Wed, 16 Jul 2008 16:04:39 -0700
6723160: Nightly failure: Error: meet not symmetric
Summary: Add missing _instance_id settings and other EA fixes.
Reviewed-by: rasbold
1 /*
2 * Copyright 1997-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 // Portions of code courtesy of Clifford Click
27 // Optimization - Graph Style
29 #include "incls/_precompiled.incl"
30 #include "incls/_memnode.cpp.incl"
32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
34 //=============================================================================
35 uint MemNode::size_of() const { return sizeof(*this); }
37 const TypePtr *MemNode::adr_type() const {
38 Node* adr = in(Address);
39 const TypePtr* cross_check = NULL;
40 DEBUG_ONLY(cross_check = _adr_type);
41 return calculate_adr_type(adr->bottom_type(), cross_check);
42 }
44 #ifndef PRODUCT
45 void MemNode::dump_spec(outputStream *st) const {
46 if (in(Address) == NULL) return; // node is dead
47 #ifndef ASSERT
48 // fake the missing field
49 const TypePtr* _adr_type = NULL;
50 if (in(Address) != NULL)
51 _adr_type = in(Address)->bottom_type()->isa_ptr();
52 #endif
53 dump_adr_type(this, _adr_type, st);
55 Compile* C = Compile::current();
56 if( C->alias_type(_adr_type)->is_volatile() )
57 st->print(" Volatile!");
58 }
60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
61 st->print(" @");
62 if (adr_type == NULL) {
63 st->print("NULL");
64 } else {
65 adr_type->dump_on(st);
66 Compile* C = Compile::current();
67 Compile::AliasType* atp = NULL;
68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
69 if (atp == NULL)
70 st->print(", idx=?\?;");
71 else if (atp->index() == Compile::AliasIdxBot)
72 st->print(", idx=Bot;");
73 else if (atp->index() == Compile::AliasIdxTop)
74 st->print(", idx=Top;");
75 else if (atp->index() == Compile::AliasIdxRaw)
76 st->print(", idx=Raw;");
77 else {
78 ciField* field = atp->field();
79 if (field) {
80 st->print(", name=");
81 field->print_name_on(st);
82 }
83 st->print(", idx=%d;", atp->index());
84 }
85 }
86 }
88 extern void print_alias_types();
90 #endif
92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
93 const TypeOopPtr *tinst = t_adr->isa_oopptr();
94 if (tinst == NULL || !tinst->is_known_instance_field())
95 return mchain; // don't try to optimize non-instance types
96 uint instance_id = tinst->instance_id();
97 Node *prev = NULL;
98 Node *result = mchain;
99 while (prev != result) {
100 prev = result;
101 // skip over a call which does not affect this memory slice
102 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
103 Node *proj_in = result->in(0);
104 if (proj_in->is_Call()) {
105 CallNode *call = proj_in->as_Call();
106 if (!call->may_modify(t_adr, phase)) {
107 result = call->in(TypeFunc::Memory);
108 }
109 } else if (proj_in->is_Initialize()) {
110 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
111 // Stop if this is the initialization for the object instance which
112 // which contains this memory slice, otherwise skip over it.
113 if (alloc != NULL && alloc->_idx != instance_id) {
114 result = proj_in->in(TypeFunc::Memory);
115 }
116 } else if (proj_in->is_MemBar()) {
117 result = proj_in->in(TypeFunc::Memory);
118 }
119 } else if (result->is_MergeMem()) {
120 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
121 }
122 }
123 return result;
124 }
126 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
127 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
128 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
129 PhaseIterGVN *igvn = phase->is_IterGVN();
130 Node *result = mchain;
131 result = optimize_simple_memory_chain(result, t_adr, phase);
132 if (is_instance && igvn != NULL && result->is_Phi()) {
133 PhiNode *mphi = result->as_Phi();
134 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
135 const TypePtr *t = mphi->adr_type();
136 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
137 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
138 t->is_oopptr()->cast_to_exactness(true)
139 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
140 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
141 // clone the Phi with our address type
142 result = mphi->split_out_instance(t_adr, igvn);
143 } else {
144 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
145 }
146 }
147 return result;
148 }
150 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
151 uint alias_idx = phase->C->get_alias_index(tp);
152 Node *mem = mmem;
153 #ifdef ASSERT
154 {
155 // Check that current type is consistent with the alias index used during graph construction
156 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
157 bool consistent = adr_check == NULL || adr_check->empty() ||
158 phase->C->must_alias(adr_check, alias_idx );
159 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
160 if( !consistent && adr_check != NULL && !adr_check->empty() &&
161 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
162 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
163 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
164 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
165 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
166 // don't assert if it is dead code.
167 consistent = true;
168 }
169 if( !consistent ) {
170 st->print("alias_idx==%d, adr_check==", alias_idx);
171 if( adr_check == NULL ) {
172 st->print("NULL");
173 } else {
174 adr_check->dump();
175 }
176 st->cr();
177 print_alias_types();
178 assert(consistent, "adr_check must match alias idx");
179 }
180 }
181 #endif
182 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
183 // means an array I have not precisely typed yet. Do not do any
184 // alias stuff with it any time soon.
185 const TypeOopPtr *tinst = tp->isa_oopptr();
186 if( tp->base() != Type::AnyPtr &&
187 !(tinst &&
188 tinst->klass()->is_java_lang_Object() &&
189 tinst->offset() == Type::OffsetBot) ) {
190 // compress paths and change unreachable cycles to TOP
191 // If not, we can update the input infinitely along a MergeMem cycle
192 // Equivalent code in PhiNode::Ideal
193 Node* m = phase->transform(mmem);
194 // If tranformed to a MergeMem, get the desired slice
195 // Otherwise the returned node represents memory for every slice
196 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
197 // Update input if it is progress over what we have now
198 }
199 return mem;
200 }
202 //--------------------------Ideal_common---------------------------------------
203 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
204 // Unhook non-raw memories from complete (macro-expanded) initializations.
205 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
206 // If our control input is a dead region, kill all below the region
207 Node *ctl = in(MemNode::Control);
208 if (ctl && remove_dead_region(phase, can_reshape))
209 return this;
211 // Ignore if memory is dead, or self-loop
212 Node *mem = in(MemNode::Memory);
213 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
214 assert( mem != this, "dead loop in MemNode::Ideal" );
216 Node *address = in(MemNode::Address);
217 const Type *t_adr = phase->type( address );
218 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
220 // Avoid independent memory operations
221 Node* old_mem = mem;
223 // The code which unhooks non-raw memories from complete (macro-expanded)
224 // initializations was removed. After macro-expansion all stores catched
225 // by Initialize node became raw stores and there is no information
226 // which memory slices they modify. So it is unsafe to move any memory
227 // operation above these stores. Also in most cases hooked non-raw memories
228 // were already unhooked by using information from detect_ptr_independence()
229 // and find_previous_store().
231 if (mem->is_MergeMem()) {
232 MergeMemNode* mmem = mem->as_MergeMem();
233 const TypePtr *tp = t_adr->is_ptr();
235 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
236 }
238 if (mem != old_mem) {
239 set_req(MemNode::Memory, mem);
240 return this;
241 }
243 // let the subclass continue analyzing...
244 return NULL;
245 }
247 // Helper function for proving some simple control dominations.
248 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
249 // Already assumes that 'dom' is available at 'sub', and that 'sub'
250 // is not a constant (dominated by the method's StartNode).
251 // Used by MemNode::find_previous_store to prove that the
252 // control input of a memory operation predates (dominates)
253 // an allocation it wants to look past.
254 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
255 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
256 return false; // Conservative answer for dead code
258 // Check 'dom'. Skip Proj and CatchProj nodes.
259 dom = dom->find_exact_control(dom);
260 if (dom == NULL || dom->is_top())
261 return false; // Conservative answer for dead code
263 if (dom == sub) {
264 // For the case when, for example, 'sub' is Initialize and the original
265 // 'dom' is Proj node of the 'sub'.
266 return false;
267 }
269 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
270 return true;
272 // 'dom' dominates 'sub' if its control edge and control edges
273 // of all its inputs dominate or equal to sub's control edge.
275 // Currently 'sub' is either Allocate, Initialize or Start nodes.
276 // Or Region for the check in LoadNode::Ideal();
277 // 'sub' should have sub->in(0) != NULL.
278 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
279 sub->is_Region(), "expecting only these nodes");
281 // Get control edge of 'sub'.
282 Node* orig_sub = sub;
283 sub = sub->find_exact_control(sub->in(0));
284 if (sub == NULL || sub->is_top())
285 return false; // Conservative answer for dead code
287 assert(sub->is_CFG(), "expecting control");
289 if (sub == dom)
290 return true;
292 if (sub->is_Start() || sub->is_Root())
293 return false;
295 {
296 // Check all control edges of 'dom'.
298 ResourceMark rm;
299 Arena* arena = Thread::current()->resource_area();
300 Node_List nlist(arena);
301 Unique_Node_List dom_list(arena);
303 dom_list.push(dom);
304 bool only_dominating_controls = false;
306 for (uint next = 0; next < dom_list.size(); next++) {
307 Node* n = dom_list.at(next);
308 if (n == orig_sub)
309 return false; // One of dom's inputs dominated by sub.
310 if (!n->is_CFG() && n->pinned()) {
311 // Check only own control edge for pinned non-control nodes.
312 n = n->find_exact_control(n->in(0));
313 if (n == NULL || n->is_top())
314 return false; // Conservative answer for dead code
315 assert(n->is_CFG(), "expecting control");
316 dom_list.push(n);
317 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
318 only_dominating_controls = true;
319 } else if (n->is_CFG()) {
320 if (n->dominates(sub, nlist))
321 only_dominating_controls = true;
322 else
323 return false;
324 } else {
325 // First, own control edge.
326 Node* m = n->find_exact_control(n->in(0));
327 if (m != NULL) {
328 if (m->is_top())
329 return false; // Conservative answer for dead code
330 dom_list.push(m);
331 }
332 // Now, the rest of edges.
333 uint cnt = n->req();
334 for (uint i = 1; i < cnt; i++) {
335 m = n->find_exact_control(n->in(i));
336 if (m == NULL || m->is_top())
337 continue;
338 dom_list.push(m);
339 }
340 }
341 }
342 return only_dominating_controls;
343 }
344 }
346 //---------------------detect_ptr_independence---------------------------------
347 // Used by MemNode::find_previous_store to prove that two base
348 // pointers are never equal.
349 // The pointers are accompanied by their associated allocations,
350 // if any, which have been previously discovered by the caller.
351 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
352 Node* p2, AllocateNode* a2,
353 PhaseTransform* phase) {
354 // Attempt to prove that these two pointers cannot be aliased.
355 // They may both manifestly be allocations, and they should differ.
356 // Or, if they are not both allocations, they can be distinct constants.
357 // Otherwise, one is an allocation and the other a pre-existing value.
358 if (a1 == NULL && a2 == NULL) { // neither an allocation
359 return (p1 != p2) && p1->is_Con() && p2->is_Con();
360 } else if (a1 != NULL && a2 != NULL) { // both allocations
361 return (a1 != a2);
362 } else if (a1 != NULL) { // one allocation a1
363 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
364 return all_controls_dominate(p2, a1);
365 } else { //(a2 != NULL) // one allocation a2
366 return all_controls_dominate(p1, a2);
367 }
368 return false;
369 }
372 // The logic for reordering loads and stores uses four steps:
373 // (a) Walk carefully past stores and initializations which we
374 // can prove are independent of this load.
375 // (b) Observe that the next memory state makes an exact match
376 // with self (load or store), and locate the relevant store.
377 // (c) Ensure that, if we were to wire self directly to the store,
378 // the optimizer would fold it up somehow.
379 // (d) Do the rewiring, and return, depending on some other part of
380 // the optimizer to fold up the load.
381 // This routine handles steps (a) and (b). Steps (c) and (d) are
382 // specific to loads and stores, so they are handled by the callers.
383 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
384 //
385 Node* MemNode::find_previous_store(PhaseTransform* phase) {
386 Node* ctrl = in(MemNode::Control);
387 Node* adr = in(MemNode::Address);
388 intptr_t offset = 0;
389 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
390 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
392 if (offset == Type::OffsetBot)
393 return NULL; // cannot unalias unless there are precise offsets
395 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
397 intptr_t size_in_bytes = memory_size();
399 Node* mem = in(MemNode::Memory); // start searching here...
401 int cnt = 50; // Cycle limiter
402 for (;;) { // While we can dance past unrelated stores...
403 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
405 if (mem->is_Store()) {
406 Node* st_adr = mem->in(MemNode::Address);
407 intptr_t st_offset = 0;
408 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
409 if (st_base == NULL)
410 break; // inscrutable pointer
411 if (st_offset != offset && st_offset != Type::OffsetBot) {
412 const int MAX_STORE = BytesPerLong;
413 if (st_offset >= offset + size_in_bytes ||
414 st_offset <= offset - MAX_STORE ||
415 st_offset <= offset - mem->as_Store()->memory_size()) {
416 // Success: The offsets are provably independent.
417 // (You may ask, why not just test st_offset != offset and be done?
418 // The answer is that stores of different sizes can co-exist
419 // in the same sequence of RawMem effects. We sometimes initialize
420 // a whole 'tile' of array elements with a single jint or jlong.)
421 mem = mem->in(MemNode::Memory);
422 continue; // (a) advance through independent store memory
423 }
424 }
425 if (st_base != base &&
426 detect_ptr_independence(base, alloc,
427 st_base,
428 AllocateNode::Ideal_allocation(st_base, phase),
429 phase)) {
430 // Success: The bases are provably independent.
431 mem = mem->in(MemNode::Memory);
432 continue; // (a) advance through independent store memory
433 }
435 // (b) At this point, if the bases or offsets do not agree, we lose,
436 // since we have not managed to prove 'this' and 'mem' independent.
437 if (st_base == base && st_offset == offset) {
438 return mem; // let caller handle steps (c), (d)
439 }
441 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
442 InitializeNode* st_init = mem->in(0)->as_Initialize();
443 AllocateNode* st_alloc = st_init->allocation();
444 if (st_alloc == NULL)
445 break; // something degenerated
446 bool known_identical = false;
447 bool known_independent = false;
448 if (alloc == st_alloc)
449 known_identical = true;
450 else if (alloc != NULL)
451 known_independent = true;
452 else if (all_controls_dominate(this, st_alloc))
453 known_independent = true;
455 if (known_independent) {
456 // The bases are provably independent: Either they are
457 // manifestly distinct allocations, or else the control
458 // of this load dominates the store's allocation.
459 int alias_idx = phase->C->get_alias_index(adr_type());
460 if (alias_idx == Compile::AliasIdxRaw) {
461 mem = st_alloc->in(TypeFunc::Memory);
462 } else {
463 mem = st_init->memory(alias_idx);
464 }
465 continue; // (a) advance through independent store memory
466 }
468 // (b) at this point, if we are not looking at a store initializing
469 // the same allocation we are loading from, we lose.
470 if (known_identical) {
471 // From caller, can_see_stored_value will consult find_captured_store.
472 return mem; // let caller handle steps (c), (d)
473 }
475 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
476 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
477 if (mem->is_Proj() && mem->in(0)->is_Call()) {
478 CallNode *call = mem->in(0)->as_Call();
479 if (!call->may_modify(addr_t, phase)) {
480 mem = call->in(TypeFunc::Memory);
481 continue; // (a) advance through independent call memory
482 }
483 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
484 mem = mem->in(0)->in(TypeFunc::Memory);
485 continue; // (a) advance through independent MemBar memory
486 } else if (mem->is_MergeMem()) {
487 int alias_idx = phase->C->get_alias_index(adr_type());
488 mem = mem->as_MergeMem()->memory_at(alias_idx);
489 continue; // (a) advance through independent MergeMem memory
490 }
491 }
493 // Unless there is an explicit 'continue', we must bail out here,
494 // because 'mem' is an inscrutable memory state (e.g., a call).
495 break;
496 }
498 return NULL; // bail out
499 }
501 //----------------------calculate_adr_type-------------------------------------
502 // Helper function. Notices when the given type of address hits top or bottom.
503 // Also, asserts a cross-check of the type against the expected address type.
504 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
505 if (t == Type::TOP) return NULL; // does not touch memory any more?
