Wed, 17 Sep 2008 08:29:17 -0700
6711100: 64bit fastdebug server vm crashes with assert(_base == Int,"Not an Int")
Summary: insert CastII nodes to narrow type of load_array_length() node
Reviewed-by: never, kvn
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 *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
98 Node *prev = NULL;
99 Node *result = mchain;
100 while (prev != result) {
101 prev = result;
102 if (result == start_mem)
103 break; // hit one of our sentinals
104 // skip over a call which does not affect this memory slice
105 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
106 Node *proj_in = result->in(0);
107 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
108 break; // hit one of our sentinals
109 } else if (proj_in->is_Call()) {
110 CallNode *call = proj_in->as_Call();
111 if (!call->may_modify(t_adr, phase)) {
112 result = call->in(TypeFunc::Memory);
113 }
114 } else if (proj_in->is_Initialize()) {
115 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
116 // Stop if this is the initialization for the object instance which
117 // which contains this memory slice, otherwise skip over it.
118 if (alloc != NULL && alloc->_idx != instance_id) {
119 result = proj_in->in(TypeFunc::Memory);
120 }
121 } else if (proj_in->is_MemBar()) {
122 result = proj_in->in(TypeFunc::Memory);
123 } else {
124 assert(false, "unexpected projection");
125 }
126 } else if (result->is_MergeMem()) {
127 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
128 }
129 }
130 return result;
131 }
133 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
134 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
135 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
136 PhaseIterGVN *igvn = phase->is_IterGVN();
137 Node *result = mchain;
138 result = optimize_simple_memory_chain(result, t_adr, phase);
139 if (is_instance && igvn != NULL && result->is_Phi()) {
140 PhiNode *mphi = result->as_Phi();
141 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
142 const TypePtr *t = mphi->adr_type();
143 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
144 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
145 t->is_oopptr()->cast_to_exactness(true)
146 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
147 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
148 // clone the Phi with our address type
149 result = mphi->split_out_instance(t_adr, igvn);
150 } else {
151 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
152 }
153 }
154 return result;
155 }
157 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
158 uint alias_idx = phase->C->get_alias_index(tp);
159 Node *mem = mmem;
160 #ifdef ASSERT
161 {
162 // Check that current type is consistent with the alias index used during graph construction
163 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
164 bool consistent = adr_check == NULL || adr_check->empty() ||
165 phase->C->must_alias(adr_check, alias_idx );
166 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
167 if( !consistent && adr_check != NULL && !adr_check->empty() &&
168 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
169 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
170 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
171 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
172 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
173 // don't assert if it is dead code.
174 consistent = true;
175 }
176 if( !consistent ) {
177 st->print("alias_idx==%d, adr_check==", alias_idx);
178 if( adr_check == NULL ) {
179 st->print("NULL");
180 } else {
181 adr_check->dump();
182 }
183 st->cr();
184 print_alias_types();
185 assert(consistent, "adr_check must match alias idx");
186 }
187 }
188 #endif
189 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
190 // means an array I have not precisely typed yet. Do not do any
191 // alias stuff with it any time soon.
192 const TypeOopPtr *tinst = tp->isa_oopptr();
193 if( tp->base() != Type::AnyPtr &&
194 !(tinst &&
195 tinst->klass()->is_java_lang_Object() &&
196 tinst->offset() == Type::OffsetBot) ) {
197 // compress paths and change unreachable cycles to TOP
198 // If not, we can update the input infinitely along a MergeMem cycle
199 // Equivalent code in PhiNode::Ideal
200 Node* m = phase->transform(mmem);
201 // If tranformed to a MergeMem, get the desired slice
202 // Otherwise the returned node represents memory for every slice
203 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
204 // Update input if it is progress over what we have now
205 }
206 return mem;
207 }
209 //--------------------------Ideal_common---------------------------------------
210 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
211 // Unhook non-raw memories from complete (macro-expanded) initializations.
212 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
213 // If our control input is a dead region, kill all below the region
214 Node *ctl = in(MemNode::Control);
215 if (ctl && remove_dead_region(phase, can_reshape))
216 return this;
217 ctl = in(MemNode::Control);
218 // Don't bother trying to transform a dead node
219 if( ctl && ctl->is_top() ) return NodeSentinel;
221 // Ignore if memory is dead, or self-loop
222 Node *mem = in(MemNode::Memory);
223 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
224 assert( mem != this, "dead loop in MemNode::Ideal" );
226 Node *address = in(MemNode::Address);
227 const Type *t_adr = phase->type( address );
228 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
230 // Avoid independent memory operations
231 Node* old_mem = mem;
233 // The code which unhooks non-raw memories from complete (macro-expanded)
234 // initializations was removed. After macro-expansion all stores catched
235 // by Initialize node became raw stores and there is no information
236 // which memory slices they modify. So it is unsafe to move any memory
237 // operation above these stores. Also in most cases hooked non-raw memories
238 // were already unhooked by using information from detect_ptr_independence()
239 // and find_previous_store().
241 if (mem->is_MergeMem()) {
242 MergeMemNode* mmem = mem->as_MergeMem();
243 const TypePtr *tp = t_adr->is_ptr();
245 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
246 }
248 if (mem != old_mem) {
249 set_req(MemNode::Memory, mem);
250 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
251 return this;
252 }
254 // let the subclass continue analyzing...
255 return NULL;
256 }
258 // Helper function for proving some simple control dominations.
259 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
260 // Already assumes that 'dom' is available at 'sub', and that 'sub'
261 // is not a constant (dominated by the method's StartNode).
262 // Used by MemNode::find_previous_store to prove that the
263 // control input of a memory operation predates (dominates)
264 // an allocation it wants to look past.
265 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
266 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
267 return false; // Conservative answer for dead code
269 // Check 'dom'. Skip Proj and CatchProj nodes.
270 dom = dom->find_exact_control(dom);
271 if (dom == NULL || dom->is_top())
272 return false; // Conservative answer for dead code
274 if (dom == sub) {
275 // For the case when, for example, 'sub' is Initialize and the original
276 // 'dom' is Proj node of the 'sub'.
277 return false;
278 }
280 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
281 return true;
283 // 'dom' dominates 'sub' if its control edge and control edges
284 // of all its inputs dominate or equal to sub's control edge.
286 // Currently 'sub' is either Allocate, Initialize or Start nodes.
287 // Or Region for the check in LoadNode::Ideal();
288 // 'sub' should have sub->in(0) != NULL.
289 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
290 sub->is_Region(), "expecting only these nodes");
292 // Get control edge of 'sub'.
293 Node* orig_sub = sub;
294 sub = sub->find_exact_control(sub->in(0));
295 if (sub == NULL || sub->is_top())
296 return false; // Conservative answer for dead code
298 assert(sub->is_CFG(), "expecting control");
300 if (sub == dom)
301 return true;
303 if (sub->is_Start() || sub->is_Root())
304 return false;
306 {
307 // Check all control edges of 'dom'.
309 ResourceMark rm;
310 Arena* arena = Thread::current()->resource_area();
311 Node_List nlist(arena);
312 Unique_Node_List dom_list(arena);
314 dom_list.push(dom);
315 bool only_dominating_controls = false;
317 for (uint next = 0; next < dom_list.size(); next++) {
318 Node* n = dom_list.at(next);
319 if (n == orig_sub)
320 return false; // One of dom's inputs dominated by sub.
321 if (!n->is_CFG() && n->pinned()) {
322 // Check only own control edge for pinned non-control nodes.
323 n = n->find_exact_control(n->in(0));
324 if (n == NULL || n->is_top())
325 return false; // Conservative answer for dead code
326 assert(n->is_CFG(), "expecting control");
327 dom_list.push(n);
328 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
329 only_dominating_controls = true;
330 } else if (n->is_CFG()) {
331 if (n->dominates(sub, nlist))
332 only_dominating_controls = true;
333 else
334 return false;
335 } else {
336 // First, own control edge.
337 Node* m = n->find_exact_control(n->in(0));
338 if (m != NULL) {
339 if (m->is_top())
340 return false; // Conservative answer for dead code
341 dom_list.push(m);
342 }
343 // Now, the rest of edges.
344 uint cnt = n->req();
345 for (uint i = 1; i < cnt; i++) {
346 m = n->find_exact_control(n->in(i));
347 if (m == NULL || m->is_top())
348 continue;
349 dom_list.push(m);
350 }
351 }
352 }
353 return only_dominating_controls;
354 }
355 }
357 //---------------------detect_ptr_independence---------------------------------
358 // Used by MemNode::find_previous_store to prove that two base
359 // pointers are never equal.
360 // The pointers are accompanied by their associated allocations,
361 // if any, which have been previously discovered by the caller.
362 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
363 Node* p2, AllocateNode* a2,
364 PhaseTransform* phase) {
365 // Attempt to prove that these two pointers cannot be aliased.
366 // They may both manifestly be allocations, and they should differ.
367 // Or, if they are not both allocations, they can be distinct constants.
368 // Otherwise, one is an allocation and the other a pre-existing value.
369 if (a1 == NULL && a2 == NULL) { // neither an allocation
370 return (p1 != p2) && p1->is_Con() && p2->is_Con();
371 } else if (a1 != NULL && a2 != NULL) { // both allocations
372 return (a1 != a2);
373 } else if (a1 != NULL) { // one allocation a1
374 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
375 return all_controls_dominate(p2, a1);
376 } else { //(a2 != NULL) // one allocation a2
377 return all_controls_dominate(p1, a2);
378 }
379 return false;
380 }
383 // The logic for reordering loads and stores uses four steps:
384 // (a) Walk carefully past stores and initializations which we
385 // can prove are independent of this load.
386 // (b) Observe that the next memory state makes an exact match
387 // with self (load or store), and locate the relevant store.
388 // (c) Ensure that, if we were to wire self directly to the store,
389 // the optimizer would fold it up somehow.
390 // (d) Do the rewiring, and return, depending on some other part of
391 // the optimizer to fold up the load.
392 // This routine handles steps (a) and (b). Steps (c) and (d) are
393 // specific to loads and stores, so they are handled by the callers.
394 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
395 //
396 Node* MemNode::find_previous_store(PhaseTransform* phase) {
397 Node* ctrl = in(MemNode::Control);
398 Node* adr = in(MemNode::Address);
399 intptr_t offset = 0;
400 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
401 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
403 if (offset == Type::OffsetBot)
404 return NULL; // cannot unalias unless there are precise offsets
406 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
408 intptr_t size_in_bytes = memory_size();
410 Node* mem = in(MemNode::Memory); // start searching here...
412 int cnt = 50; // Cycle limiter
413 for (;;) { // While we can dance past unrelated stores...
414 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
416 if (mem->is_Store()) {
417 Node* st_adr = mem->in(MemNode::Address);
418 intptr_t st_offset = 0;
419 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
420 if (st_base == NULL)
421 break; // inscrutable pointer
422 if (st_offset != offset && st_offset != Type::OffsetBot) {
423 const int MAX_STORE = BytesPerLong;
424 if (st_offset >= offset + size_in_bytes ||
425 st_offset <= offset - MAX_STORE ||
426 st_offset <= offset - mem->as_Store()->memory_size()) {
427 // Success: The offsets are provably independent.
428 // (You may ask, why not just test st_offset != offset and be done?
429 // The answer is that stores of different sizes can co-exist
430 // in the same sequence of RawMem effects. We sometimes initialize
431 // a whole 'tile' of array elements with a single jint or jlong.)
432 mem = mem->in(MemNode::Memory);
433 continue; // (a) advance through independent store memory
434 }
435 }
436 if (st_base != base &&
437 detect_ptr_independence(base, alloc,
438 st_base,
439 AllocateNode::Ideal_allocation(st_base, phase),
440 phase)) {
441 // Success: The bases are provably independent.
442 mem = mem->in(MemNode::Memory);
443 continue; // (a) advance through independent store memory
444 }
446 // (b) At this point, if the bases or offsets do not agree, we lose,
447 // since we have not managed to prove 'this' and 'mem' independent.
448 if (st_base == base && st_offset == offset) {
449 return mem; // let caller handle steps (c), (d)
450 }
452 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
453 InitializeNode* st_init = mem->in(0)->as_Initialize();
454 AllocateNode* st_alloc = st_init->allocation();
455 if (st_alloc == NULL)
456 break; // something degenerated
457 bool known_identical = false;
458 bool known_independent = false;
459 if (alloc == st_alloc)
460 known_identical = true;
461 else if (alloc != NULL)
462 known_independent = true;
463 else if (all_controls_dominate(this, st_alloc))
464 known_independent = true;
466 if (known_independent) {
467 // The bases are provably independent: Either they are
468 // manifestly distinct allocations, or else the control
469 // of this load dominates the store's allocation.
470 int alias_idx = phase->C->get_alias_index(adr_type());
471 if (alias_idx == Compile::AliasIdxRaw) {
472 mem = st_alloc->in(TypeFunc::Memory);
473 } else {
474 mem = st_init->memory(alias_idx);
475 }
476 continue; // (a) advance through independent store memory
477 }
479 // (b) at this point, if we are not looking at a store initializing
480 // the same allocation we are loading from, we lose.
481 if (known_identical) {
482 // From caller, can_see_stored_value will consult find_captured_store.
483 return mem; // let caller handle steps (c), (d)
484 }
486 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
487 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
488 if (mem->is_Proj() && mem->in(0)->is_Call()) {
489 CallNode *call = mem->in(0)->as_Call();
490 if (!call->may_modify(addr_t, phase)) {
491 mem = call->in(TypeFunc::Memory);
492 continue; // (a) advance through independent call memory
493 }
494 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
495 mem = mem->in(0)->in(TypeFunc::Memory);
496 continue; // (a) advance through independent MemBar memory
497 } else if (mem->is_MergeMem()) {
498 int alias_idx = phase->C->get_alias_index(adr_type());
499 mem = mem->as_MergeMem()->memory_at(alias_idx);
500 continue; // (a) advance through independent MergeMem memory
501 }
502 }
504 // Unless there is an explicit 'continue', we must bail out here,
505 // because 'mem' is an inscrutable memory state (e.g., a call).
506 break;
507 }
509 return NULL; // bail out
510 }
512 //----------------------calculate_adr_type-------------------------------------
513 // Helper function. Notices when the given type of address hits top or bottom.
514 // Also, asserts a cross-check of the type against the expected address type.
515 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
516 if (t == Type::TOP) return NULL; // does not touch memory any more?
517 #ifdef PRODUCT
518 cross_check = NULL;
519 #else
520 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
521 #endif
522 const TypePtr* tp = t->isa_ptr();
523 if (tp == NULL) {
524 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
525 return TypePtr::BOTTOM; // touches lots of memory
526 } else {
527 #ifdef ASSERT
528 // %%%% [phh] We don't check the alias index if cross_check is
529 // TypeRawPtr::BOTTOM. Needs to be investigated.
530 if (cross_check != NULL &&
531 cross_check != TypePtr::BOTTOM &&
532 cross_check != TypeRawPtr::BOTTOM) {
533 // Recheck the alias index, to see if it has changed (due to a bug).
534 Compile* C = Compile::current();
535 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
536 "must stay in the original alias category");
537 // The type of the address must be contained in the adr_type,
538 // disregarding "null"-ness.
539 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
540 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
541 assert(cross_check->meet(tp_notnull) == cross_check,
542 "real address must not escape from expected memory type");
543 }
544 #endif
545 return tp;
546 }
547 }
549 //------------------------adr_phi_is_loop_invariant----------------------------
550 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
551 // loop is loop invariant. Make a quick traversal of Phi and associated
552 // CastPP nodes, looking to see if they are a closed group within the loop.
553 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
554 // The idea is that the phi-nest must boil down to only CastPP nodes
555 // with the same data. This implies that any path into the loop already
556 // includes such a CastPP, and so the original cast, whatever its input,
557 // must be covered by an equivalent cast, with an earlier control input.
558 ResourceMark rm;
560 // The loop entry input of the phi should be the unique dominating
561 // node for every Phi/CastPP in the loop.
562 Unique_Node_List closure;
563 closure.push(adr_phi->in(LoopNode::EntryControl));
565 // Add the phi node and the cast to the worklist.
