Mon, 28 Apr 2008 08:08:12 -0700
Merge
1 /*
2 * Copyright 1997-2007 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_instance_field())
95 return mchain; // don't try to optimize non-instance types
96 uint instance_id = tinst->instance_id();
97 Node *prev = NULL;
98 Node *result = mchain;
99 while (prev != result) {
100 prev = result;
101 // skip over a call which does not affect this memory slice
102 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
103 Node *proj_in = result->in(0);
104 if (proj_in->is_Call()) {
105 CallNode *call = proj_in->as_Call();
106 if (!call->may_modify(t_adr, phase)) {
107 result = call->in(TypeFunc::Memory);
108 }
109 } else if (proj_in->is_Initialize()) {
110 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
111 // Stop if this is the initialization for the object instance which
112 // which contains this memory slice, otherwise skip over it.
113 if (alloc != NULL && alloc->_idx != instance_id) {
114 result = proj_in->in(TypeFunc::Memory);
115 }
116 } else if (proj_in->is_MemBar()) {
117 result = proj_in->in(TypeFunc::Memory);
118 }
119 } else if (result->is_MergeMem()) {
120 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
121 }
122 }
123 return result;
124 }
126 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
127 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
128 bool is_instance = (t_oop != NULL) && t_oop->is_instance_field();
129 PhaseIterGVN *igvn = phase->is_IterGVN();
130 Node *result = mchain;
131 result = optimize_simple_memory_chain(result, t_adr, phase);
132 if (is_instance && igvn != NULL && result->is_Phi()) {
133 PhiNode *mphi = result->as_Phi();
134 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
135 const TypePtr *t = mphi->adr_type();
136 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM) {
137 // clone the Phi with our address type
138 result = mphi->split_out_instance(t_adr, igvn);
139 } else {
140 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
141 }
142 }
143 return result;
144 }
146 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
147 uint alias_idx = phase->C->get_alias_index(tp);
148 Node *mem = mmem;
149 #ifdef ASSERT
150 {
151 // Check that current type is consistent with the alias index used during graph construction
152 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
153 bool consistent = adr_check == NULL || adr_check->empty() ||
154 phase->C->must_alias(adr_check, alias_idx );
155 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
156 if( !consistent && adr_check != NULL && !adr_check->empty() &&
157 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
158 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
159 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
160 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
161 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
162 // don't assert if it is dead code.
163 consistent = true;
164 }
165 if( !consistent ) {
166 st->print("alias_idx==%d, adr_check==", alias_idx);
167 if( adr_check == NULL ) {
168 st->print("NULL");
169 } else {
170 adr_check->dump();
171 }
172 st->cr();
173 print_alias_types();
174 assert(consistent, "adr_check must match alias idx");
175 }
176 }
177 #endif
178 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
179 // means an array I have not precisely typed yet. Do not do any
180 // alias stuff with it any time soon.
181 const TypeOopPtr *tinst = tp->isa_oopptr();
182 if( tp->base() != Type::AnyPtr &&
183 !(tinst &&
184 tinst->klass()->is_java_lang_Object() &&
185 tinst->offset() == Type::OffsetBot) ) {
186 // compress paths and change unreachable cycles to TOP
187 // If not, we can update the input infinitely along a MergeMem cycle
188 // Equivalent code in PhiNode::Ideal
189 Node* m = phase->transform(mmem);
190 // If tranformed to a MergeMem, get the desired slice
191 // Otherwise the returned node represents memory for every slice
192 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
193 // Update input if it is progress over what we have now
194 }
195 return mem;
196 }
198 //--------------------------Ideal_common---------------------------------------
199 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
200 // Unhook non-raw memories from complete (macro-expanded) initializations.
201 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
202 // If our control input is a dead region, kill all below the region
203 Node *ctl = in(MemNode::Control);
204 if (ctl && remove_dead_region(phase, can_reshape))
205 return this;
207 // Ignore if memory is dead, or self-loop
208 Node *mem = in(MemNode::Memory);
209 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
210 assert( mem != this, "dead loop in MemNode::Ideal" );
212 Node *address = in(MemNode::Address);
213 const Type *t_adr = phase->type( address );
214 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
216 // Avoid independent memory operations
217 Node* old_mem = mem;
219 // The code which unhooks non-raw memories from complete (macro-expanded)
220 // initializations was removed. After macro-expansion all stores catched
221 // by Initialize node became raw stores and there is no information
222 // which memory slices they modify. So it is unsafe to move any memory
223 // operation above these stores. Also in most cases hooked non-raw memories
224 // were already unhooked by using information from detect_ptr_independence()
225 // and find_previous_store().
227 if (mem->is_MergeMem()) {
228 MergeMemNode* mmem = mem->as_MergeMem();
229 const TypePtr *tp = t_adr->is_ptr();
231 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
232 }
234 if (mem != old_mem) {
235 set_req(MemNode::Memory, mem);
236 return this;
237 }
239 // let the subclass continue analyzing...
240 return NULL;
241 }
243 // Helper function for proving some simple control dominations.
244 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
245 // Already assumes that 'dom' is available at 'sub', and that 'sub'
246 // is not a constant (dominated by the method's StartNode).
247 // Used by MemNode::find_previous_store to prove that the
248 // control input of a memory operation predates (dominates)
249 // an allocation it wants to look past.
250 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
251 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
252 return false; // Conservative answer for dead code
254 // Check 'dom'.
255 dom = dom->find_exact_control(dom);
256 if (dom == NULL || dom->is_top())
257 return false; // Conservative answer for dead code
259 if (dom->is_Start() || dom->is_Root() || dom == sub)
260 return true;
262 // 'dom' dominates 'sub' if its control edge and control edges
263 // of all its inputs dominate or equal to sub's control edge.
265 // Currently 'sub' is either Allocate, Initialize or Start nodes.
266 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start(), "expecting only these nodes");
268 // Get control edge of 'sub'.
269 sub = sub->find_exact_control(sub->in(0));
270 if (sub == NULL || sub->is_top())
271 return false; // Conservative answer for dead code
273 assert(sub->is_CFG(), "expecting control");
275 if (sub == dom)
276 return true;
278 if (sub->is_Start() || sub->is_Root())
279 return false;
281 {
282 // Check all control edges of 'dom'.
284 ResourceMark rm;
285 Arena* arena = Thread::current()->resource_area();
286 Node_List nlist(arena);
287 Unique_Node_List dom_list(arena);
289 dom_list.push(dom);
290 bool only_dominating_controls = false;
292 for (uint next = 0; next < dom_list.size(); next++) {
293 Node* n = dom_list.at(next);
294 if (!n->is_CFG() && n->pinned()) {
295 // Check only own control edge for pinned non-control nodes.
296 n = n->find_exact_control(n->in(0));
297 if (n == NULL || n->is_top())
298 return false; // Conservative answer for dead code
299 assert(n->is_CFG(), "expecting control");
300 }
301 if (n->is_Start() || n->is_Root()) {
302 only_dominating_controls = true;
303 } else if (n->is_CFG()) {
304 if (n->dominates(sub, nlist))
305 only_dominating_controls = true;
306 else
307 return false;
308 } else {
309 // First, own control edge.
310 Node* m = n->find_exact_control(n->in(0));
311 if (m == NULL)
312 continue;
313 if (m->is_top())
314 return false; // Conservative answer for dead code
315 dom_list.push(m);
317 // Now, the rest of edges.
318 uint cnt = n->req();
319 for (uint i = 1; i < cnt; i++) {
320 m = n->find_exact_control(n->in(i));
321 if (m == NULL || m->is_top())
322 continue;
323 dom_list.push(m);
324 }
325 }
326 }
327 return only_dominating_controls;
328 }
329 }
331 //---------------------detect_ptr_independence---------------------------------
332 // Used by MemNode::find_previous_store to prove that two base
333 // pointers are never equal.
334 // The pointers are accompanied by their associated allocations,
335 // if any, which have been previously discovered by the caller.
336 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
337 Node* p2, AllocateNode* a2,
338 PhaseTransform* phase) {
339 // Attempt to prove that these two pointers cannot be aliased.
340 // They may both manifestly be allocations, and they should differ.
341 // Or, if they are not both allocations, they can be distinct constants.
342 // Otherwise, one is an allocation and the other a pre-existing value.
343 if (a1 == NULL && a2 == NULL) { // neither an allocation
344 return (p1 != p2) && p1->is_Con() && p2->is_Con();
345 } else if (a1 != NULL && a2 != NULL) { // both allocations
346 return (a1 != a2);
347 } else if (a1 != NULL) { // one allocation a1
348 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
349 return all_controls_dominate(p2, a1);
350 } else { //(a2 != NULL) // one allocation a2
351 return all_controls_dominate(p1, a2);
352 }
353 return false;
354 }
357 // The logic for reordering loads and stores uses four steps:
358 // (a) Walk carefully past stores and initializations which we
359 // can prove are independent of this load.
360 // (b) Observe that the next memory state makes an exact match
361 // with self (load or store), and locate the relevant store.
362 // (c) Ensure that, if we were to wire self directly to the store,
363 // the optimizer would fold it up somehow.
364 // (d) Do the rewiring, and return, depending on some other part of
365 // the optimizer to fold up the load.
366 // This routine handles steps (a) and (b). Steps (c) and (d) are
367 // specific to loads and stores, so they are handled by the callers.
368 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
369 //
370 Node* MemNode::find_previous_store(PhaseTransform* phase) {
371 Node* ctrl = in(MemNode::Control);
372 Node* adr = in(MemNode::Address);
373 intptr_t offset = 0;
374 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
375 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
377 if (offset == Type::OffsetBot)
378 return NULL; // cannot unalias unless there are precise offsets
380 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
382 intptr_t size_in_bytes = memory_size();
384 Node* mem = in(MemNode::Memory); // start searching here...
386 int cnt = 50; // Cycle limiter
387 for (;;) { // While we can dance past unrelated stores...
388 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
390 if (mem->is_Store()) {
391 Node* st_adr = mem->in(MemNode::Address);
392 intptr_t st_offset = 0;
393 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
394 if (st_base == NULL)
395 break; // inscrutable pointer
396 if (st_offset != offset && st_offset != Type::OffsetBot) {
397 const int MAX_STORE = BytesPerLong;
398 if (st_offset >= offset + size_in_bytes ||
399 st_offset <= offset - MAX_STORE ||
400 st_offset <= offset - mem->as_Store()->memory_size()) {
401 // Success: The offsets are provably independent.
402 // (You may ask, why not just test st_offset != offset and be done?
403 // The answer is that stores of different sizes can co-exist
404 // in the same sequence of RawMem effects. We sometimes initialize
405 // a whole 'tile' of array elements with a single jint or jlong.)
406 mem = mem->in(MemNode::Memory);
407 continue; // (a) advance through independent store memory
408 }
409 }
410 if (st_base != base &&
411 detect_ptr_independence(base, alloc,
412 st_base,
413 AllocateNode::Ideal_allocation(st_base, phase),
414 phase)) {
415 // Success: The bases are provably independent.
416 mem = mem->in(MemNode::Memory);
417 continue; // (a) advance through independent store memory
418 }
420 // (b) At this point, if the bases or offsets do not agree, we lose,
421 // since we have not managed to prove 'this' and 'mem' independent.
422 if (st_base == base && st_offset == offset) {
423 return mem; // let caller handle steps (c), (d)
424 }
426 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
427 InitializeNode* st_init = mem->in(0)->as_Initialize();
428 AllocateNode* st_alloc = st_init->allocation();
429 if (st_alloc == NULL)
430 break; // something degenerated
431 bool known_identical = false;
432 bool known_independent = false;
433 if (alloc == st_alloc)
434 known_identical = true;
435 else if (alloc != NULL)
436 known_independent = true;
437 else if (all_controls_dominate(this, st_alloc))
438 known_independent = true;
440 if (known_independent) {
441 // The bases are provably independent: Either they are
442 // manifestly distinct allocations, or else the control
443 // of this load dominates the store's allocation.
444 int alias_idx = phase->C->get_alias_index(adr_type());
445 if (alias_idx == Compile::AliasIdxRaw) {
446 mem = st_alloc->in(TypeFunc::Memory);
447 } else {
448 mem = st_init->memory(alias_idx);
449 }
450 continue; // (a) advance through independent store memory
451 }
453 // (b) at this point, if we are not looking at a store initializing
454 // the same allocation we are loading from, we lose.
455 if (known_identical) {
456 // From caller, can_see_stored_value will consult find_captured_store.
457 return mem; // let caller handle steps (c), (d)
458 }
460 } else if (addr_t != NULL && addr_t->is_instance_field()) {
461 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
462 if (mem->is_Proj() && mem->in(0)->is_Call()) {
463 CallNode *call = mem->in(0)->as_Call();
464 if (!call->may_modify(addr_t, phase)) {
465 mem = call->in(TypeFunc::Memory);
466 continue; // (a) advance through independent call memory
467 }
468 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
469 mem = mem->in(0)->in(TypeFunc::Memory);
470 continue; // (a) advance through independent MemBar memory
471 } else if (mem->is_MergeMem()) {
472 int alias_idx = phase->C->get_alias_index(adr_type());
473 mem = mem->as_MergeMem()->memory_at(alias_idx);
474 continue; // (a) advance through independent MergeMem memory
475 }
476 }
478 // Unless there is an explicit 'continue', we must bail out here,
479 // because 'mem' is an inscrutable memory state (e.g., a call).
480 break;
481 }
483 return NULL; // bail out
484 }
486 //----------------------calculate_adr_type-------------------------------------
487 // Helper function. Notices when the given type of address hits top or bottom.
488 // Also, asserts a cross-check of the type against the expected address type.
489 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
490 if (t == Type::TOP) return NULL; // does not touch memory any more?
491 #ifdef PRODUCT
492 cross_check = NULL;
493 #else
494 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
495 #endif
496 const TypePtr* tp = t->isa_ptr();
497 if (tp == NULL) {
498 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
499 return TypePtr::BOTTOM; // touches lots of memory
500 } else {
501 #ifdef ASSERT
502 // %%%% [phh] We don't check the alias index if cross_check is
503 // TypeRawPtr::BOTTOM. Needs to be investigated.
504 if (cross_check != NULL &&
505 cross_check != TypePtr::BOTTOM &&
506 cross_check != TypeRawPtr::BOTTOM) {
507 // Recheck the alias index, to see if it has changed (due to a bug).
508 Compile* C = Compile::current();
509 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
510 "must stay in the original alias category");
511 // The type of the address must be contained in the adr_type,
512 // disregarding "null"-ness.
513 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
514 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
515 assert(cross_check->meet(tp_notnull) == cross_check,
516 "real address must not escape from expected memory type");
517 }
518 #endif
519 return tp;
520 }
521 }
523 //------------------------adr_phi_is_loop_invariant----------------------------
524 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
525 // loop is loop invariant. Make a quick traversal of Phi and associated
526 // CastPP nodes, looking to see if they are a closed group within the loop.
527 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
528 // The idea is that the phi-nest must boil down to only CastPP nodes
529 // with the same data. This implies that any path into the loop already
530 // includes such a CastPP, and so the original cast, whatever its input,
531 // must be covered by an equivalent cast, with an earlier control input.
532 ResourceMark rm;
534 // The loop entry input of the phi should be the unique dominating
535 // node for every Phi/CastPP in the loop.
536 Unique_Node_List closure;
537 closure.push(adr_phi->in(LoopNode::EntryControl));
539 // Add the phi node and the cast to the worklist.
540 Unique_Node_List worklist;
541 worklist.push(adr_phi);
542 if( cast != NULL ){
543 if( !cast->is_ConstraintCast() ) return false;
544 worklist.push(cast);
545 }
547 // Begin recursive walk of phi nodes.
548 while( worklist.size() ){
549 // Take a node off the worklist
550 Node *n = worklist.pop();
551 if( !closure.member(n) ){
552 // Add it to the closure.