506 #ifdef PRODUCT
507 cross_check = NULL;
508 #else
509 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
510 #endif
511 const TypePtr* tp = t->isa_ptr();
512 if (tp == NULL) {
513 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
514 return TypePtr::BOTTOM; // touches lots of memory
515 } else {
516 #ifdef ASSERT
517 // %%%% [phh] We don't check the alias index if cross_check is
518 // TypeRawPtr::BOTTOM. Needs to be investigated.
519 if (cross_check != NULL &&
520 cross_check != TypePtr::BOTTOM &&
521 cross_check != TypeRawPtr::BOTTOM) {
522 // Recheck the alias index, to see if it has changed (due to a bug).
523 Compile* C = Compile::current();
524 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
525 "must stay in the original alias category");
526 // The type of the address must be contained in the adr_type,
527 // disregarding "null"-ness.
528 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
529 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
530 assert(cross_check->meet(tp_notnull) == cross_check,
531 "real address must not escape from expected memory type");
532 }
533 #endif
534 return tp;
535 }
536 }
538 //------------------------adr_phi_is_loop_invariant----------------------------
539 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
540 // loop is loop invariant. Make a quick traversal of Phi and associated
541 // CastPP nodes, looking to see if they are a closed group within the loop.
542 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
543 // The idea is that the phi-nest must boil down to only CastPP nodes
544 // with the same data. This implies that any path into the loop already
545 // includes such a CastPP, and so the original cast, whatever its input,
546 // must be covered by an equivalent cast, with an earlier control input.
547 ResourceMark rm;
549 // The loop entry input of the phi should be the unique dominating
550 // node for every Phi/CastPP in the loop.
551 Unique_Node_List closure;
552 closure.push(adr_phi->in(LoopNode::EntryControl));
554 // Add the phi node and the cast to the worklist.
555 Unique_Node_List worklist;
556 worklist.push(adr_phi);
557 if( cast != NULL ){
558 if( !cast->is_ConstraintCast() ) return false;
559 worklist.push(cast);
560 }
562 // Begin recursive walk of phi nodes.
563 while( worklist.size() ){
564 // Take a node off the worklist
565 Node *n = worklist.pop();
566 if( !closure.member(n) ){
567 // Add it to the closure.
568 closure.push(n);
569 // Make a sanity check to ensure we don't waste too much time here.
570 if( closure.size() > 20) return false;
571 // This node is OK if:
572 // - it is a cast of an identical value
573 // - or it is a phi node (then we add its inputs to the worklist)
574 // Otherwise, the node is not OK, and we presume the cast is not invariant
575 if( n->is_ConstraintCast() ){
576 worklist.push(n->in(1));
577 } else if( n->is_Phi() ) {
578 for( uint i = 1; i < n->req(); i++ ) {
579 worklist.push(n->in(i));
580 }
581 } else {
582 return false;
583 }
584 }
585 }
587 // Quit when the worklist is empty, and we've found no offending nodes.
588 return true;
589 }
591 //------------------------------Ideal_DU_postCCP-------------------------------
592 // Find any cast-away of null-ness and keep its control. Null cast-aways are
593 // going away in this pass and we need to make this memory op depend on the
594 // gating null check.
595 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
596 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
597 }
599 // I tried to leave the CastPP's in. This makes the graph more accurate in
600 // some sense; we get to keep around the knowledge that an oop is not-null
601 // after some test. Alas, the CastPP's interfere with GVN (some values are
602 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
603 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
604 // some of the more trivial cases in the optimizer. Removing more useless
605 // Phi's started allowing Loads to illegally float above null checks. I gave
606 // up on this approach. CNC 10/20/2000
607 // This static method may be called not from MemNode (EncodePNode calls it).
608 // Only the control edge of the node 'n' might be updated.
609 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
610 Node *skipped_cast = NULL;
611 // Need a null check? Regular static accesses do not because they are
612 // from constant addresses. Array ops are gated by the range check (which
613 // always includes a NULL check). Just check field ops.
614 if( n->in(MemNode::Control) == NULL ) {
615 // Scan upwards for the highest location we can place this memory op.
616 while( true ) {
617 switch( adr->Opcode() ) {
619 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
620 adr = adr->in(AddPNode::Base);
621 continue;
623 case Op_DecodeN: // No change to NULL-ness, so peek thru
624 adr = adr->in(1);
625 continue;
627 case Op_CastPP:
628 // If the CastPP is useless, just peek on through it.
629 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
630 // Remember the cast that we've peeked though. If we peek
631 // through more than one, then we end up remembering the highest
632 // one, that is, if in a loop, the one closest to the top.
633 skipped_cast = adr;
634 adr = adr->in(1);
635 continue;
636 }
637 // CastPP is going away in this pass! We need this memory op to be
638 // control-dependent on the test that is guarding the CastPP.
639 ccp->hash_delete(n);
640 n->set_req(MemNode::Control, adr->in(0));
641 ccp->hash_insert(n);
642 return n;
644 case Op_Phi:
645 // Attempt to float above a Phi to some dominating point.
646 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
647 // If we've already peeked through a Cast (which could have set the
648 // control), we can't float above a Phi, because the skipped Cast
649 // may not be loop invariant.
650 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
651 adr = adr->in(1);
652 continue;
653 }
654 }
656 // Intentional fallthrough!
658 // No obvious dominating point. The mem op is pinned below the Phi
659 // by the Phi itself. If the Phi goes away (no true value is merged)
660 // then the mem op can float, but not indefinitely. It must be pinned
661 // behind the controls leading to the Phi.
662 case Op_CheckCastPP:
663 // These usually stick around to change address type, however a
664 // useless one can be elided and we still need to pick up a control edge
665 if (adr->in(0) == NULL) {
666 // This CheckCastPP node has NO control and is likely useless. But we
667 // need check further up the ancestor chain for a control input to keep
668 // the node in place. 4959717.
669 skipped_cast = adr;
670 adr = adr->in(1);
671 continue;
672 }
673 ccp->hash_delete(n);
674 n->set_req(MemNode::Control, adr->in(0));
675 ccp->hash_insert(n);
676 return n;
678 // List of "safe" opcodes; those that implicitly block the memory
679 // op below any null check.
680 case Op_CastX2P: // no null checks on native pointers
681 case Op_Parm: // 'this' pointer is not null
682 case Op_LoadP: // Loading from within a klass
683 case Op_LoadN: // Loading from within a klass
684 case Op_LoadKlass: // Loading from within a klass
685 case Op_LoadNKlass: // Loading from within a klass
686 case Op_ConP: // Loading from a klass
687 case Op_ConN: // Loading from a klass
688 case Op_CreateEx: // Sucking up the guts of an exception oop
689 case Op_Con: // Reading from TLS
690 case Op_CMoveP: // CMoveP is pinned
691 case Op_CMoveN: // CMoveN is pinned
692 break; // No progress
694 case Op_Proj: // Direct call to an allocation routine
695 case Op_SCMemProj: // Memory state from store conditional ops
696 #ifdef ASSERT
697 {
698 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
699 const Node* call = adr->in(0);
700 if (call->is_CallJava()) {
701 const CallJavaNode* call_java = call->as_CallJava();
702 const TypeTuple *r = call_java->tf()->range();
703 assert(r->cnt() > TypeFunc::Parms, "must return value");
704 const Type* ret_type = r->field_at(TypeFunc::Parms);
705 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
706 // We further presume that this is one of
707 // new_instance_Java, new_array_Java, or
708 // the like, but do not assert for this.
709 } else if (call->is_Allocate()) {
710 // similar case to new_instance_Java, etc.
711 } else if (!call->is_CallLeaf()) {
712 // Projections from fetch_oop (OSR) are allowed as well.
713 ShouldNotReachHere();
714 }
715 }
716 #endif
717 break;
718 default:
719 ShouldNotReachHere();
720 }
721 break;
722 }
723 }
725 return NULL; // No progress
726 }
729 //=============================================================================
730 uint LoadNode::size_of() const { return sizeof(*this); }
731 uint LoadNode::cmp( const Node &n ) const
732 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
733 const Type *LoadNode::bottom_type() const { return _type; }
734 uint LoadNode::ideal_reg() const {
735 return Matcher::base2reg[_type->base()];
736 }
738 #ifndef PRODUCT
739 void LoadNode::dump_spec(outputStream *st) const {
740 MemNode::dump_spec(st);
741 if( !Verbose && !WizardMode ) {
742 // standard dump does this in Verbose and WizardMode
743 st->print(" #"); _type->dump_on(st);
744 }
745 }
746 #endif
749 //----------------------------LoadNode::make-----------------------------------
750 // Polymorphic factory method:
751 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
752 Compile* C = gvn.C;
754 // sanity check the alias category against the created node type
755 assert(!(adr_type->isa_oopptr() &&
756 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
757 "use LoadKlassNode instead");
758 assert(!(adr_type->isa_aryptr() &&
759 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
760 "use LoadRangeNode instead");
761 switch (bt) {
762 case T_BOOLEAN:
763 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
764 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
765 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
766 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
767 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
768 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
769 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
770 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
771 case T_OBJECT:
772 #ifdef _LP64
773 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
774 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
775 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
776 } else
777 #endif
778 {
779 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
780 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
781 }
782 }
783 ShouldNotReachHere();
784 return (LoadNode*)NULL;
785 }
787 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
788 bool require_atomic = true;
789 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
790 }
795 //------------------------------hash-------------------------------------------
796 uint LoadNode::hash() const {
797 // unroll addition of interesting fields
798 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
799 }
801 //---------------------------can_see_stored_value------------------------------
802 // This routine exists to make sure this set of tests is done the same
803 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
804 // will change the graph shape in a way which makes memory alive twice at the
805 // same time (uses the Oracle model of aliasing), then some
806 // LoadXNode::Identity will fold things back to the equivalence-class model
807 // of aliasing.
808 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
809 Node* ld_adr = in(MemNode::Address);
811 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
812 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
813 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
814 atp->field() != NULL && !atp->field()->is_volatile()) {
815 uint alias_idx = atp->index();
816 bool final = atp->field()->is_final();
817 Node* result = NULL;
818 Node* current = st;
819 // Skip through chains of MemBarNodes checking the MergeMems for
820 // new states for the slice of this load. Stop once any other
821 // kind of node is encountered. Loads from final memory can skip
822 // through any kind of MemBar but normal loads shouldn't skip
823 // through MemBarAcquire since the could allow them to move out of
824 // a synchronized region.
825 while (current->is_Proj()) {
826 int opc = current->in(0)->Opcode();
827 if ((final && opc == Op_MemBarAcquire) ||
828 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
829 Node* mem = current->in(0)->in(TypeFunc::Memory);
830 if (mem->is_MergeMem()) {
831 MergeMemNode* merge = mem->as_MergeMem();
832 Node* new_st = merge->memory_at(alias_idx);
833 if (new_st == merge->base_memory()) {
834 // Keep searching
835 current = merge->base_memory();
836 continue;
837 }
838 // Save the new memory state for the slice and fall through
839 // to exit.
840 result = new_st;
841 }
842 }
843 break;
844 }
845 if (result != NULL) {
846 st = result;
847 }
848 }
851 // Loop around twice in the case Load -> Initialize -> Store.
852 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
853 for (int trip = 0; trip <= 1; trip++) {
855 if (st->is_Store()) {
856 Node* st_adr = st->in(MemNode::Address);
857 if (!phase->eqv(st_adr, ld_adr)) {
858 // Try harder before giving up... Match raw and non-raw pointers.
859 intptr_t st_off = 0;
860 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
861 if (alloc == NULL) return NULL;
862 intptr_t ld_off = 0;
863 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
864 if (alloc != allo2) return NULL;
865 if (ld_off != st_off) return NULL;
866 // At this point we have proven something like this setup:
867 // A = Allocate(...)
868 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
869 // S = StoreQ(, AddP(, A.Parm , #Off), V)
870 // (Actually, we haven't yet proven the Q's are the same.)
871 // In other words, we are loading from a casted version of
872 // the same pointer-and-offset that we stored to.
873 // Thus, we are able to replace L by V.
874 }
875 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
876 if (store_Opcode() != st->Opcode())
877 return NULL;
878 return st->in(MemNode::ValueIn);
879 }
881 intptr_t offset = 0; // scratch
883 // A load from a freshly-created object always returns zero.
884 // (This can happen after LoadNode::Ideal resets the load's memory input
885 // to find_captured_store, which returned InitializeNode::zero_memory.)
886 if (st->is_Proj() && st->in(0)->is_Allocate() &&
887 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
888 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
889 // return a zero value for the load's basic type
890 // (This is one of the few places where a generic PhaseTransform
891 // can create new nodes. Think of it as lazily manifesting
892 // virtually pre-existing constants.)
893 return phase->zerocon(memory_type());
894 }
896 // A load from an initialization barrier can match a captured store.
897 if (st->is_Proj() && st->in(0)->is_Initialize()) {
898 InitializeNode* init = st->in(0)->as_Initialize();
899 AllocateNode* alloc = init->allocation();
900 if (alloc != NULL &&
901 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
902 // examine a captured store value
903 st = init->find_captured_store(offset, memory_size(), phase);
904 if (st != NULL)
905 continue; // take one more trip around
906 }
907 }
909 break;
910 }
912 return NULL;
913 }
915 //----------------------is_instance_field_load_with_local_phi------------------
916 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
917 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
918 in(MemNode::Address)->is_AddP() ) {
919 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
920 // Only instances.
921 if( t_oop != NULL && t_oop->is_known_instance_field() &&
922 t_oop->offset() != Type::OffsetBot &&
923 t_oop->offset() != Type::OffsetTop) {
924 return true;
925 }
926 }
927 return false;
928 }
930 //------------------------------Identity---------------------------------------
931 // Loads are identity if previous store is to same address
932 Node *LoadNode::Identity( PhaseTransform *phase ) {
933 // If the previous store-maker is the right kind of Store, and the store is
934 // to the same address, then we are equal to the value stored.
935 Node* mem = in(MemNode::Memory);
936 Node* value = can_see_stored_value(mem, phase);
937 if( value ) {
938 // byte, short & char stores truncate naturally.
939 // A load has to load the truncated value which requires
940 // some sort of masking operation and that requires an
941 // Ideal call instead of an Identity call.
942 if (memory_size() < BytesPerInt) {
943 // If the input to the store does not fit with the load's result type,
944 // it must be truncated via an Ideal call.
945 if (!phase->type(value)->higher_equal(phase->type(this)))
946 return this;
947 }
948 // (This works even when value is a Con, but LoadNode::Value
949 // usually runs first, producing the singleton type of the Con.)
950 return value;
951 }
953 // Search for an existing data phi which was generated before for the same
954 // instance's field to avoid infinite genertion of phis in a loop.
955 Node *region = mem->in(0);
956 if (is_instance_field_load_with_local_phi(region)) {
957 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
958 int this_index = phase->C->get_alias_index(addr_t);
959 int this_offset = addr_t->offset();
960 int this_id = addr_t->is_oopptr()->instance_id();
961 const Type* this_type = bottom_type();
962 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
963 Node* phi = region->fast_out(i);
964 if (phi->is_Phi() && phi != mem &&
965 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
966 return phi;
967 }
968 }
969 }
971 return this;
972 }
975 // Returns true if the AliasType refers to the field that holds the
976 // cached box array. Currently only handles the IntegerCache case.
977 static bool is_autobox_cache(Compile::AliasType* atp) {
978 if (atp != NULL && atp->field() != NULL) {
979 ciField* field = atp->field();
980 ciSymbol* klass = field->holder()->name();
981 if (field->name() == ciSymbol::cache_field_name() &&
982 field->holder()->uses_default_loader() &&
983 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
984 return true;
985 }
986 }
987 return false;
988 }
990 // Fetch the base value in the autobox array
991 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
992 if (atp != NULL && atp->field() != NULL) {
993 ciField* field = atp->field();
994 ciSymbol* klass = field->holder()->name();
995 if (field->name() == ciSymbol::cache_field_name() &&
996 field->holder()->uses_default_loader() &&
997 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
998 assert(field->is_constant(), "what?");
999 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
1000 // Fetch the box object at the base of the array and get its value
1001 ciInstance* box = array->obj_at(0)->as_instance();
1002 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1003 if (ik->nof_nonstatic_fields() == 1) {
1004 // This should be true nonstatic_field_at requires calling
1005 // nof_nonstatic_fields so check it anyway
1006 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1007 cache_offset = c.as_int();
1008 }
1009 return true;
1010 }
1011 }
1012 return false;
1013 }
1015 // Returns true if the AliasType refers to the value field of an
1016 // autobox object. Currently only handles Integer.
