Thu, 20 Mar 2008 15:11:44 -0700
6674600: (Escape Analysis) Optimize memory graph for instance's fields
Summary: EA gives opportunite to do more aggressive memory optimizations.
Reviewed-by: never, jrose
1 /*
2 * Copyright 1997-2007 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 // Portions of code courtesy of Clifford Click
27 // Optimization - Graph Style
29 #include "incls/_precompiled.incl"
30 #include "incls/_memnode.cpp.incl"
32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
34 //=============================================================================
35 uint MemNode::size_of() const { return sizeof(*this); }
37 const TypePtr *MemNode::adr_type() const {
38 Node* adr = in(Address);
39 const TypePtr* cross_check = NULL;
40 DEBUG_ONLY(cross_check = _adr_type);
41 return calculate_adr_type(adr->bottom_type(), cross_check);
42 }
44 #ifndef PRODUCT
45 void MemNode::dump_spec(outputStream *st) const {
46 if (in(Address) == NULL) return; // node is dead
47 #ifndef ASSERT
48 // fake the missing field
49 const TypePtr* _adr_type = NULL;
50 if (in(Address) != NULL)
51 _adr_type = in(Address)->bottom_type()->isa_ptr();
52 #endif
53 dump_adr_type(this, _adr_type, st);
55 Compile* C = Compile::current();
56 if( C->alias_type(_adr_type)->is_volatile() )
57 st->print(" Volatile!");
58 }
60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
61 st->print(" @");
62 if (adr_type == NULL) {
63 st->print("NULL");
64 } else {
65 adr_type->dump_on(st);
66 Compile* C = Compile::current();
67 Compile::AliasType* atp = NULL;
68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
69 if (atp == NULL)
70 st->print(", idx=?\?;");
71 else if (atp->index() == Compile::AliasIdxBot)
72 st->print(", idx=Bot;");
73 else if (atp->index() == Compile::AliasIdxTop)
74 st->print(", idx=Top;");
75 else if (atp->index() == Compile::AliasIdxRaw)
76 st->print(", idx=Raw;");
77 else {
78 ciField* field = atp->field();
79 if (field) {
80 st->print(", name=");
81 field->print_name_on(st);
82 }
83 st->print(", idx=%d;", atp->index());
84 }
85 }
86 }
88 extern void print_alias_types();
90 #endif
92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
93 const TypeOopPtr *tinst = t_adr->isa_oopptr();
94 if (tinst == NULL || !tinst->is_instance_field())
95 return mchain; // don't try to optimize non-instance types
96 uint instance_id = tinst->instance_id();
97 Node *prev = NULL;
98 Node *result = mchain;
99 while (prev != result) {
100 prev = result;
101 // skip over a call which does not affect this memory slice
102 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
103 Node *proj_in = result->in(0);
104 if (proj_in->is_Call()) {
105 CallNode *call = proj_in->as_Call();
106 if (!call->may_modify(t_adr, phase)) {
107 result = call->in(TypeFunc::Memory);
108 }
109 } else if (proj_in->is_Initialize()) {
110 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
111 // Stop if this is the initialization for the object instance which
112 // which contains this memory slice, otherwise skip over it.
113 if (alloc != NULL && alloc->_idx != instance_id) {
114 result = proj_in->in(TypeFunc::Memory);
115 }
116 } else if (proj_in->is_MemBar()) {
117 result = proj_in->in(TypeFunc::Memory);
118 }
119 } else if (result->is_MergeMem()) {
120 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
121 }
122 }
123 return result;
124 }
126 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
127 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
128 bool is_instance = (t_oop != NULL) && t_oop->is_instance_field();
129 PhaseIterGVN *igvn = phase->is_IterGVN();
130 Node *result = mchain;
131 result = optimize_simple_memory_chain(result, t_adr, phase);
132 if (is_instance && igvn != NULL && result->is_Phi()) {
133 PhiNode *mphi = result->as_Phi();
134 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
135 const TypePtr *t = mphi->adr_type();
136 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM) {
137 // 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 control input 'dom' dominates (or equals) '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::detect_dominating_control(Node* dom, Node* sub) {
251 if (dom == NULL) return false;
252 if (dom->is_Proj()) dom = dom->in(0);
253 if (dom->is_Start()) return true; // anything inside the method
254 if (dom->is_Root()) return true; // dom 'controls' a constant
255 int cnt = 20; // detect cycle or too much effort
256 while (sub != NULL) { // walk 'sub' up the chain to 'dom'
257 if (--cnt < 0) return false; // in a cycle or too complex
258 if (sub == dom) return true;
259 if (sub->is_Start()) return false;
260 if (sub->is_Root()) return false;
261 Node* up = sub->in(0);
262 if (sub == up && sub->is_Region()) {
263 for (uint i = 1; i < sub->req(); i++) {
264 Node* in = sub->in(i);
265 if (in != NULL && !in->is_top() && in != sub) {
266 up = in; break; // take any path on the way up to 'dom'
267 }
268 }
269 }
270 if (sub == up) return false; // some kind of tight cycle
271 sub = up;
272 }
273 return false;
274 }
276 //---------------------detect_ptr_independence---------------------------------
277 // Used by MemNode::find_previous_store to prove that two base
278 // pointers are never equal.
279 // The pointers are accompanied by their associated allocations,
280 // if any, which have been previously discovered by the caller.
281 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
282 Node* p2, AllocateNode* a2,
283 PhaseTransform* phase) {
284 // Attempt to prove that these two pointers cannot be aliased.
285 // They may both manifestly be allocations, and they should differ.
286 // Or, if they are not both allocations, they can be distinct constants.
287 // Otherwise, one is an allocation and the other a pre-existing value.
288 if (a1 == NULL && a2 == NULL) { // neither an allocation
289 return (p1 != p2) && p1->is_Con() && p2->is_Con();
290 } else if (a1 != NULL && a2 != NULL) { // both allocations
291 return (a1 != a2);
292 } else if (a1 != NULL) { // one allocation a1
293 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
294 return detect_dominating_control(p2->in(0), a1->in(0));
295 } else { //(a2 != NULL) // one allocation a2
296 return detect_dominating_control(p1->in(0), a2->in(0));
297 }
298 return false;
299 }
302 // The logic for reordering loads and stores uses four steps:
303 // (a) Walk carefully past stores and initializations which we
304 // can prove are independent of this load.
305 // (b) Observe that the next memory state makes an exact match
306 // with self (load or store), and locate the relevant store.
307 // (c) Ensure that, if we were to wire self directly to the store,
308 // the optimizer would fold it up somehow.
309 // (d) Do the rewiring, and return, depending on some other part of
310 // the optimizer to fold up the load.
311 // This routine handles steps (a) and (b). Steps (c) and (d) are
312 // specific to loads and stores, so they are handled by the callers.
313 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
314 //
315 Node* MemNode::find_previous_store(PhaseTransform* phase) {
316 Node* ctrl = in(MemNode::Control);
317 Node* adr = in(MemNode::Address);
318 intptr_t offset = 0;
319 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
320 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
322 if (offset == Type::OffsetBot)
323 return NULL; // cannot unalias unless there are precise offsets
325 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
327 intptr_t size_in_bytes = memory_size();
329 Node* mem = in(MemNode::Memory); // start searching here...
331 int cnt = 50; // Cycle limiter
332 for (;;) { // While we can dance past unrelated stores...
333 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
335 if (mem->is_Store()) {
336 Node* st_adr = mem->in(MemNode::Address);
337 intptr_t st_offset = 0;
338 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
339 if (st_base == NULL)
340 break; // inscrutable pointer
341 if (st_offset != offset && st_offset != Type::OffsetBot) {
342 const int MAX_STORE = BytesPerLong;
343 if (st_offset >= offset + size_in_bytes ||
344 st_offset <= offset - MAX_STORE ||
345 st_offset <= offset - mem->as_Store()->memory_size()) {
346 // Success: The offsets are provably independent.
347 // (You may ask, why not just test st_offset != offset and be done?
348 // The answer is that stores of different sizes can co-exist
349 // in the same sequence of RawMem effects. We sometimes initialize
350 // a whole 'tile' of array elements with a single jint or jlong.)
351 mem = mem->in(MemNode::Memory);
352 continue; // (a) advance through independent store memory
353 }
354 }
355 if (st_base != base &&
356 detect_ptr_independence(base, alloc,
357 st_base,
358 AllocateNode::Ideal_allocation(st_base, phase),
359 phase)) {
360 // Success: The bases are provably independent.
361 mem = mem->in(MemNode::Memory);
362 continue; // (a) advance through independent store memory
363 }
365 // (b) At this point, if the bases or offsets do not agree, we lose,
366 // since we have not managed to prove 'this' and 'mem' independent.
367 if (st_base == base && st_offset == offset) {
368 return mem; // let caller handle steps (c), (d)
369 }
371 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
372 InitializeNode* st_init = mem->in(0)->as_Initialize();
373 AllocateNode* st_alloc = st_init->allocation();
374 if (st_alloc == NULL)
375 break; // something degenerated
376 bool known_identical = false;
377 bool known_independent = false;
378 if (alloc == st_alloc)
379 known_identical = true;
380 else if (alloc != NULL)
381 known_independent = true;
382 else if (ctrl != NULL &&
383 detect_dominating_control(ctrl, st_alloc->in(0)))
384 known_independent = true;
386 if (known_independent) {
387 // The bases are provably independent: Either they are
388 // manifestly distinct allocations, or else the control
389 // of this load dominates the store's allocation.
390 int alias_idx = phase->C->get_alias_index(adr_type());
391 if (alias_idx == Compile::AliasIdxRaw) {
392 mem = st_alloc->in(TypeFunc::Memory);
393 } else {
394 mem = st_init->memory(alias_idx);
395 }
396 continue; // (a) advance through independent store memory
397 }
399 // (b) at this point, if we are not looking at a store initializing
400 // the same allocation we are loading from, we lose.
401 if (known_identical) {
402 // From caller, can_see_stored_value will consult find_captured_store.
403 return mem; // let caller handle steps (c), (d)
404 }
406 } else if (addr_t != NULL && addr_t->is_instance_field()) {
407 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
408 if (mem->is_Proj() && mem->in(0)->is_Call()) {
409 CallNode *call = mem->in(0)->as_Call();
410 if (!call->may_modify(addr_t, phase)) {
411 mem = call->in(TypeFunc::Memory);
412 continue; // (a) advance through independent call memory
413 }
414 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
415 mem = mem->in(0)->in(TypeFunc::Memory);
416 continue; // (a) advance through independent MemBar memory
417 } else if (mem->is_MergeMem()) {
418 int alias_idx = phase->C->get_alias_index(adr_type());
419 mem = mem->as_MergeMem()->memory_at(alias_idx);
420 continue; // (a) advance through independent MergeMem memory
421 }
422 }
424 // Unless there is an explicit 'continue', we must bail out here,
425 // because 'mem' is an inscrutable memory state (e.g., a call).
426 break;
427 }
429 return NULL; // bail out
430 }
432 //----------------------calculate_adr_type-------------------------------------
433 // Helper function. Notices when the given type of address hits top or bottom.
434 // Also, asserts a cross-check of the type against the expected address type.
435 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
436 if (t == Type::TOP) return NULL; // does not touch memory any more?
437 #ifdef PRODUCT
438 cross_check = NULL;
439 #else
440 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
441 #endif
442 const TypePtr* tp = t->isa_ptr();
443 if (tp == NULL) {
444 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
445 return TypePtr::BOTTOM; // touches lots of memory
446 } else {
447 #ifdef ASSERT
448 // %%%% [phh] We don't check the alias index if cross_check is
449 // TypeRawPtr::BOTTOM. Needs to be investigated.
450 if (cross_check != NULL &&
451 cross_check != TypePtr::BOTTOM &&
452 cross_check != TypeRawPtr::BOTTOM) {
453 // Recheck the alias index, to see if it has changed (due to a bug).
454 Compile* C = Compile::current();
455 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
456 "must stay in the original alias category");
457 // The type of the address must be contained in the adr_type,
458 // disregarding "null"-ness.
459 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
460 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
461 assert(cross_check->meet(tp_notnull) == cross_check,
462 "real address must not escape from expected memory type");
463 }
464 #endif
465 return tp;
466 }
467 }
469 //------------------------adr_phi_is_loop_invariant----------------------------
470 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
471 // loop is loop invariant. Make a quick traversal of Phi and associated
472 // CastPP nodes, looking to see if they are a closed group within the loop.
473 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
474 // The idea is that the phi-nest must boil down to only CastPP nodes
475 // with the same data. This implies that any path into the loop already
476 // includes such a CastPP, and so the original cast, whatever its input,
477 // must be covered by an equivalent cast, with an earlier control input.
478 ResourceMark rm;
480 // The loop entry input of the phi should be the unique dominating
481 // node for every Phi/CastPP in the loop.
482 Unique_Node_List closure;
483 closure.push(adr_phi->in(LoopNode::EntryControl));
485 // Add the phi node and the cast to the worklist.
486 Unique_Node_List worklist;
487 worklist.push(adr_phi);
488 if( cast != NULL ){
489 if( !cast->is_ConstraintCast() ) return false;
490 worklist.push(cast);
491 }
493 // Begin recursive walk of phi nodes.
494 while( worklist.size() ){
495 // Take a node off the worklist
496 Node *n = worklist.pop();
497 if( !closure.member(n) ){
498 // Add it to the closure.
499 closure.push(n);
500 // Make a sanity check to ensure we don't waste too much time here.
501 if( closure.size() > 20) return false;
502 // This node is OK if:
503 // - it is a cast of an identical value
504 // - or it is a phi node (then we add its inputs to the worklist)
505 // Otherwise, the node is not OK, and we presume the cast is not invariant
506 if( n->is_ConstraintCast() ){
507 worklist.push(n->in(1));
508 } else if( n->is_Phi() ) {
509 for( uint i = 1; i < n->req(); i++ ) {
510 worklist.push(n->in(i));
511 }
512 } else {
513 return false;
514 }
515 }
516 }
518 // Quit when the worklist is empty, and we've found no offending nodes.
519 return true;
520 }
522 //------------------------------Ideal_DU_postCCP-------------------------------
523 // Find any cast-away of null-ness and keep its control. Null cast-aways are
524 // going away in this pass and we need to make this memory op depend on the
525 // gating null check.
527 // I tried to leave the CastPP's in. This makes the graph more accurate in
528 // some sense; we get to keep around the knowledge that an oop is not-null
529 // after some test. Alas, the CastPP's interfere with GVN (some values are
530 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
531 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
532 // some of the more trivial cases in the optimizer. Removing more useless
533 // Phi's started allowing Loads to illegally float above null checks. I gave
534 // up on this approach. CNC 10/20/2000
535 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
536 Node *ctr = in(MemNode::Control);
537 Node *mem = in(MemNode::Memory);
538 Node *adr = in(MemNode::Address);
539 Node *skipped_cast = NULL;
540 // Need a null check? Regular static accesses do not because they are
541 // from constant addresses. Array ops are gated by the range check (which
542 // always includes a NULL check). Just check field ops.
543 if( !ctr ) {
544 // Scan upwards for the highest location we can place this memory op.
545 while( true ) {
546 switch( adr->Opcode() ) {
548 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
549 adr = adr->in(AddPNode::Base);
550 continue;
552 case Op_CastPP:
553 // If the CastPP is useless, just peek on through it.