553 closure.push(n);
554 // Make a sanity check to ensure we don't waste too much time here.
555 if( closure.size() > 20) return false;
556 // This node is OK if:
557 // - it is a cast of an identical value
558 // - or it is a phi node (then we add its inputs to the worklist)
559 // Otherwise, the node is not OK, and we presume the cast is not invariant
560 if( n->is_ConstraintCast() ){
561 worklist.push(n->in(1));
562 } else if( n->is_Phi() ) {
563 for( uint i = 1; i < n->req(); i++ ) {
564 worklist.push(n->in(i));
565 }
566 } else {
567 return false;
568 }
569 }
570 }
572 // Quit when the worklist is empty, and we've found no offending nodes.
573 return true;
574 }
576 //------------------------------Ideal_DU_postCCP-------------------------------
577 // Find any cast-away of null-ness and keep its control. Null cast-aways are
578 // going away in this pass and we need to make this memory op depend on the
579 // gating null check.
581 // I tried to leave the CastPP's in. This makes the graph more accurate in
582 // some sense; we get to keep around the knowledge that an oop is not-null
583 // after some test. Alas, the CastPP's interfere with GVN (some values are
584 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
585 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
586 // some of the more trivial cases in the optimizer. Removing more useless
587 // Phi's started allowing Loads to illegally float above null checks. I gave
588 // up on this approach. CNC 10/20/2000
589 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
590 Node *ctr = in(MemNode::Control);
591 Node *mem = in(MemNode::Memory);
592 Node *adr = in(MemNode::Address);
593 Node *skipped_cast = NULL;
594 // Need a null check? Regular static accesses do not because they are
595 // from constant addresses. Array ops are gated by the range check (which
596 // always includes a NULL check). Just check field ops.
597 if( !ctr ) {
598 // Scan upwards for the highest location we can place this memory op.
599 while( true ) {
600 switch( adr->Opcode() ) {
602 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
603 adr = adr->in(AddPNode::Base);
604 continue;
606 case Op_DecodeN: // No change to NULL-ness, so peek thru
607 adr = adr->in(1);
608 continue;
610 case Op_CastPP:
611 // If the CastPP is useless, just peek on through it.
612 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
613 // Remember the cast that we've peeked though. If we peek
614 // through more than one, then we end up remembering the highest
615 // one, that is, if in a loop, the one closest to the top.
616 skipped_cast = adr;
617 adr = adr->in(1);
618 continue;
619 }
620 // CastPP is going away in this pass! We need this memory op to be
621 // control-dependent on the test that is guarding the CastPP.
622 ccp->hash_delete(this);
623 set_req(MemNode::Control, adr->in(0));
624 ccp->hash_insert(this);
625 return this;
627 case Op_Phi:
628 // Attempt to float above a Phi to some dominating point.
629 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
630 // If we've already peeked through a Cast (which could have set the
631 // control), we can't float above a Phi, because the skipped Cast
632 // may not be loop invariant.
633 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
634 adr = adr->in(1);
635 continue;
636 }
637 }
639 // Intentional fallthrough!
641 // No obvious dominating point. The mem op is pinned below the Phi
642 // by the Phi itself. If the Phi goes away (no true value is merged)
643 // then the mem op can float, but not indefinitely. It must be pinned
644 // behind the controls leading to the Phi.
645 case Op_CheckCastPP:
646 // These usually stick around to change address type, however a
647 // useless one can be elided and we still need to pick up a control edge
648 if (adr->in(0) == NULL) {
649 // This CheckCastPP node has NO control and is likely useless. But we
650 // need check further up the ancestor chain for a control input to keep
651 // the node in place. 4959717.
652 skipped_cast = adr;
653 adr = adr->in(1);
654 continue;
655 }
656 ccp->hash_delete(this);
657 set_req(MemNode::Control, adr->in(0));
658 ccp->hash_insert(this);
659 return this;
661 // List of "safe" opcodes; those that implicitly block the memory
662 // op below any null check.
663 case Op_CastX2P: // no null checks on native pointers
664 case Op_Parm: // 'this' pointer is not null
665 case Op_LoadP: // Loading from within a klass
666 case Op_LoadN: // Loading from within a klass
667 case Op_LoadKlass: // Loading from within a klass
668 case Op_ConP: // Loading from a klass
669 case Op_CreateEx: // Sucking up the guts of an exception oop
670 case Op_Con: // Reading from TLS
671 case Op_CMoveP: // CMoveP is pinned
672 break; // No progress
674 case Op_Proj: // Direct call to an allocation routine
675 case Op_SCMemProj: // Memory state from store conditional ops
676 #ifdef ASSERT
677 {
678 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
679 const Node* call = adr->in(0);
680 if (call->is_CallStaticJava()) {
681 const CallStaticJavaNode* call_java = call->as_CallStaticJava();
682 const TypeTuple *r = call_java->tf()->range();
683 assert(r->cnt() > TypeFunc::Parms, "must return value");
684 const Type* ret_type = r->field_at(TypeFunc::Parms);
685 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
686 // We further presume that this is one of
687 // new_instance_Java, new_array_Java, or
688 // the like, but do not assert for this.
689 } else if (call->is_Allocate()) {
690 // similar case to new_instance_Java, etc.
691 } else if (!call->is_CallLeaf()) {
692 // Projections from fetch_oop (OSR) are allowed as well.
693 ShouldNotReachHere();
694 }
695 }
696 #endif
697 break;
698 default:
699 ShouldNotReachHere();
700 }
701 break;
702 }
703 }
705 return NULL; // No progress
706 }
709 //=============================================================================
710 uint LoadNode::size_of() const { return sizeof(*this); }
711 uint LoadNode::cmp( const Node &n ) const
712 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
713 const Type *LoadNode::bottom_type() const { return _type; }
714 uint LoadNode::ideal_reg() const {
715 return Matcher::base2reg[_type->base()];
716 }
718 #ifndef PRODUCT
719 void LoadNode::dump_spec(outputStream *st) const {
720 MemNode::dump_spec(st);
721 if( !Verbose && !WizardMode ) {
722 // standard dump does this in Verbose and WizardMode
723 st->print(" #"); _type->dump_on(st);
724 }
725 }
726 #endif
729 //----------------------------LoadNode::make-----------------------------------
730 // Polymorphic factory method:
731 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
732 Compile* C = gvn.C;
734 // sanity check the alias category against the created node type
735 assert(!(adr_type->isa_oopptr() &&
736 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
737 "use LoadKlassNode instead");
738 assert(!(adr_type->isa_aryptr() &&
739 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
740 "use LoadRangeNode instead");
741 switch (bt) {
742 case T_BOOLEAN:
743 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
744 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
745 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
746 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
747 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
748 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
749 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
750 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
751 case T_OBJECT:
752 #ifdef _LP64
753 if (adr->bottom_type()->is_narrow()) {
754 const TypeNarrowOop* narrowtype;
755 if (rt->isa_narrowoop()) {
756 narrowtype = rt->is_narrowoop();
757 } else {
758 narrowtype = rt->is_oopptr()->make_narrowoop();
759 }
760 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, narrowtype));
762 return DecodeNNode::decode(&gvn, load);
763 } else
764 #endif
765 {
766 assert(!adr->bottom_type()->is_narrow(), "should have got back a narrow oop");
767 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
768 }
769 }
770 ShouldNotReachHere();
771 return (LoadNode*)NULL;
772 }
774 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
775 bool require_atomic = true;
776 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
777 }
782 //------------------------------hash-------------------------------------------
783 uint LoadNode::hash() const {
784 // unroll addition of interesting fields
785 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
786 }
788 //---------------------------can_see_stored_value------------------------------
789 // This routine exists to make sure this set of tests is done the same
790 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
791 // will change the graph shape in a way which makes memory alive twice at the
792 // same time (uses the Oracle model of aliasing), then some
793 // LoadXNode::Identity will fold things back to the equivalence-class model
794 // of aliasing.
795 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
796 Node* ld_adr = in(MemNode::Address);
798 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
799 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
800 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
801 atp->field() != NULL && !atp->field()->is_volatile()) {
802 uint alias_idx = atp->index();
803 bool final = atp->field()->is_final();
804 Node* result = NULL;
805 Node* current = st;
806 // Skip through chains of MemBarNodes checking the MergeMems for
807 // new states for the slice of this load. Stop once any other
808 // kind of node is encountered. Loads from final memory can skip
809 // through any kind of MemBar but normal loads shouldn't skip
810 // through MemBarAcquire since the could allow them to move out of
811 // a synchronized region.
812 while (current->is_Proj()) {
813 int opc = current->in(0)->Opcode();
814 if ((final && opc == Op_MemBarAcquire) ||
815 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
816 Node* mem = current->in(0)->in(TypeFunc::Memory);
817 if (mem->is_MergeMem()) {
818 MergeMemNode* merge = mem->as_MergeMem();
819 Node* new_st = merge->memory_at(alias_idx);
820 if (new_st == merge->base_memory()) {
821 // Keep searching
822 current = merge->base_memory();
823 continue;
824 }
825 // Save the new memory state for the slice and fall through
826 // to exit.
827 result = new_st;
828 }
829 }
830 break;
831 }
832 if (result != NULL) {
833 st = result;
834 }
835 }
838 // Loop around twice in the case Load -> Initialize -> Store.
839 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
840 for (int trip = 0; trip <= 1; trip++) {
842 if (st->is_Store()) {
843 Node* st_adr = st->in(MemNode::Address);
844 if (!phase->eqv(st_adr, ld_adr)) {
845 // Try harder before giving up... Match raw and non-raw pointers.
846 intptr_t st_off = 0;
847 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
848 if (alloc == NULL) return NULL;
849 intptr_t ld_off = 0;
850 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
851 if (alloc != allo2) return NULL;
852 if (ld_off != st_off) return NULL;
853 // At this point we have proven something like this setup:
854 // A = Allocate(...)
855 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
856 // S = StoreQ(, AddP(, A.Parm , #Off), V)
857 // (Actually, we haven't yet proven the Q's are the same.)
858 // In other words, we are loading from a casted version of
859 // the same pointer-and-offset that we stored to.
860 // Thus, we are able to replace L by V.
861 }
862 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
863 if (store_Opcode() != st->Opcode())
864 return NULL;
865 return st->in(MemNode::ValueIn);
866 }
868 intptr_t offset = 0; // scratch
870 // A load from a freshly-created object always returns zero.
871 // (This can happen after LoadNode::Ideal resets the load's memory input
872 // to find_captured_store, which returned InitializeNode::zero_memory.)
873 if (st->is_Proj() && st->in(0)->is_Allocate() &&
874 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
875 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
876 // return a zero value for the load's basic type
877 // (This is one of the few places where a generic PhaseTransform
878 // can create new nodes. Think of it as lazily manifesting
879 // virtually pre-existing constants.)
880 return phase->zerocon(memory_type());
881 }
883 // A load from an initialization barrier can match a captured store.
884 if (st->is_Proj() && st->in(0)->is_Initialize()) {
885 InitializeNode* init = st->in(0)->as_Initialize();
886 AllocateNode* alloc = init->allocation();
887 if (alloc != NULL &&
888 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
889 // examine a captured store value
890 st = init->find_captured_store(offset, memory_size(), phase);
891 if (st != NULL)
892 continue; // take one more trip around
893 }
894 }
896 break;
897 }
899 return NULL;
900 }
902 //----------------------is_instance_field_load_with_local_phi------------------
903 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
904 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
905 in(MemNode::Address)->is_AddP() ) {
906 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
907 // Only instances.
908 if( t_oop != NULL && t_oop->is_instance_field() &&
909 t_oop->offset() != Type::OffsetBot &&
910 t_oop->offset() != Type::OffsetTop) {
911 return true;
912 }
913 }
914 return false;
915 }
917 //------------------------------Identity---------------------------------------
918 // Loads are identity if previous store is to same address
919 Node *LoadNode::Identity( PhaseTransform *phase ) {
920 // If the previous store-maker is the right kind of Store, and the store is
921 // to the same address, then we are equal to the value stored.
922 Node* mem = in(MemNode::Memory);
923 Node* value = can_see_stored_value(mem, phase);
924 if( value ) {
925 // byte, short & char stores truncate naturally.
926 // A load has to load the truncated value which requires
927 // some sort of masking operation and that requires an
928 // Ideal call instead of an Identity call.
929 if (memory_size() < BytesPerInt) {
930 // If the input to the store does not fit with the load's result type,
931 // it must be truncated via an Ideal call.
932 if (!phase->type(value)->higher_equal(phase->type(this)))
933 return this;
934 }
935 // (This works even when value is a Con, but LoadNode::Value
936 // usually runs first, producing the singleton type of the Con.)
937 return value;
938 }
940 // Search for an existing data phi which was generated before for the same
941 // instance's field to avoid infinite genertion of phis in a loop.
942 Node *region = mem->in(0);
943 if (is_instance_field_load_with_local_phi(region)) {
944 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
945 int this_index = phase->C->get_alias_index(addr_t);
946 int this_offset = addr_t->offset();
947 int this_id = addr_t->is_oopptr()->instance_id();
948 const Type* this_type = bottom_type();
949 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
950 Node* phi = region->fast_out(i);
951 if (phi->is_Phi() && phi != mem &&
952 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
953 return phi;
954 }
955 }
956 }
958 return this;
959 }
962 // Returns true if the AliasType refers to the field that holds the
963 // cached box array. Currently only handles the IntegerCache case.
964 static bool is_autobox_cache(Compile::AliasType* atp) {
965 if (atp != NULL && atp->field() != NULL) {
966 ciField* field = atp->field();
967 ciSymbol* klass = field->holder()->name();
968 if (field->name() == ciSymbol::cache_field_name() &&
969 field->holder()->uses_default_loader() &&
970 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
971 return true;
972 }
973 }
974 return false;
975 }
977 // Fetch the base value in the autobox array
978 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
979 if (atp != NULL && atp->field() != NULL) {
980 ciField* field = atp->field();
981 ciSymbol* klass = field->holder()->name();
982 if (field->name() == ciSymbol::cache_field_name() &&
983 field->holder()->uses_default_loader() &&
984 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
985 assert(field->is_constant(), "what?");
986 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
987 // Fetch the box object at the base of the array and get its value
988 ciInstance* box = array->obj_at(0)->as_instance();
989 ciInstanceKlass* ik = box->klass()->as_instance_klass();
990 if (ik->nof_nonstatic_fields() == 1) {
991 // This should be true nonstatic_field_at requires calling
992 // nof_nonstatic_fields so check it anyway
993 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
994 cache_offset = c.as_int();
995 }
996 return true;
997 }
998 }
999 return false;
1000 }
1002 // Returns true if the AliasType refers to the value field of an
1003 // autobox object. Currently only handles Integer.
1004 static bool is_autobox_object(Compile::AliasType* atp) {
1005 if (atp != NULL && atp->field() != NULL) {
1006 ciField* field = atp->field();
1007 ciSymbol* klass = field->holder()->name();
1008 if (field->name() == ciSymbol::value_name() &&
1009 field->holder()->uses_default_loader() &&
1010 klass == ciSymbol::java_lang_Integer()) {
1011 return true;
1012 }
1013 }
1014 return false;
1015 }
1018 // We're loading from an object which has autobox behaviour.
1019 // If this object is result of a valueOf call we'll have a phi
1020 // merging a newly allocated object and a load from the cache.
1021 // We want to replace this load with the original incoming
1022 // argument to the valueOf call.