1017 static bool is_autobox_object(Compile::AliasType* atp) {
1018 if (atp != NULL && atp->field() != NULL) {
1019 ciField* field = atp->field();
1020 ciSymbol* klass = field->holder()->name();
1021 if (field->name() == ciSymbol::value_name() &&
1022 field->holder()->uses_default_loader() &&
1023 klass == ciSymbol::java_lang_Integer()) {
1024 return true;
1025 }
1026 }
1027 return false;
1028 }
1031 // We're loading from an object which has autobox behaviour.
1032 // If this object is result of a valueOf call we'll have a phi
1033 // merging a newly allocated object and a load from the cache.
1034 // We want to replace this load with the original incoming
1035 // argument to the valueOf call.
1036 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1037 Node* base = in(Address)->in(AddPNode::Base);
1038 if (base->is_Phi() && base->req() == 3) {
1039 AllocateNode* allocation = NULL;
1040 int allocation_index = -1;
1041 int load_index = -1;
1042 for (uint i = 1; i < base->req(); i++) {
1043 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1044 if (allocation != NULL) {
1045 allocation_index = i;
1046 load_index = 3 - allocation_index;
1047 break;
1048 }
1049 }
1050 LoadNode* load = NULL;
1051 if (allocation != NULL && base->in(load_index)->is_Load()) {
1052 load = base->in(load_index)->as_Load();
1053 }
1054 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1055 // Push the loads from the phi that comes from valueOf up
1056 // through it to allow elimination of the loads and the recovery
1057 // of the original value.
1058 Node* mem_phi = in(Memory);
1059 Node* offset = in(Address)->in(AddPNode::Offset);
1061 Node* in1 = clone();
1062 Node* in1_addr = in1->in(Address)->clone();
1063 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1064 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1065 in1_addr->set_req(AddPNode::Offset, offset);
1066 in1->set_req(0, base->in(allocation_index));
1067 in1->set_req(Address, in1_addr);
1068 in1->set_req(Memory, mem_phi->in(allocation_index));
1070 Node* in2 = clone();
1071 Node* in2_addr = in2->in(Address)->clone();
1072 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1073 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1074 in2_addr->set_req(AddPNode::Offset, offset);
1075 in2->set_req(0, base->in(load_index));
1076 in2->set_req(Address, in2_addr);
1077 in2->set_req(Memory, mem_phi->in(load_index));
1079 in1_addr = phase->transform(in1_addr);
1080 in1 = phase->transform(in1);
1081 in2_addr = phase->transform(in2_addr);
1082 in2 = phase->transform(in2);
1084 PhiNode* result = PhiNode::make_blank(base->in(0), this);
1085 result->set_req(allocation_index, in1);
1086 result->set_req(load_index, in2);
1087 return result;
1088 }
1089 } else if (base->is_Load()) {
1090 // Eliminate the load of Integer.value for integers from the cache
1091 // array by deriving the value from the index into the array.
1092 // Capture the offset of the load and then reverse the computation.
1093 Node* load_base = base->in(Address)->in(AddPNode::Base);
1094 if (load_base != NULL) {
1095 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1096 intptr_t cache_offset;
1097 int shift = -1;
1098 Node* cache = NULL;
1099 if (is_autobox_cache(atp)) {
1100 shift = exact_log2(type2aelembytes(T_OBJECT));
1101 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1102 }
1103 if (cache != NULL && base->in(Address)->is_AddP()) {
1104 Node* elements[4];
1105 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1106 int cache_low;
1107 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1108 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1109 // Add up all the offsets making of the address of the load
1110 Node* result = elements[0];
1111 for (int i = 1; i < count; i++) {
1112 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1113 }
1114 // Remove the constant offset from the address and then
1115 // remove the scaling of the offset to recover the original index.
1116 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1117 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1118 // Peel the shift off directly but wrap it in a dummy node
1119 // since Ideal can't return existing nodes
1120 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1121 } else {
1122 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1123 }
1124 #ifdef _LP64
1125 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1126 #endif
1127 return result;
1128 }
1129 }
1130 }
1131 }
1132 return NULL;
1133 }
1135 //------------------------------split_through_phi------------------------------
1136 // Split instance field load through Phi.
1137 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1138 Node* mem = in(MemNode::Memory);
1139 Node* address = in(MemNode::Address);
1140 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1141 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1143 assert(mem->is_Phi() && (t_oop != NULL) &&
1144 t_oop->is_known_instance_field(), "invalide conditions");
1146 Node *region = mem->in(0);
1147 if (region == NULL) {
1148 return NULL; // Wait stable graph
1149 }
1150 uint cnt = mem->req();
1151 for( uint i = 1; i < cnt; i++ ) {
1152 Node *in = mem->in(i);
1153 if( in == NULL ) {
1154 return NULL; // Wait stable graph
1155 }
1156 }
1157 // Check for loop invariant.
1158 if (cnt == 3) {
1159 for( uint i = 1; i < cnt; i++ ) {
1160 Node *in = mem->in(i);
1161 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1162 if (m == mem) {
1163 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
1164 return this;
1165 }
1166 }
1167 }
1168 // Split through Phi (see original code in loopopts.cpp).
1169 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1171 // Do nothing here if Identity will find a value
1172 // (to avoid infinite chain of value phis generation).
1173 if ( !phase->eqv(this, this->Identity(phase)) )
1174 return NULL;
1176 // Skip the split if the region dominates some control edge of the address.
1177 if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
1178 return NULL;
1180 const Type* this_type = this->bottom_type();
1181 int this_index = phase->C->get_alias_index(addr_t);
1182 int this_offset = addr_t->offset();
1183 int this_iid = addr_t->is_oopptr()->instance_id();
1184 int wins = 0;
1185 PhaseIterGVN *igvn = phase->is_IterGVN();
1186 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1187 for( uint i = 1; i < region->req(); i++ ) {
1188 Node *x;
1189 Node* the_clone = NULL;
1190 if( region->in(i) == phase->C->top() ) {
1191 x = phase->C->top(); // Dead path? Use a dead data op
1192 } else {
1193 x = this->clone(); // Else clone up the data op
1194 the_clone = x; // Remember for possible deletion.
1195 // Alter data node to use pre-phi inputs
1196 if( this->in(0) == region ) {
1197 x->set_req( 0, region->in(i) );
1198 } else {
1199 x->set_req( 0, NULL );
1200 }
1201 for( uint j = 1; j < this->req(); j++ ) {
1202 Node *in = this->in(j);
1203 if( in->is_Phi() && in->in(0) == region )
1204 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1205 }
1206 }
1207 // Check for a 'win' on some paths
1208 const Type *t = x->Value(igvn);
1210 bool singleton = t->singleton();
1212 // See comments in PhaseIdealLoop::split_thru_phi().
1213 if( singleton && t == Type::TOP ) {
1214 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1215 }
1217 if( singleton ) {
1218 wins++;
1219 x = igvn->makecon(t);
1220 } else {
1221 // We now call Identity to try to simplify the cloned node.
1222 // Note that some Identity methods call phase->type(this).
1223 // Make sure that the type array is big enough for
1224 // our new node, even though we may throw the node away.
1225 // (This tweaking with igvn only works because x is a new node.)
1226 igvn->set_type(x, t);
1227 Node *y = x->Identity(igvn);
1228 if( y != x ) {
1229 wins++;
1230 x = y;
1231 } else {
1232 y = igvn->hash_find(x);
1233 if( y ) {
1234 wins++;
1235 x = y;
1236 } else {
1237 // Else x is a new node we are keeping
1238 // We do not need register_new_node_with_optimizer
1239 // because set_type has already been called.
1240 igvn->_worklist.push(x);
1241 }
1242 }
1243 }
1244 if (x != the_clone && the_clone != NULL)
1245 igvn->remove_dead_node(the_clone);
1246 phi->set_req(i, x);
1247 }
1248 if( wins > 0 ) {
1249 // Record Phi
1250 igvn->register_new_node_with_optimizer(phi);
1251 return phi;
1252 }
1253 igvn->remove_dead_node(phi);
1254 return NULL;
1255 }
1257 //------------------------------Ideal------------------------------------------
1258 // If the load is from Field memory and the pointer is non-null, we can
1259 // zero out the control input.
1260 // If the offset is constant and the base is an object allocation,
1261 // try to hook me up to the exact initializing store.
1262 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1263 Node* p = MemNode::Ideal_common(phase, can_reshape);
1264 if (p) return (p == NodeSentinel) ? NULL : p;
1266 Node* ctrl = in(MemNode::Control);
1267 Node* address = in(MemNode::Address);
1269 // Skip up past a SafePoint control. Cannot do this for Stores because
1270 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1271 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1272 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1273 ctrl = ctrl->in(0);
1274 set_req(MemNode::Control,ctrl);
1275 }
1277 // Check for useless control edge in some common special cases
1278 if (in(MemNode::Control) != NULL) {
1279 intptr_t ignore = 0;
1280 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1281 if (base != NULL
1282 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1283 && all_controls_dominate(base, phase->C->start())) {
1284 // A method-invariant, non-null address (constant or 'this' argument).
1285 set_req(MemNode::Control, NULL);
1286 }
1287 }
1289 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
1290 Node* base = in(Address)->in(AddPNode::Base);
1291 if (base != NULL) {
1292 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1293 if (is_autobox_object(atp)) {
1294 Node* result = eliminate_autobox(phase);
1295 if (result != NULL) return result;
1296 }
1297 }
1298 }
1300 Node* mem = in(MemNode::Memory);
1301 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1303 if (addr_t != NULL) {
1304 // try to optimize our memory input
1305 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1306 if (opt_mem != mem) {
1307 set_req(MemNode::Memory, opt_mem);
1308 return this;
1309 }
1310 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1311 if (can_reshape && opt_mem->is_Phi() &&
1312 (t_oop != NULL) && t_oop->is_known_instance_field()) {
1313 // Split instance field load through Phi.
1314 Node* result = split_through_phi(phase);
1315 if (result != NULL) return result;
1316 }
1317 }
1319 // Check for prior store with a different base or offset; make Load
1320 // independent. Skip through any number of them. Bail out if the stores
1321 // are in an endless dead cycle and report no progress. This is a key
1322 // transform for Reflection. However, if after skipping through the Stores
1323 // we can't then fold up against a prior store do NOT do the transform as
1324 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1325 // array memory alive twice: once for the hoisted Load and again after the
1326 // bypassed Store. This situation only works if EVERYBODY who does
1327 // anti-dependence work knows how to bypass. I.e. we need all
1328 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1329 // the alias index stuff. So instead, peek through Stores and IFF we can
1330 // fold up, do so.
1331 Node* prev_mem = find_previous_store(phase);
1332 // Steps (a), (b): Walk past independent stores to find an exact match.
1333 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1334 // (c) See if we can fold up on the spot, but don't fold up here.
1335 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
1336 // just return a prior value, which is done by Identity calls.
1337 if (can_see_stored_value(prev_mem, phase)) {
1338 // Make ready for step (d):
1339 set_req(MemNode::Memory, prev_mem);
1340 return this;
1341 }
1342 }
1344 return NULL; // No further progress
1345 }
1347 // Helper to recognize certain Klass fields which are invariant across
1348 // some group of array types (e.g., int[] or all T[] where T < Object).
1349 const Type*
1350 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1351 ciKlass* klass) const {
1352 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1353 // The field is Klass::_modifier_flags. Return its (constant) value.
1354 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1355 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1356 return TypeInt::make(klass->modifier_flags());
1357 }
1358 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1359 // The field is Klass::_access_flags. Return its (constant) value.
1360 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1361 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1362 return TypeInt::make(klass->access_flags());
1363 }
1364 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1365 // The field is Klass::_layout_helper. Return its constant value if known.
1366 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1367 return TypeInt::make(klass->layout_helper());
1368 }
1370 // No match.
1371 return NULL;
1372 }
1374 //------------------------------Value-----------------------------------------
1375 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1376 // Either input is TOP ==> the result is TOP
1377 Node* mem = in(MemNode::Memory);
1378 const Type *t1 = phase->type(mem);
1379 if (t1 == Type::TOP) return Type::TOP;
1380 Node* adr = in(MemNode::Address);
1381 const TypePtr* tp = phase->type(adr)->isa_ptr();
1382 if (tp == NULL || tp->empty()) return Type::TOP;
1383 int off = tp->offset();
1384 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1386 // Try to guess loaded type from pointer type
1387 if (tp->base() == Type::AryPtr) {
1388 const Type *t = tp->is_aryptr()->elem();
1389 // Don't do this for integer types. There is only potential profit if
1390 // the element type t is lower than _type; that is, for int types, if _type is
1391 // more restrictive than t. This only happens here if one is short and the other
1392 // char (both 16 bits), and in those cases we've made an intentional decision
1393 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1394 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1395 //
1396 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1397 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1398 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1399 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1400 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1401 // In fact, that could have been the original type of p1, and p1 could have
1402 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1403 // expression (LShiftL quux 3) independently optimized to the constant 8.
1404 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1405 && Opcode() != Op_LoadKlass) {
1406 // t might actually be lower than _type, if _type is a unique
1407 // concrete subclass of abstract class t.
1408 // Make sure the reference is not into the header, by comparing
1409 // the offset against the offset of the start of the array's data.
1410 // Different array types begin at slightly different offsets (12 vs. 16).
1411 // We choose T_BYTE as an example base type that is least restrictive
1412 // as to alignment, which will therefore produce the smallest
1413 // possible base offset.
1414 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1415 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1416 const Type* jt = t->join(_type);
1417 // In any case, do not allow the join, per se, to empty out the type.
1418 if (jt->empty() && !t->empty()) {
1419 // This can happen if a interface-typed array narrows to a class type.
1420 jt = _type;
1421 }
1423 if (EliminateAutoBox) {
1424 // The pointers in the autobox arrays are always non-null
1425 Node* base = in(Address)->in(AddPNode::Base);
1426 if (base != NULL) {
1427 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1428 if (is_autobox_cache(atp)) {
1429 return jt->join(TypePtr::NOTNULL)->is_ptr();
1430 }
1431 }
1432 }
1433 return jt;
1434 }
1435 }
1436 } else if (tp->base() == Type::InstPtr) {
1437 assert( off != Type::OffsetBot ||
1438 // arrays can be cast to Objects
1439 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1440 // unsafe field access may not have a constant offset
1441 phase->C->has_unsafe_access(),
1442 "Field accesses must be precise" );
1443 // For oop loads, we expect the _type to be precise
1444 } else if (tp->base() == Type::KlassPtr) {
1445 assert( off != Type::OffsetBot ||
1446 // arrays can be cast to Objects
1447 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1448 // also allow array-loading from the primary supertype
1449 // array during subtype checks
1450 Opcode() == Op_LoadKlass,
1451 "Field accesses must be precise" );
1452 // For klass/static loads, we expect the _type to be precise
1453 }
1455 const TypeKlassPtr *tkls = tp->isa_klassptr();
1456 if (tkls != NULL && !StressReflectiveCode) {
1457 ciKlass* klass = tkls->klass();
1458 if (klass->is_loaded() && tkls->klass_is_exact()) {
1459 // We are loading a field from a Klass metaobject whose identity
1460 // is known at compile time (the type is "exact" or "precise").
1461 // Check for fields we know are maintained as constants by the VM.
1462 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1463 // The field is Klass::_super_check_offset. Return its (constant) value.
1464 // (Folds up type checking code.)
1465 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1466 return TypeInt::make(klass->super_check_offset());
1467 }
1468 // Compute index into primary_supers array
1469 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1470 // Check for overflowing; use unsigned compare to handle the negative case.
1471 if( depth < ciKlass::primary_super_limit() ) {
1472 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1473 // (Folds up type checking code.)
1474 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1475 ciKlass *ss = klass->super_of_depth(depth);
1476 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1477 }
1478 const Type* aift = load_array_final_field(tkls, klass);
1479 if (aift != NULL) return aift;
1480 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1481 && klass->is_array_klass()) {
1482 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1483 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1484 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1485 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1486 }
1487 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1488 // The field is Klass::_java_mirror. Return its (constant) value.
