Wed, 21 May 2008 10:45:07 -0700
6695810: null oop passed to encode_heap_oop_not_null
Summary: fix several problems in C2 related to Escape Analysis and Compressed Oops.
Reviewed-by: never, jrose
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 t->isa_oopptr() && !t->is_oopptr()->is_instance() &&
138 t->is_oopptr()->cast_to_instance(t_oop->instance_id()) == t_oop) {
139 // clone the Phi with our address type
140 result = mphi->split_out_instance(t_adr, igvn);
141 } else {
142 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
143 }
144 }
145 return result;
146 }
148 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
149 uint alias_idx = phase->C->get_alias_index(tp);
150 Node *mem = mmem;
151 #ifdef ASSERT
152 {
153 // Check that current type is consistent with the alias index used during graph construction
154 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
155 bool consistent = adr_check == NULL || adr_check->empty() ||
156 phase->C->must_alias(adr_check, alias_idx );
157 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
158 if( !consistent && adr_check != NULL && !adr_check->empty() &&
159 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
160 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
161 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
162 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
163 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
164 // don't assert if it is dead code.
165 consistent = true;
166 }
167 if( !consistent ) {
168 st->print("alias_idx==%d, adr_check==", alias_idx);
169 if( adr_check == NULL ) {
170 st->print("NULL");
171 } else {
172 adr_check->dump();
173 }
174 st->cr();
175 print_alias_types();
176 assert(consistent, "adr_check must match alias idx");
177 }
178 }
179 #endif
180 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
181 // means an array I have not precisely typed yet. Do not do any
182 // alias stuff with it any time soon.
183 const TypeOopPtr *tinst = tp->isa_oopptr();
184 if( tp->base() != Type::AnyPtr &&
185 !(tinst &&
186 tinst->klass()->is_java_lang_Object() &&
187 tinst->offset() == Type::OffsetBot) ) {
188 // compress paths and change unreachable cycles to TOP
189 // If not, we can update the input infinitely along a MergeMem cycle
190 // Equivalent code in PhiNode::Ideal
191 Node* m = phase->transform(mmem);
192 // If tranformed to a MergeMem, get the desired slice
193 // Otherwise the returned node represents memory for every slice
194 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
195 // Update input if it is progress over what we have now
196 }
197 return mem;
198 }
200 //--------------------------Ideal_common---------------------------------------
201 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
202 // Unhook non-raw memories from complete (macro-expanded) initializations.
203 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
204 // If our control input is a dead region, kill all below the region
205 Node *ctl = in(MemNode::Control);
206 if (ctl && remove_dead_region(phase, can_reshape))
207 return this;
209 // Ignore if memory is dead, or self-loop
210 Node *mem = in(MemNode::Memory);
211 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
212 assert( mem != this, "dead loop in MemNode::Ideal" );
214 Node *address = in(MemNode::Address);
215 const Type *t_adr = phase->type( address );
216 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
218 // Avoid independent memory operations
219 Node* old_mem = mem;
221 // The code which unhooks non-raw memories from complete (macro-expanded)
222 // initializations was removed. After macro-expansion all stores catched
223 // by Initialize node became raw stores and there is no information
224 // which memory slices they modify. So it is unsafe to move any memory
225 // operation above these stores. Also in most cases hooked non-raw memories
226 // were already unhooked by using information from detect_ptr_independence()
227 // and find_previous_store().
229 if (mem->is_MergeMem()) {
230 MergeMemNode* mmem = mem->as_MergeMem();
231 const TypePtr *tp = t_adr->is_ptr();
233 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
234 }
236 if (mem != old_mem) {
237 set_req(MemNode::Memory, mem);
238 return this;
239 }
241 // let the subclass continue analyzing...
242 return NULL;
243 }
245 // Helper function for proving some simple control dominations.
246 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
247 // Already assumes that 'dom' is available at 'sub', and that 'sub'
248 // is not a constant (dominated by the method's StartNode).
249 // Used by MemNode::find_previous_store to prove that the
250 // control input of a memory operation predates (dominates)
251 // an allocation it wants to look past.
252 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
253 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
254 return false; // Conservative answer for dead code
256 // Check 'dom'.
257 dom = dom->find_exact_control(dom);
258 if (dom == NULL || dom->is_top())
259 return false; // Conservative answer for dead code
261 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
262 return true;
264 // 'dom' dominates 'sub' if its control edge and control edges
265 // of all its inputs dominate or equal to sub's control edge.
267 // Currently 'sub' is either Allocate, Initialize or Start nodes.
268 // Or Region for the check in LoadNode::Ideal();
269 // 'sub' should have sub->in(0) != NULL.
270 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
271 sub->is_Region(), "expecting only these nodes");
273 // Get control edge of 'sub'.
274 sub = sub->find_exact_control(sub->in(0));
275 if (sub == NULL || sub->is_top())
276 return false; // Conservative answer for dead code
278 assert(sub->is_CFG(), "expecting control");
280 if (sub == dom)
281 return true;
283 if (sub->is_Start() || sub->is_Root())
284 return false;
286 {
287 // Check all control edges of 'dom'.
289 ResourceMark rm;
290 Arena* arena = Thread::current()->resource_area();
291 Node_List nlist(arena);
292 Unique_Node_List dom_list(arena);
294 dom_list.push(dom);
295 bool only_dominating_controls = false;
297 for (uint next = 0; next < dom_list.size(); next++) {
298 Node* n = dom_list.at(next);
299 if (!n->is_CFG() && n->pinned()) {
300 // Check only own control edge for pinned non-control nodes.
301 n = n->find_exact_control(n->in(0));
302 if (n == NULL || n->is_top())
303 return false; // Conservative answer for dead code
304 assert(n->is_CFG(), "expecting control");
305 }
306 if (n->is_Con() || n->is_Start() || n->is_Root()) {
307 only_dominating_controls = true;
308 } else if (n->is_CFG()) {
309 if (n->dominates(sub, nlist))
310 only_dominating_controls = true;
311 else
312 return false;
313 } else {
314 // First, own control edge.
315 Node* m = n->find_exact_control(n->in(0));
316 if (m != NULL) {
317 if (m->is_top())
318 return false; // Conservative answer for dead code
319 dom_list.push(m);
320 }
321 // Now, the rest of edges.
322 uint cnt = n->req();
323 for (uint i = 1; i < cnt; i++) {
324 m = n->find_exact_control(n->in(i));
325 if (m == NULL || m->is_top())
326 continue;
327 dom_list.push(m);
328 }
329 }
330 }
331 return only_dominating_controls;
332 }
333 }
335 //---------------------detect_ptr_independence---------------------------------
336 // Used by MemNode::find_previous_store to prove that two base
337 // pointers are never equal.
338 // The pointers are accompanied by their associated allocations,
339 // if any, which have been previously discovered by the caller.
340 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
341 Node* p2, AllocateNode* a2,
342 PhaseTransform* phase) {
343 // Attempt to prove that these two pointers cannot be aliased.
344 // They may both manifestly be allocations, and they should differ.
345 // Or, if they are not both allocations, they can be distinct constants.
346 // Otherwise, one is an allocation and the other a pre-existing value.
347 if (a1 == NULL && a2 == NULL) { // neither an allocation
348 return (p1 != p2) && p1->is_Con() && p2->is_Con();
349 } else if (a1 != NULL && a2 != NULL) { // both allocations
350 return (a1 != a2);
351 } else if (a1 != NULL) { // one allocation a1
352 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
353 return all_controls_dominate(p2, a1);
354 } else { //(a2 != NULL) // one allocation a2
355 return all_controls_dominate(p1, a2);
356 }
357 return false;
358 }
361 // The logic for reordering loads and stores uses four steps:
362 // (a) Walk carefully past stores and initializations which we
363 // can prove are independent of this load.
364 // (b) Observe that the next memory state makes an exact match
365 // with self (load or store), and locate the relevant store.
366 // (c) Ensure that, if we were to wire self directly to the store,
367 // the optimizer would fold it up somehow.
368 // (d) Do the rewiring, and return, depending on some other part of
369 // the optimizer to fold up the load.
370 // This routine handles steps (a) and (b). Steps (c) and (d) are
371 // specific to loads and stores, so they are handled by the callers.
372 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
373 //
374 Node* MemNode::find_previous_store(PhaseTransform* phase) {
375 Node* ctrl = in(MemNode::Control);
376 Node* adr = in(MemNode::Address);
377 intptr_t offset = 0;
378 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
379 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
381 if (offset == Type::OffsetBot)
382 return NULL; // cannot unalias unless there are precise offsets
384 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
386 intptr_t size_in_bytes = memory_size();
388 Node* mem = in(MemNode::Memory); // start searching here...
390 int cnt = 50; // Cycle limiter
391 for (;;) { // While we can dance past unrelated stores...
392 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
394 if (mem->is_Store()) {
395 Node* st_adr = mem->in(MemNode::Address);
396 intptr_t st_offset = 0;
397 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
398 if (st_base == NULL)
399 break; // inscrutable pointer
400 if (st_offset != offset && st_offset != Type::OffsetBot) {
401 const int MAX_STORE = BytesPerLong;
402 if (st_offset >= offset + size_in_bytes ||
403 st_offset <= offset - MAX_STORE ||
404 st_offset <= offset - mem->as_Store()->memory_size()) {
405 // Success: The offsets are provably independent.
406 // (You may ask, why not just test st_offset != offset and be done?
407 // The answer is that stores of different sizes can co-exist
408 // in the same sequence of RawMem effects. We sometimes initialize
409 // a whole 'tile' of array elements with a single jint or jlong.)
410 mem = mem->in(MemNode::Memory);
411 continue; // (a) advance through independent store memory
412 }
413 }
414 if (st_base != base &&
415 detect_ptr_independence(base, alloc,
416 st_base,
417 AllocateNode::Ideal_allocation(st_base, phase),
418 phase)) {
419 // Success: The bases are provably independent.
420 mem = mem->in(MemNode::Memory);
421 continue; // (a) advance through independent store memory
422 }
424 // (b) At this point, if the bases or offsets do not agree, we lose,
425 // since we have not managed to prove 'this' and 'mem' independent.
426 if (st_base == base && st_offset == offset) {
427 return mem; // let caller handle steps (c), (d)
428 }
430 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
431 InitializeNode* st_init = mem->in(0)->as_Initialize();
432 AllocateNode* st_alloc = st_init->allocation();
433 if (st_alloc == NULL)
434 break; // something degenerated
435 bool known_identical = false;
436 bool known_independent = false;
437 if (alloc == st_alloc)
438 known_identical = true;
439 else if (alloc != NULL)
440 known_independent = true;
441 else if (all_controls_dominate(this, st_alloc))
442 known_independent = true;
444 if (known_independent) {
445 // The bases are provably independent: Either they are
446 // manifestly distinct allocations, or else the control
447 // of this load dominates the store's allocation.
448 int alias_idx = phase->C->get_alias_index(adr_type());
449 if (alias_idx == Compile::AliasIdxRaw) {
450 mem = st_alloc->in(TypeFunc::Memory);
451 } else {
452 mem = st_init->memory(alias_idx);
453 }
454 continue; // (a) advance through independent store memory
455 }
457 // (b) at this point, if we are not looking at a store initializing
458 // the same allocation we are loading from, we lose.
459 if (known_identical) {
460 // From caller, can_see_stored_value will consult find_captured_store.
461 return mem; // let caller handle steps (c), (d)
462 }
464 } else if (addr_t != NULL && addr_t->is_instance_field()) {
465 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
466 if (mem->is_Proj() && mem->in(0)->is_Call()) {
467 CallNode *call = mem->in(0)->as_Call();
468 if (!call->may_modify(addr_t, phase)) {
469 mem = call->in(TypeFunc::Memory);
470 continue; // (a) advance through independent call memory
471 }
472 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
473 mem = mem->in(0)->in(TypeFunc::Memory);
474 continue; // (a) advance through independent MemBar memory
475 } else if (mem->is_MergeMem()) {
476 int alias_idx = phase->C->get_alias_index(adr_type());
477 mem = mem->as_MergeMem()->memory_at(alias_idx);
478 continue; // (a) advance through independent MergeMem memory
479 }
480 }
482 // Unless there is an explicit 'continue', we must bail out here,
483 // because 'mem' is an inscrutable memory state (e.g., a call).
484 break;
485 }
487 return NULL; // bail out
488 }
490 //----------------------calculate_adr_type-------------------------------------
491 // Helper function. Notices when the given type of address hits top or bottom.
492 // Also, asserts a cross-check of the type against the expected address type.
493 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
494 if (t == Type::TOP) return NULL; // does not touch memory any more?
495 #ifdef PRODUCT
496 cross_check = NULL;
497 #else
498 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
499 #endif
500 const TypePtr* tp = t->isa_ptr();
501 if (tp == NULL) {
502 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
503 return TypePtr::BOTTOM; // touches lots of memory
504 } else {
505 #ifdef ASSERT
506 // %%%% [phh] We don't check the alias index if cross_check is
507 // TypeRawPtr::BOTTOM. Needs to be investigated.
508 if (cross_check != NULL &&
509 cross_check != TypePtr::BOTTOM &&
510 cross_check != TypeRawPtr::BOTTOM) {
511 // Recheck the alias index, to see if it has changed (due to a bug).
512 Compile* C = Compile::current();
513 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
514 "must stay in the original alias category");
515 // The type of the address must be contained in the adr_type,
516 // disregarding "null"-ness.
517 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
518 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
519 assert(cross_check->meet(tp_notnull) == cross_check,
520 "real address must not escape from expected memory type");
521 }
522 #endif
523 return tp;
524 }
525 }
527 //------------------------adr_phi_is_loop_invariant----------------------------
528 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
529 // loop is loop invariant. Make a quick traversal of Phi and associated
530 // CastPP nodes, looking to see if they are a closed group within the loop.
531 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
532 // The idea is that the phi-nest must boil down to only CastPP nodes
533 // with the same data. This implies that any path into the loop already
534 // includes such a CastPP, and so the original cast, whatever its input,
535 // must be covered by an equivalent cast, with an earlier control input.
536 ResourceMark rm;
538 // The loop entry input of the phi should be the unique dominating
539 // node for every Phi/CastPP in the loop.
540 Unique_Node_List closure;
541 closure.push(adr_phi->in(LoopNode::EntryControl));
543 // Add the phi node and the cast to the worklist.
544 Unique_Node_List worklist;
545 worklist.push(adr_phi);
546 if( cast != NULL ){
547 if( !cast->is_ConstraintCast() ) return false;
548 worklist.push(cast);
549 }
551 // Begin recursive walk of phi nodes.
552 while( worklist.size() ){
553 // Take a node off the worklist
554 Node *n = worklist.pop();
555 if( !closure.member(n) ){
556 // Add it to the closure.
557 closure.push(n);
558 // Make a sanity check to ensure we don't waste too much time here.
559 if( closure.size() > 20) return false;
560 // This node is OK if:
561 // - it is a cast of an identical value
562 // - or it is a phi node (then we add its inputs to the worklist)
563 // Otherwise, the node is not OK, and we presume the cast is not invariant
564 if( n->is_ConstraintCast() ){
565 worklist.push(n->in(1));
566 } else if( n->is_Phi() ) {
567 for( uint i = 1; i < n->req(); i++ ) {
568 worklist.push(n->in(i));
569 }
570 } else {
571 return false;
572 }
573 }
574 }
576 // Quit when the worklist is empty, and we've found no offending nodes.
577 return true;
578 }
580 //------------------------------Ideal_DU_postCCP-------------------------------
581 // Find any cast-away of null-ness and keep its control. Null cast-aways are
582 // going away in this pass and we need to make this memory op depend on the
583 // gating null check.