1023 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1024 Node* base = in(Address)->in(AddPNode::Base);
1025 if (base->is_Phi() && base->req() == 3) {
1026 AllocateNode* allocation = NULL;
1027 int allocation_index = -1;
1028 int load_index = -1;
1029 for (uint i = 1; i < base->req(); i++) {
1030 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1031 if (allocation != NULL) {
1032 allocation_index = i;
1033 load_index = 3 - allocation_index;
1034 break;
1035 }
1036 }
1037 LoadNode* load = NULL;
1038 if (allocation != NULL && base->in(load_index)->is_Load()) {
1039 load = base->in(load_index)->as_Load();
1040 }
1041 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1042 // Push the loads from the phi that comes from valueOf up
1043 // through it to allow elimination of the loads and the recovery
1044 // of the original value.
1045 Node* mem_phi = in(Memory);
1046 Node* offset = in(Address)->in(AddPNode::Offset);
1048 Node* in1 = clone();
1049 Node* in1_addr = in1->in(Address)->clone();
1050 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1051 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1052 in1_addr->set_req(AddPNode::Offset, offset);
1053 in1->set_req(0, base->in(allocation_index));
1054 in1->set_req(Address, in1_addr);
1055 in1->set_req(Memory, mem_phi->in(allocation_index));
1057 Node* in2 = clone();
1058 Node* in2_addr = in2->in(Address)->clone();
1059 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1060 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1061 in2_addr->set_req(AddPNode::Offset, offset);
1062 in2->set_req(0, base->in(load_index));
1063 in2->set_req(Address, in2_addr);
1064 in2->set_req(Memory, mem_phi->in(load_index));
1066 in1_addr = phase->transform(in1_addr);
1067 in1 = phase->transform(in1);
1068 in2_addr = phase->transform(in2_addr);
1069 in2 = phase->transform(in2);
1071 PhiNode* result = PhiNode::make_blank(base->in(0), this);
1072 result->set_req(allocation_index, in1);
1073 result->set_req(load_index, in2);
1074 return result;
1075 }
1076 } else if (base->is_Load()) {
1077 // Eliminate the load of Integer.value for integers from the cache
1078 // array by deriving the value from the index into the array.
1079 // Capture the offset of the load and then reverse the computation.
1080 Node* load_base = base->in(Address)->in(AddPNode::Base);
1081 if (load_base != NULL) {
1082 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1083 intptr_t cache_offset;
1084 int shift = -1;
1085 Node* cache = NULL;
1086 if (is_autobox_cache(atp)) {
1087 shift = exact_log2(type2aelembytes(T_OBJECT));
1088 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1089 }
1090 if (cache != NULL && base->in(Address)->is_AddP()) {
1091 Node* elements[4];
1092 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1093 int cache_low;
1094 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1095 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1096 // Add up all the offsets making of the address of the load
1097 Node* result = elements[0];
1098 for (int i = 1; i < count; i++) {
1099 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1100 }
1101 // Remove the constant offset from the address and then
1102 // remove the scaling of the offset to recover the original index.
1103 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1104 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1105 // Peel the shift off directly but wrap it in a dummy node
1106 // since Ideal can't return existing nodes
1107 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1108 } else {
1109 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1110 }
1111 #ifdef _LP64
1112 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1113 #endif
1114 return result;
1115 }
1116 }
1117 }
1118 }
1119 return NULL;
1120 }
1123 //------------------------------Ideal------------------------------------------
1124 // If the load is from Field memory and the pointer is non-null, we can
1125 // zero out the control input.
1126 // If the offset is constant and the base is an object allocation,
1127 // try to hook me up to the exact initializing store.
1128 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1129 Node* p = MemNode::Ideal_common(phase, can_reshape);
1130 if (p) return (p == NodeSentinel) ? NULL : p;
1132 Node* ctrl = in(MemNode::Control);
1133 Node* address = in(MemNode::Address);
1135 // Skip up past a SafePoint control. Cannot do this for Stores because
1136 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1137 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1138 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1139 ctrl = ctrl->in(0);
1140 set_req(MemNode::Control,ctrl);
1141 }
1143 // Check for useless control edge in some common special cases
1144 if (in(MemNode::Control) != NULL) {
1145 intptr_t ignore = 0;
1146 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1147 if (base != NULL
1148 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1149 && all_controls_dominate(base, phase->C->start())) {
1150 // A method-invariant, non-null address (constant or 'this' argument).
1151 set_req(MemNode::Control, NULL);
1152 }
1153 }
1155 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
1156 Node* base = in(Address)->in(AddPNode::Base);
1157 if (base != NULL) {
1158 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1159 if (is_autobox_object(atp)) {
1160 Node* result = eliminate_autobox(phase);
1161 if (result != NULL) return result;
1162 }
1163 }
1164 }
1166 Node* mem = in(MemNode::Memory);
1167 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1169 if (addr_t != NULL) {
1170 // try to optimize our memory input
1171 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1172 if (opt_mem != mem) {
1173 set_req(MemNode::Memory, opt_mem);
1174 return this;
1175 }
1176 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1177 if (can_reshape && opt_mem->is_Phi() &&
1178 (t_oop != NULL) && t_oop->is_instance_field()) {
1179 assert(t_oop->offset() != Type::OffsetBot && t_oop->offset() != Type::OffsetTop, "");
1180 Node *region = opt_mem->in(0);
1181 uint cnt = opt_mem->req();
1182 for( uint i = 1; i < cnt; i++ ) {
1183 Node *in = opt_mem->in(i);
1184 if( in == NULL ) {
1185 region = NULL; // Wait stable graph
1186 break;
1187 }
1188 }
1189 if (region != NULL) {
1190 // Check for loop invariant.
1191 if (cnt == 3) {
1192 for( uint i = 1; i < cnt; i++ ) {
1193 Node *in = opt_mem->in(i);
1194 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1195 if (m == opt_mem) {
1196 set_req(MemNode::Memory, opt_mem->in(cnt - i)); // Skip this phi.
1197 return this;
1198 }
1199 }
1200 }
1201 // Split through Phi (see original code in loopopts.cpp).
1202 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1204 // Do nothing here if Identity will find a value
1205 // (to avoid infinite chain of value phis generation).
1206 if ( !phase->eqv(this, this->Identity(phase)) )
1207 return NULL;
1209 const Type* this_type = this->bottom_type();
1210 int this_index = phase->C->get_alias_index(addr_t);
1211 int this_offset = addr_t->offset();
1212 int this_iid = addr_t->is_oopptr()->instance_id();
1213 int wins = 0;
1214 PhaseIterGVN *igvn = phase->is_IterGVN();
1215 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1216 for( uint i = 1; i < region->req(); i++ ) {
1217 Node *x;
1218 Node* the_clone = NULL;
1219 if( region->in(i) == phase->C->top() ) {
1220 x = phase->C->top(); // Dead path? Use a dead data op
1221 } else {
1222 x = this->clone(); // Else clone up the data op
1223 the_clone = x; // Remember for possible deletion.
1224 // Alter data node to use pre-phi inputs
1225 if( this->in(0) == region ) {
1226 x->set_req( 0, region->in(i) );
1227 } else {
1228 x->set_req( 0, NULL );
1229 }
1230 for( uint j = 1; j < this->req(); j++ ) {
1231 Node *in = this->in(j);
1232 if( in->is_Phi() && in->in(0) == region )
1233 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1234 }
1235 }
1236 // Check for a 'win' on some paths
1237 const Type *t = x->Value(igvn);
1239 bool singleton = t->singleton();
1241 // See comments in PhaseIdealLoop::split_thru_phi().
1242 if( singleton && t == Type::TOP ) {
1243 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1244 }
1246 if( singleton ) {
1247 wins++;
1248 x = igvn->makecon(t);
1249 } else {
1250 // We now call Identity to try to simplify the cloned node.
1251 // Note that some Identity methods call phase->type(this).
1252 // Make sure that the type array is big enough for
1253 // our new node, even though we may throw the node away.
1254 // (This tweaking with igvn only works because x is a new node.)
1255 igvn->set_type(x, t);
1256 Node *y = x->Identity(igvn);
1257 if( y != x ) {
1258 wins++;
1259 x = y;
1260 } else {
1261 y = igvn->hash_find(x);
1262 if( y ) {
1263 wins++;
1264 x = y;
1265 } else {
1266 // Else x is a new node we are keeping
1267 // We do not need register_new_node_with_optimizer
1268 // because set_type has already been called.
1269 igvn->_worklist.push(x);
1270 }
1271 }
1272 }
1273 if (x != the_clone && the_clone != NULL)
1274 igvn->remove_dead_node(the_clone);
1275 phi->set_req(i, x);
1276 }
1277 if( wins > 0 ) {
1278 // Record Phi
1279 igvn->register_new_node_with_optimizer(phi);
1280 return phi;
1281 } else {
1282 igvn->remove_dead_node(phi);
1283 }
1284 }
1285 }
1286 }
1288 // Check for prior store with a different base or offset; make Load
1289 // independent. Skip through any number of them. Bail out if the stores
1290 // are in an endless dead cycle and report no progress. This is a key
1291 // transform for Reflection. However, if after skipping through the Stores
1292 // we can't then fold up against a prior store do NOT do the transform as
1293 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1294 // array memory alive twice: once for the hoisted Load and again after the
1295 // bypassed Store. This situation only works if EVERYBODY who does
1296 // anti-dependence work knows how to bypass. I.e. we need all
1297 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1298 // the alias index stuff. So instead, peek through Stores and IFF we can
1299 // fold up, do so.
1300 Node* prev_mem = find_previous_store(phase);
1301 // Steps (a), (b): Walk past independent stores to find an exact match.
1302 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1303 // (c) See if we can fold up on the spot, but don't fold up here.
1304 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
1305 // just return a prior value, which is done by Identity calls.
1306 if (can_see_stored_value(prev_mem, phase)) {
1307 // Make ready for step (d):
1308 set_req(MemNode::Memory, prev_mem);
1309 return this;
1310 }
1311 }
1313 return NULL; // No further progress
1314 }
1316 // Helper to recognize certain Klass fields which are invariant across
1317 // some group of array types (e.g., int[] or all T[] where T < Object).
1318 const Type*
1319 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1320 ciKlass* klass) const {
1321 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1322 // The field is Klass::_modifier_flags. Return its (constant) value.
1323 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1324 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1325 return TypeInt::make(klass->modifier_flags());
1326 }
1327 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1328 // The field is Klass::_access_flags. Return its (constant) value.
1329 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1330 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1331 return TypeInt::make(klass->access_flags());
1332 }
1333 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1334 // The field is Klass::_layout_helper. Return its constant value if known.
1335 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1336 return TypeInt::make(klass->layout_helper());
1337 }
1339 // No match.
1340 return NULL;
1341 }
1343 //------------------------------Value-----------------------------------------
1344 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1345 // Either input is TOP ==> the result is TOP
1346 Node* mem = in(MemNode::Memory);
1347 const Type *t1 = phase->type(mem);
1348 if (t1 == Type::TOP) return Type::TOP;
1349 Node* adr = in(MemNode::Address);
1350 const TypePtr* tp = phase->type(adr)->isa_ptr();
1351 if (tp == NULL || tp->empty()) return Type::TOP;
1352 int off = tp->offset();
1353 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1355 // Try to guess loaded type from pointer type
1356 if (tp->base() == Type::AryPtr) {
1357 const Type *t = tp->is_aryptr()->elem();
1358 // Don't do this for integer types. There is only potential profit if
1359 // the element type t is lower than _type; that is, for int types, if _type is
1360 // more restrictive than t. This only happens here if one is short and the other
1361 // char (both 16 bits), and in those cases we've made an intentional decision
1362 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1363 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1364 //
1365 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1366 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1367 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1368 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1369 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1370 // In fact, that could have been the original type of p1, and p1 could have
1371 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1372 // expression (LShiftL quux 3) independently optimized to the constant 8.
1373 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1374 && Opcode() != Op_LoadKlass) {
1375 // t might actually be lower than _type, if _type is a unique
1376 // concrete subclass of abstract class t.
1377 // Make sure the reference is not into the header, by comparing
1378 // the offset against the offset of the start of the array's data.
1379 // Different array types begin at slightly different offsets (12 vs. 16).
1380 // We choose T_BYTE as an example base type that is least restrictive
1381 // as to alignment, which will therefore produce the smallest
1382 // possible base offset.
1383 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1384 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1385 const Type* jt = t->join(_type);
1386 // In any case, do not allow the join, per se, to empty out the type.
1387 if (jt->empty() && !t->empty()) {
1388 // This can happen if a interface-typed array narrows to a class type.
1389 jt = _type;
1390 }
1392 if (EliminateAutoBox) {
1393 // The pointers in the autobox arrays are always non-null
1394 Node* base = in(Address)->in(AddPNode::Base);
1395 if (base != NULL) {
1396 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1397 if (is_autobox_cache(atp)) {
1398 return jt->join(TypePtr::NOTNULL)->is_ptr();
1399 }
1400 }
1401 }
1402 return jt;
1403 }
1404 }
1405 } else if (tp->base() == Type::InstPtr) {
1406 assert( off != Type::OffsetBot ||
1407 // arrays can be cast to Objects
1408 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1409 // unsafe field access may not have a constant offset
1410 phase->C->has_unsafe_access(),
1411 "Field accesses must be precise" );
1412 // For oop loads, we expect the _type to be precise
1413 } else if (tp->base() == Type::KlassPtr) {
1414 assert( off != Type::OffsetBot ||
1415 // arrays can be cast to Objects
1416 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1417 // also allow array-loading from the primary supertype
1418 // array during subtype checks
1419 Opcode() == Op_LoadKlass,
1420 "Field accesses must be precise" );
1421 // For klass/static loads, we expect the _type to be precise
1422 }
1424 const TypeKlassPtr *tkls = tp->isa_klassptr();
1425 if (tkls != NULL && !StressReflectiveCode) {
1426 ciKlass* klass = tkls->klass();
1427 if (klass->is_loaded() && tkls->klass_is_exact()) {
1428 // We are loading a field from a Klass metaobject whose identity
1429 // is known at compile time (the type is "exact" or "precise").
1430 // Check for fields we know are maintained as constants by the VM.
1431 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1432 // The field is Klass::_super_check_offset. Return its (constant) value.
1433 // (Folds up type checking code.)
1434 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1435 return TypeInt::make(klass->super_check_offset());
1436 }
1437 // Compute index into primary_supers array
1438 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1439 // Check for overflowing; use unsigned compare to handle the negative case.
1440 if( depth < ciKlass::primary_super_limit() ) {
1441 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1442 // (Folds up type checking code.)
1443 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1444 ciKlass *ss = klass->super_of_depth(depth);
1445 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1446 }
1447 const Type* aift = load_array_final_field(tkls, klass);
1448 if (aift != NULL) return aift;
1449 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1450 && klass->is_array_klass()) {
1451 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1452 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1453 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1454 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1455 }
1456 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1457 // The field is Klass::_java_mirror. Return its (constant) value.
1458 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1459 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1460 return TypeInstPtr::make(klass->java_mirror());
1461 }
1462 }
1464 // We can still check if we are loading from the primary_supers array at a
1465 // shallow enough depth. Even though the klass is not exact, entries less
1466 // than or equal to its super depth are correct.
1467 if (klass->is_loaded() ) {
1468 ciType *inner = klass->klass();
1469 while( inner->is_obj_array_klass() )
1470 inner = inner->as_obj_array_klass()->base_element_type();
1471 if( inner->is_instance_klass() &&
1472 !inner->as_instance_klass()->flags().is_interface() ) {
1473 // Compute index into primary_supers array
1474 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1475 // Check for overflowing; use unsigned compare to handle the negative case.
1476 if( depth < ciKlass::primary_super_limit() &&
1477 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1478 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1479 // (Folds up type checking code.)
1480 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1481 ciKlass *ss = klass->super_of_depth(depth);
1482 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1483 }
1484 }
1485 }
1487 // If the type is enough to determine that the thing is not an array,
1488 // we can give the layout_helper a positive interval type.
1489 // This will help short-circuit some reflective code.
