Tue, 15 Apr 2008 10:49:32 -0700
6692301: Side effect in NumberFormat tests with -server -Xcomp
Summary: Optimization in CmpPNode::sub() removed the valid compare instruction because of false positive answer from detect_dominating_control().
Reviewed-by: jrose, sgoldman
1 /*
2 * Copyright 1997-2007 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 // Portions of code courtesy of Clifford Click
27 // Optimization - Graph Style
29 #include "incls/_precompiled.incl"
30 #include "incls/_memnode.cpp.incl"
32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
34 //=============================================================================
35 uint MemNode::size_of() const { return sizeof(*this); }
37 const TypePtr *MemNode::adr_type() const {
38 Node* adr = in(Address);
39 const TypePtr* cross_check = NULL;
40 DEBUG_ONLY(cross_check = _adr_type);
41 return calculate_adr_type(adr->bottom_type(), cross_check);
42 }
44 #ifndef PRODUCT
45 void MemNode::dump_spec(outputStream *st) const {
46 if (in(Address) == NULL) return; // node is dead
47 #ifndef ASSERT
48 // fake the missing field
49 const TypePtr* _adr_type = NULL;
50 if (in(Address) != NULL)
51 _adr_type = in(Address)->bottom_type()->isa_ptr();
52 #endif
53 dump_adr_type(this, _adr_type, st);
55 Compile* C = Compile::current();
56 if( C->alias_type(_adr_type)->is_volatile() )
57 st->print(" Volatile!");
58 }
60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
61 st->print(" @");
62 if (adr_type == NULL) {
63 st->print("NULL");
64 } else {
65 adr_type->dump_on(st);
66 Compile* C = Compile::current();
67 Compile::AliasType* atp = NULL;
68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
69 if (atp == NULL)
70 st->print(", idx=?\?;");
71 else if (atp->index() == Compile::AliasIdxBot)
72 st->print(", idx=Bot;");
73 else if (atp->index() == Compile::AliasIdxTop)
74 st->print(", idx=Top;");
75 else if (atp->index() == Compile::AliasIdxRaw)
76 st->print(", idx=Raw;");
77 else {
78 ciField* field = atp->field();
79 if (field) {
80 st->print(", name=");
81 field->print_name_on(st);
82 }
83 st->print(", idx=%d;", atp->index());
84 }
85 }
86 }
88 extern void print_alias_types();
90 #endif
92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
93 const TypeOopPtr *tinst = t_adr->isa_oopptr();
94 if (tinst == NULL || !tinst->is_instance_field())
95 return mchain; // don't try to optimize non-instance types
96 uint instance_id = tinst->instance_id();
97 Node *prev = NULL;
98 Node *result = mchain;
99 while (prev != result) {
100 prev = result;
101 // skip over a call which does not affect this memory slice
102 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
103 Node *proj_in = result->in(0);
104 if (proj_in->is_Call()) {
105 CallNode *call = proj_in->as_Call();
106 if (!call->may_modify(t_adr, phase)) {
107 result = call->in(TypeFunc::Memory);
108 }
109 } else if (proj_in->is_Initialize()) {
110 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
111 // Stop if this is the initialization for the object instance which
112 // which contains this memory slice, otherwise skip over it.
113 if (alloc != NULL && alloc->_idx != instance_id) {
114 result = proj_in->in(TypeFunc::Memory);
115 }
116 } else if (proj_in->is_MemBar()) {
117 result = proj_in->in(TypeFunc::Memory);
118 }
119 } else if (result->is_MergeMem()) {
120 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
121 }
122 }
123 return result;
124 }
126 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
127 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
128 bool is_instance = (t_oop != NULL) && t_oop->is_instance_field();
129 PhaseIterGVN *igvn = phase->is_IterGVN();
130 Node *result = mchain;
131 result = optimize_simple_memory_chain(result, t_adr, phase);
132 if (is_instance && igvn != NULL && result->is_Phi()) {
133 PhiNode *mphi = result->as_Phi();
134 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
135 const TypePtr *t = mphi->adr_type();
136 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM) {
137 // clone the Phi with our address type
138 result = mphi->split_out_instance(t_adr, igvn);
139 } else {
140 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
141 }
142 }
143 return result;
144 }
146 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
147 uint alias_idx = phase->C->get_alias_index(tp);
148 Node *mem = mmem;
149 #ifdef ASSERT
150 {
151 // Check that current type is consistent with the alias index used during graph construction
152 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
153 bool consistent = adr_check == NULL || adr_check->empty() ||
154 phase->C->must_alias(adr_check, alias_idx );
155 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
156 if( !consistent && adr_check != NULL && !adr_check->empty() &&
157 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
158 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
159 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
160 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
161 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
162 // don't assert if it is dead code.
163 consistent = true;
164 }
165 if( !consistent ) {
166 st->print("alias_idx==%d, adr_check==", alias_idx);
167 if( adr_check == NULL ) {
168 st->print("NULL");
169 } else {
170 adr_check->dump();
171 }
172 st->cr();
173 print_alias_types();
174 assert(consistent, "adr_check must match alias idx");
175 }
176 }
177 #endif
178 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
179 // means an array I have not precisely typed yet. Do not do any
180 // alias stuff with it any time soon.
181 const TypeOopPtr *tinst = tp->isa_oopptr();
182 if( tp->base() != Type::AnyPtr &&
183 !(tinst &&
184 tinst->klass()->is_java_lang_Object() &&
185 tinst->offset() == Type::OffsetBot) ) {
186 // compress paths and change unreachable cycles to TOP
187 // If not, we can update the input infinitely along a MergeMem cycle
188 // Equivalent code in PhiNode::Ideal
189 Node* m = phase->transform(mmem);
190 // If tranformed to a MergeMem, get the desired slice
191 // Otherwise the returned node represents memory for every slice
192 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
193 // Update input if it is progress over what we have now
194 }
195 return mem;
196 }
198 //--------------------------Ideal_common---------------------------------------
199 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
200 // Unhook non-raw memories from complete (macro-expanded) initializations.
201 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
202 // If our control input is a dead region, kill all below the region
203 Node *ctl = in(MemNode::Control);
204 if (ctl && remove_dead_region(phase, can_reshape))
205 return this;
207 // Ignore if memory is dead, or self-loop
208 Node *mem = in(MemNode::Memory);
209 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
210 assert( mem != this, "dead loop in MemNode::Ideal" );
212 Node *address = in(MemNode::Address);
213 const Type *t_adr = phase->type( address );
214 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
216 // Avoid independent memory operations
217 Node* old_mem = mem;
219 // The code which unhooks non-raw memories from complete (macro-expanded)
220 // initializations was removed. After macro-expansion all stores catched
221 // by Initialize node became raw stores and there is no information
222 // which memory slices they modify. So it is unsafe to move any memory
223 // operation above these stores. Also in most cases hooked non-raw memories
224 // were already unhooked by using information from detect_ptr_independence()
225 // and find_previous_store().
227 if (mem->is_MergeMem()) {
228 MergeMemNode* mmem = mem->as_MergeMem();
229 const TypePtr *tp = t_adr->is_ptr();
231 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
232 }
234 if (mem != old_mem) {
235 set_req(MemNode::Memory, mem);
236 return this;
237 }
239 // let the subclass continue analyzing...
240 return NULL;
241 }
243 // Helper function for proving some simple control dominations.
244 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
245 // Already assumes that 'dom' is available at 'sub', and that 'sub'
246 // is not a constant (dominated by the method's StartNode).
247 // Used by MemNode::find_previous_store to prove that the
248 // control input of a memory operation predates (dominates)
249 // an allocation it wants to look past.
250 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
251 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
252 return false; // Conservative answer for dead code
254 // Check 'dom'.
255 dom = dom->find_exact_control(dom);
256 if (dom == NULL || dom->is_top())
257 return false; // Conservative answer for dead code
259 if (dom->is_Start() || dom->is_Root() || dom == sub)
260 return true;
262 // 'dom' dominates 'sub' if its control edge and control edges
263 // of all its inputs dominate or equal to sub's control edge.
265 // Currently 'sub' is either Allocate, Initialize or Start nodes.
266 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start(), "expecting only these nodes");
268 // Get control edge of 'sub'.
269 sub = sub->find_exact_control(sub->in(0));
270 if (sub == NULL || sub->is_top())
271 return false; // Conservative answer for dead code
273 assert(sub->is_CFG(), "expecting control");
275 if (sub == dom)
276 return true;
278 if (sub->is_Start() || sub->is_Root())
279 return false;
281 {
282 // Check all control edges of 'dom'.
284 ResourceMark rm;
285 Arena* arena = Thread::current()->resource_area();
286 Node_List nlist(arena);
287 Unique_Node_List dom_list(arena);
289 dom_list.push(dom);
290 bool only_dominating_controls = false;
292 for (uint next = 0; next < dom_list.size(); next++) {
293 Node* n = dom_list.at(next);
294 if (!n->is_CFG() && n->pinned()) {
295 // Check only own control edge for pinned non-control nodes.
296 n = n->find_exact_control(n->in(0));
297 if (n == NULL || n->is_top())
298 return false; // Conservative answer for dead code
299 assert(n->is_CFG(), "expecting control");
300 }
301 if (n->is_Start() || n->is_Root()) {
302 only_dominating_controls = true;
303 } else if (n->is_CFG()) {
304 if (n->dominates(sub, nlist))
305 only_dominating_controls = true;
306 else
307 return false;
308 } else {
309 // First, own control edge.
310 Node* m = n->find_exact_control(n->in(0));
311 if (m == NULL)
312 continue;
313 if (m->is_top())
314 return false; // Conservative answer for dead code
315 dom_list.push(m);
317 // Now, the rest of edges.
318 uint cnt = n->req();
319 for (uint i = 1; i < cnt; i++) {
320 m = n->find_exact_control(n->in(i));
321 if (m == NULL || m->is_top())
322 continue;
323 dom_list.push(m);
324 }
325 }
326 }
327 return only_dominating_controls;
328 }
329 }
331 //---------------------detect_ptr_independence---------------------------------
332 // Used by MemNode::find_previous_store to prove that two base
333 // pointers are never equal.
334 // The pointers are accompanied by their associated allocations,
335 // if any, which have been previously discovered by the caller.
336 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
337 Node* p2, AllocateNode* a2,
338 PhaseTransform* phase) {
339 // Attempt to prove that these two pointers cannot be aliased.
340 // They may both manifestly be allocations, and they should differ.
341 // Or, if they are not both allocations, they can be distinct constants.
342 // Otherwise, one is an allocation and the other a pre-existing value.
343 if (a1 == NULL && a2 == NULL) { // neither an allocation
344 return (p1 != p2) && p1->is_Con() && p2->is_Con();
345 } else if (a1 != NULL && a2 != NULL) { // both allocations
346 return (a1 != a2);
347 } else if (a1 != NULL) { // one allocation a1
348 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
349 return all_controls_dominate(p2, a1);
350 } else { //(a2 != NULL) // one allocation a2
351 return all_controls_dominate(p1, a2);
352 }
353 return false;
354 }
357 // The logic for reordering loads and stores uses four steps:
358 // (a) Walk carefully past stores and initializations which we
359 // can prove are independent of this load.
360 // (b) Observe that the next memory state makes an exact match
361 // with self (load or store), and locate the relevant store.
362 // (c) Ensure that, if we were to wire self directly to the store,
363 // the optimizer would fold it up somehow.
364 // (d) Do the rewiring, and return, depending on some other part of
365 // the optimizer to fold up the load.
366 // This routine handles steps (a) and (b). Steps (c) and (d) are
367 // specific to loads and stores, so they are handled by the callers.
368 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
369 //
370 Node* MemNode::find_previous_store(PhaseTransform* phase) {
371 Node* ctrl = in(MemNode::Control);
372 Node* adr = in(MemNode::Address);
373 intptr_t offset = 0;
374 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
375 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
377 if (offset == Type::OffsetBot)
378 return NULL; // cannot unalias unless there are precise offsets
380 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
382 intptr_t size_in_bytes = memory_size();
384 Node* mem = in(MemNode::Memory); // start searching here...
386 int cnt = 50; // Cycle limiter
387 for (;;) { // While we can dance past unrelated stores...
388 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
390 if (mem->is_Store()) {
391 Node* st_adr = mem->in(MemNode::Address);
392 intptr_t st_offset = 0;
393 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
394 if (st_base == NULL)
395 break; // inscrutable pointer
396 if (st_offset != offset && st_offset != Type::OffsetBot) {
397 const int MAX_STORE = BytesPerLong;
398 if (st_offset >= offset + size_in_bytes ||
399 st_offset <= offset - MAX_STORE ||
400 st_offset <= offset - mem->as_Store()->memory_size()) {
401 // Success: The offsets are provably independent.
402 // (You may ask, why not just test st_offset != offset and be done?
403 // The answer is that stores of different sizes can co-exist
404 // in the same sequence of RawMem effects. We sometimes initialize
405 // a whole 'tile' of array elements with a single jint or jlong.)
406 mem = mem->in(MemNode::Memory);
407 continue; // (a) advance through independent store memory
408 }
409 }
410 if (st_base != base &&
411 detect_ptr_independence(base, alloc,
412 st_base,
413 AllocateNode::Ideal_allocation(st_base, phase),
414 phase)) {
415 // Success: The bases are provably independent.
416 mem = mem->in(MemNode::Memory);
417 continue; // (a) advance through independent store memory
418 }
420 // (b) At this point, if the bases or offsets do not agree, we lose,
421 // since we have not managed to prove 'this' and 'mem' independent.
422 if (st_base == base && st_offset == offset) {
423 return mem; // let caller handle steps (c), (d)
424 }
426 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
427 InitializeNode* st_init = mem->in(0)->as_Initialize();
428 AllocateNode* st_alloc = st_init->allocation();
429 if (st_alloc == NULL)
430 break; // something degenerated
431 bool known_identical = false;
432 bool known_independent = false;
433 if (alloc == st_alloc)
434 known_identical = true;
435 else if (alloc != NULL)
436 known_independent = true;
437 else if (all_controls_dominate(this, st_alloc))
438 known_independent = true;
440 if (known_independent) {
441 // The bases are provably independent: Either they are
442 // manifestly distinct allocations, or else the control
443 // of this load dominates the store's allocation.
444 int alias_idx = phase->C->get_alias_index(adr_type());
445 if (alias_idx == Compile::AliasIdxRaw) {
446 mem = st_alloc->in(TypeFunc::Memory);
447 } else {
448 mem = st_init->memory(alias_idx);
449 }
450 continue; // (a) advance through independent store memory
451 }
453 // (b) at this point, if we are not looking at a store initializing
454 // the same allocation we are loading from, we lose.
455 if (known_identical) {
456 // From caller, can_see_stored_value will consult find_captured_store.
457 return mem; // let caller handle steps (c), (d)
458 }
460 } else if (addr_t != NULL && addr_t->is_instance_field()) {
461 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
462 if (mem->is_Proj() && mem->in(0)->is_Call()) {
463 CallNode *call = mem->in(0)->as_Call();
464 if (!call->may_modify(addr_t, phase)) {
465 mem = call->in(TypeFunc::Memory);
466 continue; // (a) advance through independent call memory
467 }
468 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
469 mem = mem->in(0)->in(TypeFunc::Memory);
470 continue; // (a) advance through independent MemBar memory
471 } else if (mem->is_MergeMem()) {
472 int alias_idx = phase->C->get_alias_index(adr_type());
473 mem = mem->as_MergeMem()->memory_at(alias_idx);
474 continue; // (a) advance through independent MergeMem memory
475 }
476 }
478 // Unless there is an explicit 'continue', we must bail out here,
479 // because 'mem' is an inscrutable memory state (e.g., a call).
480 break;
481 }
483 return NULL; // bail out
484 }
486 //----------------------calculate_adr_type-------------------------------------
487 // Helper function. Notices when the given type of address hits top or bottom.
488 // Also, asserts a cross-check of the type against the expected address type.