584 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
585 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
586 }
588 // I tried to leave the CastPP's in. This makes the graph more accurate in
589 // some sense; we get to keep around the knowledge that an oop is not-null
590 // after some test. Alas, the CastPP's interfere with GVN (some values are
591 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
592 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
593 // some of the more trivial cases in the optimizer. Removing more useless
594 // Phi's started allowing Loads to illegally float above null checks. I gave
595 // up on this approach. CNC 10/20/2000
596 // This static method may be called not from MemNode (EncodePNode calls it).
597 // Only the control edge of the node 'n' might be updated.
598 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
599 Node *skipped_cast = NULL;
600 // Need a null check? Regular static accesses do not because they are
601 // from constant addresses. Array ops are gated by the range check (which
602 // always includes a NULL check). Just check field ops.
603 if( n->in(MemNode::Control) == NULL ) {
604 // Scan upwards for the highest location we can place this memory op.
605 while( true ) {
606 switch( adr->Opcode() ) {
608 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
609 adr = adr->in(AddPNode::Base);
610 continue;
612 case Op_DecodeN: // No change to NULL-ness, so peek thru
613 adr = adr->in(1);
614 continue;
616 case Op_CastPP:
617 // If the CastPP is useless, just peek on through it.
618 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
619 // Remember the cast that we've peeked though. If we peek
620 // through more than one, then we end up remembering the highest
621 // one, that is, if in a loop, the one closest to the top.
622 skipped_cast = adr;
623 adr = adr->in(1);
624 continue;
625 }
626 // CastPP is going away in this pass! We need this memory op to be
627 // control-dependent on the test that is guarding the CastPP.
628 ccp->hash_delete(n);
629 n->set_req(MemNode::Control, adr->in(0));
630 ccp->hash_insert(n);
631 return n;
633 case Op_Phi:
634 // Attempt to float above a Phi to some dominating point.
635 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
636 // If we've already peeked through a Cast (which could have set the
637 // control), we can't float above a Phi, because the skipped Cast
638 // may not be loop invariant.
639 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
640 adr = adr->in(1);
641 continue;
642 }
643 }
645 // Intentional fallthrough!
647 // No obvious dominating point. The mem op is pinned below the Phi
648 // by the Phi itself. If the Phi goes away (no true value is merged)
649 // then the mem op can float, but not indefinitely. It must be pinned
650 // behind the controls leading to the Phi.
651 case Op_CheckCastPP:
652 // These usually stick around to change address type, however a
653 // useless one can be elided and we still need to pick up a control edge
654 if (adr->in(0) == NULL) {
655 // This CheckCastPP node has NO control and is likely useless. But we
656 // need check further up the ancestor chain for a control input to keep
657 // the node in place. 4959717.
658 skipped_cast = adr;
659 adr = adr->in(1);
660 continue;
661 }
662 ccp->hash_delete(n);
663 n->set_req(MemNode::Control, adr->in(0));
664 ccp->hash_insert(n);
665 return n;
667 // List of "safe" opcodes; those that implicitly block the memory
668 // op below any null check.
669 case Op_CastX2P: // no null checks on native pointers
670 case Op_Parm: // 'this' pointer is not null
671 case Op_LoadP: // Loading from within a klass
672 case Op_LoadN: // Loading from within a klass
673 case Op_LoadKlass: // Loading from within a klass
674 case Op_ConP: // Loading from a klass
675 case Op_ConN: // Loading from a klass
676 case Op_CreateEx: // Sucking up the guts of an exception oop
677 case Op_Con: // Reading from TLS
678 case Op_CMoveP: // CMoveP is pinned
679 break; // No progress
681 case Op_Proj: // Direct call to an allocation routine
682 case Op_SCMemProj: // Memory state from store conditional ops
683 #ifdef ASSERT
684 {
685 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
686 const Node* call = adr->in(0);
687 if (call->is_CallJava()) {
688 const CallJavaNode* call_java = call->as_CallJava();
689 const TypeTuple *r = call_java->tf()->range();
690 assert(r->cnt() > TypeFunc::Parms, "must return value");
691 const Type* ret_type = r->field_at(TypeFunc::Parms);
692 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
693 // We further presume that this is one of
694 // new_instance_Java, new_array_Java, or
695 // the like, but do not assert for this.
696 } else if (call->is_Allocate()) {
697 // similar case to new_instance_Java, etc.
698 } else if (!call->is_CallLeaf()) {
699 // Projections from fetch_oop (OSR) are allowed as well.
700 ShouldNotReachHere();
701 }
702 }
703 #endif
704 break;
705 default:
706 ShouldNotReachHere();
707 }
708 break;
709 }
710 }
712 return NULL; // No progress
713 }
716 //=============================================================================
717 uint LoadNode::size_of() const { return sizeof(*this); }
718 uint LoadNode::cmp( const Node &n ) const
719 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
720 const Type *LoadNode::bottom_type() const { return _type; }
721 uint LoadNode::ideal_reg() const {
722 return Matcher::base2reg[_type->base()];
723 }
725 #ifndef PRODUCT
726 void LoadNode::dump_spec(outputStream *st) const {
727 MemNode::dump_spec(st);
728 if( !Verbose && !WizardMode ) {
729 // standard dump does this in Verbose and WizardMode
730 st->print(" #"); _type->dump_on(st);
731 }
732 }
733 #endif
736 //----------------------------LoadNode::make-----------------------------------
737 // Polymorphic factory method:
738 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
739 Compile* C = gvn.C;
741 // sanity check the alias category against the created node type
742 assert(!(adr_type->isa_oopptr() &&
743 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
744 "use LoadKlassNode instead");
745 assert(!(adr_type->isa_aryptr() &&
746 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
747 "use LoadRangeNode instead");
748 switch (bt) {
749 case T_BOOLEAN:
750 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
751 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
752 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
753 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
754 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
755 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
756 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
757 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
758 case T_OBJECT:
759 #ifdef _LP64
760 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
761 const TypeNarrowOop* narrowtype;
762 if (rt->isa_narrowoop()) {
763 narrowtype = rt->is_narrowoop();
764 } else {
765 narrowtype = rt->is_oopptr()->make_narrowoop();
766 }
767 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, narrowtype));
769 return DecodeNNode::decode(&gvn, load);
770 } else
771 #endif
772 {
773 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
774 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
775 }
776 }
777 ShouldNotReachHere();
778 return (LoadNode*)NULL;
779 }
781 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
782 bool require_atomic = true;
783 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
784 }
789 //------------------------------hash-------------------------------------------
790 uint LoadNode::hash() const {
791 // unroll addition of interesting fields
792 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
793 }
795 //---------------------------can_see_stored_value------------------------------
796 // This routine exists to make sure this set of tests is done the same
797 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
798 // will change the graph shape in a way which makes memory alive twice at the
799 // same time (uses the Oracle model of aliasing), then some
800 // LoadXNode::Identity will fold things back to the equivalence-class model
801 // of aliasing.
802 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
803 Node* ld_adr = in(MemNode::Address);
805 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
806 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
807 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
808 atp->field() != NULL && !atp->field()->is_volatile()) {
809 uint alias_idx = atp->index();
810 bool final = atp->field()->is_final();
811 Node* result = NULL;
812 Node* current = st;
813 // Skip through chains of MemBarNodes checking the MergeMems for
814 // new states for the slice of this load. Stop once any other
815 // kind of node is encountered. Loads from final memory can skip
816 // through any kind of MemBar but normal loads shouldn't skip
817 // through MemBarAcquire since the could allow them to move out of
818 // a synchronized region.
819 while (current->is_Proj()) {
820 int opc = current->in(0)->Opcode();
821 if ((final && opc == Op_MemBarAcquire) ||
822 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
823 Node* mem = current->in(0)->in(TypeFunc::Memory);
824 if (mem->is_MergeMem()) {
825 MergeMemNode* merge = mem->as_MergeMem();
826 Node* new_st = merge->memory_at(alias_idx);
827 if (new_st == merge->base_memory()) {
828 // Keep searching
829 current = merge->base_memory();
830 continue;
831 }
832 // Save the new memory state for the slice and fall through
833 // to exit.
834 result = new_st;
835 }
836 }
837 break;
838 }
839 if (result != NULL) {
840 st = result;
841 }
842 }
845 // Loop around twice in the case Load -> Initialize -> Store.
846 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
847 for (int trip = 0; trip <= 1; trip++) {
849 if (st->is_Store()) {
850 Node* st_adr = st->in(MemNode::Address);
851 if (!phase->eqv(st_adr, ld_adr)) {
852 // Try harder before giving up... Match raw and non-raw pointers.
853 intptr_t st_off = 0;
854 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
855 if (alloc == NULL) return NULL;
856 intptr_t ld_off = 0;
857 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
858 if (alloc != allo2) return NULL;
859 if (ld_off != st_off) return NULL;
860 // At this point we have proven something like this setup:
861 // A = Allocate(...)
862 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
863 // S = StoreQ(, AddP(, A.Parm , #Off), V)
864 // (Actually, we haven't yet proven the Q's are the same.)
865 // In other words, we are loading from a casted version of
866 // the same pointer-and-offset that we stored to.
867 // Thus, we are able to replace L by V.
868 }
869 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
870 if (store_Opcode() != st->Opcode())
871 return NULL;
872 return st->in(MemNode::ValueIn);
873 }
875 intptr_t offset = 0; // scratch
877 // A load from a freshly-created object always returns zero.
878 // (This can happen after LoadNode::Ideal resets the load's memory input
879 // to find_captured_store, which returned InitializeNode::zero_memory.)
880 if (st->is_Proj() && st->in(0)->is_Allocate() &&
881 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
882 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
883 // return a zero value for the load's basic type
884 // (This is one of the few places where a generic PhaseTransform
885 // can create new nodes. Think of it as lazily manifesting
886 // virtually pre-existing constants.)
887 return phase->zerocon(memory_type());
888 }
890 // A load from an initialization barrier can match a captured store.
891 if (st->is_Proj() && st->in(0)->is_Initialize()) {
892 InitializeNode* init = st->in(0)->as_Initialize();
893 AllocateNode* alloc = init->allocation();
894 if (alloc != NULL &&
895 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
896 // examine a captured store value
897 st = init->find_captured_store(offset, memory_size(), phase);
898 if (st != NULL)
899 continue; // take one more trip around
900 }
901 }
903 break;
904 }
906 return NULL;
907 }
909 //----------------------is_instance_field_load_with_local_phi------------------
910 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
911 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
912 in(MemNode::Address)->is_AddP() ) {
913 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
914 // Only instances.
915 if( t_oop != NULL && t_oop->is_instance_field() &&
916 t_oop->offset() != Type::OffsetBot &&
917 t_oop->offset() != Type::OffsetTop) {
918 return true;
919 }
920 }
921 return false;
922 }
924 //------------------------------Identity---------------------------------------
925 // Loads are identity if previous store is to same address
926 Node *LoadNode::Identity( PhaseTransform *phase ) {
927 // If the previous store-maker is the right kind of Store, and the store is
928 // to the same address, then we are equal to the value stored.
929 Node* mem = in(MemNode::Memory);
930 Node* value = can_see_stored_value(mem, phase);
931 if( value ) {
932 // byte, short & char stores truncate naturally.
933 // A load has to load the truncated value which requires
934 // some sort of masking operation and that requires an
935 // Ideal call instead of an Identity call.
936 if (memory_size() < BytesPerInt) {
937 // If the input to the store does not fit with the load's result type,
938 // it must be truncated via an Ideal call.
939 if (!phase->type(value)->higher_equal(phase->type(this)))
940 return this;
941 }
942 // (This works even when value is a Con, but LoadNode::Value
943 // usually runs first, producing the singleton type of the Con.)
944 return value;
945 }
947 // Search for an existing data phi which was generated before for the same
948 // instance's field to avoid infinite genertion of phis in a loop.
949 Node *region = mem->in(0);
950 if (is_instance_field_load_with_local_phi(region)) {
951 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
952 int this_index = phase->C->get_alias_index(addr_t);
953 int this_offset = addr_t->offset();
954 int this_id = addr_t->is_oopptr()->instance_id();
955 const Type* this_type = bottom_type();
956 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
957 Node* phi = region->fast_out(i);
958 if (phi->is_Phi() && phi != mem &&
959 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
960 return phi;
961 }
962 }
963 }
965 return this;
966 }
969 // Returns true if the AliasType refers to the field that holds the
970 // cached box array. Currently only handles the IntegerCache case.
971 static bool is_autobox_cache(Compile::AliasType* atp) {
972 if (atp != NULL && atp->field() != NULL) {
973 ciField* field = atp->field();
974 ciSymbol* klass = field->holder()->name();
975 if (field->name() == ciSymbol::cache_field_name() &&
976 field->holder()->uses_default_loader() &&
977 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
978 return true;
979 }
980 }
981 return false;
982 }
984 // Fetch the base value in the autobox array
985 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
986 if (atp != NULL && atp->field() != NULL) {
987 ciField* field = atp->field();
988 ciSymbol* klass = field->holder()->name();
989 if (field->name() == ciSymbol::cache_field_name() &&
990 field->holder()->uses_default_loader() &&
991 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
992 assert(field->is_constant(), "what?");
993 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
994 // Fetch the box object at the base of the array and get its value
995 ciInstance* box = array->obj_at(0)->as_instance();
996 ciInstanceKlass* ik = box->klass()->as_instance_klass();
997 if (ik->nof_nonstatic_fields() == 1) {
998 // This should be true nonstatic_field_at requires calling
999 // nof_nonstatic_fields so check it anyway
1000 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1001 cache_offset = c.as_int();
1002 }
1003 return true;
1004 }
1005 }
1006 return false;
1007 }
1009 // Returns true if the AliasType refers to the value field of an
1010 // autobox object. Currently only handles Integer.
1011 static bool is_autobox_object(Compile::AliasType* atp) {
1012 if (atp != NULL && atp->field() != NULL) {
1013 ciField* field = atp->field();
1014 ciSymbol* klass = field->holder()->name();
1015 if (field->name() == ciSymbol::value_name() &&
1016 field->holder()->uses_default_loader() &&
1017 klass == ciSymbol::java_lang_Integer()) {
1018 return true;
1019 }
1020 }
1021 return false;
1022 }
1025 // We're loading from an object which has autobox behaviour.
1026 // If this object is result of a valueOf call we'll have a phi
1027 // merging a newly allocated object and a load from the cache.
1028 // We want to replace this load with the original incoming
1029 // argument to the valueOf call.
1030 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1031 Node* base = in(Address)->in(AddPNode::Base);
1032 if (base->is_Phi() && base->req() == 3) {
1033 AllocateNode* allocation = NULL;
1034 int allocation_index = -1;
1035 int load_index = -1;
1036 for (uint i = 1; i < base->req(); i++) {
1037 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1038 if (allocation != NULL) {
1039 allocation_index = i;
1040 load_index = 3 - allocation_index;
1041 break;
1042 }
1043 }
1044 LoadNode* load = NULL;
1045 if (allocation != NULL && base->in(load_index)->is_Load()) {
1046 load = base->in(load_index)->as_Load();
1047 }
1048 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1049 // Push the loads from the phi that comes from valueOf up
1050 // through it to allow elimination of the loads and the recovery
1051 // of the original value.