1490 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1491 && !klass->is_array_klass() // not directly typed as an array
1492 && !klass->is_interface() // specifically not Serializable & Cloneable
1493 && !klass->is_java_lang_Object() // not the supertype of all T[]
1494 ) {
1495 // Note: When interfaces are reliable, we can narrow the interface
1496 // test to (klass != Serializable && klass != Cloneable).
1497 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1498 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1499 // The key property of this type is that it folds up tests
1500 // for array-ness, since it proves that the layout_helper is positive.
1501 // Thus, a generic value like the basic object layout helper works fine.
1502 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1503 }
1504 }
1506 // If we are loading from a freshly-allocated object, produce a zero,
1507 // if the load is provably beyond the header of the object.
1508 // (Also allow a variable load from a fresh array to produce zero.)
1509 if (ReduceFieldZeroing) {
1510 Node* value = can_see_stored_value(mem,phase);
1511 if (value != NULL && value->is_Con())
1512 return value->bottom_type();
1513 }
1515 const TypeOopPtr *tinst = tp->isa_oopptr();
1516 if (tinst != NULL && tinst->is_instance_field()) {
1517 // If we have an instance type and our memory input is the
1518 // programs's initial memory state, there is no matching store,
1519 // so just return a zero of the appropriate type
1520 Node *mem = in(MemNode::Memory);
1521 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1522 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1523 return Type::get_zero_type(_type->basic_type());
1524 }
1525 }
1526 return _type;
1527 }
1529 //------------------------------match_edge-------------------------------------
1530 // Do we Match on this edge index or not? Match only the address.
1531 uint LoadNode::match_edge(uint idx) const {
1532 return idx == MemNode::Address;
1533 }
1535 //--------------------------LoadBNode::Ideal--------------------------------------
1536 //
1537 // If the previous store is to the same address as this load,
1538 // and the value stored was larger than a byte, replace this load
1539 // with the value stored truncated to a byte. If no truncation is
1540 // needed, the replacement is done in LoadNode::Identity().
1541 //
1542 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1543 Node* mem = in(MemNode::Memory);
1544 Node* value = can_see_stored_value(mem,phase);
1545 if( value && !phase->type(value)->higher_equal( _type ) ) {
1546 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1547 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1548 }
1549 // Identity call will handle the case where truncation is not needed.
1550 return LoadNode::Ideal(phase, can_reshape);
1551 }
1553 //--------------------------LoadCNode::Ideal--------------------------------------
1554 //
1555 // If the previous store is to the same address as this load,
1556 // and the value stored was larger than a char, replace this load
1557 // with the value stored truncated to a char. If no truncation is
1558 // needed, the replacement is done in LoadNode::Identity().
1559 //
1560 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1561 Node* mem = in(MemNode::Memory);
1562 Node* value = can_see_stored_value(mem,phase);
1563 if( value && !phase->type(value)->higher_equal( _type ) )
1564 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1565 // Identity call will handle the case where truncation is not needed.
1566 return LoadNode::Ideal(phase, can_reshape);
1567 }
1569 //--------------------------LoadSNode::Ideal--------------------------------------
1570 //
1571 // If the previous store is to the same address as this load,
1572 // and the value stored was larger than a short, replace this load
1573 // with the value stored truncated to a short. If no truncation is
1574 // needed, the replacement is done in LoadNode::Identity().
1575 //
1576 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1577 Node* mem = in(MemNode::Memory);
1578 Node* value = can_see_stored_value(mem,phase);
1579 if( value && !phase->type(value)->higher_equal( _type ) ) {
1580 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1581 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1582 }
1583 // Identity call will handle the case where truncation is not needed.
1584 return LoadNode::Ideal(phase, can_reshape);
1585 }
1587 //=============================================================================
1588 //------------------------------Value------------------------------------------
1589 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1590 // Either input is TOP ==> the result is TOP
1591 const Type *t1 = phase->type( in(MemNode::Memory) );
1592 if (t1 == Type::TOP) return Type::TOP;
1593 Node *adr = in(MemNode::Address);
1594 const Type *t2 = phase->type( adr );
1595 if (t2 == Type::TOP) return Type::TOP;
1596 const TypePtr *tp = t2->is_ptr();
1597 if (TypePtr::above_centerline(tp->ptr()) ||
1598 tp->ptr() == TypePtr::Null) return Type::TOP;
1600 // Return a more precise klass, if possible
1601 const TypeInstPtr *tinst = tp->isa_instptr();
1602 if (tinst != NULL) {
1603 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1604 int offset = tinst->offset();
1605 if (ik == phase->C->env()->Class_klass()
1606 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1607 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1608 // We are loading a special hidden field from a Class mirror object,
1609 // the field which points to the VM's Klass metaobject.
1610 ciType* t = tinst->java_mirror_type();
1611 // java_mirror_type returns non-null for compile-time Class constants.
1612 if (t != NULL) {
1613 // constant oop => constant klass
1614 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1615 return TypeKlassPtr::make(ciArrayKlass::make(t));
1616 }
1617 if (!t->is_klass()) {
1618 // a primitive Class (e.g., int.class) has NULL for a klass field
1619 return TypePtr::NULL_PTR;
1620 }
1621 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1622 return TypeKlassPtr::make(t->as_klass());
1623 }
1624 // non-constant mirror, so we can't tell what's going on
1625 }
1626 if( !ik->is_loaded() )
1627 return _type; // Bail out if not loaded
1628 if (offset == oopDesc::klass_offset_in_bytes()) {
1629 if (tinst->klass_is_exact()) {
1630 return TypeKlassPtr::make(ik);
1631 }
1632 // See if we can become precise: no subklasses and no interface
1633 // (Note: We need to support verified interfaces.)
1634 if (!ik->is_interface() && !ik->has_subklass()) {
1635 //assert(!UseExactTypes, "this code should be useless with exact types");
1636 // Add a dependence; if any subclass added we need to recompile
1637 if (!ik->is_final()) {
1638 // %%% should use stronger assert_unique_concrete_subtype instead
1639 phase->C->dependencies()->assert_leaf_type(ik);
1640 }
1641 // Return precise klass
1642 return TypeKlassPtr::make(ik);
1643 }
1645 // Return root of possible klass
1646 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1647 }
1648 }
1650 // Check for loading klass from an array
1651 const TypeAryPtr *tary = tp->isa_aryptr();
1652 if( tary != NULL ) {
1653 ciKlass *tary_klass = tary->klass();
1654 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1655 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1656 if (tary->klass_is_exact()) {
1657 return TypeKlassPtr::make(tary_klass);
1658 }
1659 ciArrayKlass *ak = tary->klass()->as_array_klass();
1660 // If the klass is an object array, we defer the question to the
1661 // array component klass.
1662 if( ak->is_obj_array_klass() ) {
1663 assert( ak->is_loaded(), "" );
1664 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1665 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1666 ciInstanceKlass* ik = base_k->as_instance_klass();
1667 // See if we can become precise: no subklasses and no interface
1668 if (!ik->is_interface() && !ik->has_subklass()) {
1669 //assert(!UseExactTypes, "this code should be useless with exact types");
1670 // Add a dependence; if any subclass added we need to recompile
1671 if (!ik->is_final()) {
1672 phase->C->dependencies()->assert_leaf_type(ik);
1673 }
1674 // Return precise array klass
1675 return TypeKlassPtr::make(ak);
1676 }
1677 }
1678 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1679 } else { // Found a type-array?
1680 //assert(!UseExactTypes, "this code should be useless with exact types");
1681 assert( ak->is_type_array_klass(), "" );
1682 return TypeKlassPtr::make(ak); // These are always precise
1683 }
1684 }
1685 }
1687 // Check for loading klass from an array klass
1688 const TypeKlassPtr *tkls = tp->isa_klassptr();
1689 if (tkls != NULL && !StressReflectiveCode) {
1690 ciKlass* klass = tkls->klass();
1691 if( !klass->is_loaded() )
1692 return _type; // Bail out if not loaded
1693 if( klass->is_obj_array_klass() &&
1694 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1695 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1696 // // Always returning precise element type is incorrect,
1697 // // e.g., element type could be object and array may contain strings
1698 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1700 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1701 // according to the element type's subclassing.
1702 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1703 }
1704 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1705 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1706 ciKlass* sup = klass->as_instance_klass()->super();
1707 // The field is Klass::_super. Return its (constant) value.
1708 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1709 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1710 }
1711 }
1713 // Bailout case
1714 return LoadNode::Value(phase);
1715 }
1717 //------------------------------Identity---------------------------------------
1718 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1719 // Also feed through the klass in Allocate(...klass...)._klass.
1720 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1721 Node* x = LoadNode::Identity(phase);
1722 if (x != this) return x;
1724 // Take apart the address into an oop and and offset.
1725 // Return 'this' if we cannot.
1726 Node* adr = in(MemNode::Address);
1727 intptr_t offset = 0;
1728 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1729 if (base == NULL) return this;
1730 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1731 if (toop == NULL) return this;
1733 // We can fetch the klass directly through an AllocateNode.
1734 // This works even if the klass is not constant (clone or newArray).
1735 if (offset == oopDesc::klass_offset_in_bytes()) {
1736 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1737 if (allocated_klass != NULL) {
1738 return allocated_klass;
1739 }
1740 }
1742 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1743 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1744 // See inline_native_Class_query for occurrences of these patterns.
1745 // Java Example: x.getClass().isAssignableFrom(y)
1746 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1747 //
1748 // This improves reflective code, often making the Class
1749 // mirror go completely dead. (Current exception: Class
1750 // mirrors may appear in debug info, but we could clean them out by
1751 // introducing a new debug info operator for klassOop.java_mirror).
1752 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1753 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1754 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1755 // We are loading a special hidden field from a Class mirror,
1756 // the field which points to its Klass or arrayKlass metaobject.
1757 if (base->is_Load()) {
1758 Node* adr2 = base->in(MemNode::Address);
1759 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1760 if (tkls != NULL && !tkls->empty()
1761 && (tkls->klass()->is_instance_klass() ||
1762 tkls->klass()->is_array_klass())
1763 && adr2->is_AddP()
1764 ) {
1765 int mirror_field = Klass::java_mirror_offset_in_bytes();
1766 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1767 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1768 }
1769 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1770 return adr2->in(AddPNode::Base);
1771 }
1772 }
1773 }
1774 }
1776 return this;
1777 }
1779 //------------------------------Value-----------------------------------------
1780 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1781 // Either input is TOP ==> the result is TOP
1782 const Type *t1 = phase->type( in(MemNode::Memory) );
1783 if( t1 == Type::TOP ) return Type::TOP;
1784 Node *adr = in(MemNode::Address);
1785 const Type *t2 = phase->type( adr );
1786 if( t2 == Type::TOP ) return Type::TOP;
1787 const TypePtr *tp = t2->is_ptr();
1788 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1789 const TypeAryPtr *tap = tp->isa_aryptr();
1790 if( !tap ) return _type;
1791 return tap->size();
1792 }
1794 //------------------------------Identity---------------------------------------
1795 // Feed through the length in AllocateArray(...length...)._length.
1796 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1797 Node* x = LoadINode::Identity(phase);
1798 if (x != this) return x;
1800 // Take apart the address into an oop and and offset.
1801 // Return 'this' if we cannot.
1802 Node* adr = in(MemNode::Address);
1803 intptr_t offset = 0;
1804 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1805 if (base == NULL) return this;
1806 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1807 if (tary == NULL) return this;
1809 // We can fetch the length directly through an AllocateArrayNode.
1810 // This works even if the length is not constant (clone or newArray).
1811 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1812 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
1813 if (allocated_length != NULL) {
1814 return allocated_length;
1815 }
1816 }
1818 return this;
1820 }
1821 //=============================================================================
1822 //---------------------------StoreNode::make-----------------------------------
1823 // Polymorphic factory method:
1824 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1825 Compile* C = gvn.C;
1827 switch (bt) {
1828 case T_BOOLEAN:
1829 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1830 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1831 case T_CHAR:
1832 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1833 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1834 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1835 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1836 case T_ADDRESS:
1837 case T_OBJECT:
1838 #ifdef _LP64
1839 if (adr->bottom_type()->is_narrow() ||
1840 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
1841 adr->bottom_type()->isa_rawptr())) {
1842 const TypePtr* type = val->bottom_type()->is_ptr();
1843 Node* cp = EncodePNode::encode(&gvn, val);
1844 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, cp);
1845 } else
1846 #endif
1847 {
1848 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1849 }
1850 }
1851 ShouldNotReachHere();
1852 return (StoreNode*)NULL;
1853 }
1855 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1856 bool require_atomic = true;
1857 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1858 }
1861 //--------------------------bottom_type----------------------------------------
1862 const Type *StoreNode::bottom_type() const {
1863 return Type::MEMORY;
1864 }
1866 //------------------------------hash-------------------------------------------
1867 uint StoreNode::hash() const {
1868 // unroll addition of interesting fields
1869 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1871 // Since they are not commoned, do not hash them:
1872 return NO_HASH;
1873 }
1875 //------------------------------Ideal------------------------------------------
1876 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1877 // When a store immediately follows a relevant allocation/initialization,
1878 // try to capture it into the initialization, or hoist it above.
1879 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1880 Node* p = MemNode::Ideal_common(phase, can_reshape);
1881 if (p) return (p == NodeSentinel) ? NULL : p;
1883 Node* mem = in(MemNode::Memory);
1884 Node* address = in(MemNode::Address);
1886 // Back-to-back stores to same address? Fold em up.
1887 // Generally unsafe if I have intervening uses...
1888 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
1889 // Looking at a dead closed cycle of memory?
1890 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1892 assert(Opcode() == mem->Opcode() ||
1893 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1894 "no mismatched stores, except on raw memory");
1896 if (mem->outcnt() == 1 && // check for intervening uses
1897 mem->as_Store()->memory_size() <= this->memory_size()) {
1898 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1899 // For example, 'mem' might be the final state at a conditional return.
1900 // Or, 'mem' might be used by some node which is live at the same time
1901 // 'this' is live, which might be unschedulable. So, require exactly
1902 // ONE user, the 'this' store, until such time as we clone 'mem' for
1903 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1904 if (can_reshape) { // (%%% is this an anachronism?)
1905 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1906 phase->is_IterGVN());
1907 } else {
1908 // It's OK to do this in the parser, since DU info is always accurate,
1909 // and the parser always refers to nodes via SafePointNode maps.
1910 set_req(MemNode::Memory, mem->in(MemNode::Memory));
1911 }
1912 return this;
1913 }
1914 }
1916 // Capture an unaliased, unconditional, simple store into an initializer.
1917 // Or, if it is independent of the allocation, hoist it above the allocation.
1918 if (ReduceFieldZeroing && /*can_reshape &&*/
1919 mem->is_Proj() && mem->in(0)->is_Initialize()) {
1920 InitializeNode* init = mem->in(0)->as_Initialize();
1921 intptr_t offset = init->can_capture_store(this, phase);
1922 if (offset > 0) {
1923 Node* moved = init->capture_store(this, offset, phase);
1924 // If the InitializeNode captured me, it made a raw copy of me,
1925 // and I need to disappear.
1926 if (moved != NULL) {
1927 // %%% hack to ensure that Ideal returns a new node:
1928 mem = MergeMemNode::make(phase->C, mem);
1929 return mem; // fold me away
1930 }
1931 }
1932 }
1934 return NULL; // No further progress
1935 }
1937 //------------------------------Value-----------------------------------------
1938 const Type *StoreNode::Value( PhaseTransform *phase ) const {
1939 // Either input is TOP ==> the result is TOP
1940 const Type *t1 = phase->type( in(MemNode::Memory) );
1941 if( t1 == Type::TOP ) return Type::TOP;
1942 const Type *t2 = phase->type( in(MemNode::Address) );
1943 if( t2 == Type::TOP ) return Type::TOP;
1944 const Type *t3 = phase->type( in(MemNode::ValueIn) );
1945 if( t3 == Type::TOP ) return Type::TOP;
1946 return Type::MEMORY;
1947 }
1949 //------------------------------Identity---------------------------------------
1950 // Remove redundant stores:
1951 // Store(m, p, Load(m, p)) changes to m.
