //
4584 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4585 // (It is currently unimplemented.) As you can see, the resulting merge is
4586 // actually a disjoint union of memory states, rather than an overlay.
4587 //
4588
4589 //------------------------------MergeMemNode-----------------------------------
4590 Node* MergeMemNode::make_empty_memory() {
4591 Node* empty_memory = (Node*) Compile::current()->top();
4592 assert(empty_memory->is_top(), "correct sentinel identity");
4593 return empty_memory;
4594 }
4595
4596 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4597 init_class_id(Class_MergeMem);
4598 // all inputs are nullified in Node::Node(int)
4599 // set_input(0, NULL); // no control input
4600
4601 // Initialize the edges uniformly to top, for starters.
4602 Node* empty_mem = make_empty_memory();
4603 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4604 init_req(i,empty_mem);
4605 }
4606 assert(empty_memory() == empty_mem, "");
4607
4608 if( new_base != NULL && new_base->is_MergeMem() ) {
4609 MergeMemNode* mdef = new_base->as_MergeMem();
4610 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4611 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4612 mms.set_memory(mms.memory2());
4613 }
4614 assert(base_memory() == mdef->base_memory(), "");
4615 } else {
4616 set_base_memory(new_base);
4617 }
4618 }
4619
4620 // Make a new, untransformed MergeMem with the same base as 'mem'.
4621 // If mem is itself a MergeMem, populate the result with the same edges.
4622 MergeMemNode* MergeMemNode::make(Node* mem) {
4623 return new MergeMemNode(mem);
4624 }
4625
4626 //------------------------------cmp--------------------------------------------
4627 uint MergeMemNode::hash() const { return NO_HASH; }
4628 bool MergeMemNode::cmp( const Node &n ) const {
4629 return (&n == this); // Always fail except on self
4630 }
4631
4632 //------------------------------Identity---------------------------------------
4633 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4634 // Identity if this merge point does not record any interesting memory
4635 // disambiguations.
4636 Node* base_mem = base_memory();
4637 Node* empty_mem = empty_memory();
4638 if (base_mem != empty_mem) { // Memory path is not dead?
4639 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4640 Node* mem = in(i);
4641 if (mem != empty_mem && mem != base_mem) {
4642 return this; // Many memory splits; no change
4643 }
4644 }
4645 }
4646 return base_mem; // No memory splits; ID on the one true input
4647 }
4648
4649 //------------------------------Ideal------------------------------------------
4650 // This method is invoked recursively on chains of MergeMem nodes
4651 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4652 // Remove chain'd MergeMems
4653 //
4654 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4655 // relative to the "in(Bot)". Since we are patching both at the same time,
4656 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4657 // but rewrite each "in(i)" relative to the new "in(Bot)".
4658 Node *progress = NULL;
4659
4660
4661 Node* old_base = base_memory();
4662 Node* empty_mem = empty_memory();
4663 if (old_base == empty_mem)
4664 return NULL; // Dead memory path.
4665
4666 MergeMemNode* old_mbase;
4667 if (old_base != NULL && old_base->is_MergeMem())
4668 old_mbase = old_base->as_MergeMem();
4669 else
4670 old_mbase = NULL;
4671 Node* new_base = old_base;
4672
4673 // simplify stacked MergeMems in base memory
4674 if (old_mbase) new_base = old_mbase->base_memory();
4675
4676 // the base memory might contribute new slices beyond my req()
4677 if (old_mbase) grow_to_match(old_mbase);
4678
4679 // Look carefully at the base node if it is a phi.
4680 PhiNode* phi_base;
4681 if (new_base != NULL && new_base->is_Phi())
4682 phi_base = new_base->as_Phi();
4683 else
4684 phi_base = NULL;
4685
4686 Node* phi_reg = NULL;
4687 uint phi_len = (uint)-1;
4688 if (phi_base != NULL && !phi_base->is_copy()) {
4689 // do not examine phi if degraded to a copy
4690 phi_reg = phi_base->region();
4691 phi_len = phi_base->req();
4692 // see if the phi is unfinished
4693 for (uint i = 1; i < phi_len; i++) {
4694 if (phi_base->in(i) == NULL) {
4695 // incomplete phi; do not look at it yet!
