1 /*
2 * Copyright (c) 1997, 2017, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/compile.hpp"
35 #include "opto/connode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/loopnode.hpp"
38 #include "opto/machnode.hpp"
39 #include "opto/matcher.hpp"
40 #include "opto/memnode.hpp"
41 #include "opto/mulnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/phaseX.hpp"
44 #include "opto/regmask.hpp"
45 #include "utilities/align.hpp"
46 #include "utilities/copy.hpp"
47 #include "utilities/vmError.hpp"
48
49 // Portions of code courtesy of Clifford Click
50
51 // Optimization - Graph Style
52
53 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
54
55 //=============================================================================
56 uint MemNode::size_of() const { return sizeof(*this); }
57
58 const TypePtr *MemNode::adr_type() const {
59 Node* adr = in(Address);
60 if (adr == NULL) return NULL; // node is dead
61 const TypePtr* cross_check = NULL;
62 DEBUG_ONLY(cross_check = _adr_type);
63 return calculate_adr_type(adr->bottom_type(), cross_check);
64 }
65
66 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
67 if (adr != NULL) {
68 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
69 return true;
70 }
71 }
72 return false;
73 }
74
75 #ifndef PRODUCT
76 void MemNode::dump_spec(outputStream *st) const {
77 if (in(Address) == NULL) return; // node is dead
78 #ifndef ASSERT
79 // fake the missing field
80 const TypePtr* _adr_type = NULL;
81 if (in(Address) != NULL)
82 _adr_type = in(Address)->bottom_type()->isa_ptr();
83 #endif
84 dump_adr_type(this, _adr_type, st);
85
86 Compile* C = Compile::current();
87 if (C->alias_type(_adr_type)->is_volatile()) {
88 st->print(" Volatile!");
89 }
90 if (_unaligned_access) {
91 st->print(" unaligned");
92 }
93 if (_mismatched_access) {
94 st->print(" mismatched");
95 }
96 }
97
98 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
99 st->print(" @");
100 if (adr_type == NULL) {
101 st->print("NULL");
102 } else {
103 adr_type->dump_on(st);
104 Compile* C = Compile::current();
105 Compile::AliasType* atp = NULL;
106 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
107 if (atp == NULL)
108 st->print(", idx=?\?;");
109 else if (atp->index() == Compile::AliasIdxBot)
110 st->print(", idx=Bot;");
111 else if (atp->index() == Compile::AliasIdxTop)
112 st->print(", idx=Top;");
113 else if (atp->index() == Compile::AliasIdxRaw)
114 st->print(", idx=Raw;");
115 else {
116 ciField* field = atp->field();
117 if (field) {
118 st->print(", name=");
119 field->print_name_on(st);
120 }
121 st->print(", idx=%d;", atp->index());
122 }
123 }
124 }
125
126 extern void print_alias_types();
127
128 #endif
129
130 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
131 assert((t_oop != NULL), "sanity");
132 bool is_instance = t_oop->is_known_instance_field();
133 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
134 (load != NULL) && load->is_Load() &&
135 (phase->is_IterGVN() != NULL);
136 if (!(is_instance || is_boxed_value_load))
137 return mchain; // don't try to optimize non-instance types
138 uint instance_id = t_oop->instance_id();
139 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
140 Node *prev = NULL;
141 Node *result = mchain;
142 while (prev != result) {
143 prev = result;
144 if (result == start_mem)
145 break; // hit one of our sentinels
146 // skip over a call which does not affect this memory slice
147 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
148 Node *proj_in = result->in(0);
149 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
150 break; // hit one of our sentinels
151 } else if (proj_in->is_Call()) {
152 // ArrayCopyNodes processed here as well
153 CallNode *call = proj_in->as_Call();
154 if (!call->may_modify(t_oop, phase)) { // returns false for instances
155 result = call->in(TypeFunc::Memory);
156 }
157 } else if (proj_in->is_Initialize()) {
158 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
159 // Stop if this is the initialization for the object instance which
160 // contains this memory slice, otherwise skip over it.
161 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
162 break;
163 }
164 if (is_instance) {
165 result = proj_in->in(TypeFunc::Memory);
166 } else if (is_boxed_value_load) {
167 Node* klass = alloc->in(AllocateNode::KlassNode);
168 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
169 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
170 result = proj_in->in(TypeFunc::Memory); // not related allocation
171 }
172 }
173 } else if (proj_in->is_MemBar()) {
174 ArrayCopyNode* ac = NULL;
175 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
176 break;
177 }
178 result = proj_in->in(TypeFunc::Memory);
179 } else {
180 assert(false, "unexpected projection");
181 }
182 } else if (result->is_ClearArray()) {
183 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
184 // Can not bypass initialization of the instance
185 // we are looking for.
186 break;
187 }
188 // Otherwise skip it (the call updated 'result' value).
189 } else if (result->is_MergeMem()) {
190 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
191 }
192 }
193 return result;
194 }
195
196 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
197 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
198 if (t_oop == NULL)
199 return mchain; // don't try to optimize non-oop types
200 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
201 bool is_instance = t_oop->is_known_instance_field();
202 PhaseIterGVN *igvn = phase->is_IterGVN();
203 if (is_instance && igvn != NULL && result->is_Phi()) {
204 PhiNode *mphi = result->as_Phi();
205 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
206 const TypePtr *t = mphi->adr_type();
207 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
208 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
209 t->is_oopptr()->cast_to_exactness(true)
210 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
211 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
212 // clone the Phi with our address type
213 result = mphi->split_out_instance(t_adr, igvn);
214 } else {
215 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
216 }
217 }
218 return result;
219 }
220
221 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
222 uint alias_idx = phase->C->get_alias_index(tp);
223 Node *mem = mmem;
224 #ifdef ASSERT
225 {
226 // Check that current type is consistent with the alias index used during graph construction
227 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
228 bool consistent = adr_check == NULL || adr_check->empty() ||
229 phase->C->must_alias(adr_check, alias_idx );
230 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
231 if( !consistent && adr_check != NULL && !adr_check->empty() &&
232 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
233 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
234 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
235 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
236 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
237 // don't assert if it is dead code.
238 consistent = true;
239 }
240 if( !consistent ) {
241 st->print("alias_idx==%d, adr_check==", alias_idx);
242 if( adr_check == NULL ) {
243 st->print("NULL");
244 } else {
245 adr_check->dump();
246 }
247 st->cr();
248 print_alias_types();
249 assert(consistent, "adr_check must match alias idx");
250 }
251 }
252 #endif
253 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
254 // means an array I have not precisely typed yet. Do not do any
255 // alias stuff with it any time soon.
256 const TypeOopPtr *toop = tp->isa_oopptr();
257 if( tp->base() != Type::AnyPtr &&
258 !(toop &&
259 toop->klass() != NULL &&
260 toop->klass()->is_java_lang_Object() &&
261 toop->offset() == Type::OffsetBot) ) {
262 // compress paths and change unreachable cycles to TOP
263 // If not, we can update the input infinitely along a MergeMem cycle
264 // Equivalent code in PhiNode::Ideal
265 Node* m = phase->transform(mmem);
266 // If transformed to a MergeMem, get the desired slice
267 // Otherwise the returned node represents memory for every slice
268 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
269 // Update input if it is progress over what we have now
270 }
271 return mem;
272 }
273
274 //--------------------------Ideal_common---------------------------------------
275 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
276 // Unhook non-raw memories from complete (macro-expanded) initializations.
277 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
278 // If our control input is a dead region, kill all below the region
279 Node *ctl = in(MemNode::Control);
280 if (ctl && remove_dead_region(phase, can_reshape))
281 return this;
282 ctl = in(MemNode::Control);
283 // Don't bother trying to transform a dead node
284 if (ctl && ctl->is_top()) return NodeSentinel;
285
286 PhaseIterGVN *igvn = phase->is_IterGVN();
287 // Wait if control on the worklist.
288 if (ctl && can_reshape && igvn != NULL) {
289 Node* bol = NULL;
290 Node* cmp = NULL;
291 if (ctl->in(0)->is_If()) {
292 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
293 bol = ctl->in(0)->in(1);
294 if (bol->is_Bool())
295 cmp = ctl->in(0)->in(1)->in(1);
296 }
297 if (igvn->_worklist.member(ctl) ||
298 (bol != NULL && igvn->_worklist.member(bol)) ||
299 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
300 // This control path may be dead.
301 // Delay this memory node transformation until the control is processed.
302 phase->is_IterGVN()->_worklist.push(this);
303 return NodeSentinel; // caller will return NULL
304 }
305 }
306 // Ignore if memory is dead, or self-loop
307 Node *mem = in(MemNode::Memory);
308 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
309 assert(mem != this, "dead loop in MemNode::Ideal");
310
311 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
312 // This memory slice may be dead.
313 // Delay this mem node transformation until the memory is processed.
314 phase->is_IterGVN()->_worklist.push(this);
315 return NodeSentinel; // caller will return NULL
316 }
317
318 Node *address = in(MemNode::Address);
319 const Type *t_adr = phase->type(address);
320 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
321
322 if (can_reshape && igvn != NULL &&
323 (igvn->_worklist.member(address) ||
324 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
325 // The address's base and type may change when the address is processed.
326 // Delay this mem node transformation until the address is processed.
327 phase->is_IterGVN()->_worklist.push(this);
328 return NodeSentinel; // caller will return NULL
329 }
330
331 // Do NOT remove or optimize the next lines: ensure a new alias index
332 // is allocated for an oop pointer type before Escape Analysis.
333 // Note: C++ will not remove it since the call has side effect.
334 if (t_adr->isa_oopptr()) {
335 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
336 }
337
338 Node* base = NULL;
339 if (address->is_AddP()) {
340 base = address->in(AddPNode::Base);
341 }
342 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
343 !t_adr->isa_rawptr()) {
344 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
345 // Skip this node optimization if its address has TOP base.
346 return NodeSentinel; // caller will return NULL
347 }
348
349 // Avoid independent memory operations
350 Node* old_mem = mem;
351
352 // The code which unhooks non-raw memories from complete (macro-expanded)
353 // initializations was removed. After macro-expansion all stores catched
354 // by Initialize node became raw stores and there is no information
355 // which memory slices they modify. So it is unsafe to move any memory
356 // operation above these stores. Also in most cases hooked non-raw memories
357 // were already unhooked by using information from detect_ptr_independence()
358 // and find_previous_store().
359
360 if (mem->is_MergeMem()) {
361 MergeMemNode* mmem = mem->as_MergeMem();
362 const TypePtr *tp = t_adr->is_ptr();
363
364 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
365 }
366
367 if (mem != old_mem) {
368 set_req(MemNode::Memory, mem);
369 if (can_reshape && old_mem->outcnt() == 0) {
370 igvn->_worklist.push(old_mem);
371 }
372 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
373 return this;
374 }
375
376 // let the subclass continue analyzing...
377 return NULL;
378 }
379
380 // Helper function for proving some simple control dominations.
381 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
382 // Already assumes that 'dom' is available at 'sub', and that 'sub'
383 // is not a constant (dominated by the method's StartNode).
384 // Used by MemNode::find_previous_store to prove that the
385 // control input of a memory operation predates (dominates)
386 // an allocation it wants to look past.
387 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
388 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
389 return false; // Conservative answer for dead code
390
391 // Check 'dom'. Skip Proj and CatchProj nodes.
392 dom = dom->find_exact_control(dom);
393 if (dom == NULL || dom->is_top())
394 return false; // Conservative answer for dead code
395
396 if (dom == sub) {
397 // For the case when, for example, 'sub' is Initialize and the original
398 // 'dom' is Proj node of the 'sub'.
399 return false;
400 }
401
402 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
403 return true;
404
405 // 'dom' dominates 'sub' if its control edge and control edges
406 // of all its inputs dominate or equal to sub's control edge.
407
408 // Currently 'sub' is either Allocate, Initialize or Start nodes.
409 // Or Region for the check in LoadNode::Ideal();
410 // 'sub' should have sub->in(0) != NULL.
411 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
412 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
413
414 // Get control edge of 'sub'.
415 Node* orig_sub = sub;
416 sub = sub->find_exact_control(sub->in(0));
417 if (sub == NULL || sub->is_top())
418 return false; // Conservative answer for dead code
419
420 assert(sub->is_CFG(), "expecting control");
421
422 if (sub == dom)
423 return true;
424
425 if (sub->is_Start() || sub->is_Root())
426 return false;
427
428 {
429 // Check all control edges of 'dom'.
430
431 ResourceMark rm;
432 Arena* arena = Thread::current()->resource_area();
433 Node_List nlist(arena);
434 Unique_Node_List dom_list(arena);
435
436 dom_list.push(dom);
437 bool only_dominating_controls = false;
438
439 for (uint next = 0; next < dom_list.size(); next++) {
440 Node* n = dom_list.at(next);
441 if (n == orig_sub)
442 return false; // One of dom's inputs dominated by sub.
443 if (!n->is_CFG() && n->pinned()) {
444 // Check only own control edge for pinned non-control nodes.
445 n = n->find_exact_control(n->in(0));
446 if (n == NULL || n->is_top())
447 return false; // Conservative answer for dead code
448 assert(n->is_CFG(), "expecting control");
449 dom_list.push(n);
450 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
451 only_dominating_controls = true;
452 } else if (n->is_CFG()) {
453 if (n->dominates(sub, nlist))
454 only_dominating_controls = true;
455 else
456 return false;
457 } else {
458 // First, own control edge.
459 Node* m = n->find_exact_control(n->in(0));
460 if (m != NULL) {
461 if (m->is_top())
462 return false; // Conservative answer for dead code
463 dom_list.push(m);
464 }
465 // Now, the rest of edges.
466 uint cnt = n->req();
467 for (uint i = 1; i < cnt; i++) {
468 m = n->find_exact_control(n->in(i));
469 if (m == NULL || m->is_top())
470 continue;
471 dom_list.push(m);
472 }
473 }
474 }
475 return only_dominating_controls;
476 }
477 }
478
479 //---------------------detect_ptr_independence---------------------------------
480 // Used by MemNode::find_previous_store to prove that two base
481 // pointers are never equal.
482 // The pointers are accompanied by their associated allocations,
483 // if any, which have been previously discovered by the caller.
484 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
485 Node* p2, AllocateNode* a2,
486 PhaseTransform* phase) {
487 // Attempt to prove that these two pointers cannot be aliased.
488 // They may both manifestly be allocations, and they should differ.
489 // Or, if they are not both allocations, they can be distinct constants.
490 // Otherwise, one is an allocation and the other a pre-existing value.
491 if (a1 == NULL && a2 == NULL) { // neither an allocation
492 return (p1 != p2) && p1->is_Con() && p2->is_Con();
493 } else if (a1 != NULL && a2 != NULL) { // both allocations
494 return (a1 != a2);
495 } else if (a1 != NULL) { // one allocation a1
496 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
497 return all_controls_dominate(p2, a1);
498 } else { //(a2 != NULL) // one allocation a2
499 return all_controls_dominate(p1, a2);
500 }
501 return false;
502 }
503
504
505 // Find an arraycopy that must have set (can_see_stored_value=true) or
506 // could have set (can_see_stored_value=false) the value for this load
507 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
508 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
509 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
510 Node* mb = mem->in(0);
511 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
512 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
513 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
514 if (ac->is_clonebasic()) {
515 intptr_t offset;
516 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
517 assert(alloc != NULL && (!ReduceBulkZeroing || alloc->initialization()->is_complete_with_arraycopy()), "broken allocation");
518 if (alloc == ld_alloc) {
519 return ac;
520 }
521 }
522 }
523 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
524 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
525
526 if (ac->is_arraycopy_validated() ||
527 ac->is_copyof_validated() ||
528 ac->is_copyofrange_validated()) {
529 Node* ld_addp = in(MemNode::Address);
530 if (ld_addp->is_AddP()) {
531 Node* ld_base = ld_addp->in(AddPNode::Address);
532 Node* ld_offs = ld_addp->in(AddPNode::Offset);
533
534 Node* dest = ac->in(ArrayCopyNode::Dest);
535
536 if (dest == ld_base) {
537 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
538 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
539 return ac;
540 }
541 if (!can_see_stored_value) {
542 mem = ac->in(TypeFunc::Memory);
543 }
544 }
545 }
546 }
547 }
548 return NULL;
549 }
550
551 // The logic for reordering loads and stores uses four steps:
552 // (a) Walk carefully past stores and initializations which we
553 // can prove are independent of this load.
554 // (b) Observe that the next memory state makes an exact match
555 // with self (load or store), and locate the relevant store.
556 // (c) Ensure that, if we were to wire self directly to the store,
557 // the optimizer would fold it up somehow.
558 // (d) Do the rewiring, and return, depending on some other part of
559 // the optimizer to fold up the load.
560 // This routine handles steps (a) and (b). Steps (c) and (d) are
561 // specific to loads and stores, so they are handled by the callers.
562 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
563 //
564 Node* MemNode::find_previous_store(PhaseTransform* phase) {
565 Node* ctrl = in(MemNode::Control);
566 Node* adr = in(MemNode::Address);
567 intptr_t offset = 0;
568 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
569 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
570
571 if (offset == Type::OffsetBot)
572 return NULL; // cannot unalias unless there are precise offsets
573
574 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
575 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
576
577 intptr_t size_in_bytes = memory_size();
578
579 Node* mem = in(MemNode::Memory); // start searching here...
580
581 int cnt = 50; // Cycle limiter
582 for (;;) { // While we can dance past unrelated stores...
583 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
584
585 Node* prev = mem;
586 if (mem->is_Store()) {
587 Node* st_adr = mem->in(MemNode::Address);
588 intptr_t st_offset = 0;
589 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
590 if (st_base == NULL)
591 break; // inscrutable pointer
592
593 // For raw accesses it's not enough to prove that constant offsets don't intersect.
594 // We need the bases to be the equal in order for the offset check to make sense.
595 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
596 break;
597 }
598
599 if (st_offset != offset && st_offset != Type::OffsetBot) {
600 const int MAX_STORE = BytesPerLong;
601 if (st_offset >= offset + size_in_bytes ||
602 st_offset <= offset - MAX_STORE ||
603 st_offset <= offset - mem->as_Store()->memory_size()) {
604 // Success: The offsets are provably independent.
605 // (You may ask, why not just test st_offset != offset and be done?
606 // The answer is that stores of different sizes can co-exist
607 // in the same sequence of RawMem effects. We sometimes initialize
608 // a whole 'tile' of array elements with a single jint or jlong.)
609 mem = mem->in(MemNode::Memory);
610 continue; // (a) advance through independent store memory
611 }
612 }
613 if (st_base != base &&
614 detect_ptr_independence(base, alloc,
615 st_base,
616 AllocateNode::Ideal_allocation(st_base, phase),
617 phase)) {
618 // Success: The bases are provably independent.
619 mem = mem->in(MemNode::Memory);
620 continue; // (a) advance through independent store memory
621 }
622
623 // (b) At this point, if the bases or offsets do not agree, we lose,
624 // since we have not managed to prove 'this' and 'mem' independent.