1952 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
1953 Node *StoreNode::Identity( PhaseTransform *phase ) {
1954 Node* mem = in(MemNode::Memory);
1955 Node* adr = in(MemNode::Address);
1956 Node* val = in(MemNode::ValueIn);
1958 // Load then Store? Then the Store is useless
1959 if (val->is_Load() &&
1960 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
1961 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
1962 val->as_Load()->store_Opcode() == Opcode()) {
1963 return mem;
1964 }
1966 // Two stores in a row of the same value?
1967 if (mem->is_Store() &&
1968 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
1969 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
1970 mem->Opcode() == Opcode()) {
1971 return mem;
1972 }
1974 // Store of zero anywhere into a freshly-allocated object?
1975 // Then the store is useless.
1976 // (It must already have been captured by the InitializeNode.)
1977 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
1978 // a newly allocated object is already all-zeroes everywhere
1979 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
1980 return mem;
1981 }
1983 // the store may also apply to zero-bits in an earlier object
1984 Node* prev_mem = find_previous_store(phase);
1985 // Steps (a), (b): Walk past independent stores to find an exact match.
1986 if (prev_mem != NULL) {
1987 Node* prev_val = can_see_stored_value(prev_mem, phase);
1988 if (prev_val != NULL && phase->eqv(prev_val, val)) {
1989 // prev_val and val might differ by a cast; it would be good
1990 // to keep the more informative of the two.
1991 return mem;
1992 }
1993 }
1994 }
1996 return this;
1997 }
1999 //------------------------------match_edge-------------------------------------
2000 // Do we Match on this edge index or not? Match only memory & value
2001 uint StoreNode::match_edge(uint idx) const {
2002 return idx == MemNode::Address || idx == MemNode::ValueIn;
2003 }
2005 //------------------------------cmp--------------------------------------------
2006 // Do not common stores up together. They generally have to be split
2007 // back up anyways, so do not bother.
2008 uint StoreNode::cmp( const Node &n ) const {
2009 return (&n == this); // Always fail except on self
2010 }
2012 //------------------------------Ideal_masked_input-----------------------------
2013 // Check for a useless mask before a partial-word store
2014 // (StoreB ... (AndI valIn conIa) )
2015 // If (conIa & mask == mask) this simplifies to
2016 // (StoreB ... (valIn) )
2017 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2018 Node *val = in(MemNode::ValueIn);
2019 if( val->Opcode() == Op_AndI ) {
2020 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2021 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2022 set_req(MemNode::ValueIn, val->in(1));
2023 return this;
2024 }
2025 }
2026 return NULL;
2027 }
2030 //------------------------------Ideal_sign_extended_input----------------------
2031 // Check for useless sign-extension before a partial-word store
2032 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2033 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2034 // (StoreB ... (valIn) )
2035 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2036 Node *val = in(MemNode::ValueIn);
2037 if( val->Opcode() == Op_RShiftI ) {
2038 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2039 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2040 Node *shl = val->in(1);
2041 if( shl->Opcode() == Op_LShiftI ) {
2042 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2043 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2044 set_req(MemNode::ValueIn, shl->in(1));
2045 return this;
2046 }
2047 }
2048 }
2049 }
2050 return NULL;
2051 }
2053 //------------------------------value_never_loaded-----------------------------------
2054 // Determine whether there are any possible loads of the value stored.
2055 // For simplicity, we actually check if there are any loads from the
2056 // address stored to, not just for loads of the value stored by this node.
2057 //
2058 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2059 Node *adr = in(Address);
2060 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2061 if (adr_oop == NULL)
2062 return false;
2063 if (!adr_oop->is_instance_field())
2064 return false; // if not a distinct instance, there may be aliases of the address
2065 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2066 Node *use = adr->fast_out(i);
2067 int opc = use->Opcode();
2068 if (use->is_Load() || use->is_LoadStore()) {
2069 return false;
2070 }
2071 }
2072 return true;
2073 }
2075 //=============================================================================
2076 //------------------------------Ideal------------------------------------------
2077 // If the store is from an AND mask that leaves the low bits untouched, then
2078 // we can skip the AND operation. If the store is from a sign-extension
2079 // (a left shift, then right shift) we can skip both.
2080 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2081 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2082 if( progress != NULL ) return progress;
2084 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2085 if( progress != NULL ) return progress;
2087 // Finally check the default case
2088 return StoreNode::Ideal(phase, can_reshape);
2089 }
2091 //=============================================================================
2092 //------------------------------Ideal------------------------------------------
2093 // If the store is from an AND mask that leaves the low bits untouched, then
2094 // we can skip the AND operation
2095 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2096 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2097 if( progress != NULL ) return progress;
2099 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2100 if( progress != NULL ) return progress;
2102 // Finally check the default case
2103 return StoreNode::Ideal(phase, can_reshape);
2104 }
2106 //=============================================================================
2107 //------------------------------Identity---------------------------------------
2108 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2109 // No need to card mark when storing a null ptr
2110 Node* my_store = in(MemNode::OopStore);
2111 if (my_store->is_Store()) {
2112 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2113 if( t1 == TypePtr::NULL_PTR ) {
2114 return in(MemNode::Memory);
2115 }
2116 }
2117 return this;
2118 }
2120 //------------------------------Value-----------------------------------------
2121 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2122 // Either input is TOP ==> the result is TOP
2123 const Type *t = phase->type( in(MemNode::Memory) );
2124 if( t == Type::TOP ) return Type::TOP;
2125 t = phase->type( in(MemNode::Address) );
2126 if( t == Type::TOP ) return Type::TOP;
2127 t = phase->type( in(MemNode::ValueIn) );
2128 if( t == Type::TOP ) return Type::TOP;
2129 // If extra input is TOP ==> the result is TOP
2130 t = phase->type( in(MemNode::OopStore) );
2131 if( t == Type::TOP ) return Type::TOP;
2133 return StoreNode::Value( phase );
2134 }
2137 //=============================================================================
2138 //----------------------------------SCMemProjNode------------------------------
2139 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2140 {
2141 return bottom_type();
2142 }
2144 //=============================================================================
2145 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2146 init_req(MemNode::Control, c );
2147 init_req(MemNode::Memory , mem);
2148 init_req(MemNode::Address, adr);
2149 init_req(MemNode::ValueIn, val);
2150 init_req( ExpectedIn, ex );
2151 init_class_id(Class_LoadStore);
2153 }
2155 //=============================================================================
2156 //-------------------------------adr_type--------------------------------------
2157 // Do we Match on this edge index or not? Do not match memory
2158 const TypePtr* ClearArrayNode::adr_type() const {
2159 Node *adr = in(3);
2160 return MemNode::calculate_adr_type(adr->bottom_type());
2161 }
2163 //------------------------------match_edge-------------------------------------
2164 // Do we Match on this edge index or not? Do not match memory
2165 uint ClearArrayNode::match_edge(uint idx) const {
2166 return idx > 1;
2167 }
2169 //------------------------------Identity---------------------------------------
2170 // Clearing a zero length array does nothing
2171 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2172 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2173 }
2175 //------------------------------Idealize---------------------------------------
2176 // Clearing a short array is faster with stores
2177 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2178 const int unit = BytesPerLong;
2179 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2180 if (!t) return NULL;
2181 if (!t->is_con()) return NULL;
2182 intptr_t raw_count = t->get_con();
2183 intptr_t size = raw_count;
2184 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2185 // Clearing nothing uses the Identity call.
2186 // Negative clears are possible on dead ClearArrays
2187 // (see jck test stmt114.stmt11402.val).
2188 if (size <= 0 || size % unit != 0) return NULL;
2189 intptr_t count = size / unit;
2190 // Length too long; use fast hardware clear
2191 if (size > Matcher::init_array_short_size) return NULL;
2192 Node *mem = in(1);
2193 if( phase->type(mem)==Type::TOP ) return NULL;
2194 Node *adr = in(3);
2195 const Type* at = phase->type(adr);
2196 if( at==Type::TOP ) return NULL;
2197 const TypePtr* atp = at->isa_ptr();
2198 // adjust atp to be the correct array element address type
2199 if (atp == NULL) atp = TypePtr::BOTTOM;
2200 else atp = atp->add_offset(Type::OffsetBot);
2201 // Get base for derived pointer purposes
2202 if( adr->Opcode() != Op_AddP ) Unimplemented();
2203 Node *base = adr->in(1);
2205 Node *zero = phase->makecon(TypeLong::ZERO);
2206 Node *off = phase->MakeConX(BytesPerLong);
2207 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2208 count--;
2209 while( count-- ) {
2210 mem = phase->transform(mem);
2211 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2212 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2213 }
2214 return mem;
2215 }
2217 //----------------------------clear_memory-------------------------------------
2218 // Generate code to initialize object storage to zero.
2219 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2220 intptr_t start_offset,
2221 Node* end_offset,
2222 PhaseGVN* phase) {
2223 Compile* C = phase->C;
2224 intptr_t offset = start_offset;
2226 int unit = BytesPerLong;
2227 if ((offset % unit) != 0) {
2228 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2229 adr = phase->transform(adr);
2230 const TypePtr* atp = TypeRawPtr::BOTTOM;
2231 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2232 mem = phase->transform(mem);
2233 offset += BytesPerInt;
2234 }
2235 assert((offset % unit) == 0, "");
2237 // Initialize the remaining stuff, if any, with a ClearArray.
2238 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2239 }
2241 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2242 Node* start_offset,
2243 Node* end_offset,
2244 PhaseGVN* phase) {
2245 if (start_offset == end_offset) {
2246 // nothing to do
2247 return mem;
2248 }
2250 Compile* C = phase->C;
2251 int unit = BytesPerLong;
2252 Node* zbase = start_offset;
2253 Node* zend = end_offset;
2255 // Scale to the unit required by the CPU:
2256 if (!Matcher::init_array_count_is_in_bytes) {
2257 Node* shift = phase->intcon(exact_log2(unit));
2258 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2259 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2260 }
2262 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2263 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2265 // Bulk clear double-words
2266 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2267 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2268 return phase->transform(mem);
2269 }
2271 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2272 intptr_t start_offset,
2273 intptr_t end_offset,
2274 PhaseGVN* phase) {
2275 if (start_offset == end_offset) {
2276 // nothing to do
2277 return mem;
2278 }
2280 Compile* C = phase->C;
2281 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2282 intptr_t done_offset = end_offset;
2283 if ((done_offset % BytesPerLong) != 0) {
2284 done_offset -= BytesPerInt;
2285 }
2286 if (done_offset > start_offset) {
2287 mem = clear_memory(ctl, mem, dest,
2288 start_offset, phase->MakeConX(done_offset), phase);
2289 }
2290 if (done_offset < end_offset) { // emit the final 32-bit store
2291 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2292 adr = phase->transform(adr);
2293 const TypePtr* atp = TypeRawPtr::BOTTOM;
2294 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2295 mem = phase->transform(mem);
2296 done_offset += BytesPerInt;
2297 }
2298 assert(done_offset == end_offset, "");
2299 return mem;
2300 }
2302 //=============================================================================
2303 // Do we match on this edge? No memory edges
2304 uint StrCompNode::match_edge(uint idx) const {
2305 return idx == 5 || idx == 6;
2306 }
2308 //------------------------------Ideal------------------------------------------
2309 // Return a node which is more "ideal" than the current node. Strip out
2310 // control copies
2311 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2312 return remove_dead_region(phase, can_reshape) ? this : NULL;
2313 }
2316 //=============================================================================
2317 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2318 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2319 _adr_type(C->get_adr_type(alias_idx))
2320 {
2321 init_class_id(Class_MemBar);
2322 Node* top = C->top();
2323 init_req(TypeFunc::I_O,top);
2324 init_req(TypeFunc::FramePtr,top);
2325 init_req(TypeFunc::ReturnAdr,top);
2326 if (precedent != NULL)
2327 init_req(TypeFunc::Parms, precedent);
2328 }
2330 //------------------------------cmp--------------------------------------------
2331 uint MemBarNode::hash() const { return NO_HASH; }
2332 uint MemBarNode::cmp( const Node &n ) const {
2333 return (&n == this); // Always fail except on self
2334 }
2336 //------------------------------make-------------------------------------------
2337 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2338 int len = Precedent + (pn == NULL? 0: 1);
2339 switch (opcode) {
2340 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2341 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2342 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2343 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2344 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2345 default: ShouldNotReachHere(); return NULL;
2346 }
2347 }
2349 //------------------------------Ideal------------------------------------------
2350 // Return a node which is more "ideal" than the current node. Strip out
2351 // control copies
2352 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2353 if (remove_dead_region(phase, can_reshape)) return this;
2354 return NULL;
2355 }
2357 //------------------------------Value------------------------------------------
2358 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2359 if( !in(0) ) return Type::TOP;
2360 if( phase->type(in(0)) == Type::TOP )
2361 return Type::TOP;
2362 return TypeTuple::MEMBAR;
2363 }
2365 //------------------------------match------------------------------------------
2366 // Construct projections for memory.
2367 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2368 switch (proj->_con) {
2369 case TypeFunc::Control:
2370 case TypeFunc::Memory:
2371 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2372 }
2373 ShouldNotReachHere();
2374 return NULL;
2375 }
2377 //===========================InitializeNode====================================
2378 // SUMMARY:
2379 // This node acts as a memory barrier on raw memory, after some raw stores.
2380 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2381 // The Initialize can 'capture' suitably constrained stores as raw inits.
2382 // It can coalesce related raw stores into larger units (called 'tiles').
2383 // It can avoid zeroing new storage for memory units which have raw inits.
2384 // At macro-expansion, it is marked 'complete', and does not optimize further.
2385 //
2386 // EXAMPLE:
2387 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2388 // ctl = incoming control; mem* = incoming memory
2389 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2390 // First allocate uninitialized memory and fill in the header:
2391 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2392 // ctl := alloc.Control; mem* := alloc.Memory*
2393 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2394 // Then initialize to zero the non-header parts of the raw memory block:
2395 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2396 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2397 // After the initialize node executes, the object is ready for service:
2398 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2399 // Suppose its body is immediately initialized as {1,2}:
2400 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2401 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2402 // mem.SLICE(#short[*]) := store2
2403 //
2404 // DETAILS:
2405 // An InitializeNode collects and isolates object initialization after
2406 // an AllocateNode and before the next possible safepoint. As a
2407 // memory barrier (MemBarNode), it keeps critical stores from drifting
2408 // down past any safepoint or any publication of the allocation.
2409 // Before this barrier, a newly-allocated object may have uninitialized bits.
2410 // After this barrier, it may be treated as a real oop, and GC is allowed.
2411 //
2412 // The semantics of the InitializeNode include an implicit zeroing of
2413 // the new object from object header to the end of the object.
2414 // (The object header and end are determined by the AllocateNode.)
2415 //
2416 // Certain stores may be added as direct inputs to the InitializeNode.
2417 // These stores must update raw memory, and they must be to addresses
2418 // derived from the raw address produced by AllocateNode, and with
2419 // a constant offset. They must be ordered by increasing offset.
2420 // The first one is at in(RawStores), the last at in(req()-1).
2421 // Unlike most memory operations, they are not linked in a chain,
2422 // but are displayed in parallel as users of the rawmem output of
2423 // the allocation.
2424 //
2425 // (See comments in InitializeNode::capture_store, which continue
2426 // the example given above.)
2427 //
2428 // When the associated Allocate is macro-expanded, the InitializeNode
2429 // may be rewritten to optimize collected stores. A ClearArrayNode
2430 // may also be created at that point to represent any required zeroing.