1052 Node* mem_phi = in(Memory);
1053 Node* offset = in(Address)->in(AddPNode::Offset);
1055 Node* in1 = clone();
1056 Node* in1_addr = in1->in(Address)->clone();
1057 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1058 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1059 in1_addr->set_req(AddPNode::Offset, offset);
1060 in1->set_req(0, base->in(allocation_index));
1061 in1->set_req(Address, in1_addr);
1062 in1->set_req(Memory, mem_phi->in(allocation_index));
1064 Node* in2 = clone();
1065 Node* in2_addr = in2->in(Address)->clone();
1066 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1067 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1068 in2_addr->set_req(AddPNode::Offset, offset);
1069 in2->set_req(0, base->in(load_index));
1070 in2->set_req(Address, in2_addr);
1071 in2->set_req(Memory, mem_phi->in(load_index));
1073 in1_addr = phase->transform(in1_addr);
1074 in1 = phase->transform(in1);
1075 in2_addr = phase->transform(in2_addr);
1076 in2 = phase->transform(in2);
1078 PhiNode* result = PhiNode::make_blank(base->in(0), this);
1079 result->set_req(allocation_index, in1);
1080 result->set_req(load_index, in2);
1081 return result;
1082 }
1083 } else if (base->is_Load()) {
1084 // Eliminate the load of Integer.value for integers from the cache
1085 // array by deriving the value from the index into the array.
1086 // Capture the offset of the load and then reverse the computation.
1087 Node* load_base = base->in(Address)->in(AddPNode::Base);
1088 if (load_base != NULL) {
1089 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1090 intptr_t cache_offset;
1091 int shift = -1;
1092 Node* cache = NULL;
1093 if (is_autobox_cache(atp)) {
1094 shift = exact_log2(type2aelembytes(T_OBJECT));
1095 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1096 }
1097 if (cache != NULL && base->in(Address)->is_AddP()) {
1098 Node* elements[4];
1099 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1100 int cache_low;
1101 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1102 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1103 // Add up all the offsets making of the address of the load
1104 Node* result = elements[0];
1105 for (int i = 1; i < count; i++) {
1106 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1107 }
1108 // Remove the constant offset from the address and then
1109 // remove the scaling of the offset to recover the original index.
1110 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1111 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1112 // Peel the shift off directly but wrap it in a dummy node
1113 // since Ideal can't return existing nodes
1114 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1115 } else {
1116 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1117 }
1118 #ifdef _LP64
1119 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1120 #endif
1121 return result;
1122 }
1123 }
1124 }
1125 }
1126 return NULL;
1127 }
1129 //------------------------------split_through_phi------------------------------
1130 // Split instance field load through Phi.
1131 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1132 Node* mem = in(MemNode::Memory);
1133 Node* address = in(MemNode::Address);
1134 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1135 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1137 assert(mem->is_Phi() && (t_oop != NULL) &&
1138 t_oop->is_instance_field(), "invalide conditions");
1140 Node *region = mem->in(0);
1141 if (region == NULL) {
1142 return NULL; // Wait stable graph
1143 }
1144 uint cnt = mem->req();
1145 for( uint i = 1; i < cnt; i++ ) {
1146 Node *in = mem->in(i);
1147 if( in == NULL ) {
1148 return NULL; // Wait stable graph
1149 }
1150 }
1151 // Check for loop invariant.
1152 if (cnt == 3) {
1153 for( uint i = 1; i < cnt; i++ ) {
1154 Node *in = mem->in(i);
1155 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1156 if (m == mem) {
1157 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
1158 return this;
1159 }
1160 }
1161 }
1162 // Split through Phi (see original code in loopopts.cpp).
1163 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1165 // Do nothing here if Identity will find a value
1166 // (to avoid infinite chain of value phis generation).
1167 if ( !phase->eqv(this, this->Identity(phase)) )
1168 return NULL;
1170 // Skip the split if the region dominates some control edge of the address.
1171 if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
1172 return NULL;
1174 const Type* this_type = this->bottom_type();
1175 int this_index = phase->C->get_alias_index(addr_t);
1176 int this_offset = addr_t->offset();
1177 int this_iid = addr_t->is_oopptr()->instance_id();
1178 int wins = 0;
1179 PhaseIterGVN *igvn = phase->is_IterGVN();
1180 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1181 for( uint i = 1; i < region->req(); i++ ) {
1182 Node *x;
1183 Node* the_clone = NULL;
1184 if( region->in(i) == phase->C->top() ) {
1185 x = phase->C->top(); // Dead path? Use a dead data op
1186 } else {
1187 x = this->clone(); // Else clone up the data op
1188 the_clone = x; // Remember for possible deletion.
1189 // Alter data node to use pre-phi inputs
1190 if( this->in(0) == region ) {
1191 x->set_req( 0, region->in(i) );
1192 } else {
1193 x->set_req( 0, NULL );
1194 }
1195 for( uint j = 1; j < this->req(); j++ ) {
1196 Node *in = this->in(j);
1197 if( in->is_Phi() && in->in(0) == region )
1198 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1199 }
1200 }
1201 // Check for a 'win' on some paths
1202 const Type *t = x->Value(igvn);
1204 bool singleton = t->singleton();
1206 // See comments in PhaseIdealLoop::split_thru_phi().
1207 if( singleton && t == Type::TOP ) {
1208 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1209 }
1211 if( singleton ) {
1212 wins++;
1213 x = igvn->makecon(t);
1214 } else {
1215 // We now call Identity to try to simplify the cloned node.
1216 // Note that some Identity methods call phase->type(this).
1217 // Make sure that the type array is big enough for
1218 // our new node, even though we may throw the node away.
1219 // (This tweaking with igvn only works because x is a new node.)
1220 igvn->set_type(x, t);
1221 Node *y = x->Identity(igvn);
1222 if( y != x ) {
1223 wins++;
1224 x = y;
1225 } else {
1226 y = igvn->hash_find(x);
1227 if( y ) {
1228 wins++;
1229 x = y;
1230 } else {
1231 // Else x is a new node we are keeping
1232 // We do not need register_new_node_with_optimizer
1233 // because set_type has already been called.
1234 igvn->_worklist.push(x);
1235 }
1236 }
1237 }
1238 if (x != the_clone && the_clone != NULL)
1239 igvn->remove_dead_node(the_clone);
1240 phi->set_req(i, x);
1241 }
1242 if( wins > 0 ) {
1243 // Record Phi
1244 igvn->register_new_node_with_optimizer(phi);
1245 return phi;
1246 }
1247 igvn->remove_dead_node(phi);
1248 return NULL;
1249 }
1251 //------------------------------Ideal------------------------------------------
1252 // If the load is from Field memory and the pointer is non-null, we can
1253 // zero out the control input.
1254 // If the offset is constant and the base is an object allocation,
1255 // try to hook me up to the exact initializing store.
1256 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1257 Node* p = MemNode::Ideal_common(phase, can_reshape);
1258 if (p) return (p == NodeSentinel) ? NULL : p;
1260 Node* ctrl = in(MemNode::Control);
1261 Node* address = in(MemNode::Address);
1263 // Skip up past a SafePoint control. Cannot do this for Stores because
1264 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1265 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1266 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1267 ctrl = ctrl->in(0);
1268 set_req(MemNode::Control,ctrl);
1269 }
1271 // Check for useless control edge in some common special cases
1272 if (in(MemNode::Control) != NULL) {
1273 intptr_t ignore = 0;
1274 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1275 if (base != NULL
1276 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1277 && all_controls_dominate(base, phase->C->start())) {
1278 // A method-invariant, non-null address (constant or 'this' argument).
1279 set_req(MemNode::Control, NULL);
1280 }
1281 }
1283 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
1284 Node* base = in(Address)->in(AddPNode::Base);
1285 if (base != NULL) {
1286 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1287 if (is_autobox_object(atp)) {
1288 Node* result = eliminate_autobox(phase);
1289 if (result != NULL) return result;
1290 }
1291 }
1292 }
1294 Node* mem = in(MemNode::Memory);
1295 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1297 if (addr_t != NULL) {
1298 // try to optimize our memory input
1299 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1300 if (opt_mem != mem) {
1301 set_req(MemNode::Memory, opt_mem);
1302 return this;
1303 }
1304 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1305 if (can_reshape && opt_mem->is_Phi() &&
1306 (t_oop != NULL) && t_oop->is_instance_field()) {
1307 // Split instance field load through Phi.
1308 Node* result = split_through_phi(phase);
1309 if (result != NULL) return result;
1310 }
1311 }
1313 // Check for prior store with a different base or offset; make Load
1314 // independent. Skip through any number of them. Bail out if the stores
1315 // are in an endless dead cycle and report no progress. This is a key
1316 // transform for Reflection. However, if after skipping through the Stores
1317 // we can't then fold up against a prior store do NOT do the transform as
1318 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1319 // array memory alive twice: once for the hoisted Load and again after the
1320 // bypassed Store. This situation only works if EVERYBODY who does
1321 // anti-dependence work knows how to bypass. I.e. we need all
1322 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1323 // the alias index stuff. So instead, peek through Stores and IFF we can
1324 // fold up, do so.
1325 Node* prev_mem = find_previous_store(phase);
1326 // Steps (a), (b): Walk past independent stores to find an exact match.
1327 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1328 // (c) See if we can fold up on the spot, but don't fold up here.
1329 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
1330 // just return a prior value, which is done by Identity calls.
1331 if (can_see_stored_value(prev_mem, phase)) {
1332 // Make ready for step (d):
1333 set_req(MemNode::Memory, prev_mem);
1334 return this;
1335 }
1336 }
1338 return NULL; // No further progress
1339 }
1341 // Helper to recognize certain Klass fields which are invariant across
1342 // some group of array types (e.g., int[] or all T[] where T < Object).
1343 const Type*
1344 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1345 ciKlass* klass) const {
1346 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1347 // The field is Klass::_modifier_flags. Return its (constant) value.
1348 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1349 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1350 return TypeInt::make(klass->modifier_flags());
1351 }
1352 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1353 // The field is Klass::_access_flags. Return its (constant) value.
1354 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1355 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1356 return TypeInt::make(klass->access_flags());
1357 }
1358 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1359 // The field is Klass::_layout_helper. Return its constant value if known.
1360 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1361 return TypeInt::make(klass->layout_helper());
1362 }
1364 // No match.
1365 return NULL;
1366 }
1368 //------------------------------Value-----------------------------------------
1369 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1370 // Either input is TOP ==> the result is TOP
1371 Node* mem = in(MemNode::Memory);
1372 const Type *t1 = phase->type(mem);
1373 if (t1 == Type::TOP) return Type::TOP;
1374 Node* adr = in(MemNode::Address);
1375 const TypePtr* tp = phase->type(adr)->isa_ptr();
1376 if (tp == NULL || tp->empty()) return Type::TOP;
1377 int off = tp->offset();
1378 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1380 // Try to guess loaded type from pointer type
1381 if (tp->base() == Type::AryPtr) {
1382 const Type *t = tp->is_aryptr()->elem();
1383 // Don't do this for integer types. There is only potential profit if
1384 // the element type t is lower than _type; that is, for int types, if _type is
1385 // more restrictive than t. This only happens here if one is short and the other
1386 // char (both 16 bits), and in those cases we've made an intentional decision
1387 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1388 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1389 //
1390 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1391 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1392 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1393 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1394 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1395 // In fact, that could have been the original type of p1, and p1 could have
1396 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1397 // expression (LShiftL quux 3) independently optimized to the constant 8.
1398 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1399 && Opcode() != Op_LoadKlass) {
1400 // t might actually be lower than _type, if _type is a unique
1401 // concrete subclass of abstract class t.
1402 // Make sure the reference is not into the header, by comparing
1403 // the offset against the offset of the start of the array's data.
1404 // Different array types begin at slightly different offsets (12 vs. 16).
1405 // We choose T_BYTE as an example base type that is least restrictive
1406 // as to alignment, which will therefore produce the smallest
1407 // possible base offset.
1408 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1409 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1410 const Type* jt = t->join(_type);
1411 // In any case, do not allow the join, per se, to empty out the type.
1412 if (jt->empty() && !t->empty()) {
1413 // This can happen if a interface-typed array narrows to a class type.
1414 jt = _type;
1415 }
1417 if (EliminateAutoBox) {
1418 // The pointers in the autobox arrays are always non-null
1419 Node* base = in(Address)->in(AddPNode::Base);
1420 if (base != NULL) {
1421 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1422 if (is_autobox_cache(atp)) {
1423 return jt->join(TypePtr::NOTNULL)->is_ptr();
1424 }
1425 }
1426 }
1427 return jt;
1428 }
1429 }
1430 } else if (tp->base() == Type::InstPtr) {
1431 assert( off != Type::OffsetBot ||
1432 // arrays can be cast to Objects
1433 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1434 // unsafe field access may not have a constant offset
1435 phase->C->has_unsafe_access(),
1436 "Field accesses must be precise" );
1437 // For oop loads, we expect the _type to be precise
1438 } else if (tp->base() == Type::KlassPtr) {
1439 assert( off != Type::OffsetBot ||
1440 // arrays can be cast to Objects
1441 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1442 // also allow array-loading from the primary supertype
1443 // array during subtype checks
1444 Opcode() == Op_LoadKlass,
1445 "Field accesses must be precise" );
1446 // For klass/static loads, we expect the _type to be precise
1447 }
1449 const TypeKlassPtr *tkls = tp->isa_klassptr();
1450 if (tkls != NULL && !StressReflectiveCode) {
1451 ciKlass* klass = tkls->klass();
1452 if (klass->is_loaded() && tkls->klass_is_exact()) {
1453 // We are loading a field from a Klass metaobject whose identity
1454 // is known at compile time (the type is "exact" or "precise").
1455 // Check for fields we know are maintained as constants by the VM.
1456 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1457 // The field is Klass::_super_check_offset. Return its (constant) value.
1458 // (Folds up type checking code.)
1459 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1460 return TypeInt::make(klass->super_check_offset());
1461 }
1462 // Compute index into primary_supers array
1463 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1464 // Check for overflowing; use unsigned compare to handle the negative case.
1465 if( depth < ciKlass::primary_super_limit() ) {
1466 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1467 // (Folds up type checking code.)
1468 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1469 ciKlass *ss = klass->super_of_depth(depth);
1470 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1471 }
1472 const Type* aift = load_array_final_field(tkls, klass);
1473 if (aift != NULL) return aift;
1474 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1475 && klass->is_array_klass()) {
1476 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1477 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1478 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1479 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1480 }
1481 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1482 // The field is Klass::_java_mirror. Return its (constant) value.
1483 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1484 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1485 return TypeInstPtr::make(klass->java_mirror());
1486 }
1487 }
1489 // We can still check if we are loading from the primary_supers array at a
1490 // shallow enough depth. Even though the klass is not exact, entries less
1491 // than or equal to its super depth are correct.
1492 if (klass->is_loaded() ) {
1493 ciType *inner = klass->klass();
1494 while( inner->is_obj_array_klass() )
1495 inner = inner->as_obj_array_klass()->base_element_type();
1496 if( inner->is_instance_klass() &&
1497 !inner->as_instance_klass()->flags().is_interface() ) {
1498 // Compute index into primary_supers array
1499 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1500 // Check for overflowing; use unsigned compare to handle the negative case.
1501 if( depth < ciKlass::primary_super_limit() &&
1502 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1503 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1504 // (Folds up type checking code.)
1505 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1506 ciKlass *ss = klass->super_of_depth(depth);
1507 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1508 }
1509 }
1510 }
1512 // If the type is enough to determine that the thing is not an array,
1513 // we can give the layout_helper a positive interval type.
1514 // This will help short-circuit some reflective code.
1515 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1516 && !klass->is_array_klass() // not directly typed as an array
1517 && !klass->is_interface() // specifically not Serializable & Cloneable
1518 && !klass->is_java_lang_Object() // not the supertype of all T[]
1519 ) {
1520 // Note: When interfaces are reliable, we can narrow the interface
1521 // test to (klass != Serializable && klass != Cloneable).
1522 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1523 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1524 // The key property of this type is that it folds up tests
1525 // for array-ness, since it proves that the layout_helper is positive.