1489 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1490 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1491 return TypeInstPtr::make(klass->java_mirror());
1492 }
1493 }
1495 // We can still check if we are loading from the primary_supers array at a
1496 // shallow enough depth. Even though the klass is not exact, entries less
1497 // than or equal to its super depth are correct.
1498 if (klass->is_loaded() ) {
1499 ciType *inner = klass->klass();
1500 while( inner->is_obj_array_klass() )
1501 inner = inner->as_obj_array_klass()->base_element_type();
1502 if( inner->is_instance_klass() &&
1503 !inner->as_instance_klass()->flags().is_interface() ) {
1504 // Compute index into primary_supers array
1505 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1506 // Check for overflowing; use unsigned compare to handle the negative case.
1507 if( depth < ciKlass::primary_super_limit() &&
1508 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1509 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1510 // (Folds up type checking code.)
1511 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1512 ciKlass *ss = klass->super_of_depth(depth);
1513 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1514 }
1515 }
1516 }
1518 // If the type is enough to determine that the thing is not an array,
1519 // we can give the layout_helper a positive interval type.
1520 // This will help short-circuit some reflective code.
1521 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1522 && !klass->is_array_klass() // not directly typed as an array
1523 && !klass->is_interface() // specifically not Serializable & Cloneable
1524 && !klass->is_java_lang_Object() // not the supertype of all T[]
1525 ) {
1526 // Note: When interfaces are reliable, we can narrow the interface
1527 // test to (klass != Serializable && klass != Cloneable).
1528 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1529 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1530 // The key property of this type is that it folds up tests
1531 // for array-ness, since it proves that the layout_helper is positive.
1532 // Thus, a generic value like the basic object layout helper works fine.
1533 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1534 }
1535 }
1537 // If we are loading from a freshly-allocated object, produce a zero,
1538 // if the load is provably beyond the header of the object.
1539 // (Also allow a variable load from a fresh array to produce zero.)
1540 if (ReduceFieldZeroing) {
1541 Node* value = can_see_stored_value(mem,phase);
1542 if (value != NULL && value->is_Con())
1543 return value->bottom_type();
1544 }
1546 const TypeOopPtr *tinst = tp->isa_oopptr();
1547 if (tinst != NULL && tinst->is_known_instance_field()) {
1548 // If we have an instance type and our memory input is the
1549 // programs's initial memory state, there is no matching store,
1550 // so just return a zero of the appropriate type
1551 Node *mem = in(MemNode::Memory);
1552 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1553 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1554 return Type::get_zero_type(_type->basic_type());
1555 }
1556 }
1557 return _type;
1558 }
1560 //------------------------------match_edge-------------------------------------
1561 // Do we Match on this edge index or not? Match only the address.
1562 uint LoadNode::match_edge(uint idx) const {
1563 return idx == MemNode::Address;
1564 }
1566 //--------------------------LoadBNode::Ideal--------------------------------------
1567 //
1568 // If the previous store is to the same address as this load,
1569 // and the value stored was larger than a byte, replace this load
1570 // with the value stored truncated to a byte. If no truncation is
1571 // needed, the replacement is done in LoadNode::Identity().
1572 //
1573 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1574 Node* mem = in(MemNode::Memory);
1575 Node* value = can_see_stored_value(mem,phase);
1576 if( value && !phase->type(value)->higher_equal( _type ) ) {
1577 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1578 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1579 }
1580 // Identity call will handle the case where truncation is not needed.
1581 return LoadNode::Ideal(phase, can_reshape);
1582 }
1584 //--------------------------LoadCNode::Ideal--------------------------------------
1585 //
1586 // If the previous store is to the same address as this load,
1587 // and the value stored was larger than a char, replace this load
1588 // with the value stored truncated to a char. If no truncation is
1589 // needed, the replacement is done in LoadNode::Identity().
1590 //
1591 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1592 Node* mem = in(MemNode::Memory);
1593 Node* value = can_see_stored_value(mem,phase);
1594 if( value && !phase->type(value)->higher_equal( _type ) )
1595 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1596 // Identity call will handle the case where truncation is not needed.
1597 return LoadNode::Ideal(phase, can_reshape);
1598 }
1600 //--------------------------LoadSNode::Ideal--------------------------------------
1601 //
1602 // If the previous store is to the same address as this load,
1603 // and the value stored was larger than a short, replace this load
1604 // with the value stored truncated to a short. If no truncation is
1605 // needed, the replacement is done in LoadNode::Identity().
1606 //
1607 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1608 Node* mem = in(MemNode::Memory);
1609 Node* value = can_see_stored_value(mem,phase);
1610 if( value && !phase->type(value)->higher_equal( _type ) ) {
1611 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1612 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1613 }
1614 // Identity call will handle the case where truncation is not needed.
1615 return LoadNode::Ideal(phase, can_reshape);
1616 }
1618 //=============================================================================
1619 //----------------------------LoadKlassNode::make------------------------------
1620 // Polymorphic factory method:
1621 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1622 Compile* C = gvn.C;
1623 Node *ctl = NULL;
1624 // sanity check the alias category against the created node type
1625 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
1626 assert(adr_type != NULL, "expecting TypeOopPtr");
1627 #ifdef _LP64
1628 if (adr_type->is_ptr_to_narrowoop()) {
1629 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
1630 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
1631 }
1632 #endif
1633 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1634 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
1635 }
1637 //------------------------------Value------------------------------------------
1638 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1639 return klass_value_common(phase);
1640 }
1642 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1643 // Either input is TOP ==> the result is TOP
1644 const Type *t1 = phase->type( in(MemNode::Memory) );
1645 if (t1 == Type::TOP) return Type::TOP;
1646 Node *adr = in(MemNode::Address);
1647 const Type *t2 = phase->type( adr );
1648 if (t2 == Type::TOP) return Type::TOP;
1649 const TypePtr *tp = t2->is_ptr();
1650 if (TypePtr::above_centerline(tp->ptr()) ||
1651 tp->ptr() == TypePtr::Null) return Type::TOP;
1653 // Return a more precise klass, if possible
1654 const TypeInstPtr *tinst = tp->isa_instptr();
1655 if (tinst != NULL) {
1656 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1657 int offset = tinst->offset();
1658 if (ik == phase->C->env()->Class_klass()
1659 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1660 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1661 // We are loading a special hidden field from a Class mirror object,
1662 // the field which points to the VM's Klass metaobject.
1663 ciType* t = tinst->java_mirror_type();
1664 // java_mirror_type returns non-null for compile-time Class constants.
1665 if (t != NULL) {
1666 // constant oop => constant klass
1667 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1668 return TypeKlassPtr::make(ciArrayKlass::make(t));
1669 }
1670 if (!t->is_klass()) {
1671 // a primitive Class (e.g., int.class) has NULL for a klass field
1672 return TypePtr::NULL_PTR;
1673 }
1674 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1675 return TypeKlassPtr::make(t->as_klass());
1676 }
1677 // non-constant mirror, so we can't tell what's going on
1678 }
1679 if( !ik->is_loaded() )
1680 return _type; // Bail out if not loaded
1681 if (offset == oopDesc::klass_offset_in_bytes()) {
1682 if (tinst->klass_is_exact()) {
1683 return TypeKlassPtr::make(ik);
1684 }
1685 // See if we can become precise: no subklasses and no interface
1686 // (Note: We need to support verified interfaces.)
1687 if (!ik->is_interface() && !ik->has_subklass()) {
1688 //assert(!UseExactTypes, "this code should be useless with exact types");
1689 // Add a dependence; if any subclass added we need to recompile
1690 if (!ik->is_final()) {
1691 // %%% should use stronger assert_unique_concrete_subtype instead
1692 phase->C->dependencies()->assert_leaf_type(ik);
1693 }
1694 // Return precise klass
1695 return TypeKlassPtr::make(ik);
1696 }
1698 // Return root of possible klass
1699 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1700 }
1701 }
1703 // Check for loading klass from an array
1704 const TypeAryPtr *tary = tp->isa_aryptr();
1705 if( tary != NULL ) {
1706 ciKlass *tary_klass = tary->klass();
1707 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1708 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1709 if (tary->klass_is_exact()) {
1710 return TypeKlassPtr::make(tary_klass);
1711 }
1712 ciArrayKlass *ak = tary->klass()->as_array_klass();
1713 // If the klass is an object array, we defer the question to the
1714 // array component klass.
1715 if( ak->is_obj_array_klass() ) {
1716 assert( ak->is_loaded(), "" );
1717 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1718 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1719 ciInstanceKlass* ik = base_k->as_instance_klass();
1720 // See if we can become precise: no subklasses and no interface
1721 if (!ik->is_interface() && !ik->has_subklass()) {
1722 //assert(!UseExactTypes, "this code should be useless with exact types");
1723 // Add a dependence; if any subclass added we need to recompile
1724 if (!ik->is_final()) {
1725 phase->C->dependencies()->assert_leaf_type(ik);
1726 }
1727 // Return precise array klass
1728 return TypeKlassPtr::make(ak);
1729 }
1730 }
1731 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1732 } else { // Found a type-array?
1733 //assert(!UseExactTypes, "this code should be useless with exact types");
1734 assert( ak->is_type_array_klass(), "" );
1735 return TypeKlassPtr::make(ak); // These are always precise
1736 }
1737 }
1738 }
1740 // Check for loading klass from an array klass
1741 const TypeKlassPtr *tkls = tp->isa_klassptr();
1742 if (tkls != NULL && !StressReflectiveCode) {
1743 ciKlass* klass = tkls->klass();
1744 if( !klass->is_loaded() )
1745 return _type; // Bail out if not loaded
1746 if( klass->is_obj_array_klass() &&
1747 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1748 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1749 // // Always returning precise element type is incorrect,
1750 // // e.g., element type could be object and array may contain strings
1751 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1753 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1754 // according to the element type's subclassing.
1755 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1756 }
1757 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1758 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1759 ciKlass* sup = klass->as_instance_klass()->super();
1760 // The field is Klass::_super. Return its (constant) value.
1761 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1762 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1763 }
1764 }
1766 // Bailout case
1767 return LoadNode::Value(phase);
1768 }
1770 //------------------------------Identity---------------------------------------
1771 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1772 // Also feed through the klass in Allocate(...klass...)._klass.
1773 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1774 return klass_identity_common(phase);
1775 }
1777 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
1778 Node* x = LoadNode::Identity(phase);
1779 if (x != this) return x;
1781 // Take apart the address into an oop and and offset.
1782 // Return 'this' if we cannot.
1783 Node* adr = in(MemNode::Address);
1784 intptr_t offset = 0;
1785 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1786 if (base == NULL) return this;
1787 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1788 if (toop == NULL) return this;
1790 // We can fetch the klass directly through an AllocateNode.
1791 // This works even if the klass is not constant (clone or newArray).
1792 if (offset == oopDesc::klass_offset_in_bytes()) {
1793 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1794 if (allocated_klass != NULL) {
1795 return allocated_klass;
1796 }
1797 }
1799 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1800 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1801 // See inline_native_Class_query for occurrences of these patterns.
1802 // Java Example: x.getClass().isAssignableFrom(y)
1803 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1804 //
1805 // This improves reflective code, often making the Class
1806 // mirror go completely dead. (Current exception: Class
1807 // mirrors may appear in debug info, but we could clean them out by
1808 // introducing a new debug info operator for klassOop.java_mirror).
1809 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1810 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1811 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1812 // We are loading a special hidden field from a Class mirror,
1813 // the field which points to its Klass or arrayKlass metaobject.
1814 if (base->is_Load()) {
1815 Node* adr2 = base->in(MemNode::Address);
1816 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1817 if (tkls != NULL && !tkls->empty()
1818 && (tkls->klass()->is_instance_klass() ||
1819 tkls->klass()->is_array_klass())
1820 && adr2->is_AddP()
1821 ) {
1822 int mirror_field = Klass::java_mirror_offset_in_bytes();
1823 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1824 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1825 }
1826 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1827 return adr2->in(AddPNode::Base);
1828 }
1829 }
1830 }
1831 }
1833 return this;
1834 }
1837 //------------------------------Value------------------------------------------
1838 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
1839 const Type *t = klass_value_common(phase);
1840 if (t == Type::TOP)
1841 return t;
1843 return t->make_narrowoop();
1844 }
1846 //------------------------------Identity---------------------------------------
1847 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
1848 // Also feed through the klass in Allocate(...klass...)._klass.
1849 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
1850 Node *x = klass_identity_common(phase);
1852 const Type *t = phase->type( x );
1853 if( t == Type::TOP ) return x;
1854 if( t->isa_narrowoop()) return x;
1856 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
1857 }
1859 //------------------------------Value-----------------------------------------
1860 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1861 // Either input is TOP ==> the result is TOP
1862 const Type *t1 = phase->type( in(MemNode::Memory) );
1863 if( t1 == Type::TOP ) return Type::TOP;
1864 Node *adr = in(MemNode::Address);
1865 const Type *t2 = phase->type( adr );
1866 if( t2 == Type::TOP ) return Type::TOP;
1867 const TypePtr *tp = t2->is_ptr();
1868 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1869 const TypeAryPtr *tap = tp->isa_aryptr();
1870 if( !tap ) return _type;
1871 return tap->size();
1872 }
1874 //------------------------------Identity---------------------------------------
1875 // Feed through the length in AllocateArray(...length...)._length.
1876 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1877 Node* x = LoadINode::Identity(phase);
1878 if (x != this) return x;
1880 // Take apart the address into an oop and and offset.
1881 // Return 'this' if we cannot.
1882 Node* adr = in(MemNode::Address);
1883 intptr_t offset = 0;
1884 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1885 if (base == NULL) return this;
1886 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1887 if (tary == NULL) return this;
1889 // We can fetch the length directly through an AllocateArrayNode.
1890 // This works even if the length is not constant (clone or newArray).
1891 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1892 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
1893 if (allocated_length != NULL) {
1894 return allocated_length;
1895 }
1896 }
1898 return this;
1900 }
1901 //=============================================================================
1902 //---------------------------StoreNode::make-----------------------------------
1903 // Polymorphic factory method:
1904 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1905 Compile* C = gvn.C;
1907 switch (bt) {
1908 case T_BOOLEAN:
1909 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1910 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1911 case T_CHAR:
1912 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1913 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1914 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1915 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1916 case T_ADDRESS:
1917 case T_OBJECT:
1918 #ifdef _LP64
1919 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
1920 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
1921 adr->bottom_type()->isa_rawptr())) {
1922 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
1923 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
1924 } else
1925 #endif
1926 {
1927 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1928 }
1929 }
1930 ShouldNotReachHere();
1931 return (StoreNode*)NULL;
1932 }
1934 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1935 bool require_atomic = true;
1936 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1937 }
1940 //--------------------------bottom_type----------------------------------------
1941 const Type *StoreNode::bottom_type() const {
1942 return Type::MEMORY;
1943 }
1945 //------------------------------hash-------------------------------------------
1946 uint StoreNode::hash() const {
1947 // unroll addition of interesting fields
1948 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1950 // Since they are not commoned, do not hash them:
1951 return NO_HASH;
1952 }
1954 //------------------------------Ideal------------------------------------------
1955 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1956 // When a store immediately follows a relevant allocation/initialization,
1957 // try to capture it into the initialization, or hoist it above.
1958 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1959 Node* p = MemNode::Ideal_common(phase, can_reshape);
1960 if (p) return (p == NodeSentinel) ? NULL : p;
1962 Node* mem = in(MemNode::Memory);
1963 Node* address = in(MemNode::Address);
1965 // Back-to-back stores to same address? Fold em up.
1966 // Generally unsafe if I have intervening uses...
1967 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
1968 // Looking at a dead closed cycle of memory?
1969 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1971 assert(Opcode() == mem->Opcode() ||
1972 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1973 "no mismatched stores, except on raw memory");
1975 if (mem->outcnt() == 1 && // check for intervening uses
1976 mem->as_Store()->memory_size() <= this->memory_size()) {
1977 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1978 // For example, 'mem' might be the final state at a conditional return.