554 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
555 // Remember the cast that we've peeked though. If we peek
556 // through more than one, then we end up remembering the highest
557 // one, that is, if in a loop, the one closest to the top.
558 skipped_cast = adr;
559 adr = adr->in(1);
560 continue;
561 }
562 // CastPP is going away in this pass! We need this memory op to be
563 // control-dependent on the test that is guarding the CastPP.
564 ccp->hash_delete(this);
565 set_req(MemNode::Control, adr->in(0));
566 ccp->hash_insert(this);
567 return this;
569 case Op_Phi:
570 // Attempt to float above a Phi to some dominating point.
571 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
572 // If we've already peeked through a Cast (which could have set the
573 // control), we can't float above a Phi, because the skipped Cast
574 // may not be loop invariant.
575 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
576 adr = adr->in(1);
577 continue;
578 }
579 }
581 // Intentional fallthrough!
583 // No obvious dominating point. The mem op is pinned below the Phi
584 // by the Phi itself. If the Phi goes away (no true value is merged)
585 // then the mem op can float, but not indefinitely. It must be pinned
586 // behind the controls leading to the Phi.
587 case Op_CheckCastPP:
588 // These usually stick around to change address type, however a
589 // useless one can be elided and we still need to pick up a control edge
590 if (adr->in(0) == NULL) {
591 // This CheckCastPP node has NO control and is likely useless. But we
592 // need check further up the ancestor chain for a control input to keep
593 // the node in place. 4959717.
594 skipped_cast = adr;
595 adr = adr->in(1);
596 continue;
597 }
598 ccp->hash_delete(this);
599 set_req(MemNode::Control, adr->in(0));
600 ccp->hash_insert(this);
601 return this;
603 // List of "safe" opcodes; those that implicitly block the memory
604 // op below any null check.
605 case Op_CastX2P: // no null checks on native pointers
606 case Op_Parm: // 'this' pointer is not null
607 case Op_LoadP: // Loading from within a klass
608 case Op_LoadKlass: // Loading from within a klass
609 case Op_ConP: // Loading from a klass
610 case Op_CreateEx: // Sucking up the guts of an exception oop
611 case Op_Con: // Reading from TLS
612 case Op_CMoveP: // CMoveP is pinned
613 break; // No progress
615 case Op_Proj: // Direct call to an allocation routine
616 case Op_SCMemProj: // Memory state from store conditional ops
617 #ifdef ASSERT
618 {
619 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
620 const Node* call = adr->in(0);
621 if (call->is_CallStaticJava()) {
622 const CallStaticJavaNode* call_java = call->as_CallStaticJava();
623 const TypeTuple *r = call_java->tf()->range();
624 assert(r->cnt() > TypeFunc::Parms, "must return value");
625 const Type* ret_type = r->field_at(TypeFunc::Parms);
626 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
627 // We further presume that this is one of
628 // new_instance_Java, new_array_Java, or
629 // the like, but do not assert for this.
630 } else if (call->is_Allocate()) {
631 // similar case to new_instance_Java, etc.
632 } else if (!call->is_CallLeaf()) {
633 // Projections from fetch_oop (OSR) are allowed as well.
634 ShouldNotReachHere();
635 }
636 }
637 #endif
638 break;
639 default:
640 ShouldNotReachHere();
641 }
642 break;
643 }
644 }
646 return NULL; // No progress
647 }
650 //=============================================================================
651 uint LoadNode::size_of() const { return sizeof(*this); }
652 uint LoadNode::cmp( const Node &n ) const
653 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
654 const Type *LoadNode::bottom_type() const { return _type; }
655 uint LoadNode::ideal_reg() const {
656 return Matcher::base2reg[_type->base()];
657 }
659 #ifndef PRODUCT
660 void LoadNode::dump_spec(outputStream *st) const {
661 MemNode::dump_spec(st);
662 if( !Verbose && !WizardMode ) {
663 // standard dump does this in Verbose and WizardMode
664 st->print(" #"); _type->dump_on(st);
665 }
666 }
667 #endif
670 //----------------------------LoadNode::make-----------------------------------
671 // Polymorphic factory method:
672 LoadNode *LoadNode::make( Compile *C, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
673 // sanity check the alias category against the created node type
674 assert(!(adr_type->isa_oopptr() &&
675 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
676 "use LoadKlassNode instead");
677 assert(!(adr_type->isa_aryptr() &&
678 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
679 "use LoadRangeNode instead");
680 switch (bt) {
681 case T_BOOLEAN:
682 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
683 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
684 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
685 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
686 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
687 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
688 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
689 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
690 case T_OBJECT: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
691 }
692 ShouldNotReachHere();
693 return (LoadNode*)NULL;
694 }
696 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
697 bool require_atomic = true;
698 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
699 }
704 //------------------------------hash-------------------------------------------
705 uint LoadNode::hash() const {
706 // unroll addition of interesting fields
707 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
708 }
710 //---------------------------can_see_stored_value------------------------------
711 // This routine exists to make sure this set of tests is done the same
712 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
713 // will change the graph shape in a way which makes memory alive twice at the
714 // same time (uses the Oracle model of aliasing), then some
715 // LoadXNode::Identity will fold things back to the equivalence-class model
716 // of aliasing.
717 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
718 Node* ld_adr = in(MemNode::Address);
720 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
721 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
722 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
723 atp->field() != NULL && !atp->field()->is_volatile()) {
724 uint alias_idx = atp->index();
725 bool final = atp->field()->is_final();
726 Node* result = NULL;
727 Node* current = st;
728 // Skip through chains of MemBarNodes checking the MergeMems for
729 // new states for the slice of this load. Stop once any other
730 // kind of node is encountered. Loads from final memory can skip
731 // through any kind of MemBar but normal loads shouldn't skip
732 // through MemBarAcquire since the could allow them to move out of
733 // a synchronized region.
734 while (current->is_Proj()) {
735 int opc = current->in(0)->Opcode();
736 if ((final && opc == Op_MemBarAcquire) ||
737 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
738 Node* mem = current->in(0)->in(TypeFunc::Memory);
739 if (mem->is_MergeMem()) {
740 MergeMemNode* merge = mem->as_MergeMem();
741 Node* new_st = merge->memory_at(alias_idx);
742 if (new_st == merge->base_memory()) {
743 // Keep searching
744 current = merge->base_memory();
745 continue;
746 }
747 // Save the new memory state for the slice and fall through
748 // to exit.
749 result = new_st;
750 }
751 }
752 break;
753 }
754 if (result != NULL) {
755 st = result;
756 }
757 }
760 // Loop around twice in the case Load -> Initialize -> Store.
761 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
762 for (int trip = 0; trip <= 1; trip++) {
764 if (st->is_Store()) {
765 Node* st_adr = st->in(MemNode::Address);
766 if (!phase->eqv(st_adr, ld_adr)) {
767 // Try harder before giving up... Match raw and non-raw pointers.
768 intptr_t st_off = 0;
769 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
770 if (alloc == NULL) return NULL;
771 intptr_t ld_off = 0;
772 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
773 if (alloc != allo2) return NULL;
774 if (ld_off != st_off) return NULL;
775 // At this point we have proven something like this setup:
776 // A = Allocate(...)
777 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
778 // S = StoreQ(, AddP(, A.Parm , #Off), V)
779 // (Actually, we haven't yet proven the Q's are the same.)
780 // In other words, we are loading from a casted version of
781 // the same pointer-and-offset that we stored to.
782 // Thus, we are able to replace L by V.
783 }
784 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
785 if (store_Opcode() != st->Opcode())
786 return NULL;
787 return st->in(MemNode::ValueIn);
788 }
790 intptr_t offset = 0; // scratch
792 // A load from a freshly-created object always returns zero.
793 // (This can happen after LoadNode::Ideal resets the load's memory input
794 // to find_captured_store, which returned InitializeNode::zero_memory.)
795 if (st->is_Proj() && st->in(0)->is_Allocate() &&
796 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
797 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
798 // return a zero value for the load's basic type
799 // (This is one of the few places where a generic PhaseTransform
800 // can create new nodes. Think of it as lazily manifesting
801 // virtually pre-existing constants.)
802 return phase->zerocon(memory_type());
803 }
805 // A load from an initialization barrier can match a captured store.
806 if (st->is_Proj() && st->in(0)->is_Initialize()) {
807 InitializeNode* init = st->in(0)->as_Initialize();
808 AllocateNode* alloc = init->allocation();
809 if (alloc != NULL &&
810 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
811 // examine a captured store value
812 st = init->find_captured_store(offset, memory_size(), phase);
813 if (st != NULL)
814 continue; // take one more trip around
815 }
816 }
818 break;
819 }
821 return NULL;
822 }
824 //----------------------is_instance_field_load_with_local_phi------------------
825 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
826 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
827 in(MemNode::Address)->is_AddP() ) {
828 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
829 // Only instances.
830 if( t_oop != NULL && t_oop->is_instance_field() &&
831 t_oop->offset() != Type::OffsetBot &&
832 t_oop->offset() != Type::OffsetTop) {
833 return true;
834 }
835 }
836 return false;
837 }
839 //------------------------------Identity---------------------------------------
840 // Loads are identity if previous store is to same address
841 Node *LoadNode::Identity( PhaseTransform *phase ) {
842 // If the previous store-maker is the right kind of Store, and the store is
843 // to the same address, then we are equal to the value stored.
844 Node* mem = in(MemNode::Memory);
845 Node* value = can_see_stored_value(mem, phase);
846 if( value ) {
847 // byte, short & char stores truncate naturally.
848 // A load has to load the truncated value which requires
849 // some sort of masking operation and that requires an
850 // Ideal call instead of an Identity call.
851 if (memory_size() < BytesPerInt) {
852 // If the input to the store does not fit with the load's result type,
853 // it must be truncated via an Ideal call.
854 if (!phase->type(value)->higher_equal(phase->type(this)))
855 return this;
856 }
857 // (This works even when value is a Con, but LoadNode::Value
858 // usually runs first, producing the singleton type of the Con.)
859 return value;
860 }
862 // Search for an existing data phi which was generated before for the same
863 // instance's field to avoid infinite genertion of phis in a loop.
864 Node *region = mem->in(0);
865 if (is_instance_field_load_with_local_phi(region)) {
866 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
867 int this_index = phase->C->get_alias_index(addr_t);
868 int this_offset = addr_t->offset();
869 int this_id = addr_t->is_oopptr()->instance_id();
870 const Type* this_type = bottom_type();
871 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
872 Node* phi = region->fast_out(i);
873 if (phi->is_Phi() && phi != mem &&
874 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
875 return phi;
876 }
877 }
878 }
880 return this;
881 }
884 // Returns true if the AliasType refers to the field that holds the
885 // cached box array. Currently only handles the IntegerCache case.
886 static bool is_autobox_cache(Compile::AliasType* atp) {
887 if (atp != NULL && atp->field() != NULL) {
888 ciField* field = atp->field();
889 ciSymbol* klass = field->holder()->name();
890 if (field->name() == ciSymbol::cache_field_name() &&
891 field->holder()->uses_default_loader() &&
892 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
893 return true;
894 }
895 }
896 return false;
897 }
899 // Fetch the base value in the autobox array
900 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
901 if (atp != NULL && atp->field() != NULL) {
902 ciField* field = atp->field();
903 ciSymbol* klass = field->holder()->name();
904 if (field->name() == ciSymbol::cache_field_name() &&
905 field->holder()->uses_default_loader() &&
906 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
907 assert(field->is_constant(), "what?");
908 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
909 // Fetch the box object at the base of the array and get its value
910 ciInstance* box = array->obj_at(0)->as_instance();
911 ciInstanceKlass* ik = box->klass()->as_instance_klass();
912 if (ik->nof_nonstatic_fields() == 1) {
913 // This should be true nonstatic_field_at requires calling
914 // nof_nonstatic_fields so check it anyway
915 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
916 cache_offset = c.as_int();
917 }
918 return true;
919 }
920 }
921 return false;
922 }
924 // Returns true if the AliasType refers to the value field of an
925 // autobox object. Currently only handles Integer.
926 static bool is_autobox_object(Compile::AliasType* atp) {
927 if (atp != NULL && atp->field() != NULL) {
928 ciField* field = atp->field();
929 ciSymbol* klass = field->holder()->name();
930 if (field->name() == ciSymbol::value_name() &&
931 field->holder()->uses_default_loader() &&
932 klass == ciSymbol::java_lang_Integer()) {
933 return true;
934 }
935 }
936 return false;
937 }
940 // We're loading from an object which has autobox behaviour.
941 // If this object is result of a valueOf call we'll have a phi
942 // merging a newly allocated object and a load from the cache.
943 // We want to replace this load with the original incoming
944 // argument to the valueOf call.
945 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
946 Node* base = in(Address)->in(AddPNode::Base);
947 if (base->is_Phi() && base->req() == 3) {
948 AllocateNode* allocation = NULL;
949 int allocation_index = -1;
950 int load_index = -1;
951 for (uint i = 1; i < base->req(); i++) {
952 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
953 if (allocation != NULL) {
954 allocation_index = i;
955 load_index = 3 - allocation_index;
956 break;
957 }
958 }
959 LoadNode* load = NULL;
960 if (allocation != NULL && base->in(load_index)->is_Load()) {
961 load = base->in(load_index)->as_Load();
962 }
963 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
964 // Push the loads from the phi that comes from valueOf up
965 // through it to allow elimination of the loads and the recovery
966 // of the original value.
967 Node* mem_phi = in(Memory);
968 Node* offset = in(Address)->in(AddPNode::Offset);
970 Node* in1 = clone();
971 Node* in1_addr = in1->in(Address)->clone();
972 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
973 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
974 in1_addr->set_req(AddPNode::Offset, offset);
975 in1->set_req(0, base->in(allocation_index));
976 in1->set_req(Address, in1_addr);
977 in1->set_req(Memory, mem_phi->in(allocation_index));
979 Node* in2 = clone();
980 Node* in2_addr = in2->in(Address)->clone();
981 in2_addr->set_req(AddPNode::Base, base->in(load_index));
982 in2_addr->set_req(AddPNode::Address, base->in(load_index));
983 in2_addr->set_req(AddPNode::Offset, offset);
984 in2->set_req(0, base->in(load_index));
985 in2->set_req(Address, in2_addr);
986 in2->set_req(Memory, mem_phi->in(load_index));
988 in1_addr = phase->transform(in1_addr);
989 in1 = phase->transform(in1);
990 in2_addr = phase->transform(in2_addr);
991 in2 = phase->transform(in2);
993 PhiNode* result = PhiNode::make_blank(base->in(0), this);
994 result->set_req(allocation_index, in1);
995 result->set_req(load_index, in2);
996 return result;
997 }
998 } else if (base->is_Load()) {
999 // Eliminate the load of Integer.value for integers from the cache
1000 // array by deriving the value from the index into the array.
1001 // Capture the offset of the load and then reverse the computation.
1002 Node* load_base = base->in(Address)->in(AddPNode::Base);
1003 if (load_base != NULL) {
1004 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1005 intptr_t cache_offset;
1006 int shift = -1;
1007 Node* cache = NULL;
1008 if (is_autobox_cache(atp)) {
1009 shift = exact_log2(type2aelembytes(T_OBJECT));
1010 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1011 }
1012 if (cache != NULL && base->in(Address)->is_AddP()) {
1013 Node* elements[4];
1014 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1015 int cache_low;
1016 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1017 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1018 // Add up all the offsets making of the address of the load
1019 Node* result = elements[0];
1020 for (int i = 1; i < count; i++) {
1021 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1022 }
1023 // Remove the constant offset from the address and then
1024 // remove the scaling of the offset to recover the original index.