625 if (st_base == base && st_offset == offset) {
626 return mem; // let caller handle steps (c), (d)
627 }
628
629 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
630 InitializeNode* st_init = mem->in(0)->as_Initialize();
631 AllocateNode* st_alloc = st_init->allocation();
632 if (st_alloc == NULL)
633 break; // something degenerated
634 bool known_identical = false;
635 bool known_independent = false;
636 if (alloc == st_alloc)
637 known_identical = true;
638 else if (alloc != NULL)
639 known_independent = true;
640 else if (all_controls_dominate(this, st_alloc))
641 known_independent = true;
642
643 if (known_independent) {
644 // The bases are provably independent: Either they are
645 // manifestly distinct allocations, or else the control
646 // of this load dominates the store's allocation.
647 int alias_idx = phase->C->get_alias_index(adr_type());
648 if (alias_idx == Compile::AliasIdxRaw) {
649 mem = st_alloc->in(TypeFunc::Memory);
650 } else {
651 mem = st_init->memory(alias_idx);
652 }
653 continue; // (a) advance through independent store memory
654 }
655
656 // (b) at this point, if we are not looking at a store initializing
657 // the same allocation we are loading from, we lose.
658 if (known_identical) {
659 // From caller, can_see_stored_value will consult find_captured_store.
660 return mem; // let caller handle steps (c), (d)
661 }
662
663 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
664 if (prev != mem) {
665 // Found an arraycopy but it doesn't affect that load
666 continue;
667 }
668 // Found an arraycopy that may affect that load
669 return mem;
670 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
671 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
672 if (mem->is_Proj() && mem->in(0)->is_Call()) {
673 // ArrayCopyNodes processed here as well.
674 CallNode *call = mem->in(0)->as_Call();
675 if (!call->may_modify(addr_t, phase)) {
676 mem = call->in(TypeFunc::Memory);
677 continue; // (a) advance through independent call memory
678 }
679 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
680 ArrayCopyNode* ac = NULL;
681 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
682 break;
683 }
684 mem = mem->in(0)->in(TypeFunc::Memory);
685 continue; // (a) advance through independent MemBar memory
686 } else if (mem->is_ClearArray()) {
687 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
688 // (the call updated 'mem' value)
689 continue; // (a) advance through independent allocation memory
690 } else {
691 // Can not bypass initialization of the instance
692 // we are looking for.
693 return mem;
694 }
695 } else if (mem->is_MergeMem()) {
696 int alias_idx = phase->C->get_alias_index(adr_type());
697 mem = mem->as_MergeMem()->memory_at(alias_idx);
698 continue; // (a) advance through independent MergeMem memory
699 }
700 }
701
702 // Unless there is an explicit 'continue', we must bail out here,
703 // because 'mem' is an inscrutable memory state (e.g., a call).
704 break;
705 }
706
707 return NULL; // bail out
708 }
709
710 //----------------------calculate_adr_type-------------------------------------
711 // Helper function. Notices when the given type of address hits top or bottom.
712 // Also, asserts a cross-check of the type against the expected address type.
713 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
714 if (t == Type::TOP) return NULL; // does not touch memory any more?
715 #ifdef PRODUCT
716 cross_check = NULL;
717 #else
718 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
719 #endif
720 const TypePtr* tp = t->isa_ptr();
721 if (tp == NULL) {
722 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
723 return TypePtr::BOTTOM; // touches lots of memory
724 } else {
725 #ifdef ASSERT
726 // %%%% [phh] We don't check the alias index if cross_check is
727 // TypeRawPtr::BOTTOM. Needs to be investigated.
728 if (cross_check != NULL &&
729 cross_check != TypePtr::BOTTOM &&
730 cross_check != TypeRawPtr::BOTTOM) {
731 // Recheck the alias index, to see if it has changed (due to a bug).
732 Compile* C = Compile::current();
733 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
734 "must stay in the original alias category");
735 // The type of the address must be contained in the adr_type,
736 // disregarding "null"-ness.
737 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
738 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
739 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
740 "real address must not escape from expected memory type");
741 }
742 #endif
743 return tp;
744 }
745 }
746
747 //=============================================================================
748 // Should LoadNode::Ideal() attempt to remove control edges?
749 bool LoadNode::can_remove_control() const {
750 return true;
751 }
752 uint LoadNode::size_of() const { return sizeof(*this); }
753 uint LoadNode::cmp( const Node &n ) const
754 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
755 const Type *LoadNode::bottom_type() const { return _type; }
756 uint LoadNode::ideal_reg() const {
757 return _type->ideal_reg();
758 }
759
760 #ifndef PRODUCT
761 void LoadNode::dump_spec(outputStream *st) const {
762 MemNode::dump_spec(st);
763 if( !Verbose && !WizardMode ) {
764 // standard dump does this in Verbose and WizardMode
765 st->print(" #"); _type->dump_on(st);
766 }
767 if (!depends_only_on_test()) {
768 st->print(" (does not depend only on test)");
769 }
770 }
771 #endif
772
773 #ifdef ASSERT
774 //----------------------------is_immutable_value-------------------------------
775 // Helper function to allow a raw load without control edge for some cases
776 bool LoadNode::is_immutable_value(Node* adr) {
777 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
778 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
779 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
780 in_bytes(JavaThread::osthread_offset())));
781 }
782 #endif
783
784 //----------------------------LoadNode::make-----------------------------------
785 // Polymorphic factory method:
786 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
787 ControlDependency control_dependency, bool unaligned, bool mismatched) {
788 Compile* C = gvn.C;
789
790 // sanity check the alias category against the created node type
791 assert(!(adr_type->isa_oopptr() &&
792 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
793 "use LoadKlassNode instead");
794 assert(!(adr_type->isa_aryptr() &&
795 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
796 "use LoadRangeNode instead");
797 // Check control edge of raw loads
798 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
799 // oop will be recorded in oop map if load crosses safepoint
800 rt->isa_oopptr() || is_immutable_value(adr),
801 "raw memory operations should have control edge");
802 LoadNode* load = NULL;
803 switch (bt) {
804 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
805 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
806 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
807 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
808 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
809 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
810 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
811 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
812 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
813 case T_OBJECT:
814 #ifdef _LP64
815 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
816 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
817 } else
818 #endif
819 {
820 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
821 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
822 }
823 break;
824 }
825 assert(load != NULL, "LoadNode should have been created");
826 if (unaligned) {
827 load->set_unaligned_access();
828 }
829 if (mismatched) {
830 load->set_mismatched_access();
831 }
832 if (load->Opcode() == Op_LoadN) {
833 Node* ld = gvn.transform(load);
834 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
835 }
836
837 return load;
838 }
839
840 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
841 ControlDependency control_dependency, bool unaligned, bool mismatched) {
842 bool require_atomic = true;
843 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
844 if (unaligned) {
845 load->set_unaligned_access();
846 }
847 if (mismatched) {
848 load->set_mismatched_access();
849 }
850 return load;
851 }
852
853 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
854 ControlDependency control_dependency, bool unaligned, bool mismatched) {
855 bool require_atomic = true;
856 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
857 if (unaligned) {
858 load->set_unaligned_access();
859 }
860 if (mismatched) {
861 load->set_mismatched_access();
862 }
863 return load;
864 }
865
866
867
868 //------------------------------hash-------------------------------------------
869 uint LoadNode::hash() const {
870 // unroll addition of interesting fields
871 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
872 }
873
874 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
875 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
876 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
877 bool is_stable_ary = FoldStableValues &&
878 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
879 tp->isa_aryptr()->is_stable();
880
881 return (eliminate_boxing && non_volatile) || is_stable_ary;
882 }
883
884 return false;
885 }
886
887 // Is the value loaded previously stored by an arraycopy? If so return
888 // a load node that reads from the source array so we may be able to
889 // optimize out the ArrayCopy node later.
890 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseTransform* phase) const {
891 Node* ld_adr = in(MemNode::Address);
892 intptr_t ld_off = 0;
893 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
894 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
895 if (ac != NULL) {
896 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
897
898 Node* ld = clone();
899 if (ac->as_ArrayCopy()->is_clonebasic()) {
900 assert(ld_alloc != NULL, "need an alloc");
901 Node* addp = in(MemNode::Address)->clone();
902 assert(addp->is_AddP(), "address must be addp");
903 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern");
904 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern");
905 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)->in(AddPNode::Base));
906 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)->in(AddPNode::Address));
907 ld->set_req(MemNode::Address, phase->transform(addp));
908 if (in(0) != NULL) {
909 assert(ld_alloc->in(0) != NULL, "alloc must have control");
910 ld->set_req(0, ld_alloc->in(0));
911 }
912 } else {
913 Node* src = ac->in(ArrayCopyNode::Src);
914 Node* addp = in(MemNode::Address)->clone();
915 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
916 addp->set_req(AddPNode::Base, src);
917 addp->set_req(AddPNode::Address, src);
918
919 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
920 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
921 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
922 uint shift = exact_log2(type2aelembytes(ary_elem));
923
924 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
925 #ifdef _LP64
926 diff = phase->transform(new ConvI2LNode(diff));
927 #endif
928 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
929
930 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
931 addp->set_req(AddPNode::Offset, offset);
932 ld->set_req(MemNode::Address, phase->transform(addp));
933
934 const TypeX *ld_offs_t = phase->type(offset)->isa_intptr_t();
935
936 if (!ac->as_ArrayCopy()->can_replace_dest_load_with_src_load(ld_offs_t->_lo, ld_offs_t->_hi, phase)) {
937 return NULL;
938 }
939
940 if (in(0) != NULL) {
941 assert(ac->in(0) != NULL, "alloc must have control");
942 ld->set_req(0, ac->in(0));
943 }
944 }
945 // load depends on the tests that validate the arraycopy
946 ld->as_Load()->_control_dependency = Pinned;
947 return ld;
948 }
949 return NULL;
950 }
951
952
953 //---------------------------can_see_stored_value------------------------------
954 // This routine exists to make sure this set of tests is done the same
955 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
956 // will change the graph shape in a way which makes memory alive twice at the
957 // same time (uses the Oracle model of aliasing), then some
958 // LoadXNode::Identity will fold things back to the equivalence-class model
959 // of aliasing.
960 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
961 Node* ld_adr = in(MemNode::Address);
962 intptr_t ld_off = 0;
963 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
964 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
965 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
966 // This is more general than load from boxing objects.
967 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
968 uint alias_idx = atp->index();
969 bool final = !atp->is_rewritable();
970 Node* result = NULL;
971 Node* current = st;
972 // Skip through chains of MemBarNodes checking the MergeMems for
973 // new states for the slice of this load. Stop once any other
974 // kind of node is encountered. Loads from final memory can skip
975 // through any kind of MemBar but normal loads shouldn't skip
976 // through MemBarAcquire since the could allow them to move out of
977 // a synchronized region.
978 while (current->is_Proj()) {
979 int opc = current->in(0)->Opcode();
980 if ((final && (opc == Op_MemBarAcquire ||
981 opc == Op_MemBarAcquireLock ||
982 opc == Op_LoadFence)) ||
983 opc == Op_MemBarRelease ||
984 opc == Op_StoreFence ||
985 opc == Op_MemBarReleaseLock ||
986 opc == Op_MemBarStoreStore ||
987 opc == Op_MemBarCPUOrder) {
988 Node* mem = current->in(0)->in(TypeFunc::Memory);
989 if (mem->is_MergeMem()) {
990 MergeMemNode* merge = mem->as_MergeMem();
991 Node* new_st = merge->memory_at(alias_idx);
992 if (new_st == merge->base_memory()) {
993 // Keep searching
994 current = new_st;
995 continue;
996 }
997 // Save the new memory state for the slice and fall through
998 // to exit.
999 result = new_st;
1000 }
1001 }
1002 break;
1003 }
1004 if (result != NULL) {
1005 st = result;
1006 }
1007 }
1008
1009 // Loop around twice in the case Load -> Initialize -> Store.
1010 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1011 for (int trip = 0; trip <= 1; trip++) {
1012
1013 if (st->is_Store()) {
1014 Node* st_adr = st->in(MemNode::Address);
1015 if (!phase->eqv(st_adr, ld_adr)) {
1016 // Try harder before giving up... Match raw and non-raw pointers.
1017 intptr_t st_off = 0;
1018 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1019 if (alloc == NULL) return NULL;
1020 if (alloc != ld_alloc) return NULL;
1021 if (ld_off != st_off) return NULL;
1022 // At this point we have proven something like this setup:
1023 // A = Allocate(...)
1024 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
1025 // S = StoreQ(, AddP(, A.Parm , #Off), V)
1026 // (Actually, we haven't yet proven the Q's are the same.)
1027 // In other words, we are loading from a casted version of
1028 // the same pointer-and-offset that we stored to.
1029 // Thus, we are able to replace L by V.
1030 }
1031 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1032 if (store_Opcode() != st->Opcode())
1033 return NULL;
1034 return st->in(MemNode::ValueIn);
1035 }
1036
1037 // A load from a freshly-created object always returns zero.
1038 // (This can happen after LoadNode::Ideal resets the load's memory input
1039 // to find_captured_store, which returned InitializeNode::zero_memory.)
1040 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1041 (st->in(0) == ld_alloc) &&
1042 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1043 // return a zero value for the load's basic type
1044 // (This is one of the few places where a generic PhaseTransform
1045 // can create new nodes. Think of it as lazily manifesting
1046 // virtually pre-existing constants.)
1047 return phase->zerocon(memory_type());
1048 }
1049
1050 // A load from an initialization barrier can match a captured store.
1051 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1052 InitializeNode* init = st->in(0)->as_Initialize();
1053 AllocateNode* alloc = init->allocation();
1054 if ((alloc != NULL) && (alloc == ld_alloc)) {
1055 // examine a captured store value
1056 st = init->find_captured_store(ld_off, memory_size(), phase);
1057 if (st != NULL) {
1058 continue; // take one more trip around
1059 }
1060 }
1061 }
1062
1063 // Load boxed value from result of valueOf() call is input parameter.
1064 if (this->is_Load() && ld_adr->is_AddP() &&
1065 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1066 intptr_t ignore = 0;
1067 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1068 if (base != NULL && base->is_Proj() &&
1069 base->as_Proj()->_con == TypeFunc::Parms &&
1070 base->in(0)->is_CallStaticJava() &&
1071 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1072 return base->in(0)->in(TypeFunc::Parms);
1073 }
1074 }
1075
1076 break;
1077 }
1078
1079 return NULL;
1080 }
1081
1082 //----------------------is_instance_field_load_with_local_phi------------------
1083 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1084 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1085 in(Address)->is_AddP() ) {
1086 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1087 // Only instances and boxed values.
1088 if( t_oop != NULL &&
1089 (t_oop->is_ptr_to_boxed_value() ||
1090 t_oop->is_known_instance_field()) &&
1091 t_oop->offset() != Type::OffsetBot &&
1092 t_oop->offset() != Type::OffsetTop) {
1093 return true;
1094 }
1095 }
1096 return false;
1097 }
1098
1099 //------------------------------Identity---------------------------------------
1100 // Loads are identity if previous store is to same address
1101 Node* LoadNode::Identity(PhaseGVN* phase) {
1102 // If the previous store-maker is the right kind of Store, and the store is
1103 // to the same address, then we are equal to the value stored.
1104 Node* mem = in(Memory);
1105 Node* value = can_see_stored_value(mem, phase);
1106 if( value ) {
1107 // byte, short & char stores truncate naturally.
1108 // A load has to load the truncated value which requires
1109 // some sort of masking operation and that requires an
1110 // Ideal call instead of an Identity call.
1111 if (memory_size() < BytesPerInt) {
1112 // If the input to the store does not fit with the load's result type,
1113 // it must be truncated via an Ideal call.
1114 if (!phase->type(value)->higher_equal(phase->type(this)))
1115 return this;
1116 }
1117 // (This works even when value is a Con, but LoadNode::Value
1118 // usually runs first, producing the singleton type of the Con.)
1119 return value;
1120 }
1121
1122 // Search for an existing data phi which was generated before for the same
1123 // instance's field to avoid infinite generation of phis in a loop.
1124 Node *region = mem->in(0);
1125 if (is_instance_field_load_with_local_phi(region)) {
1126 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1127 int this_index = phase->C->get_alias_index(addr_t);
1128 int this_offset = addr_t->offset();
1129 int this_iid = addr_t->instance_id();
1130 if (!addr_t->is_known_instance() &&
1131 addr_t->is_ptr_to_boxed_value()) {
1132 // Use _idx of address base (could be Phi node) for boxed values.
1133 intptr_t ignore = 0;
1134 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1135 if (base == NULL) {
1136 return this;
1137 }
1138 this_iid = base->_idx;
1139 }
1140 const Type* this_type = bottom_type();
1141 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1142 Node* phi = region->fast_out(i);
1143 if (phi->is_Phi() && phi != mem &&
1144 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1145 return phi;
1146 }
1147 }
1148 }
1149
1150 return this;
1151 }
1152
1153 // Construct an equivalent unsigned load.
1154 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1155 BasicType bt = T_ILLEGAL;
1156 const Type* rt = NULL;
1157 switch (Opcode()) {
1158 case Op_LoadUB: return this;
1159 case Op_LoadUS: return this;
1160 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1161 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1162 default:
1163 assert(false, "no unsigned variant: %s", Name());
1164 return NULL;
1165 }
1166 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1167 raw_adr_type(), rt, bt, _mo, _control_dependency,
1168 is_unaligned_access(), is_mismatched_access());
1169 }
1170
1171 // Construct an equivalent signed load.
1172 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1173 BasicType bt = T_ILLEGAL;
1174 const Type* rt = NULL;
1175 switch (Opcode()) {
1176 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1177 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1178 case Op_LoadB: // fall through
1179 case Op_LoadS: // fall through
1180 case Op_LoadI: // fall through
1181 case Op_LoadL: return this;
1182 default:
1183 assert(false, "no signed variant: %s", Name());
1184 return NULL;
1185 }
1186 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1187 raw_adr_type(), rt, bt, _mo, _control_dependency,
1188 is_unaligned_access(), is_mismatched_access());
1189 }
1190
1191 // We're loading from an object which has autobox behaviour.
1192 // If this object is result of a valueOf call we'll have a phi
1193 // merging a newly allocated object and a load from the cache.
1194 // We want to replace this load with the original incoming
1195 // argument to the valueOf call.
1196 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1197 assert(phase->C->eliminate_boxing(), "sanity");
1198 intptr_t ignore = 0;
1199 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1200 if ((base == NULL) || base->is_Phi()) {
1201 // Push the loads from the phi that comes from valueOf up
1202 // through it to allow elimination of the loads and the recovery
1203 // of the original value. It is done in split_through_phi().
1204 return NULL;
1205 } else if (base->is_Load() ||
1206 base->is_DecodeN() && base->in(1)->is_Load()) {
1207 // Eliminate the load of boxed value for integer types from the cache
1208 // array by deriving the value from the index into the array.
1209 // Capture the offset of the load and then reverse the computation.
1210
1211 // Get LoadN node which loads a boxing object from 'cache' array.
1212 if (base->is_DecodeN()) {
1213 base = base->in(1);
1214 }
1215 if (!base->in(Address)->is_AddP()) {
1216 return NULL; // Complex address
1217 }
1218 AddPNode* address = base->in(Address)->as_AddP();
1219 Node* cache_base = address->in(AddPNode::Base);
1220 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1221 // Get ConP node which is static 'cache' field.