1979 // Or, 'mem' might be used by some node which is live at the same time
1980 // 'this' is live, which might be unschedulable. So, require exactly
1981 // ONE user, the 'this' store, until such time as we clone 'mem' for
1982 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1983 if (can_reshape) { // (%%% is this an anachronism?)
1984 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1985 phase->is_IterGVN());
1986 } else {
1987 // It's OK to do this in the parser, since DU info is always accurate,
1988 // and the parser always refers to nodes via SafePointNode maps.
1989 set_req(MemNode::Memory, mem->in(MemNode::Memory));
1990 }
1991 return this;
1992 }
1993 }
1995 // Capture an unaliased, unconditional, simple store into an initializer.
1996 // Or, if it is independent of the allocation, hoist it above the allocation.
1997 if (ReduceFieldZeroing && /*can_reshape &&*/
1998 mem->is_Proj() && mem->in(0)->is_Initialize()) {
1999 InitializeNode* init = mem->in(0)->as_Initialize();
2000 intptr_t offset = init->can_capture_store(this, phase);
2001 if (offset > 0) {
2002 Node* moved = init->capture_store(this, offset, phase);
2003 // If the InitializeNode captured me, it made a raw copy of me,
2004 // and I need to disappear.
2005 if (moved != NULL) {
2006 // %%% hack to ensure that Ideal returns a new node:
2007 mem = MergeMemNode::make(phase->C, mem);
2008 return mem; // fold me away
2009 }
2010 }
2011 }
2013 return NULL; // No further progress
2014 }
2016 //------------------------------Value-----------------------------------------
2017 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2018 // Either input is TOP ==> the result is TOP
2019 const Type *t1 = phase->type( in(MemNode::Memory) );
2020 if( t1 == Type::TOP ) return Type::TOP;
2021 const Type *t2 = phase->type( in(MemNode::Address) );
2022 if( t2 == Type::TOP ) return Type::TOP;
2023 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2024 if( t3 == Type::TOP ) return Type::TOP;
2025 return Type::MEMORY;
2026 }
2028 //------------------------------Identity---------------------------------------
2029 // Remove redundant stores:
2030 // Store(m, p, Load(m, p)) changes to m.
2031 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2032 Node *StoreNode::Identity( PhaseTransform *phase ) {
2033 Node* mem = in(MemNode::Memory);
2034 Node* adr = in(MemNode::Address);
2035 Node* val = in(MemNode::ValueIn);
2037 // Load then Store? Then the Store is useless
2038 if (val->is_Load() &&
2039 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
2040 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
2041 val->as_Load()->store_Opcode() == Opcode()) {
2042 return mem;
2043 }
2045 // Two stores in a row of the same value?
2046 if (mem->is_Store() &&
2047 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
2048 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
2049 mem->Opcode() == Opcode()) {
2050 return mem;
2051 }
2053 // Store of zero anywhere into a freshly-allocated object?
2054 // Then the store is useless.
2055 // (It must already have been captured by the InitializeNode.)
2056 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2057 // a newly allocated object is already all-zeroes everywhere
2058 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2059 return mem;
2060 }
2062 // the store may also apply to zero-bits in an earlier object
2063 Node* prev_mem = find_previous_store(phase);
2064 // Steps (a), (b): Walk past independent stores to find an exact match.
2065 if (prev_mem != NULL) {
2066 Node* prev_val = can_see_stored_value(prev_mem, phase);
2067 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2068 // prev_val and val might differ by a cast; it would be good
2069 // to keep the more informative of the two.
2070 return mem;
2071 }
2072 }
2073 }
2075 return this;
2076 }
2078 //------------------------------match_edge-------------------------------------
2079 // Do we Match on this edge index or not? Match only memory & value
2080 uint StoreNode::match_edge(uint idx) const {
2081 return idx == MemNode::Address || idx == MemNode::ValueIn;
2082 }
2084 //------------------------------cmp--------------------------------------------
2085 // Do not common stores up together. They generally have to be split
2086 // back up anyways, so do not bother.
2087 uint StoreNode::cmp( const Node &n ) const {
2088 return (&n == this); // Always fail except on self
2089 }
2091 //------------------------------Ideal_masked_input-----------------------------
2092 // Check for a useless mask before a partial-word store
2093 // (StoreB ... (AndI valIn conIa) )
2094 // If (conIa & mask == mask) this simplifies to
2095 // (StoreB ... (valIn) )
2096 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2097 Node *val = in(MemNode::ValueIn);
2098 if( val->Opcode() == Op_AndI ) {
2099 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2100 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2101 set_req(MemNode::ValueIn, val->in(1));
2102 return this;
2103 }
2104 }
2105 return NULL;
2106 }
2109 //------------------------------Ideal_sign_extended_input----------------------
2110 // Check for useless sign-extension before a partial-word store
2111 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2112 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2113 // (StoreB ... (valIn) )
2114 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2115 Node *val = in(MemNode::ValueIn);
2116 if( val->Opcode() == Op_RShiftI ) {
2117 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2118 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2119 Node *shl = val->in(1);
2120 if( shl->Opcode() == Op_LShiftI ) {
2121 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2122 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2123 set_req(MemNode::ValueIn, shl->in(1));
2124 return this;
2125 }
2126 }
2127 }
2128 }
2129 return NULL;
2130 }
2132 //------------------------------value_never_loaded-----------------------------------
2133 // Determine whether there are any possible loads of the value stored.
2134 // For simplicity, we actually check if there are any loads from the
2135 // address stored to, not just for loads of the value stored by this node.
2136 //
2137 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2138 Node *adr = in(Address);
2139 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2140 if (adr_oop == NULL)
2141 return false;
2142 if (!adr_oop->is_known_instance_field())
2143 return false; // if not a distinct instance, there may be aliases of the address
2144 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2145 Node *use = adr->fast_out(i);
2146 int opc = use->Opcode();
2147 if (use->is_Load() || use->is_LoadStore()) {
2148 return false;
2149 }
2150 }
2151 return true;
2152 }
2154 //=============================================================================
2155 //------------------------------Ideal------------------------------------------
2156 // If the store is from an AND mask that leaves the low bits untouched, then
2157 // we can skip the AND operation. If the store is from a sign-extension
2158 // (a left shift, then right shift) we can skip both.
2159 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2160 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2161 if( progress != NULL ) return progress;
2163 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2164 if( progress != NULL ) return progress;
2166 // Finally check the default case
2167 return StoreNode::Ideal(phase, can_reshape);
2168 }
2170 //=============================================================================
2171 //------------------------------Ideal------------------------------------------
2172 // If the store is from an AND mask that leaves the low bits untouched, then
2173 // we can skip the AND operation
2174 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2175 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2176 if( progress != NULL ) return progress;
2178 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2179 if( progress != NULL ) return progress;
2181 // Finally check the default case
2182 return StoreNode::Ideal(phase, can_reshape);
2183 }
2185 //=============================================================================
2186 //------------------------------Identity---------------------------------------
2187 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2188 // No need to card mark when storing a null ptr
2189 Node* my_store = in(MemNode::OopStore);
2190 if (my_store->is_Store()) {
2191 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2192 if( t1 == TypePtr::NULL_PTR ) {
2193 return in(MemNode::Memory);
2194 }
2195 }
2196 return this;
2197 }
2199 //------------------------------Value-----------------------------------------
2200 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2201 // Either input is TOP ==> the result is TOP
2202 const Type *t = phase->type( in(MemNode::Memory) );
2203 if( t == Type::TOP ) return Type::TOP;
2204 t = phase->type( in(MemNode::Address) );
2205 if( t == Type::TOP ) return Type::TOP;
2206 t = phase->type( in(MemNode::ValueIn) );
2207 if( t == Type::TOP ) return Type::TOP;
2208 // If extra input is TOP ==> the result is TOP
2209 t = phase->type( in(MemNode::OopStore) );
2210 if( t == Type::TOP ) return Type::TOP;
2212 return StoreNode::Value( phase );
2213 }
2216 //=============================================================================
2217 //----------------------------------SCMemProjNode------------------------------
2218 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2219 {
2220 return bottom_type();
2221 }
2223 //=============================================================================
2224 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2225 init_req(MemNode::Control, c );
2226 init_req(MemNode::Memory , mem);
2227 init_req(MemNode::Address, adr);
2228 init_req(MemNode::ValueIn, val);
2229 init_req( ExpectedIn, ex );
2230 init_class_id(Class_LoadStore);
2232 }
2234 //=============================================================================
2235 //-------------------------------adr_type--------------------------------------
2236 // Do we Match on this edge index or not? Do not match memory
2237 const TypePtr* ClearArrayNode::adr_type() const {
2238 Node *adr = in(3);
2239 return MemNode::calculate_adr_type(adr->bottom_type());
2240 }
2242 //------------------------------match_edge-------------------------------------
2243 // Do we Match on this edge index or not? Do not match memory
2244 uint ClearArrayNode::match_edge(uint idx) const {
2245 return idx > 1;
2246 }
2248 //------------------------------Identity---------------------------------------
2249 // Clearing a zero length array does nothing
2250 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2251 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2252 }
2254 //------------------------------Idealize---------------------------------------
2255 // Clearing a short array is faster with stores
2256 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2257 const int unit = BytesPerLong;
2258 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2259 if (!t) return NULL;
2260 if (!t->is_con()) return NULL;
2261 intptr_t raw_count = t->get_con();
2262 intptr_t size = raw_count;
2263 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2264 // Clearing nothing uses the Identity call.
2265 // Negative clears are possible on dead ClearArrays
2266 // (see jck test stmt114.stmt11402.val).
2267 if (size <= 0 || size % unit != 0) return NULL;
2268 intptr_t count = size / unit;
2269 // Length too long; use fast hardware clear
2270 if (size > Matcher::init_array_short_size) return NULL;
2271 Node *mem = in(1);
2272 if( phase->type(mem)==Type::TOP ) return NULL;
2273 Node *adr = in(3);
2274 const Type* at = phase->type(adr);
2275 if( at==Type::TOP ) return NULL;
2276 const TypePtr* atp = at->isa_ptr();
2277 // adjust atp to be the correct array element address type
2278 if (atp == NULL) atp = TypePtr::BOTTOM;
2279 else atp = atp->add_offset(Type::OffsetBot);
2280 // Get base for derived pointer purposes
2281 if( adr->Opcode() != Op_AddP ) Unimplemented();
2282 Node *base = adr->in(1);
2284 Node *zero = phase->makecon(TypeLong::ZERO);
2285 Node *off = phase->MakeConX(BytesPerLong);
2286 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2287 count--;
2288 while( count-- ) {
2289 mem = phase->transform(mem);
2290 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2291 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2292 }
2293 return mem;
2294 }
2296 //----------------------------clear_memory-------------------------------------
2297 // Generate code to initialize object storage to zero.
2298 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2299 intptr_t start_offset,
2300 Node* end_offset,
2301 PhaseGVN* phase) {
2302 Compile* C = phase->C;
2303 intptr_t offset = start_offset;
2305 int unit = BytesPerLong;
2306 if ((offset % unit) != 0) {
2307 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2308 adr = phase->transform(adr);
2309 const TypePtr* atp = TypeRawPtr::BOTTOM;
2310 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2311 mem = phase->transform(mem);
2312 offset += BytesPerInt;
2313 }
2314 assert((offset % unit) == 0, "");
2316 // Initialize the remaining stuff, if any, with a ClearArray.
2317 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2318 }
2320 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2321 Node* start_offset,
2322 Node* end_offset,
2323 PhaseGVN* phase) {
2324 if (start_offset == end_offset) {
2325 // nothing to do
2326 return mem;
2327 }
2329 Compile* C = phase->C;
2330 int unit = BytesPerLong;
2331 Node* zbase = start_offset;
2332 Node* zend = end_offset;
2334 // Scale to the unit required by the CPU:
2335 if (!Matcher::init_array_count_is_in_bytes) {
2336 Node* shift = phase->intcon(exact_log2(unit));
2337 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2338 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2339 }
2341 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2342 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2344 // Bulk clear double-words
2345 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2346 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2347 return phase->transform(mem);
2348 }
2350 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2351 intptr_t start_offset,
2352 intptr_t end_offset,
2353 PhaseGVN* phase) {
2354 if (start_offset == end_offset) {
2355 // nothing to do
2356 return mem;
2357 }
2359 Compile* C = phase->C;
2360 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2361 intptr_t done_offset = end_offset;
2362 if ((done_offset % BytesPerLong) != 0) {
2363 done_offset -= BytesPerInt;
2364 }
2365 if (done_offset > start_offset) {
2366 mem = clear_memory(ctl, mem, dest,
2367 start_offset, phase->MakeConX(done_offset), phase);
2368 }
2369 if (done_offset < end_offset) { // emit the final 32-bit store
2370 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2371 adr = phase->transform(adr);
2372 const TypePtr* atp = TypeRawPtr::BOTTOM;
2373 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2374 mem = phase->transform(mem);
2375 done_offset += BytesPerInt;
2376 }
2377 assert(done_offset == end_offset, "");
2378 return mem;
2379 }
2381 //=============================================================================
2382 // Do we match on this edge? No memory edges
2383 uint StrCompNode::match_edge(uint idx) const {
2384 return idx == 5 || idx == 6;
2385 }
2387 //------------------------------Ideal------------------------------------------
2388 // Return a node which is more "ideal" than the current node. Strip out
2389 // control copies
2390 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2391 return remove_dead_region(phase, can_reshape) ? this : NULL;
2392 }
2394 //------------------------------Ideal------------------------------------------
2395 // Return a node which is more "ideal" than the current node. Strip out
2396 // control copies
2397 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){
2398 return remove_dead_region(phase, can_reshape) ? this : NULL;
2399 }
2402 //=============================================================================
2403 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2404 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2405 _adr_type(C->get_adr_type(alias_idx))
2406 {
2407 init_class_id(Class_MemBar);
2408 Node* top = C->top();
2409 init_req(TypeFunc::I_O,top);
2410 init_req(TypeFunc::FramePtr,top);
2411 init_req(TypeFunc::ReturnAdr,top);
2412 if (precedent != NULL)
2413 init_req(TypeFunc::Parms, precedent);
2414 }
2416 //------------------------------cmp--------------------------------------------
2417 uint MemBarNode::hash() const { return NO_HASH; }
2418 uint MemBarNode::cmp( const Node &n ) const {
2419 return (&n == this); // Always fail except on self
2420 }
2422 //------------------------------make-------------------------------------------
2423 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2424 int len = Precedent + (pn == NULL? 0: 1);
2425 switch (opcode) {
2426 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2427 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2428 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2429 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2430 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2431 default: ShouldNotReachHere(); return NULL;
2432 }
2433 }
2435 //------------------------------Ideal------------------------------------------
2436 // Return a node which is more "ideal" than the current node. Strip out
2437 // control copies
2438 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2439 if (remove_dead_region(phase, can_reshape)) return this;
2440 return NULL;
2441 }
2443 //------------------------------Value------------------------------------------
2444 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2445 if( !in(0) ) return Type::TOP;
2446 if( phase->type(in(0)) == Type::TOP )
2447 return Type::TOP;
2448 return TypeTuple::MEMBAR;
2449 }
2451 //------------------------------match------------------------------------------
2452 // Construct projections for memory.
2453 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2454 switch (proj->_con) {
2455 case TypeFunc::Control:
2456 case TypeFunc::Memory:
2457 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2458 }
2459 ShouldNotReachHere();
2460 return NULL;
2461 }
2463 //===========================InitializeNode====================================
2464 // SUMMARY:
2465 // This node acts as a memory barrier on raw memory, after some raw stores.
2466 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2467 // The Initialize can 'capture' suitably constrained stores as raw inits.
2468 // It can coalesce related raw stores into larger units (called 'tiles').
2469 // It can avoid zeroing new storage for memory units which have raw inits.
2470 // At macro-expansion, it is marked 'complete', and does not optimize further.
2471 //
2472 // EXAMPLE:
2473 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2474 // ctl = incoming control; mem* = incoming memory
2475 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2476 // First allocate uninitialized memory and fill in the header:
2477 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2478 // ctl := alloc.Control; mem* := alloc.Memory*
2479 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2480 // Then initialize to zero the non-header parts of the raw memory block:
2481 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2482 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2483 // After the initialize node executes, the object is ready for service:
2484 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2485 // Suppose its body is immediately initialized as {1,2}:
2486 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2487 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2488 // mem.SLICE(#short[*]) := store2
2489 //
2490 // DETAILS:
2491 // An InitializeNode collects and isolates object initialization after
2492 // an AllocateNode and before the next possible safepoint. As a
2493 // memory barrier (MemBarNode), it keeps critical stores from drifting
2494 // down past any safepoint or any publication of the allocation.