1025 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1026 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1027 // Peel the shift off directly but wrap it in a dummy node
1028 // since Ideal can't return existing nodes
1029 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1030 } else {
1031 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1032 }
1033 #ifdef _LP64
1034 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1035 #endif
1036 return result;
1037 }
1038 }
1039 }
1040 }
1041 return NULL;
1042 }
1045 //------------------------------Ideal------------------------------------------
1046 // If the load is from Field memory and the pointer is non-null, we can
1047 // zero out the control input.
1048 // If the offset is constant and the base is an object allocation,
1049 // try to hook me up to the exact initializing store.
1050 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1051 Node* p = MemNode::Ideal_common(phase, can_reshape);
1052 if (p) return (p == NodeSentinel) ? NULL : p;
1054 Node* ctrl = in(MemNode::Control);
1055 Node* address = in(MemNode::Address);
1057 // Skip up past a SafePoint control. Cannot do this for Stores because
1058 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1059 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1060 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1061 ctrl = ctrl->in(0);
1062 set_req(MemNode::Control,ctrl);
1063 }
1065 // Check for useless control edge in some common special cases
1066 if (in(MemNode::Control) != NULL) {
1067 intptr_t ignore = 0;
1068 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1069 if (base != NULL
1070 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1071 && detect_dominating_control(base->in(0), phase->C->start())) {
1072 // A method-invariant, non-null address (constant or 'this' argument).
1073 set_req(MemNode::Control, NULL);
1074 }
1075 }
1077 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
1078 Node* base = in(Address)->in(AddPNode::Base);
1079 if (base != NULL) {
1080 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1081 if (is_autobox_object(atp)) {
1082 Node* result = eliminate_autobox(phase);
1083 if (result != NULL) return result;
1084 }
1085 }
1086 }
1088 Node* mem = in(MemNode::Memory);
1089 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1091 if (addr_t != NULL) {
1092 // try to optimize our memory input
1093 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1094 if (opt_mem != mem) {
1095 set_req(MemNode::Memory, opt_mem);
1096 return this;
1097 }
1098 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1099 if (can_reshape && opt_mem->is_Phi() &&
1100 (t_oop != NULL) && t_oop->is_instance_field()) {
1101 assert(t_oop->offset() != Type::OffsetBot && t_oop->offset() != Type::OffsetTop, "");
1102 Node *region = opt_mem->in(0);
1103 uint cnt = opt_mem->req();
1104 for( uint i = 1; i < cnt; i++ ) {
1105 Node *in = opt_mem->in(i);
1106 if( in == NULL ) {
1107 region = NULL; // Wait stable graph
1108 break;
1109 }
1110 }
1111 if (region != NULL) {
1112 // Check for loop invariant.
1113 if (cnt == 3) {
1114 for( uint i = 1; i < cnt; i++ ) {
1115 Node *in = opt_mem->in(i);
1116 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1117 if (m == opt_mem) {
1118 set_req(MemNode::Memory, opt_mem->in(cnt - i)); // Skip this phi.
1119 return this;
1120 }
1121 }
1122 }
1123 // Split through Phi (see original code in loopopts.cpp).
1124 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1125 const Type* this_type = this->bottom_type();
1126 int this_index = phase->C->get_alias_index(addr_t);
1127 int this_offset = addr_t->offset();
1128 int this_iid = addr_t->is_oopptr()->instance_id();
1129 int wins = 0;
1130 PhaseIterGVN *igvn = phase->is_IterGVN();
1131 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1132 for( uint i = 1; i < region->req(); i++ ) {
1133 Node *x;
1134 Node* the_clone = NULL;
1135 if( region->in(i) == phase->C->top() ) {
1136 x = phase->C->top(); // Dead path? Use a dead data op
1137 } else {
1138 x = this->clone(); // Else clone up the data op
1139 the_clone = x; // Remember for possible deletion.
1140 // Alter data node to use pre-phi inputs
1141 if( this->in(0) == region ) {
1142 x->set_req( 0, region->in(i) );
1143 } else {
1144 x->set_req( 0, NULL );
1145 }
1146 for( uint j = 1; j < this->req(); j++ ) {
1147 Node *in = this->in(j);
1148 if( in->is_Phi() && in->in(0) == region )
1149 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1150 }
1151 }
1152 // Check for a 'win' on some paths
1153 const Type *t = x->Value(igvn);
1155 bool singleton = t->singleton();
1157 // See comments in PhaseIdealLoop::split_thru_phi().
1158 if( singleton && t == Type::TOP ) {
1159 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1160 }
1162 if( singleton ) {
1163 wins++;
1164 x = igvn->makecon(t);
1165 } else {
1166 // We now call Identity to try to simplify the cloned node.
1167 // Note that some Identity methods call phase->type(this).
1168 // Make sure that the type array is big enough for
1169 // our new node, even though we may throw the node away.
1170 // (This tweaking with igvn only works because x is a new node.)
1171 igvn->set_type(x, t);
1172 Node *y = x->Identity(igvn);
1173 if( y != x ) {
1174 wins++;
1175 x = y;
1176 } else {
1177 y = igvn->hash_find(x);
1178 if( y ) {
1179 wins++;
1180 x = y;
1181 } else {
1182 // Else x is a new node we are keeping
1183 // We do not need register_new_node_with_optimizer
1184 // because set_type has already been called.
1185 igvn->_worklist.push(x);
1186 }
1187 }
1188 }
1189 if (x != the_clone && the_clone != NULL)
1190 igvn->remove_dead_node(the_clone);
1191 phi->set_req(i, x);
1192 }
1193 if( wins > 0 ) {
1194 // Record Phi
1195 igvn->register_new_node_with_optimizer(phi);
1196 return phi;
1197 } else {
1198 igvn->remove_dead_node(phi);
1199 }
1200 }
1201 }
1202 }
1204 // Check for prior store with a different base or offset; make Load
1205 // independent. Skip through any number of them. Bail out if the stores
1206 // are in an endless dead cycle and report no progress. This is a key
1207 // transform for Reflection. However, if after skipping through the Stores
1208 // we can't then fold up against a prior store do NOT do the transform as
1209 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1210 // array memory alive twice: once for the hoisted Load and again after the
1211 // bypassed Store. This situation only works if EVERYBODY who does
1212 // anti-dependence work knows how to bypass. I.e. we need all
1213 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1214 // the alias index stuff. So instead, peek through Stores and IFF we can
1215 // fold up, do so.
1216 Node* prev_mem = find_previous_store(phase);
1217 // Steps (a), (b): Walk past independent stores to find an exact match.
1218 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1219 // (c) See if we can fold up on the spot, but don't fold up here.
1220 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
1221 // just return a prior value, which is done by Identity calls.
1222 if (can_see_stored_value(prev_mem, phase)) {
1223 // Make ready for step (d):
1224 set_req(MemNode::Memory, prev_mem);
1225 return this;
1226 }
1227 }
1229 return NULL; // No further progress
1230 }
1232 // Helper to recognize certain Klass fields which are invariant across
1233 // some group of array types (e.g., int[] or all T[] where T < Object).
1234 const Type*
1235 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1236 ciKlass* klass) const {
1237 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1238 // The field is Klass::_modifier_flags. Return its (constant) value.
1239 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1240 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1241 return TypeInt::make(klass->modifier_flags());
1242 }
1243 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1244 // The field is Klass::_access_flags. Return its (constant) value.
1245 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1246 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1247 return TypeInt::make(klass->access_flags());
1248 }
1249 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1250 // The field is Klass::_layout_helper. Return its constant value if known.
1251 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1252 return TypeInt::make(klass->layout_helper());
1253 }
1255 // No match.
1256 return NULL;
1257 }
1259 //------------------------------Value-----------------------------------------
1260 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1261 // Either input is TOP ==> the result is TOP
1262 Node* mem = in(MemNode::Memory);
1263 const Type *t1 = phase->type(mem);
1264 if (t1 == Type::TOP) return Type::TOP;
1265 Node* adr = in(MemNode::Address);
1266 const TypePtr* tp = phase->type(adr)->isa_ptr();
1267 if (tp == NULL || tp->empty()) return Type::TOP;
1268 int off = tp->offset();
1269 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1271 // Try to guess loaded type from pointer type
1272 if (tp->base() == Type::AryPtr) {
1273 const Type *t = tp->is_aryptr()->elem();
1274 // Don't do this for integer types. There is only potential profit if
1275 // the element type t is lower than _type; that is, for int types, if _type is
1276 // more restrictive than t. This only happens here if one is short and the other
1277 // char (both 16 bits), and in those cases we've made an intentional decision
1278 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1279 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1280 //
1281 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1282 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1283 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1284 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1285 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1286 // In fact, that could have been the original type of p1, and p1 could have
1287 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1288 // expression (LShiftL quux 3) independently optimized to the constant 8.
1289 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1290 && Opcode() != Op_LoadKlass) {
1291 // t might actually be lower than _type, if _type is a unique
1292 // concrete subclass of abstract class t.
1293 // Make sure the reference is not into the header, by comparing
1294 // the offset against the offset of the start of the array's data.
1295 // Different array types begin at slightly different offsets (12 vs. 16).
1296 // We choose T_BYTE as an example base type that is least restrictive
1297 // as to alignment, which will therefore produce the smallest
1298 // possible base offset.
1299 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1300 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1301 const Type* jt = t->join(_type);
1302 // In any case, do not allow the join, per se, to empty out the type.
1303 if (jt->empty() && !t->empty()) {
1304 // This can happen if a interface-typed array narrows to a class type.
1305 jt = _type;
1306 }
1308 if (EliminateAutoBox) {
1309 // The pointers in the autobox arrays are always non-null
1310 Node* base = in(Address)->in(AddPNode::Base);
1311 if (base != NULL) {
1312 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1313 if (is_autobox_cache(atp)) {
1314 return jt->join(TypePtr::NOTNULL)->is_ptr();
1315 }
1316 }
1317 }
1318 return jt;
1319 }
1320 }
1321 } else if (tp->base() == Type::InstPtr) {
1322 assert( off != Type::OffsetBot ||
1323 // arrays can be cast to Objects
1324 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1325 // unsafe field access may not have a constant offset
1326 phase->C->has_unsafe_access(),
1327 "Field accesses must be precise" );
1328 // For oop loads, we expect the _type to be precise
1329 } else if (tp->base() == Type::KlassPtr) {
1330 assert( off != Type::OffsetBot ||
1331 // arrays can be cast to Objects
1332 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1333 // also allow array-loading from the primary supertype
1334 // array during subtype checks
1335 Opcode() == Op_LoadKlass,
1336 "Field accesses must be precise" );
1337 // For klass/static loads, we expect the _type to be precise
1338 }
1340 const TypeKlassPtr *tkls = tp->isa_klassptr();
1341 if (tkls != NULL && !StressReflectiveCode) {
1342 ciKlass* klass = tkls->klass();
1343 if (klass->is_loaded() && tkls->klass_is_exact()) {
1344 // We are loading a field from a Klass metaobject whose identity
1345 // is known at compile time (the type is "exact" or "precise").
1346 // Check for fields we know are maintained as constants by the VM.
1347 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1348 // The field is Klass::_super_check_offset. Return its (constant) value.
1349 // (Folds up type checking code.)
1350 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1351 return TypeInt::make(klass->super_check_offset());
1352 }
1353 // Compute index into primary_supers array
1354 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1355 // Check for overflowing; use unsigned compare to handle the negative case.
1356 if( depth < ciKlass::primary_super_limit() ) {
1357 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1358 // (Folds up type checking code.)
1359 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1360 ciKlass *ss = klass->super_of_depth(depth);
1361 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1362 }
1363 const Type* aift = load_array_final_field(tkls, klass);
1364 if (aift != NULL) return aift;
1365 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1366 && klass->is_array_klass()) {
1367 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1368 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1369 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1370 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1371 }
1372 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1373 // The field is Klass::_java_mirror. Return its (constant) value.
1374 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1375 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1376 return TypeInstPtr::make(klass->java_mirror());
1377 }
1378 }
1380 // We can still check if we are loading from the primary_supers array at a
1381 // shallow enough depth. Even though the klass is not exact, entries less
1382 // than or equal to its super depth are correct.
1383 if (klass->is_loaded() ) {
1384 ciType *inner = klass->klass();
1385 while( inner->is_obj_array_klass() )
1386 inner = inner->as_obj_array_klass()->base_element_type();
1387 if( inner->is_instance_klass() &&
1388 !inner->as_instance_klass()->flags().is_interface() ) {
1389 // Compute index into primary_supers array
1390 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1391 // Check for overflowing; use unsigned compare to handle the negative case.
1392 if( depth < ciKlass::primary_super_limit() &&
1393 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1394 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1395 // (Folds up type checking code.)
1396 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1397 ciKlass *ss = klass->super_of_depth(depth);
1398 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1399 }
1400 }
1401 }
1403 // If the type is enough to determine that the thing is not an array,
1404 // we can give the layout_helper a positive interval type.
1405 // This will help short-circuit some reflective code.
1406 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1407 && !klass->is_array_klass() // not directly typed as an array
1408 && !klass->is_interface() // specifically not Serializable & Cloneable
1409 && !klass->is_java_lang_Object() // not the supertype of all T[]
1410 ) {
1411 // Note: When interfaces are reliable, we can narrow the interface
1412 // test to (klass != Serializable && klass != Cloneable).
1413 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1414 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1415 // The key property of this type is that it folds up tests
1416 // for array-ness, since it proves that the layout_helper is positive.
1417 // Thus, a generic value like the basic object layout helper works fine.
1418 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1419 }
1420 }
1422 // If we are loading from a freshly-allocated object, produce a zero,
1423 // if the load is provably beyond the header of the object.
1424 // (Also allow a variable load from a fresh array to produce zero.)
1425 if (ReduceFieldZeroing) {
1426 Node* value = can_see_stored_value(mem,phase);
1427 if (value != NULL && value->is_Con())
1428 return value->bottom_type();
1429 }
1431 const TypeOopPtr *tinst = tp->isa_oopptr();
1432 if (tinst != NULL && tinst->is_instance_field()) {
1433 // If we have an instance type and our memory input is the
1434 // programs's initial memory state, there is no matching store,
1435 // so just return a zero of the appropriate type
1436 Node *mem = in(MemNode::Memory);
1437 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1438 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1439 return Type::get_zero_type(_type->basic_type());
1440 }
1441 }
1442 return _type;
1443 }
1445 //------------------------------match_edge-------------------------------------
1446 // Do we Match on this edge index or not? Match only the address.
1447 uint LoadNode::match_edge(uint idx) const {
1448 return idx == MemNode::Address;
1449 }
1451 //--------------------------LoadBNode::Ideal--------------------------------------
1452 //
1453 // If the previous store is to the same address as this load,
1454 // and the value stored was larger than a byte, replace this load
1455 // with the value stored truncated to a byte. If no truncation is
1456 // needed, the replacement is done in LoadNode::Identity().
1457 //
1458 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1459 Node* mem = in(MemNode::Memory);
1460 Node* value = can_see_stored_value(mem,phase);
1461 if( value && !phase->type(value)->higher_equal( _type ) ) {
1462 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1463 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1464 }
1465 // Identity call will handle the case where truncation is not needed.