1222 cache_base = cache_base->in(1);
1223 }
1224 if ((cache_base != NULL) && cache_base->is_Con()) {
1225 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1226 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1227 Node* elements[4];
1228 int shift = exact_log2(type2aelembytes(T_OBJECT));
1229 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1230 if ((count > 0) && elements[0]->is_Con() &&
1231 ((count == 1) ||
1232 (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1233 elements[1]->in(2) == phase->intcon(shift))) {
1234 ciObjArray* array = base_type->const_oop()->as_obj_array();
1235 // Fetch the box object cache[0] at the base of the array and get its value
1236 ciInstance* box = array->obj_at(0)->as_instance();
1237 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1238 assert(ik->is_box_klass(), "sanity");
1239 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1240 if (ik->nof_nonstatic_fields() == 1) {
1241 // This should be true nonstatic_field_at requires calling
1242 // nof_nonstatic_fields so check it anyway
1243 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1244 BasicType bt = c.basic_type();
1245 // Only integer types have boxing cache.
1246 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1247 bt == T_BYTE || bt == T_SHORT ||
1248 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1249 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1250 if (cache_low != (int)cache_low) {
1251 return NULL; // should not happen since cache is array indexed by value
1252 }
1253 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1254 if (offset != (int)offset) {
1255 return NULL; // should not happen since cache is array indexed by value
1256 }
1257 // Add up all the offsets making of the address of the load
1258 Node* result = elements[0];
1259 for (int i = 1; i < count; i++) {
1260 result = phase->transform(new AddXNode(result, elements[i]));
1261 }
1262 // Remove the constant offset from the address and then
1263 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1264 // remove the scaling of the offset to recover the original index.
1265 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1266 // Peel the shift off directly but wrap it in a dummy node
1267 // since Ideal can't return existing nodes
1268 result = new RShiftXNode(result->in(1), phase->intcon(0));
1269 } else if (result->is_Add() && result->in(2)->is_Con() &&
1270 result->in(1)->Opcode() == Op_LShiftX &&
1271 result->in(1)->in(2) == phase->intcon(shift)) {
1272 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1273 // but for boxing cache access we know that X<<Z will not overflow
1274 // (there is range check) so we do this optimizatrion by hand here.
1275 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1276 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1277 } else {
1278 result = new RShiftXNode(result, phase->intcon(shift));
1279 }
1280 #ifdef _LP64
1281 if (bt != T_LONG) {
1282 result = new ConvL2INode(phase->transform(result));
1283 }
1284 #else
1285 if (bt == T_LONG) {
1286 result = new ConvI2LNode(phase->transform(result));
1287 }
1288 #endif
1289 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1290 // Need to preserve unboxing load type if it is unsigned.
1291 switch(this->Opcode()) {
1292 case Op_LoadUB:
1293 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1294 break;
1295 case Op_LoadUS:
1296 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1297 break;
1298 }
1299 return result;
1300 }
1301 }
1302 }
1303 }
1304 }
1305 return NULL;
1306 }
1307
1308 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1309 Node* region = phi->in(0);
1310 if (region == NULL) {
1311 return false; // Wait stable graph
1312 }
1313 uint cnt = phi->req();
1314 for (uint i = 1; i < cnt; i++) {
1315 Node* rc = region->in(i);
1316 if (rc == NULL || phase->type(rc) == Type::TOP)
1317 return false; // Wait stable graph
1318 Node* in = phi->in(i);
1319 if (in == NULL || phase->type(in) == Type::TOP)
1320 return false; // Wait stable graph
1321 }
1322 return true;
1323 }
1324 //------------------------------split_through_phi------------------------------
1325 // Split instance or boxed field load through Phi.
1326 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1327 Node* mem = in(Memory);
1328 Node* address = in(Address);
1329 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1330
1331 assert((t_oop != NULL) &&
1332 (t_oop->is_known_instance_field() ||
1333 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1334
1335 Compile* C = phase->C;
1336 intptr_t ignore = 0;
1337 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1338 bool base_is_phi = (base != NULL) && base->is_Phi();
1339 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1340 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1341 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1342
1343 if (!((mem->is_Phi() || base_is_phi) &&
1344 (load_boxed_values || t_oop->is_known_instance_field()))) {
1345 return NULL; // memory is not Phi
1346 }
1347
1348 if (mem->is_Phi()) {
1349 if (!stable_phi(mem->as_Phi(), phase)) {
1350 return NULL; // Wait stable graph
1351 }
1352 uint cnt = mem->req();
1353 // Check for loop invariant memory.
1354 if (cnt == 3) {
1355 for (uint i = 1; i < cnt; i++) {
1356 Node* in = mem->in(i);
1357 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1358 if (m == mem) {
1359 set_req(Memory, mem->in(cnt - i));
1360 return this; // made change
1361 }
1362 }
1363 }
1364 }
1365 if (base_is_phi) {
1366 if (!stable_phi(base->as_Phi(), phase)) {
1367 return NULL; // Wait stable graph
1368 }
1369 uint cnt = base->req();
1370 // Check for loop invariant memory.
1371 if (cnt == 3) {
1372 for (uint i = 1; i < cnt; i++) {
1373 if (base->in(i) == base) {
1374 return NULL; // Wait stable graph
1375 }
1376 }
1377 }
1378 }
1379
1380 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1381
1382 // Split through Phi (see original code in loopopts.cpp).
1383 assert(C->have_alias_type(t_oop), "instance should have alias type");
1384
1385 // Do nothing here if Identity will find a value
1386 // (to avoid infinite chain of value phis generation).
1387 if (!phase->eqv(this, this->Identity(phase)))
1388 return NULL;
1389
1390 // Select Region to split through.
1391 Node* region;
1392 if (!base_is_phi) {
1393 assert(mem->is_Phi(), "sanity");
1394 region = mem->in(0);
1395 // Skip if the region dominates some control edge of the address.
1396 if (!MemNode::all_controls_dominate(address, region))
1397 return NULL;
1398 } else if (!mem->is_Phi()) {
1399 assert(base_is_phi, "sanity");
1400 region = base->in(0);
1401 // Skip if the region dominates some control edge of the memory.
1402 if (!MemNode::all_controls_dominate(mem, region))
1403 return NULL;
1404 } else if (base->in(0) != mem->in(0)) {
1405 assert(base_is_phi && mem->is_Phi(), "sanity");
1406 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1407 region = base->in(0);
1408 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1409 region = mem->in(0);
1410 } else {
1411 return NULL; // complex graph
1412 }
1413 } else {
1414 assert(base->in(0) == mem->in(0), "sanity");
1415 region = mem->in(0);
1416 }
1417
1418 const Type* this_type = this->bottom_type();
1419 int this_index = C->get_alias_index(t_oop);
1420 int this_offset = t_oop->offset();
1421 int this_iid = t_oop->instance_id();
1422 if (!t_oop->is_known_instance() && load_boxed_values) {
1423 // Use _idx of address base for boxed values.
1424 this_iid = base->_idx;
1425 }
1426 PhaseIterGVN* igvn = phase->is_IterGVN();
1427 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1428 for (uint i = 1; i < region->req(); i++) {
1429 Node* x;
1430 Node* the_clone = NULL;
1431 if (region->in(i) == C->top()) {
1432 x = C->top(); // Dead path? Use a dead data op
1433 } else {
1434 x = this->clone(); // Else clone up the data op
1435 the_clone = x; // Remember for possible deletion.
1436 // Alter data node to use pre-phi inputs
1437 if (this->in(0) == region) {
1438 x->set_req(0, region->in(i));
1439 } else {
1440 x->set_req(0, NULL);
1441 }
1442 if (mem->is_Phi() && (mem->in(0) == region)) {
1443 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1444 }
1445 if (address->is_Phi() && address->in(0) == region) {
1446 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1447 }
1448 if (base_is_phi && (base->in(0) == region)) {
1449 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1450 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1451 x->set_req(Address, adr_x);
1452 }
1453 }
1454 // Check for a 'win' on some paths
1455 const Type *t = x->Value(igvn);
1456
1457 bool singleton = t->singleton();
1458
1459 // See comments in PhaseIdealLoop::split_thru_phi().
1460 if (singleton && t == Type::TOP) {
1461 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1462 }
1463
1464 if (singleton) {
1465 x = igvn->makecon(t);
1466 } else {
1467 // We now call Identity to try to simplify the cloned node.
1468 // Note that some Identity methods call phase->type(this).
1469 // Make sure that the type array is big enough for
1470 // our new node, even though we may throw the node away.
1471 // (This tweaking with igvn only works because x is a new node.)
1472 igvn->set_type(x, t);
1473 // If x is a TypeNode, capture any more-precise type permanently into Node
1474 // otherwise it will be not updated during igvn->transform since
1475 // igvn->type(x) is set to x->Value() already.
1476 x->raise_bottom_type(t);
1477 Node *y = x->Identity(igvn);
1478 if (y != x) {
1479 x = y;
1480 } else {
1481 y = igvn->hash_find_insert(x);
1482 if (y) {
1483 x = y;
1484 } else {
1485 // Else x is a new node we are keeping
1486 // We do not need register_new_node_with_optimizer
1487 // because set_type has already been called.
1488 igvn->_worklist.push(x);
1489 }
1490 }
1491 }
1492 if (x != the_clone && the_clone != NULL) {
1493 igvn->remove_dead_node(the_clone);
1494 }
1495 phi->set_req(i, x);
1496 }
1497 // Record Phi
1498 igvn->register_new_node_with_optimizer(phi);
1499 return phi;
1500 }
1501
1502 //------------------------------Ideal------------------------------------------
1503 // If the load is from Field memory and the pointer is non-null, it might be possible to
1504 // zero out the control input.
1505 // If the offset is constant and the base is an object allocation,
1506 // try to hook me up to the exact initializing store.
1507 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1508 Node* p = MemNode::Ideal_common(phase, can_reshape);
1509 if (p) return (p == NodeSentinel) ? NULL : p;
1510
1511 Node* ctrl = in(MemNode::Control);
1512 Node* address = in(MemNode::Address);
1513 bool progress = false;
1514
1515 // Skip up past a SafePoint control. Cannot do this for Stores because
1516 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1517 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1518 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1519 ctrl = ctrl->in(0);
1520 set_req(MemNode::Control,ctrl);
1521 progress = true;
1522 }
1523
1524 intptr_t ignore = 0;
1525 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1526 if (base != NULL
1527 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1528 // Check for useless control edge in some common special cases
1529 if (in(MemNode::Control) != NULL
1530 && can_remove_control()
1531 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1532 && all_controls_dominate(base, phase->C->start())) {
1533 // A method-invariant, non-null address (constant or 'this' argument).
1534 set_req(MemNode::Control, NULL);
1535 progress = true;
1536 }
1537 }
1538
1539 Node* mem = in(MemNode::Memory);
1540 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1541
1542 if (can_reshape && (addr_t != NULL)) {
1543 // try to optimize our memory input
1544 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1545 if (opt_mem != mem) {
1546 set_req(MemNode::Memory, opt_mem);
1547 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1548 return this;
1549 }
1550 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1551 if ((t_oop != NULL) &&
1552 (t_oop->is_known_instance_field() ||
1553 t_oop->is_ptr_to_boxed_value())) {
1554 PhaseIterGVN *igvn = phase->is_IterGVN();
1555 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1556 // Delay this transformation until memory Phi is processed.
1557 phase->is_IterGVN()->_worklist.push(this);
1558 return NULL;
1559 }
1560 // Split instance field load through Phi.
1561 Node* result = split_through_phi(phase);
1562 if (result != NULL) return result;
1563
1564 if (t_oop->is_ptr_to_boxed_value()) {
1565 Node* result = eliminate_autobox(phase);
1566 if (result != NULL) return result;
1567 }
1568 }
1569 }
1570
1571 // Is there a dominating load that loads the same value? Leave
1572 // anything that is not a load of a field/array element (like
1573 // barriers etc.) alone
1574 if (in(0) != NULL && adr_type() != TypeRawPtr::BOTTOM && can_reshape) {
1575 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1576 Node *use = mem->fast_out(i);
1577 if (use != this &&
1578 use->Opcode() == Opcode() &&
1579 use->in(0) != NULL &&
1580 use->in(0) != in(0) &&
1581 use->in(Address) == in(Address)) {
1582 Node* ctl = in(0);
1583 for (int i = 0; i < 10 && ctl != NULL; i++) {
1584 ctl = IfNode::up_one_dom(ctl);
1585 if (ctl == use->in(0)) {
1586 set_req(0, use->in(0));
1587 return this;
1588 }
1589 }
1590 }
1591 }
1592 }
1593
1594 // Check for prior store with a different base or offset; make Load
1595 // independent. Skip through any number of them. Bail out if the stores
1596 // are in an endless dead cycle and report no progress. This is a key
1597 // transform for Reflection. However, if after skipping through the Stores
1598 // we can't then fold up against a prior store do NOT do the transform as
1599 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1600 // array memory alive twice: once for the hoisted Load and again after the
1601 // bypassed Store. This situation only works if EVERYBODY who does
1602 // anti-dependence work knows how to bypass. I.e. we need all
1603 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1604 // the alias index stuff. So instead, peek through Stores and IFF we can
1605 // fold up, do so.
1606 Node* prev_mem = find_previous_store(phase);
1607 if (prev_mem != NULL) {
1608 Node* value = can_see_arraycopy_value(prev_mem, phase);
1609 if (value != NULL) {
1610 return value;
1611 }
1612 }
1613 // Steps (a), (b): Walk past independent stores to find an exact match.
1614 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1615 // (c) See if we can fold up on the spot, but don't fold up here.
1616 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1617 // just return a prior value, which is done by Identity calls.
1618 if (can_see_stored_value(prev_mem, phase)) {
1619 // Make ready for step (d):
1620 set_req(MemNode::Memory, prev_mem);
1621 return this;
1622 }
1623 }
1624
1625 return progress ? this : NULL;
1626 }
1627
1628 // Helper to recognize certain Klass fields which are invariant across
1629 // some group of array types (e.g., int[] or all T[] where T < Object).
1630 const Type*
1631 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1632 ciKlass* klass) const {
1633 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1634 // The field is Klass::_modifier_flags. Return its (constant) value.
1635 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1636 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1637 return TypeInt::make(klass->modifier_flags());
1638 }
1639 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1640 // The field is Klass::_access_flags. Return its (constant) value.
1641 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1642 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1643 return TypeInt::make(klass->access_flags());
1644 }
1645 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1646 // The field is Klass::_layout_helper. Return its constant value if known.
1647 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1648 return TypeInt::make(klass->layout_helper());
1649 }
1650
1651 // No match.
1652 return NULL;
1653 }
1654
1655 //------------------------------Value-----------------------------------------
1656 const Type* LoadNode::Value(PhaseGVN* phase) const {
1657 // Either input is TOP ==> the result is TOP
1658 Node* mem = in(MemNode::Memory);
1659 const Type *t1 = phase->type(mem);
1660 if (t1 == Type::TOP) return Type::TOP;
1661 Node* adr = in(MemNode::Address);
1662 const TypePtr* tp = phase->type(adr)->isa_ptr();
1663 if (tp == NULL || tp->empty()) return Type::TOP;
1664 int off = tp->offset();
1665 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1666 Compile* C = phase->C;
1667
1668 // Try to guess loaded type from pointer type
1669 if (tp->isa_aryptr()) {
1670 const TypeAryPtr* ary = tp->is_aryptr();
1671 const Type* t = ary->elem();
1672
1673 // Determine whether the reference is beyond the header or not, by comparing
1674 // the offset against the offset of the start of the array's data.
1675 // Different array types begin at slightly different offsets (12 vs. 16).
1676 // We choose T_BYTE as an example base type that is least restrictive
1677 // as to alignment, which will therefore produce the smallest
1678 // possible base offset.
1679 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1680 const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1681
1682 // Try to constant-fold a stable array element.
1683 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1684 // Make sure the reference is not into the header and the offset is constant
1685 ciObject* aobj = ary->const_oop();
1686 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1687 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1688 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1689 stable_dimension,
1690 memory_type(), is_unsigned());
1691 if (con_type != NULL) {
1692 return con_type;
1693 }
1694 }
1695 }
1696
1697 // Don't do this for integer types. There is only potential profit if
1698 // the element type t is lower than _type; that is, for int types, if _type is
1699 // more restrictive than t. This only happens here if one is short and the other
1700 // char (both 16 bits), and in those cases we've made an intentional decision
1701 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1702 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1703 //
1704 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1705 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1706 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1707 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1708 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1709 // In fact, that could have been the original type of p1, and p1 could have
1710 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1711 // expression (LShiftL quux 3) independently optimized to the constant 8.
1712 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1713 && (_type->isa_vect() == NULL)
1714 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1715 // t might actually be lower than _type, if _type is a unique
1716 // concrete subclass of abstract class t.
1717 if (off_beyond_header) { // is the offset beyond the header?
1718 const Type* jt = t->join_speculative(_type);
1719 // In any case, do not allow the join, per se, to empty out the type.
1720 if (jt->empty() && !t->empty()) {
1721 // This can happen if a interface-typed array narrows to a class type.
1722 jt = _type;
1723 }
1724 #ifdef ASSERT
1725 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1726 // The pointers in the autobox arrays are always non-null
1727 Node* base = adr->in(AddPNode::Base);
1728 if ((base != NULL) && base->is_DecodeN()) {
1729 // Get LoadN node which loads IntegerCache.cache field
1730 base = base->in(1);
1731 }
1732 if ((base != NULL) && base->is_Con()) {
1733 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1734 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1735 // It could be narrow oop
1736 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1737 }
1738 }
1739 }
1740 #endif
1741 return jt;
1742 }
1743 }
1744 } else if (tp->base() == Type::InstPtr) {
1745 assert( off != Type::OffsetBot ||
1746 // arrays can be cast to Objects
1747 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1748 // unsafe field access may not have a constant offset
1749 C->has_unsafe_access(),
1750 "Field accesses must be precise" );
1751 // For oop loads, we expect the _type to be precise.
1752
1753 // Optimize loads from constant fields.
1754 const TypeInstPtr* tinst = tp->is_instptr();
1755 ciObject* const_oop = tinst->const_oop();
1756 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1757 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1758 if (con_type != NULL) {
1759 return con_type;
1760 }
1761 }
1762 } else if (tp->base() == Type::KlassPtr) {
1763 assert( off != Type::OffsetBot ||
1764 // arrays can be cast to Objects
1765 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1766 // also allow array-loading from the primary supertype
1767 // array during subtype checks
1768 Opcode() == Op_LoadKlass,
1769 "Field accesses must be precise" );
1770 // For klass/static loads, we expect the _type to be precise
1771 }
1772
1773 const TypeKlassPtr *tkls = tp->isa_klassptr();
1774 if (tkls != NULL && !StressReflectiveCode) {
1775 ciKlass* klass = tkls->klass();
1776 if (klass->is_loaded() && tkls->klass_is_exact()) {
1777 // We are loading a field from a Klass metaobject whose identity
1778 // is known at compile time (the type is "exact" or "precise").
1779 // Check for fields we know are maintained as constants by the VM.