4696 phi_reg = NULL;
4697 phi_len = (uint)-1;
4698 break;
4699 }
4700 }
4701 }
4702
4703 // Note: We do not call verify_sparse on entry, because inputs
4704 // can normalize to the base_memory via subsume_node or similar
4705 // mechanisms. This method repairs that damage.
4706
4707 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4708
4709 // Look at each slice.
4710 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4711 Node* old_in = in(i);
4712 // calculate the old memory value
4713 Node* old_mem = old_in;
4714 if (old_mem == empty_mem) old_mem = old_base;
4715 assert(old_mem == memory_at(i), "");
4716
4717 // maybe update (reslice) the old memory value
4718
4719 // simplify stacked MergeMems
4720 Node* new_mem = old_mem;
4721 MergeMemNode* old_mmem;
4722 if (old_mem != NULL && old_mem->is_MergeMem())
4723 old_mmem = old_mem->as_MergeMem();
4724 else
4725 old_mmem = NULL;
4726 if (old_mmem == this) {
4727 // This can happen if loops break up and safepoints disappear.
4728 // A merge of BotPtr (default) with a RawPtr memory derived from a
4729 // safepoint can be rewritten to a merge of the same BotPtr with
4730 // the BotPtr phi coming into the loop. If that phi disappears
4731 // also, we can end up with a self-loop of the mergemem.
4732 // In general, if loops degenerate and memory effects disappear,
4733 // a mergemem can be left looking at itself. This simply means
4734 // that the mergemem's default should be used, since there is
4735 // no longer any apparent effect on this slice.
4736 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4737 // from start. Update the input to TOP.
4738 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4739 }
4740 else if (old_mmem != NULL) {
4741 new_mem = old_mmem->memory_at(i);
4742 }
4743 // else preceding memory was not a MergeMem
4744
4745 // replace equivalent phis (unfortunately, they do not GVN together)
4746 if (new_mem != NULL && new_mem != new_base &&
4747 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4748 if (new_mem->is_Phi()) {
4749 PhiNode* phi_mem = new_mem->as_Phi();
4750 for (uint i = 1; i < phi_len; i++) {
4751 if (phi_base->in(i) != phi_mem->in(i)) {
4752 phi_mem = NULL;
4753 break;
4754 }
4755 }
4756 if (phi_mem != NULL) {
4757 // equivalent phi nodes; revert to the def
4758 new_mem = new_base;
4759 }
4760 }
4761 }
4762
4763 // maybe store down a new value
4764 Node* new_in = new_mem;
4765 if (new_in == new_base) new_in = empty_mem;
4766
4767 if (new_in != old_in) {
4768 // Warning: Do not combine this "if" with the previous "if"
4769 // A memory slice might have be be rewritten even if it is semantically
4770 // unchanged, if the base_memory value has changed.
4771 set_req(i, new_in);
4772 progress = this; // Report progress
4773 }
4774 }
4775
4776 if (new_base != old_base) {
4777 set_req(Compile::AliasIdxBot, new_base);
4778 // Don't use set_base_memory(new_base), because we need to update du.
4779 assert(base_memory() == new_base, "");
4780 progress = this;
4781 }
4782
4783 if( base_memory() == this ) {
4784 // a self cycle indicates this memory path is dead
4785 set_req(Compile::AliasIdxBot, empty_mem);
4786 }
4787
4788 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4789 // Recursion must occur after the self cycle check above
4790 if( base_memory()->is_MergeMem() ) {
4791 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4792 Node *m = phase->transform(new_mbase); // Rollup any cycles
4793 if( m != NULL &&
4794 (m->is_top() ||
4795 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4796 // propagate rollup of dead cycle to self
4797 set_req(Compile::AliasIdxBot, empty_mem);
4798 }
4799 }
4800
4801 if( base_memory() == empty_mem ) {
4802 progress = this;
4803 // Cut inputs during Parse phase only.