1466 return LoadNode::Ideal(phase, can_reshape);
1467 }
1469 //--------------------------LoadCNode::Ideal--------------------------------------
1470 //
1471 // If the previous store is to the same address as this load,
1472 // and the value stored was larger than a char, replace this load
1473 // with the value stored truncated to a char. If no truncation is
1474 // needed, the replacement is done in LoadNode::Identity().
1475 //
1476 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1477 Node* mem = in(MemNode::Memory);
1478 Node* value = can_see_stored_value(mem,phase);
1479 if( value && !phase->type(value)->higher_equal( _type ) )
1480 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1481 // Identity call will handle the case where truncation is not needed.
1482 return LoadNode::Ideal(phase, can_reshape);
1483 }
1485 //--------------------------LoadSNode::Ideal--------------------------------------
1486 //
1487 // If the previous store is to the same address as this load,
1488 // and the value stored was larger than a short, replace this load
1489 // with the value stored truncated to a short. If no truncation is
1490 // needed, the replacement is done in LoadNode::Identity().
1491 //
1492 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1493 Node* mem = in(MemNode::Memory);
1494 Node* value = can_see_stored_value(mem,phase);
1495 if( value && !phase->type(value)->higher_equal( _type ) ) {
1496 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1497 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1498 }
1499 // Identity call will handle the case where truncation is not needed.
1500 return LoadNode::Ideal(phase, can_reshape);
1501 }
1503 //=============================================================================
1504 //------------------------------Value------------------------------------------
1505 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1506 // Either input is TOP ==> the result is TOP
1507 const Type *t1 = phase->type( in(MemNode::Memory) );
1508 if (t1 == Type::TOP) return Type::TOP;
1509 Node *adr = in(MemNode::Address);
1510 const Type *t2 = phase->type( adr );
1511 if (t2 == Type::TOP) return Type::TOP;
1512 const TypePtr *tp = t2->is_ptr();
1513 if (TypePtr::above_centerline(tp->ptr()) ||
1514 tp->ptr() == TypePtr::Null) return Type::TOP;
1516 // Return a more precise klass, if possible
1517 const TypeInstPtr *tinst = tp->isa_instptr();
1518 if (tinst != NULL) {
1519 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1520 int offset = tinst->offset();
1521 if (ik == phase->C->env()->Class_klass()
1522 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1523 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1524 // We are loading a special hidden field from a Class mirror object,
1525 // the field which points to the VM's Klass metaobject.
1526 ciType* t = tinst->java_mirror_type();
1527 // java_mirror_type returns non-null for compile-time Class constants.
1528 if (t != NULL) {
1529 // constant oop => constant klass
1530 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1531 return TypeKlassPtr::make(ciArrayKlass::make(t));
1532 }
1533 if (!t->is_klass()) {
1534 // a primitive Class (e.g., int.class) has NULL for a klass field
1535 return TypePtr::NULL_PTR;
1536 }
1537 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1538 return TypeKlassPtr::make(t->as_klass());
1539 }
1540 // non-constant mirror, so we can't tell what's going on
1541 }
1542 if( !ik->is_loaded() )
1543 return _type; // Bail out if not loaded
1544 if (offset == oopDesc::klass_offset_in_bytes()) {
1545 if (tinst->klass_is_exact()) {
1546 return TypeKlassPtr::make(ik);
1547 }
1548 // See if we can become precise: no subklasses and no interface
1549 // (Note: We need to support verified interfaces.)
1550 if (!ik->is_interface() && !ik->has_subklass()) {
1551 //assert(!UseExactTypes, "this code should be useless with exact types");
1552 // Add a dependence; if any subclass added we need to recompile
1553 if (!ik->is_final()) {
1554 // %%% should use stronger assert_unique_concrete_subtype instead
1555 phase->C->dependencies()->assert_leaf_type(ik);
1556 }
1557 // Return precise klass
1558 return TypeKlassPtr::make(ik);
1559 }
1561 // Return root of possible klass
1562 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1563 }
1564 }
1566 // Check for loading klass from an array
1567 const TypeAryPtr *tary = tp->isa_aryptr();
1568 if( tary != NULL ) {
1569 ciKlass *tary_klass = tary->klass();
1570 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1571 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1572 if (tary->klass_is_exact()) {
1573 return TypeKlassPtr::make(tary_klass);
1574 }
1575 ciArrayKlass *ak = tary->klass()->as_array_klass();
1576 // If the klass is an object array, we defer the question to the
1577 // array component klass.
1578 if( ak->is_obj_array_klass() ) {
1579 assert( ak->is_loaded(), "" );
1580 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1581 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1582 ciInstanceKlass* ik = base_k->as_instance_klass();
1583 // See if we can become precise: no subklasses and no interface
1584 if (!ik->is_interface() && !ik->has_subklass()) {
1585 //assert(!UseExactTypes, "this code should be useless with exact types");
1586 // Add a dependence; if any subclass added we need to recompile
1587 if (!ik->is_final()) {
1588 phase->C->dependencies()->assert_leaf_type(ik);
1589 }
1590 // Return precise array klass
1591 return TypeKlassPtr::make(ak);
1592 }
1593 }
1594 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1595 } else { // Found a type-array?
1596 //assert(!UseExactTypes, "this code should be useless with exact types");
1597 assert( ak->is_type_array_klass(), "" );
1598 return TypeKlassPtr::make(ak); // These are always precise
1599 }
1600 }
1601 }
1603 // Check for loading klass from an array klass
1604 const TypeKlassPtr *tkls = tp->isa_klassptr();
1605 if (tkls != NULL && !StressReflectiveCode) {
1606 ciKlass* klass = tkls->klass();
1607 if( !klass->is_loaded() )
1608 return _type; // Bail out if not loaded
1609 if( klass->is_obj_array_klass() &&
1610 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1611 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1612 // // Always returning precise element type is incorrect,
1613 // // e.g., element type could be object and array may contain strings
1614 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1616 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1617 // according to the element type's subclassing.
1618 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1619 }
1620 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1621 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1622 ciKlass* sup = klass->as_instance_klass()->super();
1623 // The field is Klass::_super. Return its (constant) value.
1624 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1625 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1626 }
1627 }
1629 // Bailout case
1630 return LoadNode::Value(phase);
1631 }
1633 //------------------------------Identity---------------------------------------
1634 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1635 // Also feed through the klass in Allocate(...klass...)._klass.
1636 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1637 Node* x = LoadNode::Identity(phase);
1638 if (x != this) return x;
1640 // Take apart the address into an oop and and offset.
1641 // Return 'this' if we cannot.
1642 Node* adr = in(MemNode::Address);
1643 intptr_t offset = 0;
1644 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1645 if (base == NULL) return this;
1646 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1647 if (toop == NULL) return this;
1649 // We can fetch the klass directly through an AllocateNode.
1650 // This works even if the klass is not constant (clone or newArray).
1651 if (offset == oopDesc::klass_offset_in_bytes()) {
1652 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1653 if (allocated_klass != NULL) {
1654 return allocated_klass;
1655 }
1656 }
1658 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1659 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1660 // See inline_native_Class_query for occurrences of these patterns.
1661 // Java Example: x.getClass().isAssignableFrom(y)
1662 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1663 //
1664 // This improves reflective code, often making the Class
1665 // mirror go completely dead. (Current exception: Class
1666 // mirrors may appear in debug info, but we could clean them out by
1667 // introducing a new debug info operator for klassOop.java_mirror).
1668 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1669 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1670 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1671 // We are loading a special hidden field from a Class mirror,
1672 // the field which points to its Klass or arrayKlass metaobject.
1673 if (base->is_Load()) {
1674 Node* adr2 = base->in(MemNode::Address);
1675 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1676 if (tkls != NULL && !tkls->empty()
1677 && (tkls->klass()->is_instance_klass() ||
1678 tkls->klass()->is_array_klass())
1679 && adr2->is_AddP()
1680 ) {
1681 int mirror_field = Klass::java_mirror_offset_in_bytes();
1682 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1683 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1684 }
1685 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1686 return adr2->in(AddPNode::Base);
1687 }
1688 }
1689 }
1690 }
1692 return this;
1693 }
1695 //------------------------------Value-----------------------------------------
1696 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1697 // Either input is TOP ==> the result is TOP
1698 const Type *t1 = phase->type( in(MemNode::Memory) );
1699 if( t1 == Type::TOP ) return Type::TOP;
1700 Node *adr = in(MemNode::Address);
1701 const Type *t2 = phase->type( adr );
1702 if( t2 == Type::TOP ) return Type::TOP;
1703 const TypePtr *tp = t2->is_ptr();
1704 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1705 const TypeAryPtr *tap = tp->isa_aryptr();
1706 if( !tap ) return _type;
1707 return tap->size();
1708 }
1710 //------------------------------Identity---------------------------------------
1711 // Feed through the length in AllocateArray(...length...)._length.
1712 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1713 Node* x = LoadINode::Identity(phase);
1714 if (x != this) return x;
1716 // Take apart the address into an oop and and offset.
1717 // Return 'this' if we cannot.
1718 Node* adr = in(MemNode::Address);
1719 intptr_t offset = 0;
1720 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1721 if (base == NULL) return this;
1722 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1723 if (tary == NULL) return this;
1725 // We can fetch the length directly through an AllocateArrayNode.
1726 // This works even if the length is not constant (clone or newArray).
1727 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1728 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
1729 if (allocated_length != NULL) {
1730 return allocated_length;
1731 }
1732 }
1734 return this;
1736 }
1737 //=============================================================================
1738 //---------------------------StoreNode::make-----------------------------------
1739 // Polymorphic factory method:
1740 StoreNode* StoreNode::make( Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1741 switch (bt) {
1742 case T_BOOLEAN:
1743 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1744 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1745 case T_CHAR:
1746 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1747 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1748 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1749 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1750 case T_ADDRESS:
1751 case T_OBJECT: return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1752 }
1753 ShouldNotReachHere();
1754 return (StoreNode*)NULL;
1755 }
1757 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1758 bool require_atomic = true;
1759 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1760 }
1763 //--------------------------bottom_type----------------------------------------
1764 const Type *StoreNode::bottom_type() const {
1765 return Type::MEMORY;
1766 }
1768 //------------------------------hash-------------------------------------------
1769 uint StoreNode::hash() const {
1770 // unroll addition of interesting fields
1771 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1773 // Since they are not commoned, do not hash them:
1774 return NO_HASH;
1775 }
1777 //------------------------------Ideal------------------------------------------
1778 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1779 // When a store immediately follows a relevant allocation/initialization,
1780 // try to capture it into the initialization, or hoist it above.
1781 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1782 Node* p = MemNode::Ideal_common(phase, can_reshape);
1783 if (p) return (p == NodeSentinel) ? NULL : p;
1785 Node* mem = in(MemNode::Memory);
1786 Node* address = in(MemNode::Address);
1788 // Back-to-back stores to same address? Fold em up.
1789 // Generally unsafe if I have intervening uses...
1790 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
1791 // Looking at a dead closed cycle of memory?
1792 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1794 assert(Opcode() == mem->Opcode() ||
1795 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1796 "no mismatched stores, except on raw memory");
1798 if (mem->outcnt() == 1 && // check for intervening uses
1799 mem->as_Store()->memory_size() <= this->memory_size()) {
1800 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1801 // For example, 'mem' might be the final state at a conditional return.
1802 // Or, 'mem' might be used by some node which is live at the same time
1803 // 'this' is live, which might be unschedulable. So, require exactly
1804 // ONE user, the 'this' store, until such time as we clone 'mem' for
1805 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1806 if (can_reshape) { // (%%% is this an anachronism?)
1807 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1808 phase->is_IterGVN());
1809 } else {
1810 // It's OK to do this in the parser, since DU info is always accurate,
1811 // and the parser always refers to nodes via SafePointNode maps.
1812 set_req(MemNode::Memory, mem->in(MemNode::Memory));
1813 }
1814 return this;
1815 }
1816 }
1818 // Capture an unaliased, unconditional, simple store into an initializer.
1819 // Or, if it is independent of the allocation, hoist it above the allocation.
1820 if (ReduceFieldZeroing && /*can_reshape &&*/
1821 mem->is_Proj() && mem->in(0)->is_Initialize()) {
1822 InitializeNode* init = mem->in(0)->as_Initialize();
1823 intptr_t offset = init->can_capture_store(this, phase);
1824 if (offset > 0) {
1825 Node* moved = init->capture_store(this, offset, phase);
1826 // If the InitializeNode captured me, it made a raw copy of me,
1827 // and I need to disappear.
1828 if (moved != NULL) {
1829 // %%% hack to ensure that Ideal returns a new node:
1830 mem = MergeMemNode::make(phase->C, mem);
1831 return mem; // fold me away
1832 }
1833 }
1834 }
1836 return NULL; // No further progress
1837 }
1839 //------------------------------Value-----------------------------------------
1840 const Type *StoreNode::Value( PhaseTransform *phase ) const {
1841 // Either input is TOP ==> the result is TOP
1842 const Type *t1 = phase->type( in(MemNode::Memory) );
1843 if( t1 == Type::TOP ) return Type::TOP;
1844 const Type *t2 = phase->type( in(MemNode::Address) );
1845 if( t2 == Type::TOP ) return Type::TOP;
1846 const Type *t3 = phase->type( in(MemNode::ValueIn) );
1847 if( t3 == Type::TOP ) return Type::TOP;
1848 return Type::MEMORY;
1849 }
1851 //------------------------------Identity---------------------------------------
1852 // Remove redundant stores:
1853 // Store(m, p, Load(m, p)) changes to m.
1854 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
1855 Node *StoreNode::Identity( PhaseTransform *phase ) {
1856 Node* mem = in(MemNode::Memory);
1857 Node* adr = in(MemNode::Address);
1858 Node* val = in(MemNode::ValueIn);
1860 // Load then Store? Then the Store is useless
1861 if (val->is_Load() &&
1862 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
1863 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
1864 val->as_Load()->store_Opcode() == Opcode()) {
1865 return mem;
1866 }
1868 // Two stores in a row of the same value?
1869 if (mem->is_Store() &&
1870 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
1871 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
1872 mem->Opcode() == Opcode()) {
1873 return mem;
1874 }
1876 // Store of zero anywhere into a freshly-allocated object?
1877 // Then the store is useless.
1878 // (It must already have been captured by the InitializeNode.)
1879 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
1880 // a newly allocated object is already all-zeroes everywhere
1881 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
1882 return mem;
1883 }
1885 // the store may also apply to zero-bits in an earlier object
1886 Node* prev_mem = find_previous_store(phase);
1887 // Steps (a), (b): Walk past independent stores to find an exact match.
1888 if (prev_mem != NULL) {
1889 Node* prev_val = can_see_stored_value(prev_mem, phase);
1890 if (prev_val != NULL && phase->eqv(prev_val, val)) {
1891 // prev_val and val might differ by a cast; it would be good
1892 // to keep the more informative of the two.
1893 return mem;
1894 }
1895 }
1896 }
1898 return this;
1899 }
1901 //------------------------------match_edge-------------------------------------
1902 // Do we Match on this edge index or not? Match only memory & value
1903 uint StoreNode::match_edge(uint idx) const {
1904 return idx == MemNode::Address || idx == MemNode::ValueIn;
1905 }
1907 //------------------------------cmp--------------------------------------------
1908 // Do not common stores up together. They generally have to be split
1909 // back up anyways, so do not bother.