1780 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1781 // The field is Klass::_super_check_offset. Return its (constant) value.
1782 // (Folds up type checking code.)
1783 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1784 return TypeInt::make(klass->super_check_offset());
1785 }
1786 // Compute index into primary_supers array
1787 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1788 // Check for overflowing; use unsigned compare to handle the negative case.
1789 if( depth < ciKlass::primary_super_limit() ) {
1790 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1791 // (Folds up type checking code.)
1792 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1793 ciKlass *ss = klass->super_of_depth(depth);
1794 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1795 }
1796 const Type* aift = load_array_final_field(tkls, klass);
1797 if (aift != NULL) return aift;
1798 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1799 // The field is Klass::_java_mirror. Return its (constant) value.
1800 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1801 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1802 return TypeInstPtr::make(klass->java_mirror());
1803 }
1804 }
1805
1806 // We can still check if we are loading from the primary_supers array at a
1807 // shallow enough depth. Even though the klass is not exact, entries less
1808 // than or equal to its super depth are correct.
1809 if (klass->is_loaded() ) {
1810 ciType *inner = klass;
1811 while( inner->is_obj_array_klass() )
1812 inner = inner->as_obj_array_klass()->base_element_type();
1813 if( inner->is_instance_klass() &&
1814 !inner->as_instance_klass()->flags().is_interface() ) {
1815 // Compute index into primary_supers array
1816 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1817 // Check for overflowing; use unsigned compare to handle the negative case.
1818 if( depth < ciKlass::primary_super_limit() &&
1819 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1820 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1821 // (Folds up type checking code.)
1822 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1823 ciKlass *ss = klass->super_of_depth(depth);
1824 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1825 }
1826 }
1827 }
1828
1829 // If the type is enough to determine that the thing is not an array,
1830 // we can give the layout_helper a positive interval type.
1831 // This will help short-circuit some reflective code.
1832 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1833 && !klass->is_array_klass() // not directly typed as an array
1834 && !klass->is_interface() // specifically not Serializable & Cloneable
1835 && !klass->is_java_lang_Object() // not the supertype of all T[]
1836 ) {
1837 // Note: When interfaces are reliable, we can narrow the interface
1838 // test to (klass != Serializable && klass != Cloneable).
1839 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1840 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1841 // The key property of this type is that it folds up tests
1842 // for array-ness, since it proves that the layout_helper is positive.
1843 // Thus, a generic value like the basic object layout helper works fine.
1844 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1845 }
1846 }
1847
1848 // If we are loading from a freshly-allocated object, produce a zero,
1849 // if the load is provably beyond the header of the object.
1850 // (Also allow a variable load from a fresh array to produce zero.)
1851 const TypeOopPtr *tinst = tp->isa_oopptr();
1852 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1853 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1854 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1855 Node* value = can_see_stored_value(mem,phase);
1856 if (value != NULL && value->is_Con()) {
1857 assert(value->bottom_type()->higher_equal(_type),"sanity");
1858 return value->bottom_type();
1859 }
1860 }
1861
1862 if (is_instance) {
1863 // If we have an instance type and our memory input is the
1864 // programs's initial memory state, there is no matching store,
1865 // so just return a zero of the appropriate type
1866 Node *mem = in(MemNode::Memory);
1867 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1868 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1869 return Type::get_zero_type(_type->basic_type());
1870 }
1871 }
1872 return _type;
1873 }
1874
1875 //------------------------------match_edge-------------------------------------
1876 // Do we Match on this edge index or not? Match only the address.
1877 uint LoadNode::match_edge(uint idx) const {
1878 return idx == MemNode::Address;
1879 }
1880
1881 //--------------------------LoadBNode::Ideal--------------------------------------
1882 //
1883 // If the previous store is to the same address as this load,
1884 // and the value stored was larger than a byte, replace this load
1885 // with the value stored truncated to a byte. If no truncation is
1886 // needed, the replacement is done in LoadNode::Identity().
1887 //
1888 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1889 Node* mem = in(MemNode::Memory);
1890 Node* value = can_see_stored_value(mem,phase);
1891 if( value && !phase->type(value)->higher_equal( _type ) ) {
1892 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1893 return new RShiftINode(result, phase->intcon(24));
1894 }
1895 // Identity call will handle the case where truncation is not needed.
1896 return LoadNode::Ideal(phase, can_reshape);
1897 }
1898
1899 const Type* LoadBNode::Value(PhaseGVN* phase) const {
1900 Node* mem = in(MemNode::Memory);
1901 Node* value = can_see_stored_value(mem,phase);
1902 if (value != NULL && value->is_Con() &&
1903 !value->bottom_type()->higher_equal(_type)) {
1904 // If the input to the store does not fit with the load's result type,
1905 // it must be truncated. We can't delay until Ideal call since
1906 // a singleton Value is needed for split_thru_phi optimization.
1907 int con = value->get_int();
1908 return TypeInt::make((con << 24) >> 24);
1909 }
1910 return LoadNode::Value(phase);
1911 }
1912
1913 //--------------------------LoadUBNode::Ideal-------------------------------------
1914 //
1915 // If the previous store is to the same address as this load,
1916 // and the value stored was larger than a byte, replace this load
1917 // with the value stored truncated to a byte. If no truncation is
1918 // needed, the replacement is done in LoadNode::Identity().
1919 //
1920 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1921 Node* mem = in(MemNode::Memory);
1922 Node* value = can_see_stored_value(mem, phase);
1923 if (value && !phase->type(value)->higher_equal(_type))
1924 return new AndINode(value, phase->intcon(0xFF));
1925 // Identity call will handle the case where truncation is not needed.
1926 return LoadNode::Ideal(phase, can_reshape);
1927 }
1928
1929 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
1930 Node* mem = in(MemNode::Memory);
1931 Node* value = can_see_stored_value(mem,phase);
1932 if (value != NULL && value->is_Con() &&
1933 !value->bottom_type()->higher_equal(_type)) {
1934 // If the input to the store does not fit with the load's result type,
1935 // it must be truncated. We can't delay until Ideal call since
1936 // a singleton Value is needed for split_thru_phi optimization.
1937 int con = value->get_int();
1938 return TypeInt::make(con & 0xFF);
1939 }
1940 return LoadNode::Value(phase);
1941 }
1942
1943 //--------------------------LoadUSNode::Ideal-------------------------------------
1944 //
1945 // If the previous store is to the same address as this load,
1946 // and the value stored was larger than a char, replace this load
1947 // with the value stored truncated to a char. If no truncation is
1948 // needed, the replacement is done in LoadNode::Identity().
1949 //
1950 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1951 Node* mem = in(MemNode::Memory);
1952 Node* value = can_see_stored_value(mem,phase);
1953 if( value && !phase->type(value)->higher_equal( _type ) )
1954 return new AndINode(value,phase->intcon(0xFFFF));
1955 // Identity call will handle the case where truncation is not needed.
1956 return LoadNode::Ideal(phase, can_reshape);
1957 }
1958
1959 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
1960 Node* mem = in(MemNode::Memory);
1961 Node* value = can_see_stored_value(mem,phase);
1962 if (value != NULL && value->is_Con() &&
1963 !value->bottom_type()->higher_equal(_type)) {
1964 // If the input to the store does not fit with the load's result type,
1965 // it must be truncated. We can't delay until Ideal call since
1966 // a singleton Value is needed for split_thru_phi optimization.
1967 int con = value->get_int();
1968 return TypeInt::make(con & 0xFFFF);
1969 }
1970 return LoadNode::Value(phase);
1971 }
1972
1973 //--------------------------LoadSNode::Ideal--------------------------------------
1974 //
1975 // If the previous store is to the same address as this load,
1976 // and the value stored was larger than a short, replace this load
1977 // with the value stored truncated to a short. If no truncation is
1978 // needed, the replacement is done in LoadNode::Identity().
1979 //
1980 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1981 Node* mem = in(MemNode::Memory);
1982 Node* value = can_see_stored_value(mem,phase);
1983 if( value && !phase->type(value)->higher_equal( _type ) ) {
1984 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
1985 return new RShiftINode(result, phase->intcon(16));
1986 }
1987 // Identity call will handle the case where truncation is not needed.
1988 return LoadNode::Ideal(phase, can_reshape);
1989 }
1990
1991 const Type* LoadSNode::Value(PhaseGVN* phase) const {
1992 Node* mem = in(MemNode::Memory);
1993 Node* value = can_see_stored_value(mem,phase);
1994 if (value != NULL && value->is_Con() &&
1995 !value->bottom_type()->higher_equal(_type)) {
1996 // If the input to the store does not fit with the load's result type,
1997 // it must be truncated. We can't delay until Ideal call since
1998 // a singleton Value is needed for split_thru_phi optimization.
1999 int con = value->get_int();
2000 return TypeInt::make((con << 16) >> 16);
2001 }
2002 return LoadNode::Value(phase);
2003 }
2004
2005 //=============================================================================
2006 //----------------------------LoadKlassNode::make------------------------------
2007 // Polymorphic factory method:
2008 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2009 // sanity check the alias category against the created node type
2010 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2011 assert(adr_type != NULL, "expecting TypeKlassPtr");
2012 #ifdef _LP64
2013 if (adr_type->is_ptr_to_narrowklass()) {
2014 assert(UseCompressedClassPointers, "no compressed klasses");
2015 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2016 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2017 }
2018 #endif
2019 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2020 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2021 }
2022
2023 //------------------------------Value------------------------------------------
2024 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2025 return klass_value_common(phase);
2026 }
2027
2028 // In most cases, LoadKlassNode does not have the control input set. If the control
2029 // input is set, it must not be removed (by LoadNode::Ideal()).
2030 bool LoadKlassNode::can_remove_control() const {
2031 return false;
2032 }
2033
2034 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2035 // Either input is TOP ==> the result is TOP
2036 const Type *t1 = phase->type( in(MemNode::Memory) );
2037 if (t1 == Type::TOP) return Type::TOP;
2038 Node *adr = in(MemNode::Address);
2039 const Type *t2 = phase->type( adr );
2040 if (t2 == Type::TOP) return Type::TOP;
2041 const TypePtr *tp = t2->is_ptr();
2042 if (TypePtr::above_centerline(tp->ptr()) ||
2043 tp->ptr() == TypePtr::Null) return Type::TOP;
2044
2045 // Return a more precise klass, if possible
2046 const TypeInstPtr *tinst = tp->isa_instptr();
2047 if (tinst != NULL) {
2048 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2049 int offset = tinst->offset();
2050 if (ik == phase->C->env()->Class_klass()
2051 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2052 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2053 // We are loading a special hidden field from a Class mirror object,
2054 // the field which points to the VM's Klass metaobject.
2055 ciType* t = tinst->java_mirror_type();
2056 // java_mirror_type returns non-null for compile-time Class constants.
2057 if (t != NULL) {
2058 // constant oop => constant klass
2059 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2060 if (t->is_void()) {
2061 // We cannot create a void array. Since void is a primitive type return null
2062 // klass. Users of this result need to do a null check on the returned klass.
2063 return TypePtr::NULL_PTR;
2064 }
2065 return TypeKlassPtr::make(ciArrayKlass::make(t));
2066 }
2067 if (!t->is_klass()) {
2068 // a primitive Class (e.g., int.class) has NULL for a klass field
2069 return TypePtr::NULL_PTR;
2070 }
2071 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2072 return TypeKlassPtr::make(t->as_klass());
2073 }
2074 // non-constant mirror, so we can't tell what's going on
2075 }
2076 if( !ik->is_loaded() )
2077 return _type; // Bail out if not loaded
2078 if (offset == oopDesc::klass_offset_in_bytes()) {
2079 if (tinst->klass_is_exact()) {
2080 return TypeKlassPtr::make(ik);
2081 }
2082 // See if we can become precise: no subklasses and no interface
2083 // (Note: We need to support verified interfaces.)
2084 if (!ik->is_interface() && !ik->has_subklass()) {
2085 //assert(!UseExactTypes, "this code should be useless with exact types");
2086 // Add a dependence; if any subclass added we need to recompile
2087 if (!ik->is_final()) {
2088 // %%% should use stronger assert_unique_concrete_subtype instead
2089 phase->C->dependencies()->assert_leaf_type(ik);
2090 }
2091 // Return precise klass
2092 return TypeKlassPtr::make(ik);
2093 }
2094
2095 // Return root of possible klass
2096 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2097 }
2098 }
2099
2100 // Check for loading klass from an array
2101 const TypeAryPtr *tary = tp->isa_aryptr();
2102 if( tary != NULL ) {
2103 ciKlass *tary_klass = tary->klass();
2104 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2105 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2106 if (tary->klass_is_exact()) {
2107 return TypeKlassPtr::make(tary_klass);
2108 }
2109 ciArrayKlass *ak = tary->klass()->as_array_klass();
2110 // If the klass is an object array, we defer the question to the
2111 // array component klass.
2112 if( ak->is_obj_array_klass() ) {
2113 assert( ak->is_loaded(), "" );
2114 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2115 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2116 ciInstanceKlass* ik = base_k->as_instance_klass();
2117 // See if we can become precise: no subklasses and no interface
2118 if (!ik->is_interface() && !ik->has_subklass()) {
2119 //assert(!UseExactTypes, "this code should be useless with exact types");
2120 // Add a dependence; if any subclass added we need to recompile
2121 if (!ik->is_final()) {
2122 phase->C->dependencies()->assert_leaf_type(ik);
2123 }
2124 // Return precise array klass
2125 return TypeKlassPtr::make(ak);
2126 }
2127 }
2128 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2129 } else { // Found a type-array?
2130 //assert(!UseExactTypes, "this code should be useless with exact types");
2131 assert( ak->is_type_array_klass(), "" );
2132 return TypeKlassPtr::make(ak); // These are always precise
2133 }
2134 }
2135 }
2136
2137 // Check for loading klass from an array klass
2138 const TypeKlassPtr *tkls = tp->isa_klassptr();
2139 if (tkls != NULL && !StressReflectiveCode) {
2140 ciKlass* klass = tkls->klass();
2141 if( !klass->is_loaded() )
2142 return _type; // Bail out if not loaded
2143 if( klass->is_obj_array_klass() &&
2144 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2145 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2146 // // Always returning precise element type is incorrect,
2147 // // e.g., element type could be object and array may contain strings
2148 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2149
2150 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2151 // according to the element type's subclassing.
2152 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2153 }
2154 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2155 tkls->offset() == in_bytes(Klass::super_offset())) {
2156 ciKlass* sup = klass->as_instance_klass()->super();
2157 // The field is Klass::_super. Return its (constant) value.
2158 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2159 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2160 }
2161 }
2162
2163 // Bailout case
2164 return LoadNode::Value(phase);
2165 }
2166
2167 //------------------------------Identity---------------------------------------
2168 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2169 // Also feed through the klass in Allocate(...klass...)._klass.
2170 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2171 return klass_identity_common(phase);
2172 }
2173
2174 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2175 Node* x = LoadNode::Identity(phase);
2176 if (x != this) return x;
2177
2178 // Take apart the address into an oop and and offset.
2179 // Return 'this' if we cannot.
2180 Node* adr = in(MemNode::Address);
2181 intptr_t offset = 0;
2182 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2183 if (base == NULL) return this;
2184 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2185 if (toop == NULL) return this;
2186
2187 // We can fetch the klass directly through an AllocateNode.
2188 // This works even if the klass is not constant (clone or newArray).
2189 if (offset == oopDesc::klass_offset_in_bytes()) {
2190 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2191 if (allocated_klass != NULL) {
2192 return allocated_klass;
2193 }
2194 }
2195
2196 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2197 // See inline_native_Class_query for occurrences of these patterns.
2198 // Java Example: x.getClass().isAssignableFrom(y)
2199 //
2200 // This improves reflective code, often making the Class
2201 // mirror go completely dead. (Current exception: Class
2202 // mirrors may appear in debug info, but we could clean them out by
2203 // introducing a new debug info operator for Klass*.java_mirror).
2204 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2205 && offset == java_lang_Class::klass_offset_in_bytes()) {
2206 // We are loading a special hidden field from a Class mirror,
2207 // the field which points to its Klass or ArrayKlass metaobject.
2208 if (base->is_Load()) {
2209 Node* adr2 = base->in(MemNode::Address);
2210 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2211 if (tkls != NULL && !tkls->empty()
2212 && (tkls->klass()->is_instance_klass() ||
2213 tkls->klass()->is_array_klass())
2214 && adr2->is_AddP()
2215 ) {
2216 int mirror_field = in_bytes(Klass::java_mirror_offset());
2217 if (tkls->offset() == mirror_field) {
2218 return adr2->in(AddPNode::Base);
2219 }
2220 }
2221 }
2222 }
2223
2224 return this;
2225 }
2226
2227
2228 //------------------------------Value------------------------------------------
2229 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2230 const Type *t = klass_value_common(phase);
2231 if (t == Type::TOP)
2232 return t;
2233
2234 return t->make_narrowklass();
2235 }
2236
2237 //------------------------------Identity---------------------------------------
2238 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2239 // Also feed through the klass in Allocate(...klass...)._klass.
2240 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2241 Node *x = klass_identity_common(phase);
2242
2243 const Type *t = phase->type( x );
2244 if( t == Type::TOP ) return x;
2245 if( t->isa_narrowklass()) return x;
2246 assert (!t->isa_narrowoop(), "no narrow oop here");
2247
2248 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2249 }
2250
2251 //------------------------------Value-----------------------------------------
2252 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2253 // Either input is TOP ==> the result is TOP
2254 const Type *t1 = phase->type( in(MemNode::Memory) );
2255 if( t1 == Type::TOP ) return Type::TOP;
2256 Node *adr = in(MemNode::Address);
2257 const Type *t2 = phase->type( adr );
2258 if( t2 == Type::TOP ) return Type::TOP;
2259 const TypePtr *tp = t2->is_ptr();
2260 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2261 const TypeAryPtr *tap = tp->isa_aryptr();
2262 if( !tap ) return _type;
2263 return tap->size();
2264 }
2265
2266 //-------------------------------Ideal---------------------------------------
2267 // Feed through the length in AllocateArray(...length...)._length.
2268 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2269 Node* p = MemNode::Ideal_common(phase, can_reshape);
2270 if (p) return (p == NodeSentinel) ? NULL : p;
2271
2272 // Take apart the address into an oop and and offset.
2273 // Return 'this' if we cannot.
2274 Node* adr = in(MemNode::Address);
2275 intptr_t offset = 0;
2276 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2277 if (base == NULL) return NULL;
2278 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2279 if (tary == NULL) return NULL;
2280
2281 // We can fetch the length directly through an AllocateArrayNode.
2282 // This works even if the length is not constant (clone or newArray).
2283 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2284 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2285 if (alloc != NULL) {
2286 Node* allocated_length = alloc->Ideal_length();
2287 Node* len = alloc->make_ideal_length(tary, phase);
2288 if (allocated_length != len) {
2289 // New CastII improves on this.
2290 return len;
2291 }
2292 }
2293 }
2294
2295 return NULL;
2296 }
2297
2298 //------------------------------Identity---------------------------------------
2299 // Feed through the length in AllocateArray(...length...)._length.
2300 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2301 Node* x = LoadINode::Identity(phase);
2302 if (x != this) return x;
2303
2304 // Take apart the address into an oop and and offset.