2431 // The InitializeNode is then marked 'complete', prohibiting further
2432 // capturing of nearby memory operations.
2433 //
2434 // During macro-expansion, all captured initializations which store
2435 // constant values of 32 bits or smaller are coalesced (if advantagous)
2436 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2437 // initialized in fewer memory operations. Memory words which are
2438 // covered by neither tiles nor non-constant stores are pre-zeroed
2439 // by explicit stores of zero. (The code shape happens to do all
2440 // zeroing first, then all other stores, with both sequences occurring
2441 // in order of ascending offsets.)
2442 //
2443 // Alternatively, code may be inserted between an AllocateNode and its
2444 // InitializeNode, to perform arbitrary initialization of the new object.
2445 // E.g., the object copying intrinsics insert complex data transfers here.
2446 // The initialization must then be marked as 'complete' disable the
2447 // built-in zeroing semantics and the collection of initializing stores.
2448 //
2449 // While an InitializeNode is incomplete, reads from the memory state
2450 // produced by it are optimizable if they match the control edge and
2451 // new oop address associated with the allocation/initialization.
2452 // They return a stored value (if the offset matches) or else zero.
2453 // A write to the memory state, if it matches control and address,
2454 // and if it is to a constant offset, may be 'captured' by the
2455 // InitializeNode. It is cloned as a raw memory operation and rewired
2456 // inside the initialization, to the raw oop produced by the allocation.
2457 // Operations on addresses which are provably distinct (e.g., to
2458 // other AllocateNodes) are allowed to bypass the initialization.
2459 //
2460 // The effect of all this is to consolidate object initialization
2461 // (both arrays and non-arrays, both piecewise and bulk) into a
2462 // single location, where it can be optimized as a unit.
2463 //
2464 // Only stores with an offset less than TrackedInitializationLimit words
2465 // will be considered for capture by an InitializeNode. This puts a
2466 // reasonable limit on the complexity of optimized initializations.
2468 //---------------------------InitializeNode------------------------------------
2469 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2470 : _is_complete(false),
2471 MemBarNode(C, adr_type, rawoop)
2472 {
2473 init_class_id(Class_Initialize);
2475 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2476 assert(in(RawAddress) == rawoop, "proper init");
2477 // Note: allocation() can be NULL, for secondary initialization barriers
2478 }
2480 // Since this node is not matched, it will be processed by the
2481 // register allocator. Declare that there are no constraints
2482 // on the allocation of the RawAddress edge.
2483 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2484 // This edge should be set to top, by the set_complete. But be conservative.
2485 if (idx == InitializeNode::RawAddress)
2486 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2487 return RegMask::Empty;
2488 }
2490 Node* InitializeNode::memory(uint alias_idx) {
2491 Node* mem = in(Memory);
2492 if (mem->is_MergeMem()) {
2493 return mem->as_MergeMem()->memory_at(alias_idx);
2494 } else {
2495 // incoming raw memory is not split
2496 return mem;
2497 }
2498 }
2500 bool InitializeNode::is_non_zero() {
2501 if (is_complete()) return false;
2502 remove_extra_zeroes();
2503 return (req() > RawStores);
2504 }
2506 void InitializeNode::set_complete(PhaseGVN* phase) {
2507 assert(!is_complete(), "caller responsibility");
2508 _is_complete = true;
2510 // After this node is complete, it contains a bunch of
2511 // raw-memory initializations. There is no need for
2512 // it to have anything to do with non-raw memory effects.
2513 // Therefore, tell all non-raw users to re-optimize themselves,
2514 // after skipping the memory effects of this initialization.
2515 PhaseIterGVN* igvn = phase->is_IterGVN();
2516 if (igvn) igvn->add_users_to_worklist(this);
2517 }
2519 // convenience function
2520 // return false if the init contains any stores already
2521 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2522 InitializeNode* init = initialization();
2523 if (init == NULL || init->is_complete()) return false;
2524 init->remove_extra_zeroes();
2525 // for now, if this allocation has already collected any inits, bail:
2526 if (init->is_non_zero()) return false;
2527 init->set_complete(phase);
2528 return true;
2529 }
2531 void InitializeNode::remove_extra_zeroes() {
2532 if (req() == RawStores) return;
2533 Node* zmem = zero_memory();
2534 uint fill = RawStores;
2535 for (uint i = fill; i < req(); i++) {
2536 Node* n = in(i);
2537 if (n->is_top() || n == zmem) continue; // skip
2538 if (fill < i) set_req(fill, n); // compact
2539 ++fill;
2540 }
2541 // delete any empty spaces created:
2542 while (fill < req()) {
2543 del_req(fill);
2544 }
2545 }
2547 // Helper for remembering which stores go with which offsets.
2548 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2549 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2550 intptr_t offset = -1;
2551 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2552 phase, offset);
2553 if (base == NULL) return -1; // something is dead,
2554 if (offset < 0) return -1; // dead, dead
2555 return offset;
2556 }
2558 // Helper for proving that an initialization expression is
2559 // "simple enough" to be folded into an object initialization.
2560 // Attempts to prove that a store's initial value 'n' can be captured
2561 // within the initialization without creating a vicious cycle, such as:
2562 // { Foo p = new Foo(); p.next = p; }
2563 // True for constants and parameters and small combinations thereof.
2564 bool InitializeNode::detect_init_independence(Node* n,
2565 bool st_is_pinned,
2566 int& count) {
2567 if (n == NULL) return true; // (can this really happen?)
2568 if (n->is_Proj()) n = n->in(0);
2569 if (n == this) return false; // found a cycle
2570 if (n->is_Con()) return true;
2571 if (n->is_Start()) return true; // params, etc., are OK
2572 if (n->is_Root()) return true; // even better
2574 Node* ctl = n->in(0);
2575 if (ctl != NULL && !ctl->is_top()) {
2576 if (ctl->is_Proj()) ctl = ctl->in(0);
2577 if (ctl == this) return false;
2579 // If we already know that the enclosing memory op is pinned right after
2580 // the init, then any control flow that the store has picked up
2581 // must have preceded the init, or else be equal to the init.
2582 // Even after loop optimizations (which might change control edges)
2583 // a store is never pinned *before* the availability of its inputs.
2584 if (!MemNode::all_controls_dominate(n, this))
2585 return false; // failed to prove a good control
2587 }
2589 // Check data edges for possible dependencies on 'this'.
2590 if ((count += 1) > 20) return false; // complexity limit
2591 for (uint i = 1; i < n->req(); i++) {
2592 Node* m = n->in(i);
2593 if (m == NULL || m == n || m->is_top()) continue;
2594 uint first_i = n->find_edge(m);
2595 if (i != first_i) continue; // process duplicate edge just once
2596 if (!detect_init_independence(m, st_is_pinned, count)) {
2597 return false;
2598 }
2599 }
2601 return true;
2602 }
2604 // Here are all the checks a Store must pass before it can be moved into
2605 // an initialization. Returns zero if a check fails.
2606 // On success, returns the (constant) offset to which the store applies,
2607 // within the initialized memory.
2608 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2609 const int FAIL = 0;
2610 if (st->req() != MemNode::ValueIn + 1)
2611 return FAIL; // an inscrutable StoreNode (card mark?)
2612 Node* ctl = st->in(MemNode::Control);
2613 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2614 return FAIL; // must be unconditional after the initialization
2615 Node* mem = st->in(MemNode::Memory);
2616 if (!(mem->is_Proj() && mem->in(0) == this))
2617 return FAIL; // must not be preceded by other stores
2618 Node* adr = st->in(MemNode::Address);
2619 intptr_t offset;
2620 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2621 if (alloc == NULL)
2622 return FAIL; // inscrutable address
2623 if (alloc != allocation())
2624 return FAIL; // wrong allocation! (store needs to float up)
2625 Node* val = st->in(MemNode::ValueIn);
2626 int complexity_count = 0;
2627 if (!detect_init_independence(val, true, complexity_count))
2628 return FAIL; // stored value must be 'simple enough'
2630 return offset; // success
2631 }
2633 // Find the captured store in(i) which corresponds to the range
2634 // [start..start+size) in the initialized object.
2635 // If there is one, return its index i. If there isn't, return the
2636 // negative of the index where it should be inserted.
2637 // Return 0 if the queried range overlaps an initialization boundary
2638 // or if dead code is encountered.
2639 // If size_in_bytes is zero, do not bother with overlap checks.
2640 int InitializeNode::captured_store_insertion_point(intptr_t start,
2641 int size_in_bytes,
2642 PhaseTransform* phase) {
2643 const int FAIL = 0, MAX_STORE = BytesPerLong;
2645 if (is_complete())
2646 return FAIL; // arraycopy got here first; punt
2648 assert(allocation() != NULL, "must be present");
2650 // no negatives, no header fields:
2651 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2653 // after a certain size, we bail out on tracking all the stores:
2654 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2655 if (start >= ti_limit) return FAIL;
2657 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2658 if (i >= limit) return -(int)i; // not found; here is where to put it
2660 Node* st = in(i);
2661 intptr_t st_off = get_store_offset(st, phase);
2662 if (st_off < 0) {
2663 if (st != zero_memory()) {
2664 return FAIL; // bail out if there is dead garbage
2665 }
2666 } else if (st_off > start) {
2667 // ...we are done, since stores are ordered
2668 if (st_off < start + size_in_bytes) {
2669 return FAIL; // the next store overlaps
2670 }
2671 return -(int)i; // not found; here is where to put it
2672 } else if (st_off < start) {
2673 if (size_in_bytes != 0 &&
2674 start < st_off + MAX_STORE &&
2675 start < st_off + st->as_Store()->memory_size()) {
2676 return FAIL; // the previous store overlaps
2677 }
2678 } else {
2679 if (size_in_bytes != 0 &&
2680 st->as_Store()->memory_size() != size_in_bytes) {
2681 return FAIL; // mismatched store size
2682 }
2683 return i;
2684 }
2686 ++i;
2687 }
2688 }
2690 // Look for a captured store which initializes at the offset 'start'
2691 // with the given size. If there is no such store, and no other
2692 // initialization interferes, then return zero_memory (the memory
2693 // projection of the AllocateNode).
2694 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2695 PhaseTransform* phase) {
2696 assert(stores_are_sane(phase), "");
2697 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2698 if (i == 0) {
2699 return NULL; // something is dead
2700 } else if (i < 0) {
2701 return zero_memory(); // just primordial zero bits here
2702 } else {
2703 Node* st = in(i); // here is the store at this position
2704 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2705 return st;
2706 }
2707 }
2709 // Create, as a raw pointer, an address within my new object at 'offset'.
2710 Node* InitializeNode::make_raw_address(intptr_t offset,
2711 PhaseTransform* phase) {
2712 Node* addr = in(RawAddress);
2713 if (offset != 0) {
2714 Compile* C = phase->C;
2715 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2716 phase->MakeConX(offset)) );
2717 }
2718 return addr;
2719 }
2721 // Clone the given store, converting it into a raw store
2722 // initializing a field or element of my new object.
2723 // Caller is responsible for retiring the original store,
2724 // with subsume_node or the like.
2725 //
2726 // From the example above InitializeNode::InitializeNode,
2727 // here are the old stores to be captured:
2728 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2729 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2730 //
2731 // Here is the changed code; note the extra edges on init:
2732 // alloc = (Allocate ...)
2733 // rawoop = alloc.RawAddress
2734 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2735 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2736 // init = (Initialize alloc.Control alloc.Memory rawoop
2737 // rawstore1 rawstore2)
2738 //
2739 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2740 PhaseTransform* phase) {
2741 assert(stores_are_sane(phase), "");
2743 if (start < 0) return NULL;
2744 assert(can_capture_store(st, phase) == start, "sanity");
2746 Compile* C = phase->C;
2747 int size_in_bytes = st->memory_size();
2748 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2749 if (i == 0) return NULL; // bail out
2750 Node* prev_mem = NULL; // raw memory for the captured store
2751 if (i > 0) {
2752 prev_mem = in(i); // there is a pre-existing store under this one
2753 set_req(i, C->top()); // temporarily disconnect it
2754 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2755 } else {
2756 i = -i; // no pre-existing store
2757 prev_mem = zero_memory(); // a slice of the newly allocated object
2758 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2759 set_req(--i, C->top()); // reuse this edge; it has been folded away
2760 else
2761 ins_req(i, C->top()); // build a new edge
2762 }
2763 Node* new_st = st->clone();
2764 new_st->set_req(MemNode::Control, in(Control));
2765 new_st->set_req(MemNode::Memory, prev_mem);
2766 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2767 new_st = phase->transform(new_st);
2769 // At this point, new_st might have swallowed a pre-existing store
2770 // at the same offset, or perhaps new_st might have disappeared,
2771 // if it redundantly stored the same value (or zero to fresh memory).
2773 // In any case, wire it in:
2774 set_req(i, new_st);
2776 // The caller may now kill the old guy.
2777 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2778 assert(check_st == new_st || check_st == NULL, "must be findable");
2779 assert(!is_complete(), "");
2780 return new_st;
2781 }
2783 static bool store_constant(jlong* tiles, int num_tiles,
2784 intptr_t st_off, int st_size,
2785 jlong con) {
2786 if ((st_off & (st_size-1)) != 0)
2787 return false; // strange store offset (assume size==2**N)
2788 address addr = (address)tiles + st_off;
2789 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2790 switch (st_size) {
2791 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2792 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2793 case sizeof(jint): *(jint*) addr = (jint) con; break;
2794 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2795 default: return false; // strange store size (detect size!=2**N here)
2796 }
2797 return true; // return success to caller
2798 }
2800 // Coalesce subword constants into int constants and possibly
2801 // into long constants. The goal, if the CPU permits,
2802 // is to initialize the object with a small number of 64-bit tiles.
2803 // Also, convert floating-point constants to bit patterns.
2804 // Non-constants are not relevant to this pass.
2805 //
2806 // In terms of the running example on InitializeNode::InitializeNode
2807 // and InitializeNode::capture_store, here is the transformation
2808 // of rawstore1 and rawstore2 into rawstore12:
2809 // alloc = (Allocate ...)
2810 // rawoop = alloc.RawAddress
2811 // tile12 = 0x00010002
2812 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2813 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2814 //
2815 void
2816 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2817 Node* size_in_bytes,
2818 PhaseGVN* phase) {
2819 Compile* C = phase->C;
2821 assert(stores_are_sane(phase), "");
2822 // Note: After this pass, they are not completely sane,
2823 // since there may be some overlaps.
2825 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2827 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2828 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2829 size_limit = MIN2(size_limit, ti_limit);
2830 size_limit = align_size_up(size_limit, BytesPerLong);
2831 int num_tiles = size_limit / BytesPerLong;
2833 // allocate space for the tile map:
2834 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2835 jlong tiles_buf[small_len];
2836 Node* nodes_buf[small_len];
2837 jlong inits_buf[small_len];
2838 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2839 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2840 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2841 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2842 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2843 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2844 // tiles: exact bitwise model of all primitive constants
2845 // nodes: last constant-storing node subsumed into the tiles model
2846 // inits: which bytes (in each tile) are touched by any initializations
2848 //// Pass A: Fill in the tile model with any relevant stores.
2850 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2851 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2852 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2853 Node* zmem = zero_memory(); // initially zero memory state
2854 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2855 Node* st = in(i);
2856 intptr_t st_off = get_store_offset(st, phase);
2858 // Figure out the store's offset and constant value:
2859 if (st_off < header_size) continue; //skip (ignore header)
2860 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
2861 int st_size = st->as_Store()->memory_size();
2862 if (st_off + st_size > size_limit) break;
2864 // Record which bytes are touched, whether by constant or not.