1910 uint StoreNode::cmp( const Node &n ) const {
1911 return (&n == this); // Always fail except on self
1912 }
1914 //------------------------------Ideal_masked_input-----------------------------
1915 // Check for a useless mask before a partial-word store
1916 // (StoreB ... (AndI valIn conIa) )
1917 // If (conIa & mask == mask) this simplifies to
1918 // (StoreB ... (valIn) )
1919 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
1920 Node *val = in(MemNode::ValueIn);
1921 if( val->Opcode() == Op_AndI ) {
1922 const TypeInt *t = phase->type( val->in(2) )->isa_int();
1923 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
1924 set_req(MemNode::ValueIn, val->in(1));
1925 return this;
1926 }
1927 }
1928 return NULL;
1929 }
1932 //------------------------------Ideal_sign_extended_input----------------------
1933 // Check for useless sign-extension before a partial-word store
1934 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
1935 // If (conIL == conIR && conIR <= num_bits) this simplifies to
1936 // (StoreB ... (valIn) )
1937 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
1938 Node *val = in(MemNode::ValueIn);
1939 if( val->Opcode() == Op_RShiftI ) {
1940 const TypeInt *t = phase->type( val->in(2) )->isa_int();
1941 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
1942 Node *shl = val->in(1);
1943 if( shl->Opcode() == Op_LShiftI ) {
1944 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
1945 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
1946 set_req(MemNode::ValueIn, shl->in(1));
1947 return this;
1948 }
1949 }
1950 }
1951 }
1952 return NULL;
1953 }
1955 //------------------------------value_never_loaded-----------------------------------
1956 // Determine whether there are any possible loads of the value stored.
1957 // For simplicity, we actually check if there are any loads from the
1958 // address stored to, not just for loads of the value stored by this node.
1959 //
1960 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
1961 Node *adr = in(Address);
1962 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
1963 if (adr_oop == NULL)
1964 return false;
1965 if (!adr_oop->is_instance_field())
1966 return false; // if not a distinct instance, there may be aliases of the address
1967 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
1968 Node *use = adr->fast_out(i);
1969 int opc = use->Opcode();
1970 if (use->is_Load() || use->is_LoadStore()) {
1971 return false;
1972 }
1973 }
1974 return true;
1975 }
1977 //=============================================================================
1978 //------------------------------Ideal------------------------------------------
1979 // If the store is from an AND mask that leaves the low bits untouched, then
1980 // we can skip the AND operation. If the store is from a sign-extension
1981 // (a left shift, then right shift) we can skip both.
1982 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
1983 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
1984 if( progress != NULL ) return progress;
1986 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
1987 if( progress != NULL ) return progress;
1989 // Finally check the default case
1990 return StoreNode::Ideal(phase, can_reshape);
1991 }
1993 //=============================================================================
1994 //------------------------------Ideal------------------------------------------
1995 // If the store is from an AND mask that leaves the low bits untouched, then
1996 // we can skip the AND operation
1997 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
1998 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
1999 if( progress != NULL ) return progress;
2001 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2002 if( progress != NULL ) return progress;
2004 // Finally check the default case
2005 return StoreNode::Ideal(phase, can_reshape);
2006 }
2008 //=============================================================================
2009 //------------------------------Identity---------------------------------------
2010 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2011 // No need to card mark when storing a null ptr
2012 Node* my_store = in(MemNode::OopStore);
2013 if (my_store->is_Store()) {
2014 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2015 if( t1 == TypePtr::NULL_PTR ) {
2016 return in(MemNode::Memory);
2017 }
2018 }
2019 return this;
2020 }
2022 //------------------------------Value-----------------------------------------
2023 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2024 // Either input is TOP ==> the result is TOP
2025 const Type *t = phase->type( in(MemNode::Memory) );
2026 if( t == Type::TOP ) return Type::TOP;
2027 t = phase->type( in(MemNode::Address) );
2028 if( t == Type::TOP ) return Type::TOP;
2029 t = phase->type( in(MemNode::ValueIn) );
2030 if( t == Type::TOP ) return Type::TOP;
2031 // If extra input is TOP ==> the result is TOP
2032 t = phase->type( in(MemNode::OopStore) );
2033 if( t == Type::TOP ) return Type::TOP;
2035 return StoreNode::Value( phase );
2036 }
2039 //=============================================================================
2040 //----------------------------------SCMemProjNode------------------------------
2041 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2042 {
2043 return bottom_type();
2044 }
2046 //=============================================================================
2047 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2048 init_req(MemNode::Control, c );
2049 init_req(MemNode::Memory , mem);
2050 init_req(MemNode::Address, adr);
2051 init_req(MemNode::ValueIn, val);
2052 init_req( ExpectedIn, ex );
2053 init_class_id(Class_LoadStore);
2055 }
2057 //=============================================================================
2058 //-------------------------------adr_type--------------------------------------
2059 // Do we Match on this edge index or not? Do not match memory
2060 const TypePtr* ClearArrayNode::adr_type() const {
2061 Node *adr = in(3);
2062 return MemNode::calculate_adr_type(adr->bottom_type());
2063 }
2065 //------------------------------match_edge-------------------------------------
2066 // Do we Match on this edge index or not? Do not match memory
2067 uint ClearArrayNode::match_edge(uint idx) const {
2068 return idx > 1;
2069 }
2071 //------------------------------Identity---------------------------------------
2072 // Clearing a zero length array does nothing
2073 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2074 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2075 }
2077 //------------------------------Idealize---------------------------------------
2078 // Clearing a short array is faster with stores
2079 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2080 const int unit = BytesPerLong;
2081 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2082 if (!t) return NULL;
2083 if (!t->is_con()) return NULL;
2084 intptr_t raw_count = t->get_con();
2085 intptr_t size = raw_count;
2086 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2087 // Clearing nothing uses the Identity call.
2088 // Negative clears are possible on dead ClearArrays
2089 // (see jck test stmt114.stmt11402.val).
2090 if (size <= 0 || size % unit != 0) return NULL;
2091 intptr_t count = size / unit;
2092 // Length too long; use fast hardware clear
2093 if (size > Matcher::init_array_short_size) return NULL;
2094 Node *mem = in(1);
2095 if( phase->type(mem)==Type::TOP ) return NULL;
2096 Node *adr = in(3);
2097 const Type* at = phase->type(adr);
2098 if( at==Type::TOP ) return NULL;
2099 const TypePtr* atp = at->isa_ptr();
2100 // adjust atp to be the correct array element address type
2101 if (atp == NULL) atp = TypePtr::BOTTOM;
2102 else atp = atp->add_offset(Type::OffsetBot);
2103 // Get base for derived pointer purposes
2104 if( adr->Opcode() != Op_AddP ) Unimplemented();
2105 Node *base = adr->in(1);
2107 Node *zero = phase->makecon(TypeLong::ZERO);
2108 Node *off = phase->MakeConX(BytesPerLong);
2109 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2110 count--;
2111 while( count-- ) {
2112 mem = phase->transform(mem);
2113 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2114 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2115 }
2116 return mem;
2117 }
2119 //----------------------------clear_memory-------------------------------------
2120 // Generate code to initialize object storage to zero.
2121 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2122 intptr_t start_offset,
2123 Node* end_offset,
2124 PhaseGVN* phase) {
2125 Compile* C = phase->C;
2126 intptr_t offset = start_offset;
2128 int unit = BytesPerLong;
2129 if ((offset % unit) != 0) {
2130 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2131 adr = phase->transform(adr);
2132 const TypePtr* atp = TypeRawPtr::BOTTOM;
2133 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2134 mem = phase->transform(mem);
2135 offset += BytesPerInt;
2136 }
2137 assert((offset % unit) == 0, "");
2139 // Initialize the remaining stuff, if any, with a ClearArray.
2140 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2141 }
2143 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2144 Node* start_offset,
2145 Node* end_offset,
2146 PhaseGVN* phase) {
2147 if (start_offset == end_offset) {
2148 // nothing to do
2149 return mem;
2150 }
2152 Compile* C = phase->C;
2153 int unit = BytesPerLong;
2154 Node* zbase = start_offset;
2155 Node* zend = end_offset;
2157 // Scale to the unit required by the CPU:
2158 if (!Matcher::init_array_count_is_in_bytes) {
2159 Node* shift = phase->intcon(exact_log2(unit));
2160 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2161 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2162 }
2164 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2165 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2167 // Bulk clear double-words
2168 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2169 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2170 return phase->transform(mem);
2171 }
2173 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2174 intptr_t start_offset,
2175 intptr_t end_offset,
2176 PhaseGVN* phase) {
2177 if (start_offset == end_offset) {
2178 // nothing to do
2179 return mem;
2180 }
2182 Compile* C = phase->C;
2183 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2184 intptr_t done_offset = end_offset;
2185 if ((done_offset % BytesPerLong) != 0) {
2186 done_offset -= BytesPerInt;
2187 }
2188 if (done_offset > start_offset) {
2189 mem = clear_memory(ctl, mem, dest,
2190 start_offset, phase->MakeConX(done_offset), phase);
2191 }
2192 if (done_offset < end_offset) { // emit the final 32-bit store
2193 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2194 adr = phase->transform(adr);
2195 const TypePtr* atp = TypeRawPtr::BOTTOM;
2196 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2197 mem = phase->transform(mem);
2198 done_offset += BytesPerInt;
2199 }
2200 assert(done_offset == end_offset, "");
2201 return mem;
2202 }
2204 //=============================================================================
2205 // Do we match on this edge? No memory edges
2206 uint StrCompNode::match_edge(uint idx) const {
2207 return idx == 5 || idx == 6;
2208 }
2210 //------------------------------Ideal------------------------------------------
2211 // Return a node which is more "ideal" than the current node. Strip out
2212 // control copies
2213 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2214 return remove_dead_region(phase, can_reshape) ? this : NULL;
2215 }
2218 //=============================================================================
2219 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2220 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2221 _adr_type(C->get_adr_type(alias_idx))
2222 {
2223 init_class_id(Class_MemBar);
2224 Node* top = C->top();
2225 init_req(TypeFunc::I_O,top);
2226 init_req(TypeFunc::FramePtr,top);
2227 init_req(TypeFunc::ReturnAdr,top);
2228 if (precedent != NULL)
2229 init_req(TypeFunc::Parms, precedent);
2230 }
2232 //------------------------------cmp--------------------------------------------
2233 uint MemBarNode::hash() const { return NO_HASH; }
2234 uint MemBarNode::cmp( const Node &n ) const {
2235 return (&n == this); // Always fail except on self
2236 }
2238 //------------------------------make-------------------------------------------
2239 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2240 int len = Precedent + (pn == NULL? 0: 1);
2241 switch (opcode) {
2242 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2243 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2244 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2245 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2246 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2247 default: ShouldNotReachHere(); return NULL;
2248 }
2249 }
2251 //------------------------------Ideal------------------------------------------
2252 // Return a node which is more "ideal" than the current node. Strip out
2253 // control copies
2254 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2255 if (remove_dead_region(phase, can_reshape)) return this;
2256 return NULL;
2257 }
2259 //------------------------------Value------------------------------------------
2260 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2261 if( !in(0) ) return Type::TOP;
2262 if( phase->type(in(0)) == Type::TOP )
2263 return Type::TOP;
2264 return TypeTuple::MEMBAR;
2265 }
2267 //------------------------------match------------------------------------------
2268 // Construct projections for memory.
2269 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2270 switch (proj->_con) {
2271 case TypeFunc::Control:
2272 case TypeFunc::Memory:
2273 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2274 }
2275 ShouldNotReachHere();
2276 return NULL;
2277 }
2279 //===========================InitializeNode====================================
2280 // SUMMARY:
2281 // This node acts as a memory barrier on raw memory, after some raw stores.
2282 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2283 // The Initialize can 'capture' suitably constrained stores as raw inits.
2284 // It can coalesce related raw stores into larger units (called 'tiles').
2285 // It can avoid zeroing new storage for memory units which have raw inits.
2286 // At macro-expansion, it is marked 'complete', and does not optimize further.
2287 //
2288 // EXAMPLE:
2289 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2290 // ctl = incoming control; mem* = incoming memory
2291 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2292 // First allocate uninitialized memory and fill in the header:
2293 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2294 // ctl := alloc.Control; mem* := alloc.Memory*
2295 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2296 // Then initialize to zero the non-header parts of the raw memory block:
2297 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2298 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2299 // After the initialize node executes, the object is ready for service:
2300 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2301 // Suppose its body is immediately initialized as {1,2}:
2302 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2303 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2304 // mem.SLICE(#short[*]) := store2
2305 //
2306 // DETAILS:
2307 // An InitializeNode collects and isolates object initialization after
2308 // an AllocateNode and before the next possible safepoint. As a
2309 // memory barrier (MemBarNode), it keeps critical stores from drifting
2310 // down past any safepoint or any publication of the allocation.
2311 // Before this barrier, a newly-allocated object may have uninitialized bits.
2312 // After this barrier, it may be treated as a real oop, and GC is allowed.
2313 //
2314 // The semantics of the InitializeNode include an implicit zeroing of
2315 // the new object from object header to the end of the object.
2316 // (The object header and end are determined by the AllocateNode.)
2317 //
2318 // Certain stores may be added as direct inputs to the InitializeNode.
2319 // These stores must update raw memory, and they must be to addresses
2320 // derived from the raw address produced by AllocateNode, and with
2321 // a constant offset. They must be ordered by increasing offset.
2322 // The first one is at in(RawStores), the last at in(req()-1).
2323 // Unlike most memory operations, they are not linked in a chain,
2324 // but are displayed in parallel as users of the rawmem output of
2325 // the allocation.
2326 //
2327 // (See comments in InitializeNode::capture_store, which continue
2328 // the example given above.)
2329 //
2330 // When the associated Allocate is macro-expanded, the InitializeNode
2331 // may be rewritten to optimize collected stores. A ClearArrayNode
2332 // may also be created at that point to represent any required zeroing.
2333 // The InitializeNode is then marked 'complete', prohibiting further
2334 // capturing of nearby memory operations.
2335 //
2336 // During macro-expansion, all captured initializations which store
2337 // constant values of 32 bits or smaller are coalesced (if advantagous)
2338 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2339 // initialized in fewer memory operations. Memory words which are
2340 // covered by neither tiles nor non-constant stores are pre-zeroed
2341 // by explicit stores of zero. (The code shape happens to do all
2342 // zeroing first, then all other stores, with both sequences occurring
2343 // in order of ascending offsets.)
2344 //
2345 // Alternatively, code may be inserted between an AllocateNode and its
2346 // InitializeNode, to perform arbitrary initialization of the new object.
2347 // E.g., the object copying intrinsics insert complex data transfers here.
2348 // The initialization must then be marked as 'complete' disable the
2349 // built-in zeroing semantics and the collection of initializing stores.
2350 //
2351 // While an InitializeNode is incomplete, reads from the memory state
2352 // produced by it are optimizable if they match the control edge and
2353 // new oop address associated with the allocation/initialization.
2354 // They return a stored value (if the offset matches) or else zero.
2355 // A write to the memory state, if it matches control and address,
2356 // and if it is to a constant offset, may be 'captured' by the
2357 // InitializeNode. It is cloned as a raw memory operation and rewired
2358 // inside the initialization, to the raw oop produced by the allocation.