489 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
490 if (t == Type::TOP) return NULL; // does not touch memory any more?
491 #ifdef PRODUCT
492 cross_check = NULL;
493 #else
494 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
495 #endif
496 const TypePtr* tp = t->isa_ptr();
497 if (tp == NULL) {
498 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
499 return TypePtr::BOTTOM; // touches lots of memory
500 } else {
501 #ifdef ASSERT
502 // %%%% [phh] We don't check the alias index if cross_check is
503 // TypeRawPtr::BOTTOM. Needs to be investigated.
504 if (cross_check != NULL &&
505 cross_check != TypePtr::BOTTOM &&
506 cross_check != TypeRawPtr::BOTTOM) {
507 // Recheck the alias index, to see if it has changed (due to a bug).
508 Compile* C = Compile::current();
509 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
510 "must stay in the original alias category");
511 // The type of the address must be contained in the adr_type,
512 // disregarding "null"-ness.
513 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
514 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
515 assert(cross_check->meet(tp_notnull) == cross_check,
516 "real address must not escape from expected memory type");
517 }
518 #endif
519 return tp;
520 }
521 }
523 //------------------------adr_phi_is_loop_invariant----------------------------
524 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
525 // loop is loop invariant. Make a quick traversal of Phi and associated
526 // CastPP nodes, looking to see if they are a closed group within the loop.
527 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
528 // The idea is that the phi-nest must boil down to only CastPP nodes
529 // with the same data. This implies that any path into the loop already
530 // includes such a CastPP, and so the original cast, whatever its input,
531 // must be covered by an equivalent cast, with an earlier control input.
532 ResourceMark rm;
534 // The loop entry input of the phi should be the unique dominating
535 // node for every Phi/CastPP in the loop.
536 Unique_Node_List closure;
537 closure.push(adr_phi->in(LoopNode::EntryControl));
539 // Add the phi node and the cast to the worklist.
540 Unique_Node_List worklist;
541 worklist.push(adr_phi);
542 if( cast != NULL ){
543 if( !cast->is_ConstraintCast() ) return false;
544 worklist.push(cast);
545 }
547 // Begin recursive walk of phi nodes.
548 while( worklist.size() ){
549 // Take a node off the worklist
550 Node *n = worklist.pop();
551 if( !closure.member(n) ){
552 // Add it to the closure.
553 closure.push(n);
554 // Make a sanity check to ensure we don't waste too much time here.
555 if( closure.size() > 20) return false;
556 // This node is OK if:
557 // - it is a cast of an identical value
558 // - or it is a phi node (then we add its inputs to the worklist)
559 // Otherwise, the node is not OK, and we presume the cast is not invariant
560 if( n->is_ConstraintCast() ){
561 worklist.push(n->in(1));
562 } else if( n->is_Phi() ) {
563 for( uint i = 1; i < n->req(); i++ ) {
564 worklist.push(n->in(i));
565 }
566 } else {
567 return false;
568 }
569 }
570 }
572 // Quit when the worklist is empty, and we've found no offending nodes.
573 return true;
574 }
576 //------------------------------Ideal_DU_postCCP-------------------------------
577 // Find any cast-away of null-ness and keep its control. Null cast-aways are
578 // going away in this pass and we need to make this memory op depend on the
579 // gating null check.
581 // I tried to leave the CastPP's in. This makes the graph more accurate in
582 // some sense; we get to keep around the knowledge that an oop is not-null
583 // after some test. Alas, the CastPP's interfere with GVN (some values are
584 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
585 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
586 // some of the more trivial cases in the optimizer. Removing more useless
587 // Phi's started allowing Loads to illegally float above null checks. I gave
588 // up on this approach. CNC 10/20/2000
589 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
590 Node *ctr = in(MemNode::Control);
591 Node *mem = in(MemNode::Memory);
592 Node *adr = in(MemNode::Address);
593 Node *skipped_cast = NULL;
594 // Need a null check? Regular static accesses do not because they are
595 // from constant addresses. Array ops are gated by the range check (which
596 // always includes a NULL check). Just check field ops.
597 if( !ctr ) {
598 // Scan upwards for the highest location we can place this memory op.
599 while( true ) {
600 switch( adr->Opcode() ) {
602 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
603 adr = adr->in(AddPNode::Base);
604 continue;
606 case Op_CastPP:
607 // If the CastPP is useless, just peek on through it.
608 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
609 // Remember the cast that we've peeked though. If we peek
610 // through more than one, then we end up remembering the highest
611 // one, that is, if in a loop, the one closest to the top.
612 skipped_cast = adr;
613 adr = adr->in(1);
614 continue;
615 }
616 // CastPP is going away in this pass! We need this memory op to be
617 // control-dependent on the test that is guarding the CastPP.
618 ccp->hash_delete(this);
619 set_req(MemNode::Control, adr->in(0));
620 ccp->hash_insert(this);
621 return this;
623 case Op_Phi:
624 // Attempt to float above a Phi to some dominating point.
625 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
626 // If we've already peeked through a Cast (which could have set the
627 // control), we can't float above a Phi, because the skipped Cast
628 // may not be loop invariant.
629 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
630 adr = adr->in(1);
631 continue;
632 }
633 }
635 // Intentional fallthrough!
637 // No obvious dominating point. The mem op is pinned below the Phi
638 // by the Phi itself. If the Phi goes away (no true value is merged)
639 // then the mem op can float, but not indefinitely. It must be pinned
640 // behind the controls leading to the Phi.
641 case Op_CheckCastPP:
642 // These usually stick around to change address type, however a
643 // useless one can be elided and we still need to pick up a control edge
644 if (adr->in(0) == NULL) {
645 // This CheckCastPP node has NO control and is likely useless. But we
646 // need check further up the ancestor chain for a control input to keep
647 // the node in place. 4959717.
648 skipped_cast = adr;
649 adr = adr->in(1);
650 continue;
651 }
652 ccp->hash_delete(this);
653 set_req(MemNode::Control, adr->in(0));
654 ccp->hash_insert(this);
655 return this;
657 // List of "safe" opcodes; those that implicitly block the memory
658 // op below any null check.
659 case Op_CastX2P: // no null checks on native pointers
660 case Op_Parm: // 'this' pointer is not null
661 case Op_LoadP: // Loading from within a klass
662 case Op_LoadKlass: // Loading from within a klass
663 case Op_ConP: // Loading from a klass
664 case Op_CreateEx: // Sucking up the guts of an exception oop
665 case Op_Con: // Reading from TLS
666 case Op_CMoveP: // CMoveP is pinned
667 break; // No progress
669 case Op_Proj: // Direct call to an allocation routine
670 case Op_SCMemProj: // Memory state from store conditional ops
671 #ifdef ASSERT
672 {
673 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
674 const Node* call = adr->in(0);
675 if (call->is_CallStaticJava()) {
676 const CallStaticJavaNode* call_java = call->as_CallStaticJava();
677 const TypeTuple *r = call_java->tf()->range();
678 assert(r->cnt() > TypeFunc::Parms, "must return value");
679 const Type* ret_type = r->field_at(TypeFunc::Parms);
680 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
681 // We further presume that this is one of
682 // new_instance_Java, new_array_Java, or
683 // the like, but do not assert for this.
684 } else if (call->is_Allocate()) {
685 // similar case to new_instance_Java, etc.
686 } else if (!call->is_CallLeaf()) {
687 // Projections from fetch_oop (OSR) are allowed as well.
688 ShouldNotReachHere();
689 }
690 }
691 #endif
692 break;
693 default:
694 ShouldNotReachHere();
695 }
696 break;
697 }
698 }
700 return NULL; // No progress
701 }
704 //=============================================================================
705 uint LoadNode::size_of() const { return sizeof(*this); }
706 uint LoadNode::cmp( const Node &n ) const
707 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
708 const Type *LoadNode::bottom_type() const { return _type; }
709 uint LoadNode::ideal_reg() const {
710 return Matcher::base2reg[_type->base()];
711 }
713 #ifndef PRODUCT
714 void LoadNode::dump_spec(outputStream *st) const {
715 MemNode::dump_spec(st);
716 if( !Verbose && !WizardMode ) {
717 // standard dump does this in Verbose and WizardMode
718 st->print(" #"); _type->dump_on(st);
719 }
720 }
721 #endif
724 //----------------------------LoadNode::make-----------------------------------
725 // Polymorphic factory method:
726 LoadNode *LoadNode::make( Compile *C, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
727 // sanity check the alias category against the created node type
728 assert(!(adr_type->isa_oopptr() &&
729 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
730 "use LoadKlassNode instead");
731 assert(!(adr_type->isa_aryptr() &&
732 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
733 "use LoadRangeNode instead");
734 switch (bt) {
735 case T_BOOLEAN:
736 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
737 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
738 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
739 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
740 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
741 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
742 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
743 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
744 case T_OBJECT: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
745 }
746 ShouldNotReachHere();
747 return (LoadNode*)NULL;
748 }
750 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
751 bool require_atomic = true;
752 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
753 }
758 //------------------------------hash-------------------------------------------
759 uint LoadNode::hash() const {
760 // unroll addition of interesting fields
761 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
762 }
764 //---------------------------can_see_stored_value------------------------------
765 // This routine exists to make sure this set of tests is done the same
766 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
767 // will change the graph shape in a way which makes memory alive twice at the
768 // same time (uses the Oracle model of aliasing), then some
769 // LoadXNode::Identity will fold things back to the equivalence-class model
770 // of aliasing.
771 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
772 Node* ld_adr = in(MemNode::Address);
774 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
775 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
776 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
777 atp->field() != NULL && !atp->field()->is_volatile()) {
778 uint alias_idx = atp->index();
779 bool final = atp->field()->is_final();
780 Node* result = NULL;
781 Node* current = st;
782 // Skip through chains of MemBarNodes checking the MergeMems for
783 // new states for the slice of this load. Stop once any other
784 // kind of node is encountered. Loads from final memory can skip
785 // through any kind of MemBar but normal loads shouldn't skip
786 // through MemBarAcquire since the could allow them to move out of
787 // a synchronized region.
788 while (current->is_Proj()) {
789 int opc = current->in(0)->Opcode();
790 if ((final && opc == Op_MemBarAcquire) ||
791 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
792 Node* mem = current->in(0)->in(TypeFunc::Memory);
793 if (mem->is_MergeMem()) {
794 MergeMemNode* merge = mem->as_MergeMem();
795 Node* new_st = merge->memory_at(alias_idx);
796 if (new_st == merge->base_memory()) {
797 // Keep searching
798 current = merge->base_memory();
799 continue;
800 }
801 // Save the new memory state for the slice and fall through
802 // to exit.
803 result = new_st;
804 }
805 }
806 break;
807 }
808 if (result != NULL) {
809 st = result;
810 }
811 }
814 // Loop around twice in the case Load -> Initialize -> Store.
815 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
816 for (int trip = 0; trip <= 1; trip++) {
818 if (st->is_Store()) {
819 Node* st_adr = st->in(MemNode::Address);
820 if (!phase->eqv(st_adr, ld_adr)) {
821 // Try harder before giving up... Match raw and non-raw pointers.
822 intptr_t st_off = 0;
823 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
824 if (alloc == NULL) return NULL;
825 intptr_t ld_off = 0;
826 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
827 if (alloc != allo2) return NULL;
828 if (ld_off != st_off) return NULL;
829 // At this point we have proven something like this setup:
830 // A = Allocate(...)
831 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
832 // S = StoreQ(, AddP(, A.Parm , #Off), V)
833 // (Actually, we haven't yet proven the Q's are the same.)
834 // In other words, we are loading from a casted version of
835 // the same pointer-and-offset that we stored to.
836 // Thus, we are able to replace L by V.
837 }
838 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
839 if (store_Opcode() != st->Opcode())
840 return NULL;
841 return st->in(MemNode::ValueIn);
842 }
844 intptr_t offset = 0; // scratch
846 // A load from a freshly-created object always returns zero.
847 // (This can happen after LoadNode::Ideal resets the load's memory input
848 // to find_captured_store, which returned InitializeNode::zero_memory.)
849 if (st->is_Proj() && st->in(0)->is_Allocate() &&
850 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
851 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
852 // return a zero value for the load's basic type
853 // (This is one of the few places where a generic PhaseTransform
854 // can create new nodes. Think of it as lazily manifesting
855 // virtually pre-existing constants.)
856 return phase->zerocon(memory_type());
857 }
859 // A load from an initialization barrier can match a captured store.
860 if (st->is_Proj() && st->in(0)->is_Initialize()) {
861 InitializeNode* init = st->in(0)->as_Initialize();
862 AllocateNode* alloc = init->allocation();
863 if (alloc != NULL &&
864 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
865 // examine a captured store value
866 st = init->find_captured_store(offset, memory_size(), phase);
867 if (st != NULL)
868 continue; // take one more trip around
869 }
870 }
872 break;
873 }
875 return NULL;
876 }
878 //----------------------is_instance_field_load_with_local_phi------------------
879 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
880 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
881 in(MemNode::Address)->is_AddP() ) {
882 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
883 // Only instances.
884 if( t_oop != NULL && t_oop->is_instance_field() &&
885 t_oop->offset() != Type::OffsetBot &&
886 t_oop->offset() != Type::OffsetTop) {
887 return true;
888 }
889 }
890 return false;
891 }
893 //------------------------------Identity---------------------------------------
894 // Loads are identity if previous store is to same address
895 Node *LoadNode::Identity( PhaseTransform *phase ) {
896 // If the previous store-maker is the right kind of Store, and the store is
897 // to the same address, then we are equal to the value stored.
898 Node* mem = in(MemNode::Memory);
899 Node* value = can_see_stored_value(mem, phase);
900 if( value ) {
901 // byte, short & char stores truncate naturally.
902 // A load has to load the truncated value which requires
903 // some sort of masking operation and that requires an
904 // Ideal call instead of an Identity call.
905 if (memory_size() < BytesPerInt) {
906 // If the input to the store does not fit with the load's result type,
907 // it must be truncated via an Ideal call.
908 if (!phase->type(value)->higher_equal(phase->type(this)))
909 return this;
910 }
911 // (This works even when value is a Con, but LoadNode::Value
912 // usually runs first, producing the singleton type of the Con.)
913 return value;
914 }
916 // Search for an existing data phi which was generated before for the same
917 // instance's field to avoid infinite genertion of phis in a loop.
918 Node *region = mem->in(0);
919 if (is_instance_field_load_with_local_phi(region)) {
920 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
921 int this_index = phase->C->get_alias_index(addr_t);
922 int this_offset = addr_t->offset();
923 int this_id = addr_t->is_oopptr()->instance_id();
924 const Type* this_type = bottom_type();
925 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
926 Node* phi = region->fast_out(i);
927 if (phi->is_Phi() && phi != mem &&
928 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
929 return phi;
930 }
931 }
932 }
934 return this;
935 }
938 // Returns true if the AliasType refers to the field that holds the
939 // cached box array. Currently only handles the IntegerCache case.
940 static bool is_autobox_cache(Compile::AliasType* atp) {
941 if (atp != NULL && atp->field() != NULL) {
942 ciField* field = atp->field();
943 ciSymbol* klass = field->holder()->name();
944 if (field->name() == ciSymbol::cache_field_name() &&
945 field->holder()->uses_default_loader() &&
946 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
947 return true;
948 }
949 }
950 return false;
951 }
953 // Fetch the base value in the autobox array
954 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
955 if (atp != NULL && atp->field() != NULL) {
956 ciField* field = atp->field();
957 ciSymbol* klass = field->holder()->name();
958 if (field->name() == ciSymbol::cache_field_name() &&
959 field->holder()->uses_default_loader() &&
960 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
961 assert(field->is_constant(), "what?");
962 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
963 // Fetch the box object at the base of the array and get its value
964 ciInstance* box = array->obj_at(0)->as_instance();
965 ciInstanceKlass* ik = box->klass()->as_instance_klass();
966 if (ik->nof_nonstatic_fields() == 1) {
967 // This should be true nonstatic_field_at requires calling
968 // nof_nonstatic_fields so check it anyway
969 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
970 cache_offset = c.as_int();
971 }
972 return true;
973 }
974 }
975 return false;
976 }
978 // Returns true if the AliasType refers to the value field of an
979 // autobox object. Currently only handles Integer.