1526 // Thus, a generic value like the basic object layout helper works fine.
1527 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1528 }
1529 }
1531 // If we are loading from a freshly-allocated object, produce a zero,
1532 // if the load is provably beyond the header of the object.
1533 // (Also allow a variable load from a fresh array to produce zero.)
1534 if (ReduceFieldZeroing) {
1535 Node* value = can_see_stored_value(mem,phase);
1536 if (value != NULL && value->is_Con())
1537 return value->bottom_type();
1538 }
1540 const TypeOopPtr *tinst = tp->isa_oopptr();
1541 if (tinst != NULL && tinst->is_instance_field()) {
1542 // If we have an instance type and our memory input is the
1543 // programs's initial memory state, there is no matching store,
1544 // so just return a zero of the appropriate type
1545 Node *mem = in(MemNode::Memory);
1546 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1547 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1548 return Type::get_zero_type(_type->basic_type());
1549 }
1550 }
1551 return _type;
1552 }
1554 //------------------------------match_edge-------------------------------------
1555 // Do we Match on this edge index or not? Match only the address.
1556 uint LoadNode::match_edge(uint idx) const {
1557 return idx == MemNode::Address;
1558 }
1560 //--------------------------LoadBNode::Ideal--------------------------------------
1561 //
1562 // If the previous store is to the same address as this load,
1563 // and the value stored was larger than a byte, replace this load
1564 // with the value stored truncated to a byte. If no truncation is
1565 // needed, the replacement is done in LoadNode::Identity().
1566 //
1567 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1568 Node* mem = in(MemNode::Memory);
1569 Node* value = can_see_stored_value(mem,phase);
1570 if( value && !phase->type(value)->higher_equal( _type ) ) {
1571 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1572 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1573 }
1574 // Identity call will handle the case where truncation is not needed.
1575 return LoadNode::Ideal(phase, can_reshape);
1576 }
1578 //--------------------------LoadCNode::Ideal--------------------------------------
1579 //
1580 // If the previous store is to the same address as this load,
1581 // and the value stored was larger than a char, replace this load
1582 // with the value stored truncated to a char. If no truncation is
1583 // needed, the replacement is done in LoadNode::Identity().
1584 //
1585 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1586 Node* mem = in(MemNode::Memory);
1587 Node* value = can_see_stored_value(mem,phase);
1588 if( value && !phase->type(value)->higher_equal( _type ) )
1589 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1590 // Identity call will handle the case where truncation is not needed.
1591 return LoadNode::Ideal(phase, can_reshape);
1592 }
1594 //--------------------------LoadSNode::Ideal--------------------------------------
1595 //
1596 // If the previous store is to the same address as this load,
1597 // and the value stored was larger than a short, replace this load
1598 // with the value stored truncated to a short. If no truncation is
1599 // needed, the replacement is done in LoadNode::Identity().
1600 //
1601 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1602 Node* mem = in(MemNode::Memory);
1603 Node* value = can_see_stored_value(mem,phase);
1604 if( value && !phase->type(value)->higher_equal( _type ) ) {
1605 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1606 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1607 }
1608 // Identity call will handle the case where truncation is not needed.
1609 return LoadNode::Ideal(phase, can_reshape);
1610 }
1612 //=============================================================================
1613 //------------------------------Value------------------------------------------
1614 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1615 // Either input is TOP ==> the result is TOP
1616 const Type *t1 = phase->type( in(MemNode::Memory) );
1617 if (t1 == Type::TOP) return Type::TOP;
1618 Node *adr = in(MemNode::Address);
1619 const Type *t2 = phase->type( adr );
1620 if (t2 == Type::TOP) return Type::TOP;
1621 const TypePtr *tp = t2->is_ptr();
1622 if (TypePtr::above_centerline(tp->ptr()) ||
1623 tp->ptr() == TypePtr::Null) return Type::TOP;
1625 // Return a more precise klass, if possible
1626 const TypeInstPtr *tinst = tp->isa_instptr();
1627 if (tinst != NULL) {
1628 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1629 int offset = tinst->offset();
1630 if (ik == phase->C->env()->Class_klass()
1631 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1632 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1633 // We are loading a special hidden field from a Class mirror object,
1634 // the field which points to the VM's Klass metaobject.
1635 ciType* t = tinst->java_mirror_type();
1636 // java_mirror_type returns non-null for compile-time Class constants.
1637 if (t != NULL) {
1638 // constant oop => constant klass
1639 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1640 return TypeKlassPtr::make(ciArrayKlass::make(t));
1641 }
1642 if (!t->is_klass()) {
1643 // a primitive Class (e.g., int.class) has NULL for a klass field
1644 return TypePtr::NULL_PTR;
1645 }
1646 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1647 return TypeKlassPtr::make(t->as_klass());
1648 }
1649 // non-constant mirror, so we can't tell what's going on
1650 }
1651 if( !ik->is_loaded() )
1652 return _type; // Bail out if not loaded
1653 if (offset == oopDesc::klass_offset_in_bytes()) {
1654 if (tinst->klass_is_exact()) {
1655 return TypeKlassPtr::make(ik);
1656 }
1657 // See if we can become precise: no subklasses and no interface
1658 // (Note: We need to support verified interfaces.)
1659 if (!ik->is_interface() && !ik->has_subklass()) {
1660 //assert(!UseExactTypes, "this code should be useless with exact types");
1661 // Add a dependence; if any subclass added we need to recompile
1662 if (!ik->is_final()) {
1663 // %%% should use stronger assert_unique_concrete_subtype instead
1664 phase->C->dependencies()->assert_leaf_type(ik);
1665 }
1666 // Return precise klass
1667 return TypeKlassPtr::make(ik);
1668 }
1670 // Return root of possible klass
1671 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1672 }
1673 }
1675 // Check for loading klass from an array
1676 const TypeAryPtr *tary = tp->isa_aryptr();
1677 if( tary != NULL ) {
1678 ciKlass *tary_klass = tary->klass();
1679 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1680 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1681 if (tary->klass_is_exact()) {
1682 return TypeKlassPtr::make(tary_klass);
1683 }
1684 ciArrayKlass *ak = tary->klass()->as_array_klass();
1685 // If the klass is an object array, we defer the question to the
1686 // array component klass.
1687 if( ak->is_obj_array_klass() ) {
1688 assert( ak->is_loaded(), "" );
1689 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1690 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1691 ciInstanceKlass* ik = base_k->as_instance_klass();
1692 // See if we can become precise: no subklasses and no interface
1693 if (!ik->is_interface() && !ik->has_subklass()) {
1694 //assert(!UseExactTypes, "this code should be useless with exact types");
1695 // Add a dependence; if any subclass added we need to recompile
1696 if (!ik->is_final()) {
1697 phase->C->dependencies()->assert_leaf_type(ik);
1698 }
1699 // Return precise array klass
1700 return TypeKlassPtr::make(ak);
1701 }
1702 }
1703 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1704 } else { // Found a type-array?
1705 //assert(!UseExactTypes, "this code should be useless with exact types");
1706 assert( ak->is_type_array_klass(), "" );
1707 return TypeKlassPtr::make(ak); // These are always precise
1708 }
1709 }
1710 }
1712 // Check for loading klass from an array klass
1713 const TypeKlassPtr *tkls = tp->isa_klassptr();
1714 if (tkls != NULL && !StressReflectiveCode) {
1715 ciKlass* klass = tkls->klass();
1716 if( !klass->is_loaded() )
1717 return _type; // Bail out if not loaded
1718 if( klass->is_obj_array_klass() &&
1719 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1720 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1721 // // Always returning precise element type is incorrect,
1722 // // e.g., element type could be object and array may contain strings
1723 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1725 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1726 // according to the element type's subclassing.
1727 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1728 }
1729 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1730 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1731 ciKlass* sup = klass->as_instance_klass()->super();
1732 // The field is Klass::_super. Return its (constant) value.
1733 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1734 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1735 }
1736 }
1738 // Bailout case
1739 return LoadNode::Value(phase);
1740 }
1742 //------------------------------Identity---------------------------------------
1743 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1744 // Also feed through the klass in Allocate(...klass...)._klass.
1745 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1746 Node* x = LoadNode::Identity(phase);
1747 if (x != this) return x;
1749 // Take apart the address into an oop and and offset.
1750 // Return 'this' if we cannot.
1751 Node* adr = in(MemNode::Address);
1752 intptr_t offset = 0;
1753 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1754 if (base == NULL) return this;
1755 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1756 if (toop == NULL) return this;
1758 // We can fetch the klass directly through an AllocateNode.
1759 // This works even if the klass is not constant (clone or newArray).
1760 if (offset == oopDesc::klass_offset_in_bytes()) {
1761 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1762 if (allocated_klass != NULL) {
1763 return allocated_klass;
1764 }
1765 }
1767 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1768 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1769 // See inline_native_Class_query for occurrences of these patterns.
1770 // Java Example: x.getClass().isAssignableFrom(y)
1771 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1772 //
1773 // This improves reflective code, often making the Class
1774 // mirror go completely dead. (Current exception: Class
1775 // mirrors may appear in debug info, but we could clean them out by
1776 // introducing a new debug info operator for klassOop.java_mirror).
1777 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1778 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1779 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1780 // We are loading a special hidden field from a Class mirror,
1781 // the field which points to its Klass or arrayKlass metaobject.
1782 if (base->is_Load()) {
1783 Node* adr2 = base->in(MemNode::Address);
1784 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1785 if (tkls != NULL && !tkls->empty()
1786 && (tkls->klass()->is_instance_klass() ||
1787 tkls->klass()->is_array_klass())
1788 && adr2->is_AddP()
1789 ) {
1790 int mirror_field = Klass::java_mirror_offset_in_bytes();
1791 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1792 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1793 }
1794 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1795 return adr2->in(AddPNode::Base);
1796 }
1797 }
1798 }
1799 }
1801 return this;
1802 }
1804 //------------------------------Value-----------------------------------------
1805 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1806 // Either input is TOP ==> the result is TOP
1807 const Type *t1 = phase->type( in(MemNode::Memory) );
1808 if( t1 == Type::TOP ) return Type::TOP;
1809 Node *adr = in(MemNode::Address);
1810 const Type *t2 = phase->type( adr );
1811 if( t2 == Type::TOP ) return Type::TOP;
1812 const TypePtr *tp = t2->is_ptr();
1813 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1814 const TypeAryPtr *tap = tp->isa_aryptr();
1815 if( !tap ) return _type;
1816 return tap->size();
1817 }
1819 //------------------------------Identity---------------------------------------
1820 // Feed through the length in AllocateArray(...length...)._length.
1821 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1822 Node* x = LoadINode::Identity(phase);
1823 if (x != this) return x;
1825 // Take apart the address into an oop and and offset.
1826 // Return 'this' if we cannot.
1827 Node* adr = in(MemNode::Address);
1828 intptr_t offset = 0;
1829 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1830 if (base == NULL) return this;
1831 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1832 if (tary == NULL) return this;
1834 // We can fetch the length directly through an AllocateArrayNode.
1835 // This works even if the length is not constant (clone or newArray).
1836 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1837 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
1838 if (allocated_length != NULL) {
1839 return allocated_length;
1840 }
1841 }
1843 return this;
1845 }
1846 //=============================================================================
1847 //---------------------------StoreNode::make-----------------------------------
1848 // Polymorphic factory method:
1849 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1850 Compile* C = gvn.C;
1852 switch (bt) {
1853 case T_BOOLEAN:
1854 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1855 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1856 case T_CHAR:
1857 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1858 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1859 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1860 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1861 case T_ADDRESS:
1862 case T_OBJECT:
1863 #ifdef _LP64
1864 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
1865 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
1866 adr->bottom_type()->isa_rawptr())) {
1867 const TypePtr* type = val->bottom_type()->is_ptr();
1868 Node* cp = EncodePNode::encode(&gvn, val);
1869 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, cp);
1870 } else
1871 #endif
1872 {
1873 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1874 }
1875 }
1876 ShouldNotReachHere();
1877 return (StoreNode*)NULL;
1878 }
1880 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1881 bool require_atomic = true;
1882 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1883 }
1886 //--------------------------bottom_type----------------------------------------
1887 const Type *StoreNode::bottom_type() const {
1888 return Type::MEMORY;
1889 }
1891 //------------------------------hash-------------------------------------------
1892 uint StoreNode::hash() const {
1893 // unroll addition of interesting fields
1894 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1896 // Since they are not commoned, do not hash them:
1897 return NO_HASH;
1898 }
1900 //------------------------------Ideal------------------------------------------
1901 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1902 // When a store immediately follows a relevant allocation/initialization,
1903 // try to capture it into the initialization, or hoist it above.
1904 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1905 Node* p = MemNode::Ideal_common(phase, can_reshape);
1906 if (p) return (p == NodeSentinel) ? NULL : p;
1908 Node* mem = in(MemNode::Memory);
1909 Node* address = in(MemNode::Address);
1911 // Back-to-back stores to same address? Fold em up.
1912 // Generally unsafe if I have intervening uses...
1913 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
1914 // Looking at a dead closed cycle of memory?
1915 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1917 assert(Opcode() == mem->Opcode() ||
1918 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1919 "no mismatched stores, except on raw memory");
1921 if (mem->outcnt() == 1 && // check for intervening uses
1922 mem->as_Store()->memory_size() <= this->memory_size()) {
1923 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1924 // For example, 'mem' might be the final state at a conditional return.
1925 // Or, 'mem' might be used by some node which is live at the same time
1926 // 'this' is live, which might be unschedulable. So, require exactly
1927 // ONE user, the 'this' store, until such time as we clone 'mem' for
1928 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1929 if (can_reshape) { // (%%% is this an anachronism?)
1930 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1931 phase->is_IterGVN());
1932 } else {
1933 // It's OK to do this in the parser, since DU info is always accurate,
1934 // and the parser always refers to nodes via SafePointNode maps.
1935 set_req(MemNode::Memory, mem->in(MemNode::Memory));
1936 }
1937 return this;
1938 }
1939 }
1941 // Capture an unaliased, unconditional, simple store into an initializer.
1942 // Or, if it is independent of the allocation, hoist it above the allocation.
1943 if (ReduceFieldZeroing && /*can_reshape &&*/
1944 mem->is_Proj() && mem->in(0)->is_Initialize()) {
1945 InitializeNode* init = mem->in(0)->as_Initialize();
1946 intptr_t offset = init->can_capture_store(this, phase);
1947 if (offset > 0) {
1948 Node* moved = init->capture_store(this, offset, phase);
1949 // If the InitializeNode captured me, it made a raw copy of me,
1950 // and I need to disappear.
1951 if (moved != NULL) {
1952 // %%% hack to ensure that Ideal returns a new node:
1953 mem = MergeMemNode::make(phase->C, mem);
1954 return mem; // fold me away
1955 }
1956 }
1957 }
1959 return NULL; // No further progress
1960 }
1962 //------------------------------Value-----------------------------------------
1963 const Type *StoreNode::Value( PhaseTransform *phase ) const {
1964 // Either input is TOP ==> the result is TOP
1965 const Type *t1 = phase->type( in(MemNode::Memory) );
1966 if( t1 == Type::TOP ) return Type::TOP;
1967 const Type *t2 = phase->type( in(MemNode::Address) );
1968 if( t2 == Type::TOP ) return Type::TOP;
1969 const Type *t3 = phase->type( in(MemNode::ValueIn) );
1970 if( t3 == Type::TOP ) return Type::TOP;
1971 return Type::MEMORY;
1972 }
1974 //------------------------------Identity---------------------------------------
1975 // Remove redundant stores:
1976 // Store(m, p, Load(m, p)) changes to m.