2305 // Return 'this' if we cannot.
2306 Node* adr = in(MemNode::Address);
2307 intptr_t offset = 0;
2308 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2309 if (base == NULL) return this;
2310 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2311 if (tary == NULL) return this;
2312
2313 // We can fetch the length directly through an AllocateArrayNode.
2314 // This works even if the length is not constant (clone or newArray).
2315 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2316 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2317 if (alloc != NULL) {
2318 Node* allocated_length = alloc->Ideal_length();
2319 // Do not allow make_ideal_length to allocate a CastII node.
2320 Node* len = alloc->make_ideal_length(tary, phase, false);
2321 if (allocated_length == len) {
2322 // Return allocated_length only if it would not be improved by a CastII.
2323 return allocated_length;
2324 }
2325 }
2326 }
2327
2328 return this;
2329
2330 }
2331
2332 //=============================================================================
2333 //---------------------------StoreNode::make-----------------------------------
2334 // Polymorphic factory method:
2335 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2336 assert((mo == unordered || mo == release), "unexpected");
2337 Compile* C = gvn.C;
2338 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2339 ctl != NULL, "raw memory operations should have control edge");
2340
2341 switch (bt) {
2342 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2343 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2344 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2345 case T_CHAR:
2346 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2347 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2348 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2349 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2350 case T_METADATA:
2351 case T_ADDRESS:
2352 case T_OBJECT:
2353 #ifdef _LP64
2354 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2355 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2356 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2357 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2358 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2359 adr->bottom_type()->isa_rawptr())) {
2360 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2361 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2362 }
2363 #endif
2364 {
2365 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2366 }
2367 }
2368 ShouldNotReachHere();
2369 return (StoreNode*)NULL;
2370 }
2371
2372 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2373 bool require_atomic = true;
2374 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2375 }
2376
2377 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2378 bool require_atomic = true;
2379 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2380 }
2381
2382
2383 //--------------------------bottom_type----------------------------------------
2384 const Type *StoreNode::bottom_type() const {
2385 return Type::MEMORY;
2386 }
2387
2388 //------------------------------hash-------------------------------------------
2389 uint StoreNode::hash() const {
2390 // unroll addition of interesting fields
2391 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2392
2393 // Since they are not commoned, do not hash them:
2394 return NO_HASH;
2395 }
2396
2397 //------------------------------Ideal------------------------------------------
2398 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2399 // When a store immediately follows a relevant allocation/initialization,
2400 // try to capture it into the initialization, or hoist it above.
2401 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2402 Node* p = MemNode::Ideal_common(phase, can_reshape);
2403 if (p) return (p == NodeSentinel) ? NULL : p;
2404
2405 Node* mem = in(MemNode::Memory);
2406 Node* address = in(MemNode::Address);
2407 // Back-to-back stores to same address? Fold em up. Generally
2408 // unsafe if I have intervening uses... Also disallowed for StoreCM
2409 // since they must follow each StoreP operation. Redundant StoreCMs
2410 // are eliminated just before matching in final_graph_reshape.
2411 {
2412 Node* st = mem;
2413 // If Store 'st' has more than one use, we cannot fold 'st' away.
2414 // For example, 'st' might be the final state at a conditional
2415 // return. Or, 'st' might be used by some node which is live at
2416 // the same time 'st' is live, which might be unschedulable. So,
2417 // require exactly ONE user until such time as we clone 'mem' for
2418 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2419 // true).
2420 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2421 // Looking at a dead closed cycle of memory?
2422 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2423 assert(Opcode() == st->Opcode() ||
2424 st->Opcode() == Op_StoreVector ||
2425 Opcode() == Op_StoreVector ||
2426 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2427 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2428 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2429 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2430
2431 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2432 st->as_Store()->memory_size() <= this->memory_size()) {
2433 Node* use = st->raw_out(0);
2434 phase->igvn_rehash_node_delayed(use);
2435 if (can_reshape) {
2436 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2437 } else {
2438 // It's OK to do this in the parser, since DU info is always accurate,
2439 // and the parser always refers to nodes via SafePointNode maps.
2440 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2441 }
2442 return this;
2443 }
2444 st = st->in(MemNode::Memory);
2445 }
2446 }
2447
2448
2449 // Capture an unaliased, unconditional, simple store into an initializer.
2450 // Or, if it is independent of the allocation, hoist it above the allocation.
2451 if (ReduceFieldZeroing && /*can_reshape &&*/
2452 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2453 InitializeNode* init = mem->in(0)->as_Initialize();
2454 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2455 if (offset > 0) {
2456 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2457 // If the InitializeNode captured me, it made a raw copy of me,
2458 // and I need to disappear.
2459 if (moved != NULL) {
2460 // %%% hack to ensure that Ideal returns a new node:
2461 mem = MergeMemNode::make(mem);
2462 return mem; // fold me away
2463 }
2464 }
2465 }
2466
2467 return NULL; // No further progress
2468 }
2469
2470 //------------------------------Value-----------------------------------------
2471 const Type* StoreNode::Value(PhaseGVN* phase) const {
2472 // Either input is TOP ==> the result is TOP
2473 const Type *t1 = phase->type( in(MemNode::Memory) );
2474 if( t1 == Type::TOP ) return Type::TOP;
2475 const Type *t2 = phase->type( in(MemNode::Address) );
2476 if( t2 == Type::TOP ) return Type::TOP;
2477 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2478 if( t3 == Type::TOP ) return Type::TOP;
2479 return Type::MEMORY;
2480 }
2481
2482 //------------------------------Identity---------------------------------------
2483 // Remove redundant stores:
2484 // Store(m, p, Load(m, p)) changes to m.
2485 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2486 Node* StoreNode::Identity(PhaseGVN* phase) {
2487 Node* mem = in(MemNode::Memory);
2488 Node* adr = in(MemNode::Address);
2489 Node* val = in(MemNode::ValueIn);
2490
2491 // Load then Store? Then the Store is useless
2492 if (val->is_Load() &&
2493 val->in(MemNode::Address)->eqv_uncast(adr) &&
2494 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2495 val->as_Load()->store_Opcode() == Opcode()) {
2496 return mem;
2497 }
2498
2499 // Two stores in a row of the same value?
2500 if (mem->is_Store() &&
2501 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2502 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2503 mem->Opcode() == Opcode()) {
2504 return mem;
2505 }
2506
2507 // Store of zero anywhere into a freshly-allocated object?
2508 // Then the store is useless.
2509 // (It must already have been captured by the InitializeNode.)
2510 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2511 // a newly allocated object is already all-zeroes everywhere
2512 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2513 return mem;
2514 }
2515
2516 // the store may also apply to zero-bits in an earlier object
2517 Node* prev_mem = find_previous_store(phase);
2518 // Steps (a), (b): Walk past independent stores to find an exact match.
2519 if (prev_mem != NULL) {
2520 Node* prev_val = can_see_stored_value(prev_mem, phase);
2521 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2522 // prev_val and val might differ by a cast; it would be good
2523 // to keep the more informative of the two.
2524 return mem;
2525 }
2526 }
2527 }
2528
2529 return this;
2530 }
2531
2532 //------------------------------match_edge-------------------------------------
2533 // Do we Match on this edge index or not? Match only memory & value
2534 uint StoreNode::match_edge(uint idx) const {
2535 return idx == MemNode::Address || idx == MemNode::ValueIn;
2536 }
2537
2538 //------------------------------cmp--------------------------------------------
2539 // Do not common stores up together. They generally have to be split
2540 // back up anyways, so do not bother.
2541 uint StoreNode::cmp( const Node &n ) const {
2542 return (&n == this); // Always fail except on self
2543 }
2544
2545 //------------------------------Ideal_masked_input-----------------------------
2546 // Check for a useless mask before a partial-word store
2547 // (StoreB ... (AndI valIn conIa) )
2548 // If (conIa & mask == mask) this simplifies to
2549 // (StoreB ... (valIn) )
2550 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2551 Node *val = in(MemNode::ValueIn);
2552 if( val->Opcode() == Op_AndI ) {
2553 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2554 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2555 set_req(MemNode::ValueIn, val->in(1));
2556 return this;
2557 }
2558 }
2559 return NULL;
2560 }
2561
2562
2563 //------------------------------Ideal_sign_extended_input----------------------
2564 // Check for useless sign-extension before a partial-word store
2565 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2566 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2567 // (StoreB ... (valIn) )
2568 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2569 Node *val = in(MemNode::ValueIn);
2570 if( val->Opcode() == Op_RShiftI ) {
2571 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2572 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2573 Node *shl = val->in(1);
2574 if( shl->Opcode() == Op_LShiftI ) {
2575 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2576 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2577 set_req(MemNode::ValueIn, shl->in(1));
2578 return this;
2579 }
2580 }
2581 }
2582 }
2583 return NULL;
2584 }
2585
2586 //------------------------------value_never_loaded-----------------------------------
2587 // Determine whether there are any possible loads of the value stored.
2588 // For simplicity, we actually check if there are any loads from the
2589 // address stored to, not just for loads of the value stored by this node.
2590 //
2591 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2592 Node *adr = in(Address);
2593 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2594 if (adr_oop == NULL)
2595 return false;
2596 if (!adr_oop->is_known_instance_field())
2597 return false; // if not a distinct instance, there may be aliases of the address
2598 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2599 Node *use = adr->fast_out(i);
2600 if (use->is_Load() || use->is_LoadStore()) {
2601 return false;
2602 }
2603 }
2604 return true;
2605 }
2606
2607 //=============================================================================
2608 //------------------------------Ideal------------------------------------------
2609 // If the store is from an AND mask that leaves the low bits untouched, then
2610 // we can skip the AND operation. If the store is from a sign-extension
2611 // (a left shift, then right shift) we can skip both.
2612 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2613 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2614 if( progress != NULL ) return progress;
2615
2616 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2617 if( progress != NULL ) return progress;
2618
2619 // Finally check the default case
2620 return StoreNode::Ideal(phase, can_reshape);
2621 }
2622
2623 //=============================================================================
2624 //------------------------------Ideal------------------------------------------
2625 // If the store is from an AND mask that leaves the low bits untouched, then
2626 // we can skip the AND operation
2627 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2628 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2629 if( progress != NULL ) return progress;
2630
2631 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2632 if( progress != NULL ) return progress;
2633
2634 // Finally check the default case
2635 return StoreNode::Ideal(phase, can_reshape);
2636 }
2637
2638 //=============================================================================
2639 //------------------------------Identity---------------------------------------
2640 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2641 // No need to card mark when storing a null ptr
2642 Node* my_store = in(MemNode::OopStore);
2643 if (my_store->is_Store()) {
2644 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2645 if( t1 == TypePtr::NULL_PTR ) {
2646 return in(MemNode::Memory);
2647 }
2648 }
2649 return this;
2650 }
2651
2652 //=============================================================================
2653 //------------------------------Ideal---------------------------------------
2654 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2655 Node* progress = StoreNode::Ideal(phase, can_reshape);
2656 if (progress != NULL) return progress;
2657
2658 Node* my_store = in(MemNode::OopStore);
2659 if (my_store->is_MergeMem()) {
2660 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2661 set_req(MemNode::OopStore, mem);
2662 return this;
2663 }
2664
2665 return NULL;
2666 }
2667
2668 //------------------------------Value-----------------------------------------
2669 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2670 // Either input is TOP ==> the result is TOP
2671 const Type *t = phase->type( in(MemNode::Memory) );
2672 if( t == Type::TOP ) return Type::TOP;
2673 t = phase->type( in(MemNode::Address) );
2674 if( t == Type::TOP ) return Type::TOP;
2675 t = phase->type( in(MemNode::ValueIn) );
2676 if( t == Type::TOP ) return Type::TOP;
2677 // If extra input is TOP ==> the result is TOP
2678 t = phase->type( in(MemNode::OopStore) );
2679 if( t == Type::TOP ) return Type::TOP;
2680
2681 return StoreNode::Value( phase );
2682 }
2683
2684
2685 //=============================================================================
2686 //----------------------------------SCMemProjNode------------------------------
2687 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2688 {
2689 return bottom_type();
2690 }
2691
2692 //=============================================================================
2693 //----------------------------------LoadStoreNode------------------------------
2694 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2695 : Node(required),
2696 _type(rt),
2697 _adr_type(at)
2698 {
2699 init_req(MemNode::Control, c );
2700 init_req(MemNode::Memory , mem);
2701 init_req(MemNode::Address, adr);
2702 init_req(MemNode::ValueIn, val);
2703 init_class_id(Class_LoadStore);
2704 }
2705
2706 uint LoadStoreNode::ideal_reg() const {
2707 return _type->ideal_reg();
2708 }
2709
2710 bool LoadStoreNode::result_not_used() const {
2711 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2712 Node *x = fast_out(i);
2713 if (x->Opcode() == Op_SCMemProj) continue;
2714 return false;
2715 }
2716 return true;
2717 }
2718
2719 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2720
2721 //=============================================================================
2722 //----------------------------------LoadStoreConditionalNode--------------------
2723 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2724 init_req(ExpectedIn, ex );
2725 }
2726
2727 //=============================================================================
2728 //-------------------------------adr_type--------------------------------------
2729 const TypePtr* ClearArrayNode::adr_type() const {
2730 Node *adr = in(3);
2731 if (adr == NULL) return NULL; // node is dead
2732 return MemNode::calculate_adr_type(adr->bottom_type());
2733 }
2734
2735 //------------------------------match_edge-------------------------------------
2736 // Do we Match on this edge index or not? Do not match memory
2737 uint ClearArrayNode::match_edge(uint idx) const {
2738 return idx > 1;
2739 }
2740
2741 //------------------------------Identity---------------------------------------
2742 // Clearing a zero length array does nothing
2743 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2744 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2745 }
2746
2747 //------------------------------Idealize---------------------------------------
2748 // Clearing a short array is faster with stores
2749 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2750 // Already know this is a large node, do not try to ideal it
2751 if (!IdealizeClearArrayNode || _is_large) return NULL;
2752
2753 const int unit = BytesPerLong;
2754 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2755 if (!t) return NULL;
2756 if (!t->is_con()) return NULL;
2757 intptr_t raw_count = t->get_con();
2758 intptr_t size = raw_count;
2759 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2760 // Clearing nothing uses the Identity call.
2761 // Negative clears are possible on dead ClearArrays
2762 // (see jck test stmt114.stmt11402.val).
2763 if (size <= 0 || size % unit != 0) return NULL;
2764 intptr_t count = size / unit;
2765 // Length too long; communicate this to matchers and assemblers.
2766 // Assemblers are responsible to produce fast hardware clears for it.
2767 if (size > InitArrayShortSize) {
2768 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
2769 }
2770 Node *mem = in(1);
2771 if( phase->type(mem)==Type::TOP ) return NULL;
2772 Node *adr = in(3);
2773 const Type* at = phase->type(adr);
2774 if( at==Type::TOP ) return NULL;
2775 const TypePtr* atp = at->isa_ptr();
2776 // adjust atp to be the correct array element address type
2777 if (atp == NULL) atp = TypePtr::BOTTOM;
2778 else atp = atp->add_offset(Type::OffsetBot);
2779 // Get base for derived pointer purposes
2780 if( adr->Opcode() != Op_AddP ) Unimplemented();
2781 Node *base = adr->in(1);
2782
2783 Node *zero = phase->makecon(TypeLong::ZERO);
2784 Node *off = phase->MakeConX(BytesPerLong);
2785 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2786 count--;
2787 while( count-- ) {
2788 mem = phase->transform(mem);
2789 adr = phase->transform(new AddPNode(base,adr,off));
2790 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2791 }
2792 return mem;
2793 }
2794
2795 //----------------------------step_through----------------------------------
2796 // Return allocation input memory edge if it is different instance
2797 // or itself if it is the one we are looking for.
2798 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2799 Node* n = *np;
2800 assert(n->is_ClearArray(), "sanity");
2801 intptr_t offset;
2802 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2803 // This method is called only before Allocate nodes are expanded
2804 // during macro nodes expansion. Before that ClearArray nodes are
2805 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2806 // Allocate nodes are expanded) which follows allocations.
2807 assert(alloc != NULL, "should have allocation");
2808 if (alloc->_idx == instance_id) {
2809 // Can not bypass initialization of the instance we are looking for.
2810 return false;
2811 }
2812 // Otherwise skip it.
2813 InitializeNode* init = alloc->initialization();
2814 if (init != NULL)
2815 *np = init->in(TypeFunc::Memory);
2816 else
2817 *np = alloc->in(TypeFunc::Memory);
2818 return true;
2819 }
2820
2821 //----------------------------clear_memory-------------------------------------
2822 // Generate code to initialize object storage to zero.
2823 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2824 intptr_t start_offset,
2825 Node* end_offset,
2826 PhaseGVN* phase) {
2827 intptr_t offset = start_offset;
2828
2829 int unit = BytesPerLong;
2830 if ((offset % unit) != 0) {
2831 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
2832 adr = phase->transform(adr);
2833 const TypePtr* atp = TypeRawPtr::BOTTOM;
2834 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2835 mem = phase->transform(mem);
2836 offset += BytesPerInt;
2837 }
2838 assert((offset % unit) == 0, "");
2839
2840 // Initialize the remaining stuff, if any, with a ClearArray.
2841 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2842 }
2843
2844 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2845 Node* start_offset,
2846 Node* end_offset,
2847 PhaseGVN* phase) {
2848 if (start_offset == end_offset) {
2849 // nothing to do
2850 return mem;
2851 }
2852
2853 int unit = BytesPerLong;
2854 Node* zbase = start_offset;
2855 Node* zend = end_offset;
2856
2857 // Scale to the unit required by the CPU:
2858 if (!Matcher::init_array_count_is_in_bytes) {
2859 Node* shift = phase->intcon(exact_log2(unit));
2860 zbase = phase->transform(new URShiftXNode(zbase, shift) );
2861 zend = phase->transform(new URShiftXNode(zend, shift) );
2862 }
2863
2864 // Bulk clear double-words
2865 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
2866 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
2867 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
2868 return phase->transform(mem);
2869 }
2870
2871 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2872 intptr_t start_offset,
2873 intptr_t end_offset,
2874 PhaseGVN* phase) {
2875 if (start_offset == end_offset) {
2876 // nothing to do
2877 return mem;
2878 }
2879
2880 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2881 intptr_t done_offset = end_offset;
2882 if ((done_offset % BytesPerLong) != 0) {
2883 done_offset -= BytesPerInt;
2884 }
2885 if (done_offset > start_offset) {
2886 mem = clear_memory(ctl, mem, dest,
2887 start_offset, phase->MakeConX(done_offset), phase);
2888 }
2889 if (done_offset < end_offset) { // emit the final 32-bit store
2890 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
2891 adr = phase->transform(adr);
2892 const TypePtr* atp = TypeRawPtr::BOTTOM;
2893 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2894 mem = phase->transform(mem);
2895 done_offset += BytesPerInt;
2896 }
2897 assert(done_offset == end_offset, "");
2898 return mem;
2899 }
2900
2901 //=============================================================================
2902 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2903 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2904 _adr_type(C->get_adr_type(alias_idx))
2905 {
2906 init_class_id(Class_MemBar);
2907 Node* top = C->top();
2908 init_req(TypeFunc::I_O,top);
2909 init_req(TypeFunc::FramePtr,top);
2910 init_req(TypeFunc::ReturnAdr,top);
2911 if (precedent != NULL)
2912 init_req(TypeFunc::Parms, precedent);
2913 }
2914
2915 //------------------------------cmp--------------------------------------------
2916 uint MemBarNode::hash() const { return NO_HASH; }
2917 uint MemBarNode::cmp( const Node &n ) const {
2918 return (&n == this); // Always fail except on self
2919 }
2920
2921 //------------------------------make-------------------------------------------
2922 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2923 switch (opcode) {
2924 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
2925 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
2926 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
2927 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
2928 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
2929 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
2930 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
2931 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
2932 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
2933 case Op_Initialize: return new InitializeNode(C, atp, pn);
2934 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
2935 default: ShouldNotReachHere(); return NULL;
2936 }
2937 }
2938
2939 //------------------------------Ideal------------------------------------------
2940 // Return a node which is more "ideal" than the current node. Strip out
2941 // control copies
2942 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2943 if (remove_dead_region(phase, can_reshape)) return this;
2944 // Don't bother trying to transform a dead node
2945 if (in(0) && in(0)->is_top()) {
2946 return NULL;
2947 }
2948
2949 bool progress = false;
2950 // Eliminate volatile MemBars for scalar replaced objects.