566 Unique_Node_List worklist;
567 worklist.push(adr_phi);
568 if( cast != NULL ){
569 if( !cast->is_ConstraintCast() ) return false;
570 worklist.push(cast);
571 }
573 // Begin recursive walk of phi nodes.
574 while( worklist.size() ){
575 // Take a node off the worklist
576 Node *n = worklist.pop();
577 if( !closure.member(n) ){
578 // Add it to the closure.
579 closure.push(n);
580 // Make a sanity check to ensure we don't waste too much time here.
581 if( closure.size() > 20) return false;
582 // This node is OK if:
583 // - it is a cast of an identical value
584 // - or it is a phi node (then we add its inputs to the worklist)
585 // Otherwise, the node is not OK, and we presume the cast is not invariant
586 if( n->is_ConstraintCast() ){
587 worklist.push(n->in(1));
588 } else if( n->is_Phi() ) {
589 for( uint i = 1; i < n->req(); i++ ) {
590 worklist.push(n->in(i));
591 }
592 } else {
593 return false;
594 }
595 }
596 }
598 // Quit when the worklist is empty, and we've found no offending nodes.
599 return true;
600 }
602 //------------------------------Ideal_DU_postCCP-------------------------------
603 // Find any cast-away of null-ness and keep its control. Null cast-aways are
604 // going away in this pass and we need to make this memory op depend on the
605 // gating null check.
606 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
607 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
608 }
610 // I tried to leave the CastPP's in. This makes the graph more accurate in
611 // some sense; we get to keep around the knowledge that an oop is not-null
612 // after some test. Alas, the CastPP's interfere with GVN (some values are
613 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
614 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
615 // some of the more trivial cases in the optimizer. Removing more useless
616 // Phi's started allowing Loads to illegally float above null checks. I gave
617 // up on this approach. CNC 10/20/2000
618 // This static method may be called not from MemNode (EncodePNode calls it).
619 // Only the control edge of the node 'n' might be updated.
620 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
621 Node *skipped_cast = NULL;
622 // Need a null check? Regular static accesses do not because they are
623 // from constant addresses. Array ops are gated by the range check (which
624 // always includes a NULL check). Just check field ops.
625 if( n->in(MemNode::Control) == NULL ) {
626 // Scan upwards for the highest location we can place this memory op.
627 while( true ) {
628 switch( adr->Opcode() ) {
630 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
631 adr = adr->in(AddPNode::Base);
632 continue;
634 case Op_DecodeN: // No change to NULL-ness, so peek thru
635 adr = adr->in(1);
636 continue;
638 case Op_CastPP:
639 // If the CastPP is useless, just peek on through it.
640 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
641 // Remember the cast that we've peeked though. If we peek
642 // through more than one, then we end up remembering the highest
643 // one, that is, if in a loop, the one closest to the top.
644 skipped_cast = adr;
645 adr = adr->in(1);
646 continue;
647 }
648 // CastPP is going away in this pass! We need this memory op to be
649 // control-dependent on the test that is guarding the CastPP.
650 ccp->hash_delete(n);
651 n->set_req(MemNode::Control, adr->in(0));
652 ccp->hash_insert(n);
653 return n;
655 case Op_Phi:
656 // Attempt to float above a Phi to some dominating point.
657 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
658 // If we've already peeked through a Cast (which could have set the
659 // control), we can't float above a Phi, because the skipped Cast
660 // may not be loop invariant.
661 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
662 adr = adr->in(1);
663 continue;
664 }
665 }
667 // Intentional fallthrough!
669 // No obvious dominating point. The mem op is pinned below the Phi
670 // by the Phi itself. If the Phi goes away (no true value is merged)
671 // then the mem op can float, but not indefinitely. It must be pinned
672 // behind the controls leading to the Phi.
673 case Op_CheckCastPP:
674 // These usually stick around to change address type, however a
675 // useless one can be elided and we still need to pick up a control edge
676 if (adr->in(0) == NULL) {
677 // This CheckCastPP node has NO control and is likely useless. But we
678 // need check further up the ancestor chain for a control input to keep
679 // the node in place. 4959717.
680 skipped_cast = adr;
681 adr = adr->in(1);
682 continue;
683 }
684 ccp->hash_delete(n);
685 n->set_req(MemNode::Control, adr->in(0));
686 ccp->hash_insert(n);
687 return n;
689 // List of "safe" opcodes; those that implicitly block the memory
690 // op below any null check.
691 case Op_CastX2P: // no null checks on native pointers
692 case Op_Parm: // 'this' pointer is not null
693 case Op_LoadP: // Loading from within a klass
694 case Op_LoadN: // Loading from within a klass
695 case Op_LoadKlass: // Loading from within a klass
696 case Op_LoadNKlass: // Loading from within a klass
697 case Op_ConP: // Loading from a klass
698 case Op_ConN: // Loading from a klass
699 case Op_CreateEx: // Sucking up the guts of an exception oop
700 case Op_Con: // Reading from TLS
701 case Op_CMoveP: // CMoveP is pinned
702 case Op_CMoveN: // CMoveN is pinned
703 break; // No progress
705 case Op_Proj: // Direct call to an allocation routine
706 case Op_SCMemProj: // Memory state from store conditional ops
707 #ifdef ASSERT
708 {
709 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
710 const Node* call = adr->in(0);
711 if (call->is_CallJava()) {
712 const CallJavaNode* call_java = call->as_CallJava();
713 const TypeTuple *r = call_java->tf()->range();
714 assert(r->cnt() > TypeFunc::Parms, "must return value");
715 const Type* ret_type = r->field_at(TypeFunc::Parms);
716 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
717 // We further presume that this is one of
718 // new_instance_Java, new_array_Java, or
719 // the like, but do not assert for this.
720 } else if (call->is_Allocate()) {
721 // similar case to new_instance_Java, etc.
722 } else if (!call->is_CallLeaf()) {
723 // Projections from fetch_oop (OSR) are allowed as well.
724 ShouldNotReachHere();
725 }
726 }
727 #endif
728 break;
729 default:
730 ShouldNotReachHere();
731 }
732 break;
733 }
734 }
736 return NULL; // No progress
737 }
740 //=============================================================================
741 uint LoadNode::size_of() const { return sizeof(*this); }
742 uint LoadNode::cmp( const Node &n ) const
743 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
744 const Type *LoadNode::bottom_type() const { return _type; }
745 uint LoadNode::ideal_reg() const {
746 return Matcher::base2reg[_type->base()];
747 }
749 #ifndef PRODUCT
750 void LoadNode::dump_spec(outputStream *st) const {
751 MemNode::dump_spec(st);
752 if( !Verbose && !WizardMode ) {
753 // standard dump does this in Verbose and WizardMode
754 st->print(" #"); _type->dump_on(st);
755 }
756 }
757 #endif
760 //----------------------------LoadNode::make-----------------------------------
761 // Polymorphic factory method:
762 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
763 Compile* C = gvn.C;
765 // sanity check the alias category against the created node type
766 assert(!(adr_type->isa_oopptr() &&
767 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
768 "use LoadKlassNode instead");
769 assert(!(adr_type->isa_aryptr() &&
770 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
771 "use LoadRangeNode instead");
772 switch (bt) {
773 case T_BOOLEAN:
774 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
775 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
776 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
777 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
778 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
779 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
780 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
781 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
782 case T_OBJECT:
783 #ifdef _LP64
784 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
785 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
786 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
787 } else
788 #endif
789 {
790 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
791 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
792 }
793 }
794 ShouldNotReachHere();
795 return (LoadNode*)NULL;
796 }
798 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
799 bool require_atomic = true;
800 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
801 }
806 //------------------------------hash-------------------------------------------
807 uint LoadNode::hash() const {
808 // unroll addition of interesting fields
809 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
810 }
812 //---------------------------can_see_stored_value------------------------------
813 // This routine exists to make sure this set of tests is done the same
814 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
815 // will change the graph shape in a way which makes memory alive twice at the
816 // same time (uses the Oracle model of aliasing), then some
817 // LoadXNode::Identity will fold things back to the equivalence-class model
818 // of aliasing.
819 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
820 Node* ld_adr = in(MemNode::Address);
822 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
823 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
824 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
825 atp->field() != NULL && !atp->field()->is_volatile()) {
826 uint alias_idx = atp->index();
827 bool final = atp->field()->is_final();
828 Node* result = NULL;
829 Node* current = st;
830 // Skip through chains of MemBarNodes checking the MergeMems for
831 // new states for the slice of this load. Stop once any other
832 // kind of node is encountered. Loads from final memory can skip
833 // through any kind of MemBar but normal loads shouldn't skip
834 // through MemBarAcquire since the could allow them to move out of
835 // a synchronized region.
836 while (current->is_Proj()) {
837 int opc = current->in(0)->Opcode();
838 if ((final && opc == Op_MemBarAcquire) ||
839 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
840 Node* mem = current->in(0)->in(TypeFunc::Memory);
841 if (mem->is_MergeMem()) {
842 MergeMemNode* merge = mem->as_MergeMem();
843 Node* new_st = merge->memory_at(alias_idx);
844 if (new_st == merge->base_memory()) {
845 // Keep searching
846 current = merge->base_memory();
847 continue;
848 }
849 // Save the new memory state for the slice and fall through
850 // to exit.
851 result = new_st;
852 }
853 }
854 break;
855 }
856 if (result != NULL) {
857 st = result;
858 }
859 }
862 // Loop around twice in the case Load -> Initialize -> Store.
863 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
864 for (int trip = 0; trip <= 1; trip++) {
866 if (st->is_Store()) {
867 Node* st_adr = st->in(MemNode::Address);
868 if (!phase->eqv(st_adr, ld_adr)) {
869 // Try harder before giving up... Match raw and non-raw pointers.
870 intptr_t st_off = 0;
871 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
872 if (alloc == NULL) return NULL;
873 intptr_t ld_off = 0;
874 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
875 if (alloc != allo2) return NULL;
876 if (ld_off != st_off) return NULL;
877 // At this point we have proven something like this setup:
878 // A = Allocate(...)
879 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
880 // S = StoreQ(, AddP(, A.Parm , #Off), V)
881 // (Actually, we haven't yet proven the Q's are the same.)
882 // In other words, we are loading from a casted version of
883 // the same pointer-and-offset that we stored to.
884 // Thus, we are able to replace L by V.
885 }
886 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
887 if (store_Opcode() != st->Opcode())
888 return NULL;
889 return st->in(MemNode::ValueIn);
890 }
892 intptr_t offset = 0; // scratch
894 // A load from a freshly-created object always returns zero.
895 // (This can happen after LoadNode::Ideal resets the load's memory input
896 // to find_captured_store, which returned InitializeNode::zero_memory.)
897 if (st->is_Proj() && st->in(0)->is_Allocate() &&
898 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
899 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
900 // return a zero value for the load's basic type
901 // (This is one of the few places where a generic PhaseTransform
902 // can create new nodes. Think of it as lazily manifesting
903 // virtually pre-existing constants.)
904 return phase->zerocon(memory_type());
905 }
907 // A load from an initialization barrier can match a captured store.
908 if (st->is_Proj() && st->in(0)->is_Initialize()) {
909 InitializeNode* init = st->in(0)->as_Initialize();
910 AllocateNode* alloc = init->allocation();
911 if (alloc != NULL &&
912 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
913 // examine a captured store value
914 st = init->find_captured_store(offset, memory_size(), phase);
915 if (st != NULL)
916 continue; // take one more trip around
917 }
918 }
920 break;
921 }
923 return NULL;
924 }
926 //----------------------is_instance_field_load_with_local_phi------------------
927 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
928 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
929 in(MemNode::Address)->is_AddP() ) {
930 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
931 // Only instances.
932 if( t_oop != NULL && t_oop->is_known_instance_field() &&
933 t_oop->offset() != Type::OffsetBot &&
934 t_oop->offset() != Type::OffsetTop) {
935 return true;
936 }
937 }
938 return false;
939 }
941 //------------------------------Identity---------------------------------------
942 // Loads are identity if previous store is to same address
943 Node *LoadNode::Identity( PhaseTransform *phase ) {
944 // If the previous store-maker is the right kind of Store, and the store is
945 // to the same address, then we are equal to the value stored.
946 Node* mem = in(MemNode::Memory);
947 Node* value = can_see_stored_value(mem, phase);
948 if( value ) {
949 // byte, short & char stores truncate naturally.
950 // A load has to load the truncated value which requires
951 // some sort of masking operation and that requires an
952 // Ideal call instead of an Identity call.
953 if (memory_size() < BytesPerInt) {
954 // If the input to the store does not fit with the load's result type,
955 // it must be truncated via an Ideal call.
956 if (!phase->type(value)->higher_equal(phase->type(this)))
957 return this;
958 }
959 // (This works even when value is a Con, but LoadNode::Value
960 // usually runs first, producing the singleton type of the Con.)
961 return value;
962 }
964 // Search for an existing data phi which was generated before for the same
965 // instance's field to avoid infinite genertion of phis in a loop.
966 Node *region = mem->in(0);
967 if (is_instance_field_load_with_local_phi(region)) {
968 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
969 int this_index = phase->C->get_alias_index(addr_t);
970 int this_offset = addr_t->offset();
971 int this_id = addr_t->is_oopptr()->instance_id();
972 const Type* this_type = bottom_type();
973 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
974 Node* phi = region->fast_out(i);
975 if (phi->is_Phi() && phi != mem &&
976 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
977 return phi;
978 }
979 }
980 }
982 return this;
983 }
986 // Returns true if the AliasType refers to the field that holds the
987 // cached box array. Currently only handles the IntegerCache case.
988 static bool is_autobox_cache(Compile::AliasType* atp) {
989 if (atp != NULL && atp->field() != NULL) {
990 ciField* field = atp->field();
991 ciSymbol* klass = field->holder()->name();
992 if (field->name() == ciSymbol::cache_field_name() &&
993 field->holder()->uses_default_loader() &&
994 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
995 return true;
996 }
997 }
998 return false;
999 }
1001 // Fetch the base value in the autobox array
1002 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
1003 if (atp != NULL && atp->field() != NULL) {
1004 ciField* field = atp->field();
1005 ciSymbol* klass = field->holder()->name();
1006 if (field->name() == ciSymbol::cache_field_name() &&
1007 field->holder()->uses_default_loader() &&
1008 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1009 assert(field->is_constant(), "what?");
1010 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
1011 // Fetch the box object at the base of the array and get its value
1012 ciInstance* box = array->obj_at(0)->as_instance();
1013 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1014 if (ik->nof_nonstatic_fields() == 1) {
1015 // This should be true nonstatic_field_at requires calling
1016 // nof_nonstatic_fields so check it anyway
1017 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1018 cache_offset = c.as_int();
1019 }
1020 return true;
1021 }
1022 }
1023 return false;
1024 }
1026 // Returns true if the AliasType refers to the value field of an
1027 // autobox object. Currently only handles Integer.
1028 static bool is_autobox_object(Compile::AliasType* atp) {
1029 if (atp != NULL && atp->field() != NULL) {
1030 ciField* field = atp->field();
1031 ciSymbol* klass = field->holder()->name();
1032 if (field->name() == ciSymbol::value_name() &&
1033 field->holder()->uses_default_loader() &&
1034 klass == ciSymbol::java_lang_Integer()) {
1035 return true;
1036 }
1037 }
1038 return false;
1039 }
1042 // We're loading from an object which has autobox behaviour.
1043 // If this object is result of a valueOf call we'll have a phi
1044 // merging a newly allocated object and a load from the cache.
1045 // We want to replace this load with the original incoming
1046 // argument to the valueOf call.
1047 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1048 Node* base = in(Address)->in(AddPNode::Base);
1049 if (base->is_Phi() && base->req() == 3) {
1050 AllocateNode* allocation = NULL;
1051 int allocation_index = -1;
1052 int load_index = -1;
1053 for (uint i = 1; i < base->req(); i++) {
1054 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1055 if (allocation != NULL) {
1056 allocation_index = i;
1057 load_index = 3 - allocation_index;
1058 break;
1059 }
1060 }
1061 LoadNode* load = NULL;
1062 if (allocation != NULL && base->in(load_index)->is_Load()) {
1063 load = base->in(load_index)->as_Load();
1064 }
1065 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1066 // Push the loads from the phi that comes from valueOf up
1067 // through it to allow elimination of the loads and the recovery
1068 // of the original value.