2359 // Operations on addresses which are provably distinct (e.g., to
2360 // other AllocateNodes) are allowed to bypass the initialization.
2361 //
2362 // The effect of all this is to consolidate object initialization
2363 // (both arrays and non-arrays, both piecewise and bulk) into a
2364 // single location, where it can be optimized as a unit.
2365 //
2366 // Only stores with an offset less than TrackedInitializationLimit words
2367 // will be considered for capture by an InitializeNode. This puts a
2368 // reasonable limit on the complexity of optimized initializations.
2370 //---------------------------InitializeNode------------------------------------
2371 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2372 : _is_complete(false),
2373 MemBarNode(C, adr_type, rawoop)
2374 {
2375 init_class_id(Class_Initialize);
2377 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2378 assert(in(RawAddress) == rawoop, "proper init");
2379 // Note: allocation() can be NULL, for secondary initialization barriers
2380 }
2382 // Since this node is not matched, it will be processed by the
2383 // register allocator. Declare that there are no constraints
2384 // on the allocation of the RawAddress edge.
2385 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2386 // This edge should be set to top, by the set_complete. But be conservative.
2387 if (idx == InitializeNode::RawAddress)
2388 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2389 return RegMask::Empty;
2390 }
2392 Node* InitializeNode::memory(uint alias_idx) {
2393 Node* mem = in(Memory);
2394 if (mem->is_MergeMem()) {
2395 return mem->as_MergeMem()->memory_at(alias_idx);
2396 } else {
2397 // incoming raw memory is not split
2398 return mem;
2399 }
2400 }
2402 bool InitializeNode::is_non_zero() {
2403 if (is_complete()) return false;
2404 remove_extra_zeroes();
2405 return (req() > RawStores);
2406 }
2408 void InitializeNode::set_complete(PhaseGVN* phase) {
2409 assert(!is_complete(), "caller responsibility");
2410 _is_complete = true;
2412 // After this node is complete, it contains a bunch of
2413 // raw-memory initializations. There is no need for
2414 // it to have anything to do with non-raw memory effects.
2415 // Therefore, tell all non-raw users to re-optimize themselves,
2416 // after skipping the memory effects of this initialization.
2417 PhaseIterGVN* igvn = phase->is_IterGVN();
2418 if (igvn) igvn->add_users_to_worklist(this);
2419 }
2421 // convenience function
2422 // return false if the init contains any stores already
2423 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2424 InitializeNode* init = initialization();
2425 if (init == NULL || init->is_complete()) return false;
2426 init->remove_extra_zeroes();
2427 // for now, if this allocation has already collected any inits, bail:
2428 if (init->is_non_zero()) return false;
2429 init->set_complete(phase);
2430 return true;
2431 }
2433 void InitializeNode::remove_extra_zeroes() {
2434 if (req() == RawStores) return;
2435 Node* zmem = zero_memory();
2436 uint fill = RawStores;
2437 for (uint i = fill; i < req(); i++) {
2438 Node* n = in(i);
2439 if (n->is_top() || n == zmem) continue; // skip
2440 if (fill < i) set_req(fill, n); // compact
2441 ++fill;
2442 }
2443 // delete any empty spaces created:
2444 while (fill < req()) {
2445 del_req(fill);
2446 }
2447 }
2449 // Helper for remembering which stores go with which offsets.
2450 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2451 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2452 intptr_t offset = -1;
2453 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2454 phase, offset);
2455 if (base == NULL) return -1; // something is dead,
2456 if (offset < 0) return -1; // dead, dead
2457 return offset;
2458 }
2460 // Helper for proving that an initialization expression is
2461 // "simple enough" to be folded into an object initialization.
2462 // Attempts to prove that a store's initial value 'n' can be captured
2463 // within the initialization without creating a vicious cycle, such as:
2464 // { Foo p = new Foo(); p.next = p; }
2465 // True for constants and parameters and small combinations thereof.
2466 bool InitializeNode::detect_init_independence(Node* n,
2467 bool st_is_pinned,
2468 int& count) {
2469 if (n == NULL) return true; // (can this really happen?)
2470 if (n->is_Proj()) n = n->in(0);
2471 if (n == this) return false; // found a cycle
2472 if (n->is_Con()) return true;
2473 if (n->is_Start()) return true; // params, etc., are OK
2474 if (n->is_Root()) return true; // even better
2476 Node* ctl = n->in(0);
2477 if (ctl != NULL && !ctl->is_top()) {
2478 if (ctl->is_Proj()) ctl = ctl->in(0);
2479 if (ctl == this) return false;
2481 // If we already know that the enclosing memory op is pinned right after
2482 // the init, then any control flow that the store has picked up
2483 // must have preceded the init, or else be equal to the init.
2484 // Even after loop optimizations (which might change control edges)
2485 // a store is never pinned *before* the availability of its inputs.
2486 if (!MemNode::detect_dominating_control(ctl, this->in(0)))
2487 return false; // failed to prove a good control
2489 }
2491 // Check data edges for possible dependencies on 'this'.
2492 if ((count += 1) > 20) return false; // complexity limit
2493 for (uint i = 1; i < n->req(); i++) {
2494 Node* m = n->in(i);
2495 if (m == NULL || m == n || m->is_top()) continue;
2496 uint first_i = n->find_edge(m);
2497 if (i != first_i) continue; // process duplicate edge just once
2498 if (!detect_init_independence(m, st_is_pinned, count)) {
2499 return false;
2500 }
2501 }
2503 return true;
2504 }
2506 // Here are all the checks a Store must pass before it can be moved into
2507 // an initialization. Returns zero if a check fails.
2508 // On success, returns the (constant) offset to which the store applies,
2509 // within the initialized memory.
2510 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2511 const int FAIL = 0;
2512 if (st->req() != MemNode::ValueIn + 1)
2513 return FAIL; // an inscrutable StoreNode (card mark?)
2514 Node* ctl = st->in(MemNode::Control);
2515 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2516 return FAIL; // must be unconditional after the initialization
2517 Node* mem = st->in(MemNode::Memory);
2518 if (!(mem->is_Proj() && mem->in(0) == this))
2519 return FAIL; // must not be preceded by other stores
2520 Node* adr = st->in(MemNode::Address);
2521 intptr_t offset;
2522 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2523 if (alloc == NULL)
2524 return FAIL; // inscrutable address
2525 if (alloc != allocation())
2526 return FAIL; // wrong allocation! (store needs to float up)
2527 Node* val = st->in(MemNode::ValueIn);
2528 int complexity_count = 0;
2529 if (!detect_init_independence(val, true, complexity_count))
2530 return FAIL; // stored value must be 'simple enough'
2532 return offset; // success
2533 }
2535 // Find the captured store in(i) which corresponds to the range
2536 // [start..start+size) in the initialized object.
2537 // If there is one, return its index i. If there isn't, return the
2538 // negative of the index where it should be inserted.
2539 // Return 0 if the queried range overlaps an initialization boundary
2540 // or if dead code is encountered.
2541 // If size_in_bytes is zero, do not bother with overlap checks.
2542 int InitializeNode::captured_store_insertion_point(intptr_t start,
2543 int size_in_bytes,
2544 PhaseTransform* phase) {
2545 const int FAIL = 0, MAX_STORE = BytesPerLong;
2547 if (is_complete())
2548 return FAIL; // arraycopy got here first; punt
2550 assert(allocation() != NULL, "must be present");
2552 // no negatives, no header fields:
2553 if (start < (intptr_t) sizeof(oopDesc)) return FAIL;
2554 if (start < (intptr_t) sizeof(arrayOopDesc) &&
2555 start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2557 // after a certain size, we bail out on tracking all the stores:
2558 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2559 if (start >= ti_limit) return FAIL;
2561 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2562 if (i >= limit) return -(int)i; // not found; here is where to put it
2564 Node* st = in(i);
2565 intptr_t st_off = get_store_offset(st, phase);
2566 if (st_off < 0) {
2567 if (st != zero_memory()) {
2568 return FAIL; // bail out if there is dead garbage
2569 }
2570 } else if (st_off > start) {
2571 // ...we are done, since stores are ordered
2572 if (st_off < start + size_in_bytes) {
2573 return FAIL; // the next store overlaps
2574 }
2575 return -(int)i; // not found; here is where to put it
2576 } else if (st_off < start) {
2577 if (size_in_bytes != 0 &&
2578 start < st_off + MAX_STORE &&
2579 start < st_off + st->as_Store()->memory_size()) {
2580 return FAIL; // the previous store overlaps
2581 }
2582 } else {
2583 if (size_in_bytes != 0 &&
2584 st->as_Store()->memory_size() != size_in_bytes) {
2585 return FAIL; // mismatched store size
2586 }
2587 return i;
2588 }
2590 ++i;
2591 }
2592 }
2594 // Look for a captured store which initializes at the offset 'start'
2595 // with the given size. If there is no such store, and no other
2596 // initialization interferes, then return zero_memory (the memory
2597 // projection of the AllocateNode).
2598 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2599 PhaseTransform* phase) {
2600 assert(stores_are_sane(phase), "");
2601 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2602 if (i == 0) {
2603 return NULL; // something is dead
2604 } else if (i < 0) {
2605 return zero_memory(); // just primordial zero bits here
2606 } else {
2607 Node* st = in(i); // here is the store at this position
2608 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2609 return st;
2610 }
2611 }
2613 // Create, as a raw pointer, an address within my new object at 'offset'.
2614 Node* InitializeNode::make_raw_address(intptr_t offset,
2615 PhaseTransform* phase) {
2616 Node* addr = in(RawAddress);
2617 if (offset != 0) {
2618 Compile* C = phase->C;
2619 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2620 phase->MakeConX(offset)) );
2621 }
2622 return addr;
2623 }
2625 // Clone the given store, converting it into a raw store
2626 // initializing a field or element of my new object.
2627 // Caller is responsible for retiring the original store,
2628 // with subsume_node or the like.
2629 //
2630 // From the example above InitializeNode::InitializeNode,
2631 // here are the old stores to be captured:
2632 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2633 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2634 //
2635 // Here is the changed code; note the extra edges on init:
2636 // alloc = (Allocate ...)
2637 // rawoop = alloc.RawAddress
2638 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2639 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2640 // init = (Initialize alloc.Control alloc.Memory rawoop
2641 // rawstore1 rawstore2)
2642 //
2643 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2644 PhaseTransform* phase) {
2645 assert(stores_are_sane(phase), "");
2647 if (start < 0) return NULL;
2648 assert(can_capture_store(st, phase) == start, "sanity");
2650 Compile* C = phase->C;
2651 int size_in_bytes = st->memory_size();
2652 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2653 if (i == 0) return NULL; // bail out
2654 Node* prev_mem = NULL; // raw memory for the captured store
2655 if (i > 0) {
2656 prev_mem = in(i); // there is a pre-existing store under this one
2657 set_req(i, C->top()); // temporarily disconnect it
2658 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2659 } else {
2660 i = -i; // no pre-existing store
2661 prev_mem = zero_memory(); // a slice of the newly allocated object
2662 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2663 set_req(--i, C->top()); // reuse this edge; it has been folded away
2664 else
2665 ins_req(i, C->top()); // build a new edge
2666 }
2667 Node* new_st = st->clone();
2668 new_st->set_req(MemNode::Control, in(Control));
2669 new_st->set_req(MemNode::Memory, prev_mem);
2670 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2671 new_st = phase->transform(new_st);
2673 // At this point, new_st might have swallowed a pre-existing store
2674 // at the same offset, or perhaps new_st might have disappeared,
2675 // if it redundantly stored the same value (or zero to fresh memory).
2677 // In any case, wire it in:
2678 set_req(i, new_st);
2680 // The caller may now kill the old guy.
2681 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2682 assert(check_st == new_st || check_st == NULL, "must be findable");
2683 assert(!is_complete(), "");
2684 return new_st;
2685 }
2687 static bool store_constant(jlong* tiles, int num_tiles,
2688 intptr_t st_off, int st_size,
2689 jlong con) {
2690 if ((st_off & (st_size-1)) != 0)
2691 return false; // strange store offset (assume size==2**N)
2692 address addr = (address)tiles + st_off;
2693 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2694 switch (st_size) {
2695 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2696 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2697 case sizeof(jint): *(jint*) addr = (jint) con; break;
2698 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2699 default: return false; // strange store size (detect size!=2**N here)
2700 }
2701 return true; // return success to caller
2702 }
2704 // Coalesce subword constants into int constants and possibly
2705 // into long constants. The goal, if the CPU permits,
2706 // is to initialize the object with a small number of 64-bit tiles.
2707 // Also, convert floating-point constants to bit patterns.
2708 // Non-constants are not relevant to this pass.
2709 //
2710 // In terms of the running example on InitializeNode::InitializeNode
2711 // and InitializeNode::capture_store, here is the transformation
2712 // of rawstore1 and rawstore2 into rawstore12:
2713 // alloc = (Allocate ...)
2714 // rawoop = alloc.RawAddress
2715 // tile12 = 0x00010002
2716 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2717 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2718 //
2719 void
2720 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2721 Node* size_in_bytes,
2722 PhaseGVN* phase) {
2723 Compile* C = phase->C;
2725 assert(stores_are_sane(phase), "");
2726 // Note: After this pass, they are not completely sane,
2727 // since there may be some overlaps.
2729 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2731 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2732 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2733 size_limit = MIN2(size_limit, ti_limit);
2734 size_limit = align_size_up(size_limit, BytesPerLong);
2735 int num_tiles = size_limit / BytesPerLong;
2737 // allocate space for the tile map:
2738 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2739 jlong tiles_buf[small_len];
2740 Node* nodes_buf[small_len];
2741 jlong inits_buf[small_len];
2742 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2743 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2744 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2745 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2746 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2747 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2748 // tiles: exact bitwise model of all primitive constants
2749 // nodes: last constant-storing node subsumed into the tiles model
2750 // inits: which bytes (in each tile) are touched by any initializations
2752 //// Pass A: Fill in the tile model with any relevant stores.
2754 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2755 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2756 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2757 Node* zmem = zero_memory(); // initially zero memory state
2758 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2759 Node* st = in(i);
2760 intptr_t st_off = get_store_offset(st, phase);
2762 // Figure out the store's offset and constant value:
2763 if (st_off < header_size) continue; //skip (ignore header)
2764 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
2765 int st_size = st->as_Store()->memory_size();
2766 if (st_off + st_size > size_limit) break;
2768 // Record which bytes are touched, whether by constant or not.
2769 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
2770 continue; // skip (strange store size)
2772 const Type* val = phase->type(st->in(MemNode::ValueIn));
2773 if (!val->singleton()) continue; //skip (non-con store)
2774 BasicType type = val->basic_type();
2776 jlong con = 0;
2777 switch (type) {
2778 case T_INT: con = val->is_int()->get_con(); break;
2779 case T_LONG: con = val->is_long()->get_con(); break;
2780 case T_FLOAT: con = jint_cast(val->getf()); break;
2781 case T_DOUBLE: con = jlong_cast(val->getd()); break;
2782 default: continue; //skip (odd store type)
2783 }
2785 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
2786 st->Opcode() == Op_StoreL) {
2787 continue; // This StoreL is already optimal.
2788 }
2790 // Store down the constant.