1977 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
1978 Node *StoreNode::Identity( PhaseTransform *phase ) {
1979 Node* mem = in(MemNode::Memory);
1980 Node* adr = in(MemNode::Address);
1981 Node* val = in(MemNode::ValueIn);
1983 // Load then Store? Then the Store is useless
1984 if (val->is_Load() &&
1985 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
1986 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
1987 val->as_Load()->store_Opcode() == Opcode()) {
1988 return mem;
1989 }
1991 // Two stores in a row of the same value?
1992 if (mem->is_Store() &&
1993 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
1994 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
1995 mem->Opcode() == Opcode()) {
1996 return mem;
1997 }
1999 // Store of zero anywhere into a freshly-allocated object?
2000 // Then the store is useless.
2001 // (It must already have been captured by the InitializeNode.)
2002 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2003 // a newly allocated object is already all-zeroes everywhere
2004 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2005 return mem;
2006 }
2008 // the store may also apply to zero-bits in an earlier object
2009 Node* prev_mem = find_previous_store(phase);
2010 // Steps (a), (b): Walk past independent stores to find an exact match.
2011 if (prev_mem != NULL) {
2012 Node* prev_val = can_see_stored_value(prev_mem, phase);
2013 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2014 // prev_val and val might differ by a cast; it would be good
2015 // to keep the more informative of the two.
2016 return mem;
2017 }
2018 }
2019 }
2021 return this;
2022 }
2024 //------------------------------match_edge-------------------------------------
2025 // Do we Match on this edge index or not? Match only memory & value
2026 uint StoreNode::match_edge(uint idx) const {
2027 return idx == MemNode::Address || idx == MemNode::ValueIn;
2028 }
2030 //------------------------------cmp--------------------------------------------
2031 // Do not common stores up together. They generally have to be split
2032 // back up anyways, so do not bother.
2033 uint StoreNode::cmp( const Node &n ) const {
2034 return (&n == this); // Always fail except on self
2035 }
2037 //------------------------------Ideal_masked_input-----------------------------
2038 // Check for a useless mask before a partial-word store
2039 // (StoreB ... (AndI valIn conIa) )
2040 // If (conIa & mask == mask) this simplifies to
2041 // (StoreB ... (valIn) )
2042 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2043 Node *val = in(MemNode::ValueIn);
2044 if( val->Opcode() == Op_AndI ) {
2045 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2046 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2047 set_req(MemNode::ValueIn, val->in(1));
2048 return this;
2049 }
2050 }
2051 return NULL;
2052 }
2055 //------------------------------Ideal_sign_extended_input----------------------
2056 // Check for useless sign-extension before a partial-word store
2057 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2058 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2059 // (StoreB ... (valIn) )
2060 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2061 Node *val = in(MemNode::ValueIn);
2062 if( val->Opcode() == Op_RShiftI ) {
2063 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2064 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2065 Node *shl = val->in(1);
2066 if( shl->Opcode() == Op_LShiftI ) {
2067 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2068 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2069 set_req(MemNode::ValueIn, shl->in(1));
2070 return this;
2071 }
2072 }
2073 }
2074 }
2075 return NULL;
2076 }
2078 //------------------------------value_never_loaded-----------------------------------
2079 // Determine whether there are any possible loads of the value stored.
2080 // For simplicity, we actually check if there are any loads from the
2081 // address stored to, not just for loads of the value stored by this node.
2082 //
2083 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2084 Node *adr = in(Address);
2085 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2086 if (adr_oop == NULL)
2087 return false;
2088 if (!adr_oop->is_instance_field())
2089 return false; // if not a distinct instance, there may be aliases of the address
2090 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2091 Node *use = adr->fast_out(i);
2092 int opc = use->Opcode();
2093 if (use->is_Load() || use->is_LoadStore()) {
2094 return false;
2095 }
2096 }
2097 return true;
2098 }
2100 //=============================================================================
2101 //------------------------------Ideal------------------------------------------
2102 // If the store is from an AND mask that leaves the low bits untouched, then
2103 // we can skip the AND operation. If the store is from a sign-extension
2104 // (a left shift, then right shift) we can skip both.
2105 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2106 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2107 if( progress != NULL ) return progress;
2109 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2110 if( progress != NULL ) return progress;
2112 // Finally check the default case
2113 return StoreNode::Ideal(phase, can_reshape);
2114 }
2116 //=============================================================================
2117 //------------------------------Ideal------------------------------------------
2118 // If the store is from an AND mask that leaves the low bits untouched, then
2119 // we can skip the AND operation
2120 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2121 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2122 if( progress != NULL ) return progress;
2124 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2125 if( progress != NULL ) return progress;
2127 // Finally check the default case
2128 return StoreNode::Ideal(phase, can_reshape);
2129 }
2131 //=============================================================================
2132 //------------------------------Identity---------------------------------------
2133 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2134 // No need to card mark when storing a null ptr
2135 Node* my_store = in(MemNode::OopStore);
2136 if (my_store->is_Store()) {
2137 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2138 if( t1 == TypePtr::NULL_PTR ) {
2139 return in(MemNode::Memory);
2140 }
2141 }
2142 return this;
2143 }
2145 //------------------------------Value-----------------------------------------
2146 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2147 // Either input is TOP ==> the result is TOP
2148 const Type *t = phase->type( in(MemNode::Memory) );
2149 if( t == Type::TOP ) return Type::TOP;
2150 t = phase->type( in(MemNode::Address) );
2151 if( t == Type::TOP ) return Type::TOP;
2152 t = phase->type( in(MemNode::ValueIn) );
2153 if( t == Type::TOP ) return Type::TOP;
2154 // If extra input is TOP ==> the result is TOP
2155 t = phase->type( in(MemNode::OopStore) );
2156 if( t == Type::TOP ) return Type::TOP;
2158 return StoreNode::Value( phase );
2159 }
2162 //=============================================================================
2163 //----------------------------------SCMemProjNode------------------------------
2164 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2165 {
2166 return bottom_type();
2167 }
2169 //=============================================================================
2170 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2171 init_req(MemNode::Control, c );
2172 init_req(MemNode::Memory , mem);
2173 init_req(MemNode::Address, adr);
2174 init_req(MemNode::ValueIn, val);
2175 init_req( ExpectedIn, ex );
2176 init_class_id(Class_LoadStore);
2178 }
2180 //=============================================================================
2181 //-------------------------------adr_type--------------------------------------
2182 // Do we Match on this edge index or not? Do not match memory
2183 const TypePtr* ClearArrayNode::adr_type() const {
2184 Node *adr = in(3);
2185 return MemNode::calculate_adr_type(adr->bottom_type());
2186 }
2188 //------------------------------match_edge-------------------------------------
2189 // Do we Match on this edge index or not? Do not match memory
2190 uint ClearArrayNode::match_edge(uint idx) const {
2191 return idx > 1;
2192 }
2194 //------------------------------Identity---------------------------------------
2195 // Clearing a zero length array does nothing
2196 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2197 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2198 }
2200 //------------------------------Idealize---------------------------------------
2201 // Clearing a short array is faster with stores
2202 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2203 const int unit = BytesPerLong;
2204 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2205 if (!t) return NULL;
2206 if (!t->is_con()) return NULL;
2207 intptr_t raw_count = t->get_con();
2208 intptr_t size = raw_count;
2209 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2210 // Clearing nothing uses the Identity call.
2211 // Negative clears are possible on dead ClearArrays
2212 // (see jck test stmt114.stmt11402.val).
2213 if (size <= 0 || size % unit != 0) return NULL;
2214 intptr_t count = size / unit;
2215 // Length too long; use fast hardware clear
2216 if (size > Matcher::init_array_short_size) return NULL;
2217 Node *mem = in(1);
2218 if( phase->type(mem)==Type::TOP ) return NULL;
2219 Node *adr = in(3);
2220 const Type* at = phase->type(adr);
2221 if( at==Type::TOP ) return NULL;
2222 const TypePtr* atp = at->isa_ptr();
2223 // adjust atp to be the correct array element address type
2224 if (atp == NULL) atp = TypePtr::BOTTOM;
2225 else atp = atp->add_offset(Type::OffsetBot);
2226 // Get base for derived pointer purposes
2227 if( adr->Opcode() != Op_AddP ) Unimplemented();
2228 Node *base = adr->in(1);
2230 Node *zero = phase->makecon(TypeLong::ZERO);
2231 Node *off = phase->MakeConX(BytesPerLong);
2232 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2233 count--;
2234 while( count-- ) {
2235 mem = phase->transform(mem);
2236 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2237 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2238 }
2239 return mem;
2240 }
2242 //----------------------------clear_memory-------------------------------------
2243 // Generate code to initialize object storage to zero.
2244 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2245 intptr_t start_offset,
2246 Node* end_offset,
2247 PhaseGVN* phase) {
2248 Compile* C = phase->C;
2249 intptr_t offset = start_offset;
2251 int unit = BytesPerLong;
2252 if ((offset % unit) != 0) {
2253 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2254 adr = phase->transform(adr);
2255 const TypePtr* atp = TypeRawPtr::BOTTOM;
2256 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2257 mem = phase->transform(mem);
2258 offset += BytesPerInt;
2259 }
2260 assert((offset % unit) == 0, "");
2262 // Initialize the remaining stuff, if any, with a ClearArray.
2263 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2264 }
2266 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2267 Node* start_offset,
2268 Node* end_offset,
2269 PhaseGVN* phase) {
2270 if (start_offset == end_offset) {
2271 // nothing to do
2272 return mem;
2273 }
2275 Compile* C = phase->C;
2276 int unit = BytesPerLong;
2277 Node* zbase = start_offset;
2278 Node* zend = end_offset;
2280 // Scale to the unit required by the CPU:
2281 if (!Matcher::init_array_count_is_in_bytes) {
2282 Node* shift = phase->intcon(exact_log2(unit));
2283 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2284 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2285 }
2287 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2288 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2290 // Bulk clear double-words
2291 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2292 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2293 return phase->transform(mem);
2294 }
2296 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2297 intptr_t start_offset,
2298 intptr_t end_offset,
2299 PhaseGVN* phase) {
2300 if (start_offset == end_offset) {
2301 // nothing to do
2302 return mem;
2303 }
2305 Compile* C = phase->C;
2306 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2307 intptr_t done_offset = end_offset;
2308 if ((done_offset % BytesPerLong) != 0) {
2309 done_offset -= BytesPerInt;
2310 }
2311 if (done_offset > start_offset) {
2312 mem = clear_memory(ctl, mem, dest,
2313 start_offset, phase->MakeConX(done_offset), phase);
2314 }
2315 if (done_offset < end_offset) { // emit the final 32-bit store
2316 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2317 adr = phase->transform(adr);
2318 const TypePtr* atp = TypeRawPtr::BOTTOM;
2319 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2320 mem = phase->transform(mem);
2321 done_offset += BytesPerInt;
2322 }
2323 assert(done_offset == end_offset, "");
2324 return mem;
2325 }
2327 //=============================================================================
2328 // Do we match on this edge? No memory edges
2329 uint StrCompNode::match_edge(uint idx) const {
2330 return idx == 5 || idx == 6;
2331 }
2333 //------------------------------Ideal------------------------------------------
2334 // Return a node which is more "ideal" than the current node. Strip out
2335 // control copies
2336 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2337 return remove_dead_region(phase, can_reshape) ? this : NULL;
2338 }
2341 //=============================================================================
2342 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2343 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2344 _adr_type(C->get_adr_type(alias_idx))
2345 {
2346 init_class_id(Class_MemBar);
2347 Node* top = C->top();
2348 init_req(TypeFunc::I_O,top);
2349 init_req(TypeFunc::FramePtr,top);
2350 init_req(TypeFunc::ReturnAdr,top);
2351 if (precedent != NULL)
2352 init_req(TypeFunc::Parms, precedent);
2353 }
2355 //------------------------------cmp--------------------------------------------
2356 uint MemBarNode::hash() const { return NO_HASH; }
2357 uint MemBarNode::cmp( const Node &n ) const {
2358 return (&n == this); // Always fail except on self
2359 }
2361 //------------------------------make-------------------------------------------
2362 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2363 int len = Precedent + (pn == NULL? 0: 1);
2364 switch (opcode) {
2365 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2366 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2367 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2368 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2369 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2370 default: ShouldNotReachHere(); return NULL;
2371 }
2372 }
2374 //------------------------------Ideal------------------------------------------
2375 // Return a node which is more "ideal" than the current node. Strip out
2376 // control copies
2377 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2378 if (remove_dead_region(phase, can_reshape)) return this;
2379 return NULL;
2380 }
2382 //------------------------------Value------------------------------------------
2383 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2384 if( !in(0) ) return Type::TOP;
2385 if( phase->type(in(0)) == Type::TOP )
2386 return Type::TOP;
2387 return TypeTuple::MEMBAR;
2388 }
2390 //------------------------------match------------------------------------------
2391 // Construct projections for memory.
2392 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2393 switch (proj->_con) {
2394 case TypeFunc::Control:
2395 case TypeFunc::Memory:
2396 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2397 }
2398 ShouldNotReachHere();
2399 return NULL;
2400 }
2402 //===========================InitializeNode====================================
2403 // SUMMARY:
2404 // This node acts as a memory barrier on raw memory, after some raw stores.
2405 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2406 // The Initialize can 'capture' suitably constrained stores as raw inits.
2407 // It can coalesce related raw stores into larger units (called 'tiles').
2408 // It can avoid zeroing new storage for memory units which have raw inits.
2409 // At macro-expansion, it is marked 'complete', and does not optimize further.
2410 //
2411 // EXAMPLE:
2412 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2413 // ctl = incoming control; mem* = incoming memory
2414 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2415 // First allocate uninitialized memory and fill in the header:
2416 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2417 // ctl := alloc.Control; mem* := alloc.Memory*
2418 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2419 // Then initialize to zero the non-header parts of the raw memory block:
2420 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2421 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2422 // After the initialize node executes, the object is ready for service:
2423 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2424 // Suppose its body is immediately initialized as {1,2}:
2425 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2426 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2427 // mem.SLICE(#short[*]) := store2
2428 //
2429 // DETAILS:
2430 // An InitializeNode collects and isolates object initialization after
2431 // an AllocateNode and before the next possible safepoint. As a
2432 // memory barrier (MemBarNode), it keeps critical stores from drifting
2433 // down past any safepoint or any publication of the allocation.
2434 // Before this barrier, a newly-allocated object may have uninitialized bits.
2435 // After this barrier, it may be treated as a real oop, and GC is allowed.
2436 //
2437 // The semantics of the InitializeNode include an implicit zeroing of
2438 // the new object from object header to the end of the object.
2439 // (The object header and end are determined by the AllocateNode.)
2440 //
2441 // Certain stores may be added as direct inputs to the InitializeNode.
2442 // These stores must update raw memory, and they must be to addresses
2443 // derived from the raw address produced by AllocateNode, and with
2444 // a constant offset. They must be ordered by increasing offset.
2445 // The first one is at in(RawStores), the last at in(req()-1).
2446 // Unlike most memory operations, they are not linked in a chain,
2447 // but are displayed in parallel as users of the rawmem output of
2448 // the allocation.
2449 //
2450 // (See comments in InitializeNode::capture_store, which continue
2451 // the example given above.)
2452 //
2453 // When the associated Allocate is macro-expanded, the InitializeNode
2454 // may be rewritten to optimize collected stores. A ClearArrayNode
2455 // may also be created at that point to represent any required zeroing.
2456 // The InitializeNode is then marked 'complete', prohibiting further
2457 // capturing of nearby memory operations.