1069 Node* mem_phi = in(Memory);
1070 Node* offset = in(Address)->in(AddPNode::Offset);
1072 Node* in1 = clone();
1073 Node* in1_addr = in1->in(Address)->clone();
1074 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1075 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1076 in1_addr->set_req(AddPNode::Offset, offset);
1077 in1->set_req(0, base->in(allocation_index));
1078 in1->set_req(Address, in1_addr);
1079 in1->set_req(Memory, mem_phi->in(allocation_index));
1081 Node* in2 = clone();
1082 Node* in2_addr = in2->in(Address)->clone();
1083 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1084 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1085 in2_addr->set_req(AddPNode::Offset, offset);
1086 in2->set_req(0, base->in(load_index));
1087 in2->set_req(Address, in2_addr);
1088 in2->set_req(Memory, mem_phi->in(load_index));
1090 in1_addr = phase->transform(in1_addr);
1091 in1 = phase->transform(in1);
1092 in2_addr = phase->transform(in2_addr);
1093 in2 = phase->transform(in2);
1095 PhiNode* result = PhiNode::make_blank(base->in(0), this);
1096 result->set_req(allocation_index, in1);
1097 result->set_req(load_index, in2);
1098 return result;
1099 }
1100 } else if (base->is_Load()) {
1101 // Eliminate the load of Integer.value for integers from the cache
1102 // array by deriving the value from the index into the array.
1103 // Capture the offset of the load and then reverse the computation.
1104 Node* load_base = base->in(Address)->in(AddPNode::Base);
1105 if (load_base != NULL) {
1106 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1107 intptr_t cache_offset;
1108 int shift = -1;
1109 Node* cache = NULL;
1110 if (is_autobox_cache(atp)) {
1111 shift = exact_log2(type2aelembytes(T_OBJECT));
1112 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1113 }
1114 if (cache != NULL && base->in(Address)->is_AddP()) {
1115 Node* elements[4];
1116 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1117 int cache_low;
1118 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1119 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1120 // Add up all the offsets making of the address of the load
1121 Node* result = elements[0];
1122 for (int i = 1; i < count; i++) {
1123 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1124 }
1125 // Remove the constant offset from the address and then
1126 // remove the scaling of the offset to recover the original index.
1127 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1128 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1129 // Peel the shift off directly but wrap it in a dummy node
1130 // since Ideal can't return existing nodes
1131 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1132 } else {
1133 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1134 }
1135 #ifdef _LP64
1136 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1137 #endif
1138 return result;
1139 }
1140 }
1141 }
1142 }
1143 return NULL;
1144 }
1146 //------------------------------split_through_phi------------------------------
1147 // Split instance field load through Phi.
1148 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1149 Node* mem = in(MemNode::Memory);
1150 Node* address = in(MemNode::Address);
1151 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1152 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1154 assert(mem->is_Phi() && (t_oop != NULL) &&
1155 t_oop->is_known_instance_field(), "invalide conditions");
1157 Node *region = mem->in(0);
1158 if (region == NULL) {
1159 return NULL; // Wait stable graph
1160 }
1161 uint cnt = mem->req();
1162 for( uint i = 1; i < cnt; i++ ) {
1163 Node *in = mem->in(i);
1164 if( in == NULL ) {
1165 return NULL; // Wait stable graph
1166 }
1167 }
1168 // Check for loop invariant.
1169 if (cnt == 3) {
1170 for( uint i = 1; i < cnt; i++ ) {
1171 Node *in = mem->in(i);
1172 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1173 if (m == mem) {
1174 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
1175 return this;
1176 }
1177 }
1178 }
1179 // Split through Phi (see original code in loopopts.cpp).
1180 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1182 // Do nothing here if Identity will find a value
1183 // (to avoid infinite chain of value phis generation).
1184 if ( !phase->eqv(this, this->Identity(phase)) )
1185 return NULL;
1187 // Skip the split if the region dominates some control edge of the address.
1188 if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
1189 return NULL;
1191 const Type* this_type = this->bottom_type();
1192 int this_index = phase->C->get_alias_index(addr_t);
1193 int this_offset = addr_t->offset();
1194 int this_iid = addr_t->is_oopptr()->instance_id();
1195 int wins = 0;
1196 PhaseIterGVN *igvn = phase->is_IterGVN();
1197 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1198 for( uint i = 1; i < region->req(); i++ ) {
1199 Node *x;
1200 Node* the_clone = NULL;
1201 if( region->in(i) == phase->C->top() ) {
1202 x = phase->C->top(); // Dead path? Use a dead data op
1203 } else {
1204 x = this->clone(); // Else clone up the data op
1205 the_clone = x; // Remember for possible deletion.
1206 // Alter data node to use pre-phi inputs
1207 if( this->in(0) == region ) {
1208 x->set_req( 0, region->in(i) );
1209 } else {
1210 x->set_req( 0, NULL );
1211 }
1212 for( uint j = 1; j < this->req(); j++ ) {
1213 Node *in = this->in(j);
1214 if( in->is_Phi() && in->in(0) == region )
1215 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1216 }
1217 }
1218 // Check for a 'win' on some paths
1219 const Type *t = x->Value(igvn);
1221 bool singleton = t->singleton();
1223 // See comments in PhaseIdealLoop::split_thru_phi().
1224 if( singleton && t == Type::TOP ) {
1225 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1226 }
1228 if( singleton ) {
1229 wins++;
1230 x = igvn->makecon(t);
1231 } else {
1232 // We now call Identity to try to simplify the cloned node.
1233 // Note that some Identity methods call phase->type(this).
1234 // Make sure that the type array is big enough for
1235 // our new node, even though we may throw the node away.
1236 // (This tweaking with igvn only works because x is a new node.)
1237 igvn->set_type(x, t);
1238 // If x is a TypeNode, capture any more-precise type permanently into Node
1239 // othewise it will be not updated during igvn->transform since
1240 // igvn->type(x) is set to x->Value() already.
1241 x->raise_bottom_type(t);
1242 Node *y = x->Identity(igvn);
1243 if( y != x ) {
1244 wins++;
1245 x = y;
1246 } else {
1247 y = igvn->hash_find(x);
1248 if( y ) {
1249 wins++;
1250 x = y;
1251 } else {
1252 // Else x is a new node we are keeping
1253 // We do not need register_new_node_with_optimizer
1254 // because set_type has already been called.
1255 igvn->_worklist.push(x);
1256 }
1257 }
1258 }
1259 if (x != the_clone && the_clone != NULL)
1260 igvn->remove_dead_node(the_clone);
1261 phi->set_req(i, x);
1262 }
1263 if( wins > 0 ) {
1264 // Record Phi
1265 igvn->register_new_node_with_optimizer(phi);
1266 return phi;
1267 }
1268 igvn->remove_dead_node(phi);
1269 return NULL;
1270 }
1272 //------------------------------Ideal------------------------------------------
1273 // If the load is from Field memory and the pointer is non-null, we can
1274 // zero out the control input.
1275 // If the offset is constant and the base is an object allocation,
1276 // try to hook me up to the exact initializing store.
1277 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1278 Node* p = MemNode::Ideal_common(phase, can_reshape);
1279 if (p) return (p == NodeSentinel) ? NULL : p;
1281 Node* ctrl = in(MemNode::Control);
1282 Node* address = in(MemNode::Address);
1284 // Skip up past a SafePoint control. Cannot do this for Stores because
1285 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1286 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1287 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1288 ctrl = ctrl->in(0);
1289 set_req(MemNode::Control,ctrl);
1290 }
1292 // Check for useless control edge in some common special cases
1293 if (in(MemNode::Control) != NULL) {
1294 intptr_t ignore = 0;
1295 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1296 if (base != NULL
1297 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1298 && all_controls_dominate(base, phase->C->start())) {
1299 // A method-invariant, non-null address (constant or 'this' argument).
1300 set_req(MemNode::Control, NULL);
1301 }
1302 }
1304 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
1305 Node* base = in(Address)->in(AddPNode::Base);
1306 if (base != NULL) {
1307 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1308 if (is_autobox_object(atp)) {
1309 Node* result = eliminate_autobox(phase);
1310 if (result != NULL) return result;
1311 }
1312 }
1313 }
1315 Node* mem = in(MemNode::Memory);
1316 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1318 if (addr_t != NULL) {
1319 // try to optimize our memory input
1320 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1321 if (opt_mem != mem) {
1322 set_req(MemNode::Memory, opt_mem);
1323 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1324 return this;
1325 }
1326 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1327 if (can_reshape && opt_mem->is_Phi() &&
1328 (t_oop != NULL) && t_oop->is_known_instance_field()) {
1329 // Split instance field load through Phi.
1330 Node* result = split_through_phi(phase);
1331 if (result != NULL) return result;
1332 }
1333 }
1335 // Check for prior store with a different base or offset; make Load
1336 // independent. Skip through any number of them. Bail out if the stores
1337 // are in an endless dead cycle and report no progress. This is a key
1338 // transform for Reflection. However, if after skipping through the Stores
1339 // we can't then fold up against a prior store do NOT do the transform as
1340 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1341 // array memory alive twice: once for the hoisted Load and again after the
1342 // bypassed Store. This situation only works if EVERYBODY who does
1343 // anti-dependence work knows how to bypass. I.e. we need all
1344 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1345 // the alias index stuff. So instead, peek through Stores and IFF we can
1346 // fold up, do so.
1347 Node* prev_mem = find_previous_store(phase);
1348 // Steps (a), (b): Walk past independent stores to find an exact match.
1349 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1350 // (c) See if we can fold up on the spot, but don't fold up here.
1351 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
1352 // just return a prior value, which is done by Identity calls.
1353 if (can_see_stored_value(prev_mem, phase)) {
1354 // Make ready for step (d):
1355 set_req(MemNode::Memory, prev_mem);
1356 return this;
1357 }
1358 }
1360 return NULL; // No further progress
1361 }
1363 // Helper to recognize certain Klass fields which are invariant across
1364 // some group of array types (e.g., int[] or all T[] where T < Object).
1365 const Type*
1366 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1367 ciKlass* klass) const {
1368 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1369 // The field is Klass::_modifier_flags. Return its (constant) value.
1370 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1371 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1372 return TypeInt::make(klass->modifier_flags());
1373 }
1374 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1375 // The field is Klass::_access_flags. Return its (constant) value.
1376 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1377 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1378 return TypeInt::make(klass->access_flags());
1379 }
1380 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1381 // The field is Klass::_layout_helper. Return its constant value if known.
1382 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1383 return TypeInt::make(klass->layout_helper());
1384 }
1386 // No match.
1387 return NULL;
1388 }
1390 //------------------------------Value-----------------------------------------
1391 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1392 // Either input is TOP ==> the result is TOP
1393 Node* mem = in(MemNode::Memory);
1394 const Type *t1 = phase->type(mem);
1395 if (t1 == Type::TOP) return Type::TOP;
1396 Node* adr = in(MemNode::Address);
1397 const TypePtr* tp = phase->type(adr)->isa_ptr();
1398 if (tp == NULL || tp->empty()) return Type::TOP;
1399 int off = tp->offset();
1400 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1402 // Try to guess loaded type from pointer type
1403 if (tp->base() == Type::AryPtr) {
1404 const Type *t = tp->is_aryptr()->elem();
1405 // Don't do this for integer types. There is only potential profit if
1406 // the element type t is lower than _type; that is, for int types, if _type is
1407 // more restrictive than t. This only happens here if one is short and the other
1408 // char (both 16 bits), and in those cases we've made an intentional decision
1409 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1410 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1411 //
1412 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1413 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1414 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1415 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1416 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1417 // In fact, that could have been the original type of p1, and p1 could have
1418 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1419 // expression (LShiftL quux 3) independently optimized to the constant 8.
1420 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1421 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1422 // t might actually be lower than _type, if _type is a unique
1423 // concrete subclass of abstract class t.
1424 // Make sure the reference is not into the header, by comparing
1425 // the offset against the offset of the start of the array's data.
1426 // Different array types begin at slightly different offsets (12 vs. 16).
1427 // We choose T_BYTE as an example base type that is least restrictive
1428 // as to alignment, which will therefore produce the smallest
1429 // possible base offset.
1430 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1431 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1432 const Type* jt = t->join(_type);
1433 // In any case, do not allow the join, per se, to empty out the type.
1434 if (jt->empty() && !t->empty()) {
1435 // This can happen if a interface-typed array narrows to a class type.
1436 jt = _type;
1437 }
1439 if (EliminateAutoBox) {
1440 // The pointers in the autobox arrays are always non-null
1441 Node* base = in(Address)->in(AddPNode::Base);
1442 if (base != NULL) {
1443 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1444 if (is_autobox_cache(atp)) {
1445 return jt->join(TypePtr::NOTNULL)->is_ptr();
1446 }
1447 }
1448 }
1449 return jt;
1450 }
1451 }
1452 } else if (tp->base() == Type::InstPtr) {
1453 assert( off != Type::OffsetBot ||
1454 // arrays can be cast to Objects
1455 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1456 // unsafe field access may not have a constant offset
1457 phase->C->has_unsafe_access(),
1458 "Field accesses must be precise" );
1459 // For oop loads, we expect the _type to be precise
1460 } else if (tp->base() == Type::KlassPtr) {
1461 assert( off != Type::OffsetBot ||
1462 // arrays can be cast to Objects
1463 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1464 // also allow array-loading from the primary supertype
1465 // array during subtype checks
1466 Opcode() == Op_LoadKlass,
1467 "Field accesses must be precise" );
1468 // For klass/static loads, we expect the _type to be precise
1469 }
1471 const TypeKlassPtr *tkls = tp->isa_klassptr();
1472 if (tkls != NULL && !StressReflectiveCode) {
1473 ciKlass* klass = tkls->klass();
1474 if (klass->is_loaded() && tkls->klass_is_exact()) {
1475 // We are loading a field from a Klass metaobject whose identity
1476 // is known at compile time (the type is "exact" or "precise").
1477 // Check for fields we know are maintained as constants by the VM.
1478 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1479 // The field is Klass::_super_check_offset. Return its (constant) value.
1480 // (Folds up type checking code.)
1481 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1482 return TypeInt::make(klass->super_check_offset());
1483 }
1484 // Compute index into primary_supers array
1485 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1486 // Check for overflowing; use unsigned compare to handle the negative case.
1487 if( depth < ciKlass::primary_super_limit() ) {
1488 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1489 // (Folds up type checking code.)
1490 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1491 ciKlass *ss = klass->super_of_depth(depth);
1492 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1493 }
1494 const Type* aift = load_array_final_field(tkls, klass);
1495 if (aift != NULL) return aift;
1496 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1497 && klass->is_array_klass()) {
1498 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1499 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1500 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1501 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1502 }
1503 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1504 // The field is Klass::_java_mirror. Return its (constant) value.
1505 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1506 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1507 return TypeInstPtr::make(klass->java_mirror());
1508 }
1509 }
1511 // We can still check if we are loading from the primary_supers array at a
1512 // shallow enough depth. Even though the klass is not exact, entries less
1513 // than or equal to its super depth are correct.
1514 if (klass->is_loaded() ) {
1515 ciType *inner = klass->klass();
1516 while( inner->is_obj_array_klass() )
1517 inner = inner->as_obj_array_klass()->base_element_type();
1518 if( inner->is_instance_klass() &&
1519 !inner->as_instance_klass()->flags().is_interface() ) {
1520 // Compute index into primary_supers array
1521 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1522 // Check for overflowing; use unsigned compare to handle the negative case.
1523 if( depth < ciKlass::primary_super_limit() &&
1524 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1525 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1526 // (Folds up type checking code.)
1527 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1528 ciKlass *ss = klass->super_of_depth(depth);
1529 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1530 }
1531 }
1532 }
1534 // If the type is enough to determine that the thing is not an array,
1535 // we can give the layout_helper a positive interval type.
1536 // This will help short-circuit some reflective code.