980 static bool is_autobox_object(Compile::AliasType* atp) {
981 if (atp != NULL && atp->field() != NULL) {
982 ciField* field = atp->field();
983 ciSymbol* klass = field->holder()->name();
984 if (field->name() == ciSymbol::value_name() &&
985 field->holder()->uses_default_loader() &&
986 klass == ciSymbol::java_lang_Integer()) {
987 return true;
988 }
989 }
990 return false;
991 }
994 // We're loading from an object which has autobox behaviour.
995 // If this object is result of a valueOf call we'll have a phi
996 // merging a newly allocated object and a load from the cache.
997 // We want to replace this load with the original incoming
998 // argument to the valueOf call.
999 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1000 Node* base = in(Address)->in(AddPNode::Base);
1001 if (base->is_Phi() && base->req() == 3) {
1002 AllocateNode* allocation = NULL;
1003 int allocation_index = -1;
1004 int load_index = -1;
1005 for (uint i = 1; i < base->req(); i++) {
1006 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1007 if (allocation != NULL) {
1008 allocation_index = i;
1009 load_index = 3 - allocation_index;
1010 break;
1011 }
1012 }
1013 LoadNode* load = NULL;
1014 if (allocation != NULL && base->in(load_index)->is_Load()) {
1015 load = base->in(load_index)->as_Load();
1016 }
1017 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1018 // Push the loads from the phi that comes from valueOf up
1019 // through it to allow elimination of the loads and the recovery
1020 // of the original value.
1021 Node* mem_phi = in(Memory);
1022 Node* offset = in(Address)->in(AddPNode::Offset);
1024 Node* in1 = clone();
1025 Node* in1_addr = in1->in(Address)->clone();
1026 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1027 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1028 in1_addr->set_req(AddPNode::Offset, offset);
1029 in1->set_req(0, base->in(allocation_index));
1030 in1->set_req(Address, in1_addr);
1031 in1->set_req(Memory, mem_phi->in(allocation_index));
1033 Node* in2 = clone();
1034 Node* in2_addr = in2->in(Address)->clone();
1035 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1036 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1037 in2_addr->set_req(AddPNode::Offset, offset);
1038 in2->set_req(0, base->in(load_index));
1039 in2->set_req(Address, in2_addr);
1040 in2->set_req(Memory, mem_phi->in(load_index));
1042 in1_addr = phase->transform(in1_addr);
1043 in1 = phase->transform(in1);
1044 in2_addr = phase->transform(in2_addr);
1045 in2 = phase->transform(in2);
1047 PhiNode* result = PhiNode::make_blank(base->in(0), this);
1048 result->set_req(allocation_index, in1);
1049 result->set_req(load_index, in2);
1050 return result;
1051 }
1052 } else if (base->is_Load()) {
1053 // Eliminate the load of Integer.value for integers from the cache
1054 // array by deriving the value from the index into the array.
1055 // Capture the offset of the load and then reverse the computation.
1056 Node* load_base = base->in(Address)->in(AddPNode::Base);
1057 if (load_base != NULL) {
1058 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1059 intptr_t cache_offset;
1060 int shift = -1;
1061 Node* cache = NULL;
1062 if (is_autobox_cache(atp)) {
1063 shift = exact_log2(type2aelembytes(T_OBJECT));
1064 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1065 }
1066 if (cache != NULL && base->in(Address)->is_AddP()) {
1067 Node* elements[4];
1068 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1069 int cache_low;
1070 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1071 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1072 // Add up all the offsets making of the address of the load
1073 Node* result = elements[0];
1074 for (int i = 1; i < count; i++) {
1075 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1076 }
1077 // Remove the constant offset from the address and then
1078 // remove the scaling of the offset to recover the original index.
1079 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1080 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1081 // Peel the shift off directly but wrap it in a dummy node
1082 // since Ideal can't return existing nodes
1083 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1084 } else {
1085 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1086 }
1087 #ifdef _LP64
1088 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1089 #endif
1090 return result;
1091 }
1092 }
1093 }
1094 }
1095 return NULL;
1096 }
1099 //------------------------------Ideal------------------------------------------
1100 // If the load is from Field memory and the pointer is non-null, we can
1101 // zero out the control input.
1102 // If the offset is constant and the base is an object allocation,
1103 // try to hook me up to the exact initializing store.
1104 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1105 Node* p = MemNode::Ideal_common(phase, can_reshape);
1106 if (p) return (p == NodeSentinel) ? NULL : p;
1108 Node* ctrl = in(MemNode::Control);
1109 Node* address = in(MemNode::Address);
1111 // Skip up past a SafePoint control. Cannot do this for Stores because
1112 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1113 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1114 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1115 ctrl = ctrl->in(0);
1116 set_req(MemNode::Control,ctrl);
1117 }
1119 // Check for useless control edge in some common special cases
1120 if (in(MemNode::Control) != NULL) {
1121 intptr_t ignore = 0;
1122 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1123 if (base != NULL
1124 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1125 && all_controls_dominate(base, phase->C->start())) {
1126 // A method-invariant, non-null address (constant or 'this' argument).
1127 set_req(MemNode::Control, NULL);
1128 }
1129 }
1131 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
1132 Node* base = in(Address)->in(AddPNode::Base);
1133 if (base != NULL) {
1134 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1135 if (is_autobox_object(atp)) {
1136 Node* result = eliminate_autobox(phase);
1137 if (result != NULL) return result;
1138 }
1139 }
1140 }
1142 Node* mem = in(MemNode::Memory);
1143 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1145 if (addr_t != NULL) {
1146 // try to optimize our memory input
1147 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1148 if (opt_mem != mem) {
1149 set_req(MemNode::Memory, opt_mem);
1150 return this;
1151 }
1152 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1153 if (can_reshape && opt_mem->is_Phi() &&
1154 (t_oop != NULL) && t_oop->is_instance_field()) {
1155 assert(t_oop->offset() != Type::OffsetBot && t_oop->offset() != Type::OffsetTop, "");
1156 Node *region = opt_mem->in(0);
1157 uint cnt = opt_mem->req();
1158 for( uint i = 1; i < cnt; i++ ) {
1159 Node *in = opt_mem->in(i);
1160 if( in == NULL ) {
1161 region = NULL; // Wait stable graph
1162 break;
1163 }
1164 }
1165 if (region != NULL) {
1166 // Check for loop invariant.
1167 if (cnt == 3) {
1168 for( uint i = 1; i < cnt; i++ ) {
1169 Node *in = opt_mem->in(i);
1170 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1171 if (m == opt_mem) {
1172 set_req(MemNode::Memory, opt_mem->in(cnt - i)); // Skip this phi.
1173 return this;
1174 }
1175 }
1176 }
1177 // Split through Phi (see original code in loopopts.cpp).
1178 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1180 // Do nothing here if Identity will find a value
1181 // (to avoid infinite chain of value phis generation).
1182 if ( !phase->eqv(this, this->Identity(phase)) )
1183 return NULL;
1185 const Type* this_type = this->bottom_type();
1186 int this_index = phase->C->get_alias_index(addr_t);
1187 int this_offset = addr_t->offset();
1188 int this_iid = addr_t->is_oopptr()->instance_id();
1189 int wins = 0;
1190 PhaseIterGVN *igvn = phase->is_IterGVN();
1191 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1192 for( uint i = 1; i < region->req(); i++ ) {
1193 Node *x;
1194 Node* the_clone = NULL;
1195 if( region->in(i) == phase->C->top() ) {
1196 x = phase->C->top(); // Dead path? Use a dead data op
1197 } else {
1198 x = this->clone(); // Else clone up the data op
1199 the_clone = x; // Remember for possible deletion.
1200 // Alter data node to use pre-phi inputs
1201 if( this->in(0) == region ) {
1202 x->set_req( 0, region->in(i) );
1203 } else {
1204 x->set_req( 0, NULL );
1205 }
1206 for( uint j = 1; j < this->req(); j++ ) {
1207 Node *in = this->in(j);
1208 if( in->is_Phi() && in->in(0) == region )
1209 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1210 }
1211 }
1212 // Check for a 'win' on some paths
1213 const Type *t = x->Value(igvn);
1215 bool singleton = t->singleton();
1217 // See comments in PhaseIdealLoop::split_thru_phi().
1218 if( singleton && t == Type::TOP ) {
1219 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1220 }
1222 if( singleton ) {
1223 wins++;
1224 x = igvn->makecon(t);
1225 } else {
1226 // We now call Identity to try to simplify the cloned node.
1227 // Note that some Identity methods call phase->type(this).
1228 // Make sure that the type array is big enough for
1229 // our new node, even though we may throw the node away.
1230 // (This tweaking with igvn only works because x is a new node.)
1231 igvn->set_type(x, t);
1232 Node *y = x->Identity(igvn);
1233 if( y != x ) {
1234 wins++;
1235 x = y;
1236 } else {
1237 y = igvn->hash_find(x);
1238 if( y ) {
1239 wins++;
1240 x = y;
1241 } else {
1242 // Else x is a new node we are keeping
1243 // We do not need register_new_node_with_optimizer
1244 // because set_type has already been called.
1245 igvn->_worklist.push(x);
1246 }
1247 }
1248 }
1249 if (x != the_clone && the_clone != NULL)
1250 igvn->remove_dead_node(the_clone);
1251 phi->set_req(i, x);
1252 }
1253 if( wins > 0 ) {
1254 // Record Phi
1255 igvn->register_new_node_with_optimizer(phi);
1256 return phi;
1257 } else {
1258 igvn->remove_dead_node(phi);
1259 }
1260 }
1261 }
1262 }
1264 // Check for prior store with a different base or offset; make Load
1265 // independent. Skip through any number of them. Bail out if the stores
1266 // are in an endless dead cycle and report no progress. This is a key
1267 // transform for Reflection. However, if after skipping through the Stores
1268 // we can't then fold up against a prior store do NOT do the transform as
1269 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1270 // array memory alive twice: once for the hoisted Load and again after the
1271 // bypassed Store. This situation only works if EVERYBODY who does
1272 // anti-dependence work knows how to bypass. I.e. we need all
1273 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1274 // the alias index stuff. So instead, peek through Stores and IFF we can
1275 // fold up, do so.
1276 Node* prev_mem = find_previous_store(phase);
1277 // Steps (a), (b): Walk past independent stores to find an exact match.
1278 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1279 // (c) See if we can fold up on the spot, but don't fold up here.
1280 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
1281 // just return a prior value, which is done by Identity calls.
1282 if (can_see_stored_value(prev_mem, phase)) {
1283 // Make ready for step (d):
1284 set_req(MemNode::Memory, prev_mem);
1285 return this;
1286 }
1287 }
1289 return NULL; // No further progress
1290 }
1292 // Helper to recognize certain Klass fields which are invariant across
1293 // some group of array types (e.g., int[] or all T[] where T < Object).
1294 const Type*
1295 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1296 ciKlass* klass) const {
1297 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1298 // The field is Klass::_modifier_flags. Return its (constant) value.
1299 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1300 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1301 return TypeInt::make(klass->modifier_flags());
1302 }
1303 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1304 // The field is Klass::_access_flags. Return its (constant) value.
1305 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1306 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1307 return TypeInt::make(klass->access_flags());
1308 }
1309 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1310 // The field is Klass::_layout_helper. Return its constant value if known.
1311 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1312 return TypeInt::make(klass->layout_helper());
1313 }
1315 // No match.
1316 return NULL;
1317 }
1319 //------------------------------Value-----------------------------------------
1320 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1321 // Either input is TOP ==> the result is TOP
1322 Node* mem = in(MemNode::Memory);
1323 const Type *t1 = phase->type(mem);
1324 if (t1 == Type::TOP) return Type::TOP;
1325 Node* adr = in(MemNode::Address);
1326 const TypePtr* tp = phase->type(adr)->isa_ptr();
1327 if (tp == NULL || tp->empty()) return Type::TOP;
1328 int off = tp->offset();
1329 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1331 // Try to guess loaded type from pointer type
1332 if (tp->base() == Type::AryPtr) {
1333 const Type *t = tp->is_aryptr()->elem();
1334 // Don't do this for integer types. There is only potential profit if
1335 // the element type t is lower than _type; that is, for int types, if _type is
1336 // more restrictive than t. This only happens here if one is short and the other
1337 // char (both 16 bits), and in those cases we've made an intentional decision
1338 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1339 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1340 //
1341 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1342 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1343 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1344 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1345 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1346 // In fact, that could have been the original type of p1, and p1 could have
1347 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1348 // expression (LShiftL quux 3) independently optimized to the constant 8.
1349 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1350 && Opcode() != Op_LoadKlass) {
1351 // t might actually be lower than _type, if _type is a unique
1352 // concrete subclass of abstract class t.
1353 // Make sure the reference is not into the header, by comparing
1354 // the offset against the offset of the start of the array's data.
1355 // Different array types begin at slightly different offsets (12 vs. 16).
1356 // We choose T_BYTE as an example base type that is least restrictive
1357 // as to alignment, which will therefore produce the smallest
1358 // possible base offset.
1359 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1360 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1361 const Type* jt = t->join(_type);
1362 // In any case, do not allow the join, per se, to empty out the type.
1363 if (jt->empty() && !t->empty()) {
1364 // This can happen if a interface-typed array narrows to a class type.
1365 jt = _type;
1366 }
1368 if (EliminateAutoBox) {
1369 // The pointers in the autobox arrays are always non-null
1370 Node* base = in(Address)->in(AddPNode::Base);
1371 if (base != NULL) {
1372 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1373 if (is_autobox_cache(atp)) {
1374 return jt->join(TypePtr::NOTNULL)->is_ptr();
1375 }
1376 }
1377 }
1378 return jt;
1379 }
1380 }
1381 } else if (tp->base() == Type::InstPtr) {
1382 assert( off != Type::OffsetBot ||
1383 // arrays can be cast to Objects
1384 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1385 // unsafe field access may not have a constant offset
1386 phase->C->has_unsafe_access(),
1387 "Field accesses must be precise" );
1388 // For oop loads, we expect the _type to be precise
1389 } else if (tp->base() == Type::KlassPtr) {
1390 assert( off != Type::OffsetBot ||
1391 // arrays can be cast to Objects
1392 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1393 // also allow array-loading from the primary supertype
1394 // array during subtype checks
1395 Opcode() == Op_LoadKlass,
1396 "Field accesses must be precise" );
1397 // For klass/static loads, we expect the _type to be precise
1398 }
1400 const TypeKlassPtr *tkls = tp->isa_klassptr();
1401 if (tkls != NULL && !StressReflectiveCode) {
1402 ciKlass* klass = tkls->klass();
1403 if (klass->is_loaded() && tkls->klass_is_exact()) {
1404 // We are loading a field from a Klass metaobject whose identity
1405 // is known at compile time (the type is "exact" or "precise").
1406 // Check for fields we know are maintained as constants by the VM.
1407 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1408 // The field is Klass::_super_check_offset. Return its (constant) value.
1409 // (Folds up type checking code.)
1410 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1411 return TypeInt::make(klass->super_check_offset());
1412 }
1413 // Compute index into primary_supers array
1414 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1415 // Check for overflowing; use unsigned compare to handle the negative case.
1416 if( depth < ciKlass::primary_super_limit() ) {
1417 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1418 // (Folds up type checking code.)
1419 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1420 ciKlass *ss = klass->super_of_depth(depth);
1421 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1422 }
1423 const Type* aift = load_array_final_field(tkls, klass);
1424 if (aift != NULL) return aift;
1425 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1426 && klass->is_array_klass()) {
1427 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1428 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1429 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1430 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1431 }
1432 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1433 // The field is Klass::_java_mirror. Return its (constant) value.