2951 if (can_reshape && req() == (Precedent+1)) {
2952 bool eliminate = false;
2953 int opc = Opcode();
2954 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
2955 // Volatile field loads and stores.
2956 Node* my_mem = in(MemBarNode::Precedent);
2957 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
2958 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
2959 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
2960 // replace this Precedent (decodeN) with the Load instead.
2961 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
2962 Node* load_node = my_mem->in(1);
2963 set_req(MemBarNode::Precedent, load_node);
2964 phase->is_IterGVN()->_worklist.push(my_mem);
2965 my_mem = load_node;
2966 } else {
2967 assert(my_mem->unique_out() == this, "sanity");
2968 del_req(Precedent);
2969 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
2970 my_mem = NULL;
2971 }
2972 progress = true;
2973 }
2974 if (my_mem != NULL && my_mem->is_Mem()) {
2975 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2976 // Check for scalar replaced object reference.
2977 if( t_oop != NULL && t_oop->is_known_instance_field() &&
2978 t_oop->offset() != Type::OffsetBot &&
2979 t_oop->offset() != Type::OffsetTop) {
2980 eliminate = true;
2981 }
2982 }
2983 } else if (opc == Op_MemBarRelease) {
2984 // Final field stores.
2985 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
2986 if ((alloc != NULL) && alloc->is_Allocate() &&
2987 alloc->as_Allocate()->does_not_escape_thread()) {
2988 // The allocated object does not escape.
2989 eliminate = true;
2990 }
2991 }
2992 if (eliminate) {
2993 // Replace MemBar projections by its inputs.
2994 PhaseIterGVN* igvn = phase->is_IterGVN();
2995 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2996 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2997 // Must return either the original node (now dead) or a new node
2998 // (Do not return a top here, since that would break the uniqueness of top.)
2999 return new ConINode(TypeInt::ZERO);
3000 }
3001 }
3002 return progress ? this : NULL;
3003 }
3004
3005 //------------------------------Value------------------------------------------
3006 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3007 if( !in(0) ) return Type::TOP;
3008 if( phase->type(in(0)) == Type::TOP )
3009 return Type::TOP;
3010 return TypeTuple::MEMBAR;
3011 }
3012
3013 //------------------------------match------------------------------------------
3014 // Construct projections for memory.
3015 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3016 switch (proj->_con) {
3017 case TypeFunc::Control:
3018 case TypeFunc::Memory:
3019 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3020 }
3021 ShouldNotReachHere();
3022 return NULL;
3023 }
3024
3025 //===========================InitializeNode====================================
3026 // SUMMARY:
3027 // This node acts as a memory barrier on raw memory, after some raw stores.
3028 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3029 // The Initialize can 'capture' suitably constrained stores as raw inits.
3030 // It can coalesce related raw stores into larger units (called 'tiles').
3031 // It can avoid zeroing new storage for memory units which have raw inits.
3032 // At macro-expansion, it is marked 'complete', and does not optimize further.
3033 //
3034 // EXAMPLE:
3035 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3036 // ctl = incoming control; mem* = incoming memory
3037 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3038 // First allocate uninitialized memory and fill in the header:
3039 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3040 // ctl := alloc.Control; mem* := alloc.Memory*
3041 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3042 // Then initialize to zero the non-header parts of the raw memory block:
3043 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3044 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3045 // After the initialize node executes, the object is ready for service:
3046 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3047 // Suppose its body is immediately initialized as {1,2}:
3048 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3049 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3050 // mem.SLICE(#short[*]) := store2
3051 //
3052 // DETAILS:
3053 // An InitializeNode collects and isolates object initialization after
3054 // an AllocateNode and before the next possible safepoint. As a
3055 // memory barrier (MemBarNode), it keeps critical stores from drifting
3056 // down past any safepoint or any publication of the allocation.
3057 // Before this barrier, a newly-allocated object may have uninitialized bits.
3058 // After this barrier, it may be treated as a real oop, and GC is allowed.
3059 //
3060 // The semantics of the InitializeNode include an implicit zeroing of
3061 // the new object from object header to the end of the object.
3062 // (The object header and end are determined by the AllocateNode.)
3063 //
3064 // Certain stores may be added as direct inputs to the InitializeNode.
3065 // These stores must update raw memory, and they must be to addresses
3066 // derived from the raw address produced by AllocateNode, and with
3067 // a constant offset. They must be ordered by increasing offset.
3068 // The first one is at in(RawStores), the last at in(req()-1).
3069 // Unlike most memory operations, they are not linked in a chain,
3070 // but are displayed in parallel as users of the rawmem output of
3071 // the allocation.
3072 //
3073 // (See comments in InitializeNode::capture_store, which continue
3074 // the example given above.)
3075 //
3076 // When the associated Allocate is macro-expanded, the InitializeNode
3077 // may be rewritten to optimize collected stores. A ClearArrayNode
3078 // may also be created at that point to represent any required zeroing.
3079 // The InitializeNode is then marked 'complete', prohibiting further
3080 // capturing of nearby memory operations.
3081 //
3082 // During macro-expansion, all captured initializations which store
3083 // constant values of 32 bits or smaller are coalesced (if advantageous)
3084 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3085 // initialized in fewer memory operations. Memory words which are
3086 // covered by neither tiles nor non-constant stores are pre-zeroed
3087 // by explicit stores of zero. (The code shape happens to do all
3088 // zeroing first, then all other stores, with both sequences occurring
3089 // in order of ascending offsets.)
3090 //
3091 // Alternatively, code may be inserted between an AllocateNode and its
3092 // InitializeNode, to perform arbitrary initialization of the new object.
3093 // E.g., the object copying intrinsics insert complex data transfers here.
3094 // The initialization must then be marked as 'complete' disable the
3095 // built-in zeroing semantics and the collection of initializing stores.
3096 //
3097 // While an InitializeNode is incomplete, reads from the memory state
3098 // produced by it are optimizable if they match the control edge and
3099 // new oop address associated with the allocation/initialization.
3100 // They return a stored value (if the offset matches) or else zero.
3101 // A write to the memory state, if it matches control and address,
3102 // and if it is to a constant offset, may be 'captured' by the
3103 // InitializeNode. It is cloned as a raw memory operation and rewired
3104 // inside the initialization, to the raw oop produced by the allocation.
3105 // Operations on addresses which are provably distinct (e.g., to
3106 // other AllocateNodes) are allowed to bypass the initialization.
3107 //
3108 // The effect of all this is to consolidate object initialization
3109 // (both arrays and non-arrays, both piecewise and bulk) into a
3110 // single location, where it can be optimized as a unit.
3111 //
3112 // Only stores with an offset less than TrackedInitializationLimit words
3113 // will be considered for capture by an InitializeNode. This puts a
3114 // reasonable limit on the complexity of optimized initializations.
3115
3116 //---------------------------InitializeNode------------------------------------
3117 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3118 : _is_complete(Incomplete), _does_not_escape(false),
3119 MemBarNode(C, adr_type, rawoop)
3120 {
3121 init_class_id(Class_Initialize);
3122
3123 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3124 assert(in(RawAddress) == rawoop, "proper init");
3125 // Note: allocation() can be NULL, for secondary initialization barriers
3126 }
3127
3128 // Since this node is not matched, it will be processed by the
3129 // register allocator. Declare that there are no constraints
3130 // on the allocation of the RawAddress edge.
3131 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3132 // This edge should be set to top, by the set_complete. But be conservative.
3133 if (idx == InitializeNode::RawAddress)
3134 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3135 return RegMask::Empty;
3136 }
3137
3138 Node* InitializeNode::memory(uint alias_idx) {
3139 Node* mem = in(Memory);
3140 if (mem->is_MergeMem()) {
3141 return mem->as_MergeMem()->memory_at(alias_idx);
3142 } else {
3143 // incoming raw memory is not split
3144 return mem;
3145 }
3146 }
3147
3148 bool InitializeNode::is_non_zero() {
3149 if (is_complete()) return false;
3150 remove_extra_zeroes();
3151 return (req() > RawStores);
3152 }
3153
3154 void InitializeNode::set_complete(PhaseGVN* phase) {
3155 assert(!is_complete(), "caller responsibility");
3156 _is_complete = Complete;
3157
3158 // After this node is complete, it contains a bunch of
3159 // raw-memory initializations. There is no need for
3160 // it to have anything to do with non-raw memory effects.
3161 // Therefore, tell all non-raw users to re-optimize themselves,
3162 // after skipping the memory effects of this initialization.
3163 PhaseIterGVN* igvn = phase->is_IterGVN();
3164 if (igvn) igvn->add_users_to_worklist(this);
3165 }
3166
3167 // convenience function
3168 // return false if the init contains any stores already
3169 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3170 InitializeNode* init = initialization();
3171 if (init == NULL || init->is_complete()) return false;
3172 init->remove_extra_zeroes();
3173 // for now, if this allocation has already collected any inits, bail:
3174 if (init->is_non_zero()) return false;
3175 init->set_complete(phase);
3176 return true;
3177 }
3178
3179 void InitializeNode::remove_extra_zeroes() {
3180 if (req() == RawStores) return;
3181 Node* zmem = zero_memory();
3182 uint fill = RawStores;
3183 for (uint i = fill; i < req(); i++) {
3184 Node* n = in(i);
3185 if (n->is_top() || n == zmem) continue; // skip
3186 if (fill < i) set_req(fill, n); // compact
3187 ++fill;
3188 }
3189 // delete any empty spaces created:
3190 while (fill < req()) {
3191 del_req(fill);
3192 }
3193 }
3194
3195 // Helper for remembering which stores go with which offsets.
3196 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3197 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3198 intptr_t offset = -1;
3199 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3200 phase, offset);
3201 if (base == NULL) return -1; // something is dead,
3202 if (offset < 0) return -1; // dead, dead
3203 return offset;
3204 }
3205
3206 // Helper for proving that an initialization expression is
3207 // "simple enough" to be folded into an object initialization.
3208 // Attempts to prove that a store's initial value 'n' can be captured
3209 // within the initialization without creating a vicious cycle, such as:
3210 // { Foo p = new Foo(); p.next = p; }
3211 // True for constants and parameters and small combinations thereof.
3212 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3213 if (n == NULL) return true; // (can this really happen?)
3214 if (n->is_Proj()) n = n->in(0);
3215 if (n == this) return false; // found a cycle
3216 if (n->is_Con()) return true;
3217 if (n->is_Start()) return true; // params, etc., are OK
3218 if (n->is_Root()) return true; // even better
3219
3220 Node* ctl = n->in(0);
3221 if (ctl != NULL && !ctl->is_top()) {
3222 if (ctl->is_Proj()) ctl = ctl->in(0);
3223 if (ctl == this) return false;
3224
3225 // If we already know that the enclosing memory op is pinned right after
3226 // the init, then any control flow that the store has picked up
3227 // must have preceded the init, or else be equal to the init.
3228 // Even after loop optimizations (which might change control edges)
3229 // a store is never pinned *before* the availability of its inputs.
3230 if (!MemNode::all_controls_dominate(n, this))
3231 return false; // failed to prove a good control
3232 }
3233
3234 // Check data edges for possible dependencies on 'this'.
3235 if ((count += 1) > 20) return false; // complexity limit
3236 for (uint i = 1; i < n->req(); i++) {
3237 Node* m = n->in(i);
3238 if (m == NULL || m == n || m->is_top()) continue;
3239 uint first_i = n->find_edge(m);
3240 if (i != first_i) continue; // process duplicate edge just once
3241 if (!detect_init_independence(m, count)) {
3242 return false;
3243 }
3244 }
3245
3246 return true;
3247 }
3248
3249 // Here are all the checks a Store must pass before it can be moved into
3250 // an initialization. Returns zero if a check fails.
3251 // On success, returns the (constant) offset to which the store applies,
3252 // within the initialized memory.
3253 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3254 const int FAIL = 0;
3255 if (st->is_unaligned_access()) {
3256 return FAIL;
3257 }
3258 if (st->req() != MemNode::ValueIn + 1)
3259 return FAIL; // an inscrutable StoreNode (card mark?)
3260 Node* ctl = st->in(MemNode::Control);
3261 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3262 return FAIL; // must be unconditional after the initialization
3263 Node* mem = st->in(MemNode::Memory);
3264 if (!(mem->is_Proj() && mem->in(0) == this))
3265 return FAIL; // must not be preceded by other stores
3266 Node* adr = st->in(MemNode::Address);
3267 intptr_t offset;
3268 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3269 if (alloc == NULL)
3270 return FAIL; // inscrutable address
3271 if (alloc != allocation())
3272 return FAIL; // wrong allocation! (store needs to float up)
3273 Node* val = st->in(MemNode::ValueIn);
3274 int complexity_count = 0;
3275 if (!detect_init_independence(val, complexity_count))
3276 return FAIL; // stored value must be 'simple enough'
3277
3278 // The Store can be captured only if nothing after the allocation
3279 // and before the Store is using the memory location that the store
3280 // overwrites.
3281 bool failed = false;
3282 // If is_complete_with_arraycopy() is true the shape of the graph is
3283 // well defined and is safe so no need for extra checks.
3284 if (!is_complete_with_arraycopy()) {
3285 // We are going to look at each use of the memory state following
3286 // the allocation to make sure nothing reads the memory that the
3287 // Store writes.
3288 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3289 int alias_idx = phase->C->get_alias_index(t_adr);
3290 ResourceMark rm;
3291 Unique_Node_List mems;
3292 mems.push(mem);
3293 Node* unique_merge = NULL;
3294 for (uint next = 0; next < mems.size(); ++next) {
3295 Node *m = mems.at(next);
3296 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3297 Node *n = m->fast_out(j);
3298 if (n->outcnt() == 0) {
3299 continue;
3300 }
3301 if (n == st) {
3302 continue;
3303 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3304 // If the control of this use is different from the control
3305 // of the Store which is right after the InitializeNode then
3306 // this node cannot be between the InitializeNode and the
3307 // Store.
3308 continue;
3309 } else if (n->is_MergeMem()) {
3310 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3311 // We can hit a MergeMemNode (that will likely go away
3312 // later) that is a direct use of the memory state
3313 // following the InitializeNode on the same slice as the
3314 // store node that we'd like to capture. We need to check
3315 // the uses of the MergeMemNode.
3316 mems.push(n);
3317 }
3318 } else if (n->is_Mem()) {
3319 Node* other_adr = n->in(MemNode::Address);
3320 if (other_adr == adr) {
3321 failed = true;
3322 break;
3323 } else {
3324 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3325 if (other_t_adr != NULL) {
3326 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3327 if (other_alias_idx == alias_idx) {
3328 // A load from the same memory slice as the store right
3329 // after the InitializeNode. We check the control of the
3330 // object/array that is loaded from. If it's the same as
3331 // the store control then we cannot capture the store.
3332 assert(!n->is_Store(), "2 stores to same slice on same control?");
3333 Node* base = other_adr;
3334 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3335 base = base->in(AddPNode::Base);
3336 if (base != NULL) {
3337 base = base->uncast();
3338 if (base->is_Proj() && base->in(0) == alloc) {
3339 failed = true;
3340 break;
3341 }
3342 }
3343 }
3344 }
3345 }
3346 } else {
3347 failed = true;
3348 break;
3349 }
3350 }
3351 }
3352 }
3353 if (failed) {
3354 if (!can_reshape) {
3355 // We decided we couldn't capture the store during parsing. We
3356 // should try again during the next IGVN once the graph is
3357 // cleaner.
3358 phase->C->record_for_igvn(st);
3359 }
3360 return FAIL;
3361 }
3362
3363 return offset; // success
3364 }
3365
3366 // Find the captured store in(i) which corresponds to the range
3367 // [start..start+size) in the initialized object.
3368 // If there is one, return its index i. If there isn't, return the
3369 // negative of the index where it should be inserted.
3370 // Return 0 if the queried range overlaps an initialization boundary
3371 // or if dead code is encountered.
3372 // If size_in_bytes is zero, do not bother with overlap checks.
3373 int InitializeNode::captured_store_insertion_point(intptr_t start,
3374 int size_in_bytes,
3375 PhaseTransform* phase) {
3376 const int FAIL = 0, MAX_STORE = BytesPerLong;
3377
3378 if (is_complete())
3379 return FAIL; // arraycopy got here first; punt
3380
3381 assert(allocation() != NULL, "must be present");
3382
3383 // no negatives, no header fields:
3384 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3385
3386 // after a certain size, we bail out on tracking all the stores:
3387 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3388 if (start >= ti_limit) return FAIL;
3389
3390 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3391 if (i >= limit) return -(int)i; // not found; here is where to put it
3392
3393 Node* st = in(i);
3394 intptr_t st_off = get_store_offset(st, phase);
3395 if (st_off < 0) {
3396 if (st != zero_memory()) {
3397 return FAIL; // bail out if there is dead garbage
3398 }
3399 } else if (st_off > start) {
3400 // ...we are done, since stores are ordered
3401 if (st_off < start + size_in_bytes) {
3402 return FAIL; // the next store overlaps
3403 }
3404 return -(int)i; // not found; here is where to put it
3405 } else if (st_off < start) {
3406 if (size_in_bytes != 0 &&
3407 start < st_off + MAX_STORE &&
3408 start < st_off + st->as_Store()->memory_size()) {
3409 return FAIL; // the previous store overlaps
3410 }
3411 } else {
3412 if (size_in_bytes != 0 &&
3413 st->as_Store()->memory_size() != size_in_bytes) {
3414 return FAIL; // mismatched store size
3415 }
3416 return i;
3417 }
3418
3419 ++i;
3420 }
3421 }
3422
3423 // Look for a captured store which initializes at the offset 'start'
3424 // with the given size. If there is no such store, and no other
3425 // initialization interferes, then return zero_memory (the memory
3426 // projection of the AllocateNode).