2495 // Before this barrier, a newly-allocated object may have uninitialized bits.
2496 // After this barrier, it may be treated as a real oop, and GC is allowed.
2497 //
2498 // The semantics of the InitializeNode include an implicit zeroing of
2499 // the new object from object header to the end of the object.
2500 // (The object header and end are determined by the AllocateNode.)
2501 //
2502 // Certain stores may be added as direct inputs to the InitializeNode.
2503 // These stores must update raw memory, and they must be to addresses
2504 // derived from the raw address produced by AllocateNode, and with
2505 // a constant offset. They must be ordered by increasing offset.
2506 // The first one is at in(RawStores), the last at in(req()-1).
2507 // Unlike most memory operations, they are not linked in a chain,
2508 // but are displayed in parallel as users of the rawmem output of
2509 // the allocation.
2510 //
2511 // (See comments in InitializeNode::capture_store, which continue
2512 // the example given above.)
2513 //
2514 // When the associated Allocate is macro-expanded, the InitializeNode
2515 // may be rewritten to optimize collected stores. A ClearArrayNode
2516 // may also be created at that point to represent any required zeroing.
2517 // The InitializeNode is then marked 'complete', prohibiting further
2518 // capturing of nearby memory operations.
2519 //
2520 // During macro-expansion, all captured initializations which store
2521 // constant values of 32 bits or smaller are coalesced (if advantagous)
2522 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2523 // initialized in fewer memory operations. Memory words which are
2524 // covered by neither tiles nor non-constant stores are pre-zeroed
2525 // by explicit stores of zero. (The code shape happens to do all
2526 // zeroing first, then all other stores, with both sequences occurring
2527 // in order of ascending offsets.)
2528 //
2529 // Alternatively, code may be inserted between an AllocateNode and its
2530 // InitializeNode, to perform arbitrary initialization of the new object.
2531 // E.g., the object copying intrinsics insert complex data transfers here.
2532 // The initialization must then be marked as 'complete' disable the
2533 // built-in zeroing semantics and the collection of initializing stores.
2534 //
2535 // While an InitializeNode is incomplete, reads from the memory state
2536 // produced by it are optimizable if they match the control edge and
2537 // new oop address associated with the allocation/initialization.
2538 // They return a stored value (if the offset matches) or else zero.
2539 // A write to the memory state, if it matches control and address,
2540 // and if it is to a constant offset, may be 'captured' by the
2541 // InitializeNode. It is cloned as a raw memory operation and rewired
2542 // inside the initialization, to the raw oop produced by the allocation.
2543 // Operations on addresses which are provably distinct (e.g., to
2544 // other AllocateNodes) are allowed to bypass the initialization.
2545 //
2546 // The effect of all this is to consolidate object initialization
2547 // (both arrays and non-arrays, both piecewise and bulk) into a
2548 // single location, where it can be optimized as a unit.
2549 //
2550 // Only stores with an offset less than TrackedInitializationLimit words
2551 // will be considered for capture by an InitializeNode. This puts a
2552 // reasonable limit on the complexity of optimized initializations.
2554 //---------------------------InitializeNode------------------------------------
2555 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2556 : _is_complete(false),
2557 MemBarNode(C, adr_type, rawoop)
2558 {
2559 init_class_id(Class_Initialize);
2561 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2562 assert(in(RawAddress) == rawoop, "proper init");
2563 // Note: allocation() can be NULL, for secondary initialization barriers
2564 }
2566 // Since this node is not matched, it will be processed by the
2567 // register allocator. Declare that there are no constraints
2568 // on the allocation of the RawAddress edge.
2569 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2570 // This edge should be set to top, by the set_complete. But be conservative.
2571 if (idx == InitializeNode::RawAddress)
2572 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2573 return RegMask::Empty;
2574 }
2576 Node* InitializeNode::memory(uint alias_idx) {
2577 Node* mem = in(Memory);
2578 if (mem->is_MergeMem()) {
2579 return mem->as_MergeMem()->memory_at(alias_idx);
2580 } else {
2581 // incoming raw memory is not split
2582 return mem;
2583 }
2584 }
2586 bool InitializeNode::is_non_zero() {
2587 if (is_complete()) return false;
2588 remove_extra_zeroes();
2589 return (req() > RawStores);
2590 }
2592 void InitializeNode::set_complete(PhaseGVN* phase) {
2593 assert(!is_complete(), "caller responsibility");
2594 _is_complete = true;
2596 // After this node is complete, it contains a bunch of
2597 // raw-memory initializations. There is no need for
2598 // it to have anything to do with non-raw memory effects.
2599 // Therefore, tell all non-raw users to re-optimize themselves,
2600 // after skipping the memory effects of this initialization.
2601 PhaseIterGVN* igvn = phase->is_IterGVN();
2602 if (igvn) igvn->add_users_to_worklist(this);
2603 }
2605 // convenience function
2606 // return false if the init contains any stores already
2607 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2608 InitializeNode* init = initialization();
2609 if (init == NULL || init->is_complete()) return false;
2610 init->remove_extra_zeroes();
2611 // for now, if this allocation has already collected any inits, bail:
2612 if (init->is_non_zero()) return false;
2613 init->set_complete(phase);
2614 return true;
2615 }
2617 void InitializeNode::remove_extra_zeroes() {
2618 if (req() == RawStores) return;
2619 Node* zmem = zero_memory();
2620 uint fill = RawStores;
2621 for (uint i = fill; i < req(); i++) {
2622 Node* n = in(i);
2623 if (n->is_top() || n == zmem) continue; // skip
2624 if (fill < i) set_req(fill, n); // compact
2625 ++fill;
2626 }
2627 // delete any empty spaces created:
2628 while (fill < req()) {
2629 del_req(fill);
2630 }
2631 }
2633 // Helper for remembering which stores go with which offsets.
2634 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2635 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2636 intptr_t offset = -1;
2637 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2638 phase, offset);
2639 if (base == NULL) return -1; // something is dead,
2640 if (offset < 0) return -1; // dead, dead
2641 return offset;
2642 }
2644 // Helper for proving that an initialization expression is
2645 // "simple enough" to be folded into an object initialization.
2646 // Attempts to prove that a store's initial value 'n' can be captured
2647 // within the initialization without creating a vicious cycle, such as:
2648 // { Foo p = new Foo(); p.next = p; }
2649 // True for constants and parameters and small combinations thereof.
2650 bool InitializeNode::detect_init_independence(Node* n,
2651 bool st_is_pinned,
2652 int& count) {
2653 if (n == NULL) return true; // (can this really happen?)
2654 if (n->is_Proj()) n = n->in(0);
2655 if (n == this) return false; // found a cycle
2656 if (n->is_Con()) return true;
2657 if (n->is_Start()) return true; // params, etc., are OK
2658 if (n->is_Root()) return true; // even better
2660 Node* ctl = n->in(0);
2661 if (ctl != NULL && !ctl->is_top()) {
2662 if (ctl->is_Proj()) ctl = ctl->in(0);
2663 if (ctl == this) return false;
2665 // If we already know that the enclosing memory op is pinned right after
2666 // the init, then any control flow that the store has picked up
2667 // must have preceded the init, or else be equal to the init.
2668 // Even after loop optimizations (which might change control edges)
2669 // a store is never pinned *before* the availability of its inputs.
2670 if (!MemNode::all_controls_dominate(n, this))
2671 return false; // failed to prove a good control
2673 }
2675 // Check data edges for possible dependencies on 'this'.
2676 if ((count += 1) > 20) return false; // complexity limit
2677 for (uint i = 1; i < n->req(); i++) {
2678 Node* m = n->in(i);
2679 if (m == NULL || m == n || m->is_top()) continue;
2680 uint first_i = n->find_edge(m);
2681 if (i != first_i) continue; // process duplicate edge just once
2682 if (!detect_init_independence(m, st_is_pinned, count)) {
2683 return false;
2684 }
2685 }
2687 return true;
2688 }
2690 // Here are all the checks a Store must pass before it can be moved into
2691 // an initialization. Returns zero if a check fails.
2692 // On success, returns the (constant) offset to which the store applies,
2693 // within the initialized memory.
2694 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2695 const int FAIL = 0;
2696 if (st->req() != MemNode::ValueIn + 1)
2697 return FAIL; // an inscrutable StoreNode (card mark?)
2698 Node* ctl = st->in(MemNode::Control);
2699 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2700 return FAIL; // must be unconditional after the initialization
2701 Node* mem = st->in(MemNode::Memory);
2702 if (!(mem->is_Proj() && mem->in(0) == this))
2703 return FAIL; // must not be preceded by other stores
2704 Node* adr = st->in(MemNode::Address);
2705 intptr_t offset;
2706 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2707 if (alloc == NULL)
2708 return FAIL; // inscrutable address
2709 if (alloc != allocation())
2710 return FAIL; // wrong allocation! (store needs to float up)
2711 Node* val = st->in(MemNode::ValueIn);
2712 int complexity_count = 0;
2713 if (!detect_init_independence(val, true, complexity_count))
2714 return FAIL; // stored value must be 'simple enough'
2716 return offset; // success
2717 }
2719 // Find the captured store in(i) which corresponds to the range
2720 // [start..start+size) in the initialized object.
2721 // If there is one, return its index i. If there isn't, return the
2722 // negative of the index where it should be inserted.
2723 // Return 0 if the queried range overlaps an initialization boundary
2724 // or if dead code is encountered.
2725 // If size_in_bytes is zero, do not bother with overlap checks.
2726 int InitializeNode::captured_store_insertion_point(intptr_t start,
2727 int size_in_bytes,
2728 PhaseTransform* phase) {
2729 const int FAIL = 0, MAX_STORE = BytesPerLong;
2731 if (is_complete())
2732 return FAIL; // arraycopy got here first; punt
2734 assert(allocation() != NULL, "must be present");
2736 // no negatives, no header fields:
2737 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2739 // after a certain size, we bail out on tracking all the stores:
2740 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2741 if (start >= ti_limit) return FAIL;
2743 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2744 if (i >= limit) return -(int)i; // not found; here is where to put it
2746 Node* st = in(i);
2747 intptr_t st_off = get_store_offset(st, phase);
2748 if (st_off < 0) {
2749 if (st != zero_memory()) {
2750 return FAIL; // bail out if there is dead garbage
2751 }
2752 } else if (st_off > start) {
2753 // ...we are done, since stores are ordered
2754 if (st_off < start + size_in_bytes) {
2755 return FAIL; // the next store overlaps
2756 }
2757 return -(int)i; // not found; here is where to put it
2758 } else if (st_off < start) {
2759 if (size_in_bytes != 0 &&
2760 start < st_off + MAX_STORE &&
2761 start < st_off + st->as_Store()->memory_size()) {
2762 return FAIL; // the previous store overlaps
2763 }
2764 } else {
2765 if (size_in_bytes != 0 &&
2766 st->as_Store()->memory_size() != size_in_bytes) {
2767 return FAIL; // mismatched store size
2768 }
2769 return i;
2770 }
2772 ++i;
2773 }
2774 }
2776 // Look for a captured store which initializes at the offset 'start'
2777 // with the given size. If there is no such store, and no other
2778 // initialization interferes, then return zero_memory (the memory
2779 // projection of the AllocateNode).
2780 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2781 PhaseTransform* phase) {
2782 assert(stores_are_sane(phase), "");
2783 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2784 if (i == 0) {
2785 return NULL; // something is dead
2786 } else if (i < 0) {
2787 return zero_memory(); // just primordial zero bits here
2788 } else {
2789 Node* st = in(i); // here is the store at this position
2790 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2791 return st;
2792 }
2793 }
2795 // Create, as a raw pointer, an address within my new object at 'offset'.
2796 Node* InitializeNode::make_raw_address(intptr_t offset,
2797 PhaseTransform* phase) {
2798 Node* addr = in(RawAddress);
2799 if (offset != 0) {
2800 Compile* C = phase->C;
2801 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2802 phase->MakeConX(offset)) );
2803 }
2804 return addr;
2805 }
2807 // Clone the given store, converting it into a raw store
2808 // initializing a field or element of my new object.
2809 // Caller is responsible for retiring the original store,
2810 // with subsume_node or the like.
2811 //
2812 // From the example above InitializeNode::InitializeNode,
2813 // here are the old stores to be captured:
2814 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2815 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2816 //
2817 // Here is the changed code; note the extra edges on init:
2818 // alloc = (Allocate ...)
2819 // rawoop = alloc.RawAddress
2820 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2821 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2822 // init = (Initialize alloc.Control alloc.Memory rawoop
2823 // rawstore1 rawstore2)
2824 //
2825 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2826 PhaseTransform* phase) {
2827 assert(stores_are_sane(phase), "");
2829 if (start < 0) return NULL;
2830 assert(can_capture_store(st, phase) == start, "sanity");
2832 Compile* C = phase->C;
2833 int size_in_bytes = st->memory_size();
2834 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2835 if (i == 0) return NULL; // bail out
2836 Node* prev_mem = NULL; // raw memory for the captured store
2837 if (i > 0) {
2838 prev_mem = in(i); // there is a pre-existing store under this one
2839 set_req(i, C->top()); // temporarily disconnect it
2840 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2841 } else {
2842 i = -i; // no pre-existing store
2843 prev_mem = zero_memory(); // a slice of the newly allocated object
2844 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2845 set_req(--i, C->top()); // reuse this edge; it has been folded away
2846 else
2847 ins_req(i, C->top()); // build a new edge
2848 }
2849 Node* new_st = st->clone();
2850 new_st->set_req(MemNode::Control, in(Control));
2851 new_st->set_req(MemNode::Memory, prev_mem);
2852 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2853 new_st = phase->transform(new_st);
2855 // At this point, new_st might have swallowed a pre-existing store
2856 // at the same offset, or perhaps new_st might have disappeared,
2857 // if it redundantly stored the same value (or zero to fresh memory).
2859 // In any case, wire it in:
2860 set_req(i, new_st);
2862 // The caller may now kill the old guy.
2863 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2864 assert(check_st == new_st || check_st == NULL, "must be findable");
2865 assert(!is_complete(), "");
2866 return new_st;
2867 }
2869 static bool store_constant(jlong* tiles, int num_tiles,
2870 intptr_t st_off, int st_size,
2871 jlong con) {
2872 if ((st_off & (st_size-1)) != 0)
2873 return false; // strange store offset (assume size==2**N)
2874 address addr = (address)tiles + st_off;
2875 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2876 switch (st_size) {
2877 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2878 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2879 case sizeof(jint): *(jint*) addr = (jint) con; break;
2880 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2881 default: return false; // strange store size (detect size!=2**N here)
2882 }
2883 return true; // return success to caller
2884 }
2886 // Coalesce subword constants into int constants and possibly
2887 // into long constants. The goal, if the CPU permits,
2888 // is to initialize the object with a small number of 64-bit tiles.
2889 // Also, convert floating-point constants to bit patterns.
2890 // Non-constants are not relevant to this pass.
2891 //
2892 // In terms of the running example on InitializeNode::InitializeNode
2893 // and InitializeNode::capture_store, here is the transformation
2894 // of rawstore1 and rawstore2 into rawstore12:
2895 // alloc = (Allocate ...)
2896 // rawoop = alloc.RawAddress
2897 // tile12 = 0x00010002
2898 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2899 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2900 //
2901 void
2902 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2903 Node* size_in_bytes,
2904 PhaseGVN* phase) {
2905 Compile* C = phase->C;
2907 assert(stores_are_sane(phase), "");
2908 // Note: After this pass, they are not completely sane,
2909 // since there may be some overlaps.