1537 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1538 && !klass->is_array_klass() // not directly typed as an array
1539 && !klass->is_interface() // specifically not Serializable & Cloneable
1540 && !klass->is_java_lang_Object() // not the supertype of all T[]
1541 ) {
1542 // Note: When interfaces are reliable, we can narrow the interface
1543 // test to (klass != Serializable && klass != Cloneable).
1544 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1545 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1546 // The key property of this type is that it folds up tests
1547 // for array-ness, since it proves that the layout_helper is positive.
1548 // Thus, a generic value like the basic object layout helper works fine.
1549 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1550 }
1551 }
1553 // If we are loading from a freshly-allocated object, produce a zero,
1554 // if the load is provably beyond the header of the object.
1555 // (Also allow a variable load from a fresh array to produce zero.)
1556 if (ReduceFieldZeroing) {
1557 Node* value = can_see_stored_value(mem,phase);
1558 if (value != NULL && value->is_Con())
1559 return value->bottom_type();
1560 }
1562 const TypeOopPtr *tinst = tp->isa_oopptr();
1563 if (tinst != NULL && tinst->is_known_instance_field()) {
1564 // If we have an instance type and our memory input is the
1565 // programs's initial memory state, there is no matching store,
1566 // so just return a zero of the appropriate type
1567 Node *mem = in(MemNode::Memory);
1568 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1569 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1570 return Type::get_zero_type(_type->basic_type());
1571 }
1572 }
1573 return _type;
1574 }
1576 //------------------------------match_edge-------------------------------------
1577 // Do we Match on this edge index or not? Match only the address.
1578 uint LoadNode::match_edge(uint idx) const {
1579 return idx == MemNode::Address;
1580 }
1582 //--------------------------LoadBNode::Ideal--------------------------------------
1583 //
1584 // If the previous store is to the same address as this load,
1585 // and the value stored was larger than a byte, replace this load
1586 // with the value stored truncated to a byte. If no truncation is
1587 // needed, the replacement is done in LoadNode::Identity().
1588 //
1589 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1590 Node* mem = in(MemNode::Memory);
1591 Node* value = can_see_stored_value(mem,phase);
1592 if( value && !phase->type(value)->higher_equal( _type ) ) {
1593 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1594 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1595 }
1596 // Identity call will handle the case where truncation is not needed.
1597 return LoadNode::Ideal(phase, can_reshape);
1598 }
1600 //--------------------------LoadCNode::Ideal--------------------------------------
1601 //
1602 // If the previous store is to the same address as this load,
1603 // and the value stored was larger than a char, replace this load
1604 // with the value stored truncated to a char. If no truncation is
1605 // needed, the replacement is done in LoadNode::Identity().
1606 //
1607 Node *LoadCNode::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 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1612 // Identity call will handle the case where truncation is not needed.
1613 return LoadNode::Ideal(phase, can_reshape);
1614 }
1616 //--------------------------LoadSNode::Ideal--------------------------------------
1617 //
1618 // If the previous store is to the same address as this load,
1619 // and the value stored was larger than a short, replace this load
1620 // with the value stored truncated to a short. If no truncation is
1621 // needed, the replacement is done in LoadNode::Identity().
1622 //
1623 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1624 Node* mem = in(MemNode::Memory);
1625 Node* value = can_see_stored_value(mem,phase);
1626 if( value && !phase->type(value)->higher_equal( _type ) ) {
1627 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1628 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1629 }
1630 // Identity call will handle the case where truncation is not needed.
1631 return LoadNode::Ideal(phase, can_reshape);
1632 }
1634 //=============================================================================
1635 //----------------------------LoadKlassNode::make------------------------------
1636 // Polymorphic factory method:
1637 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1638 Compile* C = gvn.C;
1639 Node *ctl = NULL;
1640 // sanity check the alias category against the created node type
1641 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
1642 assert(adr_type != NULL, "expecting TypeOopPtr");
1643 #ifdef _LP64
1644 if (adr_type->is_ptr_to_narrowoop()) {
1645 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
1646 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
1647 }
1648 #endif
1649 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1650 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
1651 }
1653 //------------------------------Value------------------------------------------
1654 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1655 return klass_value_common(phase);
1656 }
1658 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1659 // Either input is TOP ==> the result is TOP
1660 const Type *t1 = phase->type( in(MemNode::Memory) );
1661 if (t1 == Type::TOP) return Type::TOP;
1662 Node *adr = in(MemNode::Address);
1663 const Type *t2 = phase->type( adr );
1664 if (t2 == Type::TOP) return Type::TOP;
1665 const TypePtr *tp = t2->is_ptr();
1666 if (TypePtr::above_centerline(tp->ptr()) ||
1667 tp->ptr() == TypePtr::Null) return Type::TOP;
1669 // Return a more precise klass, if possible
1670 const TypeInstPtr *tinst = tp->isa_instptr();
1671 if (tinst != NULL) {
1672 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1673 int offset = tinst->offset();
1674 if (ik == phase->C->env()->Class_klass()
1675 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1676 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1677 // We are loading a special hidden field from a Class mirror object,
1678 // the field which points to the VM's Klass metaobject.
1679 ciType* t = tinst->java_mirror_type();
1680 // java_mirror_type returns non-null for compile-time Class constants.
1681 if (t != NULL) {
1682 // constant oop => constant klass
1683 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1684 return TypeKlassPtr::make(ciArrayKlass::make(t));
1685 }
1686 if (!t->is_klass()) {
1687 // a primitive Class (e.g., int.class) has NULL for a klass field
1688 return TypePtr::NULL_PTR;
1689 }
1690 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1691 return TypeKlassPtr::make(t->as_klass());
1692 }
1693 // non-constant mirror, so we can't tell what's going on
1694 }
1695 if( !ik->is_loaded() )
1696 return _type; // Bail out if not loaded
1697 if (offset == oopDesc::klass_offset_in_bytes()) {
1698 if (tinst->klass_is_exact()) {
1699 return TypeKlassPtr::make(ik);
1700 }
1701 // See if we can become precise: no subklasses and no interface
1702 // (Note: We need to support verified interfaces.)
1703 if (!ik->is_interface() && !ik->has_subklass()) {
1704 //assert(!UseExactTypes, "this code should be useless with exact types");
1705 // Add a dependence; if any subclass added we need to recompile
1706 if (!ik->is_final()) {
1707 // %%% should use stronger assert_unique_concrete_subtype instead
1708 phase->C->dependencies()->assert_leaf_type(ik);
1709 }
1710 // Return precise klass
1711 return TypeKlassPtr::make(ik);
1712 }
1714 // Return root of possible klass
1715 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1716 }
1717 }
1719 // Check for loading klass from an array
1720 const TypeAryPtr *tary = tp->isa_aryptr();
1721 if( tary != NULL ) {
1722 ciKlass *tary_klass = tary->klass();
1723 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1724 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1725 if (tary->klass_is_exact()) {
1726 return TypeKlassPtr::make(tary_klass);
1727 }
1728 ciArrayKlass *ak = tary->klass()->as_array_klass();
1729 // If the klass is an object array, we defer the question to the
1730 // array component klass.
1731 if( ak->is_obj_array_klass() ) {
1732 assert( ak->is_loaded(), "" );
1733 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1734 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1735 ciInstanceKlass* ik = base_k->as_instance_klass();
1736 // See if we can become precise: no subklasses and no interface
1737 if (!ik->is_interface() && !ik->has_subklass()) {
1738 //assert(!UseExactTypes, "this code should be useless with exact types");
1739 // Add a dependence; if any subclass added we need to recompile
1740 if (!ik->is_final()) {
1741 phase->C->dependencies()->assert_leaf_type(ik);
1742 }
1743 // Return precise array klass
1744 return TypeKlassPtr::make(ak);
1745 }
1746 }
1747 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1748 } else { // Found a type-array?
1749 //assert(!UseExactTypes, "this code should be useless with exact types");
1750 assert( ak->is_type_array_klass(), "" );
1751 return TypeKlassPtr::make(ak); // These are always precise
1752 }
1753 }
1754 }
1756 // Check for loading klass from an array klass
1757 const TypeKlassPtr *tkls = tp->isa_klassptr();
1758 if (tkls != NULL && !StressReflectiveCode) {
1759 ciKlass* klass = tkls->klass();
1760 if( !klass->is_loaded() )
1761 return _type; // Bail out if not loaded
1762 if( klass->is_obj_array_klass() &&
1763 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1764 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1765 // // Always returning precise element type is incorrect,
1766 // // e.g., element type could be object and array may contain strings
1767 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1769 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1770 // according to the element type's subclassing.
1771 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1772 }
1773 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1774 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1775 ciKlass* sup = klass->as_instance_klass()->super();
1776 // The field is Klass::_super. Return its (constant) value.
1777 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1778 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1779 }
1780 }
1782 // Bailout case
1783 return LoadNode::Value(phase);
1784 }
1786 //------------------------------Identity---------------------------------------
1787 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1788 // Also feed through the klass in Allocate(...klass...)._klass.
1789 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1790 return klass_identity_common(phase);
1791 }
1793 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
1794 Node* x = LoadNode::Identity(phase);
1795 if (x != this) return x;
1797 // Take apart the address into an oop and and offset.
1798 // Return 'this' if we cannot.
1799 Node* adr = in(MemNode::Address);
1800 intptr_t offset = 0;
1801 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1802 if (base == NULL) return this;
1803 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1804 if (toop == NULL) return this;
1806 // We can fetch the klass directly through an AllocateNode.
1807 // This works even if the klass is not constant (clone or newArray).
1808 if (offset == oopDesc::klass_offset_in_bytes()) {
1809 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1810 if (allocated_klass != NULL) {
1811 return allocated_klass;
1812 }
1813 }
1815 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1816 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1817 // See inline_native_Class_query for occurrences of these patterns.
1818 // Java Example: x.getClass().isAssignableFrom(y)
1819 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1820 //
1821 // This improves reflective code, often making the Class
1822 // mirror go completely dead. (Current exception: Class
1823 // mirrors may appear in debug info, but we could clean them out by
1824 // introducing a new debug info operator for klassOop.java_mirror).
1825 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1826 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1827 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1828 // We are loading a special hidden field from a Class mirror,
1829 // the field which points to its Klass or arrayKlass metaobject.
1830 if (base->is_Load()) {
1831 Node* adr2 = base->in(MemNode::Address);
1832 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1833 if (tkls != NULL && !tkls->empty()
1834 && (tkls->klass()->is_instance_klass() ||
1835 tkls->klass()->is_array_klass())
1836 && adr2->is_AddP()
1837 ) {
1838 int mirror_field = Klass::java_mirror_offset_in_bytes();
1839 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1840 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1841 }
1842 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1843 return adr2->in(AddPNode::Base);
1844 }
1845 }
1846 }
1847 }
1849 return this;
1850 }
1853 //------------------------------Value------------------------------------------
1854 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
1855 const Type *t = klass_value_common(phase);
1856 if (t == Type::TOP)
1857 return t;
1859 return t->make_narrowoop();
1860 }
1862 //------------------------------Identity---------------------------------------
1863 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
1864 // Also feed through the klass in Allocate(...klass...)._klass.
1865 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
1866 Node *x = klass_identity_common(phase);
1868 const Type *t = phase->type( x );
1869 if( t == Type::TOP ) return x;
1870 if( t->isa_narrowoop()) return x;
1872 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
1873 }
1875 //------------------------------Value-----------------------------------------
1876 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1877 // Either input is TOP ==> the result is TOP
1878 const Type *t1 = phase->type( in(MemNode::Memory) );
1879 if( t1 == Type::TOP ) return Type::TOP;
1880 Node *adr = in(MemNode::Address);
1881 const Type *t2 = phase->type( adr );
1882 if( t2 == Type::TOP ) return Type::TOP;
1883 const TypePtr *tp = t2->is_ptr();
1884 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1885 const TypeAryPtr *tap = tp->isa_aryptr();
1886 if( !tap ) return _type;
1887 return tap->size();
1888 }
1890 //-------------------------------Ideal---------------------------------------
1891 // Feed through the length in AllocateArray(...length...)._length.
1892 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1893 Node* p = MemNode::Ideal_common(phase, can_reshape);
1894 if (p) return (p == NodeSentinel) ? NULL : p;
1896 // Take apart the address into an oop and and offset.
1897 // Return 'this' if we cannot.
1898 Node* adr = in(MemNode::Address);
1899 intptr_t offset = 0;
1900 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1901 if (base == NULL) return NULL;
1902 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1903 if (tary == NULL) return NULL;
1905 // We can fetch the length directly through an AllocateArrayNode.
1906 // This works even if the length is not constant (clone or newArray).
1907 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1908 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
1909 if (alloc != NULL) {
1910 Node* allocated_length = alloc->Ideal_length();
1911 Node* len = alloc->make_ideal_length(tary, phase);
1912 if (allocated_length != len) {
1913 // New CastII improves on this.
1914 return len;
1915 }
1916 }
1917 }
1919 return NULL;
1920 }
1922 //------------------------------Identity---------------------------------------
1923 // Feed through the length in AllocateArray(...length...)._length.
1924 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1925 Node* x = LoadINode::Identity(phase);
1926 if (x != this) return x;
1928 // Take apart the address into an oop and and offset.
1929 // Return 'this' if we cannot.
1930 Node* adr = in(MemNode::Address);
1931 intptr_t offset = 0;
1932 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1933 if (base == NULL) return this;
1934 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1935 if (tary == NULL) return this;
1937 // We can fetch the length directly through an AllocateArrayNode.
1938 // This works even if the length is not constant (clone or newArray).
1939 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1940 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
1941 if (alloc != NULL) {
1942 Node* allocated_length = alloc->Ideal_length();
1943 // Do not allow make_ideal_length to allocate a CastII node.
1944 Node* len = alloc->make_ideal_length(tary, phase, false);
1945 if (allocated_length == len) {
1946 // Return allocated_length only if it would not be improved by a CastII.
1947 return allocated_length;
1948 }
1949 }
1950 }
1952 return this;
1954 }
1956 //=============================================================================
1957 //---------------------------StoreNode::make-----------------------------------
1958 // Polymorphic factory method:
1959 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1960 Compile* C = gvn.C;
1962 switch (bt) {
1963 case T_BOOLEAN:
1964 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1965 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1966 case T_CHAR:
1967 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1968 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1969 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1970 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1971 case T_ADDRESS:
1972 case T_OBJECT:
1973 #ifdef _LP64
1974 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
1975 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
1976 adr->bottom_type()->isa_rawptr())) {
1977 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
1978 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
1979 } else
1980 #endif
1981 {
1982 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1983 }
1984 }
1985 ShouldNotReachHere();
1986 return (StoreNode*)NULL;
1987 }
1989 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1990 bool require_atomic = true;
1991 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1992 }
1995 //--------------------------bottom_type----------------------------------------
1996 const Type *StoreNode::bottom_type() const {
1997 return Type::MEMORY;
1998 }
2000 //------------------------------hash-------------------------------------------
2001 uint StoreNode::hash() const {
2002 // unroll addition of interesting fields
2003 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2005 // Since they are not commoned, do not hash them:
2006 return NO_HASH;
2007 }
2009 //------------------------------Ideal------------------------------------------
2010 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2011 // When a store immediately follows a relevant allocation/initialization,
2012 // try to capture it into the initialization, or hoist it above.
2013 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2014 Node* p = MemNode::Ideal_common(phase, can_reshape);
2015 if (p) return (p == NodeSentinel) ? NULL : p;
2017 Node* mem = in(MemNode::Memory);
2018 Node* address = in(MemNode::Address);
2020 // Back-to-back stores to same address? Fold em up.
2021 // Generally unsafe if I have intervening uses...
2022 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
2023 // Looking at a dead closed cycle of memory?
2024 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2026 assert(Opcode() == mem->Opcode() ||
2027 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2028 "no mismatched stores, except on raw memory");
2030 if (mem->outcnt() == 1 && // check for intervening uses
2031 mem->as_Store()->memory_size() <= this->memory_size()) {
2032 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2033 // For example, 'mem' might be the final state at a conditional return.
2034 // Or, 'mem' might be used by some node which is live at the same time
2035 // 'this' is live, which might be unschedulable. So, require exactly
2036 // ONE user, the 'this' store, until such time as we clone 'mem' for
2037 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2038 if (can_reshape) { // (%%% is this an anachronism?)