1434 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1435 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1436 return TypeInstPtr::make(klass->java_mirror());
1437 }
1438 }
1440 // We can still check if we are loading from the primary_supers array at a
1441 // shallow enough depth. Even though the klass is not exact, entries less
1442 // than or equal to its super depth are correct.
1443 if (klass->is_loaded() ) {
1444 ciType *inner = klass->klass();
1445 while( inner->is_obj_array_klass() )
1446 inner = inner->as_obj_array_klass()->base_element_type();
1447 if( inner->is_instance_klass() &&
1448 !inner->as_instance_klass()->flags().is_interface() ) {
1449 // Compute index into primary_supers array
1450 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1451 // Check for overflowing; use unsigned compare to handle the negative case.
1452 if( depth < ciKlass::primary_super_limit() &&
1453 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1454 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1455 // (Folds up type checking code.)
1456 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1457 ciKlass *ss = klass->super_of_depth(depth);
1458 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1459 }
1460 }
1461 }
1463 // If the type is enough to determine that the thing is not an array,
1464 // we can give the layout_helper a positive interval type.
1465 // This will help short-circuit some reflective code.
1466 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1467 && !klass->is_array_klass() // not directly typed as an array
1468 && !klass->is_interface() // specifically not Serializable & Cloneable
1469 && !klass->is_java_lang_Object() // not the supertype of all T[]
1470 ) {
1471 // Note: When interfaces are reliable, we can narrow the interface
1472 // test to (klass != Serializable && klass != Cloneable).
1473 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1474 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1475 // The key property of this type is that it folds up tests
1476 // for array-ness, since it proves that the layout_helper is positive.
1477 // Thus, a generic value like the basic object layout helper works fine.
1478 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1479 }
1480 }
1482 // If we are loading from a freshly-allocated object, produce a zero,
1483 // if the load is provably beyond the header of the object.
1484 // (Also allow a variable load from a fresh array to produce zero.)
1485 if (ReduceFieldZeroing) {
1486 Node* value = can_see_stored_value(mem,phase);
1487 if (value != NULL && value->is_Con())
1488 return value->bottom_type();
1489 }
1491 const TypeOopPtr *tinst = tp->isa_oopptr();
1492 if (tinst != NULL && tinst->is_instance_field()) {
1493 // If we have an instance type and our memory input is the
1494 // programs's initial memory state, there is no matching store,
1495 // so just return a zero of the appropriate type
1496 Node *mem = in(MemNode::Memory);
1497 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1498 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1499 return Type::get_zero_type(_type->basic_type());
1500 }
1501 }
1502 return _type;
1503 }
1505 //------------------------------match_edge-------------------------------------
1506 // Do we Match on this edge index or not? Match only the address.
1507 uint LoadNode::match_edge(uint idx) const {
1508 return idx == MemNode::Address;
1509 }
1511 //--------------------------LoadBNode::Ideal--------------------------------------
1512 //
1513 // If the previous store is to the same address as this load,
1514 // and the value stored was larger than a byte, replace this load
1515 // with the value stored truncated to a byte. If no truncation is
1516 // needed, the replacement is done in LoadNode::Identity().
1517 //
1518 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1519 Node* mem = in(MemNode::Memory);
1520 Node* value = can_see_stored_value(mem,phase);
1521 if( value && !phase->type(value)->higher_equal( _type ) ) {
1522 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1523 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1524 }
1525 // Identity call will handle the case where truncation is not needed.
1526 return LoadNode::Ideal(phase, can_reshape);
1527 }
1529 //--------------------------LoadCNode::Ideal--------------------------------------
1530 //
1531 // If the previous store is to the same address as this load,
1532 // and the value stored was larger than a char, replace this load
1533 // with the value stored truncated to a char. If no truncation is
1534 // needed, the replacement is done in LoadNode::Identity().
1535 //
1536 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1537 Node* mem = in(MemNode::Memory);
1538 Node* value = can_see_stored_value(mem,phase);
1539 if( value && !phase->type(value)->higher_equal( _type ) )
1540 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1541 // Identity call will handle the case where truncation is not needed.
1542 return LoadNode::Ideal(phase, can_reshape);
1543 }
1545 //--------------------------LoadSNode::Ideal--------------------------------------
1546 //
1547 // If the previous store is to the same address as this load,
1548 // and the value stored was larger than a short, replace this load
1549 // with the value stored truncated to a short. If no truncation is
1550 // needed, the replacement is done in LoadNode::Identity().
1551 //
1552 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1553 Node* mem = in(MemNode::Memory);
1554 Node* value = can_see_stored_value(mem,phase);
1555 if( value && !phase->type(value)->higher_equal( _type ) ) {
1556 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1557 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1558 }
1559 // Identity call will handle the case where truncation is not needed.
1560 return LoadNode::Ideal(phase, can_reshape);
1561 }
1563 //=============================================================================
1564 //------------------------------Value------------------------------------------
1565 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1566 // Either input is TOP ==> the result is TOP
1567 const Type *t1 = phase->type( in(MemNode::Memory) );
1568 if (t1 == Type::TOP) return Type::TOP;
1569 Node *adr = in(MemNode::Address);
1570 const Type *t2 = phase->type( adr );
1571 if (t2 == Type::TOP) return Type::TOP;
1572 const TypePtr *tp = t2->is_ptr();
1573 if (TypePtr::above_centerline(tp->ptr()) ||
1574 tp->ptr() == TypePtr::Null) return Type::TOP;
1576 // Return a more precise klass, if possible
1577 const TypeInstPtr *tinst = tp->isa_instptr();
1578 if (tinst != NULL) {
1579 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1580 int offset = tinst->offset();
1581 if (ik == phase->C->env()->Class_klass()
1582 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1583 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1584 // We are loading a special hidden field from a Class mirror object,
1585 // the field which points to the VM's Klass metaobject.
1586 ciType* t = tinst->java_mirror_type();
1587 // java_mirror_type returns non-null for compile-time Class constants.
1588 if (t != NULL) {
1589 // constant oop => constant klass
1590 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1591 return TypeKlassPtr::make(ciArrayKlass::make(t));
1592 }
1593 if (!t->is_klass()) {
1594 // a primitive Class (e.g., int.class) has NULL for a klass field
1595 return TypePtr::NULL_PTR;
1596 }
1597 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1598 return TypeKlassPtr::make(t->as_klass());
1599 }
1600 // non-constant mirror, so we can't tell what's going on
1601 }
1602 if( !ik->is_loaded() )
1603 return _type; // Bail out if not loaded
1604 if (offset == oopDesc::klass_offset_in_bytes()) {
1605 if (tinst->klass_is_exact()) {
1606 return TypeKlassPtr::make(ik);
1607 }
1608 // See if we can become precise: no subklasses and no interface
1609 // (Note: We need to support verified interfaces.)
1610 if (!ik->is_interface() && !ik->has_subklass()) {
1611 //assert(!UseExactTypes, "this code should be useless with exact types");
1612 // Add a dependence; if any subclass added we need to recompile
1613 if (!ik->is_final()) {
1614 // %%% should use stronger assert_unique_concrete_subtype instead
1615 phase->C->dependencies()->assert_leaf_type(ik);
1616 }
1617 // Return precise klass
1618 return TypeKlassPtr::make(ik);
1619 }
1621 // Return root of possible klass
1622 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1623 }
1624 }
1626 // Check for loading klass from an array
1627 const TypeAryPtr *tary = tp->isa_aryptr();
1628 if( tary != NULL ) {
1629 ciKlass *tary_klass = tary->klass();
1630 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1631 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1632 if (tary->klass_is_exact()) {
1633 return TypeKlassPtr::make(tary_klass);
1634 }
1635 ciArrayKlass *ak = tary->klass()->as_array_klass();
1636 // If the klass is an object array, we defer the question to the
1637 // array component klass.
1638 if( ak->is_obj_array_klass() ) {
1639 assert( ak->is_loaded(), "" );
1640 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1641 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1642 ciInstanceKlass* ik = base_k->as_instance_klass();
1643 // See if we can become precise: no subklasses and no interface
1644 if (!ik->is_interface() && !ik->has_subklass()) {
1645 //assert(!UseExactTypes, "this code should be useless with exact types");
1646 // Add a dependence; if any subclass added we need to recompile
1647 if (!ik->is_final()) {
1648 phase->C->dependencies()->assert_leaf_type(ik);
1649 }
1650 // Return precise array klass
1651 return TypeKlassPtr::make(ak);
1652 }
1653 }
1654 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1655 } else { // Found a type-array?
1656 //assert(!UseExactTypes, "this code should be useless with exact types");
1657 assert( ak->is_type_array_klass(), "" );
1658 return TypeKlassPtr::make(ak); // These are always precise
1659 }
1660 }
1661 }
1663 // Check for loading klass from an array klass
1664 const TypeKlassPtr *tkls = tp->isa_klassptr();
1665 if (tkls != NULL && !StressReflectiveCode) {
1666 ciKlass* klass = tkls->klass();
1667 if( !klass->is_loaded() )
1668 return _type; // Bail out if not loaded
1669 if( klass->is_obj_array_klass() &&
1670 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1671 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1672 // // Always returning precise element type is incorrect,
1673 // // e.g., element type could be object and array may contain strings
1674 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1676 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1677 // according to the element type's subclassing.
1678 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1679 }
1680 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1681 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1682 ciKlass* sup = klass->as_instance_klass()->super();
1683 // The field is Klass::_super. Return its (constant) value.
1684 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1685 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1686 }
1687 }
1689 // Bailout case
1690 return LoadNode::Value(phase);
1691 }
1693 //------------------------------Identity---------------------------------------
1694 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1695 // Also feed through the klass in Allocate(...klass...)._klass.
1696 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1697 Node* x = LoadNode::Identity(phase);
1698 if (x != this) return x;
1700 // Take apart the address into an oop and and offset.
1701 // Return 'this' if we cannot.
1702 Node* adr = in(MemNode::Address);
1703 intptr_t offset = 0;
1704 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1705 if (base == NULL) return this;
1706 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1707 if (toop == NULL) return this;
1709 // We can fetch the klass directly through an AllocateNode.
1710 // This works even if the klass is not constant (clone or newArray).
1711 if (offset == oopDesc::klass_offset_in_bytes()) {
1712 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1713 if (allocated_klass != NULL) {
1714 return allocated_klass;
1715 }
1716 }
1718 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1719 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1720 // See inline_native_Class_query for occurrences of these patterns.
1721 // Java Example: x.getClass().isAssignableFrom(y)
1722 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1723 //
1724 // This improves reflective code, often making the Class
1725 // mirror go completely dead. (Current exception: Class
1726 // mirrors may appear in debug info, but we could clean them out by
1727 // introducing a new debug info operator for klassOop.java_mirror).
1728 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1729 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1730 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1731 // We are loading a special hidden field from a Class mirror,
1732 // the field which points to its Klass or arrayKlass metaobject.
1733 if (base->is_Load()) {
1734 Node* adr2 = base->in(MemNode::Address);
1735 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1736 if (tkls != NULL && !tkls->empty()
1737 && (tkls->klass()->is_instance_klass() ||
1738 tkls->klass()->is_array_klass())
1739 && adr2->is_AddP()
1740 ) {
1741 int mirror_field = Klass::java_mirror_offset_in_bytes();
1742 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1743 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1744 }
1745 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1746 return adr2->in(AddPNode::Base);
1747 }
1748 }
1749 }
1750 }
1752 return this;
1753 }
1755 //------------------------------Value-----------------------------------------
1756 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1757 // Either input is TOP ==> the result is TOP
1758 const Type *t1 = phase->type( in(MemNode::Memory) );
1759 if( t1 == Type::TOP ) return Type::TOP;
1760 Node *adr = in(MemNode::Address);
1761 const Type *t2 = phase->type( adr );
1762 if( t2 == Type::TOP ) return Type::TOP;
1763 const TypePtr *tp = t2->is_ptr();
1764 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1765 const TypeAryPtr *tap = tp->isa_aryptr();
1766 if( !tap ) return _type;
1767 return tap->size();
1768 }
1770 //------------------------------Identity---------------------------------------
1771 // Feed through the length in AllocateArray(...length...)._length.
1772 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1773 Node* x = LoadINode::Identity(phase);
1774 if (x != this) return x;
1776 // Take apart the address into an oop and and offset.
1777 // Return 'this' if we cannot.
1778 Node* adr = in(MemNode::Address);
1779 intptr_t offset = 0;
1780 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1781 if (base == NULL) return this;
1782 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1783 if (tary == NULL) return this;
1785 // We can fetch the length directly through an AllocateArrayNode.
1786 // This works even if the length is not constant (clone or newArray).
1787 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1788 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
1789 if (allocated_length != NULL) {
1790 return allocated_length;
1791 }
1792 }
1794 return this;
1796 }
1797 //=============================================================================
1798 //---------------------------StoreNode::make-----------------------------------
1799 // Polymorphic factory method:
1800 StoreNode* StoreNode::make( Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1801 switch (bt) {
1802 case T_BOOLEAN:
1803 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1804 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1805 case T_CHAR:
1806 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1807 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1808 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1809 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1810 case T_ADDRESS:
1811 case T_OBJECT: return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1812 }
1813 ShouldNotReachHere();
1814 return (StoreNode*)NULL;
1815 }
1817 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1818 bool require_atomic = true;
1819 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1820 }
1823 //--------------------------bottom_type----------------------------------------
1824 const Type *StoreNode::bottom_type() const {
1825 return Type::MEMORY;
1826 }
1828 //------------------------------hash-------------------------------------------
1829 uint StoreNode::hash() const {
1830 // unroll addition of interesting fields
1831 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1833 // Since they are not commoned, do not hash them:
1834 return NO_HASH;
1835 }
1837 //------------------------------Ideal------------------------------------------
1838 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1839 // When a store immediately follows a relevant allocation/initialization,
1840 // try to capture it into the initialization, or hoist it above.
1841 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1842 Node* p = MemNode::Ideal_common(phase, can_reshape);
1843 if (p) return (p == NodeSentinel) ? NULL : p;
1845 Node* mem = in(MemNode::Memory);
1846 Node* address = in(MemNode::Address);
1848 // Back-to-back stores to same address? Fold em up.
1849 // Generally unsafe if I have intervening uses...
1850 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
1851 // Looking at a dead closed cycle of memory?
1852 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1854 assert(Opcode() == mem->Opcode() ||
1855 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1856 "no mismatched stores, except on raw memory");
1858 if (mem->outcnt() == 1 && // check for intervening uses
1859 mem->as_Store()->memory_size() <= this->memory_size()) {
1860 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1861 // For example, 'mem' might be the final state at a conditional return.
1862 // Or, 'mem' might be used by some node which is live at the same time
1863 // 'this' is live, which might be unschedulable. So, require exactly
1864 // ONE user, the 'this' store, until such time as we clone 'mem' for
1865 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1866 if (can_reshape) { // (%%% is this an anachronism?)
1867 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1868 phase->is_IterGVN());
1869 } else {
1870 // It's OK to do this in the parser, since DU info is always accurate,
1871 // and the parser always refers to nodes via SafePointNode maps.
1872 set_req(MemNode::Memory, mem->in(MemNode::Memory));
1873 }
1874 return this;
1875 }
1876 }
1878 // Capture an unaliased, unconditional, simple store into an initializer.
1879 // Or, if it is independent of the allocation, hoist it above the allocation.
1880 if (ReduceFieldZeroing && /*can_reshape &&*/
1881 mem->is_Proj() && mem->in(0)->is_Initialize()) {
1882 InitializeNode* init = mem->in(0)->as_Initialize();
1883 intptr_t offset = init->can_capture_store(this, phase);
1884 if (offset > 0) {
1885 Node* moved = init->capture_store(this, offset, phase);
1886 // If the InitializeNode captured me, it made a raw copy of me,
1887 // and I need to disappear.