2911 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2913 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2914 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2915 size_limit = MIN2(size_limit, ti_limit);
2916 size_limit = align_size_up(size_limit, BytesPerLong);
2917 int num_tiles = size_limit / BytesPerLong;
2919 // allocate space for the tile map:
2920 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2921 jlong tiles_buf[small_len];
2922 Node* nodes_buf[small_len];
2923 jlong inits_buf[small_len];
2924 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2925 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2926 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2927 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2928 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2929 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2930 // tiles: exact bitwise model of all primitive constants
2931 // nodes: last constant-storing node subsumed into the tiles model
2932 // inits: which bytes (in each tile) are touched by any initializations
2934 //// Pass A: Fill in the tile model with any relevant stores.
2936 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2937 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2938 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2939 Node* zmem = zero_memory(); // initially zero memory state
2940 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2941 Node* st = in(i);
2942 intptr_t st_off = get_store_offset(st, phase);
2944 // Figure out the store's offset and constant value:
2945 if (st_off < header_size) continue; //skip (ignore header)
2946 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
2947 int st_size = st->as_Store()->memory_size();
2948 if (st_off + st_size > size_limit) break;
2950 // Record which bytes are touched, whether by constant or not.
2951 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
2952 continue; // skip (strange store size)
2954 const Type* val = phase->type(st->in(MemNode::ValueIn));
2955 if (!val->singleton()) continue; //skip (non-con store)
2956 BasicType type = val->basic_type();
2958 jlong con = 0;
2959 switch (type) {
2960 case T_INT: con = val->is_int()->get_con(); break;
2961 case T_LONG: con = val->is_long()->get_con(); break;
2962 case T_FLOAT: con = jint_cast(val->getf()); break;
2963 case T_DOUBLE: con = jlong_cast(val->getd()); break;
2964 default: continue; //skip (odd store type)
2965 }
2967 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
2968 st->Opcode() == Op_StoreL) {
2969 continue; // This StoreL is already optimal.
2970 }
2972 // Store down the constant.
2973 store_constant(tiles, num_tiles, st_off, st_size, con);
2975 intptr_t j = st_off >> LogBytesPerLong;
2977 if (type == T_INT && st_size == BytesPerInt
2978 && (st_off & BytesPerInt) == BytesPerInt) {
2979 jlong lcon = tiles[j];
2980 if (!Matcher::isSimpleConstant64(lcon) &&
2981 st->Opcode() == Op_StoreI) {
2982 // This StoreI is already optimal by itself.
2983 jint* intcon = (jint*) &tiles[j];
2984 intcon[1] = 0; // undo the store_constant()
2986 // If the previous store is also optimal by itself, back up and
2987 // undo the action of the previous loop iteration... if we can.
2988 // But if we can't, just let the previous half take care of itself.
2989 st = nodes[j];
2990 st_off -= BytesPerInt;
2991 con = intcon[0];
2992 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
2993 assert(st_off >= header_size, "still ignoring header");
2994 assert(get_store_offset(st, phase) == st_off, "must be");
2995 assert(in(i-1) == zmem, "must be");
2996 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
2997 assert(con == tcon->is_int()->get_con(), "must be");
2998 // Undo the effects of the previous loop trip, which swallowed st:
2999 intcon[0] = 0; // undo store_constant()
3000 set_req(i-1, st); // undo set_req(i, zmem)
3001 nodes[j] = NULL; // undo nodes[j] = st
3002 --old_subword; // undo ++old_subword
3003 }
3004 continue; // This StoreI is already optimal.
3005 }
3006 }
3008 // This store is not needed.
3009 set_req(i, zmem);
3010 nodes[j] = st; // record for the moment
3011 if (st_size < BytesPerLong) // something has changed
3012 ++old_subword; // includes int/float, but who's counting...
3013 else ++old_long;
3014 }
3016 if ((old_subword + old_long) == 0)
3017 return; // nothing more to do
3019 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3020 // Be sure to insert them before overlapping non-constant stores.
3021 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3022 for (int j = 0; j < num_tiles; j++) {
3023 jlong con = tiles[j];
3024 jlong init = inits[j];
3025 if (con == 0) continue;
3026 jint con0, con1; // split the constant, address-wise
3027 jint init0, init1; // split the init map, address-wise
3028 { union { jlong con; jint intcon[2]; } u;
3029 u.con = con;
3030 con0 = u.intcon[0];
3031 con1 = u.intcon[1];
3032 u.con = init;
3033 init0 = u.intcon[0];
3034 init1 = u.intcon[1];
3035 }
3037 Node* old = nodes[j];
3038 assert(old != NULL, "need the prior store");
3039 intptr_t offset = (j * BytesPerLong);
3041 bool split = !Matcher::isSimpleConstant64(con);
3043 if (offset < header_size) {
3044 assert(offset + BytesPerInt >= header_size, "second int counts");
3045 assert(*(jint*)&tiles[j] == 0, "junk in header");
3046 split = true; // only the second word counts
3047 // Example: int a[] = { 42 ... }
3048 } else if (con0 == 0 && init0 == -1) {
3049 split = true; // first word is covered by full inits
3050 // Example: int a[] = { ... foo(), 42 ... }
3051 } else if (con1 == 0 && init1 == -1) {
3052 split = true; // second word is covered by full inits
3053 // Example: int a[] = { ... 42, foo() ... }
3054 }
3056 // Here's a case where init0 is neither 0 nor -1:
3057 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3058 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3059 // In this case the tile is not split; it is (jlong)42.
3060 // The big tile is stored down, and then the foo() value is inserted.
3061 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3063 Node* ctl = old->in(MemNode::Control);
3064 Node* adr = make_raw_address(offset, phase);
3065 const TypePtr* atp = TypeRawPtr::BOTTOM;
3067 // One or two coalesced stores to plop down.
3068 Node* st[2];
3069 intptr_t off[2];
3070 int nst = 0;
3071 if (!split) {
3072 ++new_long;
3073 off[nst] = offset;
3074 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3075 phase->longcon(con), T_LONG);
3076 } else {
3077 // Omit either if it is a zero.
3078 if (con0 != 0) {
3079 ++new_int;
3080 off[nst] = offset;
3081 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3082 phase->intcon(con0), T_INT);
3083 }
3084 if (con1 != 0) {
3085 ++new_int;
3086 offset += BytesPerInt;
3087 adr = make_raw_address(offset, phase);
3088 off[nst] = offset;
3089 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3090 phase->intcon(con1), T_INT);
3091 }
3092 }
3094 // Insert second store first, then the first before the second.
3095 // Insert each one just before any overlapping non-constant stores.
3096 while (nst > 0) {
3097 Node* st1 = st[--nst];
3098 C->copy_node_notes_to(st1, old);
3099 st1 = phase->transform(st1);
3100 offset = off[nst];
3101 assert(offset >= header_size, "do not smash header");
3102 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3103 guarantee(ins_idx != 0, "must re-insert constant store");
3104 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3105 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3106 set_req(--ins_idx, st1);
3107 else
3108 ins_req(ins_idx, st1);
3109 }
3110 }
3112 if (PrintCompilation && WizardMode)
3113 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3114 old_subword, old_long, new_int, new_long);
3115 if (C->log() != NULL)
3116 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3117 old_subword, old_long, new_int, new_long);
3119 // Clean up any remaining occurrences of zmem:
3120 remove_extra_zeroes();
3121 }
3123 // Explore forward from in(start) to find the first fully initialized
3124 // word, and return its offset. Skip groups of subword stores which
3125 // together initialize full words. If in(start) is itself part of a
3126 // fully initialized word, return the offset of in(start). If there
3127 // are no following full-word stores, or if something is fishy, return
3128 // a negative value.
3129 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3130 int int_map = 0;
3131 intptr_t int_map_off = 0;
3132 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3134 for (uint i = start, limit = req(); i < limit; i++) {
3135 Node* st = in(i);
3137 intptr_t st_off = get_store_offset(st, phase);
3138 if (st_off < 0) break; // return conservative answer
3140 int st_size = st->as_Store()->memory_size();
3141 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3142 return st_off; // we found a complete word init
3143 }
3145 // update the map:
3147 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3148 if (this_int_off != int_map_off) {
3149 // reset the map:
3150 int_map = 0;
3151 int_map_off = this_int_off;
3152 }
3154 int subword_off = st_off - this_int_off;
3155 int_map |= right_n_bits(st_size) << subword_off;
3156 if ((int_map & FULL_MAP) == FULL_MAP) {
3157 return this_int_off; // we found a complete word init
3158 }
3160 // Did this store hit or cross the word boundary?
3161 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3162 if (next_int_off == this_int_off + BytesPerInt) {
3163 // We passed the current int, without fully initializing it.
3164 int_map_off = next_int_off;
3165 int_map >>= BytesPerInt;
3166 } else if (next_int_off > this_int_off + BytesPerInt) {
3167 // We passed the current and next int.
3168 return this_int_off + BytesPerInt;
3169 }
3170 }
3172 return -1;
3173 }
3176 // Called when the associated AllocateNode is expanded into CFG.
3177 // At this point, we may perform additional optimizations.
3178 // Linearize the stores by ascending offset, to make memory
3179 // activity as coherent as possible.
3180 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3181 intptr_t header_size,
3182 Node* size_in_bytes,
3183 PhaseGVN* phase) {
3184 assert(!is_complete(), "not already complete");
3185 assert(stores_are_sane(phase), "");
3186 assert(allocation() != NULL, "must be present");
3188 remove_extra_zeroes();
3190 if (ReduceFieldZeroing || ReduceBulkZeroing)
3191 // reduce instruction count for common initialization patterns
3192 coalesce_subword_stores(header_size, size_in_bytes, phase);
3194 Node* zmem = zero_memory(); // initially zero memory state
3195 Node* inits = zmem; // accumulating a linearized chain of inits
3196 #ifdef ASSERT
3197 intptr_t first_offset = allocation()->minimum_header_size();
3198 intptr_t last_init_off = first_offset; // previous init offset
3199 intptr_t last_init_end = first_offset; // previous init offset+size
3200 intptr_t last_tile_end = first_offset; // previous tile offset+size
3201 #endif
3202 intptr_t zeroes_done = header_size;
3204 bool do_zeroing = true; // we might give up if inits are very sparse
3205 int big_init_gaps = 0; // how many large gaps have we seen?
3207 if (ZeroTLAB) do_zeroing = false;
3208 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3210 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3211 Node* st = in(i);
3212 intptr_t st_off = get_store_offset(st, phase);
3213 if (st_off < 0)
3214 break; // unknown junk in the inits
3215 if (st->in(MemNode::Memory) != zmem)
3216 break; // complicated store chains somehow in list
3218 int st_size = st->as_Store()->memory_size();
3219 intptr_t next_init_off = st_off + st_size;
3221 if (do_zeroing && zeroes_done < next_init_off) {
3222 // See if this store needs a zero before it or under it.
3223 intptr_t zeroes_needed = st_off;
3225 if (st_size < BytesPerInt) {
3226 // Look for subword stores which only partially initialize words.
3227 // If we find some, we must lay down some word-level zeroes first,
3228 // underneath the subword stores.
3229 //
3230 // Examples:
3231 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3232 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3233 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3234 //
3235 // Note: coalesce_subword_stores may have already done this,
3236 // if it was prompted by constant non-zero subword initializers.
3237 // But this case can still arise with non-constant stores.
3239 intptr_t next_full_store = find_next_fullword_store(i, phase);
3241 // In the examples above:
3242 // in(i) p q r s x y z
3243 // st_off 12 13 14 15 12 13 14
3244 // st_size 1 1 1 1 1 1 1
3245 // next_full_s. 12 16 16 16 16 16 16
3246 // z's_done 12 16 16 16 12 16 12
3247 // z's_needed 12 16 16 16 16 16 16
3248 // zsize 0 0 0 0 4 0 4
3249 if (next_full_store < 0) {
3250 // Conservative tack: Zero to end of current word.
3251 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3252 } else {
3253 // Zero to beginning of next fully initialized word.
3254 // Or, don't zero at all, if we are already in that word.
3255 assert(next_full_store >= zeroes_needed, "must go forward");
3256 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3257 zeroes_needed = next_full_store;
3258 }
3259 }
3261 if (zeroes_needed > zeroes_done) {
3262 intptr_t zsize = zeroes_needed - zeroes_done;
3263 // Do some incremental zeroing on rawmem, in parallel with inits.
3264 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3265 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3266 zeroes_done, zeroes_needed,
3267 phase);
3268 zeroes_done = zeroes_needed;
3269 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3270 do_zeroing = false; // leave the hole, next time
3271 }
3272 }
3274 // Collect the store and move on:
3275 st->set_req(MemNode::Memory, inits);
3276 inits = st; // put it on the linearized chain
3277 set_req(i, zmem); // unhook from previous position
3279 if (zeroes_done == st_off)
3280 zeroes_done = next_init_off;
3282 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3284 #ifdef ASSERT
3285 // Various order invariants. Weaker than stores_are_sane because
3286 // a large constant tile can be filled in by smaller non-constant stores.
3287 assert(st_off >= last_init_off, "inits do not reverse");
3288 last_init_off = st_off;
3289 const Type* val = NULL;
3290 if (st_size >= BytesPerInt &&
3291 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3292 (int)val->basic_type() < (int)T_OBJECT) {
3293 assert(st_off >= last_tile_end, "tiles do not overlap");
3294 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3295 last_tile_end = MAX2(last_tile_end, next_init_off);
3296 } else {
3297 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3298 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3299 assert(st_off >= last_init_end, "inits do not overlap");
3300 last_init_end = next_init_off; // it's a non-tile
3301 }
3302 #endif //ASSERT
3303 }
3305 remove_extra_zeroes(); // clear out all the zmems left over
3306 add_req(inits);
3308 if (!ZeroTLAB) {
3309 // If anything remains to be zeroed, zero it all now.
3310 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3311 // if it is the last unused 4 bytes of an instance, forget about it
3312 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3313 if (zeroes_done + BytesPerLong >= size_limit) {
3314 assert(allocation() != NULL, "");
3315 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3316 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3317 if (zeroes_done == k->layout_helper())
3318 zeroes_done = size_limit;
3319 }
3320 if (zeroes_done < size_limit) {
3321 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3322 zeroes_done, size_in_bytes, phase);
3323 }
3324 }
3326 set_complete(phase);
3327 return rawmem;
3328 }
3331 #ifdef ASSERT
3332 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3333 if (is_complete())
3334 return true; // stores could be anything at this point
3335 assert(allocation() != NULL, "must be present");
3336 intptr_t last_off = allocation()->minimum_header_size();
3337 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3338 Node* st = in(i);
3339 intptr_t st_off = get_store_offset(st, phase);
3340 if (st_off < 0) continue; // ignore dead garbage
3341 if (last_off > st_off) {
3342 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3343 this->dump(2);
3344 assert(false, "ascending store offsets");
3345 return false;
3346 }
3347 last_off = st_off + st->as_Store()->memory_size();
3348 }
3349 return true;
3350 }
3351 #endif //ASSERT
3356 //============================MergeMemNode=====================================
3357 //
3358 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3359 // contributing store or call operations. Each contributor provides the memory
3360 // state for a particular "alias type" (see Compile::alias_type). For example,
3361 // if a MergeMem has an input X for alias category #6, then any memory reference
3362 // to alias category #6 may use X as its memory state input, as an exact equivalent
3363 // to using the MergeMem as a whole.
3364 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3365 //
3366 // (Here, the <N> notation gives the index of the relevant adr_type.)
3367 //
3368 // In one special case (and more cases in the future), alias categories overlap.
3369 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3370 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3371 // it is exactly equivalent to that state W:
3372 // MergeMem(<Bot>: W) <==> W
3373 //
3374 // Usually, the merge has more than one input. In that case, where inputs
3375 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3376 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3377 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3378 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3379 //
3380 // A merge can take a "wide" memory state as one of its narrow inputs.
3381 // This simply means that the merge observes out only the relevant parts of
3382 // the wide input. That is, wide memory states arriving at narrow merge inputs
3383 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3384 //
3385 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3386 // and that memory slices "leak through":
3387 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3388 //
3389 // But, in such a cascade, repeated memory slices can "block the leak":
3390 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3391 //
3392 // In the last example, Y is not part of the combined memory state of the
3393 // outermost MergeMem. The system must, of course, prevent unschedulable
3394 // memory states from arising, so you can be sure that the state Y is somehow
3395 // a precursor to state Y'.
3396 //
3397 //
3398 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3399 // of each MergeMemNode array are exactly the numerical alias indexes, including
3400 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3401 // Compile::alias_type (and kin) produce and manage these indexes.