2791 store_constant(tiles, num_tiles, st_off, st_size, con);
2793 intptr_t j = st_off >> LogBytesPerLong;
2795 if (type == T_INT && st_size == BytesPerInt
2796 && (st_off & BytesPerInt) == BytesPerInt) {
2797 jlong lcon = tiles[j];
2798 if (!Matcher::isSimpleConstant64(lcon) &&
2799 st->Opcode() == Op_StoreI) {
2800 // This StoreI is already optimal by itself.
2801 jint* intcon = (jint*) &tiles[j];
2802 intcon[1] = 0; // undo the store_constant()
2804 // If the previous store is also optimal by itself, back up and
2805 // undo the action of the previous loop iteration... if we can.
2806 // But if we can't, just let the previous half take care of itself.
2807 st = nodes[j];
2808 st_off -= BytesPerInt;
2809 con = intcon[0];
2810 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
2811 assert(st_off >= header_size, "still ignoring header");
2812 assert(get_store_offset(st, phase) == st_off, "must be");
2813 assert(in(i-1) == zmem, "must be");
2814 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
2815 assert(con == tcon->is_int()->get_con(), "must be");
2816 // Undo the effects of the previous loop trip, which swallowed st:
2817 intcon[0] = 0; // undo store_constant()
2818 set_req(i-1, st); // undo set_req(i, zmem)
2819 nodes[j] = NULL; // undo nodes[j] = st
2820 --old_subword; // undo ++old_subword
2821 }
2822 continue; // This StoreI is already optimal.
2823 }
2824 }
2826 // This store is not needed.
2827 set_req(i, zmem);
2828 nodes[j] = st; // record for the moment
2829 if (st_size < BytesPerLong) // something has changed
2830 ++old_subword; // includes int/float, but who's counting...
2831 else ++old_long;
2832 }
2834 if ((old_subword + old_long) == 0)
2835 return; // nothing more to do
2837 //// Pass B: Convert any non-zero tiles into optimal constant stores.
2838 // Be sure to insert them before overlapping non-constant stores.
2839 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
2840 for (int j = 0; j < num_tiles; j++) {
2841 jlong con = tiles[j];
2842 jlong init = inits[j];
2843 if (con == 0) continue;
2844 jint con0, con1; // split the constant, address-wise
2845 jint init0, init1; // split the init map, address-wise
2846 { union { jlong con; jint intcon[2]; } u;
2847 u.con = con;
2848 con0 = u.intcon[0];
2849 con1 = u.intcon[1];
2850 u.con = init;
2851 init0 = u.intcon[0];
2852 init1 = u.intcon[1];
2853 }
2855 Node* old = nodes[j];
2856 assert(old != NULL, "need the prior store");
2857 intptr_t offset = (j * BytesPerLong);
2859 bool split = !Matcher::isSimpleConstant64(con);
2861 if (offset < header_size) {
2862 assert(offset + BytesPerInt >= header_size, "second int counts");
2863 assert(*(jint*)&tiles[j] == 0, "junk in header");
2864 split = true; // only the second word counts
2865 // Example: int a[] = { 42 ... }
2866 } else if (con0 == 0 && init0 == -1) {
2867 split = true; // first word is covered by full inits
2868 // Example: int a[] = { ... foo(), 42 ... }
2869 } else if (con1 == 0 && init1 == -1) {
2870 split = true; // second word is covered by full inits
2871 // Example: int a[] = { ... 42, foo() ... }
2872 }
2874 // Here's a case where init0 is neither 0 nor -1:
2875 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
2876 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
2877 // In this case the tile is not split; it is (jlong)42.
2878 // The big tile is stored down, and then the foo() value is inserted.
2879 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
2881 Node* ctl = old->in(MemNode::Control);
2882 Node* adr = make_raw_address(offset, phase);
2883 const TypePtr* atp = TypeRawPtr::BOTTOM;
2885 // One or two coalesced stores to plop down.
2886 Node* st[2];
2887 intptr_t off[2];
2888 int nst = 0;
2889 if (!split) {
2890 ++new_long;
2891 off[nst] = offset;
2892 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2893 phase->longcon(con), T_LONG);
2894 } else {
2895 // Omit either if it is a zero.
2896 if (con0 != 0) {
2897 ++new_int;
2898 off[nst] = offset;
2899 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2900 phase->intcon(con0), T_INT);
2901 }
2902 if (con1 != 0) {
2903 ++new_int;
2904 offset += BytesPerInt;
2905 adr = make_raw_address(offset, phase);
2906 off[nst] = offset;
2907 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
2908 phase->intcon(con1), T_INT);
2909 }
2910 }
2912 // Insert second store first, then the first before the second.
2913 // Insert each one just before any overlapping non-constant stores.
2914 while (nst > 0) {
2915 Node* st1 = st[--nst];
2916 C->copy_node_notes_to(st1, old);
2917 st1 = phase->transform(st1);
2918 offset = off[nst];
2919 assert(offset >= header_size, "do not smash header");
2920 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
2921 guarantee(ins_idx != 0, "must re-insert constant store");
2922 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
2923 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
2924 set_req(--ins_idx, st1);
2925 else
2926 ins_req(ins_idx, st1);
2927 }
2928 }
2930 if (PrintCompilation && WizardMode)
2931 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
2932 old_subword, old_long, new_int, new_long);
2933 if (C->log() != NULL)
2934 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
2935 old_subword, old_long, new_int, new_long);
2937 // Clean up any remaining occurrences of zmem:
2938 remove_extra_zeroes();
2939 }
2941 // Explore forward from in(start) to find the first fully initialized
2942 // word, and return its offset. Skip groups of subword stores which
2943 // together initialize full words. If in(start) is itself part of a
2944 // fully initialized word, return the offset of in(start). If there
2945 // are no following full-word stores, or if something is fishy, return
2946 // a negative value.
2947 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
2948 int int_map = 0;
2949 intptr_t int_map_off = 0;
2950 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
2952 for (uint i = start, limit = req(); i < limit; i++) {
2953 Node* st = in(i);
2955 intptr_t st_off = get_store_offset(st, phase);
2956 if (st_off < 0) break; // return conservative answer
2958 int st_size = st->as_Store()->memory_size();
2959 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
2960 return st_off; // we found a complete word init
2961 }
2963 // update the map:
2965 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
2966 if (this_int_off != int_map_off) {
2967 // reset the map:
2968 int_map = 0;
2969 int_map_off = this_int_off;
2970 }
2972 int subword_off = st_off - this_int_off;
2973 int_map |= right_n_bits(st_size) << subword_off;
2974 if ((int_map & FULL_MAP) == FULL_MAP) {
2975 return this_int_off; // we found a complete word init
2976 }
2978 // Did this store hit or cross the word boundary?
2979 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
2980 if (next_int_off == this_int_off + BytesPerInt) {
2981 // We passed the current int, without fully initializing it.
2982 int_map_off = next_int_off;
2983 int_map >>= BytesPerInt;
2984 } else if (next_int_off > this_int_off + BytesPerInt) {
2985 // We passed the current and next int.
2986 return this_int_off + BytesPerInt;
2987 }
2988 }
2990 return -1;
2991 }
2994 // Called when the associated AllocateNode is expanded into CFG.
2995 // At this point, we may perform additional optimizations.
2996 // Linearize the stores by ascending offset, to make memory
2997 // activity as coherent as possible.
2998 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
2999 intptr_t header_size,
3000 Node* size_in_bytes,
3001 PhaseGVN* phase) {
3002 assert(!is_complete(), "not already complete");
3003 assert(stores_are_sane(phase), "");
3004 assert(allocation() != NULL, "must be present");
3006 remove_extra_zeroes();
3008 if (ReduceFieldZeroing || ReduceBulkZeroing)
3009 // reduce instruction count for common initialization patterns
3010 coalesce_subword_stores(header_size, size_in_bytes, phase);
3012 Node* zmem = zero_memory(); // initially zero memory state
3013 Node* inits = zmem; // accumulating a linearized chain of inits
3014 #ifdef ASSERT
3015 intptr_t last_init_off = sizeof(oopDesc); // previous init offset
3016 intptr_t last_init_end = sizeof(oopDesc); // previous init offset+size
3017 intptr_t last_tile_end = sizeof(oopDesc); // previous tile offset+size
3018 #endif
3019 intptr_t zeroes_done = header_size;
3021 bool do_zeroing = true; // we might give up if inits are very sparse
3022 int big_init_gaps = 0; // how many large gaps have we seen?
3024 if (ZeroTLAB) do_zeroing = false;
3025 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3027 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3028 Node* st = in(i);
3029 intptr_t st_off = get_store_offset(st, phase);
3030 if (st_off < 0)
3031 break; // unknown junk in the inits
3032 if (st->in(MemNode::Memory) != zmem)
3033 break; // complicated store chains somehow in list
3035 int st_size = st->as_Store()->memory_size();
3036 intptr_t next_init_off = st_off + st_size;
3038 if (do_zeroing && zeroes_done < next_init_off) {
3039 // See if this store needs a zero before it or under it.
3040 intptr_t zeroes_needed = st_off;
3042 if (st_size < BytesPerInt) {
3043 // Look for subword stores which only partially initialize words.
3044 // If we find some, we must lay down some word-level zeroes first,
3045 // underneath the subword stores.
3046 //
3047 // Examples:
3048 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3049 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3050 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3051 //
3052 // Note: coalesce_subword_stores may have already done this,
3053 // if it was prompted by constant non-zero subword initializers.
3054 // But this case can still arise with non-constant stores.
3056 intptr_t next_full_store = find_next_fullword_store(i, phase);
3058 // In the examples above:
3059 // in(i) p q r s x y z
3060 // st_off 12 13 14 15 12 13 14
3061 // st_size 1 1 1 1 1 1 1
3062 // next_full_s. 12 16 16 16 16 16 16
3063 // z's_done 12 16 16 16 12 16 12
3064 // z's_needed 12 16 16 16 16 16 16
3065 // zsize 0 0 0 0 4 0 4
3066 if (next_full_store < 0) {
3067 // Conservative tack: Zero to end of current word.
3068 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3069 } else {
3070 // Zero to beginning of next fully initialized word.
3071 // Or, don't zero at all, if we are already in that word.
3072 assert(next_full_store >= zeroes_needed, "must go forward");
3073 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3074 zeroes_needed = next_full_store;
3075 }
3076 }
3078 if (zeroes_needed > zeroes_done) {
3079 intptr_t zsize = zeroes_needed - zeroes_done;
3080 // Do some incremental zeroing on rawmem, in parallel with inits.
3081 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3082 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3083 zeroes_done, zeroes_needed,
3084 phase);
3085 zeroes_done = zeroes_needed;
3086 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3087 do_zeroing = false; // leave the hole, next time
3088 }
3089 }
3091 // Collect the store and move on:
3092 st->set_req(MemNode::Memory, inits);
3093 inits = st; // put it on the linearized chain
3094 set_req(i, zmem); // unhook from previous position
3096 if (zeroes_done == st_off)
3097 zeroes_done = next_init_off;
3099 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3101 #ifdef ASSERT
3102 // Various order invariants. Weaker than stores_are_sane because
3103 // a large constant tile can be filled in by smaller non-constant stores.
3104 assert(st_off >= last_init_off, "inits do not reverse");
3105 last_init_off = st_off;
3106 const Type* val = NULL;
3107 if (st_size >= BytesPerInt &&
3108 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3109 (int)val->basic_type() < (int)T_OBJECT) {
3110 assert(st_off >= last_tile_end, "tiles do not overlap");
3111 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3112 last_tile_end = MAX2(last_tile_end, next_init_off);
3113 } else {
3114 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3115 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3116 assert(st_off >= last_init_end, "inits do not overlap");
3117 last_init_end = next_init_off; // it's a non-tile
3118 }
3119 #endif //ASSERT
3120 }
3122 remove_extra_zeroes(); // clear out all the zmems left over
3123 add_req(inits);
3125 if (!ZeroTLAB) {
3126 // If anything remains to be zeroed, zero it all now.
3127 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3128 // if it is the last unused 4 bytes of an instance, forget about it
3129 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3130 if (zeroes_done + BytesPerLong >= size_limit) {
3131 assert(allocation() != NULL, "");
3132 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3133 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3134 if (zeroes_done == k->layout_helper())
3135 zeroes_done = size_limit;
3136 }
3137 if (zeroes_done < size_limit) {
3138 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3139 zeroes_done, size_in_bytes, phase);
3140 }
3141 }
3143 set_complete(phase);
3144 return rawmem;
3145 }
3148 #ifdef ASSERT
3149 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3150 if (is_complete())
3151 return true; // stores could be anything at this point
3152 intptr_t last_off = sizeof(oopDesc);
3153 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3154 Node* st = in(i);
3155 intptr_t st_off = get_store_offset(st, phase);
3156 if (st_off < 0) continue; // ignore dead garbage
3157 if (last_off > st_off) {
3158 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3159 this->dump(2);
3160 assert(false, "ascending store offsets");
3161 return false;
3162 }
3163 last_off = st_off + st->as_Store()->memory_size();
3164 }
3165 return true;
3166 }
3167 #endif //ASSERT
3172 //============================MergeMemNode=====================================
3173 //
3174 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3175 // contributing store or call operations. Each contributor provides the memory
3176 // state for a particular "alias type" (see Compile::alias_type). For example,
3177 // if a MergeMem has an input X for alias category #6, then any memory reference
3178 // to alias category #6 may use X as its memory state input, as an exact equivalent
3179 // to using the MergeMem as a whole.
3180 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3181 //
3182 // (Here, the <N> notation gives the index of the relevant adr_type.)
3183 //
3184 // In one special case (and more cases in the future), alias categories overlap.
3185 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3186 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3187 // it is exactly equivalent to that state W:
3188 // MergeMem(<Bot>: W) <==> W
3189 //
3190 // Usually, the merge has more than one input. In that case, where inputs
3191 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3192 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3193 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3194 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3195 //
3196 // A merge can take a "wide" memory state as one of its narrow inputs.
3197 // This simply means that the merge observes out only the relevant parts of
3198 // the wide input. That is, wide memory states arriving at narrow merge inputs
3199 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3200 //
3201 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3202 // and that memory slices "leak through":
3203 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3204 //
3205 // But, in such a cascade, repeated memory slices can "block the leak":
3206 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3207 //
3208 // In the last example, Y is not part of the combined memory state of the
3209 // outermost MergeMem. The system must, of course, prevent unschedulable
3210 // memory states from arising, so you can be sure that the state Y is somehow
3211 // a precursor to state Y'.
3212 //
3213 //
3214 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3215 // of each MergeMemNode array are exactly the numerical alias indexes, including
3216 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3217 // Compile::alias_type (and kin) produce and manage these indexes.
3218 //
3219 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3220 // (Note that this provides quick access to the top node inside MergeMem methods,
3221 // without the need to reach out via TLS to Compile::current.)
3222 //
3223 // As a consequence of what was just described, a MergeMem that represents a full
3224 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3225 // containing all alias categories.
3226 //
3227 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3228 //
3229 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3230 // a memory state for the alias type <N>, or else the top node, meaning that
3231 // there is no particular input for that alias type. Note that the length of
3232 // a MergeMem is variable, and may be extended at any time to accommodate new
3233 // memory states at larger alias indexes. When merges grow, they are of course
3234 // filled with "top" in the unused in() positions.
3235 //
3236 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3237 // (Top was chosen because it works smoothly with passes like GCM.)