2039 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2040 phase->is_IterGVN());
2041 } else {
2042 // It's OK to do this in the parser, since DU info is always accurate,
2043 // and the parser always refers to nodes via SafePointNode maps.
2044 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2045 }
2046 return this;
2047 }
2048 }
2050 // Capture an unaliased, unconditional, simple store into an initializer.
2051 // Or, if it is independent of the allocation, hoist it above the allocation.
2052 if (ReduceFieldZeroing && /*can_reshape &&*/
2053 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2054 InitializeNode* init = mem->in(0)->as_Initialize();
2055 intptr_t offset = init->can_capture_store(this, phase);
2056 if (offset > 0) {
2057 Node* moved = init->capture_store(this, offset, phase);
2058 // If the InitializeNode captured me, it made a raw copy of me,
2059 // and I need to disappear.
2060 if (moved != NULL) {
2061 // %%% hack to ensure that Ideal returns a new node:
2062 mem = MergeMemNode::make(phase->C, mem);
2063 return mem; // fold me away
2064 }
2065 }
2066 }
2068 return NULL; // No further progress
2069 }
2071 //------------------------------Value-----------------------------------------
2072 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2073 // Either input is TOP ==> the result is TOP
2074 const Type *t1 = phase->type( in(MemNode::Memory) );
2075 if( t1 == Type::TOP ) return Type::TOP;
2076 const Type *t2 = phase->type( in(MemNode::Address) );
2077 if( t2 == Type::TOP ) return Type::TOP;
2078 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2079 if( t3 == Type::TOP ) return Type::TOP;
2080 return Type::MEMORY;
2081 }
2083 //------------------------------Identity---------------------------------------
2084 // Remove redundant stores:
2085 // Store(m, p, Load(m, p)) changes to m.
2086 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2087 Node *StoreNode::Identity( PhaseTransform *phase ) {
2088 Node* mem = in(MemNode::Memory);
2089 Node* adr = in(MemNode::Address);
2090 Node* val = in(MemNode::ValueIn);
2092 // Load then Store? Then the Store is useless
2093 if (val->is_Load() &&
2094 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
2095 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
2096 val->as_Load()->store_Opcode() == Opcode()) {
2097 return mem;
2098 }
2100 // Two stores in a row of the same value?
2101 if (mem->is_Store() &&
2102 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
2103 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
2104 mem->Opcode() == Opcode()) {
2105 return mem;
2106 }
2108 // Store of zero anywhere into a freshly-allocated object?
2109 // Then the store is useless.
2110 // (It must already have been captured by the InitializeNode.)
2111 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2112 // a newly allocated object is already all-zeroes everywhere
2113 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2114 return mem;
2115 }
2117 // the store may also apply to zero-bits in an earlier object
2118 Node* prev_mem = find_previous_store(phase);
2119 // Steps (a), (b): Walk past independent stores to find an exact match.
2120 if (prev_mem != NULL) {
2121 Node* prev_val = can_see_stored_value(prev_mem, phase);
2122 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2123 // prev_val and val might differ by a cast; it would be good
2124 // to keep the more informative of the two.
2125 return mem;
2126 }
2127 }
2128 }
2130 return this;
2131 }
2133 //------------------------------match_edge-------------------------------------
2134 // Do we Match on this edge index or not? Match only memory & value
2135 uint StoreNode::match_edge(uint idx) const {
2136 return idx == MemNode::Address || idx == MemNode::ValueIn;
2137 }
2139 //------------------------------cmp--------------------------------------------
2140 // Do not common stores up together. They generally have to be split
2141 // back up anyways, so do not bother.
2142 uint StoreNode::cmp( const Node &n ) const {
2143 return (&n == this); // Always fail except on self
2144 }
2146 //------------------------------Ideal_masked_input-----------------------------
2147 // Check for a useless mask before a partial-word store
2148 // (StoreB ... (AndI valIn conIa) )
2149 // If (conIa & mask == mask) this simplifies to
2150 // (StoreB ... (valIn) )
2151 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2152 Node *val = in(MemNode::ValueIn);
2153 if( val->Opcode() == Op_AndI ) {
2154 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2155 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2156 set_req(MemNode::ValueIn, val->in(1));
2157 return this;
2158 }
2159 }
2160 return NULL;
2161 }
2164 //------------------------------Ideal_sign_extended_input----------------------
2165 // Check for useless sign-extension before a partial-word store
2166 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2167 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2168 // (StoreB ... (valIn) )
2169 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2170 Node *val = in(MemNode::ValueIn);
2171 if( val->Opcode() == Op_RShiftI ) {
2172 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2173 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2174 Node *shl = val->in(1);
2175 if( shl->Opcode() == Op_LShiftI ) {
2176 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2177 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2178 set_req(MemNode::ValueIn, shl->in(1));
2179 return this;
2180 }
2181 }
2182 }
2183 }
2184 return NULL;
2185 }
2187 //------------------------------value_never_loaded-----------------------------------
2188 // Determine whether there are any possible loads of the value stored.
2189 // For simplicity, we actually check if there are any loads from the
2190 // address stored to, not just for loads of the value stored by this node.
2191 //
2192 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2193 Node *adr = in(Address);
2194 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2195 if (adr_oop == NULL)
2196 return false;
2197 if (!adr_oop->is_known_instance_field())
2198 return false; // if not a distinct instance, there may be aliases of the address
2199 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2200 Node *use = adr->fast_out(i);
2201 int opc = use->Opcode();
2202 if (use->is_Load() || use->is_LoadStore()) {
2203 return false;
2204 }
2205 }
2206 return true;
2207 }
2209 //=============================================================================
2210 //------------------------------Ideal------------------------------------------
2211 // If the store is from an AND mask that leaves the low bits untouched, then
2212 // we can skip the AND operation. If the store is from a sign-extension
2213 // (a left shift, then right shift) we can skip both.
2214 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2215 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2216 if( progress != NULL ) return progress;
2218 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2219 if( progress != NULL ) return progress;
2221 // Finally check the default case
2222 return StoreNode::Ideal(phase, can_reshape);
2223 }
2225 //=============================================================================
2226 //------------------------------Ideal------------------------------------------
2227 // If the store is from an AND mask that leaves the low bits untouched, then
2228 // we can skip the AND operation
2229 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2230 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2231 if( progress != NULL ) return progress;
2233 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2234 if( progress != NULL ) return progress;
2236 // Finally check the default case
2237 return StoreNode::Ideal(phase, can_reshape);
2238 }
2240 //=============================================================================
2241 //------------------------------Identity---------------------------------------
2242 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2243 // No need to card mark when storing a null ptr
2244 Node* my_store = in(MemNode::OopStore);
2245 if (my_store->is_Store()) {
2246 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2247 if( t1 == TypePtr::NULL_PTR ) {
2248 return in(MemNode::Memory);
2249 }
2250 }
2251 return this;
2252 }
2254 //------------------------------Value-----------------------------------------
2255 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2256 // Either input is TOP ==> the result is TOP
2257 const Type *t = phase->type( in(MemNode::Memory) );
2258 if( t == Type::TOP ) return Type::TOP;
2259 t = phase->type( in(MemNode::Address) );
2260 if( t == Type::TOP ) return Type::TOP;
2261 t = phase->type( in(MemNode::ValueIn) );
2262 if( t == Type::TOP ) return Type::TOP;
2263 // If extra input is TOP ==> the result is TOP
2264 t = phase->type( in(MemNode::OopStore) );
2265 if( t == Type::TOP ) return Type::TOP;
2267 return StoreNode::Value( phase );
2268 }
2271 //=============================================================================
2272 //----------------------------------SCMemProjNode------------------------------
2273 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2274 {
2275 return bottom_type();
2276 }
2278 //=============================================================================
2279 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2280 init_req(MemNode::Control, c );
2281 init_req(MemNode::Memory , mem);
2282 init_req(MemNode::Address, adr);
2283 init_req(MemNode::ValueIn, val);
2284 init_req( ExpectedIn, ex );
2285 init_class_id(Class_LoadStore);
2287 }
2289 //=============================================================================
2290 //-------------------------------adr_type--------------------------------------
2291 // Do we Match on this edge index or not? Do not match memory
2292 const TypePtr* ClearArrayNode::adr_type() const {
2293 Node *adr = in(3);
2294 return MemNode::calculate_adr_type(adr->bottom_type());
2295 }
2297 //------------------------------match_edge-------------------------------------
2298 // Do we Match on this edge index or not? Do not match memory
2299 uint ClearArrayNode::match_edge(uint idx) const {
2300 return idx > 1;
2301 }
2303 //------------------------------Identity---------------------------------------
2304 // Clearing a zero length array does nothing
2305 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2306 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2307 }
2309 //------------------------------Idealize---------------------------------------
2310 // Clearing a short array is faster with stores
2311 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2312 const int unit = BytesPerLong;
2313 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2314 if (!t) return NULL;
2315 if (!t->is_con()) return NULL;
2316 intptr_t raw_count = t->get_con();
2317 intptr_t size = raw_count;
2318 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2319 // Clearing nothing uses the Identity call.
2320 // Negative clears are possible on dead ClearArrays
2321 // (see jck test stmt114.stmt11402.val).
2322 if (size <= 0 || size % unit != 0) return NULL;
2323 intptr_t count = size / unit;
2324 // Length too long; use fast hardware clear
2325 if (size > Matcher::init_array_short_size) return NULL;
2326 Node *mem = in(1);
2327 if( phase->type(mem)==Type::TOP ) return NULL;
2328 Node *adr = in(3);
2329 const Type* at = phase->type(adr);
2330 if( at==Type::TOP ) return NULL;
2331 const TypePtr* atp = at->isa_ptr();
2332 // adjust atp to be the correct array element address type
2333 if (atp == NULL) atp = TypePtr::BOTTOM;
2334 else atp = atp->add_offset(Type::OffsetBot);
2335 // Get base for derived pointer purposes
2336 if( adr->Opcode() != Op_AddP ) Unimplemented();
2337 Node *base = adr->in(1);
2339 Node *zero = phase->makecon(TypeLong::ZERO);
2340 Node *off = phase->MakeConX(BytesPerLong);
2341 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2342 count--;
2343 while( count-- ) {
2344 mem = phase->transform(mem);
2345 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2346 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2347 }
2348 return mem;
2349 }
2351 //----------------------------clear_memory-------------------------------------
2352 // Generate code to initialize object storage to zero.
2353 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2354 intptr_t start_offset,
2355 Node* end_offset,
2356 PhaseGVN* phase) {
2357 Compile* C = phase->C;
2358 intptr_t offset = start_offset;
2360 int unit = BytesPerLong;
2361 if ((offset % unit) != 0) {
2362 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2363 adr = phase->transform(adr);
2364 const TypePtr* atp = TypeRawPtr::BOTTOM;
2365 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2366 mem = phase->transform(mem);
2367 offset += BytesPerInt;
2368 }
2369 assert((offset % unit) == 0, "");
2371 // Initialize the remaining stuff, if any, with a ClearArray.
2372 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2373 }
2375 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2376 Node* start_offset,
2377 Node* end_offset,
2378 PhaseGVN* phase) {
2379 if (start_offset == end_offset) {
2380 // nothing to do
2381 return mem;
2382 }
2384 Compile* C = phase->C;
2385 int unit = BytesPerLong;
2386 Node* zbase = start_offset;
2387 Node* zend = end_offset;
2389 // Scale to the unit required by the CPU:
2390 if (!Matcher::init_array_count_is_in_bytes) {
2391 Node* shift = phase->intcon(exact_log2(unit));
2392 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2393 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2394 }
2396 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2397 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2399 // Bulk clear double-words
2400 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2401 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2402 return phase->transform(mem);
2403 }
2405 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2406 intptr_t start_offset,
2407 intptr_t end_offset,
2408 PhaseGVN* phase) {
2409 if (start_offset == end_offset) {
2410 // nothing to do
2411 return mem;
2412 }
2414 Compile* C = phase->C;
2415 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2416 intptr_t done_offset = end_offset;
2417 if ((done_offset % BytesPerLong) != 0) {
2418 done_offset -= BytesPerInt;
2419 }
2420 if (done_offset > start_offset) {
2421 mem = clear_memory(ctl, mem, dest,
2422 start_offset, phase->MakeConX(done_offset), phase);
2423 }
2424 if (done_offset < end_offset) { // emit the final 32-bit store
2425 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2426 adr = phase->transform(adr);
2427 const TypePtr* atp = TypeRawPtr::BOTTOM;
2428 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2429 mem = phase->transform(mem);
2430 done_offset += BytesPerInt;
2431 }
2432 assert(done_offset == end_offset, "");
2433 return mem;
2434 }
2436 //=============================================================================
2437 // Do we match on this edge? No memory edges
2438 uint StrCompNode::match_edge(uint idx) const {
2439 return idx == 5 || idx == 6;
2440 }
2442 //------------------------------Ideal------------------------------------------
2443 // Return a node which is more "ideal" than the current node. Strip out
2444 // control copies
2445 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2446 return remove_dead_region(phase, can_reshape) ? this : NULL;
2447 }
2449 //------------------------------Ideal------------------------------------------
2450 // Return a node which is more "ideal" than the current node. Strip out
2451 // control copies
2452 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){
2453 return remove_dead_region(phase, can_reshape) ? this : NULL;
2454 }
2457 //=============================================================================
2458 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2459 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2460 _adr_type(C->get_adr_type(alias_idx))
2461 {
2462 init_class_id(Class_MemBar);
2463 Node* top = C->top();
2464 init_req(TypeFunc::I_O,top);
2465 init_req(TypeFunc::FramePtr,top);
2466 init_req(TypeFunc::ReturnAdr,top);
2467 if (precedent != NULL)
2468 init_req(TypeFunc::Parms, precedent);
2469 }
2471 //------------------------------cmp--------------------------------------------
2472 uint MemBarNode::hash() const { return NO_HASH; }
2473 uint MemBarNode::cmp( const Node &n ) const {
2474 return (&n == this); // Always fail except on self
2475 }
2477 //------------------------------make-------------------------------------------
2478 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2479 int len = Precedent + (pn == NULL? 0: 1);
2480 switch (opcode) {
2481 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2482 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2483 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2484 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2485 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2486 default: ShouldNotReachHere(); return NULL;
2487 }
2488 }
2490 //------------------------------Ideal------------------------------------------
2491 // Return a node which is more "ideal" than the current node. Strip out
2492 // control copies
2493 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2494 return remove_dead_region(phase, can_reshape) ? this : NULL;
2495 }
2497 //------------------------------Value------------------------------------------
2498 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2499 if( !in(0) ) return Type::TOP;
2500 if( phase->type(in(0)) == Type::TOP )
2501 return Type::TOP;
2502 return TypeTuple::MEMBAR;
2503 }
2505 //------------------------------match------------------------------------------
2506 // Construct projections for memory.
2507 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2508 switch (proj->_con) {
2509 case TypeFunc::Control:
2510 case TypeFunc::Memory:
2511 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2512 }
2513 ShouldNotReachHere();
2514 return NULL;
2515 }
2517 //===========================InitializeNode====================================
2518 // SUMMARY:
2519 // This node acts as a memory barrier on raw memory, after some raw stores.
2520 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2521 // The Initialize can 'capture' suitably constrained stores as raw inits.
2522 // It can coalesce related raw stores into larger units (called 'tiles').
2523 // It can avoid zeroing new storage for memory units which have raw inits.
2524 // At macro-expansion, it is marked 'complete', and does not optimize further.
2525 //
2526 // EXAMPLE:
2527 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2528 // ctl = incoming control; mem* = incoming memory
2529 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2530 // First allocate uninitialized memory and fill in the header:
2531 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2532 // ctl := alloc.Control; mem* := alloc.Memory*
2533 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2534 // Then initialize to zero the non-header parts of the raw memory block:
2535 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2536 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2537 // After the initialize node executes, the object is ready for service:
2538 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2539 // Suppose its body is immediately initialized as {1,2}:
2540 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2541 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2542 // mem.SLICE(#short[*]) := store2
2543 //
2544 // DETAILS:
2545 // An InitializeNode collects and isolates object initialization after
2546 // an AllocateNode and before the next possible safepoint. As a
2547 // memory barrier (MemBarNode), it keeps critical stores from drifting
2548 // down past any safepoint or any publication of the allocation.
