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