1888 if (moved != NULL) {
1889 // %%% hack to ensure that Ideal returns a new node:
1890 mem = MergeMemNode::make(phase->C, mem);
1891 return mem; // fold me away
1892 }
1893 }
1894 }
1896 return NULL; // No further progress
1897 }
1899 //------------------------------Value-----------------------------------------
1900 const Type *StoreNode::Value( PhaseTransform *phase ) const {
1901 // Either input is TOP ==> the result is TOP
1902 const Type *t1 = phase->type( in(MemNode::Memory) );
1903 if( t1 == Type::TOP ) return Type::TOP;
1904 const Type *t2 = phase->type( in(MemNode::Address) );
1905 if( t2 == Type::TOP ) return Type::TOP;
1906 const Type *t3 = phase->type( in(MemNode::ValueIn) );
1907 if( t3 == Type::TOP ) return Type::TOP;
1908 return Type::MEMORY;
1909 }
1911 //------------------------------Identity---------------------------------------
1912 // Remove redundant stores:
1913 // Store(m, p, Load(m, p)) changes to m.
1914 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
1915 Node *StoreNode::Identity( PhaseTransform *phase ) {
1916 Node* mem = in(MemNode::Memory);
1917 Node* adr = in(MemNode::Address);
1918 Node* val = in(MemNode::ValueIn);
1920 // Load then Store? Then the Store is useless
1921 if (val->is_Load() &&
1922 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
1923 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
1924 val->as_Load()->store_Opcode() == Opcode()) {
1925 return mem;
1926 }
1928 // Two stores in a row of the same value?
1929 if (mem->is_Store() &&
1930 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
1931 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
1932 mem->Opcode() == Opcode()) {
1933 return mem;
1934 }
1936 // Store of zero anywhere into a freshly-allocated object?
1937 // Then the store is useless.
1938 // (It must already have been captured by the InitializeNode.)
1939 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
1940 // a newly allocated object is already all-zeroes everywhere
1941 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
1942 return mem;
1943 }
1945 // the store may also apply to zero-bits in an earlier object
1946 Node* prev_mem = find_previous_store(phase);
1947 // Steps (a), (b): Walk past independent stores to find an exact match.
1948 if (prev_mem != NULL) {
1949 Node* prev_val = can_see_stored_value(prev_mem, phase);
1950 if (prev_val != NULL && phase->eqv(prev_val, val)) {
1951 // prev_val and val might differ by a cast; it would be good
1952 // to keep the more informative of the two.
1953 return mem;
1954 }
1955 }
1956 }
1958 return this;
1959 }
1961 //------------------------------match_edge-------------------------------------
1962 // Do we Match on this edge index or not? Match only memory & value
1963 uint StoreNode::match_edge(uint idx) const {
1964 return idx == MemNode::Address || idx == MemNode::ValueIn;
1965 }
1967 //------------------------------cmp--------------------------------------------
1968 // Do not common stores up together. They generally have to be split
1969 // back up anyways, so do not bother.
1970 uint StoreNode::cmp( const Node &n ) const {
1971 return (&n == this); // Always fail except on self
1972 }
1974 //------------------------------Ideal_masked_input-----------------------------
1975 // Check for a useless mask before a partial-word store
1976 // (StoreB ... (AndI valIn conIa) )
1977 // If (conIa & mask == mask) this simplifies to
1978 // (StoreB ... (valIn) )
1979 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
1980 Node *val = in(MemNode::ValueIn);
1981 if( val->Opcode() == Op_AndI ) {
1982 const TypeInt *t = phase->type( val->in(2) )->isa_int();
1983 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
1984 set_req(MemNode::ValueIn, val->in(1));
1985 return this;
1986 }
1987 }
1988 return NULL;
1989 }
1992 //------------------------------Ideal_sign_extended_input----------------------
1993 // Check for useless sign-extension before a partial-word store
1994 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
1995 // If (conIL == conIR && conIR <= num_bits) this simplifies to
1996 // (StoreB ... (valIn) )
1997 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
1998 Node *val = in(MemNode::ValueIn);
1999 if( val->Opcode() == Op_RShiftI ) {
2000 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2001 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2002 Node *shl = val->in(1);
2003 if( shl->Opcode() == Op_LShiftI ) {
2004 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2005 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2006 set_req(MemNode::ValueIn, shl->in(1));
2007 return this;
2008 }
2009 }
2010 }
2011 }
2012 return NULL;
2013 }
2015 //------------------------------value_never_loaded-----------------------------------
2016 // Determine whether there are any possible loads of the value stored.
2017 // For simplicity, we actually check if there are any loads from the
2018 // address stored to, not just for loads of the value stored by this node.
2019 //
2020 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2021 Node *adr = in(Address);
2022 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2023 if (adr_oop == NULL)
2024 return false;
2025 if (!adr_oop->is_instance_field())
2026 return false; // if not a distinct instance, there may be aliases of the address
2027 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2028 Node *use = adr->fast_out(i);
2029 int opc = use->Opcode();
2030 if (use->is_Load() || use->is_LoadStore()) {
2031 return false;
2032 }
2033 }
2034 return true;
2035 }
2037 //=============================================================================
2038 //------------------------------Ideal------------------------------------------
2039 // If the store is from an AND mask that leaves the low bits untouched, then
2040 // we can skip the AND operation. If the store is from a sign-extension
2041 // (a left shift, then right shift) we can skip both.
2042 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2043 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2044 if( progress != NULL ) return progress;
2046 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2047 if( progress != NULL ) return progress;
2049 // Finally check the default case
2050 return StoreNode::Ideal(phase, can_reshape);
2051 }
2053 //=============================================================================
2054 //------------------------------Ideal------------------------------------------
2055 // If the store is from an AND mask that leaves the low bits untouched, then
2056 // we can skip the AND operation
2057 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2058 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2059 if( progress != NULL ) return progress;
2061 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2062 if( progress != NULL ) return progress;
2064 // Finally check the default case
2065 return StoreNode::Ideal(phase, can_reshape);
2066 }
2068 //=============================================================================
2069 //------------------------------Identity---------------------------------------
2070 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2071 // No need to card mark when storing a null ptr
2072 Node* my_store = in(MemNode::OopStore);
2073 if (my_store->is_Store()) {
2074 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2075 if( t1 == TypePtr::NULL_PTR ) {
2076 return in(MemNode::Memory);
2077 }
2078 }
2079 return this;
2080 }
2082 //------------------------------Value-----------------------------------------
2083 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2084 // Either input is TOP ==> the result is TOP
2085 const Type *t = phase->type( in(MemNode::Memory) );
2086 if( t == Type::TOP ) return Type::TOP;
2087 t = phase->type( in(MemNode::Address) );
2088 if( t == Type::TOP ) return Type::TOP;
2089 t = phase->type( in(MemNode::ValueIn) );
2090 if( t == Type::TOP ) return Type::TOP;
2091 // If extra input is TOP ==> the result is TOP
2092 t = phase->type( in(MemNode::OopStore) );
2093 if( t == Type::TOP ) return Type::TOP;
2095 return StoreNode::Value( phase );
2096 }
2099 //=============================================================================
2100 //----------------------------------SCMemProjNode------------------------------
2101 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2102 {
2103 return bottom_type();
2104 }
2106 //=============================================================================
2107 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2108 init_req(MemNode::Control, c );
2109 init_req(MemNode::Memory , mem);
2110 init_req(MemNode::Address, adr);
2111 init_req(MemNode::ValueIn, val);
2112 init_req( ExpectedIn, ex );
2113 init_class_id(Class_LoadStore);
2115 }
2117 //=============================================================================
2118 //-------------------------------adr_type--------------------------------------
2119 // Do we Match on this edge index or not? Do not match memory
2120 const TypePtr* ClearArrayNode::adr_type() const {
2121 Node *adr = in(3);
2122 return MemNode::calculate_adr_type(adr->bottom_type());
2123 }
2125 //------------------------------match_edge-------------------------------------
2126 // Do we Match on this edge index or not? Do not match memory
2127 uint ClearArrayNode::match_edge(uint idx) const {
2128 return idx > 1;
2129 }
2131 //------------------------------Identity---------------------------------------
2132 // Clearing a zero length array does nothing
2133 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2134 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2135 }
2137 //------------------------------Idealize---------------------------------------
2138 // Clearing a short array is faster with stores
2139 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2140 const int unit = BytesPerLong;
2141 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2142 if (!t) return NULL;
2143 if (!t->is_con()) return NULL;
2144 intptr_t raw_count = t->get_con();
2145 intptr_t size = raw_count;
2146 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2147 // Clearing nothing uses the Identity call.
2148 // Negative clears are possible on dead ClearArrays
2149 // (see jck test stmt114.stmt11402.val).
2150 if (size <= 0 || size % unit != 0) return NULL;
2151 intptr_t count = size / unit;
2152 // Length too long; use fast hardware clear
2153 if (size > Matcher::init_array_short_size) return NULL;
2154 Node *mem = in(1);
2155 if( phase->type(mem)==Type::TOP ) return NULL;
2156 Node *adr = in(3);
2157 const Type* at = phase->type(adr);
2158 if( at==Type::TOP ) return NULL;
2159 const TypePtr* atp = at->isa_ptr();
2160 // adjust atp to be the correct array element address type
2161 if (atp == NULL) atp = TypePtr::BOTTOM;
2162 else atp = atp->add_offset(Type::OffsetBot);
2163 // Get base for derived pointer purposes
2164 if( adr->Opcode() != Op_AddP ) Unimplemented();
2165 Node *base = adr->in(1);
2167 Node *zero = phase->makecon(TypeLong::ZERO);
2168 Node *off = phase->MakeConX(BytesPerLong);
2169 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2170 count--;
2171 while( count-- ) {
2172 mem = phase->transform(mem);
2173 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2174 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2175 }
2176 return mem;
2177 }
2179 //----------------------------clear_memory-------------------------------------
2180 // Generate code to initialize object storage to zero.
2181 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2182 intptr_t start_offset,
2183 Node* end_offset,
2184 PhaseGVN* phase) {
2185 Compile* C = phase->C;
2186 intptr_t offset = start_offset;
2188 int unit = BytesPerLong;
2189 if ((offset % unit) != 0) {
2190 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2191 adr = phase->transform(adr);
2192 const TypePtr* atp = TypeRawPtr::BOTTOM;
2193 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2194 mem = phase->transform(mem);
2195 offset += BytesPerInt;
2196 }
2197 assert((offset % unit) == 0, "");
2199 // Initialize the remaining stuff, if any, with a ClearArray.
2200 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2201 }
2203 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2204 Node* start_offset,
2205 Node* end_offset,
2206 PhaseGVN* phase) {
2207 if (start_offset == end_offset) {
2208 // nothing to do
2209 return mem;
2210 }
2212 Compile* C = phase->C;
2213 int unit = BytesPerLong;
2214 Node* zbase = start_offset;
2215 Node* zend = end_offset;
2217 // Scale to the unit required by the CPU:
2218 if (!Matcher::init_array_count_is_in_bytes) {
2219 Node* shift = phase->intcon(exact_log2(unit));
2220 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2221 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2222 }
2224 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2225 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2227 // Bulk clear double-words
2228 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2229 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2230 return phase->transform(mem);
2231 }
2233 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2234 intptr_t start_offset,
2235 intptr_t end_offset,
2236 PhaseGVN* phase) {
2237 if (start_offset == end_offset) {
2238 // nothing to do
2239 return mem;
2240 }
2242 Compile* C = phase->C;
2243 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2244 intptr_t done_offset = end_offset;
2245 if ((done_offset % BytesPerLong) != 0) {
2246 done_offset -= BytesPerInt;
2247 }
2248 if (done_offset > start_offset) {
2249 mem = clear_memory(ctl, mem, dest,
2250 start_offset, phase->MakeConX(done_offset), phase);
2251 }
2252 if (done_offset < end_offset) { // emit the final 32-bit store
2253 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2254 adr = phase->transform(adr);
2255 const TypePtr* atp = TypeRawPtr::BOTTOM;
2256 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2257 mem = phase->transform(mem);
2258 done_offset += BytesPerInt;
2259 }
2260 assert(done_offset == end_offset, "");
2261 return mem;
2262 }
2264 //=============================================================================
2265 // Do we match on this edge? No memory edges
2266 uint StrCompNode::match_edge(uint idx) const {
2267 return idx == 5 || idx == 6;
2268 }
2270 //------------------------------Ideal------------------------------------------
2271 // Return a node which is more "ideal" than the current node. Strip out
2272 // control copies
2273 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2274 return remove_dead_region(phase, can_reshape) ? this : NULL;
2275 }
2278 //=============================================================================
2279 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2280 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2281 _adr_type(C->get_adr_type(alias_idx))
2282 {
2283 init_class_id(Class_MemBar);
2284 Node* top = C->top();
2285 init_req(TypeFunc::I_O,top);
2286 init_req(TypeFunc::FramePtr,top);
2287 init_req(TypeFunc::ReturnAdr,top);
2288 if (precedent != NULL)
2289 init_req(TypeFunc::Parms, precedent);
2290 }
2292 //------------------------------cmp--------------------------------------------
2293 uint MemBarNode::hash() const { return NO_HASH; }
2294 uint MemBarNode::cmp( const Node &n ) const {
2295 return (&n == this); // Always fail except on self
2296 }
2298 //------------------------------make-------------------------------------------
2299 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2300 int len = Precedent + (pn == NULL? 0: 1);
2301 switch (opcode) {
2302 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2303 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2304 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2305 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2306 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2307 default: ShouldNotReachHere(); return NULL;
2308 }
2309 }
2311 //------------------------------Ideal------------------------------------------
2312 // Return a node which is more "ideal" than the current node. Strip out
2313 // control copies
2314 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2315 if (remove_dead_region(phase, can_reshape)) return this;
2316 return NULL;
2317 }
2319 //------------------------------Value------------------------------------------
2320 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2321 if( !in(0) ) return Type::TOP;
2322 if( phase->type(in(0)) == Type::TOP )
2323 return Type::TOP;
2324 return TypeTuple::MEMBAR;
2325 }
2327 //------------------------------match------------------------------------------
2328 // Construct projections for memory.
2329 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2330 switch (proj->_con) {
2331 case TypeFunc::Control:
2332 case TypeFunc::Memory:
2333 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2334 }
2335 ShouldNotReachHere();
2336 return NULL;
2337 }
2339 //===========================InitializeNode====================================
2340 // SUMMARY:
2341 // This node acts as a memory barrier on raw memory, after some raw stores.
2342 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2343 // The Initialize can 'capture' suitably constrained stores as raw inits.
2344 // It can coalesce related raw stores into larger units (called 'tiles').
2345 // It can avoid zeroing new storage for memory units which have raw inits.
2346 // At macro-expansion, it is marked 'complete', and does not optimize further.
2347 //
2348 // EXAMPLE:
2349 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2350 // ctl = incoming control; mem* = incoming memory
2351 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2352 // First allocate uninitialized memory and fill in the header:
2353 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2354 // ctl := alloc.Control; mem* := alloc.Memory*
2355 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2356 // Then initialize to zero the non-header parts of the raw memory block:
2357 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2358 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2359 // After the initialize node executes, the object is ready for service:
2360 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2361 // Suppose its body is immediately initialized as {1,2}:
2362 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2363 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2364 // mem.SLICE(#short[*]) := store2
2365 //
2366 // DETAILS:
2367 // An InitializeNode collects and isolates object initialization after
2368 // an AllocateNode and before the next possible safepoint. As a
2369 // memory barrier (MemBarNode), it keeps critical stores from drifting
2370 // down past any safepoint or any publication of the allocation.
2371 // Before this barrier, a newly-allocated object may have uninitialized bits.
2372 // After this barrier, it may be treated as a real oop, and GC is allowed.
2373 //
2374 // The semantics of the InitializeNode include an implicit zeroing of
2375 // the new object from object header to the end of the object.
2376 // (The object header and end are determined by the AllocateNode.)
2377 //
2378 // Certain stores may be added as direct inputs to the InitializeNode.
2379 // These stores must update raw memory, and they must be to addresses
2380 // derived from the raw address produced by AllocateNode, and with
2381 // a constant offset. They must be ordered by increasing offset.