3427 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3428 PhaseTransform* phase) {
3429 assert(stores_are_sane(phase), "");
3430 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3431 if (i == 0) {
3432 return NULL; // something is dead
3433 } else if (i < 0) {
3434 return zero_memory(); // just primordial zero bits here
3435 } else {
3436 Node* st = in(i); // here is the store at this position
3437 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3438 return st;
3439 }
3440 }
3441
3442 // Create, as a raw pointer, an address within my new object at 'offset'.
3443 Node* InitializeNode::make_raw_address(intptr_t offset,
3444 PhaseTransform* phase) {
3445 Node* addr = in(RawAddress);
3446 if (offset != 0) {
3447 Compile* C = phase->C;
3448 addr = phase->transform( new AddPNode(C->top(), addr,
3449 phase->MakeConX(offset)) );
3450 }
3451 return addr;
3452 }
3453
3454 // Clone the given store, converting it into a raw store
3455 // initializing a field or element of my new object.
3456 // Caller is responsible for retiring the original store,
3457 // with subsume_node or the like.
3458 //
3459 // From the example above InitializeNode::InitializeNode,
3460 // here are the old stores to be captured:
3461 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3462 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3463 //
3464 // Here is the changed code; note the extra edges on init:
3465 // alloc = (Allocate ...)
3466 // rawoop = alloc.RawAddress
3467 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3468 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3469 // init = (Initialize alloc.Control alloc.Memory rawoop
3470 // rawstore1 rawstore2)
3471 //
3472 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3473 PhaseTransform* phase, bool can_reshape) {
3474 assert(stores_are_sane(phase), "");
3475
3476 if (start < 0) return NULL;
3477 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3478
3479 Compile* C = phase->C;
3480 int size_in_bytes = st->memory_size();
3481 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3482 if (i == 0) return NULL; // bail out
3483 Node* prev_mem = NULL; // raw memory for the captured store
3484 if (i > 0) {
3485 prev_mem = in(i); // there is a pre-existing store under this one
3486 set_req(i, C->top()); // temporarily disconnect it
3487 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3488 } else {
3489 i = -i; // no pre-existing store
3490 prev_mem = zero_memory(); // a slice of the newly allocated object
3491 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3492 set_req(--i, C->top()); // reuse this edge; it has been folded away
3493 else
3494 ins_req(i, C->top()); // build a new edge
3495 }
3496 Node* new_st = st->clone();
3497 new_st->set_req(MemNode::Control, in(Control));
3498 new_st->set_req(MemNode::Memory, prev_mem);
3499 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3500 new_st = phase->transform(new_st);
3501
3502 // At this point, new_st might have swallowed a pre-existing store
3503 // at the same offset, or perhaps new_st might have disappeared,
3504 // if it redundantly stored the same value (or zero to fresh memory).
3505
3506 // In any case, wire it in:
3507 phase->igvn_rehash_node_delayed(this);
3508 set_req(i, new_st);
3509
3510 // The caller may now kill the old guy.
3511 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3512 assert(check_st == new_st || check_st == NULL, "must be findable");
3513 assert(!is_complete(), "");
3514 return new_st;
3515 }
3516
3517 static bool store_constant(jlong* tiles, int num_tiles,
3518 intptr_t st_off, int st_size,
3519 jlong con) {
3520 if ((st_off & (st_size-1)) != 0)
3521 return false; // strange store offset (assume size==2**N)
3522 address addr = (address)tiles + st_off;
3523 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3524 switch (st_size) {
3525 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3526 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3527 case sizeof(jint): *(jint*) addr = (jint) con; break;
3528 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3529 default: return false; // strange store size (detect size!=2**N here)
3530 }
3531 return true; // return success to caller
3532 }
3533
3534 // Coalesce subword constants into int constants and possibly
3535 // into long constants. The goal, if the CPU permits,
3536 // is to initialize the object with a small number of 64-bit tiles.
3537 // Also, convert floating-point constants to bit patterns.
3538 // Non-constants are not relevant to this pass.
3539 //
3540 // In terms of the running example on InitializeNode::InitializeNode
3541 // and InitializeNode::capture_store, here is the transformation
3542 // of rawstore1 and rawstore2 into rawstore12:
3543 // alloc = (Allocate ...)
3544 // rawoop = alloc.RawAddress
3545 // tile12 = 0x00010002
3546 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3547 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3548 //
3549 void
3550 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3551 Node* size_in_bytes,
3552 PhaseGVN* phase) {
3553 Compile* C = phase->C;
3554
3555 assert(stores_are_sane(phase), "");
3556 // Note: After this pass, they are not completely sane,
3557 // since there may be some overlaps.
3558
3559 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3560
3561 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3562 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3563 size_limit = MIN2(size_limit, ti_limit);
3564 size_limit = align_up(size_limit, BytesPerLong);
3565 int num_tiles = size_limit / BytesPerLong;
3566
3567 // allocate space for the tile map:
3568 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3569 jlong tiles_buf[small_len];
3570 Node* nodes_buf[small_len];
3571 jlong inits_buf[small_len];
3572 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3573 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3574 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3575 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3576 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3577 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3578 // tiles: exact bitwise model of all primitive constants
3579 // nodes: last constant-storing node subsumed into the tiles model
3580 // inits: which bytes (in each tile) are touched by any initializations
3581
3582 //// Pass A: Fill in the tile model with any relevant stores.
3583
3584 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3585 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3586 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3587 Node* zmem = zero_memory(); // initially zero memory state
3588 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3589 Node* st = in(i);
3590 intptr_t st_off = get_store_offset(st, phase);
3591
3592 // Figure out the store's offset and constant value:
3593 if (st_off < header_size) continue; //skip (ignore header)
3594 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3595 int st_size = st->as_Store()->memory_size();
3596 if (st_off + st_size > size_limit) break;
3597
3598 // Record which bytes are touched, whether by constant or not.
3599 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3600 continue; // skip (strange store size)
3601
3602 const Type* val = phase->type(st->in(MemNode::ValueIn));
3603 if (!val->singleton()) continue; //skip (non-con store)
3604 BasicType type = val->basic_type();
3605
3606 jlong con = 0;
3607 switch (type) {
3608 case T_INT: con = val->is_int()->get_con(); break;
3609 case T_LONG: con = val->is_long()->get_con(); break;
3610 case T_FLOAT: con = jint_cast(val->getf()); break;
3611 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3612 default: continue; //skip (odd store type)
3613 }
3614
3615 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3616 st->Opcode() == Op_StoreL) {
3617 continue; // This StoreL is already optimal.
3618 }
3619
3620 // Store down the constant.
3621 store_constant(tiles, num_tiles, st_off, st_size, con);
3622
3623 intptr_t j = st_off >> LogBytesPerLong;
3624
3625 if (type == T_INT && st_size == BytesPerInt
3626 && (st_off & BytesPerInt) == BytesPerInt) {
3627 jlong lcon = tiles[j];
3628 if (!Matcher::isSimpleConstant64(lcon) &&
3629 st->Opcode() == Op_StoreI) {
3630 // This StoreI is already optimal by itself.
3631 jint* intcon = (jint*) &tiles[j];
3632 intcon[1] = 0; // undo the store_constant()
3633
3634 // If the previous store is also optimal by itself, back up and
3635 // undo the action of the previous loop iteration... if we can.
3636 // But if we can't, just let the previous half take care of itself.
3637 st = nodes[j];
3638 st_off -= BytesPerInt;
3639 con = intcon[0];
3640 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3641 assert(st_off >= header_size, "still ignoring header");
3642 assert(get_store_offset(st, phase) == st_off, "must be");
3643 assert(in(i-1) == zmem, "must be");
3644 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3645 assert(con == tcon->is_int()->get_con(), "must be");
3646 // Undo the effects of the previous loop trip, which swallowed st:
3647 intcon[0] = 0; // undo store_constant()
3648 set_req(i-1, st); // undo set_req(i, zmem)
3649 nodes[j] = NULL; // undo nodes[j] = st
3650 --old_subword; // undo ++old_subword
3651 }
3652 continue; // This StoreI is already optimal.
3653 }
3654 }
3655
3656 // This store is not needed.
3657 set_req(i, zmem);
3658 nodes[j] = st; // record for the moment
3659 if (st_size < BytesPerLong) // something has changed
3660 ++old_subword; // includes int/float, but who's counting...
3661 else ++old_long;
3662 }
3663
3664 if ((old_subword + old_long) == 0)
3665 return; // nothing more to do
3666
3667 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3668 // Be sure to insert them before overlapping non-constant stores.
3669 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3670 for (int j = 0; j < num_tiles; j++) {
3671 jlong con = tiles[j];
3672 jlong init = inits[j];
3673 if (con == 0) continue;
3674 jint con0, con1; // split the constant, address-wise
3675 jint init0, init1; // split the init map, address-wise
3676 { union { jlong con; jint intcon[2]; } u;
3677 u.con = con;
3678 con0 = u.intcon[0];
3679 con1 = u.intcon[1];
3680 u.con = init;
3681 init0 = u.intcon[0];
3682 init1 = u.intcon[1];
3683 }
3684
3685 Node* old = nodes[j];
3686 assert(old != NULL, "need the prior store");
3687 intptr_t offset = (j * BytesPerLong);
3688
3689 bool split = !Matcher::isSimpleConstant64(con);
3690
3691 if (offset < header_size) {
3692 assert(offset + BytesPerInt >= header_size, "second int counts");
3693 assert(*(jint*)&tiles[j] == 0, "junk in header");
3694 split = true; // only the second word counts
3695 // Example: int a[] = { 42 ... }
3696 } else if (con0 == 0 && init0 == -1) {
3697 split = true; // first word is covered by full inits
3698 // Example: int a[] = { ... foo(), 42 ... }
3699 } else if (con1 == 0 && init1 == -1) {
3700 split = true; // second word is covered by full inits
3701 // Example: int a[] = { ... 42, foo() ... }
3702 }
3703
3704 // Here's a case where init0 is neither 0 nor -1:
3705 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3706 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3707 // In this case the tile is not split; it is (jlong)42.
3708 // The big tile is stored down, and then the foo() value is inserted.
3709 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3710
3711 Node* ctl = old->in(MemNode::Control);
3712 Node* adr = make_raw_address(offset, phase);
3713 const TypePtr* atp = TypeRawPtr::BOTTOM;
3714
3715 // One or two coalesced stores to plop down.
3716 Node* st[2];
3717 intptr_t off[2];
3718 int nst = 0;
3719 if (!split) {
3720 ++new_long;
3721 off[nst] = offset;
3722 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3723 phase->longcon(con), T_LONG, MemNode::unordered);
3724 } else {
3725 // Omit either if it is a zero.
3726 if (con0 != 0) {
3727 ++new_int;
3728 off[nst] = offset;
3729 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3730 phase->intcon(con0), T_INT, MemNode::unordered);
3731 }
3732 if (con1 != 0) {
3733 ++new_int;
3734 offset += BytesPerInt;
3735 adr = make_raw_address(offset, phase);
3736 off[nst] = offset;
3737 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3738 phase->intcon(con1), T_INT, MemNode::unordered);
3739 }
3740 }
3741
3742 // Insert second store first, then the first before the second.
3743 // Insert each one just before any overlapping non-constant stores.
3744 while (nst > 0) {
3745 Node* st1 = st[--nst];
3746 C->copy_node_notes_to(st1, old);
3747 st1 = phase->transform(st1);
3748 offset = off[nst];
3749 assert(offset >= header_size, "do not smash header");
3750 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3751 guarantee(ins_idx != 0, "must re-insert constant store");
3752 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3753 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3754 set_req(--ins_idx, st1);
3755 else
3756 ins_req(ins_idx, st1);
3757 }
3758 }
3759
3760 if (PrintCompilation && WizardMode)
3761 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3762 old_subword, old_long, new_int, new_long);
3763 if (C->log() != NULL)
3764 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3765 old_subword, old_long, new_int, new_long);
3766
3767 // Clean up any remaining occurrences of zmem:
3768 remove_extra_zeroes();
3769 }
3770
3771 // Explore forward from in(start) to find the first fully initialized
3772 // word, and return its offset. Skip groups of subword stores which
3773 // together initialize full words. If in(start) is itself part of a
3774 // fully initialized word, return the offset of in(start). If there
3775 // are no following full-word stores, or if something is fishy, return
3776 // a negative value.
3777 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3778 int int_map = 0;
3779 intptr_t int_map_off = 0;
3780 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3781
3782 for (uint i = start, limit = req(); i < limit; i++) {
3783 Node* st = in(i);
3784
3785 intptr_t st_off = get_store_offset(st, phase);
3786 if (st_off < 0) break; // return conservative answer
3787
3788 int st_size = st->as_Store()->memory_size();
3789 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3790 return st_off; // we found a complete word init
3791 }
3792
3793 // update the map:
3794
3795 intptr_t this_int_off = align_down(st_off, BytesPerInt);
3796 if (this_int_off != int_map_off) {
3797 // reset the map:
3798 int_map = 0;
3799 int_map_off = this_int_off;
3800 }
3801
3802 int subword_off = st_off - this_int_off;
3803 int_map |= right_n_bits(st_size) << subword_off;
3804 if ((int_map & FULL_MAP) == FULL_MAP) {
3805 return this_int_off; // we found a complete word init
3806 }
3807
3808 // Did this store hit or cross the word boundary?
3809 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
3810 if (next_int_off == this_int_off + BytesPerInt) {
3811 // We passed the current int, without fully initializing it.
3812 int_map_off = next_int_off;
3813 int_map >>= BytesPerInt;
3814 } else if (next_int_off > this_int_off + BytesPerInt) {
3815 // We passed the current and next int.
3816 return this_int_off + BytesPerInt;
3817 }
3818 }
3819
3820 return -1;
3821 }
3822
3823
3824 // Called when the associated AllocateNode is expanded into CFG.
3825 // At this point, we may perform additional optimizations.
3826 // Linearize the stores by ascending offset, to make memory
3827 // activity as coherent as possible.
3828 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3829 intptr_t header_size,
3830 Node* size_in_bytes,
3831 PhaseGVN* phase) {
3832 assert(!is_complete(), "not already complete");
3833 assert(stores_are_sane(phase), "");
3834 assert(allocation() != NULL, "must be present");
3835
3836 remove_extra_zeroes();
3837
3838 if (ReduceFieldZeroing || ReduceBulkZeroing)
3839 // reduce instruction count for common initialization patterns
3840 coalesce_subword_stores(header_size, size_in_bytes, phase);
3841
3842 Node* zmem = zero_memory(); // initially zero memory state
3843 Node* inits = zmem; // accumulating a linearized chain of inits
3844 #ifdef ASSERT
3845 intptr_t first_offset = allocation()->minimum_header_size();
3846 intptr_t last_init_off = first_offset; // previous init offset
3847 intptr_t last_init_end = first_offset; // previous init offset+size
3848 intptr_t last_tile_end = first_offset; // previous tile offset+size
3849 #endif
3850 intptr_t zeroes_done = header_size;
3851
3852 bool do_zeroing = true; // we might give up if inits are very sparse
3853 int big_init_gaps = 0; // how many large gaps have we seen?
3854
3855 if (UseTLAB && ZeroTLAB) do_zeroing = false;
3856 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3857
3858 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3859 Node* st = in(i);
3860 intptr_t st_off = get_store_offset(st, phase);
3861 if (st_off < 0)
3862 break; // unknown junk in the inits
3863 if (st->in(MemNode::Memory) != zmem)
3864 break; // complicated store chains somehow in list
3865
3866 int st_size = st->as_Store()->memory_size();
3867 intptr_t next_init_off = st_off + st_size;
3868
3869 if (do_zeroing && zeroes_done < next_init_off) {
3870 // See if this store needs a zero before it or under it.
3871 intptr_t zeroes_needed = st_off;
3872
3873 if (st_size < BytesPerInt) {
3874 // Look for subword stores which only partially initialize words.
3875 // If we find some, we must lay down some word-level zeroes first,
3876 // underneath the subword stores.
3877 //
3878 // Examples:
3879 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3880 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3881 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3882 //
3883 // Note: coalesce_subword_stores may have already done this,
3884 // if it was prompted by constant non-zero subword initializers.
3885 // But this case can still arise with non-constant stores.
3886
3887 intptr_t next_full_store = find_next_fullword_store(i, phase);
3888
3889 // In the examples above:
3890 // in(i) p q r s x y z
3891 // st_off 12 13 14 15 12 13 14
3892 // st_size 1 1 1 1 1 1 1
3893 // next_full_s. 12 16 16 16 16 16 16
3894 // z's_done 12 16 16 16 12 16 12
3895 // z's_needed 12 16 16 16 16 16 16
3896 // zsize 0 0 0 0 4 0 4
3897 if (next_full_store < 0) {
3898 // Conservative tack: Zero to end of current word.
3899 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
3900 } else {
3901 // Zero to beginning of next fully initialized word.
3902 // Or, don't zero at all, if we are already in that word.
3903 assert(next_full_store >= zeroes_needed, "must go forward");
3904 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3905 zeroes_needed = next_full_store;
3906 }
3907 }
3908
3909 if (zeroes_needed > zeroes_done) {
3910 intptr_t zsize = zeroes_needed - zeroes_done;
3911 // Do some incremental zeroing on rawmem, in parallel with inits.
3912 zeroes_done = align_down(zeroes_done, BytesPerInt);
3913 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3914 zeroes_done, zeroes_needed,
3915 phase);
3916 zeroes_done = zeroes_needed;
3917 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
3918 do_zeroing = false; // leave the hole, next time
3919 }
3920 }
3921
3922 // Collect the store and move on:
3923 st->set_req(MemNode::Memory, inits);
3924 inits = st; // put it on the linearized chain
3925 set_req(i, zmem); // unhook from previous position
3926
3927 if (zeroes_done == st_off)
3928 zeroes_done = next_init_off;
3929
3930 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3931
3932 #ifdef ASSERT
3933 // Various order invariants. Weaker than stores_are_sane because
3934 // a large constant tile can be filled in by smaller non-constant stores.
3935 assert(st_off >= last_init_off, "inits do not reverse");
3936 last_init_off = st_off;
3937 const Type* val = NULL;
3938 if (st_size >= BytesPerInt &&
3939 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3940 (int)val->basic_type() < (int)T_OBJECT) {
3941 assert(st_off >= last_tile_end, "tiles do not overlap");
3942 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3943 last_tile_end = MAX2(last_tile_end, next_init_off);
3944 } else {
3945 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
3946 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3947 assert(st_off >= last_init_end, "inits do not overlap");
3948 last_init_end = next_init_off; // it's a non-tile
3949 }
3950 #endif //ASSERT
3951 }
3952
3953 remove_extra_zeroes(); // clear out all the zmems left over
3954 add_req(inits);
3955
3956 if (!(UseTLAB && ZeroTLAB)) {
3957 // If anything remains to be zeroed, zero it all now.