2458 //
2459 // During macro-expansion, all captured initializations which store
2460 // constant values of 32 bits or smaller are coalesced (if advantagous)
2461 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2462 // initialized in fewer memory operations. Memory words which are
2463 // covered by neither tiles nor non-constant stores are pre-zeroed
2464 // by explicit stores of zero. (The code shape happens to do all
2465 // zeroing first, then all other stores, with both sequences occurring
2466 // in order of ascending offsets.)
2467 //
2468 // Alternatively, code may be inserted between an AllocateNode and its
2469 // InitializeNode, to perform arbitrary initialization of the new object.
2470 // E.g., the object copying intrinsics insert complex data transfers here.
2471 // The initialization must then be marked as 'complete' disable the
2472 // built-in zeroing semantics and the collection of initializing stores.
2473 //
2474 // While an InitializeNode is incomplete, reads from the memory state
2475 // produced by it are optimizable if they match the control edge and
2476 // new oop address associated with the allocation/initialization.
2477 // They return a stored value (if the offset matches) or else zero.
2478 // A write to the memory state, if it matches control and address,
2479 // and if it is to a constant offset, may be 'captured' by the
2480 // InitializeNode. It is cloned as a raw memory operation and rewired
2481 // inside the initialization, to the raw oop produced by the allocation.
2482 // Operations on addresses which are provably distinct (e.g., to
2483 // other AllocateNodes) are allowed to bypass the initialization.
2484 //
2485 // The effect of all this is to consolidate object initialization
2486 // (both arrays and non-arrays, both piecewise and bulk) into a
2487 // single location, where it can be optimized as a unit.
2488 //
2489 // Only stores with an offset less than TrackedInitializationLimit words
2490 // will be considered for capture by an InitializeNode. This puts a
2491 // reasonable limit on the complexity of optimized initializations.
2493 //---------------------------InitializeNode------------------------------------
2494 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2495 : _is_complete(false),
2496 MemBarNode(C, adr_type, rawoop)
2497 {
2498 init_class_id(Class_Initialize);
2500 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2501 assert(in(RawAddress) == rawoop, "proper init");
2502 // Note: allocation() can be NULL, for secondary initialization barriers
2503 }
2505 // Since this node is not matched, it will be processed by the
2506 // register allocator. Declare that there are no constraints
2507 // on the allocation of the RawAddress edge.
2508 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2509 // This edge should be set to top, by the set_complete. But be conservative.
2510 if (idx == InitializeNode::RawAddress)
2511 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2512 return RegMask::Empty;
2513 }
2515 Node* InitializeNode::memory(uint alias_idx) {
2516 Node* mem = in(Memory);
2517 if (mem->is_MergeMem()) {
2518 return mem->as_MergeMem()->memory_at(alias_idx);
2519 } else {
2520 // incoming raw memory is not split
2521 return mem;
2522 }
2523 }
2525 bool InitializeNode::is_non_zero() {
2526 if (is_complete()) return false;
2527 remove_extra_zeroes();
2528 return (req() > RawStores);
2529 }
2531 void InitializeNode::set_complete(PhaseGVN* phase) {
2532 assert(!is_complete(), "caller responsibility");
2533 _is_complete = true;
2535 // After this node is complete, it contains a bunch of
2536 // raw-memory initializations. There is no need for
2537 // it to have anything to do with non-raw memory effects.
2538 // Therefore, tell all non-raw users to re-optimize themselves,
2539 // after skipping the memory effects of this initialization.
2540 PhaseIterGVN* igvn = phase->is_IterGVN();
2541 if (igvn) igvn->add_users_to_worklist(this);
2542 }
2544 // convenience function
2545 // return false if the init contains any stores already
2546 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2547 InitializeNode* init = initialization();
2548 if (init == NULL || init->is_complete()) return false;
2549 init->remove_extra_zeroes();
2550 // for now, if this allocation has already collected any inits, bail:
2551 if (init->is_non_zero()) return false;
2552 init->set_complete(phase);
2553 return true;
2554 }
2556 void InitializeNode::remove_extra_zeroes() {
2557 if (req() == RawStores) return;
2558 Node* zmem = zero_memory();
2559 uint fill = RawStores;
2560 for (uint i = fill; i < req(); i++) {
2561 Node* n = in(i);
2562 if (n->is_top() || n == zmem) continue; // skip
2563 if (fill < i) set_req(fill, n); // compact
2564 ++fill;
2565 }
2566 // delete any empty spaces created:
2567 while (fill < req()) {
2568 del_req(fill);
2569 }
2570 }
2572 // Helper for remembering which stores go with which offsets.
2573 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2574 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2575 intptr_t offset = -1;
2576 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2577 phase, offset);
2578 if (base == NULL) return -1; // something is dead,
2579 if (offset < 0) return -1; // dead, dead
2580 return offset;
2581 }
2583 // Helper for proving that an initialization expression is
2584 // "simple enough" to be folded into an object initialization.
2585 // Attempts to prove that a store's initial value 'n' can be captured
2586 // within the initialization without creating a vicious cycle, such as:
2587 // { Foo p = new Foo(); p.next = p; }
2588 // True for constants and parameters and small combinations thereof.
2589 bool InitializeNode::detect_init_independence(Node* n,
2590 bool st_is_pinned,
2591 int& count) {
2592 if (n == NULL) return true; // (can this really happen?)
2593 if (n->is_Proj()) n = n->in(0);
2594 if (n == this) return false; // found a cycle
2595 if (n->is_Con()) return true;
2596 if (n->is_Start()) return true; // params, etc., are OK
2597 if (n->is_Root()) return true; // even better
2599 Node* ctl = n->in(0);
2600 if (ctl != NULL && !ctl->is_top()) {
2601 if (ctl->is_Proj()) ctl = ctl->in(0);
2602 if (ctl == this) return false;
2604 // If we already know that the enclosing memory op is pinned right after
2605 // the init, then any control flow that the store has picked up
2606 // must have preceded the init, or else be equal to the init.
2607 // Even after loop optimizations (which might change control edges)
2608 // a store is never pinned *before* the availability of its inputs.
2609 if (!MemNode::all_controls_dominate(n, this))
2610 return false; // failed to prove a good control
2612 }
2614 // Check data edges for possible dependencies on 'this'.
2615 if ((count += 1) > 20) return false; // complexity limit
2616 for (uint i = 1; i < n->req(); i++) {
2617 Node* m = n->in(i);
2618 if (m == NULL || m == n || m->is_top()) continue;
2619 uint first_i = n->find_edge(m);
2620 if (i != first_i) continue; // process duplicate edge just once
2621 if (!detect_init_independence(m, st_is_pinned, count)) {
2622 return false;
2623 }
2624 }
2626 return true;
2627 }
2629 // Here are all the checks a Store must pass before it can be moved into
2630 // an initialization. Returns zero if a check fails.
2631 // On success, returns the (constant) offset to which the store applies,
2632 // within the initialized memory.
2633 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2634 const int FAIL = 0;
2635 if (st->req() != MemNode::ValueIn + 1)
2636 return FAIL; // an inscrutable StoreNode (card mark?)
2637 Node* ctl = st->in(MemNode::Control);
2638 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2639 return FAIL; // must be unconditional after the initialization
2640 Node* mem = st->in(MemNode::Memory);
2641 if (!(mem->is_Proj() && mem->in(0) == this))
2642 return FAIL; // must not be preceded by other stores
2643 Node* adr = st->in(MemNode::Address);
2644 intptr_t offset;
2645 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2646 if (alloc == NULL)
2647 return FAIL; // inscrutable address
2648 if (alloc != allocation())
2649 return FAIL; // wrong allocation! (store needs to float up)
2650 Node* val = st->in(MemNode::ValueIn);
2651 int complexity_count = 0;
2652 if (!detect_init_independence(val, true, complexity_count))
2653 return FAIL; // stored value must be 'simple enough'
2655 return offset; // success
2656 }
2658 // Find the captured store in(i) which corresponds to the range
2659 // [start..start+size) in the initialized object.
2660 // If there is one, return its index i. If there isn't, return the
2661 // negative of the index where it should be inserted.
2662 // Return 0 if the queried range overlaps an initialization boundary
2663 // or if dead code is encountered.
2664 // If size_in_bytes is zero, do not bother with overlap checks.
2665 int InitializeNode::captured_store_insertion_point(intptr_t start,
2666 int size_in_bytes,
2667 PhaseTransform* phase) {
2668 const int FAIL = 0, MAX_STORE = BytesPerLong;
2670 if (is_complete())
2671 return FAIL; // arraycopy got here first; punt
2673 assert(allocation() != NULL, "must be present");
2675 // no negatives, no header fields:
2676 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2678 // after a certain size, we bail out on tracking all the stores:
2679 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2680 if (start >= ti_limit) return FAIL;
2682 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2683 if (i >= limit) return -(int)i; // not found; here is where to put it
2685 Node* st = in(i);
2686 intptr_t st_off = get_store_offset(st, phase);
2687 if (st_off < 0) {
2688 if (st != zero_memory()) {
2689 return FAIL; // bail out if there is dead garbage
2690 }
2691 } else if (st_off > start) {
2692 // ...we are done, since stores are ordered
2693 if (st_off < start + size_in_bytes) {
2694 return FAIL; // the next store overlaps
2695 }
2696 return -(int)i; // not found; here is where to put it
2697 } else if (st_off < start) {
2698 if (size_in_bytes != 0 &&
2699 start < st_off + MAX_STORE &&
2700 start < st_off + st->as_Store()->memory_size()) {
2701 return FAIL; // the previous store overlaps
2702 }
2703 } else {
2704 if (size_in_bytes != 0 &&
2705 st->as_Store()->memory_size() != size_in_bytes) {
2706 return FAIL; // mismatched store size
2707 }
2708 return i;
2709 }
2711 ++i;
2712 }
2713 }
2715 // Look for a captured store which initializes at the offset 'start'
2716 // with the given size. If there is no such store, and no other
2717 // initialization interferes, then return zero_memory (the memory
2718 // projection of the AllocateNode).
2719 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2720 PhaseTransform* phase) {
2721 assert(stores_are_sane(phase), "");
2722 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2723 if (i == 0) {
2724 return NULL; // something is dead
2725 } else if (i < 0) {
2726 return zero_memory(); // just primordial zero bits here
2727 } else {
2728 Node* st = in(i); // here is the store at this position
2729 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2730 return st;
2731 }
2732 }
2734 // Create, as a raw pointer, an address within my new object at 'offset'.
2735 Node* InitializeNode::make_raw_address(intptr_t offset,
2736 PhaseTransform* phase) {
2737 Node* addr = in(RawAddress);
2738 if (offset != 0) {
2739 Compile* C = phase->C;
2740 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2741 phase->MakeConX(offset)) );
2742 }
2743 return addr;
2744 }
2746 // Clone the given store, converting it into a raw store
2747 // initializing a field or element of my new object.
2748 // Caller is responsible for retiring the original store,
2749 // with subsume_node or the like.
2750 //
2751 // From the example above InitializeNode::InitializeNode,
2752 // here are the old stores to be captured:
2753 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2754 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2755 //
2756 // Here is the changed code; note the extra edges on init:
2757 // alloc = (Allocate ...)
2758 // rawoop = alloc.RawAddress
2759 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2760 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2761 // init = (Initialize alloc.Control alloc.Memory rawoop
2762 // rawstore1 rawstore2)
2763 //
2764 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2765 PhaseTransform* phase) {
2766 assert(stores_are_sane(phase), "");
2768 if (start < 0) return NULL;
2769 assert(can_capture_store(st, phase) == start, "sanity");
2771 Compile* C = phase->C;
2772 int size_in_bytes = st->memory_size();
2773 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2774 if (i == 0) return NULL; // bail out
2775 Node* prev_mem = NULL; // raw memory for the captured store
2776 if (i > 0) {
2777 prev_mem = in(i); // there is a pre-existing store under this one
2778 set_req(i, C->top()); // temporarily disconnect it
2779 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2780 } else {
2781 i = -i; // no pre-existing store
2782 prev_mem = zero_memory(); // a slice of the newly allocated object
2783 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2784 set_req(--i, C->top()); // reuse this edge; it has been folded away
2785 else
2786 ins_req(i, C->top()); // build a new edge
2787 }
2788 Node* new_st = st->clone();
2789 new_st->set_req(MemNode::Control, in(Control));
2790 new_st->set_req(MemNode::Memory, prev_mem);
2791 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2792 new_st = phase->transform(new_st);
2794 // At this point, new_st might have swallowed a pre-existing store
2795 // at the same offset, or perhaps new_st might have disappeared,
2796 // if it redundantly stored the same value (or zero to fresh memory).
2798 // In any case, wire it in:
2799 set_req(i, new_st);
2801 // The caller may now kill the old guy.
2802 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2803 assert(check_st == new_st || check_st == NULL, "must be findable");
2804 assert(!is_complete(), "");
2805 return new_st;
2806 }
2808 static bool store_constant(jlong* tiles, int num_tiles,
2809 intptr_t st_off, int st_size,
2810 jlong con) {
2811 if ((st_off & (st_size-1)) != 0)
2812 return false; // strange store offset (assume size==2**N)
2813 address addr = (address)tiles + st_off;
2814 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2815 switch (st_size) {
2816 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2817 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2818 case sizeof(jint): *(jint*) addr = (jint) con; break;
2819 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2820 default: return false; // strange store size (detect size!=2**N here)
2821 }
2822 return true; // return success to caller
2823 }
2825 // Coalesce subword constants into int constants and possibly
2826 // into long constants. The goal, if the CPU permits,
2827 // is to initialize the object with a small number of 64-bit tiles.
2828 // Also, convert floating-point constants to bit patterns.
2829 // Non-constants are not relevant to this pass.
2830 //
2831 // In terms of the running example on InitializeNode::InitializeNode
2832 // and InitializeNode::capture_store, here is the transformation
2833 // of rawstore1 and rawstore2 into rawstore12:
2834 // alloc = (Allocate ...)
2835 // rawoop = alloc.RawAddress
2836 // tile12 = 0x00010002
2837 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2838 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2839 //
2840 void
2841 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2842 Node* size_in_bytes,
2843 PhaseGVN* phase) {
2844 Compile* C = phase->C;
2846 assert(stores_are_sane(phase), "");
2847 // Note: After this pass, they are not completely sane,
2848 // since there may be some overlaps.
2850 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2852 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2853 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2854 size_limit = MIN2(size_limit, ti_limit);
2855 size_limit = align_size_up(size_limit, BytesPerLong);
2856 int num_tiles = size_limit / BytesPerLong;
2858 // allocate space for the tile map:
2859 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2860 jlong tiles_buf[small_len];
2861 Node* nodes_buf[small_len];
2862 jlong inits_buf[small_len];
2863 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2864 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2865 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2866 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2867 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2868 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2869 // tiles: exact bitwise model of all primitive constants
2870 // nodes: last constant-storing node subsumed into the tiles model
2871 // inits: which bytes (in each tile) are touched by any initializations
2873 //// Pass A: Fill in the tile model with any relevant stores.
2875 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2876 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2877 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2878 Node* zmem = zero_memory(); // initially zero memory state
2879 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2880 Node* st = in(i);
2881 intptr_t st_off = get_store_offset(st, phase);
2883 // Figure out the store's offset and constant value:
2884 if (st_off < header_size) continue; //skip (ignore header)
2885 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
2886 int st_size = st->as_Store()->memory_size();
2887 if (st_off + st_size > size_limit) break;
2889 // Record which bytes are touched, whether by constant or not.