2382 // The first one is at in(RawStores), the last at in(req()-1).
2383 // Unlike most memory operations, they are not linked in a chain,
2384 // but are displayed in parallel as users of the rawmem output of
2385 // the allocation.
2386 //
2387 // (See comments in InitializeNode::capture_store, which continue
2388 // the example given above.)
2389 //
2390 // When the associated Allocate is macro-expanded, the InitializeNode
2391 // may be rewritten to optimize collected stores. A ClearArrayNode
2392 // may also be created at that point to represent any required zeroing.
2393 // The InitializeNode is then marked 'complete', prohibiting further
2394 // capturing of nearby memory operations.
2395 //
2396 // During macro-expansion, all captured initializations which store
2397 // constant values of 32 bits or smaller are coalesced (if advantagous)
2398 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2399 // initialized in fewer memory operations. Memory words which are
2400 // covered by neither tiles nor non-constant stores are pre-zeroed
2401 // by explicit stores of zero. (The code shape happens to do all
2402 // zeroing first, then all other stores, with both sequences occurring
2403 // in order of ascending offsets.)
2404 //
2405 // Alternatively, code may be inserted between an AllocateNode and its
2406 // InitializeNode, to perform arbitrary initialization of the new object.
2407 // E.g., the object copying intrinsics insert complex data transfers here.
2408 // The initialization must then be marked as 'complete' disable the
2409 // built-in zeroing semantics and the collection of initializing stores.
2410 //
2411 // While an InitializeNode is incomplete, reads from the memory state
2412 // produced by it are optimizable if they match the control edge and
2413 // new oop address associated with the allocation/initialization.
2414 // They return a stored value (if the offset matches) or else zero.
2415 // A write to the memory state, if it matches control and address,
2416 // and if it is to a constant offset, may be 'captured' by the
2417 // InitializeNode. It is cloned as a raw memory operation and rewired
2418 // inside the initialization, to the raw oop produced by the allocation.
2419 // Operations on addresses which are provably distinct (e.g., to
2420 // other AllocateNodes) are allowed to bypass the initialization.
2421 //
2422 // The effect of all this is to consolidate object initialization
2423 // (both arrays and non-arrays, both piecewise and bulk) into a
2424 // single location, where it can be optimized as a unit.
2425 //
2426 // Only stores with an offset less than TrackedInitializationLimit words
2427 // will be considered for capture by an InitializeNode. This puts a
2428 // reasonable limit on the complexity of optimized initializations.
2430 //---------------------------InitializeNode------------------------------------
2431 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2432 : _is_complete(false),
2433 MemBarNode(C, adr_type, rawoop)
2434 {
2435 init_class_id(Class_Initialize);
2437 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2438 assert(in(RawAddress) == rawoop, "proper init");
2439 // Note: allocation() can be NULL, for secondary initialization barriers
2440 }
2442 // Since this node is not matched, it will be processed by the
2443 // register allocator. Declare that there are no constraints
2444 // on the allocation of the RawAddress edge.
2445 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2446 // This edge should be set to top, by the set_complete. But be conservative.
2447 if (idx == InitializeNode::RawAddress)
2448 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2449 return RegMask::Empty;
2450 }
2452 Node* InitializeNode::memory(uint alias_idx) {
2453 Node* mem = in(Memory);
2454 if (mem->is_MergeMem()) {
2455 return mem->as_MergeMem()->memory_at(alias_idx);
2456 } else {
2457 // incoming raw memory is not split
2458 return mem;
2459 }
2460 }
2462 bool InitializeNode::is_non_zero() {
2463 if (is_complete()) return false;
2464 remove_extra_zeroes();
2465 return (req() > RawStores);
2466 }
2468 void InitializeNode::set_complete(PhaseGVN* phase) {
2469 assert(!is_complete(), "caller responsibility");
2470 _is_complete = true;
2472 // After this node is complete, it contains a bunch of
2473 // raw-memory initializations. There is no need for
2474 // it to have anything to do with non-raw memory effects.
2475 // Therefore, tell all non-raw users to re-optimize themselves,
2476 // after skipping the memory effects of this initialization.
2477 PhaseIterGVN* igvn = phase->is_IterGVN();
2478 if (igvn) igvn->add_users_to_worklist(this);
2479 }
2481 // convenience function
2482 // return false if the init contains any stores already
2483 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2484 InitializeNode* init = initialization();
2485 if (init == NULL || init->is_complete()) return false;
2486 init->remove_extra_zeroes();
2487 // for now, if this allocation has already collected any inits, bail:
2488 if (init->is_non_zero()) return false;
2489 init->set_complete(phase);
2490 return true;
2491 }
2493 void InitializeNode::remove_extra_zeroes() {
2494 if (req() == RawStores) return;
2495 Node* zmem = zero_memory();
2496 uint fill = RawStores;
2497 for (uint i = fill; i < req(); i++) {
2498 Node* n = in(i);
2499 if (n->is_top() || n == zmem) continue; // skip
2500 if (fill < i) set_req(fill, n); // compact
2501 ++fill;
2502 }
2503 // delete any empty spaces created:
2504 while (fill < req()) {
2505 del_req(fill);
2506 }
2507 }
2509 // Helper for remembering which stores go with which offsets.
2510 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2511 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2512 intptr_t offset = -1;
2513 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2514 phase, offset);
2515 if (base == NULL) return -1; // something is dead,
2516 if (offset < 0) return -1; // dead, dead
2517 return offset;
2518 }
2520 // Helper for proving that an initialization expression is
2521 // "simple enough" to be folded into an object initialization.
2522 // Attempts to prove that a store's initial value 'n' can be captured
2523 // within the initialization without creating a vicious cycle, such as:
2524 // { Foo p = new Foo(); p.next = p; }
2525 // True for constants and parameters and small combinations thereof.
2526 bool InitializeNode::detect_init_independence(Node* n,
2527 bool st_is_pinned,
2528 int& count) {
2529 if (n == NULL) return true; // (can this really happen?)
2530 if (n->is_Proj()) n = n->in(0);
2531 if (n == this) return false; // found a cycle
2532 if (n->is_Con()) return true;
2533 if (n->is_Start()) return true; // params, etc., are OK
2534 if (n->is_Root()) return true; // even better
2536 Node* ctl = n->in(0);
2537 if (ctl != NULL && !ctl->is_top()) {
2538 if (ctl->is_Proj()) ctl = ctl->in(0);
2539 if (ctl == this) return false;
2541 // If we already know that the enclosing memory op is pinned right after
2542 // the init, then any control flow that the store has picked up
2543 // must have preceded the init, or else be equal to the init.
2544 // Even after loop optimizations (which might change control edges)
2545 // a store is never pinned *before* the availability of its inputs.
2546 if (!MemNode::all_controls_dominate(n, this))
2547 return false; // failed to prove a good control
2549 }
2551 // Check data edges for possible dependencies on 'this'.
2552 if ((count += 1) > 20) return false; // complexity limit
2553 for (uint i = 1; i < n->req(); i++) {
2554 Node* m = n->in(i);
2555 if (m == NULL || m == n || m->is_top()) continue;
2556 uint first_i = n->find_edge(m);
2557 if (i != first_i) continue; // process duplicate edge just once
2558 if (!detect_init_independence(m, st_is_pinned, count)) {
2559 return false;
2560 }
2561 }
2563 return true;
2564 }
2566 // Here are all the checks a Store must pass before it can be moved into
2567 // an initialization. Returns zero if a check fails.
2568 // On success, returns the (constant) offset to which the store applies,
2569 // within the initialized memory.
2570 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2571 const int FAIL = 0;
2572 if (st->req() != MemNode::ValueIn + 1)
2573 return FAIL; // an inscrutable StoreNode (card mark?)
2574 Node* ctl = st->in(MemNode::Control);
2575 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2576 return FAIL; // must be unconditional after the initialization
2577 Node* mem = st->in(MemNode::Memory);
2578 if (!(mem->is_Proj() && mem->in(0) == this))
2579 return FAIL; // must not be preceded by other stores
2580 Node* adr = st->in(MemNode::Address);
2581 intptr_t offset;
2582 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2583 if (alloc == NULL)
2584 return FAIL; // inscrutable address
2585 if (alloc != allocation())
2586 return FAIL; // wrong allocation! (store needs to float up)
2587 Node* val = st->in(MemNode::ValueIn);
2588 int complexity_count = 0;
2589 if (!detect_init_independence(val, true, complexity_count))
2590 return FAIL; // stored value must be 'simple enough'
2592 return offset; // success
2593 }
2595 // Find the captured store in(i) which corresponds to the range
2596 // [start..start+size) in the initialized object.
2597 // If there is one, return its index i. If there isn't, return the
2598 // negative of the index where it should be inserted.
2599 // Return 0 if the queried range overlaps an initialization boundary
2600 // or if dead code is encountered.
2601 // If size_in_bytes is zero, do not bother with overlap checks.
2602 int InitializeNode::captured_store_insertion_point(intptr_t start,
2603 int size_in_bytes,
2604 PhaseTransform* phase) {
2605 const int FAIL = 0, MAX_STORE = BytesPerLong;
2607 if (is_complete())
2608 return FAIL; // arraycopy got here first; punt
2610 assert(allocation() != NULL, "must be present");
2612 // no negatives, no header fields:
2613 if (start < (intptr_t) sizeof(oopDesc)) return FAIL;
2614 if (start < (intptr_t) sizeof(arrayOopDesc) &&
2615 start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2617 // after a certain size, we bail out on tracking all the stores:
2618 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2619 if (start >= ti_limit) return FAIL;
2621 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2622 if (i >= limit) return -(int)i; // not found; here is where to put it
2624 Node* st = in(i);
2625 intptr_t st_off = get_store_offset(st, phase);
2626 if (st_off < 0) {
2627 if (st != zero_memory()) {
2628 return FAIL; // bail out if there is dead garbage
2629 }
2630 } else if (st_off > start) {
2631 // ...we are done, since stores are ordered
2632 if (st_off < start + size_in_bytes) {
2633 return FAIL; // the next store overlaps
2634 }
2635 return -(int)i; // not found; here is where to put it
2636 } else if (st_off < start) {
2637 if (size_in_bytes != 0 &&
2638 start < st_off + MAX_STORE &&
2639 start < st_off + st->as_Store()->memory_size()) {
2640 return FAIL; // the previous store overlaps
2641 }
2642 } else {
2643 if (size_in_bytes != 0 &&
2644 st->as_Store()->memory_size() != size_in_bytes) {
2645 return FAIL; // mismatched store size
2646 }
2647 return i;
2648 }
2650 ++i;
2651 }
2652 }
2654 // Look for a captured store which initializes at the offset 'start'
2655 // with the given size. If there is no such store, and no other
2656 // initialization interferes, then return zero_memory (the memory
2657 // projection of the AllocateNode).
2658 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2659 PhaseTransform* phase) {
2660 assert(stores_are_sane(phase), "");
2661 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2662 if (i == 0) {
2663 return NULL; // something is dead
2664 } else if (i < 0) {
2665 return zero_memory(); // just primordial zero bits here
2666 } else {
2667 Node* st = in(i); // here is the store at this position
2668 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2669 return st;
2670 }
2671 }
2673 // Create, as a raw pointer, an address within my new object at 'offset'.
2674 Node* InitializeNode::make_raw_address(intptr_t offset,
2675 PhaseTransform* phase) {
2676 Node* addr = in(RawAddress);
2677 if (offset != 0) {
2678 Compile* C = phase->C;
2679 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2680 phase->MakeConX(offset)) );
2681 }
2682 return addr;
2683 }
2685 // Clone the given store, converting it into a raw store
2686 // initializing a field or element of my new object.
2687 // Caller is responsible for retiring the original store,
2688 // with subsume_node or the like.
2689 //
2690 // From the example above InitializeNode::InitializeNode,
2691 // here are the old stores to be captured:
2692 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2693 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2694 //
2695 // Here is the changed code; note the extra edges on init:
2696 // alloc = (Allocate ...)
2697 // rawoop = alloc.RawAddress
2698 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2699 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2700 // init = (Initialize alloc.Control alloc.Memory rawoop
2701 // rawstore1 rawstore2)
2702 //
2703 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2704 PhaseTransform* phase) {
2705 assert(stores_are_sane(phase), "");
2707 if (start < 0) return NULL;
2708 assert(can_capture_store(st, phase) == start, "sanity");
2710 Compile* C = phase->C;
2711 int size_in_bytes = st->memory_size();
2712 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2713 if (i == 0) return NULL; // bail out
2714 Node* prev_mem = NULL; // raw memory for the captured store
2715 if (i > 0) {
2716 prev_mem = in(i); // there is a pre-existing store under this one
2717 set_req(i, C->top()); // temporarily disconnect it
2718 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2719 } else {
2720 i = -i; // no pre-existing store
2721 prev_mem = zero_memory(); // a slice of the newly allocated object
2722 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2723 set_req(--i, C->top()); // reuse this edge; it has been folded away
2724 else
2725 ins_req(i, C->top()); // build a new edge
2726 }
2727 Node* new_st = st->clone();
2728 new_st->set_req(MemNode::Control, in(Control));
2729 new_st->set_req(MemNode::Memory, prev_mem);
2730 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2731 new_st = phase->transform(new_st);
2733 // At this point, new_st might have swallowed a pre-existing store
2734 // at the same offset, or perhaps new_st might have disappeared,
2735 // if it redundantly stored the same value (or zero to fresh memory).
2737 // In any case, wire it in:
2738 set_req(i, new_st);
2740 // The caller may now kill the old guy.
2741 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2742 assert(check_st == new_st || check_st == NULL, "must be findable");
2743 assert(!is_complete(), "");
2744 return new_st;
2745 }
2747 static bool store_constant(jlong* tiles, int num_tiles,
2748 intptr_t st_off, int st_size,
2749 jlong con) {
2750 if ((st_off & (st_size-1)) != 0)
2751 return false; // strange store offset (assume size==2**N)
2752 address addr = (address)tiles + st_off;
2753 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2754 switch (st_size) {
2755 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2756 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2757 case sizeof(jint): *(jint*) addr = (jint) con; break;
2758 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2759 default: return false; // strange store size (detect size!=2**N here)
2760 }
2761 return true; // return success to caller
2762 }
2764 // Coalesce subword constants into int constants and possibly
2765 // into long constants. The goal, if the CPU permits,
2766 // is to initialize the object with a small number of 64-bit tiles.
2767 // Also, convert floating-point constants to bit patterns.
2768 // Non-constants are not relevant to this pass.
2769 //
2770 // In terms of the running example on InitializeNode::InitializeNode
2771 // and InitializeNode::capture_store, here is the transformation
2772 // of rawstore1 and rawstore2 into rawstore12:
2773 // alloc = (Allocate ...)
2774 // rawoop = alloc.RawAddress
2775 // tile12 = 0x00010002
2776 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2777 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2778 //
2779 void
2780 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2781 Node* size_in_bytes,
2782 PhaseGVN* phase) {
2783 Compile* C = phase->C;
2785 assert(stores_are_sane(phase), "");
2786 // Note: After this pass, they are not completely sane,
2787 // since there may be some overlaps.
2789 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2791 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2792 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2793 size_limit = MIN2(size_limit, ti_limit);
2794 size_limit = align_size_up(size_limit, BytesPerLong);
2795 int num_tiles = size_limit / BytesPerLong;
2797 // allocate space for the tile map:
2798 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2799 jlong tiles_buf[small_len];
2800 Node* nodes_buf[small_len];
2801 jlong inits_buf[small_len];
2802 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2803 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2804 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2805 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2806 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2807 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2808 // tiles: exact bitwise model of all primitive constants
2809 // nodes: last constant-storing node subsumed into the tiles model
2810 // inits: which bytes (in each tile) are touched by any initializations
2812 //// Pass A: Fill in the tile model with any relevant stores.