3402 //
3403 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3404 // (Note that this provides quick access to the top node inside MergeMem methods,
3405 // without the need to reach out via TLS to Compile::current.)
3406 //
3407 // As a consequence of what was just described, a MergeMem that represents a full
3408 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3409 // containing all alias categories.
3410 //
3411 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3412 //
3413 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3414 // a memory state for the alias type <N>, or else the top node, meaning that
3415 // there is no particular input for that alias type. Note that the length of
3416 // a MergeMem is variable, and may be extended at any time to accommodate new
3417 // memory states at larger alias indexes. When merges grow, they are of course
3418 // filled with "top" in the unused in() positions.
3419 //
3420 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3421 // (Top was chosen because it works smoothly with passes like GCM.)
3422 //
3423 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3424 // the type of random VM bits like TLS references.) Since it is always the
3425 // first non-Bot memory slice, some low-level loops use it to initialize an
3426 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3427 //
3428 //
3429 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3430 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3431 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3432 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3433 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3434 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3435 //
3436 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3437 // really that different from the other memory inputs. An abbreviation called
3438 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3439 //
3440 //
3441 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3442 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3443 // that "emerges though" the base memory will be marked as excluding the alias types
3444 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3445 //
3446 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3447 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3448 //
3449 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3450 // (It is currently unimplemented.) As you can see, the resulting merge is
3451 // actually a disjoint union of memory states, rather than an overlay.
3452 //
3454 //------------------------------MergeMemNode-----------------------------------
3455 Node* MergeMemNode::make_empty_memory() {
3456 Node* empty_memory = (Node*) Compile::current()->top();
3457 assert(empty_memory->is_top(), "correct sentinel identity");
3458 return empty_memory;
3459 }
3461 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3462 init_class_id(Class_MergeMem);
3463 // all inputs are nullified in Node::Node(int)
3464 // set_input(0, NULL); // no control input
3466 // Initialize the edges uniformly to top, for starters.
3467 Node* empty_mem = make_empty_memory();
3468 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3469 init_req(i,empty_mem);
3470 }
3471 assert(empty_memory() == empty_mem, "");
3473 if( new_base != NULL && new_base->is_MergeMem() ) {
3474 MergeMemNode* mdef = new_base->as_MergeMem();
3475 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3476 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3477 mms.set_memory(mms.memory2());
3478 }
3479 assert(base_memory() == mdef->base_memory(), "");
3480 } else {
3481 set_base_memory(new_base);
3482 }
3483 }
3485 // Make a new, untransformed MergeMem with the same base as 'mem'.
3486 // If mem is itself a MergeMem, populate the result with the same edges.
3487 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3488 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3489 }
3491 //------------------------------cmp--------------------------------------------
3492 uint MergeMemNode::hash() const { return NO_HASH; }
3493 uint MergeMemNode::cmp( const Node &n ) const {
3494 return (&n == this); // Always fail except on self
3495 }
3497 //------------------------------Identity---------------------------------------
3498 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3499 // Identity if this merge point does not record any interesting memory
3500 // disambiguations.
3501 Node* base_mem = base_memory();
3502 Node* empty_mem = empty_memory();
3503 if (base_mem != empty_mem) { // Memory path is not dead?
3504 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3505 Node* mem = in(i);
3506 if (mem != empty_mem && mem != base_mem) {
3507 return this; // Many memory splits; no change
3508 }
3509 }
3510 }
3511 return base_mem; // No memory splits; ID on the one true input
3512 }
3514 //------------------------------Ideal------------------------------------------
3515 // This method is invoked recursively on chains of MergeMem nodes
3516 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3517 // Remove chain'd MergeMems
3518 //
3519 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3520 // relative to the "in(Bot)". Since we are patching both at the same time,
3521 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3522 // but rewrite each "in(i)" relative to the new "in(Bot)".
3523 Node *progress = NULL;
3526 Node* old_base = base_memory();
3527 Node* empty_mem = empty_memory();
3528 if (old_base == empty_mem)
3529 return NULL; // Dead memory path.
3531 MergeMemNode* old_mbase;
3532 if (old_base != NULL && old_base->is_MergeMem())
3533 old_mbase = old_base->as_MergeMem();
3534 else
3535 old_mbase = NULL;
3536 Node* new_base = old_base;
3538 // simplify stacked MergeMems in base memory
3539 if (old_mbase) new_base = old_mbase->base_memory();
3541 // the base memory might contribute new slices beyond my req()
3542 if (old_mbase) grow_to_match(old_mbase);
3544 // Look carefully at the base node if it is a phi.
3545 PhiNode* phi_base;
3546 if (new_base != NULL && new_base->is_Phi())
3547 phi_base = new_base->as_Phi();
3548 else
3549 phi_base = NULL;
3551 Node* phi_reg = NULL;
3552 uint phi_len = (uint)-1;
3553 if (phi_base != NULL && !phi_base->is_copy()) {
3554 // do not examine phi if degraded to a copy
3555 phi_reg = phi_base->region();
3556 phi_len = phi_base->req();
3557 // see if the phi is unfinished
3558 for (uint i = 1; i < phi_len; i++) {
3559 if (phi_base->in(i) == NULL) {
3560 // incomplete phi; do not look at it yet!
3561 phi_reg = NULL;
3562 phi_len = (uint)-1;
3563 break;
3564 }
3565 }
3566 }
3568 // Note: We do not call verify_sparse on entry, because inputs
3569 // can normalize to the base_memory via subsume_node or similar
3570 // mechanisms. This method repairs that damage.
3572 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3574 // Look at each slice.
3575 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3576 Node* old_in = in(i);
3577 // calculate the old memory value
3578 Node* old_mem = old_in;
3579 if (old_mem == empty_mem) old_mem = old_base;
3580 assert(old_mem == memory_at(i), "");
3582 // maybe update (reslice) the old memory value
3584 // simplify stacked MergeMems
3585 Node* new_mem = old_mem;
3586 MergeMemNode* old_mmem;
3587 if (old_mem != NULL && old_mem->is_MergeMem())
3588 old_mmem = old_mem->as_MergeMem();
3589 else
3590 old_mmem = NULL;
3591 if (old_mmem == this) {
3592 // This can happen if loops break up and safepoints disappear.
3593 // A merge of BotPtr (default) with a RawPtr memory derived from a
3594 // safepoint can be rewritten to a merge of the same BotPtr with
3595 // the BotPtr phi coming into the loop. If that phi disappears
3596 // also, we can end up with a self-loop of the mergemem.
3597 // In general, if loops degenerate and memory effects disappear,
3598 // a mergemem can be left looking at itself. This simply means
3599 // that the mergemem's default should be used, since there is
3600 // no longer any apparent effect on this slice.
3601 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3602 // from start. Update the input to TOP.
3603 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3604 }
3605 else if (old_mmem != NULL) {
3606 new_mem = old_mmem->memory_at(i);
3607 }
3608 // else preceeding memory was not a MergeMem
3610 // replace equivalent phis (unfortunately, they do not GVN together)
3611 if (new_mem != NULL && new_mem != new_base &&
3612 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3613 if (new_mem->is_Phi()) {
3614 PhiNode* phi_mem = new_mem->as_Phi();
3615 for (uint i = 1; i < phi_len; i++) {
3616 if (phi_base->in(i) != phi_mem->in(i)) {
3617 phi_mem = NULL;
3618 break;
3619 }
3620 }
3621 if (phi_mem != NULL) {
3622 // equivalent phi nodes; revert to the def
3623 new_mem = new_base;
3624 }
3625 }
3626 }
3628 // maybe store down a new value
3629 Node* new_in = new_mem;
3630 if (new_in == new_base) new_in = empty_mem;
3632 if (new_in != old_in) {
3633 // Warning: Do not combine this "if" with the previous "if"
3634 // A memory slice might have be be rewritten even if it is semantically
3635 // unchanged, if the base_memory value has changed.
3636 set_req(i, new_in);
3637 progress = this; // Report progress
3638 }
3639 }
3641 if (new_base != old_base) {
3642 set_req(Compile::AliasIdxBot, new_base);
3643 // Don't use set_base_memory(new_base), because we need to update du.
3644 assert(base_memory() == new_base, "");
3645 progress = this;
3646 }
3648 if( base_memory() == this ) {
3649 // a self cycle indicates this memory path is dead
3650 set_req(Compile::AliasIdxBot, empty_mem);
3651 }
3653 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3654 // Recursion must occur after the self cycle check above
3655 if( base_memory()->is_MergeMem() ) {
3656 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3657 Node *m = phase->transform(new_mbase); // Rollup any cycles
3658 if( m != NULL && (m->is_top() ||
3659 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3660 // propagate rollup of dead cycle to self
3661 set_req(Compile::AliasIdxBot, empty_mem);
3662 }
3663 }
3665 if( base_memory() == empty_mem ) {
3666 progress = this;
3667 // Cut inputs during Parse phase only.
3668 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3669 if( !can_reshape ) {
3670 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3671 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3672 }
3673 }
3674 }
3676 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3677 // Check if PhiNode::Ideal's "Split phis through memory merges"
3678 // transform should be attempted. Look for this->phi->this cycle.
3679 uint merge_width = req();
3680 if (merge_width > Compile::AliasIdxRaw) {
3681 PhiNode* phi = base_memory()->as_Phi();
3682 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3683 if (phi->in(i) == this) {
3684 phase->is_IterGVN()->_worklist.push(phi);
3685 break;
3686 }
3687 }
3688 }
3689 }
3691 assert(progress || verify_sparse(), "please, no dups of base");
3692 return progress;
3693 }
3695 //-------------------------set_base_memory-------------------------------------
3696 void MergeMemNode::set_base_memory(Node *new_base) {
3697 Node* empty_mem = empty_memory();
3698 set_req(Compile::AliasIdxBot, new_base);
3699 assert(memory_at(req()) == new_base, "must set default memory");
3700 // Clear out other occurrences of new_base:
3701 if (new_base != empty_mem) {
3702 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3703 if (in(i) == new_base) set_req(i, empty_mem);
3704 }
3705 }
3706 }
3708 //------------------------------out_RegMask------------------------------------
3709 const RegMask &MergeMemNode::out_RegMask() const {
3710 return RegMask::Empty;
3711 }
3713 //------------------------------dump_spec--------------------------------------
3714 #ifndef PRODUCT
3715 void MergeMemNode::dump_spec(outputStream *st) const {
3716 st->print(" {");
3717 Node* base_mem = base_memory();
3718 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3719 Node* mem = memory_at(i);
3720 if (mem == base_mem) { st->print(" -"); continue; }
3721 st->print( " N%d:", mem->_idx );
3722 Compile::current()->get_adr_type(i)->dump_on(st);
3723 }
3724 st->print(" }");
3725 }
3726 #endif // !PRODUCT
3729 #ifdef ASSERT
3730 static bool might_be_same(Node* a, Node* b) {
3731 if (a == b) return true;
3732 if (!(a->is_Phi() || b->is_Phi())) return false;
3733 // phis shift around during optimization
3734 return true; // pretty stupid...
3735 }
3737 // verify a narrow slice (either incoming or outgoing)
3738 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3739 if (!VerifyAliases) return; // don't bother to verify unless requested
3740 if (is_error_reported()) return; // muzzle asserts when debugging an error
3741 if (Node::in_dump()) return; // muzzle asserts when printing
3742 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3743 assert(n != NULL, "");
3744 // Elide intervening MergeMem's
3745 while (n->is_MergeMem()) {
3746 n = n->as_MergeMem()->memory_at(alias_idx);
3747 }
3748 Compile* C = Compile::current();
3749 const TypePtr* n_adr_type = n->adr_type();
3750 if (n == m->empty_memory()) {
3751 // Implicit copy of base_memory()
3752 } else if (n_adr_type != TypePtr::BOTTOM) {
3753 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3754 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3755 } else {
3756 // A few places like make_runtime_call "know" that VM calls are narrow,
3757 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3758 bool expected_wide_mem = false;
3759 if (n == m->base_memory()) {
3760 expected_wide_mem = true;
3761 } else if (alias_idx == Compile::AliasIdxRaw ||
3762 n == m->memory_at(Compile::AliasIdxRaw)) {
3763 expected_wide_mem = true;
3764 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3765 // memory can "leak through" calls on channels that
3766 // are write-once. Allow this also.
3767 expected_wide_mem = true;
3768 }
3769 assert(expected_wide_mem, "expected narrow slice replacement");
3770 }
3771 }
3772 #else // !ASSERT
3773 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3774 #endif
3777 //-----------------------------memory_at---------------------------------------
3778 Node* MergeMemNode::memory_at(uint alias_idx) const {
3779 assert(alias_idx >= Compile::AliasIdxRaw ||
3780 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3781 "must avoid base_memory and AliasIdxTop");
3783 // Otherwise, it is a narrow slice.
3784 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3785 Compile *C = Compile::current();
3786 if (is_empty_memory(n)) {
3787 // the array is sparse; empty slots are the "top" node
3788 n = base_memory();
3789 assert(Node::in_dump()
3790 || n == NULL || n->bottom_type() == Type::TOP
3791 || n->adr_type() == TypePtr::BOTTOM
3792 || n->adr_type() == TypeRawPtr::BOTTOM
3793 || Compile::current()->AliasLevel() == 0,
3794 "must be a wide memory");
3795 // AliasLevel == 0 if we are organizing the memory states manually.
3796 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3797 } else {
3798 // make sure the stored slice is sane
3799 #ifdef ASSERT
3800 if (is_error_reported() || Node::in_dump()) {
3801 } else if (might_be_same(n, base_memory())) {
3802 // Give it a pass: It is a mostly harmless repetition of the base.
3803 // This can arise normally from node subsumption during optimization.
3804 } else {
3805 verify_memory_slice(this, alias_idx, n);
3806 }
3807 #endif
3808 }
3809 return n;
3810 }
3812 //---------------------------set_memory_at-------------------------------------
3813 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3814 verify_memory_slice(this, alias_idx, n);
3815 Node* empty_mem = empty_memory();
3816 if (n == base_memory()) n = empty_mem; // collapse default
3817 uint need_req = alias_idx+1;
3818 if (req() < need_req) {
3819 if (n == empty_mem) return; // already the default, so do not grow me
3820 // grow the sparse array
3821 do {
3822 add_req(empty_mem);
3823 } while (req() < need_req);
3824 }
3825 set_req( alias_idx, n );
3826 }
3830 //--------------------------iteration_setup------------------------------------
3831 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3832 if (other != NULL) {
3833 grow_to_match(other);
3834 // invariant: the finite support of mm2 is within mm->req()
3835 #ifdef ASSERT
3836 for (uint i = req(); i < other->req(); i++) {
3837 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3838 }
3839 #endif
3840 }
3841 // Replace spurious copies of base_memory by top.
3842 Node* base_mem = base_memory();
3843 if (base_mem != NULL && !base_mem->is_top()) {
3844 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3845 if (in(i) == base_mem)
3846 set_req(i, empty_memory());
3847 }
3848 }
3849 }
3851 //---------------------------grow_to_match-------------------------------------
3852 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3853 Node* empty_mem = empty_memory();
3854 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3855 // look for the finite support of the other memory
3856 for (uint i = other->req(); --i >= req(); ) {
3857 if (other->in(i) != empty_mem) {
3858 uint new_len = i+1;
3859 while (req() < new_len) add_req(empty_mem);
3860 break;
3861 }
3862 }
3863 }
3865 //---------------------------verify_sparse-------------------------------------
3866 #ifndef PRODUCT
3867 bool MergeMemNode::verify_sparse() const {
3868 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3869 Node* base_mem = base_memory();
3870 // The following can happen in degenerate cases, since empty==top.
3871 if (is_empty_memory(base_mem)) return true;
3872 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3873 assert(in(i) != NULL, "sane slice");
3874 if (in(i) == base_mem) return false; // should have been the sentinel value!
3875 }
3876 return true;
3877 }
3879 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3880 Node* n;
3881 n = mm->in(idx);
3882 if (mem == n) return true; // might be empty_memory()
3883 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3884 if (mem == n) return true;
3885 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3886 if (mem == n) return true;
3887 if (n == NULL) break;
3888 }
3889 return false;
3890 }
3891 #endif // !PRODUCT