2865 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
2866 continue; // skip (strange store size)
2868 const Type* val = phase->type(st->in(MemNode::ValueIn));
2869 if (!val->singleton()) continue; //skip (non-con store)
2870 BasicType type = val->basic_type();
2872 jlong con = 0;
2873 switch (type) {
2874 case T_INT: con = val->is_int()->get_con(); break;
2875 case T_LONG: con = val->is_long()->get_con(); break;
2876 case T_FLOAT: con = jint_cast(val->getf()); break;
2877 case T_DOUBLE: con = jlong_cast(val->getd()); break;
2878 default: continue; //skip (odd store type)
2879 }
2881 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
2882 st->Opcode() == Op_StoreL) {
2883 continue; // This StoreL is already optimal.
2884 }
2886 // Store down the constant.
2887 store_constant(tiles, num_tiles, st_off, st_size, con);
2889 intptr_t j = st_off >> LogBytesPerLong;
2891 if (type == T_INT && st_size == BytesPerInt
2892 && (st_off & BytesPerInt) == BytesPerInt) {
2893 jlong lcon = tiles[j];
2894 if (!Matcher::isSimpleConstant64(lcon) &&
2895 st->Opcode() == Op_StoreI) {
2896 // This StoreI is already optimal by itself.
2897 jint* intcon = (jint*) &tiles[j];
2898 intcon[1] = 0; // undo the store_constant()
2900 // If the previous store is also optimal by itself, back up and
2901 // undo the action of the previous loop iteration... if we can.
2902 // But if we can't, just let the previous half take care of itself.
2903 st = nodes[j];
2904 st_off -= BytesPerInt;
2905 con = intcon[0];
2906 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
2907 assert(st_off >= header_size, "still ignoring header");
2908 assert(get_store_offset(st, phase) == st_off, "must be");
2909 assert(in(i-1) == zmem, "must be");
2910 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
2911 assert(con == tcon->is_int()->get_con(), "must be");
2912 // Undo the effects of the previous loop trip, which swallowed st:
2913 intcon[0] = 0; // undo store_constant()
2914 set_req(i-1, st); // undo set_req(i, zmem)
2915 nodes[j] = NULL; // undo nodes[j] = st
2916 --old_subword; // undo ++old_subword
2917 }
2918 continue; // This StoreI is already optimal.
2919 }
2920 }
2922 // This store is not needed.
2923 set_req(i, zmem);
2924 nodes[j] = st; // record for the moment
2925 if (st_size < BytesPerLong) // something has changed
2926 ++old_subword; // includes int/float, but who's counting...
2927 else ++old_long;
2928 }
2930 if ((old_subword + old_long) == 0)
2931 return; // nothing more to do
2933 //// Pass B: Convert any non-zero tiles into optimal constant stores.
2934 // Be sure to insert them before overlapping non-constant stores.
2935 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
2936 for (int j = 0; j < num_tiles; j++) {
2937 jlong con = tiles[j];
2938 jlong init = inits[j];
2939 if (con == 0) continue;
2940 jint con0, con1; // split the constant, address-wise
2941 jint init0, init1; // split the init map, address-wise
2942 { union { jlong con; jint intcon[2]; } u;
2943 u.con = con;
2944 con0 = u.intcon[0];
2945 con1 = u.intcon[1];
2946 u.con = init;
2947 init0 = u.intcon[0];
2948 init1 = u.intcon[1];
2949 }
2951 Node* old = nodes[j];
2952 assert(old != NULL, "need the prior store");
2953 intptr_t offset = (j * BytesPerLong);
2955 bool split = !Matcher::isSimpleConstant64(con);
2957 if (offset < header_size) {
2958 assert(offset + BytesPerInt >= header_size, "second int counts");
2959 assert(*(jint*)&tiles[j] == 0, "junk in header");
2960 split = true; // only the second word counts
2961 // Example: int a[] = { 42 ... }
2962 } else if (con0 == 0 && init0 == -1) {
2963 split = true; // first word is covered by full inits
2964 // Example: int a[] = { ... foo(), 42 ... }
2965 } else if (con1 == 0 && init1 == -1) {
2966 split = true; // second word is covered by full inits
2967 // Example: int a[] = { ... 42, foo() ... }
2968 }
2970 // Here's a case where init0 is neither 0 nor -1:
2971 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
2972 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
2973 // In this case the tile is not split; it is (jlong)42.
2974 // The big tile is stored down, and then the foo() value is inserted.
2975 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
2977 Node* ctl = old->in(MemNode::Control);
2978 Node* adr = make_raw_address(offset, phase);
2979 const TypePtr* atp = TypeRawPtr::BOTTOM;
2981 // One or two coalesced stores to plop down.
2982 Node* st[2];
2983 intptr_t off[2];
2984 int nst = 0;
2985 if (!split) {
2986 ++new_long;
2987 off[nst] = offset;
2988 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
2989 phase->longcon(con), T_LONG);
2990 } else {
2991 // Omit either if it is a zero.
2992 if (con0 != 0) {
2993 ++new_int;
2994 off[nst] = offset;
2995 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
2996 phase->intcon(con0), T_INT);
2997 }
2998 if (con1 != 0) {
2999 ++new_int;
3000 offset += BytesPerInt;
3001 adr = make_raw_address(offset, phase);
3002 off[nst] = offset;
3003 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3004 phase->intcon(con1), T_INT);
3005 }
3006 }
3008 // Insert second store first, then the first before the second.
3009 // Insert each one just before any overlapping non-constant stores.
3010 while (nst > 0) {
3011 Node* st1 = st[--nst];
3012 C->copy_node_notes_to(st1, old);
3013 st1 = phase->transform(st1);
3014 offset = off[nst];
3015 assert(offset >= header_size, "do not smash header");
3016 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3017 guarantee(ins_idx != 0, "must re-insert constant store");
3018 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3019 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3020 set_req(--ins_idx, st1);
3021 else
3022 ins_req(ins_idx, st1);
3023 }
3024 }
3026 if (PrintCompilation && WizardMode)
3027 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3028 old_subword, old_long, new_int, new_long);
3029 if (C->log() != NULL)
3030 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3031 old_subword, old_long, new_int, new_long);
3033 // Clean up any remaining occurrences of zmem:
3034 remove_extra_zeroes();
3035 }
3037 // Explore forward from in(start) to find the first fully initialized
3038 // word, and return its offset. Skip groups of subword stores which
3039 // together initialize full words. If in(start) is itself part of a
3040 // fully initialized word, return the offset of in(start). If there
3041 // are no following full-word stores, or if something is fishy, return
3042 // a negative value.
3043 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3044 int int_map = 0;
3045 intptr_t int_map_off = 0;
3046 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3048 for (uint i = start, limit = req(); i < limit; i++) {
3049 Node* st = in(i);
3051 intptr_t st_off = get_store_offset(st, phase);
3052 if (st_off < 0) break; // return conservative answer
3054 int st_size = st->as_Store()->memory_size();
3055 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3056 return st_off; // we found a complete word init
3057 }
3059 // update the map:
3061 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3062 if (this_int_off != int_map_off) {
3063 // reset the map:
3064 int_map = 0;
3065 int_map_off = this_int_off;
3066 }
3068 int subword_off = st_off - this_int_off;
3069 int_map |= right_n_bits(st_size) << subword_off;
3070 if ((int_map & FULL_MAP) == FULL_MAP) {
3071 return this_int_off; // we found a complete word init
3072 }
3074 // Did this store hit or cross the word boundary?
3075 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3076 if (next_int_off == this_int_off + BytesPerInt) {
3077 // We passed the current int, without fully initializing it.
3078 int_map_off = next_int_off;
3079 int_map >>= BytesPerInt;
3080 } else if (next_int_off > this_int_off + BytesPerInt) {
3081 // We passed the current and next int.
3082 return this_int_off + BytesPerInt;
3083 }
3084 }
3086 return -1;
3087 }
3090 // Called when the associated AllocateNode is expanded into CFG.
3091 // At this point, we may perform additional optimizations.
3092 // Linearize the stores by ascending offset, to make memory
3093 // activity as coherent as possible.
3094 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3095 intptr_t header_size,
3096 Node* size_in_bytes,
3097 PhaseGVN* phase) {
3098 assert(!is_complete(), "not already complete");
3099 assert(stores_are_sane(phase), "");
3100 assert(allocation() != NULL, "must be present");
3102 remove_extra_zeroes();
3104 if (ReduceFieldZeroing || ReduceBulkZeroing)
3105 // reduce instruction count for common initialization patterns
3106 coalesce_subword_stores(header_size, size_in_bytes, phase);
3108 Node* zmem = zero_memory(); // initially zero memory state
3109 Node* inits = zmem; // accumulating a linearized chain of inits
3110 #ifdef ASSERT
3111 intptr_t first_offset = allocation()->minimum_header_size();
3112 intptr_t last_init_off = first_offset; // previous init offset
3113 intptr_t last_init_end = first_offset; // previous init offset+size
3114 intptr_t last_tile_end = first_offset; // previous tile offset+size
3115 #endif
3116 intptr_t zeroes_done = header_size;
3118 bool do_zeroing = true; // we might give up if inits are very sparse
3119 int big_init_gaps = 0; // how many large gaps have we seen?
3121 if (ZeroTLAB) do_zeroing = false;
3122 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3124 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3125 Node* st = in(i);
3126 intptr_t st_off = get_store_offset(st, phase);
3127 if (st_off < 0)
3128 break; // unknown junk in the inits
3129 if (st->in(MemNode::Memory) != zmem)
3130 break; // complicated store chains somehow in list
3132 int st_size = st->as_Store()->memory_size();
3133 intptr_t next_init_off = st_off + st_size;
3135 if (do_zeroing && zeroes_done < next_init_off) {
3136 // See if this store needs a zero before it or under it.
3137 intptr_t zeroes_needed = st_off;
3139 if (st_size < BytesPerInt) {
3140 // Look for subword stores which only partially initialize words.
3141 // If we find some, we must lay down some word-level zeroes first,
3142 // underneath the subword stores.
3143 //
3144 // Examples:
3145 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3146 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3147 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3148 //
3149 // Note: coalesce_subword_stores may have already done this,
3150 // if it was prompted by constant non-zero subword initializers.
3151 // But this case can still arise with non-constant stores.
3153 intptr_t next_full_store = find_next_fullword_store(i, phase);
3155 // In the examples above:
3156 // in(i) p q r s x y z
3157 // st_off 12 13 14 15 12 13 14
3158 // st_size 1 1 1 1 1 1 1
3159 // next_full_s. 12 16 16 16 16 16 16
3160 // z's_done 12 16 16 16 12 16 12
3161 // z's_needed 12 16 16 16 16 16 16
3162 // zsize 0 0 0 0 4 0 4
3163 if (next_full_store < 0) {
3164 // Conservative tack: Zero to end of current word.
3165 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3166 } else {
3167 // Zero to beginning of next fully initialized word.
3168 // Or, don't zero at all, if we are already in that word.
3169 assert(next_full_store >= zeroes_needed, "must go forward");
3170 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3171 zeroes_needed = next_full_store;
3172 }
3173 }
3175 if (zeroes_needed > zeroes_done) {
3176 intptr_t zsize = zeroes_needed - zeroes_done;
3177 // Do some incremental zeroing on rawmem, in parallel with inits.
3178 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3179 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3180 zeroes_done, zeroes_needed,
3181 phase);
3182 zeroes_done = zeroes_needed;
3183 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3184 do_zeroing = false; // leave the hole, next time
3185 }
3186 }
3188 // Collect the store and move on:
3189 st->set_req(MemNode::Memory, inits);
3190 inits = st; // put it on the linearized chain
3191 set_req(i, zmem); // unhook from previous position
3193 if (zeroes_done == st_off)
3194 zeroes_done = next_init_off;
3196 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3198 #ifdef ASSERT
3199 // Various order invariants. Weaker than stores_are_sane because
3200 // a large constant tile can be filled in by smaller non-constant stores.
3201 assert(st_off >= last_init_off, "inits do not reverse");
3202 last_init_off = st_off;
3203 const Type* val = NULL;
3204 if (st_size >= BytesPerInt &&
3205 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3206 (int)val->basic_type() < (int)T_OBJECT) {
3207 assert(st_off >= last_tile_end, "tiles do not overlap");
3208 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3209 last_tile_end = MAX2(last_tile_end, next_init_off);
3210 } else {
3211 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3212 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3213 assert(st_off >= last_init_end, "inits do not overlap");
3214 last_init_end = next_init_off; // it's a non-tile
3215 }
3216 #endif //ASSERT
3217 }
3219 remove_extra_zeroes(); // clear out all the zmems left over
3220 add_req(inits);
3222 if (!ZeroTLAB) {
3223 // If anything remains to be zeroed, zero it all now.
3224 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3225 // if it is the last unused 4 bytes of an instance, forget about it
3226 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3227 if (zeroes_done + BytesPerLong >= size_limit) {
3228 assert(allocation() != NULL, "");
3229 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3230 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3231 if (zeroes_done == k->layout_helper())
3232 zeroes_done = size_limit;
3233 }
3234 if (zeroes_done < size_limit) {
3235 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3236 zeroes_done, size_in_bytes, phase);
3237 }
3238 }
3240 set_complete(phase);
3241 return rawmem;
3242 }
3245 #ifdef ASSERT
3246 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3247 if (is_complete())
3248 return true; // stores could be anything at this point
3249 assert(allocation() != NULL, "must be present");
3250 intptr_t last_off = allocation()->minimum_header_size();
3251 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3252 Node* st = in(i);
3253 intptr_t st_off = get_store_offset(st, phase);
3254 if (st_off < 0) continue; // ignore dead garbage
3255 if (last_off > st_off) {
3256 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3257 this->dump(2);
3258 assert(false, "ascending store offsets");
3259 return false;
3260 }
3261 last_off = st_off + st->as_Store()->memory_size();
3262 }
3263 return true;
3264 }
3265 #endif //ASSERT
3270 //============================MergeMemNode=====================================
3271 //
3272 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3273 // contributing store or call operations. Each contributor provides the memory
3274 // state for a particular "alias type" (see Compile::alias_type). For example,
3275 // if a MergeMem has an input X for alias category #6, then any memory reference
3276 // to alias category #6 may use X as its memory state input, as an exact equivalent
3277 // to using the MergeMem as a whole.
3278 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3279 //
3280 // (Here, the <N> notation gives the index of the relevant adr_type.)
3281 //
3282 // In one special case (and more cases in the future), alias categories overlap.
3283 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3284 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3285 // it is exactly equivalent to that state W:
3286 // MergeMem(<Bot>: W) <==> W
3287 //
3288 // Usually, the merge has more than one input. In that case, where inputs
3289 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3290 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3291 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3292 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3293 //
3294 // A merge can take a "wide" memory state as one of its narrow inputs.
3295 // This simply means that the merge observes out only the relevant parts of
3296 // the wide input. That is, wide memory states arriving at narrow merge inputs
3297 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3298 //
3299 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3300 // and that memory slices "leak through":
3301 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3302 //
3303 // But, in such a cascade, repeated memory slices can "block the leak":
3304 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3305 //
3306 // In the last example, Y is not part of the combined memory state of the
3307 // outermost MergeMem. The system must, of course, prevent unschedulable
3308 // memory states from arising, so you can be sure that the state Y is somehow
3309 // a precursor to state Y'.
3310 //
3311 //
3312 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3313 // of each MergeMemNode array are exactly the numerical alias indexes, including
3314 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3315 // Compile::alias_type (and kin) produce and manage these indexes.
3316 //
3317 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3318 // (Note that this provides quick access to the top node inside MergeMem methods,
3319 // without the need to reach out via TLS to Compile::current.)
3320 //
3321 // As a consequence of what was just described, a MergeMem that represents a full
3322 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3323 // containing all alias categories.