2549 // Before this barrier, a newly-allocated object may have uninitialized bits.
2550 // After this barrier, it may be treated as a real oop, and GC is allowed.
2551 //
2552 // The semantics of the InitializeNode include an implicit zeroing of
2553 // the new object from object header to the end of the object.
2554 // (The object header and end are determined by the AllocateNode.)
2555 //
2556 // Certain stores may be added as direct inputs to the InitializeNode.
2557 // These stores must update raw memory, and they must be to addresses
2558 // derived from the raw address produced by AllocateNode, and with
2559 // a constant offset. They must be ordered by increasing offset.
2560 // The first one is at in(RawStores), the last at in(req()-1).
2561 // Unlike most memory operations, they are not linked in a chain,
2562 // but are displayed in parallel as users of the rawmem output of
2563 // the allocation.
2564 //
2565 // (See comments in InitializeNode::capture_store, which continue
2566 // the example given above.)
2567 //
2568 // When the associated Allocate is macro-expanded, the InitializeNode
2569 // may be rewritten to optimize collected stores. A ClearArrayNode
2570 // may also be created at that point to represent any required zeroing.
2571 // The InitializeNode is then marked 'complete', prohibiting further
2572 // capturing of nearby memory operations.
2573 //
2574 // During macro-expansion, all captured initializations which store
2575 // constant values of 32 bits or smaller are coalesced (if advantagous)
2576 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2577 // initialized in fewer memory operations. Memory words which are
2578 // covered by neither tiles nor non-constant stores are pre-zeroed
2579 // by explicit stores of zero. (The code shape happens to do all
2580 // zeroing first, then all other stores, with both sequences occurring
2581 // in order of ascending offsets.)
2582 //
2583 // Alternatively, code may be inserted between an AllocateNode and its
2584 // InitializeNode, to perform arbitrary initialization of the new object.
2585 // E.g., the object copying intrinsics insert complex data transfers here.
2586 // The initialization must then be marked as 'complete' disable the
2587 // built-in zeroing semantics and the collection of initializing stores.
2588 //
2589 // While an InitializeNode is incomplete, reads from the memory state
2590 // produced by it are optimizable if they match the control edge and
2591 // new oop address associated with the allocation/initialization.
2592 // They return a stored value (if the offset matches) or else zero.
2593 // A write to the memory state, if it matches control and address,
2594 // and if it is to a constant offset, may be 'captured' by the
2595 // InitializeNode. It is cloned as a raw memory operation and rewired
2596 // inside the initialization, to the raw oop produced by the allocation.
2597 // Operations on addresses which are provably distinct (e.g., to
2598 // other AllocateNodes) are allowed to bypass the initialization.
2599 //
2600 // The effect of all this is to consolidate object initialization
2601 // (both arrays and non-arrays, both piecewise and bulk) into a
2602 // single location, where it can be optimized as a unit.
2603 //
2604 // Only stores with an offset less than TrackedInitializationLimit words
2605 // will be considered for capture by an InitializeNode. This puts a
2606 // reasonable limit on the complexity of optimized initializations.
2608 //---------------------------InitializeNode------------------------------------
2609 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2610 : _is_complete(false),
2611 MemBarNode(C, adr_type, rawoop)
2612 {
2613 init_class_id(Class_Initialize);
2615 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2616 assert(in(RawAddress) == rawoop, "proper init");
2617 // Note: allocation() can be NULL, for secondary initialization barriers
2618 }
2620 // Since this node is not matched, it will be processed by the
2621 // register allocator. Declare that there are no constraints
2622 // on the allocation of the RawAddress edge.
2623 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2624 // This edge should be set to top, by the set_complete. But be conservative.
2625 if (idx == InitializeNode::RawAddress)
2626 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2627 return RegMask::Empty;
2628 }
2630 Node* InitializeNode::memory(uint alias_idx) {
2631 Node* mem = in(Memory);
2632 if (mem->is_MergeMem()) {
2633 return mem->as_MergeMem()->memory_at(alias_idx);
2634 } else {
2635 // incoming raw memory is not split
2636 return mem;
2637 }
2638 }
2640 bool InitializeNode::is_non_zero() {
2641 if (is_complete()) return false;
2642 remove_extra_zeroes();
2643 return (req() > RawStores);
2644 }
2646 void InitializeNode::set_complete(PhaseGVN* phase) {
2647 assert(!is_complete(), "caller responsibility");
2648 _is_complete = true;
2650 // After this node is complete, it contains a bunch of
2651 // raw-memory initializations. There is no need for
2652 // it to have anything to do with non-raw memory effects.
2653 // Therefore, tell all non-raw users to re-optimize themselves,
2654 // after skipping the memory effects of this initialization.
2655 PhaseIterGVN* igvn = phase->is_IterGVN();
2656 if (igvn) igvn->add_users_to_worklist(this);
2657 }
2659 // convenience function
2660 // return false if the init contains any stores already
2661 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2662 InitializeNode* init = initialization();
2663 if (init == NULL || init->is_complete()) return false;
2664 init->remove_extra_zeroes();
2665 // for now, if this allocation has already collected any inits, bail:
2666 if (init->is_non_zero()) return false;
2667 init->set_complete(phase);
2668 return true;
2669 }
2671 void InitializeNode::remove_extra_zeroes() {
2672 if (req() == RawStores) return;
2673 Node* zmem = zero_memory();
2674 uint fill = RawStores;
2675 for (uint i = fill; i < req(); i++) {
2676 Node* n = in(i);
2677 if (n->is_top() || n == zmem) continue; // skip
2678 if (fill < i) set_req(fill, n); // compact
2679 ++fill;
2680 }
2681 // delete any empty spaces created:
2682 while (fill < req()) {
2683 del_req(fill);
2684 }
2685 }
2687 // Helper for remembering which stores go with which offsets.
2688 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2689 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2690 intptr_t offset = -1;
2691 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2692 phase, offset);
2693 if (base == NULL) return -1; // something is dead,
2694 if (offset < 0) return -1; // dead, dead
2695 return offset;
2696 }
2698 // Helper for proving that an initialization expression is
2699 // "simple enough" to be folded into an object initialization.
2700 // Attempts to prove that a store's initial value 'n' can be captured
2701 // within the initialization without creating a vicious cycle, such as:
2702 // { Foo p = new Foo(); p.next = p; }
2703 // True for constants and parameters and small combinations thereof.
2704 bool InitializeNode::detect_init_independence(Node* n,
2705 bool st_is_pinned,
2706 int& count) {
2707 if (n == NULL) return true; // (can this really happen?)
2708 if (n->is_Proj()) n = n->in(0);
2709 if (n == this) return false; // found a cycle
2710 if (n->is_Con()) return true;
2711 if (n->is_Start()) return true; // params, etc., are OK
2712 if (n->is_Root()) return true; // even better
2714 Node* ctl = n->in(0);
2715 if (ctl != NULL && !ctl->is_top()) {
2716 if (ctl->is_Proj()) ctl = ctl->in(0);
2717 if (ctl == this) return false;
2719 // If we already know that the enclosing memory op is pinned right after
2720 // the init, then any control flow that the store has picked up
2721 // must have preceded the init, or else be equal to the init.
2722 // Even after loop optimizations (which might change control edges)
2723 // a store is never pinned *before* the availability of its inputs.
2724 if (!MemNode::all_controls_dominate(n, this))
2725 return false; // failed to prove a good control
2727 }
2729 // Check data edges for possible dependencies on 'this'.
2730 if ((count += 1) > 20) return false; // complexity limit
2731 for (uint i = 1; i < n->req(); i++) {
2732 Node* m = n->in(i);
2733 if (m == NULL || m == n || m->is_top()) continue;
2734 uint first_i = n->find_edge(m);
2735 if (i != first_i) continue; // process duplicate edge just once
2736 if (!detect_init_independence(m, st_is_pinned, count)) {
2737 return false;
2738 }
2739 }
2741 return true;
2742 }
2744 // Here are all the checks a Store must pass before it can be moved into
2745 // an initialization. Returns zero if a check fails.
2746 // On success, returns the (constant) offset to which the store applies,
2747 // within the initialized memory.
2748 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2749 const int FAIL = 0;
2750 if (st->req() != MemNode::ValueIn + 1)
2751 return FAIL; // an inscrutable StoreNode (card mark?)
2752 Node* ctl = st->in(MemNode::Control);
2753 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2754 return FAIL; // must be unconditional after the initialization
2755 Node* mem = st->in(MemNode::Memory);
2756 if (!(mem->is_Proj() && mem->in(0) == this))
2757 return FAIL; // must not be preceded by other stores
2758 Node* adr = st->in(MemNode::Address);
2759 intptr_t offset;
2760 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2761 if (alloc == NULL)
2762 return FAIL; // inscrutable address
2763 if (alloc != allocation())
2764 return FAIL; // wrong allocation! (store needs to float up)
2765 Node* val = st->in(MemNode::ValueIn);
2766 int complexity_count = 0;
2767 if (!detect_init_independence(val, true, complexity_count))
2768 return FAIL; // stored value must be 'simple enough'
2770 return offset; // success
2771 }
2773 // Find the captured store in(i) which corresponds to the range
2774 // [start..start+size) in the initialized object.
2775 // If there is one, return its index i. If there isn't, return the
2776 // negative of the index where it should be inserted.
2777 // Return 0 if the queried range overlaps an initialization boundary
2778 // or if dead code is encountered.
2779 // If size_in_bytes is zero, do not bother with overlap checks.
2780 int InitializeNode::captured_store_insertion_point(intptr_t start,
2781 int size_in_bytes,
2782 PhaseTransform* phase) {
2783 const int FAIL = 0, MAX_STORE = BytesPerLong;
2785 if (is_complete())
2786 return FAIL; // arraycopy got here first; punt
2788 assert(allocation() != NULL, "must be present");
2790 // no negatives, no header fields:
2791 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2793 // after a certain size, we bail out on tracking all the stores:
2794 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2795 if (start >= ti_limit) return FAIL;
2797 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2798 if (i >= limit) return -(int)i; // not found; here is where to put it
2800 Node* st = in(i);
2801 intptr_t st_off = get_store_offset(st, phase);
2802 if (st_off < 0) {
2803 if (st != zero_memory()) {
2804 return FAIL; // bail out if there is dead garbage
2805 }
2806 } else if (st_off > start) {
2807 // ...we are done, since stores are ordered
2808 if (st_off < start + size_in_bytes) {
2809 return FAIL; // the next store overlaps
2810 }
2811 return -(int)i; // not found; here is where to put it
2812 } else if (st_off < start) {
2813 if (size_in_bytes != 0 &&
2814 start < st_off + MAX_STORE &&
2815 start < st_off + st->as_Store()->memory_size()) {
2816 return FAIL; // the previous store overlaps
2817 }
2818 } else {
2819 if (size_in_bytes != 0 &&
2820 st->as_Store()->memory_size() != size_in_bytes) {
2821 return FAIL; // mismatched store size
2822 }
2823 return i;
2824 }
2826 ++i;
2827 }
2828 }
2830 // Look for a captured store which initializes at the offset 'start'
2831 // with the given size. If there is no such store, and no other
2832 // initialization interferes, then return zero_memory (the memory
2833 // projection of the AllocateNode).
2834 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2835 PhaseTransform* phase) {
2836 assert(stores_are_sane(phase), "");
2837 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2838 if (i == 0) {
2839 return NULL; // something is dead
2840 } else if (i < 0) {
2841 return zero_memory(); // just primordial zero bits here
2842 } else {
2843 Node* st = in(i); // here is the store at this position
2844 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2845 return st;
2846 }
2847 }
2849 // Create, as a raw pointer, an address within my new object at 'offset'.
2850 Node* InitializeNode::make_raw_address(intptr_t offset,
2851 PhaseTransform* phase) {
2852 Node* addr = in(RawAddress);
2853 if (offset != 0) {
2854 Compile* C = phase->C;
2855 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2856 phase->MakeConX(offset)) );
2857 }
2858 return addr;
2859 }
2861 // Clone the given store, converting it into a raw store
2862 // initializing a field or element of my new object.
2863 // Caller is responsible for retiring the original store,
2864 // with subsume_node or the like.
2865 //
2866 // From the example above InitializeNode::InitializeNode,
2867 // here are the old stores to be captured:
2868 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2869 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2870 //
2871 // Here is the changed code; note the extra edges on init:
2872 // alloc = (Allocate ...)
2873 // rawoop = alloc.RawAddress
2874 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2875 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2876 // init = (Initialize alloc.Control alloc.Memory rawoop
2877 // rawstore1 rawstore2)
2878 //
2879 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2880 PhaseTransform* phase) {
2881 assert(stores_are_sane(phase), "");
2883 if (start < 0) return NULL;
2884 assert(can_capture_store(st, phase) == start, "sanity");
2886 Compile* C = phase->C;
2887 int size_in_bytes = st->memory_size();
2888 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2889 if (i == 0) return NULL; // bail out
2890 Node* prev_mem = NULL; // raw memory for the captured store
2891 if (i > 0) {
2892 prev_mem = in(i); // there is a pre-existing store under this one
2893 set_req(i, C->top()); // temporarily disconnect it
2894 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2895 } else {
2896 i = -i; // no pre-existing store
2897 prev_mem = zero_memory(); // a slice of the newly allocated object
2898 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2899 set_req(--i, C->top()); // reuse this edge; it has been folded away
2900 else
2901 ins_req(i, C->top()); // build a new edge
2902 }
2903 Node* new_st = st->clone();
2904 new_st->set_req(MemNode::Control, in(Control));
2905 new_st->set_req(MemNode::Memory, prev_mem);
2906 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2907 new_st = phase->transform(new_st);
2909 // At this point, new_st might have swallowed a pre-existing store
2910 // at the same offset, or perhaps new_st might have disappeared,
2911 // if it redundantly stored the same value (or zero to fresh memory).
2913 // In any case, wire it in:
2914 set_req(i, new_st);
2916 // The caller may now kill the old guy.
2917 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2918 assert(check_st == new_st || check_st == NULL, "must be findable");
2919 assert(!is_complete(), "");
2920 return new_st;
2921 }
2923 static bool store_constant(jlong* tiles, int num_tiles,
2924 intptr_t st_off, int st_size,
2925 jlong con) {
2926 if ((st_off & (st_size-1)) != 0)
2927 return false; // strange store offset (assume size==2**N)
2928 address addr = (address)tiles + st_off;
2929 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2930 switch (st_size) {
2931 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2932 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2933 case sizeof(jint): *(jint*) addr = (jint) con; break;
2934 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2935 default: return false; // strange store size (detect size!=2**N here)
2936 }
2937 return true; // return success to caller
2938 }
2940 // Coalesce subword constants into int constants and possibly
2941 // into long constants. The goal, if the CPU permits,
2942 // is to initialize the object with a small number of 64-bit tiles.
2943 // Also, convert floating-point constants to bit patterns.
2944 // Non-constants are not relevant to this pass.
2945 //
2946 // In terms of the running example on InitializeNode::InitializeNode
2947 // and InitializeNode::capture_store, here is the transformation
2948 // of rawstore1 and rawstore2 into rawstore12:
2949 // alloc = (Allocate ...)
2950 // rawoop = alloc.RawAddress
2951 // tile12 = 0x00010002
2952 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2953 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2954 //
2955 void
2956 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2957 Node* size_in_bytes,
2958 PhaseGVN* phase) {
2959 Compile* C = phase->C;
2961 assert(stores_are_sane(phase), "");
2962 // Note: After this pass, they are not completely sane,
2963 // since there may be some overlaps.
2965 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2967 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2968 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2969 size_limit = MIN2(size_limit, ti_limit);
2970 size_limit = align_size_up(size_limit, BytesPerLong);
2971 int num_tiles = size_limit / BytesPerLong;
2973 // allocate space for the tile map:
2974 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2975 jlong tiles_buf[small_len];
2976 Node* nodes_buf[small_len];
2977 jlong inits_buf[small_len];
2978 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2979 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2980 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2981 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2982 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2983 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2984 // tiles: exact bitwise model of all primitive constants
2985 // nodes: last constant-storing node subsumed into the tiles model
2986 // inits: which bytes (in each tile) are touched by any initializations
2988 //// Pass A: Fill in the tile model with any relevant stores.