2814 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2815 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2816 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2817 Node* zmem = zero_memory(); // initially zero memory state
2818 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2819 Node* st = in(i);
2820 intptr_t st_off = get_store_offset(st, phase);
2822 // Figure out the store's offset and constant value:
2823 if (st_off < header_size) continue; //skip (ignore header)
2824 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
2825 int st_size = st->as_Store()->memory_size();
2826 if (st_off + st_size > size_limit) break;
2828 // Record which bytes are touched, whether by constant or not.
2829 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
2830 continue; // skip (strange store size)
2832 const Type* val = phase->type(st->in(MemNode::ValueIn));
2833 if (!val->singleton()) continue; //skip (non-con store)
2834 BasicType type = val->basic_type();
2836 jlong con = 0;
2837 switch (type) {
2838 case T_INT: con = val->is_int()->get_con(); break;
2839 case T_LONG: con = val->is_long()->get_con(); break;
2840 case T_FLOAT: con = jint_cast(val->getf()); break;
2841 case T_DOUBLE: con = jlong_cast(val->getd()); break;
2842 default: continue; //skip (odd store type)
2843 }
2845 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
2846 st->Opcode() == Op_StoreL) {
2847 continue; // This StoreL is already optimal.
2848 }
2850 // Store down the constant.
2851 store_constant(tiles, num_tiles, st_off, st_size, con);
2853 intptr_t j = st_off >> LogBytesPerLong;
2855 if (type == T_INT && st_size == BytesPerInt
2856 && (st_off & BytesPerInt) == BytesPerInt) {
2857 jlong lcon = tiles[j];
2858 if (!Matcher::isSimpleConstant64(lcon) &&
2859 st->Opcode() == Op_StoreI) {
2860 // This StoreI is already optimal by itself.
2861 jint* intcon = (jint*) &tiles[j];
2862 intcon[1] = 0; // undo the store_constant()
2864 // If the previous store is also optimal by itself, back up and
2865 // undo the action of the previous loop iteration... if we can.
2866 // But if we can't, just let the previous half take care of itself.
2867 st = nodes[j];
2868 st_off -= BytesPerInt;
2869 con = intcon[0];
2870 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
2871 assert(st_off >= header_size, "still ignoring header");
2872 assert(get_store_offset(st, phase) == st_off, "must be");
2873 assert(in(i-1) == zmem, "must be");
2874 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
2875 assert(con == tcon->is_int()->get_con(), "must be");
2876 // Undo the effects of the previous loop trip, which swallowed st:
2877 intcon[0] = 0; // undo store_constant()
2878 set_req(i-1, st); // undo set_req(i, zmem)
2879 nodes[j] = NULL; // undo nodes[j] = st
2880 --old_subword; // undo ++old_subword
2881 }
2882 continue; // This StoreI is already optimal.
2883 }
2884 }
2886 // This store is not needed.
2887 set_req(i, zmem);
2888 nodes[j] = st; // record for the moment
2889 if (st_size < BytesPerLong) // something has changed
2890 ++old_subword; // includes int/float, but who's counting...
2891 else ++old_long;
2892 }
2894 if ((old_subword + old_long) == 0)
2895 return; // nothing more to do
2897 //// Pass B: Convert any non-zero tiles into optimal constant stores.
2898 // Be sure to insert them before overlapping non-constant stores.
2899 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
2900 for (int j = 0; j < num_tiles; j++) {
2901 jlong con = tiles[j];
2902 jlong init = inits[j];
2903 if (con == 0) continue;
2904 jint con0, con1; // split the constant, address-wise
2905 jint init0, init1; // split the init map, address-wise
2906 { union { jlong con; jint intcon[2]; } u;
2907 u.con = con;
2908 con0 = u.intcon[0];
2909 con1 = u.intcon[1];
2910 u.con = init;
2911 init0 = u.intcon[0];
2912 init1 = u.intcon[1];
2913 }
2915 Node* old = nodes[j];
2916 assert(old != NULL, "need the prior store");
2917 intptr_t offset = (j * BytesPerLong);
2919 bool split = !Matcher::isSimpleConstant64(con);
2921 if (offset < header_size) {
2922 assert(offset + BytesPerInt >= header_size, "second int counts");
2923 assert(*(jint*)&tiles[j] == 0, "junk in header");
2924 split = true; // only the second word counts
2925 // Example: int a[] = { 42 ... }
2926 } else if (con0 == 0 && init0 == -1) {
2927 split = true; // first word is covered by full inits
2928 // Example: int a[] = { ... foo(), 42 ... }
2929 } else if (con1 == 0 && init1 == -1) {
2930 split = true; // second word is covered by full inits
2931 // Example: int a[] = { ... 42, foo() ... }
2932 }
2934 // Here's a case where init0 is neither 0 nor -1:
2935 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
2936 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
2937 // In this case the tile is not split; it is (jlong)42.
2938 // The big tile is stored down, and then the foo() value is inserted.
2939 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
2941 Node* ctl = old->in(MemNode::Control);
2942 Node* adr = make_raw_address(offset, phase);
2943 const TypePtr* atp = TypeRawPtr::BOTTOM;
2945 // One or two coalesced stores to plop down.
2946 Node* st[2];
2947 intptr_t off[2];
2948 int nst = 0;
2949 if (!split) {
2950 ++new_long;
2951 off[nst] = offset;
2952 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2953 phase->longcon(con), T_LONG);
2954 } else {
2955 // Omit either if it is a zero.
2956 if (con0 != 0) {
2957 ++new_int;
2958 off[nst] = offset;
2959 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2960 phase->intcon(con0), T_INT);
2961 }
2962 if (con1 != 0) {
2963 ++new_int;
2964 offset += BytesPerInt;
2965 adr = make_raw_address(offset, phase);
2966 off[nst] = offset;
2967 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2968 phase->intcon(con1), T_INT);
2969 }
2970 }
2972 // Insert second store first, then the first before the second.
2973 // Insert each one just before any overlapping non-constant stores.
2974 while (nst > 0) {
2975 Node* st1 = st[--nst];
2976 C->copy_node_notes_to(st1, old);
2977 st1 = phase->transform(st1);
2978 offset = off[nst];
2979 assert(offset >= header_size, "do not smash header");
2980 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
2981 guarantee(ins_idx != 0, "must re-insert constant store");
2982 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
2983 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
2984 set_req(--ins_idx, st1);
2985 else
2986 ins_req(ins_idx, st1);
2987 }
2988 }
2990 if (PrintCompilation && WizardMode)
2991 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
2992 old_subword, old_long, new_int, new_long);
2993 if (C->log() != NULL)
2994 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
2995 old_subword, old_long, new_int, new_long);
2997 // Clean up any remaining occurrences of zmem:
2998 remove_extra_zeroes();
2999 }
3001 // Explore forward from in(start) to find the first fully initialized
3002 // word, and return its offset. Skip groups of subword stores which
3003 // together initialize full words. If in(start) is itself part of a
3004 // fully initialized word, return the offset of in(start). If there
3005 // are no following full-word stores, or if something is fishy, return
3006 // a negative value.
3007 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3008 int int_map = 0;
3009 intptr_t int_map_off = 0;
3010 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3012 for (uint i = start, limit = req(); i < limit; i++) {
3013 Node* st = in(i);
3015 intptr_t st_off = get_store_offset(st, phase);
3016 if (st_off < 0) break; // return conservative answer
3018 int st_size = st->as_Store()->memory_size();
3019 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3020 return st_off; // we found a complete word init
3021 }
3023 // update the map:
3025 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3026 if (this_int_off != int_map_off) {
3027 // reset the map:
3028 int_map = 0;
3029 int_map_off = this_int_off;
3030 }
3032 int subword_off = st_off - this_int_off;
3033 int_map |= right_n_bits(st_size) << subword_off;
3034 if ((int_map & FULL_MAP) == FULL_MAP) {
3035 return this_int_off; // we found a complete word init
3036 }
3038 // Did this store hit or cross the word boundary?
3039 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3040 if (next_int_off == this_int_off + BytesPerInt) {
3041 // We passed the current int, without fully initializing it.
3042 int_map_off = next_int_off;
3043 int_map >>= BytesPerInt;
3044 } else if (next_int_off > this_int_off + BytesPerInt) {
3045 // We passed the current and next int.
3046 return this_int_off + BytesPerInt;
3047 }
3048 }
3050 return -1;
3051 }
3054 // Called when the associated AllocateNode is expanded into CFG.
3055 // At this point, we may perform additional optimizations.
3056 // Linearize the stores by ascending offset, to make memory
3057 // activity as coherent as possible.
3058 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3059 intptr_t header_size,
3060 Node* size_in_bytes,
3061 PhaseGVN* phase) {
3062 assert(!is_complete(), "not already complete");
3063 assert(stores_are_sane(phase), "");
3064 assert(allocation() != NULL, "must be present");
3066 remove_extra_zeroes();
3068 if (ReduceFieldZeroing || ReduceBulkZeroing)
3069 // reduce instruction count for common initialization patterns
3070 coalesce_subword_stores(header_size, size_in_bytes, phase);
3072 Node* zmem = zero_memory(); // initially zero memory state
3073 Node* inits = zmem; // accumulating a linearized chain of inits
3074 #ifdef ASSERT
3075 intptr_t last_init_off = sizeof(oopDesc); // previous init offset
3076 intptr_t last_init_end = sizeof(oopDesc); // previous init offset+size
3077 intptr_t last_tile_end = sizeof(oopDesc); // previous tile offset+size
3078 #endif
3079 intptr_t zeroes_done = header_size;
3081 bool do_zeroing = true; // we might give up if inits are very sparse
3082 int big_init_gaps = 0; // how many large gaps have we seen?
3084 if (ZeroTLAB) do_zeroing = false;
3085 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3087 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3088 Node* st = in(i);
3089 intptr_t st_off = get_store_offset(st, phase);
3090 if (st_off < 0)
3091 break; // unknown junk in the inits
3092 if (st->in(MemNode::Memory) != zmem)
3093 break; // complicated store chains somehow in list
3095 int st_size = st->as_Store()->memory_size();
3096 intptr_t next_init_off = st_off + st_size;
3098 if (do_zeroing && zeroes_done < next_init_off) {
3099 // See if this store needs a zero before it or under it.
3100 intptr_t zeroes_needed = st_off;
3102 if (st_size < BytesPerInt) {
3103 // Look for subword stores which only partially initialize words.
3104 // If we find some, we must lay down some word-level zeroes first,
3105 // underneath the subword stores.
3106 //
3107 // Examples:
3108 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3109 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3110 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3111 //
3112 // Note: coalesce_subword_stores may have already done this,
3113 // if it was prompted by constant non-zero subword initializers.
3114 // But this case can still arise with non-constant stores.
3116 intptr_t next_full_store = find_next_fullword_store(i, phase);
3118 // In the examples above:
3119 // in(i) p q r s x y z
3120 // st_off 12 13 14 15 12 13 14
3121 // st_size 1 1 1 1 1 1 1
3122 // next_full_s. 12 16 16 16 16 16 16
3123 // z's_done 12 16 16 16 12 16 12
3124 // z's_needed 12 16 16 16 16 16 16
3125 // zsize 0 0 0 0 4 0 4
3126 if (next_full_store < 0) {
3127 // Conservative tack: Zero to end of current word.
3128 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3129 } else {
3130 // Zero to beginning of next fully initialized word.
3131 // Or, don't zero at all, if we are already in that word.
3132 assert(next_full_store >= zeroes_needed, "must go forward");
3133 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3134 zeroes_needed = next_full_store;
3135 }
3136 }
3138 if (zeroes_needed > zeroes_done) {
3139 intptr_t zsize = zeroes_needed - zeroes_done;
3140 // Do some incremental zeroing on rawmem, in parallel with inits.
3141 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3142 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3143 zeroes_done, zeroes_needed,
3144 phase);
3145 zeroes_done = zeroes_needed;
3146 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3147 do_zeroing = false; // leave the hole, next time
3148 }
3149 }
3151 // Collect the store and move on:
3152 st->set_req(MemNode::Memory, inits);
3153 inits = st; // put it on the linearized chain
3154 set_req(i, zmem); // unhook from previous position
3156 if (zeroes_done == st_off)
3157 zeroes_done = next_init_off;
3159 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3161 #ifdef ASSERT
3162 // Various order invariants. Weaker than stores_are_sane because
3163 // a large constant tile can be filled in by smaller non-constant stores.
3164 assert(st_off >= last_init_off, "inits do not reverse");
3165 last_init_off = st_off;
3166 const Type* val = NULL;
3167 if (st_size >= BytesPerInt &&
3168 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3169 (int)val->basic_type() < (int)T_OBJECT) {
3170 assert(st_off >= last_tile_end, "tiles do not overlap");
3171 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3172 last_tile_end = MAX2(last_tile_end, next_init_off);
3173 } else {
3174 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3175 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3176 assert(st_off >= last_init_end, "inits do not overlap");
3177 last_init_end = next_init_off; // it's a non-tile
3178 }
3179 #endif //ASSERT
3180 }
3182 remove_extra_zeroes(); // clear out all the zmems left over
3183 add_req(inits);
3185 if (!ZeroTLAB) {
3186 // If anything remains to be zeroed, zero it all now.
3187 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3188 // if it is the last unused 4 bytes of an instance, forget about it
3189 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3190 if (zeroes_done + BytesPerLong >= size_limit) {
3191 assert(allocation() != NULL, "");
3192 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3193 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3194 if (zeroes_done == k->layout_helper())
3195 zeroes_done = size_limit;
3196 }
3197 if (zeroes_done < size_limit) {
3198 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3199 zeroes_done, size_in_bytes, phase);
3200 }
3201 }
3203 set_complete(phase);
3204 return rawmem;
3205 }
3208 #ifdef ASSERT
3209 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3210 if (is_complete())
3211 return true; // stores could be anything at this point
3212 intptr_t last_off = sizeof(oopDesc);
3213 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3214 Node* st = in(i);
3215 intptr_t st_off = get_store_offset(st, phase);
3216 if (st_off < 0) continue; // ignore dead garbage
3217 if (last_off > st_off) {
3218 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3219 this->dump(2);
3220 assert(false, "ascending store offsets");
3221 return false;
3222 }
3223 last_off = st_off + st->as_Store()->memory_size();
3224 }
3225 return true;
3226 }
3227 #endif //ASSERT
3232 //============================MergeMemNode=====================================
3233 //
3234 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3235 // contributing store or call operations. Each contributor provides the memory
3236 // state for a particular "alias type" (see Compile::alias_type). For example,
3237 // if a MergeMem has an input X for alias category #6, then any memory reference
3238 // to alias category #6 may use X as its memory state input, as an exact equivalent
3239 // to using the MergeMem as a whole.
3240 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3241 //
3242 // (Here, the <N> notation gives the index of the relevant adr_type.)
3243 //
3244 // In one special case (and more cases in the future), alias categories overlap.
3245 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3246 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3247 // it is exactly equivalent to that state W:
3248 // MergeMem(<Bot>: W) <==> W
3249 //
3250 // Usually, the merge has more than one input. In that case, where inputs
3251 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3252 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3253 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3254 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3255 //
3256 // A merge can take a "wide" memory state as one of its narrow inputs.
3257 // This simply means that the merge observes out only the relevant parts of
3258 // the wide input. That is, wide memory states arriving at narrow merge inputs
3259 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3260 //
3261 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3262 // and that memory slices "leak through":
3263 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3264 //
3265 // But, in such a cascade, repeated memory slices can "block the leak":
3266 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3267 //
3268 // In the last example, Y is not part of the combined memory state of the
3269 // outermost MergeMem. The system must, of course, prevent unschedulable
3270 // memory states from arising, so you can be sure that the state Y is somehow
3271 // a precursor to state Y'.
3272 //
3273 //
3274 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3275 // of each MergeMemNode array are exactly the numerical alias indexes, including
3276 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3277 // Compile::alias_type (and kin) produce and manage these indexes.
3278 //
3279 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3280 // (Note that this provides quick access to the top node inside MergeMem methods,
3281 // without the need to reach out via TLS to Compile::current.)
3282 //
3283 // As a consequence of what was just described, a MergeMem that represents a full
3284 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3285 // containing all alias categories.