3958 zeroes_done = align_down(zeroes_done, BytesPerInt);
3959 // if it is the last unused 4 bytes of an instance, forget about it
3960 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3961 if (zeroes_done + BytesPerLong >= size_limit) {
3962 AllocateNode* alloc = allocation();
3963 assert(alloc != NULL, "must be present");
3964 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
3965 Node* klass_node = alloc->in(AllocateNode::KlassNode);
3966 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3967 if (zeroes_done == k->layout_helper())
3968 zeroes_done = size_limit;
3969 }
3970 }
3971 if (zeroes_done < size_limit) {
3972 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3973 zeroes_done, size_in_bytes, phase);
3974 }
3975 }
3976
3977 set_complete(phase);
3978 return rawmem;
3979 }
3980
3981
3982 #ifdef ASSERT
3983 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3984 if (is_complete())
3985 return true; // stores could be anything at this point
3986 assert(allocation() != NULL, "must be present");
3987 intptr_t last_off = allocation()->minimum_header_size();
3988 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3989 Node* st = in(i);
3990 intptr_t st_off = get_store_offset(st, phase);
3991 if (st_off < 0) continue; // ignore dead garbage
3992 if (last_off > st_off) {
3993 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
3994 this->dump(2);
3995 assert(false, "ascending store offsets");
3996 return false;
3997 }
3998 last_off = st_off + st->as_Store()->memory_size();
3999 }
4000 return true;
4001 }
4002 #endif //ASSERT
4003
4004
4005
4006
4007 //============================MergeMemNode=====================================
4008 //
4009 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4010 // contributing store or call operations. Each contributor provides the memory
4011 // state for a particular "alias type" (see Compile::alias_type). For example,
4012 // if a MergeMem has an input X for alias category #6, then any memory reference
4013 // to alias category #6 may use X as its memory state input, as an exact equivalent
4014 // to using the MergeMem as a whole.
4015 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4016 //
4017 // (Here, the <N> notation gives the index of the relevant adr_type.)
4018 //
4019 // In one special case (and more cases in the future), alias categories overlap.
4020 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4021 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4022 // it is exactly equivalent to that state W:
4023 // MergeMem(<Bot>: W) <==> W
4024 //
4025 // Usually, the merge has more than one input. In that case, where inputs
4026 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4027 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4028 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4029 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4030 //
4031 // A merge can take a "wide" memory state as one of its narrow inputs.
4032 // This simply means that the merge observes out only the relevant parts of
4033 // the wide input. That is, wide memory states arriving at narrow merge inputs
4034 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4035 //
4036 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4037 // and that memory slices "leak through":
4038 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4039 //
4040 // But, in such a cascade, repeated memory slices can "block the leak":
4041 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4042 //
4043 // In the last example, Y is not part of the combined memory state of the
4044 // outermost MergeMem. The system must, of course, prevent unschedulable
4045 // memory states from arising, so you can be sure that the state Y is somehow
4046 // a precursor to state Y'.
4047 //
4048 //
4049 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4050 // of each MergeMemNode array are exactly the numerical alias indexes, including
4051 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4052 // Compile::alias_type (and kin) produce and manage these indexes.
4053 //
4054 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4055 // (Note that this provides quick access to the top node inside MergeMem methods,
4056 // without the need to reach out via TLS to Compile::current.)
4057 //
4058 // As a consequence of what was just described, a MergeMem that represents a full
4059 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4060 // containing all alias categories.
4061 //
4062 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4063 //
4064 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4065 // a memory state for the alias type <N>, or else the top node, meaning that
4066 // there is no particular input for that alias type. Note that the length of
4067 // a MergeMem is variable, and may be extended at any time to accommodate new
4068 // memory states at larger alias indexes. When merges grow, they are of course
4069 // filled with "top" in the unused in() positions.
4070 //
4071 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4072 // (Top was chosen because it works smoothly with passes like GCM.)
4073 //
4074 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4075 // the type of random VM bits like TLS references.) Since it is always the
4076 // first non-Bot memory slice, some low-level loops use it to initialize an
4077 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4078 //
4079 //
4080 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4081 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4082 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4083 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4084 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4085 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4086 //
4087 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4088 // really that different from the other memory inputs. An abbreviation called
4089 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4090 //
4091 //
4092 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4093 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4094 // that "emerges though" the base memory will be marked as excluding the alias types
4095 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4096 //
4097 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4098 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4099 //
4100 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4101 // (It is currently unimplemented.) As you can see, the resulting merge is
4102 // actually a disjoint union of memory states, rather than an overlay.
4103 //
4104
4105 //------------------------------MergeMemNode-----------------------------------
4106 Node* MergeMemNode::make_empty_memory() {
4107 Node* empty_memory = (Node*) Compile::current()->top();
4108 assert(empty_memory->is_top(), "correct sentinel identity");
4109 return empty_memory;
4110 }
4111
4112 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4113 init_class_id(Class_MergeMem);
4114 // all inputs are nullified in Node::Node(int)
4115 // set_input(0, NULL); // no control input
4116
4117 // Initialize the edges uniformly to top, for starters.
4118 Node* empty_mem = make_empty_memory();
4119 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4120 init_req(i,empty_mem);
4121 }
4122 assert(empty_memory() == empty_mem, "");
4123
4124 if( new_base != NULL && new_base->is_MergeMem() ) {
4125 MergeMemNode* mdef = new_base->as_MergeMem();
4126 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4127 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4128 mms.set_memory(mms.memory2());
4129 }
4130 assert(base_memory() == mdef->base_memory(), "");
4131 } else {
4132 set_base_memory(new_base);
4133 }
4134 }
4135
4136 // Make a new, untransformed MergeMem with the same base as 'mem'.
4137 // If mem is itself a MergeMem, populate the result with the same edges.
4138 MergeMemNode* MergeMemNode::make(Node* mem) {
4139 return new MergeMemNode(mem);
4140 }
4141
4142 //------------------------------cmp--------------------------------------------
4143 uint MergeMemNode::hash() const { return NO_HASH; }
4144 uint MergeMemNode::cmp( const Node &n ) const {
4145 return (&n == this); // Always fail except on self
4146 }
4147
4148 //------------------------------Identity---------------------------------------
4149 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4150 // Identity if this merge point does not record any interesting memory
4151 // disambiguations.
4152 Node* base_mem = base_memory();
4153 Node* empty_mem = empty_memory();
4154 if (base_mem != empty_mem) { // Memory path is not dead?
4155 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4156 Node* mem = in(i);
4157 if (mem != empty_mem && mem != base_mem) {
4158 return this; // Many memory splits; no change
4159 }
4160 }
4161 }
4162 return base_mem; // No memory splits; ID on the one true input
4163 }
4164
4165 //------------------------------Ideal------------------------------------------
4166 // This method is invoked recursively on chains of MergeMem nodes
4167 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4168 // Remove chain'd MergeMems
4169 //
4170 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4171 // relative to the "in(Bot)". Since we are patching both at the same time,
4172 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4173 // but rewrite each "in(i)" relative to the new "in(Bot)".
4174 Node *progress = NULL;
4175
4176
4177 Node* old_base = base_memory();
4178 Node* empty_mem = empty_memory();
4179 if (old_base == empty_mem)
4180 return NULL; // Dead memory path.
4181
4182 MergeMemNode* old_mbase;
4183 if (old_base != NULL && old_base->is_MergeMem())
4184 old_mbase = old_base->as_MergeMem();
4185 else
4186 old_mbase = NULL;
4187 Node* new_base = old_base;
4188
4189 // simplify stacked MergeMems in base memory
4190 if (old_mbase) new_base = old_mbase->base_memory();
4191
4192 // the base memory might contribute new slices beyond my req()
4193 if (old_mbase) grow_to_match(old_mbase);
4194
4195 // Look carefully at the base node if it is a phi.
4196 PhiNode* phi_base;
4197 if (new_base != NULL && new_base->is_Phi())
4198 phi_base = new_base->as_Phi();
4199 else
4200 phi_base = NULL;
4201
4202 Node* phi_reg = NULL;
4203 uint phi_len = (uint)-1;
4204 if (phi_base != NULL && !phi_base->is_copy()) {
4205 // do not examine phi if degraded to a copy
4206 phi_reg = phi_base->region();
4207 phi_len = phi_base->req();
4208 // see if the phi is unfinished
4209 for (uint i = 1; i < phi_len; i++) {
4210 if (phi_base->in(i) == NULL) {
4211 // incomplete phi; do not look at it yet!
4212 phi_reg = NULL;
4213 phi_len = (uint)-1;
4214 break;
4215 }
4216 }
4217 }
4218
4219 // Note: We do not call verify_sparse on entry, because inputs
4220 // can normalize to the base_memory via subsume_node or similar
4221 // mechanisms. This method repairs that damage.
4222
4223 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4224
4225 // Look at each slice.
4226 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4227 Node* old_in = in(i);
4228 // calculate the old memory value
4229 Node* old_mem = old_in;
4230 if (old_mem == empty_mem) old_mem = old_base;
4231 assert(old_mem == memory_at(i), "");
4232
4233 // maybe update (reslice) the old memory value
4234
4235 // simplify stacked MergeMems
4236 Node* new_mem = old_mem;
4237 MergeMemNode* old_mmem;
4238 if (old_mem != NULL && old_mem->is_MergeMem())
4239 old_mmem = old_mem->as_MergeMem();
4240 else
4241 old_mmem = NULL;
4242 if (old_mmem == this) {
4243 // This can happen if loops break up and safepoints disappear.
4244 // A merge of BotPtr (default) with a RawPtr memory derived from a
4245 // safepoint can be rewritten to a merge of the same BotPtr with
4246 // the BotPtr phi coming into the loop. If that phi disappears
4247 // also, we can end up with a self-loop of the mergemem.
4248 // In general, if loops degenerate and memory effects disappear,
4249 // a mergemem can be left looking at itself. This simply means
4250 // that the mergemem's default should be used, since there is
4251 // no longer any apparent effect on this slice.
4252 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4253 // from start. Update the input to TOP.
4254 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4255 }
4256 else if (old_mmem != NULL) {
4257 new_mem = old_mmem->memory_at(i);
4258 }
4259 // else preceding memory was not a MergeMem
4260
4261 // replace equivalent phis (unfortunately, they do not GVN together)
4262 if (new_mem != NULL && new_mem != new_base &&
4263 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4264 if (new_mem->is_Phi()) {
4265 PhiNode* phi_mem = new_mem->as_Phi();
4266 for (uint i = 1; i < phi_len; i++) {
4267 if (phi_base->in(i) != phi_mem->in(i)) {
4268 phi_mem = NULL;
4269 break;
4270 }
4271 }
4272 if (phi_mem != NULL) {
4273 // equivalent phi nodes; revert to the def
4274 new_mem = new_base;
4275 }
4276 }
4277 }
4278
4279 // maybe store down a new value
4280 Node* new_in = new_mem;
4281 if (new_in == new_base) new_in = empty_mem;
4282
4283 if (new_in != old_in) {
4284 // Warning: Do not combine this "if" with the previous "if"
4285 // A memory slice might have be be rewritten even if it is semantically
4286 // unchanged, if the base_memory value has changed.
4287 set_req(i, new_in);
4288 progress = this; // Report progress
4289 }
4290 }
4291
4292 if (new_base != old_base) {
4293 set_req(Compile::AliasIdxBot, new_base);
4294 // Don't use set_base_memory(new_base), because we need to update du.
4295 assert(base_memory() == new_base, "");
4296 progress = this;
4297 }
4298
4299 if( base_memory() == this ) {
4300 // a self cycle indicates this memory path is dead
4301 set_req(Compile::AliasIdxBot, empty_mem);
4302 }
4303
4304 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4305 // Recursion must occur after the self cycle check above
4306 if( base_memory()->is_MergeMem() ) {
4307 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4308 Node *m = phase->transform(new_mbase); // Rollup any cycles
4309 if( m != NULL && (m->is_top() ||
4310 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4311 // propagate rollup of dead cycle to self
4312 set_req(Compile::AliasIdxBot, empty_mem);
4313 }
4314 }
4315
4316 if( base_memory() == empty_mem ) {
4317 progress = this;
4318 // Cut inputs during Parse phase only.
4319 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4320 if( !can_reshape ) {
4321 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4322 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4323 }
4324 }
4325 }
4326
4327 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4328 // Check if PhiNode::Ideal's "Split phis through memory merges"
4329 // transform should be attempted. Look for this->phi->this cycle.
4330 uint merge_width = req();
4331 if (merge_width > Compile::AliasIdxRaw) {
4332 PhiNode* phi = base_memory()->as_Phi();
4333 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4334 if (phi->in(i) == this) {
4335 phase->is_IterGVN()->_worklist.push(phi);
4336 break;
4337 }
4338 }
4339 }
4340 }
4341
4342 assert(progress || verify_sparse(), "please, no dups of base");
4343 return progress;
4344 }
4345
4346 //-------------------------set_base_memory-------------------------------------
4347 void MergeMemNode::set_base_memory(Node *new_base) {
4348 Node* empty_mem = empty_memory();
4349 set_req(Compile::AliasIdxBot, new_base);
4350 assert(memory_at(req()) == new_base, "must set default memory");
4351 // Clear out other occurrences of new_base:
4352 if (new_base != empty_mem) {
4353 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4354 if (in(i) == new_base) set_req(i, empty_mem);
4355 }
4356 }
4357 }
4358
4359 //------------------------------out_RegMask------------------------------------
4360 const RegMask &MergeMemNode::out_RegMask() const {
4361 return RegMask::Empty;
4362 }
4363
4364 //------------------------------dump_spec--------------------------------------
4365 #ifndef PRODUCT
4366 void MergeMemNode::dump_spec(outputStream *st) const {
4367 st->print(" {");
4368 Node* base_mem = base_memory();
4369 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4370 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4371 if (mem == base_mem) { st->print(" -"); continue; }
4372 st->print( " N%d:", mem->_idx );
4373 Compile::current()->get_adr_type(i)->dump_on(st);
4374 }
4375 st->print(" }");
4376 }
4377 #endif // !PRODUCT
4378
4379
4380 #ifdef ASSERT
4381 static bool might_be_same(Node* a, Node* b) {
4382 if (a == b) return true;
4383 if (!(a->is_Phi() || b->is_Phi())) return false;
4384 // phis shift around during optimization
4385 return true; // pretty stupid...
4386 }
4387
4388 // verify a narrow slice (either incoming or outgoing)
4389 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4390 if (!VerifyAliases) return; // don't bother to verify unless requested
4391 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4392 if (Node::in_dump()) return; // muzzle asserts when printing
4393 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4394 assert(n != NULL, "");
4395 // Elide intervening MergeMem's
4396 while (n->is_MergeMem()) {
4397 n = n->as_MergeMem()->memory_at(alias_idx);
4398 }
4399 Compile* C = Compile::current();
4400 const TypePtr* n_adr_type = n->adr_type();
4401 if (n == m->empty_memory()) {
4402 // Implicit copy of base_memory()
4403 } else if (n_adr_type != TypePtr::BOTTOM) {
4404 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4405 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4406 } else {
4407 // A few places like make_runtime_call "know" that VM calls are narrow,
4408 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4409 bool expected_wide_mem = false;
4410 if (n == m->base_memory()) {
4411 expected_wide_mem = true;
4412 } else if (alias_idx == Compile::AliasIdxRaw ||
4413 n == m->memory_at(Compile::AliasIdxRaw)) {
4414 expected_wide_mem = true;
4415 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4416 // memory can "leak through" calls on channels that
4417 // are write-once. Allow this also.
4418 expected_wide_mem = true;
4419 }
4420 assert(expected_wide_mem, "expected narrow slice replacement");
4421 }
4422 }
4423 #else // !ASSERT
4424 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4425 #endif
4426
4427
4428 //-----------------------------memory_at---------------------------------------
4429 Node* MergeMemNode::memory_at(uint alias_idx) const {
4430 assert(alias_idx >= Compile::AliasIdxRaw ||
4431 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4432 "must avoid base_memory and AliasIdxTop");
4433
4434 // Otherwise, it is a narrow slice.
4435 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4436 Compile *C = Compile::current();
4437 if (is_empty_memory(n)) {
4438 // the array is sparse; empty slots are the "top" node
4439 n = base_memory();
4440 assert(Node::in_dump()
4441 || n == NULL || n->bottom_type() == Type::TOP
4442 || n->adr_type() == NULL // address is TOP
4443 || n->adr_type() == TypePtr::BOTTOM
4444 || n->adr_type() == TypeRawPtr::BOTTOM
4445 || Compile::current()->AliasLevel() == 0,
4446 "must be a wide memory");
4447 // AliasLevel == 0 if we are organizing the memory states manually.
4448 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4449 } else {
4450 // make sure the stored slice is sane
4451 #ifdef ASSERT
4452 if (VMError::is_error_reported() || Node::in_dump()) {
4453 } else if (might_be_same(n, base_memory())) {
4454 // Give it a pass: It is a mostly harmless repetition of the base.
4455 // This can arise normally from node subsumption during optimization.
4456 } else {
4457 verify_memory_slice(this, alias_idx, n);
4458 }
4459 #endif
4460 }
4461 return n;
4462 }
4463
4464 //---------------------------set_memory_at-------------------------------------
4465 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4466 verify_memory_slice(this, alias_idx, n);
4467 Node* empty_mem = empty_memory();
4468 if (n == base_memory()) n = empty_mem; // collapse default
4469 uint need_req = alias_idx+1;
4470 if (req() < need_req) {
4471 if (n == empty_mem) return; // already the default, so do not grow me
4472 // grow the sparse array
4473 do {
4474 add_req(empty_mem);
4475 } while (req() < need_req);
4476 }
4477 set_req( alias_idx, n );
4478 }
4479
4480
4481
4482 //--------------------------iteration_setup------------------------------------
4483 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4484 if (other != NULL) {
4485 grow_to_match(other);
4486 // invariant: the finite support of mm2 is within mm->req()
4487 #ifdef ASSERT
4488 for (uint i = req(); i < other->req(); i++) {
4489 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4490 }
4491 #endif
4492 }
4493 // Replace spurious copies of base_memory by top.
4494 Node* base_mem = base_memory();
4495 if (base_mem != NULL && !base_mem->is_top()) {
4496 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4497 if (in(i) == base_mem)
4498 set_req(i, empty_memory());
4499 }
4500 }
4501 }
4502
4503 //---------------------------grow_to_match-------------------------------------
4504 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4505 Node* empty_mem = empty_memory();
4506 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4507 // look for the finite support of the other memory
4508 for (uint i = other->req(); --i >= req(); ) {
4509 if (other->in(i) != empty_mem) {
4510 uint new_len = i+1;
4511 while (req() < new_len) add_req(empty_mem);
4512 break;
4513 }
4514 }
4515 }
4516
4517 //---------------------------verify_sparse-------------------------------------
4518 #ifndef PRODUCT
4519 bool MergeMemNode::verify_sparse() const {
4520 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4521 Node* base_mem = base_memory();
4522 // The following can happen in degenerate cases, since empty==top.
4523 if (is_empty_memory(base_mem)) return true;
4524 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4525 assert(in(i) != NULL, "sane slice");
4526 if (in(i) == base_mem) return false; // should have been the sentinel value!
4527 }
4528 return true;
4529 }
4530
4531 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4532 Node* n;
4533 n = mm->in(idx);
4534 if (mem == n) return true; // might be empty_memory()
4535 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4536 if (mem == n) return true;
4537 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4538 if (mem == n) return true;
4539 if (n == NULL) break;
4540 }
4541 return false;
4542 }
4543 #endif // !PRODUCT