2890 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
2891 continue; // skip (strange store size)
2893 const Type* val = phase->type(st->in(MemNode::ValueIn));
2894 if (!val->singleton()) continue; //skip (non-con store)
2895 BasicType type = val->basic_type();
2897 jlong con = 0;
2898 switch (type) {
2899 case T_INT: con = val->is_int()->get_con(); break;
2900 case T_LONG: con = val->is_long()->get_con(); break;
2901 case T_FLOAT: con = jint_cast(val->getf()); break;
2902 case T_DOUBLE: con = jlong_cast(val->getd()); break;
2903 default: continue; //skip (odd store type)
2904 }
2906 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
2907 st->Opcode() == Op_StoreL) {
2908 continue; // This StoreL is already optimal.
2909 }
2911 // Store down the constant.
2912 store_constant(tiles, num_tiles, st_off, st_size, con);
2914 intptr_t j = st_off >> LogBytesPerLong;
2916 if (type == T_INT && st_size == BytesPerInt
2917 && (st_off & BytesPerInt) == BytesPerInt) {
2918 jlong lcon = tiles[j];
2919 if (!Matcher::isSimpleConstant64(lcon) &&
2920 st->Opcode() == Op_StoreI) {
2921 // This StoreI is already optimal by itself.
2922 jint* intcon = (jint*) &tiles[j];
2923 intcon[1] = 0; // undo the store_constant()
2925 // If the previous store is also optimal by itself, back up and
2926 // undo the action of the previous loop iteration... if we can.
2927 // But if we can't, just let the previous half take care of itself.
2928 st = nodes[j];
2929 st_off -= BytesPerInt;
2930 con = intcon[0];
2931 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
2932 assert(st_off >= header_size, "still ignoring header");
2933 assert(get_store_offset(st, phase) == st_off, "must be");
2934 assert(in(i-1) == zmem, "must be");
2935 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
2936 assert(con == tcon->is_int()->get_con(), "must be");
2937 // Undo the effects of the previous loop trip, which swallowed st:
2938 intcon[0] = 0; // undo store_constant()
2939 set_req(i-1, st); // undo set_req(i, zmem)
2940 nodes[j] = NULL; // undo nodes[j] = st
2941 --old_subword; // undo ++old_subword
2942 }
2943 continue; // This StoreI is already optimal.
2944 }
2945 }
2947 // This store is not needed.
2948 set_req(i, zmem);
2949 nodes[j] = st; // record for the moment
2950 if (st_size < BytesPerLong) // something has changed
2951 ++old_subword; // includes int/float, but who's counting...
2952 else ++old_long;
2953 }
2955 if ((old_subword + old_long) == 0)
2956 return; // nothing more to do
2958 //// Pass B: Convert any non-zero tiles into optimal constant stores.
2959 // Be sure to insert them before overlapping non-constant stores.
2960 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
2961 for (int j = 0; j < num_tiles; j++) {
2962 jlong con = tiles[j];
2963 jlong init = inits[j];
2964 if (con == 0) continue;
2965 jint con0, con1; // split the constant, address-wise
2966 jint init0, init1; // split the init map, address-wise
2967 { union { jlong con; jint intcon[2]; } u;
2968 u.con = con;
2969 con0 = u.intcon[0];
2970 con1 = u.intcon[1];
2971 u.con = init;
2972 init0 = u.intcon[0];
2973 init1 = u.intcon[1];
2974 }
2976 Node* old = nodes[j];
2977 assert(old != NULL, "need the prior store");
2978 intptr_t offset = (j * BytesPerLong);
2980 bool split = !Matcher::isSimpleConstant64(con);
2982 if (offset < header_size) {
2983 assert(offset + BytesPerInt >= header_size, "second int counts");
2984 assert(*(jint*)&tiles[j] == 0, "junk in header");
2985 split = true; // only the second word counts
2986 // Example: int a[] = { 42 ... }
2987 } else if (con0 == 0 && init0 == -1) {
2988 split = true; // first word is covered by full inits
2989 // Example: int a[] = { ... foo(), 42 ... }
2990 } else if (con1 == 0 && init1 == -1) {
2991 split = true; // second word is covered by full inits
2992 // Example: int a[] = { ... 42, foo() ... }
2993 }
2995 // Here's a case where init0 is neither 0 nor -1:
2996 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
2997 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
2998 // In this case the tile is not split; it is (jlong)42.
2999 // The big tile is stored down, and then the foo() value is inserted.
3000 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3002 Node* ctl = old->in(MemNode::Control);
3003 Node* adr = make_raw_address(offset, phase);
3004 const TypePtr* atp = TypeRawPtr::BOTTOM;
3006 // One or two coalesced stores to plop down.
3007 Node* st[2];
3008 intptr_t off[2];
3009 int nst = 0;
3010 if (!split) {
3011 ++new_long;
3012 off[nst] = offset;
3013 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3014 phase->longcon(con), T_LONG);
3015 } else {
3016 // Omit either if it is a zero.
3017 if (con0 != 0) {
3018 ++new_int;
3019 off[nst] = offset;
3020 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3021 phase->intcon(con0), T_INT);
3022 }
3023 if (con1 != 0) {
3024 ++new_int;
3025 offset += BytesPerInt;
3026 adr = make_raw_address(offset, phase);
3027 off[nst] = offset;
3028 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3029 phase->intcon(con1), T_INT);
3030 }
3031 }
3033 // Insert second store first, then the first before the second.
3034 // Insert each one just before any overlapping non-constant stores.
3035 while (nst > 0) {
3036 Node* st1 = st[--nst];
3037 C->copy_node_notes_to(st1, old);
3038 st1 = phase->transform(st1);
3039 offset = off[nst];
3040 assert(offset >= header_size, "do not smash header");
3041 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3042 guarantee(ins_idx != 0, "must re-insert constant store");
3043 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3044 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3045 set_req(--ins_idx, st1);
3046 else
3047 ins_req(ins_idx, st1);
3048 }
3049 }
3051 if (PrintCompilation && WizardMode)
3052 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3053 old_subword, old_long, new_int, new_long);
3054 if (C->log() != NULL)
3055 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3056 old_subword, old_long, new_int, new_long);
3058 // Clean up any remaining occurrences of zmem:
3059 remove_extra_zeroes();
3060 }
3062 // Explore forward from in(start) to find the first fully initialized
3063 // word, and return its offset. Skip groups of subword stores which
3064 // together initialize full words. If in(start) is itself part of a
3065 // fully initialized word, return the offset of in(start). If there
3066 // are no following full-word stores, or if something is fishy, return
3067 // a negative value.
3068 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3069 int int_map = 0;
3070 intptr_t int_map_off = 0;
3071 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3073 for (uint i = start, limit = req(); i < limit; i++) {
3074 Node* st = in(i);
3076 intptr_t st_off = get_store_offset(st, phase);
3077 if (st_off < 0) break; // return conservative answer
3079 int st_size = st->as_Store()->memory_size();
3080 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3081 return st_off; // we found a complete word init
3082 }
3084 // update the map:
3086 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3087 if (this_int_off != int_map_off) {
3088 // reset the map:
3089 int_map = 0;
3090 int_map_off = this_int_off;
3091 }
3093 int subword_off = st_off - this_int_off;
3094 int_map |= right_n_bits(st_size) << subword_off;
3095 if ((int_map & FULL_MAP) == FULL_MAP) {
3096 return this_int_off; // we found a complete word init
3097 }
3099 // Did this store hit or cross the word boundary?
3100 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3101 if (next_int_off == this_int_off + BytesPerInt) {
3102 // We passed the current int, without fully initializing it.
3103 int_map_off = next_int_off;
3104 int_map >>= BytesPerInt;
3105 } else if (next_int_off > this_int_off + BytesPerInt) {
3106 // We passed the current and next int.
3107 return this_int_off + BytesPerInt;
3108 }
3109 }
3111 return -1;
3112 }
3115 // Called when the associated AllocateNode is expanded into CFG.
3116 // At this point, we may perform additional optimizations.
3117 // Linearize the stores by ascending offset, to make memory
3118 // activity as coherent as possible.
3119 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3120 intptr_t header_size,
3121 Node* size_in_bytes,
3122 PhaseGVN* phase) {
3123 assert(!is_complete(), "not already complete");
3124 assert(stores_are_sane(phase), "");
3125 assert(allocation() != NULL, "must be present");
3127 remove_extra_zeroes();
3129 if (ReduceFieldZeroing || ReduceBulkZeroing)
3130 // reduce instruction count for common initialization patterns
3131 coalesce_subword_stores(header_size, size_in_bytes, phase);
3133 Node* zmem = zero_memory(); // initially zero memory state
3134 Node* inits = zmem; // accumulating a linearized chain of inits
3135 #ifdef ASSERT
3136 intptr_t first_offset = allocation()->minimum_header_size();
3137 intptr_t last_init_off = first_offset; // previous init offset
3138 intptr_t last_init_end = first_offset; // previous init offset+size
3139 intptr_t last_tile_end = first_offset; // previous tile offset+size
3140 #endif
3141 intptr_t zeroes_done = header_size;
3143 bool do_zeroing = true; // we might give up if inits are very sparse
3144 int big_init_gaps = 0; // how many large gaps have we seen?
3146 if (ZeroTLAB) do_zeroing = false;
3147 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3149 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3150 Node* st = in(i);
3151 intptr_t st_off = get_store_offset(st, phase);
3152 if (st_off < 0)
3153 break; // unknown junk in the inits
3154 if (st->in(MemNode::Memory) != zmem)
3155 break; // complicated store chains somehow in list
3157 int st_size = st->as_Store()->memory_size();
3158 intptr_t next_init_off = st_off + st_size;
3160 if (do_zeroing && zeroes_done < next_init_off) {
3161 // See if this store needs a zero before it or under it.
3162 intptr_t zeroes_needed = st_off;
3164 if (st_size < BytesPerInt) {
3165 // Look for subword stores which only partially initialize words.
3166 // If we find some, we must lay down some word-level zeroes first,
3167 // underneath the subword stores.
3168 //
3169 // Examples:
3170 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3171 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3172 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3173 //
3174 // Note: coalesce_subword_stores may have already done this,
3175 // if it was prompted by constant non-zero subword initializers.
3176 // But this case can still arise with non-constant stores.
3178 intptr_t next_full_store = find_next_fullword_store(i, phase);
3180 // In the examples above:
3181 // in(i) p q r s x y z
3182 // st_off 12 13 14 15 12 13 14
3183 // st_size 1 1 1 1 1 1 1
3184 // next_full_s. 12 16 16 16 16 16 16
3185 // z's_done 12 16 16 16 12 16 12
3186 // z's_needed 12 16 16 16 16 16 16
3187 // zsize 0 0 0 0 4 0 4
3188 if (next_full_store < 0) {
3189 // Conservative tack: Zero to end of current word.
3190 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3191 } else {
3192 // Zero to beginning of next fully initialized word.
3193 // Or, don't zero at all, if we are already in that word.
3194 assert(next_full_store >= zeroes_needed, "must go forward");
3195 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3196 zeroes_needed = next_full_store;
3197 }
3198 }
3200 if (zeroes_needed > zeroes_done) {
3201 intptr_t zsize = zeroes_needed - zeroes_done;
3202 // Do some incremental zeroing on rawmem, in parallel with inits.
3203 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3204 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3205 zeroes_done, zeroes_needed,
3206 phase);
3207 zeroes_done = zeroes_needed;
3208 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3209 do_zeroing = false; // leave the hole, next time
3210 }
3211 }
3213 // Collect the store and move on:
3214 st->set_req(MemNode::Memory, inits);
3215 inits = st; // put it on the linearized chain
3216 set_req(i, zmem); // unhook from previous position
3218 if (zeroes_done == st_off)
3219 zeroes_done = next_init_off;
3221 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3223 #ifdef ASSERT
3224 // Various order invariants. Weaker than stores_are_sane because
3225 // a large constant tile can be filled in by smaller non-constant stores.
3226 assert(st_off >= last_init_off, "inits do not reverse");
3227 last_init_off = st_off;
3228 const Type* val = NULL;
3229 if (st_size >= BytesPerInt &&
3230 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3231 (int)val->basic_type() < (int)T_OBJECT) {
3232 assert(st_off >= last_tile_end, "tiles do not overlap");
3233 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3234 last_tile_end = MAX2(last_tile_end, next_init_off);
3235 } else {
3236 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3237 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3238 assert(st_off >= last_init_end, "inits do not overlap");
3239 last_init_end = next_init_off; // it's a non-tile
3240 }
3241 #endif //ASSERT
3242 }
3244 remove_extra_zeroes(); // clear out all the zmems left over
3245 add_req(inits);
3247 if (!ZeroTLAB) {
3248 // If anything remains to be zeroed, zero it all now.
3249 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3250 // if it is the last unused 4 bytes of an instance, forget about it
3251 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3252 if (zeroes_done + BytesPerLong >= size_limit) {
3253 assert(allocation() != NULL, "");
3254 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3255 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3256 if (zeroes_done == k->layout_helper())
3257 zeroes_done = size_limit;
3258 }
3259 if (zeroes_done < size_limit) {
3260 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3261 zeroes_done, size_in_bytes, phase);
3262 }
3263 }
3265 set_complete(phase);
3266 return rawmem;
3267 }
3270 #ifdef ASSERT
3271 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3272 if (is_complete())
3273 return true; // stores could be anything at this point
3274 assert(allocation() != NULL, "must be present");
3275 intptr_t last_off = allocation()->minimum_header_size();
3276 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3277 Node* st = in(i);
3278 intptr_t st_off = get_store_offset(st, phase);
3279 if (st_off < 0) continue; // ignore dead garbage
3280 if (last_off > st_off) {
3281 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3282 this->dump(2);
3283 assert(false, "ascending store offsets");
3284 return false;
3285 }
3286 last_off = st_off + st->as_Store()->memory_size();
3287 }
3288 return true;
3289 }
3290 #endif //ASSERT
3295 //============================MergeMemNode=====================================
3296 //
3297 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3298 // contributing store or call operations. Each contributor provides the memory
3299 // state for a particular "alias type" (see Compile::alias_type). For example,
3300 // if a MergeMem has an input X for alias category #6, then any memory reference
3301 // to alias category #6 may use X as its memory state input, as an exact equivalent
3302 // to using the MergeMem as a whole.
3303 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3304 //
3305 // (Here, the <N> notation gives the index of the relevant adr_type.)
3306 //
3307 // In one special case (and more cases in the future), alias categories overlap.
3308 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3309 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3310 // it is exactly equivalent to that state W:
3311 // MergeMem(<Bot>: W) <==> W
3312 //
3313 // Usually, the merge has more than one input. In that case, where inputs
3314 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3315 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3316 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3317 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3318 //
3319 // A merge can take a "wide" memory state as one of its narrow inputs.
3320 // This simply means that the merge observes out only the relevant parts of
3321 // the wide input. That is, wide memory states arriving at narrow merge inputs
3322 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3323 //
3324 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3325 // and that memory slices "leak through":
3326 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3327 //
3328 // But, in such a cascade, repeated memory slices can "block the leak":
3329 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3330 //
3331 // In the last example, Y is not part of the combined memory state of the
3332 // outermost MergeMem. The system must, of course, prevent unschedulable
3333 // memory states from arising, so you can be sure that the state Y is somehow
3334 // a precursor to state Y'.
3335 //
3336 //
3337 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3338 // of each MergeMemNode array are exactly the numerical alias indexes, including
3339 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3340 // Compile::alias_type (and kin) produce and manage these indexes.
3341 //
3342 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3343 // (Note that this provides quick access to the top node inside MergeMem methods,
3344 // without the need to reach out via TLS to Compile::current.)
3345 //
3346 // As a consequence of what was just described, a MergeMem that represents a full
3347 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3348 // containing all alias categories.
3349 //
3350 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3351 //
3352 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3353 // a memory state for the alias type <N>, or else the top node, meaning that
3354 // there is no particular input for that alias type. Note that the length of
3355 // a MergeMem is variable, and may be extended at any time to accommodate new
3356 // memory states at larger alias indexes. When merges grow, they are of course
3357 // filled with "top" in the unused in() positions.