3238 //
3239 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3240 // the type of random VM bits like TLS references.) Since it is always the
3241 // first non-Bot memory slice, some low-level loops use it to initialize an
3242 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3243 //
3244 //
3245 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3246 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3247 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3248 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3249 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3250 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3251 //
3252 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3253 // really that different from the other memory inputs. An abbreviation called
3254 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3255 //
3256 //
3257 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3258 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3259 // that "emerges though" the base memory will be marked as excluding the alias types
3260 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3261 //
3262 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3263 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3264 //
3265 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3266 // (It is currently unimplemented.) As you can see, the resulting merge is
3267 // actually a disjoint union of memory states, rather than an overlay.
3268 //
3270 //------------------------------MergeMemNode-----------------------------------
3271 Node* MergeMemNode::make_empty_memory() {
3272 Node* empty_memory = (Node*) Compile::current()->top();
3273 assert(empty_memory->is_top(), "correct sentinel identity");
3274 return empty_memory;
3275 }
3277 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3278 init_class_id(Class_MergeMem);
3279 // all inputs are nullified in Node::Node(int)
3280 // set_input(0, NULL); // no control input
3282 // Initialize the edges uniformly to top, for starters.
3283 Node* empty_mem = make_empty_memory();
3284 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3285 init_req(i,empty_mem);
3286 }
3287 assert(empty_memory() == empty_mem, "");
3289 if( new_base != NULL && new_base->is_MergeMem() ) {
3290 MergeMemNode* mdef = new_base->as_MergeMem();
3291 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3292 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3293 mms.set_memory(mms.memory2());
3294 }
3295 assert(base_memory() == mdef->base_memory(), "");
3296 } else {
3297 set_base_memory(new_base);
3298 }
3299 }
3301 // Make a new, untransformed MergeMem with the same base as 'mem'.
3302 // If mem is itself a MergeMem, populate the result with the same edges.
3303 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3304 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3305 }
3307 //------------------------------cmp--------------------------------------------
3308 uint MergeMemNode::hash() const { return NO_HASH; }
3309 uint MergeMemNode::cmp( const Node &n ) const {
3310 return (&n == this); // Always fail except on self
3311 }
3313 //------------------------------Identity---------------------------------------
3314 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3315 // Identity if this merge point does not record any interesting memory
3316 // disambiguations.
3317 Node* base_mem = base_memory();
3318 Node* empty_mem = empty_memory();
3319 if (base_mem != empty_mem) { // Memory path is not dead?
3320 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3321 Node* mem = in(i);
3322 if (mem != empty_mem && mem != base_mem) {
3323 return this; // Many memory splits; no change
3324 }
3325 }
3326 }
3327 return base_mem; // No memory splits; ID on the one true input
3328 }
3330 //------------------------------Ideal------------------------------------------
3331 // This method is invoked recursively on chains of MergeMem nodes
3332 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3333 // Remove chain'd MergeMems
3334 //
3335 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3336 // relative to the "in(Bot)". Since we are patching both at the same time,
3337 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3338 // but rewrite each "in(i)" relative to the new "in(Bot)".
3339 Node *progress = NULL;
3342 Node* old_base = base_memory();
3343 Node* empty_mem = empty_memory();
3344 if (old_base == empty_mem)
3345 return NULL; // Dead memory path.
3347 MergeMemNode* old_mbase;
3348 if (old_base != NULL && old_base->is_MergeMem())
3349 old_mbase = old_base->as_MergeMem();
3350 else
3351 old_mbase = NULL;
3352 Node* new_base = old_base;
3354 // simplify stacked MergeMems in base memory
3355 if (old_mbase) new_base = old_mbase->base_memory();
3357 // the base memory might contribute new slices beyond my req()
3358 if (old_mbase) grow_to_match(old_mbase);
3360 // Look carefully at the base node if it is a phi.
3361 PhiNode* phi_base;
3362 if (new_base != NULL && new_base->is_Phi())
3363 phi_base = new_base->as_Phi();
3364 else
3365 phi_base = NULL;
3367 Node* phi_reg = NULL;
3368 uint phi_len = (uint)-1;
3369 if (phi_base != NULL && !phi_base->is_copy()) {
3370 // do not examine phi if degraded to a copy
3371 phi_reg = phi_base->region();
3372 phi_len = phi_base->req();
3373 // see if the phi is unfinished
3374 for (uint i = 1; i < phi_len; i++) {
3375 if (phi_base->in(i) == NULL) {
3376 // incomplete phi; do not look at it yet!
3377 phi_reg = NULL;
3378 phi_len = (uint)-1;
3379 break;
3380 }
3381 }
3382 }
3384 // Note: We do not call verify_sparse on entry, because inputs
3385 // can normalize to the base_memory via subsume_node or similar
3386 // mechanisms. This method repairs that damage.
3388 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3390 // Look at each slice.
3391 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3392 Node* old_in = in(i);
3393 // calculate the old memory value
3394 Node* old_mem = old_in;
3395 if (old_mem == empty_mem) old_mem = old_base;
3396 assert(old_mem == memory_at(i), "");
3398 // maybe update (reslice) the old memory value
3400 // simplify stacked MergeMems
3401 Node* new_mem = old_mem;
3402 MergeMemNode* old_mmem;
3403 if (old_mem != NULL && old_mem->is_MergeMem())
3404 old_mmem = old_mem->as_MergeMem();
3405 else
3406 old_mmem = NULL;
3407 if (old_mmem == this) {
3408 // This can happen if loops break up and safepoints disappear.
3409 // A merge of BotPtr (default) with a RawPtr memory derived from a
3410 // safepoint can be rewritten to a merge of the same BotPtr with
3411 // the BotPtr phi coming into the loop. If that phi disappears
3412 // also, we can end up with a self-loop of the mergemem.
3413 // In general, if loops degenerate and memory effects disappear,
3414 // a mergemem can be left looking at itself. This simply means
3415 // that the mergemem's default should be used, since there is
3416 // no longer any apparent effect on this slice.
3417 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3418 // from start. Update the input to TOP.
3419 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3420 }
3421 else if (old_mmem != NULL) {
3422 new_mem = old_mmem->memory_at(i);
3423 }
3424 // else preceeding memory was not a MergeMem
3426 // replace equivalent phis (unfortunately, they do not GVN together)
3427 if (new_mem != NULL && new_mem != new_base &&
3428 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3429 if (new_mem->is_Phi()) {
3430 PhiNode* phi_mem = new_mem->as_Phi();
3431 for (uint i = 1; i < phi_len; i++) {
3432 if (phi_base->in(i) != phi_mem->in(i)) {
3433 phi_mem = NULL;
3434 break;
3435 }
3436 }
3437 if (phi_mem != NULL) {
3438 // equivalent phi nodes; revert to the def
3439 new_mem = new_base;
3440 }
3441 }
3442 }
3444 // maybe store down a new value
3445 Node* new_in = new_mem;
3446 if (new_in == new_base) new_in = empty_mem;
3448 if (new_in != old_in) {
3449 // Warning: Do not combine this "if" with the previous "if"
3450 // A memory slice might have be be rewritten even if it is semantically
3451 // unchanged, if the base_memory value has changed.
3452 set_req(i, new_in);
3453 progress = this; // Report progress
3454 }
3455 }
3457 if (new_base != old_base) {
3458 set_req(Compile::AliasIdxBot, new_base);
3459 // Don't use set_base_memory(new_base), because we need to update du.
3460 assert(base_memory() == new_base, "");
3461 progress = this;
3462 }
3464 if( base_memory() == this ) {
3465 // a self cycle indicates this memory path is dead
3466 set_req(Compile::AliasIdxBot, empty_mem);
3467 }
3469 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3470 // Recursion must occur after the self cycle check above
3471 if( base_memory()->is_MergeMem() ) {
3472 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3473 Node *m = phase->transform(new_mbase); // Rollup any cycles
3474 if( m != NULL && (m->is_top() ||
3475 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3476 // propagate rollup of dead cycle to self
3477 set_req(Compile::AliasIdxBot, empty_mem);
3478 }
3479 }
3481 if( base_memory() == empty_mem ) {
3482 progress = this;
3483 // Cut inputs during Parse phase only.
3484 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3485 if( !can_reshape ) {
3486 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3487 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3488 }
3489 }
3490 }
3492 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3493 // Check if PhiNode::Ideal's "Split phis through memory merges"
3494 // transform should be attempted. Look for this->phi->this cycle.
3495 uint merge_width = req();
3496 if (merge_width > Compile::AliasIdxRaw) {
3497 PhiNode* phi = base_memory()->as_Phi();
3498 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3499 if (phi->in(i) == this) {
3500 phase->is_IterGVN()->_worklist.push(phi);
3501 break;
3502 }
3503 }
3504 }
3505 }
3507 assert(progress || verify_sparse(), "please, no dups of base");
3508 return progress;
3509 }
3511 //-------------------------set_base_memory-------------------------------------
3512 void MergeMemNode::set_base_memory(Node *new_base) {
3513 Node* empty_mem = empty_memory();
3514 set_req(Compile::AliasIdxBot, new_base);
3515 assert(memory_at(req()) == new_base, "must set default memory");
3516 // Clear out other occurrences of new_base:
3517 if (new_base != empty_mem) {
3518 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3519 if (in(i) == new_base) set_req(i, empty_mem);
3520 }
3521 }
3522 }
3524 //------------------------------out_RegMask------------------------------------
3525 const RegMask &MergeMemNode::out_RegMask() const {
3526 return RegMask::Empty;
3527 }
3529 //------------------------------dump_spec--------------------------------------
3530 #ifndef PRODUCT
3531 void MergeMemNode::dump_spec(outputStream *st) const {
3532 st->print(" {");
3533 Node* base_mem = base_memory();
3534 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3535 Node* mem = memory_at(i);
3536 if (mem == base_mem) { st->print(" -"); continue; }
3537 st->print( " N%d:", mem->_idx );
3538 Compile::current()->get_adr_type(i)->dump_on(st);
3539 }
3540 st->print(" }");
3541 }
3542 #endif // !PRODUCT
3545 #ifdef ASSERT
3546 static bool might_be_same(Node* a, Node* b) {
3547 if (a == b) return true;
3548 if (!(a->is_Phi() || b->is_Phi())) return false;
3549 // phis shift around during optimization
3550 return true; // pretty stupid...
3551 }
3553 // verify a narrow slice (either incoming or outgoing)
3554 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3555 if (!VerifyAliases) return; // don't bother to verify unless requested
3556 if (is_error_reported()) return; // muzzle asserts when debugging an error
3557 if (Node::in_dump()) return; // muzzle asserts when printing
3558 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3559 assert(n != NULL, "");
3560 // Elide intervening MergeMem's
3561 while (n->is_MergeMem()) {
3562 n = n->as_MergeMem()->memory_at(alias_idx);
3563 }
3564 Compile* C = Compile::current();
3565 const TypePtr* n_adr_type = n->adr_type();
3566 if (n == m->empty_memory()) {
3567 // Implicit copy of base_memory()
3568 } else if (n_adr_type != TypePtr::BOTTOM) {
3569 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3570 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3571 } else {
3572 // A few places like make_runtime_call "know" that VM calls are narrow,
3573 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3574 bool expected_wide_mem = false;
3575 if (n == m->base_memory()) {
3576 expected_wide_mem = true;
3577 } else if (alias_idx == Compile::AliasIdxRaw ||
3578 n == m->memory_at(Compile::AliasIdxRaw)) {
3579 expected_wide_mem = true;
3580 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3581 // memory can "leak through" calls on channels that
3582 // are write-once. Allow this also.
3583 expected_wide_mem = true;
3584 }
3585 assert(expected_wide_mem, "expected narrow slice replacement");
3586 }
3587 }
3588 #else // !ASSERT
3589 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3590 #endif
3593 //-----------------------------memory_at---------------------------------------
3594 Node* MergeMemNode::memory_at(uint alias_idx) const {
3595 assert(alias_idx >= Compile::AliasIdxRaw ||
3596 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3597 "must avoid base_memory and AliasIdxTop");
3599 // Otherwise, it is a narrow slice.
3600 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3601 Compile *C = Compile::current();
3602 if (is_empty_memory(n)) {
3603 // the array is sparse; empty slots are the "top" node
3604 n = base_memory();
3605 assert(Node::in_dump()
3606 || n == NULL || n->bottom_type() == Type::TOP
3607 || n->adr_type() == TypePtr::BOTTOM
3608 || n->adr_type() == TypeRawPtr::BOTTOM
3609 || Compile::current()->AliasLevel() == 0,
3610 "must be a wide memory");
3611 // AliasLevel == 0 if we are organizing the memory states manually.
3612 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3613 } else {
3614 // make sure the stored slice is sane
3615 #ifdef ASSERT
3616 if (is_error_reported() || Node::in_dump()) {
3617 } else if (might_be_same(n, base_memory())) {
3618 // Give it a pass: It is a mostly harmless repetition of the base.
3619 // This can arise normally from node subsumption during optimization.
3620 } else {
3621 verify_memory_slice(this, alias_idx, n);
3622 }
3623 #endif
3624 }
3625 return n;
3626 }
3628 //---------------------------set_memory_at-------------------------------------
3629 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3630 verify_memory_slice(this, alias_idx, n);
3631 Node* empty_mem = empty_memory();
3632 if (n == base_memory()) n = empty_mem; // collapse default
3633 uint need_req = alias_idx+1;
3634 if (req() < need_req) {
3635 if (n == empty_mem) return; // already the default, so do not grow me
3636 // grow the sparse array
3637 do {
3638 add_req(empty_mem);
3639 } while (req() < need_req);
3640 }
3641 set_req( alias_idx, n );
3642 }
3646 //--------------------------iteration_setup------------------------------------
3647 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3648 if (other != NULL) {
3649 grow_to_match(other);
3650 // invariant: the finite support of mm2 is within mm->req()
3651 #ifdef ASSERT
3652 for (uint i = req(); i < other->req(); i++) {
3653 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3654 }
3655 #endif
3656 }
3657 // Replace spurious copies of base_memory by top.
3658 Node* base_mem = base_memory();
3659 if (base_mem != NULL && !base_mem->is_top()) {
3660 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3661 if (in(i) == base_mem)
3662 set_req(i, empty_memory());
3663 }
3664 }
3665 }
3667 //---------------------------grow_to_match-------------------------------------
3668 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3669 Node* empty_mem = empty_memory();
3670 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3671 // look for the finite support of the other memory
3672 for (uint i = other->req(); --i >= req(); ) {
3673 if (other->in(i) != empty_mem) {
3674 uint new_len = i+1;
3675 while (req() < new_len) add_req(empty_mem);
3676 break;
3677 }
3678 }
3679 }
3681 //---------------------------verify_sparse-------------------------------------
3682 #ifndef PRODUCT
3683 bool MergeMemNode::verify_sparse() const {
3684 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3685 Node* base_mem = base_memory();
3686 // The following can happen in degenerate cases, since empty==top.
3687 if (is_empty_memory(base_mem)) return true;
3688 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3689 assert(in(i) != NULL, "sane slice");
3690 if (in(i) == base_mem) return false; // should have been the sentinel value!
3691 }
3692 return true;
3693 }
3695 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3696 Node* n;
3697 n = mm->in(idx);
3698 if (mem == n) return true; // might be empty_memory()
3699 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3700 if (mem == n) return true;
3701 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3702 if (mem == n) return true;
3703 if (n == NULL) break;
3704 }
3705 return false;
3706 }
3707 #endif // !PRODUCT