4804 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4805 if( !can_reshape ) {
4806 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4807 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4808 }
4809 }
4810 }
4811
4812 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4813 // Check if PhiNode::Ideal's "Split phis through memory merges"
4814 // transform should be attempted. Look for this->phi->this cycle.
4815 uint merge_width = req();
4816 if (merge_width > Compile::AliasIdxRaw) {
4817 PhiNode* phi = base_memory()->as_Phi();
4818 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4819 if (phi->in(i) == this) {
4820 phase->is_IterGVN()->_worklist.push(phi);
4821 break;
4822 }
4823 }
4824 }
4825 }
4826
4827 assert(progress || verify_sparse(), "please, no dups of base");
4828 return progress;
4829 }
4830
4831 //-------------------------set_base_memory-------------------------------------
4832 void MergeMemNode::set_base_memory(Node *new_base) {
4833 Node* empty_mem = empty_memory();
4834 set_req(Compile::AliasIdxBot, new_base);
4835 assert(memory_at(req()) == new_base, "must set default memory");
4836 // Clear out other occurrences of new_base:
4837 if (new_base != empty_mem) {
4838 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4839 if (in(i) == new_base) set_req(i, empty_mem);
4840 }
4841 }
4842 }
4843
4844 //------------------------------out_RegMask------------------------------------
4845 const RegMask &MergeMemNode::out_RegMask() const {
4846 return RegMask::Empty;
4847 }
4848
4849 //------------------------------dump_spec--------------------------------------
4850 #ifndef PRODUCT
4851 void MergeMemNode::dump_spec(outputStream *st) const {
4852 st->print(" {");
4853 Node* base_mem = base_memory();
4854 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4855 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4856 if (mem == base_mem) { st->print(" -"); continue; }
4857 st->print( " N%d:", mem->_idx );
4858 Compile::current()->get_adr_type(i)->dump_on(st);
4859 }
4860 st->print(" }");
4861 }
4862 #endif // !PRODUCT
4863
4864
4865 #ifdef ASSERT
4866 static bool might_be_same(Node* a, Node* b) {
4867 if (a == b) return true;
4868 if (!(a->is_Phi() || b->is_Phi())) return false;
4869 // phis shift around during optimization
4870 return true; // pretty stupid...
4871 }
4872
4873 // verify a narrow slice (either incoming or outgoing)
4874 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4875 if (!VerifyAliases) return; // don't bother to verify unless requested
4876 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4877 if (Node::in_dump()) return; // muzzle asserts when printing
4878 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4879 assert(n != NULL, "");
4880 // Elide intervening MergeMem's
4881 while (n->is_MergeMem()) {
4882 n = n->as_MergeMem()->memory_at(alias_idx);
4883 }
4884 Compile* C = Compile::current();
4885 const TypePtr* n_adr_type = n->adr_type();
4886 if (n == m->empty_memory()) {
4887 // Implicit copy of base_memory()
4888 } else if (n_adr_type != TypePtr::BOTTOM) {
4889 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4890 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4891 } else {
4892 // A few places like make_runtime_call "know" that VM calls are narrow,
4893 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4894 bool expected_wide_mem = false;
4895 if (n == m->base_memory()) {
4896 expected_wide_mem = true;
4897 } else if (alias_idx == Compile::AliasIdxRaw ||
4898 n == m->memory_at(Compile::AliasIdxRaw)) {
4899 expected_wide_mem = true;
4900 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4901 // memory can "leak through" calls on channels that
4902 // are write-once. Allow this also.