2990 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2991 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2992 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2993 Node* zmem = zero_memory(); // initially zero memory state
2994 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2995 Node* st = in(i);
2996 intptr_t st_off = get_store_offset(st, phase);
2998 // Figure out the store's offset and constant value:
2999 if (st_off < header_size) continue; //skip (ignore header)
3000 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3001 int st_size = st->as_Store()->memory_size();
3002 if (st_off + st_size > size_limit) break;
3004 // Record which bytes are touched, whether by constant or not.
3005 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3006 continue; // skip (strange store size)
3008 const Type* val = phase->type(st->in(MemNode::ValueIn));
3009 if (!val->singleton()) continue; //skip (non-con store)
3010 BasicType type = val->basic_type();
3012 jlong con = 0;
3013 switch (type) {
3014 case T_INT: con = val->is_int()->get_con(); break;
3015 case T_LONG: con = val->is_long()->get_con(); break;
3016 case T_FLOAT: con = jint_cast(val->getf()); break;
3017 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3018 default: continue; //skip (odd store type)
3019 }
3021 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3022 st->Opcode() == Op_StoreL) {
3023 continue; // This StoreL is already optimal.
3024 }
3026 // Store down the constant.
3027 store_constant(tiles, num_tiles, st_off, st_size, con);
3029 intptr_t j = st_off >> LogBytesPerLong;
3031 if (type == T_INT && st_size == BytesPerInt
3032 && (st_off & BytesPerInt) == BytesPerInt) {
3033 jlong lcon = tiles[j];
3034 if (!Matcher::isSimpleConstant64(lcon) &&
3035 st->Opcode() == Op_StoreI) {
3036 // This StoreI is already optimal by itself.
3037 jint* intcon = (jint*) &tiles[j];
3038 intcon[1] = 0; // undo the store_constant()
3040 // If the previous store is also optimal by itself, back up and
3041 // undo the action of the previous loop iteration... if we can.
3042 // But if we can't, just let the previous half take care of itself.
3043 st = nodes[j];
3044 st_off -= BytesPerInt;
3045 con = intcon[0];
3046 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3047 assert(st_off >= header_size, "still ignoring header");
3048 assert(get_store_offset(st, phase) == st_off, "must be");
3049 assert(in(i-1) == zmem, "must be");
3050 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3051 assert(con == tcon->is_int()->get_con(), "must be");
3052 // Undo the effects of the previous loop trip, which swallowed st:
3053 intcon[0] = 0; // undo store_constant()
3054 set_req(i-1, st); // undo set_req(i, zmem)
3055 nodes[j] = NULL; // undo nodes[j] = st
3056 --old_subword; // undo ++old_subword
3057 }
3058 continue; // This StoreI is already optimal.
3059 }
3060 }
3062 // This store is not needed.
3063 set_req(i, zmem);
3064 nodes[j] = st; // record for the moment
3065 if (st_size < BytesPerLong) // something has changed
3066 ++old_subword; // includes int/float, but who's counting...
3067 else ++old_long;
3068 }
3070 if ((old_subword + old_long) == 0)
3071 return; // nothing more to do
3073 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3074 // Be sure to insert them before overlapping non-constant stores.
3075 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3076 for (int j = 0; j < num_tiles; j++) {
3077 jlong con = tiles[j];
3078 jlong init = inits[j];
3079 if (con == 0) continue;
3080 jint con0, con1; // split the constant, address-wise
3081 jint init0, init1; // split the init map, address-wise
3082 { union { jlong con; jint intcon[2]; } u;
3083 u.con = con;
3084 con0 = u.intcon[0];
3085 con1 = u.intcon[1];
3086 u.con = init;
3087 init0 = u.intcon[0];
3088 init1 = u.intcon[1];
3089 }
3091 Node* old = nodes[j];
3092 assert(old != NULL, "need the prior store");
3093 intptr_t offset = (j * BytesPerLong);
3095 bool split = !Matcher::isSimpleConstant64(con);
3097 if (offset < header_size) {
3098 assert(offset + BytesPerInt >= header_size, "second int counts");
3099 assert(*(jint*)&tiles[j] == 0, "junk in header");
3100 split = true; // only the second word counts
3101 // Example: int a[] = { 42 ... }
3102 } else if (con0 == 0 && init0 == -1) {
3103 split = true; // first word is covered by full inits
3104 // Example: int a[] = { ... foo(), 42 ... }
3105 } else if (con1 == 0 && init1 == -1) {
3106 split = true; // second word is covered by full inits
3107 // Example: int a[] = { ... 42, foo() ... }
3108 }
3110 // Here's a case where init0 is neither 0 nor -1:
3111 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3112 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3113 // In this case the tile is not split; it is (jlong)42.
3114 // The big tile is stored down, and then the foo() value is inserted.
3115 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3117 Node* ctl = old->in(MemNode::Control);
3118 Node* adr = make_raw_address(offset, phase);
3119 const TypePtr* atp = TypeRawPtr::BOTTOM;
3121 // One or two coalesced stores to plop down.
3122 Node* st[2];
3123 intptr_t off[2];
3124 int nst = 0;
3125 if (!split) {
3126 ++new_long;
3127 off[nst] = offset;
3128 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3129 phase->longcon(con), T_LONG);
3130 } else {
3131 // Omit either if it is a zero.
3132 if (con0 != 0) {
3133 ++new_int;
3134 off[nst] = offset;
3135 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3136 phase->intcon(con0), T_INT);
3137 }
3138 if (con1 != 0) {
3139 ++new_int;
3140 offset += BytesPerInt;
3141 adr = make_raw_address(offset, phase);
3142 off[nst] = offset;
3143 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3144 phase->intcon(con1), T_INT);
3145 }
3146 }
3148 // Insert second store first, then the first before the second.
3149 // Insert each one just before any overlapping non-constant stores.
3150 while (nst > 0) {
3151 Node* st1 = st[--nst];
3152 C->copy_node_notes_to(st1, old);
3153 st1 = phase->transform(st1);
3154 offset = off[nst];
3155 assert(offset >= header_size, "do not smash header");
3156 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3157 guarantee(ins_idx != 0, "must re-insert constant store");
3158 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3159 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3160 set_req(--ins_idx, st1);
3161 else
3162 ins_req(ins_idx, st1);
3163 }
3164 }
3166 if (PrintCompilation && WizardMode)
3167 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3168 old_subword, old_long, new_int, new_long);
3169 if (C->log() != NULL)
3170 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3171 old_subword, old_long, new_int, new_long);
3173 // Clean up any remaining occurrences of zmem:
3174 remove_extra_zeroes();
3175 }
3177 // Explore forward from in(start) to find the first fully initialized
3178 // word, and return its offset. Skip groups of subword stores which
3179 // together initialize full words. If in(start) is itself part of a
3180 // fully initialized word, return the offset of in(start). If there
3181 // are no following full-word stores, or if something is fishy, return
3182 // a negative value.
3183 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3184 int int_map = 0;
3185 intptr_t int_map_off = 0;
3186 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3188 for (uint i = start, limit = req(); i < limit; i++) {
3189 Node* st = in(i);
3191 intptr_t st_off = get_store_offset(st, phase);
3192 if (st_off < 0) break; // return conservative answer
3194 int st_size = st->as_Store()->memory_size();
3195 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3196 return st_off; // we found a complete word init
3197 }
3199 // update the map:
3201 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3202 if (this_int_off != int_map_off) {
3203 // reset the map:
3204 int_map = 0;
3205 int_map_off = this_int_off;
3206 }
3208 int subword_off = st_off - this_int_off;
3209 int_map |= right_n_bits(st_size) << subword_off;
3210 if ((int_map & FULL_MAP) == FULL_MAP) {
3211 return this_int_off; // we found a complete word init
3212 }
3214 // Did this store hit or cross the word boundary?
3215 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3216 if (next_int_off == this_int_off + BytesPerInt) {
3217 // We passed the current int, without fully initializing it.
3218 int_map_off = next_int_off;
3219 int_map >>= BytesPerInt;
3220 } else if (next_int_off > this_int_off + BytesPerInt) {
3221 // We passed the current and next int.
3222 return this_int_off + BytesPerInt;
3223 }
3224 }
3226 return -1;
3227 }
3230 // Called when the associated AllocateNode is expanded into CFG.
3231 // At this point, we may perform additional optimizations.
3232 // Linearize the stores by ascending offset, to make memory
3233 // activity as coherent as possible.
3234 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3235 intptr_t header_size,
3236 Node* size_in_bytes,
3237 PhaseGVN* phase) {
3238 assert(!is_complete(), "not already complete");
3239 assert(stores_are_sane(phase), "");
3240 assert(allocation() != NULL, "must be present");
3242 remove_extra_zeroes();
3244 if (ReduceFieldZeroing || ReduceBulkZeroing)
3245 // reduce instruction count for common initialization patterns
3246 coalesce_subword_stores(header_size, size_in_bytes, phase);
3248 Node* zmem = zero_memory(); // initially zero memory state
3249 Node* inits = zmem; // accumulating a linearized chain of inits
3250 #ifdef ASSERT
3251 intptr_t first_offset = allocation()->minimum_header_size();
3252 intptr_t last_init_off = first_offset; // previous init offset
3253 intptr_t last_init_end = first_offset; // previous init offset+size
3254 intptr_t last_tile_end = first_offset; // previous tile offset+size
3255 #endif
3256 intptr_t zeroes_done = header_size;
3258 bool do_zeroing = true; // we might give up if inits are very sparse
3259 int big_init_gaps = 0; // how many large gaps have we seen?
3261 if (ZeroTLAB) do_zeroing = false;
3262 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3264 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3265 Node* st = in(i);
3266 intptr_t st_off = get_store_offset(st, phase);
3267 if (st_off < 0)
3268 break; // unknown junk in the inits
3269 if (st->in(MemNode::Memory) != zmem)
3270 break; // complicated store chains somehow in list
3272 int st_size = st->as_Store()->memory_size();
3273 intptr_t next_init_off = st_off + st_size;
3275 if (do_zeroing && zeroes_done < next_init_off) {
3276 // See if this store needs a zero before it or under it.
3277 intptr_t zeroes_needed = st_off;
3279 if (st_size < BytesPerInt) {
3280 // Look for subword stores which only partially initialize words.
3281 // If we find some, we must lay down some word-level zeroes first,
3282 // underneath the subword stores.
3283 //
3284 // Examples:
3285 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3286 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3287 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3288 //
3289 // Note: coalesce_subword_stores may have already done this,
3290 // if it was prompted by constant non-zero subword initializers.
3291 // But this case can still arise with non-constant stores.
3293 intptr_t next_full_store = find_next_fullword_store(i, phase);
3295 // In the examples above:
3296 // in(i) p q r s x y z
3297 // st_off 12 13 14 15 12 13 14
3298 // st_size 1 1 1 1 1 1 1
3299 // next_full_s. 12 16 16 16 16 16 16
3300 // z's_done 12 16 16 16 12 16 12
3301 // z's_needed 12 16 16 16 16 16 16
3302 // zsize 0 0 0 0 4 0 4
3303 if (next_full_store < 0) {
3304 // Conservative tack: Zero to end of current word.
3305 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3306 } else {
3307 // Zero to beginning of next fully initialized word.
3308 // Or, don't zero at all, if we are already in that word.
3309 assert(next_full_store >= zeroes_needed, "must go forward");
3310 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3311 zeroes_needed = next_full_store;
3312 }
3313 }
3315 if (zeroes_needed > zeroes_done) {
3316 intptr_t zsize = zeroes_needed - zeroes_done;
3317 // Do some incremental zeroing on rawmem, in parallel with inits.
3318 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3319 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3320 zeroes_done, zeroes_needed,
3321 phase);
3322 zeroes_done = zeroes_needed;
3323 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3324 do_zeroing = false; // leave the hole, next time
3325 }
3326 }
3328 // Collect the store and move on:
3329 st->set_req(MemNode::Memory, inits);
3330 inits = st; // put it on the linearized chain
3331 set_req(i, zmem); // unhook from previous position
3333 if (zeroes_done == st_off)
3334 zeroes_done = next_init_off;
3336 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3338 #ifdef ASSERT
3339 // Various order invariants. Weaker than stores_are_sane because
3340 // a large constant tile can be filled in by smaller non-constant stores.
3341 assert(st_off >= last_init_off, "inits do not reverse");
3342 last_init_off = st_off;
3343 const Type* val = NULL;
3344 if (st_size >= BytesPerInt &&
3345 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3346 (int)val->basic_type() < (int)T_OBJECT) {
3347 assert(st_off >= last_tile_end, "tiles do not overlap");
3348 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3349 last_tile_end = MAX2(last_tile_end, next_init_off);
3350 } else {
3351 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3352 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3353 assert(st_off >= last_init_end, "inits do not overlap");
3354 last_init_end = next_init_off; // it's a non-tile
3355 }
3356 #endif //ASSERT
3357 }
3359 remove_extra_zeroes(); // clear out all the zmems left over
3360 add_req(inits);
3362 if (!ZeroTLAB) {
3363 // If anything remains to be zeroed, zero it all now.
3364 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3365 // if it is the last unused 4 bytes of an instance, forget about it
3366 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3367 if (zeroes_done + BytesPerLong >= size_limit) {
3368 assert(allocation() != NULL, "");
3369 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3370 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3371 if (zeroes_done == k->layout_helper())
3372 zeroes_done = size_limit;
3373 }
3374 if (zeroes_done < size_limit) {
3375 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3376 zeroes_done, size_in_bytes, phase);
3377 }
3378 }
3380 set_complete(phase);
3381 return rawmem;
3382 }
3385 #ifdef ASSERT
3386 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3387 if (is_complete())
3388 return true; // stores could be anything at this point
3389 assert(allocation() != NULL, "must be present");
3390 intptr_t last_off = allocation()->minimum_header_size();
3391 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3392 Node* st = in(i);
3393 intptr_t st_off = get_store_offset(st, phase);
3394 if (st_off < 0) continue; // ignore dead garbage
3395 if (last_off > st_off) {
3396 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3397 this->dump(2);
3398 assert(false, "ascending store offsets");
3399 return false;
3400 }
3401 last_off = st_off + st->as_Store()->memory_size();
3402 }
3403 return true;
3404 }
3405 #endif //ASSERT
3410 //============================MergeMemNode=====================================
3411 //
3412 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3413 // contributing store or call operations. Each contributor provides the memory
3414 // state for a particular "alias type" (see Compile::alias_type). For example,
3415 // if a MergeMem has an input X for alias category #6, then any memory reference
3416 // to alias category #6 may use X as its memory state input, as an exact equivalent
3417 // to using the MergeMem as a whole.
3418 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3419 //
3420 // (Here, the <N> notation gives the index of the relevant adr_type.)
3421 //
3422 // In one special case (and more cases in the future), alias categories overlap.
3423 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3424 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3425 // it is exactly equivalent to that state W:
3426 // MergeMem(<Bot>: W) <==> W
3427 //
3428 // Usually, the merge has more than one input. In that case, where inputs
3429 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3430 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3431 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3432 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3433 //
3434 // A merge can take a "wide" memory state as one of its narrow inputs.
3435 // This simply means that the merge observes out only the relevant parts of
3436 // the wide input. That is, wide memory states arriving at narrow merge inputs
3437 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3438 //
3439 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3440 // and that memory slices "leak through":
3441 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3442 //
3443 // But, in such a cascade, repeated memory slices can "block the leak":
3444 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3445 //
3446 // In the last example, Y is not part of the combined memory state of the
3447 // outermost MergeMem. The system must, of course, prevent unschedulable
3448 // memory states from arising, so you can be sure that the state Y is somehow
3449 // a precursor to state Y'.
3450 //
3451 //
3452 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3453 // of each MergeMemNode array are exactly the numerical alias indexes, including
3454 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3455 // Compile::alias_type (and kin) produce and manage these indexes.