3286 //
3287 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3288 //
3289 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3290 // a memory state for the alias type <N>, or else the top node, meaning that
3291 // there is no particular input for that alias type. Note that the length of
3292 // a MergeMem is variable, and may be extended at any time to accommodate new
3293 // memory states at larger alias indexes. When merges grow, they are of course
3294 // filled with "top" in the unused in() positions.
3295 //
3296 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3297 // (Top was chosen because it works smoothly with passes like GCM.)
3298 //
3299 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3300 // the type of random VM bits like TLS references.) Since it is always the
3301 // first non-Bot memory slice, some low-level loops use it to initialize an
3302 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3303 //
3304 //
3305 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3306 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3307 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3308 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3309 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3310 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3311 //
3312 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3313 // really that different from the other memory inputs. An abbreviation called
3314 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3315 //
3316 //
3317 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3318 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3319 // that "emerges though" the base memory will be marked as excluding the alias types
3320 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3321 //
3322 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3323 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3324 //
3325 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3326 // (It is currently unimplemented.) As you can see, the resulting merge is
3327 // actually a disjoint union of memory states, rather than an overlay.
3328 //
3330 //------------------------------MergeMemNode-----------------------------------
3331 Node* MergeMemNode::make_empty_memory() {
3332 Node* empty_memory = (Node*) Compile::current()->top();
3333 assert(empty_memory->is_top(), "correct sentinel identity");
3334 return empty_memory;
3335 }
3337 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3338 init_class_id(Class_MergeMem);
3339 // all inputs are nullified in Node::Node(int)
3340 // set_input(0, NULL); // no control input
3342 // Initialize the edges uniformly to top, for starters.
3343 Node* empty_mem = make_empty_memory();
3344 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3345 init_req(i,empty_mem);
3346 }
3347 assert(empty_memory() == empty_mem, "");
3349 if( new_base != NULL && new_base->is_MergeMem() ) {
3350 MergeMemNode* mdef = new_base->as_MergeMem();
3351 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3352 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3353 mms.set_memory(mms.memory2());
3354 }
3355 assert(base_memory() == mdef->base_memory(), "");
3356 } else {
3357 set_base_memory(new_base);
3358 }
3359 }
3361 // Make a new, untransformed MergeMem with the same base as 'mem'.
3362 // If mem is itself a MergeMem, populate the result with the same edges.
3363 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3364 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3365 }
3367 //------------------------------cmp--------------------------------------------
3368 uint MergeMemNode::hash() const { return NO_HASH; }
3369 uint MergeMemNode::cmp( const Node &n ) const {
3370 return (&n == this); // Always fail except on self
3371 }
3373 //------------------------------Identity---------------------------------------
3374 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3375 // Identity if this merge point does not record any interesting memory
3376 // disambiguations.
3377 Node* base_mem = base_memory();
3378 Node* empty_mem = empty_memory();
3379 if (base_mem != empty_mem) { // Memory path is not dead?
3380 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3381 Node* mem = in(i);
3382 if (mem != empty_mem && mem != base_mem) {
3383 return this; // Many memory splits; no change
3384 }
3385 }
3386 }
3387 return base_mem; // No memory splits; ID on the one true input
3388 }
3390 //------------------------------Ideal------------------------------------------
3391 // This method is invoked recursively on chains of MergeMem nodes
3392 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3393 // Remove chain'd MergeMems
3394 //
3395 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3396 // relative to the "in(Bot)". Since we are patching both at the same time,
3397 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3398 // but rewrite each "in(i)" relative to the new "in(Bot)".
3399 Node *progress = NULL;
3402 Node* old_base = base_memory();
3403 Node* empty_mem = empty_memory();
3404 if (old_base == empty_mem)
3405 return NULL; // Dead memory path.
3407 MergeMemNode* old_mbase;
3408 if (old_base != NULL && old_base->is_MergeMem())
3409 old_mbase = old_base->as_MergeMem();
3410 else
3411 old_mbase = NULL;
3412 Node* new_base = old_base;
3414 // simplify stacked MergeMems in base memory
3415 if (old_mbase) new_base = old_mbase->base_memory();
3417 // the base memory might contribute new slices beyond my req()
3418 if (old_mbase) grow_to_match(old_mbase);
3420 // Look carefully at the base node if it is a phi.
3421 PhiNode* phi_base;
3422 if (new_base != NULL && new_base->is_Phi())
3423 phi_base = new_base->as_Phi();
3424 else
3425 phi_base = NULL;
3427 Node* phi_reg = NULL;
3428 uint phi_len = (uint)-1;
3429 if (phi_base != NULL && !phi_base->is_copy()) {
3430 // do not examine phi if degraded to a copy
3431 phi_reg = phi_base->region();
3432 phi_len = phi_base->req();
3433 // see if the phi is unfinished
3434 for (uint i = 1; i < phi_len; i++) {
3435 if (phi_base->in(i) == NULL) {
3436 // incomplete phi; do not look at it yet!
3437 phi_reg = NULL;
3438 phi_len = (uint)-1;
3439 break;
3440 }
3441 }
3442 }
3444 // Note: We do not call verify_sparse on entry, because inputs
3445 // can normalize to the base_memory via subsume_node or similar
3446 // mechanisms. This method repairs that damage.
3448 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3450 // Look at each slice.
3451 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3452 Node* old_in = in(i);
3453 // calculate the old memory value
3454 Node* old_mem = old_in;
3455 if (old_mem == empty_mem) old_mem = old_base;
3456 assert(old_mem == memory_at(i), "");
3458 // maybe update (reslice) the old memory value
3460 // simplify stacked MergeMems
3461 Node* new_mem = old_mem;
3462 MergeMemNode* old_mmem;
3463 if (old_mem != NULL && old_mem->is_MergeMem())
3464 old_mmem = old_mem->as_MergeMem();
3465 else
3466 old_mmem = NULL;
3467 if (old_mmem == this) {
3468 // This can happen if loops break up and safepoints disappear.
3469 // A merge of BotPtr (default) with a RawPtr memory derived from a
3470 // safepoint can be rewritten to a merge of the same BotPtr with
3471 // the BotPtr phi coming into the loop. If that phi disappears
3472 // also, we can end up with a self-loop of the mergemem.
3473 // In general, if loops degenerate and memory effects disappear,
3474 // a mergemem can be left looking at itself. This simply means
3475 // that the mergemem's default should be used, since there is
3476 // no longer any apparent effect on this slice.
3477 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3478 // from start. Update the input to TOP.
3479 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3480 }
3481 else if (old_mmem != NULL) {
3482 new_mem = old_mmem->memory_at(i);
3483 }
3484 // else preceeding memory was not a MergeMem
3486 // replace equivalent phis (unfortunately, they do not GVN together)
3487 if (new_mem != NULL && new_mem != new_base &&
3488 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3489 if (new_mem->is_Phi()) {
3490 PhiNode* phi_mem = new_mem->as_Phi();
3491 for (uint i = 1; i < phi_len; i++) {
3492 if (phi_base->in(i) != phi_mem->in(i)) {
3493 phi_mem = NULL;
3494 break;
3495 }
3496 }
3497 if (phi_mem != NULL) {
3498 // equivalent phi nodes; revert to the def
3499 new_mem = new_base;
3500 }
3501 }
3502 }
3504 // maybe store down a new value
3505 Node* new_in = new_mem;
3506 if (new_in == new_base) new_in = empty_mem;
3508 if (new_in != old_in) {
3509 // Warning: Do not combine this "if" with the previous "if"
3510 // A memory slice might have be be rewritten even if it is semantically
3511 // unchanged, if the base_memory value has changed.
3512 set_req(i, new_in);
3513 progress = this; // Report progress
3514 }
3515 }
3517 if (new_base != old_base) {
3518 set_req(Compile::AliasIdxBot, new_base);
3519 // Don't use set_base_memory(new_base), because we need to update du.
3520 assert(base_memory() == new_base, "");
3521 progress = this;
3522 }
3524 if( base_memory() == this ) {
3525 // a self cycle indicates this memory path is dead
3526 set_req(Compile::AliasIdxBot, empty_mem);
3527 }
3529 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3530 // Recursion must occur after the self cycle check above
3531 if( base_memory()->is_MergeMem() ) {
3532 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3533 Node *m = phase->transform(new_mbase); // Rollup any cycles
3534 if( m != NULL && (m->is_top() ||
3535 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3536 // propagate rollup of dead cycle to self
3537 set_req(Compile::AliasIdxBot, empty_mem);
3538 }
3539 }
3541 if( base_memory() == empty_mem ) {
3542 progress = this;
3543 // Cut inputs during Parse phase only.
3544 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3545 if( !can_reshape ) {
3546 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3547 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3548 }
3549 }
3550 }
3552 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3553 // Check if PhiNode::Ideal's "Split phis through memory merges"
3554 // transform should be attempted. Look for this->phi->this cycle.
3555 uint merge_width = req();
3556 if (merge_width > Compile::AliasIdxRaw) {
3557 PhiNode* phi = base_memory()->as_Phi();
3558 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3559 if (phi->in(i) == this) {
3560 phase->is_IterGVN()->_worklist.push(phi);
3561 break;
3562 }
3563 }
3564 }
3565 }
3567 assert(progress || verify_sparse(), "please, no dups of base");
3568 return progress;
3569 }
3571 //-------------------------set_base_memory-------------------------------------
3572 void MergeMemNode::set_base_memory(Node *new_base) {
3573 Node* empty_mem = empty_memory();
3574 set_req(Compile::AliasIdxBot, new_base);
3575 assert(memory_at(req()) == new_base, "must set default memory");
3576 // Clear out other occurrences of new_base:
3577 if (new_base != empty_mem) {
3578 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3579 if (in(i) == new_base) set_req(i, empty_mem);
3580 }
3581 }
3582 }
3584 //------------------------------out_RegMask------------------------------------
3585 const RegMask &MergeMemNode::out_RegMask() const {
3586 return RegMask::Empty;
3587 }
3589 //------------------------------dump_spec--------------------------------------
3590 #ifndef PRODUCT
3591 void MergeMemNode::dump_spec(outputStream *st) const {
3592 st->print(" {");
3593 Node* base_mem = base_memory();
3594 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3595 Node* mem = memory_at(i);
3596 if (mem == base_mem) { st->print(" -"); continue; }
3597 st->print( " N%d:", mem->_idx );
3598 Compile::current()->get_adr_type(i)->dump_on(st);
3599 }
3600 st->print(" }");
3601 }
3602 #endif // !PRODUCT
3605 #ifdef ASSERT
3606 static bool might_be_same(Node* a, Node* b) {
3607 if (a == b) return true;
3608 if (!(a->is_Phi() || b->is_Phi())) return false;
3609 // phis shift around during optimization
3610 return true; // pretty stupid...
3611 }
3613 // verify a narrow slice (either incoming or outgoing)
3614 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3615 if (!VerifyAliases) return; // don't bother to verify unless requested
3616 if (is_error_reported()) return; // muzzle asserts when debugging an error
3617 if (Node::in_dump()) return; // muzzle asserts when printing
3618 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3619 assert(n != NULL, "");
3620 // Elide intervening MergeMem's
3621 while (n->is_MergeMem()) {
3622 n = n->as_MergeMem()->memory_at(alias_idx);
3623 }
3624 Compile* C = Compile::current();
3625 const TypePtr* n_adr_type = n->adr_type();
3626 if (n == m->empty_memory()) {
3627 // Implicit copy of base_memory()
3628 } else if (n_adr_type != TypePtr::BOTTOM) {
3629 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3630 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3631 } else {
3632 // A few places like make_runtime_call "know" that VM calls are narrow,
3633 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3634 bool expected_wide_mem = false;
3635 if (n == m->base_memory()) {
3636 expected_wide_mem = true;
3637 } else if (alias_idx == Compile::AliasIdxRaw ||
3638 n == m->memory_at(Compile::AliasIdxRaw)) {
3639 expected_wide_mem = true;
3640 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3641 // memory can "leak through" calls on channels that
3642 // are write-once. Allow this also.
3643 expected_wide_mem = true;
3644 }
3645 assert(expected_wide_mem, "expected narrow slice replacement");
3646 }
3647 }
3648 #else // !ASSERT
3649 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3650 #endif
3653 //-----------------------------memory_at---------------------------------------
3654 Node* MergeMemNode::memory_at(uint alias_idx) const {
3655 assert(alias_idx >= Compile::AliasIdxRaw ||
3656 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3657 "must avoid base_memory and AliasIdxTop");
3659 // Otherwise, it is a narrow slice.
3660 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3661 Compile *C = Compile::current();
3662 if (is_empty_memory(n)) {
3663 // the array is sparse; empty slots are the "top" node
3664 n = base_memory();
3665 assert(Node::in_dump()
3666 || n == NULL || n->bottom_type() == Type::TOP
3667 || n->adr_type() == TypePtr::BOTTOM
3668 || n->adr_type() == TypeRawPtr::BOTTOM
3669 || Compile::current()->AliasLevel() == 0,
3670 "must be a wide memory");
3671 // AliasLevel == 0 if we are organizing the memory states manually.
3672 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3673 } else {
3674 // make sure the stored slice is sane
3675 #ifdef ASSERT
3676 if (is_error_reported() || Node::in_dump()) {
3677 } else if (might_be_same(n, base_memory())) {
3678 // Give it a pass: It is a mostly harmless repetition of the base.
3679 // This can arise normally from node subsumption during optimization.
3680 } else {
3681 verify_memory_slice(this, alias_idx, n);
3682 }
3683 #endif
3684 }
3685 return n;
3686 }
3688 //---------------------------set_memory_at-------------------------------------
3689 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3690 verify_memory_slice(this, alias_idx, n);
3691 Node* empty_mem = empty_memory();
3692 if (n == base_memory()) n = empty_mem; // collapse default
3693 uint need_req = alias_idx+1;
3694 if (req() < need_req) {
3695 if (n == empty_mem) return; // already the default, so do not grow me
3696 // grow the sparse array
3697 do {
3698 add_req(empty_mem);
3699 } while (req() < need_req);
3700 }
3701 set_req( alias_idx, n );
3702 }
3706 //--------------------------iteration_setup------------------------------------
3707 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3708 if (other != NULL) {
3709 grow_to_match(other);
3710 // invariant: the finite support of mm2 is within mm->req()
3711 #ifdef ASSERT
3712 for (uint i = req(); i < other->req(); i++) {
3713 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3714 }
3715 #endif
3716 }
3717 // Replace spurious copies of base_memory by top.
3718 Node* base_mem = base_memory();
3719 if (base_mem != NULL && !base_mem->is_top()) {
3720 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3721 if (in(i) == base_mem)
3722 set_req(i, empty_memory());
3723 }
3724 }
3725 }
3727 //---------------------------grow_to_match-------------------------------------
3728 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3729 Node* empty_mem = empty_memory();
3730 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3731 // look for the finite support of the other memory
3732 for (uint i = other->req(); --i >= req(); ) {
3733 if (other->in(i) != empty_mem) {
3734 uint new_len = i+1;
3735 while (req() < new_len) add_req(empty_mem);
3736 break;
3737 }
3738 }
3739 }
3741 //---------------------------verify_sparse-------------------------------------
3742 #ifndef PRODUCT
3743 bool MergeMemNode::verify_sparse() const {
3744 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3745 Node* base_mem = base_memory();
3746 // The following can happen in degenerate cases, since empty==top.
3747 if (is_empty_memory(base_mem)) return true;
3748 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3749 assert(in(i) != NULL, "sane slice");
3750 if (in(i) == base_mem) return false; // should have been the sentinel value!
3751 }
3752 return true;
3753 }
3755 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3756 Node* n;
3757 n = mm->in(idx);
3758 if (mem == n) return true; // might be empty_memory()
3759 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3760 if (mem == n) return true;
3761 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3762 if (mem == n) return true;
3763 if (n == NULL) break;
3764 }
3765 return false;
3766 }
3767 #endif // !PRODUCT