3324 //
3325 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3326 //
3327 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3328 // a memory state for the alias type <N>, or else the top node, meaning that
3329 // there is no particular input for that alias type. Note that the length of
3330 // a MergeMem is variable, and may be extended at any time to accommodate new
3331 // memory states at larger alias indexes. When merges grow, they are of course
3332 // filled with "top" in the unused in() positions.
3333 //
3334 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3335 // (Top was chosen because it works smoothly with passes like GCM.)
3336 //
3337 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3338 // the type of random VM bits like TLS references.) Since it is always the
3339 // first non-Bot memory slice, some low-level loops use it to initialize an
3340 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3341 //
3342 //
3343 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3344 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3345 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3346 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3347 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3348 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3349 //
3350 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3351 // really that different from the other memory inputs. An abbreviation called
3352 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3353 //
3354 //
3355 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3356 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3357 // that "emerges though" the base memory will be marked as excluding the alias types
3358 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3359 //
3360 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3361 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3362 //
3363 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3364 // (It is currently unimplemented.) As you can see, the resulting merge is
3365 // actually a disjoint union of memory states, rather than an overlay.
3366 //
3368 //------------------------------MergeMemNode-----------------------------------
3369 Node* MergeMemNode::make_empty_memory() {
3370 Node* empty_memory = (Node*) Compile::current()->top();
3371 assert(empty_memory->is_top(), "correct sentinel identity");
3372 return empty_memory;
3373 }
3375 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3376 init_class_id(Class_MergeMem);
3377 // all inputs are nullified in Node::Node(int)
3378 // set_input(0, NULL); // no control input
3380 // Initialize the edges uniformly to top, for starters.
3381 Node* empty_mem = make_empty_memory();
3382 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3383 init_req(i,empty_mem);
3384 }
3385 assert(empty_memory() == empty_mem, "");
3387 if( new_base != NULL && new_base->is_MergeMem() ) {
3388 MergeMemNode* mdef = new_base->as_MergeMem();
3389 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3390 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3391 mms.set_memory(mms.memory2());
3392 }
3393 assert(base_memory() == mdef->base_memory(), "");
3394 } else {
3395 set_base_memory(new_base);
3396 }
3397 }
3399 // Make a new, untransformed MergeMem with the same base as 'mem'.
3400 // If mem is itself a MergeMem, populate the result with the same edges.
3401 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3402 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3403 }
3405 //------------------------------cmp--------------------------------------------
3406 uint MergeMemNode::hash() const { return NO_HASH; }
3407 uint MergeMemNode::cmp( const Node &n ) const {
3408 return (&n == this); // Always fail except on self
3409 }
3411 //------------------------------Identity---------------------------------------
3412 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3413 // Identity if this merge point does not record any interesting memory
3414 // disambiguations.
3415 Node* base_mem = base_memory();
3416 Node* empty_mem = empty_memory();
3417 if (base_mem != empty_mem) { // Memory path is not dead?
3418 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3419 Node* mem = in(i);
3420 if (mem != empty_mem && mem != base_mem) {
3421 return this; // Many memory splits; no change
3422 }
3423 }
3424 }
3425 return base_mem; // No memory splits; ID on the one true input
3426 }
3428 //------------------------------Ideal------------------------------------------
3429 // This method is invoked recursively on chains of MergeMem nodes
3430 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3431 // Remove chain'd MergeMems
3432 //
3433 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3434 // relative to the "in(Bot)". Since we are patching both at the same time,
3435 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3436 // but rewrite each "in(i)" relative to the new "in(Bot)".
3437 Node *progress = NULL;
3440 Node* old_base = base_memory();
3441 Node* empty_mem = empty_memory();
3442 if (old_base == empty_mem)
3443 return NULL; // Dead memory path.
3445 MergeMemNode* old_mbase;
3446 if (old_base != NULL && old_base->is_MergeMem())
3447 old_mbase = old_base->as_MergeMem();
3448 else
3449 old_mbase = NULL;
3450 Node* new_base = old_base;
3452 // simplify stacked MergeMems in base memory
3453 if (old_mbase) new_base = old_mbase->base_memory();
3455 // the base memory might contribute new slices beyond my req()
3456 if (old_mbase) grow_to_match(old_mbase);
3458 // Look carefully at the base node if it is a phi.
3459 PhiNode* phi_base;
3460 if (new_base != NULL && new_base->is_Phi())
3461 phi_base = new_base->as_Phi();
3462 else
3463 phi_base = NULL;
3465 Node* phi_reg = NULL;
3466 uint phi_len = (uint)-1;
3467 if (phi_base != NULL && !phi_base->is_copy()) {
3468 // do not examine phi if degraded to a copy
3469 phi_reg = phi_base->region();
3470 phi_len = phi_base->req();
3471 // see if the phi is unfinished
3472 for (uint i = 1; i < phi_len; i++) {
3473 if (phi_base->in(i) == NULL) {
3474 // incomplete phi; do not look at it yet!
3475 phi_reg = NULL;
3476 phi_len = (uint)-1;
3477 break;
3478 }
3479 }
3480 }
3482 // Note: We do not call verify_sparse on entry, because inputs
3483 // can normalize to the base_memory via subsume_node or similar
3484 // mechanisms. This method repairs that damage.
3486 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3488 // Look at each slice.
3489 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3490 Node* old_in = in(i);
3491 // calculate the old memory value
3492 Node* old_mem = old_in;
3493 if (old_mem == empty_mem) old_mem = old_base;
3494 assert(old_mem == memory_at(i), "");
3496 // maybe update (reslice) the old memory value
3498 // simplify stacked MergeMems
3499 Node* new_mem = old_mem;
3500 MergeMemNode* old_mmem;
3501 if (old_mem != NULL && old_mem->is_MergeMem())
3502 old_mmem = old_mem->as_MergeMem();
3503 else
3504 old_mmem = NULL;
3505 if (old_mmem == this) {
3506 // This can happen if loops break up and safepoints disappear.
3507 // A merge of BotPtr (default) with a RawPtr memory derived from a
3508 // safepoint can be rewritten to a merge of the same BotPtr with
3509 // the BotPtr phi coming into the loop. If that phi disappears
3510 // also, we can end up with a self-loop of the mergemem.
3511 // In general, if loops degenerate and memory effects disappear,
3512 // a mergemem can be left looking at itself. This simply means
3513 // that the mergemem's default should be used, since there is
3514 // no longer any apparent effect on this slice.
3515 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3516 // from start. Update the input to TOP.
3517 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3518 }
3519 else if (old_mmem != NULL) {
3520 new_mem = old_mmem->memory_at(i);
3521 }
3522 // else preceeding memory was not a MergeMem
3524 // replace equivalent phis (unfortunately, they do not GVN together)
3525 if (new_mem != NULL && new_mem != new_base &&
3526 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3527 if (new_mem->is_Phi()) {
3528 PhiNode* phi_mem = new_mem->as_Phi();
3529 for (uint i = 1; i < phi_len; i++) {
3530 if (phi_base->in(i) != phi_mem->in(i)) {
3531 phi_mem = NULL;
3532 break;
3533 }
3534 }
3535 if (phi_mem != NULL) {
3536 // equivalent phi nodes; revert to the def
3537 new_mem = new_base;
3538 }
3539 }
3540 }
3542 // maybe store down a new value
3543 Node* new_in = new_mem;
3544 if (new_in == new_base) new_in = empty_mem;
3546 if (new_in != old_in) {
3547 // Warning: Do not combine this "if" with the previous "if"
3548 // A memory slice might have be be rewritten even if it is semantically
3549 // unchanged, if the base_memory value has changed.
3550 set_req(i, new_in);
3551 progress = this; // Report progress
3552 }
3553 }
3555 if (new_base != old_base) {
3556 set_req(Compile::AliasIdxBot, new_base);
3557 // Don't use set_base_memory(new_base), because we need to update du.
3558 assert(base_memory() == new_base, "");
3559 progress = this;
3560 }
3562 if( base_memory() == this ) {
3563 // a self cycle indicates this memory path is dead
3564 set_req(Compile::AliasIdxBot, empty_mem);
3565 }
3567 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3568 // Recursion must occur after the self cycle check above
3569 if( base_memory()->is_MergeMem() ) {
3570 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3571 Node *m = phase->transform(new_mbase); // Rollup any cycles
3572 if( m != NULL && (m->is_top() ||
3573 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3574 // propagate rollup of dead cycle to self
3575 set_req(Compile::AliasIdxBot, empty_mem);
3576 }
3577 }
3579 if( base_memory() == empty_mem ) {
3580 progress = this;
3581 // Cut inputs during Parse phase only.
3582 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3583 if( !can_reshape ) {
3584 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3585 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3586 }
3587 }
3588 }
3590 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3591 // Check if PhiNode::Ideal's "Split phis through memory merges"
3592 // transform should be attempted. Look for this->phi->this cycle.
3593 uint merge_width = req();
3594 if (merge_width > Compile::AliasIdxRaw) {
3595 PhiNode* phi = base_memory()->as_Phi();
3596 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3597 if (phi->in(i) == this) {
3598 phase->is_IterGVN()->_worklist.push(phi);
3599 break;
3600 }
3601 }
3602 }
3603 }
3605 assert(progress || verify_sparse(), "please, no dups of base");
3606 return progress;
3607 }
3609 //-------------------------set_base_memory-------------------------------------
3610 void MergeMemNode::set_base_memory(Node *new_base) {
3611 Node* empty_mem = empty_memory();
3612 set_req(Compile::AliasIdxBot, new_base);
3613 assert(memory_at(req()) == new_base, "must set default memory");
3614 // Clear out other occurrences of new_base:
3615 if (new_base != empty_mem) {
3616 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3617 if (in(i) == new_base) set_req(i, empty_mem);
3618 }
3619 }
3620 }
3622 //------------------------------out_RegMask------------------------------------
3623 const RegMask &MergeMemNode::out_RegMask() const {
3624 return RegMask::Empty;
3625 }
3627 //------------------------------dump_spec--------------------------------------
3628 #ifndef PRODUCT
3629 void MergeMemNode::dump_spec(outputStream *st) const {
3630 st->print(" {");
3631 Node* base_mem = base_memory();
3632 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3633 Node* mem = memory_at(i);
3634 if (mem == base_mem) { st->print(" -"); continue; }
3635 st->print( " N%d:", mem->_idx );
3636 Compile::current()->get_adr_type(i)->dump_on(st);
3637 }
3638 st->print(" }");
3639 }
3640 #endif // !PRODUCT
3643 #ifdef ASSERT
3644 static bool might_be_same(Node* a, Node* b) {
3645 if (a == b) return true;
3646 if (!(a->is_Phi() || b->is_Phi())) return false;
3647 // phis shift around during optimization
3648 return true; // pretty stupid...
3649 }
3651 // verify a narrow slice (either incoming or outgoing)
3652 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3653 if (!VerifyAliases) return; // don't bother to verify unless requested
3654 if (is_error_reported()) return; // muzzle asserts when debugging an error
3655 if (Node::in_dump()) return; // muzzle asserts when printing
3656 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3657 assert(n != NULL, "");
3658 // Elide intervening MergeMem's
3659 while (n->is_MergeMem()) {
3660 n = n->as_MergeMem()->memory_at(alias_idx);
3661 }
3662 Compile* C = Compile::current();
3663 const TypePtr* n_adr_type = n->adr_type();
3664 if (n == m->empty_memory()) {
3665 // Implicit copy of base_memory()
3666 } else if (n_adr_type != TypePtr::BOTTOM) {
3667 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3668 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3669 } else {
3670 // A few places like make_runtime_call "know" that VM calls are narrow,
3671 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3672 bool expected_wide_mem = false;
3673 if (n == m->base_memory()) {
3674 expected_wide_mem = true;
3675 } else if (alias_idx == Compile::AliasIdxRaw ||
3676 n == m->memory_at(Compile::AliasIdxRaw)) {
3677 expected_wide_mem = true;
3678 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3679 // memory can "leak through" calls on channels that
3680 // are write-once. Allow this also.
3681 expected_wide_mem = true;
3682 }
3683 assert(expected_wide_mem, "expected narrow slice replacement");
3684 }
3685 }
3686 #else // !ASSERT
3687 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3688 #endif
3691 //-----------------------------memory_at---------------------------------------
3692 Node* MergeMemNode::memory_at(uint alias_idx) const {
3693 assert(alias_idx >= Compile::AliasIdxRaw ||
3694 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3695 "must avoid base_memory and AliasIdxTop");
3697 // Otherwise, it is a narrow slice.
3698 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3699 Compile *C = Compile::current();
3700 if (is_empty_memory(n)) {
3701 // the array is sparse; empty slots are the "top" node
3702 n = base_memory();
3703 assert(Node::in_dump()
3704 || n == NULL || n->bottom_type() == Type::TOP
3705 || n->adr_type() == TypePtr::BOTTOM
3706 || n->adr_type() == TypeRawPtr::BOTTOM
3707 || Compile::current()->AliasLevel() == 0,
3708 "must be a wide memory");
3709 // AliasLevel == 0 if we are organizing the memory states manually.
3710 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3711 } else {
3712 // make sure the stored slice is sane
3713 #ifdef ASSERT
3714 if (is_error_reported() || Node::in_dump()) {
3715 } else if (might_be_same(n, base_memory())) {
3716 // Give it a pass: It is a mostly harmless repetition of the base.
3717 // This can arise normally from node subsumption during optimization.
3718 } else {
3719 verify_memory_slice(this, alias_idx, n);
3720 }
3721 #endif
3722 }
3723 return n;
3724 }
3726 //---------------------------set_memory_at-------------------------------------
3727 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3728 verify_memory_slice(this, alias_idx, n);
3729 Node* empty_mem = empty_memory();
3730 if (n == base_memory()) n = empty_mem; // collapse default
3731 uint need_req = alias_idx+1;
3732 if (req() < need_req) {
3733 if (n == empty_mem) return; // already the default, so do not grow me
3734 // grow the sparse array
3735 do {
3736 add_req(empty_mem);
3737 } while (req() < need_req);
3738 }
3739 set_req( alias_idx, n );
3740 }
3744 //--------------------------iteration_setup------------------------------------
3745 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3746 if (other != NULL) {
3747 grow_to_match(other);
3748 // invariant: the finite support of mm2 is within mm->req()
3749 #ifdef ASSERT
3750 for (uint i = req(); i < other->req(); i++) {
3751 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3752 }
3753 #endif
3754 }
3755 // Replace spurious copies of base_memory by top.
3756 Node* base_mem = base_memory();
3757 if (base_mem != NULL && !base_mem->is_top()) {
3758 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3759 if (in(i) == base_mem)
3760 set_req(i, empty_memory());
3761 }
3762 }
3763 }
3765 //---------------------------grow_to_match-------------------------------------
3766 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3767 Node* empty_mem = empty_memory();
3768 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3769 // look for the finite support of the other memory
3770 for (uint i = other->req(); --i >= req(); ) {
3771 if (other->in(i) != empty_mem) {
3772 uint new_len = i+1;
3773 while (req() < new_len) add_req(empty_mem);
3774 break;
3775 }
3776 }
3777 }
3779 //---------------------------verify_sparse-------------------------------------
3780 #ifndef PRODUCT
3781 bool MergeMemNode::verify_sparse() const {
3782 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3783 Node* base_mem = base_memory();
3784 // The following can happen in degenerate cases, since empty==top.
3785 if (is_empty_memory(base_mem)) return true;
3786 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3787 assert(in(i) != NULL, "sane slice");
3788 if (in(i) == base_mem) return false; // should have been the sentinel value!
3789 }
3790 return true;
3791 }
3793 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3794 Node* n;
3795 n = mm->in(idx);
3796 if (mem == n) return true; // might be empty_memory()
3797 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3798 if (mem == n) return true;
3799 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3800 if (mem == n) return true;
3801 if (n == NULL) break;
3802 }
3803 return false;
3804 }
3805 #endif // !PRODUCT