4903 expected_wide_mem = true;
4904 }
4905 assert(expected_wide_mem, "expected narrow slice replacement");
4906 }
4907 }
4908 #else // !ASSERT
4909 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4910 #endif
4911
4912
4913 //-----------------------------memory_at---------------------------------------
4914 Node* MergeMemNode::memory_at(uint alias_idx) const {
4915 assert(alias_idx >= Compile::AliasIdxRaw ||
4916 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4917 "must avoid base_memory and AliasIdxTop");
4918
4919 // Otherwise, it is a narrow slice.
4920 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4921 Compile *C = Compile::current();
4922 if (is_empty_memory(n)) {
4923 // the array is sparse; empty slots are the "top" node
4924 n = base_memory();
4925 assert(Node::in_dump()
4926 || n == NULL || n->bottom_type() == Type::TOP
4927 || n->adr_type() == NULL // address is TOP
4928 || n->adr_type() == TypePtr::BOTTOM
4929 || n->adr_type() == TypeRawPtr::BOTTOM
4930 || Compile::current()->AliasLevel() == 0,
4931 "must be a wide memory");
4932 // AliasLevel == 0 if we are organizing the memory states manually.
4933 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4934 } else {
4935 // make sure the stored slice is sane
4936 #ifdef ASSERT
4937 if (VMError::is_error_reported() || Node::in_dump()) {
4938 } else if (might_be_same(n, base_memory())) {
4939 // Give it a pass: It is a mostly harmless repetition of the base.
4940 // This can arise normally from node subsumption during optimization.
4941 } else {
4942 verify_memory_slice(this, alias_idx, n);
4943 }
4944 #endif
4945 }
4946 return n;
4947 }
4948
4949 //---------------------------set_memory_at-------------------------------------
4950 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4951 verify_memory_slice(this, alias_idx, n);
4952 Node* empty_mem = empty_memory();
4953 if (n == base_memory()) n = empty_mem; // collapse default
4954 uint need_req = alias_idx+1;
4955 if (req() < need_req) {
4956 if (n == empty_mem) return; // already the default, so do not grow me
4957 // grow the sparse array
4958 do {
4959 add_req(empty_mem);
4960 } while (req() < need_req);
4961 }
4962 set_req( alias_idx, n );
4963 }
4964
4965
4966
4967 //--------------------------iteration_setup------------------------------------
4968 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4969 if (other != NULL) {
4970 grow_to_match(other);
4971 // invariant: the finite support of mm2 is within mm->req()
4972 #ifdef ASSERT
4973 for (uint i = req(); i < other->req(); i++) {
4974 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4975 }
4976 #endif
4977 }
4978 // Replace spurious copies of base_memory by top.
4979 Node* base_mem = base_memory();
4980 if (base_mem != NULL && !base_mem->is_top()) {
4981 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4982 if (in(i) == base_mem)
4983 set_req(i, empty_memory());
4984 }
4985 }
4986 }
4987
4988 //---------------------------grow_to_match-------------------------------------
4989 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4990 Node* empty_mem = empty_memory();
4991 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4992 // look for the finite support of the other memory
4993 for (uint i = other->req(); --i >= req(); ) {
4994 if (other->in(i) != empty_mem) {
4995 uint new_len = i+1;
4996 while (req() < new_len) add_req(empty_mem);
4997 break;
4998 }
4999 }
5000 }
5001
5002 //---------------------------verify_sparse-------------------------------------
5003 #ifndef PRODUCT
5004 bool MergeMemNode::verify_sparse() const {
5005 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
5006 Node* base_mem = base_memory();
5007 // The following can happen in degenerate cases, since empty==top.
5008 if (is_empty_memory(base_mem)) return true;
5009 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5010 assert(in(i) != NULL, "sane slice");
5011 if (in(i) == base_mem) return false; // should have been the sentinel value!
5012 }
5013 return true;
5014 }
5015
5016 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5017 Node* n;
5018 n = mm->in(idx);
5019 if (mem == n) return true; // might be empty_memory()
5020 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5021 if (mem == n) return true;
5022 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
5023 if (mem == n) return true;
5024 if (n == NULL) break;
5025 }
5026 return false;
5027 }
5028 #endif // !PRODUCT