3358 //
3359 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3360 // (Top was chosen because it works smoothly with passes like GCM.)
3361 //
3362 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3363 // the type of random VM bits like TLS references.) Since it is always the
3364 // first non-Bot memory slice, some low-level loops use it to initialize an
3365 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3366 //
3367 //
3368 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3369 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3370 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3371 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3372 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3373 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3374 //
3375 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3376 // really that different from the other memory inputs. An abbreviation called
3377 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3378 //
3379 //
3380 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3381 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3382 // that "emerges though" the base memory will be marked as excluding the alias types
3383 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3384 //
3385 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3386 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3387 //
3388 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3389 // (It is currently unimplemented.) As you can see, the resulting merge is
3390 // actually a disjoint union of memory states, rather than an overlay.
3391 //
3393 //------------------------------MergeMemNode-----------------------------------
3394 Node* MergeMemNode::make_empty_memory() {
3395 Node* empty_memory = (Node*) Compile::current()->top();
3396 assert(empty_memory->is_top(), "correct sentinel identity");
3397 return empty_memory;
3398 }
3400 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3401 init_class_id(Class_MergeMem);
3402 // all inputs are nullified in Node::Node(int)
3403 // set_input(0, NULL); // no control input
3405 // Initialize the edges uniformly to top, for starters.
3406 Node* empty_mem = make_empty_memory();
3407 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3408 init_req(i,empty_mem);
3409 }
3410 assert(empty_memory() == empty_mem, "");
3412 if( new_base != NULL && new_base->is_MergeMem() ) {
3413 MergeMemNode* mdef = new_base->as_MergeMem();
3414 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3415 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3416 mms.set_memory(mms.memory2());
3417 }
3418 assert(base_memory() == mdef->base_memory(), "");
3419 } else {
3420 set_base_memory(new_base);
3421 }
3422 }
3424 // Make a new, untransformed MergeMem with the same base as 'mem'.
3425 // If mem is itself a MergeMem, populate the result with the same edges.
3426 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3427 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3428 }
3430 //------------------------------cmp--------------------------------------------
3431 uint MergeMemNode::hash() const { return NO_HASH; }
3432 uint MergeMemNode::cmp( const Node &n ) const {
3433 return (&n == this); // Always fail except on self
3434 }
3436 //------------------------------Identity---------------------------------------
3437 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3438 // Identity if this merge point does not record any interesting memory
3439 // disambiguations.
3440 Node* base_mem = base_memory();
3441 Node* empty_mem = empty_memory();
3442 if (base_mem != empty_mem) { // Memory path is not dead?
3443 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3444 Node* mem = in(i);
3445 if (mem != empty_mem && mem != base_mem) {
3446 return this; // Many memory splits; no change
3447 }
3448 }
3449 }
3450 return base_mem; // No memory splits; ID on the one true input
3451 }
3453 //------------------------------Ideal------------------------------------------
3454 // This method is invoked recursively on chains of MergeMem nodes
3455 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3456 // Remove chain'd MergeMems
3457 //
3458 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3459 // relative to the "in(Bot)". Since we are patching both at the same time,
3460 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3461 // but rewrite each "in(i)" relative to the new "in(Bot)".
3462 Node *progress = NULL;
3465 Node* old_base = base_memory();
3466 Node* empty_mem = empty_memory();
3467 if (old_base == empty_mem)
3468 return NULL; // Dead memory path.
3470 MergeMemNode* old_mbase;
3471 if (old_base != NULL && old_base->is_MergeMem())
3472 old_mbase = old_base->as_MergeMem();
3473 else
3474 old_mbase = NULL;
3475 Node* new_base = old_base;
3477 // simplify stacked MergeMems in base memory
3478 if (old_mbase) new_base = old_mbase->base_memory();
3480 // the base memory might contribute new slices beyond my req()
3481 if (old_mbase) grow_to_match(old_mbase);
3483 // Look carefully at the base node if it is a phi.
3484 PhiNode* phi_base;
3485 if (new_base != NULL && new_base->is_Phi())
3486 phi_base = new_base->as_Phi();
3487 else
3488 phi_base = NULL;
3490 Node* phi_reg = NULL;
3491 uint phi_len = (uint)-1;
3492 if (phi_base != NULL && !phi_base->is_copy()) {
3493 // do not examine phi if degraded to a copy
3494 phi_reg = phi_base->region();
3495 phi_len = phi_base->req();
3496 // see if the phi is unfinished
3497 for (uint i = 1; i < phi_len; i++) {
3498 if (phi_base->in(i) == NULL) {
3499 // incomplete phi; do not look at it yet!
3500 phi_reg = NULL;
3501 phi_len = (uint)-1;
3502 break;
3503 }
3504 }
3505 }
3507 // Note: We do not call verify_sparse on entry, because inputs
3508 // can normalize to the base_memory via subsume_node or similar
3509 // mechanisms. This method repairs that damage.
3511 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3513 // Look at each slice.
3514 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3515 Node* old_in = in(i);
3516 // calculate the old memory value
3517 Node* old_mem = old_in;
3518 if (old_mem == empty_mem) old_mem = old_base;
3519 assert(old_mem == memory_at(i), "");
3521 // maybe update (reslice) the old memory value
3523 // simplify stacked MergeMems
3524 Node* new_mem = old_mem;
3525 MergeMemNode* old_mmem;
3526 if (old_mem != NULL && old_mem->is_MergeMem())
3527 old_mmem = old_mem->as_MergeMem();
3528 else
3529 old_mmem = NULL;
3530 if (old_mmem == this) {
3531 // This can happen if loops break up and safepoints disappear.
3532 // A merge of BotPtr (default) with a RawPtr memory derived from a
3533 // safepoint can be rewritten to a merge of the same BotPtr with
3534 // the BotPtr phi coming into the loop. If that phi disappears
3535 // also, we can end up with a self-loop of the mergemem.
3536 // In general, if loops degenerate and memory effects disappear,
3537 // a mergemem can be left looking at itself. This simply means
3538 // that the mergemem's default should be used, since there is
3539 // no longer any apparent effect on this slice.
3540 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3541 // from start. Update the input to TOP.
3542 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3543 }
3544 else if (old_mmem != NULL) {
3545 new_mem = old_mmem->memory_at(i);
3546 }
3547 // else preceeding memory was not a MergeMem
3549 // replace equivalent phis (unfortunately, they do not GVN together)
3550 if (new_mem != NULL && new_mem != new_base &&
3551 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3552 if (new_mem->is_Phi()) {
3553 PhiNode* phi_mem = new_mem->as_Phi();
3554 for (uint i = 1; i < phi_len; i++) {
3555 if (phi_base->in(i) != phi_mem->in(i)) {
3556 phi_mem = NULL;
3557 break;
3558 }
3559 }
3560 if (phi_mem != NULL) {
3561 // equivalent phi nodes; revert to the def
3562 new_mem = new_base;
3563 }
3564 }
3565 }
3567 // maybe store down a new value
3568 Node* new_in = new_mem;
3569 if (new_in == new_base) new_in = empty_mem;
3571 if (new_in != old_in) {
3572 // Warning: Do not combine this "if" with the previous "if"
3573 // A memory slice might have be be rewritten even if it is semantically
3574 // unchanged, if the base_memory value has changed.
3575 set_req(i, new_in);
3576 progress = this; // Report progress
3577 }
3578 }
3580 if (new_base != old_base) {
3581 set_req(Compile::AliasIdxBot, new_base);
3582 // Don't use set_base_memory(new_base), because we need to update du.
3583 assert(base_memory() == new_base, "");
3584 progress = this;
3585 }
3587 if( base_memory() == this ) {
3588 // a self cycle indicates this memory path is dead
3589 set_req(Compile::AliasIdxBot, empty_mem);
3590 }
3592 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3593 // Recursion must occur after the self cycle check above
3594 if( base_memory()->is_MergeMem() ) {
3595 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3596 Node *m = phase->transform(new_mbase); // Rollup any cycles
3597 if( m != NULL && (m->is_top() ||
3598 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3599 // propagate rollup of dead cycle to self
3600 set_req(Compile::AliasIdxBot, empty_mem);
3601 }
3602 }
3604 if( base_memory() == empty_mem ) {
3605 progress = this;
3606 // Cut inputs during Parse phase only.
3607 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3608 if( !can_reshape ) {
3609 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3610 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3611 }
3612 }
3613 }
3615 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3616 // Check if PhiNode::Ideal's "Split phis through memory merges"
3617 // transform should be attempted. Look for this->phi->this cycle.
3618 uint merge_width = req();
3619 if (merge_width > Compile::AliasIdxRaw) {
3620 PhiNode* phi = base_memory()->as_Phi();
3621 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3622 if (phi->in(i) == this) {
3623 phase->is_IterGVN()->_worklist.push(phi);
3624 break;
3625 }
3626 }
3627 }
3628 }
3630 assert(progress || verify_sparse(), "please, no dups of base");
3631 return progress;
3632 }
3634 //-------------------------set_base_memory-------------------------------------
3635 void MergeMemNode::set_base_memory(Node *new_base) {
3636 Node* empty_mem = empty_memory();
3637 set_req(Compile::AliasIdxBot, new_base);
3638 assert(memory_at(req()) == new_base, "must set default memory");
3639 // Clear out other occurrences of new_base:
3640 if (new_base != empty_mem) {
3641 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3642 if (in(i) == new_base) set_req(i, empty_mem);
3643 }
3644 }
3645 }
3647 //------------------------------out_RegMask------------------------------------
3648 const RegMask &MergeMemNode::out_RegMask() const {
3649 return RegMask::Empty;
3650 }
3652 //------------------------------dump_spec--------------------------------------
3653 #ifndef PRODUCT
3654 void MergeMemNode::dump_spec(outputStream *st) const {
3655 st->print(" {");
3656 Node* base_mem = base_memory();
3657 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3658 Node* mem = memory_at(i);
3659 if (mem == base_mem) { st->print(" -"); continue; }
3660 st->print( " N%d:", mem->_idx );
3661 Compile::current()->get_adr_type(i)->dump_on(st);
3662 }
3663 st->print(" }");
3664 }
3665 #endif // !PRODUCT
3668 #ifdef ASSERT
3669 static bool might_be_same(Node* a, Node* b) {
3670 if (a == b) return true;
3671 if (!(a->is_Phi() || b->is_Phi())) return false;
3672 // phis shift around during optimization
3673 return true; // pretty stupid...
3674 }
3676 // verify a narrow slice (either incoming or outgoing)
3677 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3678 if (!VerifyAliases) return; // don't bother to verify unless requested
3679 if (is_error_reported()) return; // muzzle asserts when debugging an error
3680 if (Node::in_dump()) return; // muzzle asserts when printing
3681 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3682 assert(n != NULL, "");
3683 // Elide intervening MergeMem's
3684 while (n->is_MergeMem()) {
3685 n = n->as_MergeMem()->memory_at(alias_idx);
3686 }
3687 Compile* C = Compile::current();
3688 const TypePtr* n_adr_type = n->adr_type();
3689 if (n == m->empty_memory()) {
3690 // Implicit copy of base_memory()
3691 } else if (n_adr_type != TypePtr::BOTTOM) {
3692 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3693 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3694 } else {
3695 // A few places like make_runtime_call "know" that VM calls are narrow,
3696 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3697 bool expected_wide_mem = false;
3698 if (n == m->base_memory()) {
3699 expected_wide_mem = true;
3700 } else if (alias_idx == Compile::AliasIdxRaw ||
3701 n == m->memory_at(Compile::AliasIdxRaw)) {
3702 expected_wide_mem = true;
3703 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3704 // memory can "leak through" calls on channels that
3705 // are write-once. Allow this also.
3706 expected_wide_mem = true;
3707 }
3708 assert(expected_wide_mem, "expected narrow slice replacement");
3709 }
3710 }
3711 #else // !ASSERT
3712 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3713 #endif
3716 //-----------------------------memory_at---------------------------------------
3717 Node* MergeMemNode::memory_at(uint alias_idx) const {
3718 assert(alias_idx >= Compile::AliasIdxRaw ||
3719 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3720 "must avoid base_memory and AliasIdxTop");
3722 // Otherwise, it is a narrow slice.
3723 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3724 Compile *C = Compile::current();
3725 if (is_empty_memory(n)) {
3726 // the array is sparse; empty slots are the "top" node
3727 n = base_memory();
3728 assert(Node::in_dump()
3729 || n == NULL || n->bottom_type() == Type::TOP
3730 || n->adr_type() == TypePtr::BOTTOM
3731 || n->adr_type() == TypeRawPtr::BOTTOM
3732 || Compile::current()->AliasLevel() == 0,
3733 "must be a wide memory");
3734 // AliasLevel == 0 if we are organizing the memory states manually.
3735 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3736 } else {
3737 // make sure the stored slice is sane
3738 #ifdef ASSERT
3739 if (is_error_reported() || Node::in_dump()) {
3740 } else if (might_be_same(n, base_memory())) {
3741 // Give it a pass: It is a mostly harmless repetition of the base.
3742 // This can arise normally from node subsumption during optimization.
3743 } else {
3744 verify_memory_slice(this, alias_idx, n);
3745 }
3746 #endif
3747 }
3748 return n;
3749 }
3751 //---------------------------set_memory_at-------------------------------------
3752 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3753 verify_memory_slice(this, alias_idx, n);
3754 Node* empty_mem = empty_memory();
3755 if (n == base_memory()) n = empty_mem; // collapse default
3756 uint need_req = alias_idx+1;
3757 if (req() < need_req) {
3758 if (n == empty_mem) return; // already the default, so do not grow me
3759 // grow the sparse array
3760 do {
3761 add_req(empty_mem);
3762 } while (req() < need_req);
3763 }
3764 set_req( alias_idx, n );
3765 }
3769 //--------------------------iteration_setup------------------------------------
3770 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3771 if (other != NULL) {
3772 grow_to_match(other);
3773 // invariant: the finite support of mm2 is within mm->req()
3774 #ifdef ASSERT
3775 for (uint i = req(); i < other->req(); i++) {
3776 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3777 }
3778 #endif
3779 }
3780 // Replace spurious copies of base_memory by top.
3781 Node* base_mem = base_memory();
3782 if (base_mem != NULL && !base_mem->is_top()) {
3783 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3784 if (in(i) == base_mem)
3785 set_req(i, empty_memory());
3786 }
3787 }
3788 }
3790 //---------------------------grow_to_match-------------------------------------
3791 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3792 Node* empty_mem = empty_memory();
3793 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3794 // look for the finite support of the other memory
3795 for (uint i = other->req(); --i >= req(); ) {
3796 if (other->in(i) != empty_mem) {
3797 uint new_len = i+1;
3798 while (req() < new_len) add_req(empty_mem);
3799 break;
3800 }
3801 }
3802 }
3804 //---------------------------verify_sparse-------------------------------------
3805 #ifndef PRODUCT
3806 bool MergeMemNode::verify_sparse() const {
3807 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3808 Node* base_mem = base_memory();
3809 // The following can happen in degenerate cases, since empty==top.
3810 if (is_empty_memory(base_mem)) return true;
3811 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3812 assert(in(i) != NULL, "sane slice");
3813 if (in(i) == base_mem) return false; // should have been the sentinel value!
3814 }
3815 return true;
3816 }
3818 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3819 Node* n;
3820 n = mm->in(idx);
3821 if (mem == n) return true; // might be empty_memory()
3822 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3823 if (mem == n) return true;
3824 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3825 if (mem == n) return true;
3826 if (n == NULL) break;
3827 }
3828 return false;
3829 }
3830 #endif // !PRODUCT