3456 //
3457 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3458 // (Note that this provides quick access to the top node inside MergeMem methods,
3459 // without the need to reach out via TLS to Compile::current.)
3460 //
3461 // As a consequence of what was just described, a MergeMem that represents a full
3462 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3463 // containing all alias categories.
3464 //
3465 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3466 //
3467 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3468 // a memory state for the alias type <N>, or else the top node, meaning that
3469 // there is no particular input for that alias type. Note that the length of
3470 // a MergeMem is variable, and may be extended at any time to accommodate new
3471 // memory states at larger alias indexes. When merges grow, they are of course
3472 // filled with "top" in the unused in() positions.
3473 //
3474 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3475 // (Top was chosen because it works smoothly with passes like GCM.)
3476 //
3477 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3478 // the type of random VM bits like TLS references.) Since it is always the
3479 // first non-Bot memory slice, some low-level loops use it to initialize an
3480 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3481 //
3482 //
3483 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3484 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3485 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3486 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3487 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3488 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3489 //
3490 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3491 // really that different from the other memory inputs. An abbreviation called
3492 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3493 //
3494 //
3495 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3496 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3497 // that "emerges though" the base memory will be marked as excluding the alias types
3498 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3499 //
3500 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3501 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3502 //
3503 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3504 // (It is currently unimplemented.) As you can see, the resulting merge is
3505 // actually a disjoint union of memory states, rather than an overlay.
3506 //
3508 //------------------------------MergeMemNode-----------------------------------
3509 Node* MergeMemNode::make_empty_memory() {
3510 Node* empty_memory = (Node*) Compile::current()->top();
3511 assert(empty_memory->is_top(), "correct sentinel identity");
3512 return empty_memory;
3513 }
3515 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3516 init_class_id(Class_MergeMem);
3517 // all inputs are nullified in Node::Node(int)
3518 // set_input(0, NULL); // no control input
3520 // Initialize the edges uniformly to top, for starters.
3521 Node* empty_mem = make_empty_memory();
3522 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3523 init_req(i,empty_mem);
3524 }
3525 assert(empty_memory() == empty_mem, "");
3527 if( new_base != NULL && new_base->is_MergeMem() ) {
3528 MergeMemNode* mdef = new_base->as_MergeMem();
3529 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3530 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3531 mms.set_memory(mms.memory2());
3532 }
3533 assert(base_memory() == mdef->base_memory(), "");
3534 } else {
3535 set_base_memory(new_base);
3536 }
3537 }
3539 // Make a new, untransformed MergeMem with the same base as 'mem'.
3540 // If mem is itself a MergeMem, populate the result with the same edges.
3541 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3542 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3543 }
3545 //------------------------------cmp--------------------------------------------
3546 uint MergeMemNode::hash() const { return NO_HASH; }
3547 uint MergeMemNode::cmp( const Node &n ) const {
3548 return (&n == this); // Always fail except on self
3549 }
3551 //------------------------------Identity---------------------------------------
3552 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3553 // Identity if this merge point does not record any interesting memory
3554 // disambiguations.
3555 Node* base_mem = base_memory();
3556 Node* empty_mem = empty_memory();
3557 if (base_mem != empty_mem) { // Memory path is not dead?
3558 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3559 Node* mem = in(i);
3560 if (mem != empty_mem && mem != base_mem) {
3561 return this; // Many memory splits; no change
3562 }
3563 }
3564 }
3565 return base_mem; // No memory splits; ID on the one true input
3566 }
3568 //------------------------------Ideal------------------------------------------
3569 // This method is invoked recursively on chains of MergeMem nodes
3570 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3571 // Remove chain'd MergeMems
3572 //
3573 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3574 // relative to the "in(Bot)". Since we are patching both at the same time,
3575 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3576 // but rewrite each "in(i)" relative to the new "in(Bot)".
3577 Node *progress = NULL;
3580 Node* old_base = base_memory();
3581 Node* empty_mem = empty_memory();
3582 if (old_base == empty_mem)
3583 return NULL; // Dead memory path.
3585 MergeMemNode* old_mbase;
3586 if (old_base != NULL && old_base->is_MergeMem())
3587 old_mbase = old_base->as_MergeMem();
3588 else
3589 old_mbase = NULL;
3590 Node* new_base = old_base;
3592 // simplify stacked MergeMems in base memory
3593 if (old_mbase) new_base = old_mbase->base_memory();
3595 // the base memory might contribute new slices beyond my req()
3596 if (old_mbase) grow_to_match(old_mbase);
3598 // Look carefully at the base node if it is a phi.
3599 PhiNode* phi_base;
3600 if (new_base != NULL && new_base->is_Phi())
3601 phi_base = new_base->as_Phi();
3602 else
3603 phi_base = NULL;
3605 Node* phi_reg = NULL;
3606 uint phi_len = (uint)-1;
3607 if (phi_base != NULL && !phi_base->is_copy()) {
3608 // do not examine phi if degraded to a copy
3609 phi_reg = phi_base->region();
3610 phi_len = phi_base->req();
3611 // see if the phi is unfinished
3612 for (uint i = 1; i < phi_len; i++) {
3613 if (phi_base->in(i) == NULL) {
3614 // incomplete phi; do not look at it yet!
3615 phi_reg = NULL;
3616 phi_len = (uint)-1;
3617 break;
3618 }
3619 }
3620 }
3622 // Note: We do not call verify_sparse on entry, because inputs
3623 // can normalize to the base_memory via subsume_node or similar
3624 // mechanisms. This method repairs that damage.
3626 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3628 // Look at each slice.
3629 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3630 Node* old_in = in(i);
3631 // calculate the old memory value
3632 Node* old_mem = old_in;
3633 if (old_mem == empty_mem) old_mem = old_base;
3634 assert(old_mem == memory_at(i), "");
3636 // maybe update (reslice) the old memory value
3638 // simplify stacked MergeMems
3639 Node* new_mem = old_mem;
3640 MergeMemNode* old_mmem;
3641 if (old_mem != NULL && old_mem->is_MergeMem())
3642 old_mmem = old_mem->as_MergeMem();
3643 else
3644 old_mmem = NULL;
3645 if (old_mmem == this) {
3646 // This can happen if loops break up and safepoints disappear.
3647 // A merge of BotPtr (default) with a RawPtr memory derived from a
3648 // safepoint can be rewritten to a merge of the same BotPtr with
3649 // the BotPtr phi coming into the loop. If that phi disappears
3650 // also, we can end up with a self-loop of the mergemem.
3651 // In general, if loops degenerate and memory effects disappear,
3652 // a mergemem can be left looking at itself. This simply means
3653 // that the mergemem's default should be used, since there is
3654 // no longer any apparent effect on this slice.
3655 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3656 // from start. Update the input to TOP.
3657 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3658 }
3659 else if (old_mmem != NULL) {
3660 new_mem = old_mmem->memory_at(i);
3661 }
3662 // else preceeding memory was not a MergeMem
3664 // replace equivalent phis (unfortunately, they do not GVN together)
3665 if (new_mem != NULL && new_mem != new_base &&
3666 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3667 if (new_mem->is_Phi()) {
3668 PhiNode* phi_mem = new_mem->as_Phi();
3669 for (uint i = 1; i < phi_len; i++) {
3670 if (phi_base->in(i) != phi_mem->in(i)) {
3671 phi_mem = NULL;
3672 break;
3673 }
3674 }
3675 if (phi_mem != NULL) {
3676 // equivalent phi nodes; revert to the def
3677 new_mem = new_base;
3678 }
3679 }
3680 }
3682 // maybe store down a new value
3683 Node* new_in = new_mem;
3684 if (new_in == new_base) new_in = empty_mem;
3686 if (new_in != old_in) {
3687 // Warning: Do not combine this "if" with the previous "if"
3688 // A memory slice might have be be rewritten even if it is semantically
3689 // unchanged, if the base_memory value has changed.
3690 set_req(i, new_in);
3691 progress = this; // Report progress
3692 }
3693 }
3695 if (new_base != old_base) {
3696 set_req(Compile::AliasIdxBot, new_base);
3697 // Don't use set_base_memory(new_base), because we need to update du.
3698 assert(base_memory() == new_base, "");
3699 progress = this;
3700 }
3702 if( base_memory() == this ) {
3703 // a self cycle indicates this memory path is dead
3704 set_req(Compile::AliasIdxBot, empty_mem);
3705 }
3707 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3708 // Recursion must occur after the self cycle check above
3709 if( base_memory()->is_MergeMem() ) {
3710 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3711 Node *m = phase->transform(new_mbase); // Rollup any cycles
3712 if( m != NULL && (m->is_top() ||
3713 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3714 // propagate rollup of dead cycle to self
3715 set_req(Compile::AliasIdxBot, empty_mem);
3716 }
3717 }
3719 if( base_memory() == empty_mem ) {
3720 progress = this;
3721 // Cut inputs during Parse phase only.
3722 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3723 if( !can_reshape ) {
3724 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3725 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3726 }
3727 }
3728 }
3730 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3731 // Check if PhiNode::Ideal's "Split phis through memory merges"
3732 // transform should be attempted. Look for this->phi->this cycle.
3733 uint merge_width = req();
3734 if (merge_width > Compile::AliasIdxRaw) {
3735 PhiNode* phi = base_memory()->as_Phi();
3736 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3737 if (phi->in(i) == this) {
3738 phase->is_IterGVN()->_worklist.push(phi);
3739 break;
3740 }
3741 }
3742 }
3743 }
3745 assert(progress || verify_sparse(), "please, no dups of base");
3746 return progress;
3747 }
3749 //-------------------------set_base_memory-------------------------------------
3750 void MergeMemNode::set_base_memory(Node *new_base) {
3751 Node* empty_mem = empty_memory();
3752 set_req(Compile::AliasIdxBot, new_base);
3753 assert(memory_at(req()) == new_base, "must set default memory");
3754 // Clear out other occurrences of new_base:
3755 if (new_base != empty_mem) {
3756 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3757 if (in(i) == new_base) set_req(i, empty_mem);
3758 }
3759 }
3760 }
3762 //------------------------------out_RegMask------------------------------------
3763 const RegMask &MergeMemNode::out_RegMask() const {
3764 return RegMask::Empty;
3765 }
3767 //------------------------------dump_spec--------------------------------------
3768 #ifndef PRODUCT
3769 void MergeMemNode::dump_spec(outputStream *st) const {
3770 st->print(" {");
3771 Node* base_mem = base_memory();
3772 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3773 Node* mem = memory_at(i);
3774 if (mem == base_mem) { st->print(" -"); continue; }
3775 st->print( " N%d:", mem->_idx );
3776 Compile::current()->get_adr_type(i)->dump_on(st);
3777 }
3778 st->print(" }");
3779 }
3780 #endif // !PRODUCT
3783 #ifdef ASSERT
3784 static bool might_be_same(Node* a, Node* b) {
3785 if (a == b) return true;
3786 if (!(a->is_Phi() || b->is_Phi())) return false;
3787 // phis shift around during optimization
3788 return true; // pretty stupid...
3789 }
3791 // verify a narrow slice (either incoming or outgoing)
3792 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3793 if (!VerifyAliases) return; // don't bother to verify unless requested
3794 if (is_error_reported()) return; // muzzle asserts when debugging an error
3795 if (Node::in_dump()) return; // muzzle asserts when printing
3796 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3797 assert(n != NULL, "");
3798 // Elide intervening MergeMem's
3799 while (n->is_MergeMem()) {
3800 n = n->as_MergeMem()->memory_at(alias_idx);
3801 }
3802 Compile* C = Compile::current();
3803 const TypePtr* n_adr_type = n->adr_type();
3804 if (n == m->empty_memory()) {
3805 // Implicit copy of base_memory()
3806 } else if (n_adr_type != TypePtr::BOTTOM) {
3807 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3808 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3809 } else {
3810 // A few places like make_runtime_call "know" that VM calls are narrow,
3811 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3812 bool expected_wide_mem = false;
3813 if (n == m->base_memory()) {
3814 expected_wide_mem = true;
3815 } else if (alias_idx == Compile::AliasIdxRaw ||
3816 n == m->memory_at(Compile::AliasIdxRaw)) {
3817 expected_wide_mem = true;
3818 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3819 // memory can "leak through" calls on channels that
3820 // are write-once. Allow this also.
3821 expected_wide_mem = true;
3822 }
3823 assert(expected_wide_mem, "expected narrow slice replacement");
3824 }
3825 }
3826 #else // !ASSERT
3827 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3828 #endif
3831 //-----------------------------memory_at---------------------------------------
3832 Node* MergeMemNode::memory_at(uint alias_idx) const {
3833 assert(alias_idx >= Compile::AliasIdxRaw ||
3834 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3835 "must avoid base_memory and AliasIdxTop");
3837 // Otherwise, it is a narrow slice.
3838 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3839 Compile *C = Compile::current();
3840 if (is_empty_memory(n)) {
3841 // the array is sparse; empty slots are the "top" node
3842 n = base_memory();
3843 assert(Node::in_dump()
3844 || n == NULL || n->bottom_type() == Type::TOP
3845 || n->adr_type() == TypePtr::BOTTOM
3846 || n->adr_type() == TypeRawPtr::BOTTOM
3847 || Compile::current()->AliasLevel() == 0,
3848 "must be a wide memory");
3849 // AliasLevel == 0 if we are organizing the memory states manually.
3850 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3851 } else {
3852 // make sure the stored slice is sane
3853 #ifdef ASSERT
3854 if (is_error_reported() || Node::in_dump()) {
3855 } else if (might_be_same(n, base_memory())) {
3856 // Give it a pass: It is a mostly harmless repetition of the base.
3857 // This can arise normally from node subsumption during optimization.
3858 } else {
3859 verify_memory_slice(this, alias_idx, n);
3860 }
3861 #endif
3862 }
3863 return n;
3864 }
3866 //---------------------------set_memory_at-------------------------------------
3867 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3868 verify_memory_slice(this, alias_idx, n);
3869 Node* empty_mem = empty_memory();
3870 if (n == base_memory()) n = empty_mem; // collapse default
3871 uint need_req = alias_idx+1;
3872 if (req() < need_req) {
3873 if (n == empty_mem) return; // already the default, so do not grow me
3874 // grow the sparse array
3875 do {
3876 add_req(empty_mem);
3877 } while (req() < need_req);
3878 }
3879 set_req( alias_idx, n );
3880 }
3884 //--------------------------iteration_setup------------------------------------
3885 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3886 if (other != NULL) {
3887 grow_to_match(other);
3888 // invariant: the finite support of mm2 is within mm->req()
3889 #ifdef ASSERT
3890 for (uint i = req(); i < other->req(); i++) {
3891 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3892 }
3893 #endif
3894 }
3895 // Replace spurious copies of base_memory by top.
3896 Node* base_mem = base_memory();
3897 if (base_mem != NULL && !base_mem->is_top()) {
3898 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3899 if (in(i) == base_mem)
3900 set_req(i, empty_memory());
3901 }
3902 }
3903 }
3905 //---------------------------grow_to_match-------------------------------------
3906 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3907 Node* empty_mem = empty_memory();
3908 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3909 // look for the finite support of the other memory
3910 for (uint i = other->req(); --i >= req(); ) {
3911 if (other->in(i) != empty_mem) {
3912 uint new_len = i+1;
3913 while (req() < new_len) add_req(empty_mem);
3914 break;
3915 }
3916 }
3917 }
3919 //---------------------------verify_sparse-------------------------------------
3920 #ifndef PRODUCT
3921 bool MergeMemNode::verify_sparse() const {
3922 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3923 Node* base_mem = base_memory();
3924 // The following can happen in degenerate cases, since empty==top.
3925 if (is_empty_memory(base_mem)) return true;
3926 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3927 assert(in(i) != NULL, "sane slice");
3928 if (in(i) == base_mem) return false; // should have been the sentinel value!
3929 }
3930 return true;
3931 }
3933 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3934 Node* n;
3935 n = mm->in(idx);
3936 if (mem == n) return true; // might be empty_memory()
3937 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3938 if (mem == n) return true;
3939 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3940 if (mem == n) return true;
3941 if (n == NULL) break;
3942 }
3943 return false;
3944 }
3945 #endif // !PRODUCT