1 /*
2 * Copyright (c) 1997, 2018, 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 "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/connode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/matcher.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/regmask.hpp"
47 #include "utilities/align.hpp"
48 #include "utilities/copy.hpp"
49 #include "utilities/macros.hpp"
50 #include "utilities/vmError.hpp"
51 #if INCLUDE_ZGC
52 #include "gc/z/c2/zBarrierSetC2.hpp"
53 #endif
54
55 // Portions of code courtesy of Clifford Click
56
57 // Optimization - Graph Style
58
59 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
60
61 //=============================================================================
62 uint MemNode::size_of() const { return sizeof(*this); }
63
64 const TypePtr *MemNode::adr_type() const {
65 Node* adr = in(Address);
66 if (adr == NULL) return NULL; // node is dead
67 const TypePtr* cross_check = NULL;
68 DEBUG_ONLY(cross_check = _adr_type);
69 return calculate_adr_type(adr->bottom_type(), cross_check);
70 }
71
72 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
73 if (adr != NULL) {
74 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
75 return true;
76 }
77 }
78 return false;
79 }
80
81 #ifndef PRODUCT
82 void MemNode::dump_spec(outputStream *st) const {
83 if (in(Address) == NULL) return; // node is dead
84 #ifndef ASSERT
85 // fake the missing field
86 const TypePtr* _adr_type = NULL;
87 if (in(Address) != NULL)
88 _adr_type = in(Address)->bottom_type()->isa_ptr();
89 #endif
90 dump_adr_type(this, _adr_type, st);
91
92 Compile* C = Compile::current();
93 if (C->alias_type(_adr_type)->is_volatile()) {
94 st->print(" Volatile!");
95 }
96 if (_unaligned_access) {
97 st->print(" unaligned");
98 }
99 if (_mismatched_access) {
100 st->print(" mismatched");
101 }
102 if (_unsafe_access) {
103 st->print(" unsafe");
104 }
105 }
106
107 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
108 st->print(" @");
109 if (adr_type == NULL) {
110 st->print("NULL");
111 } else {
112 adr_type->dump_on(st);
113 Compile* C = Compile::current();
114 Compile::AliasType* atp = NULL;
115 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
116 if (atp == NULL)
117 st->print(", idx=?\?;");
118 else if (atp->index() == Compile::AliasIdxBot)
119 st->print(", idx=Bot;");
120 else if (atp->index() == Compile::AliasIdxTop)
121 st->print(", idx=Top;");
122 else if (atp->index() == Compile::AliasIdxRaw)
123 st->print(", idx=Raw;");
124 else {
125 ciField* field = atp->field();
126 if (field) {
127 st->print(", name=");
128 field->print_name_on(st);
129 }
130 st->print(", idx=%d;", atp->index());
131 }
132 }
133 }
134
135 extern void print_alias_types();
136
137 #endif
138
139 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
140 assert((t_oop != NULL), "sanity");
141 bool is_instance = t_oop->is_known_instance_field();
142 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
143 (load != NULL) && load->is_Load() &&
144 (phase->is_IterGVN() != NULL);
145 if (!(is_instance || is_boxed_value_load))
146 return mchain; // don't try to optimize non-instance types
147 uint instance_id = t_oop->instance_id();
148 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
149 Node *prev = NULL;
150 Node *result = mchain;
151 while (prev != result) {
152 prev = result;
153 if (result == start_mem)
154 break; // hit one of our sentinels
155 // skip over a call which does not affect this memory slice
156 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
157 Node *proj_in = result->in(0);
158 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
159 break; // hit one of our sentinels
160 } else if (proj_in->is_Call()) {
161 // ArrayCopyNodes processed here as well
162 CallNode *call = proj_in->as_Call();
163 if (!call->may_modify(t_oop, phase)) { // returns false for instances
164 result = call->in(TypeFunc::Memory);
165 }
166 } else if (proj_in->is_Initialize()) {
167 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
168 // Stop if this is the initialization for the object instance which
169 // contains this memory slice, otherwise skip over it.
170 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
171 break;
172 }
173 if (is_instance) {
174 result = proj_in->in(TypeFunc::Memory);
175 } else if (is_boxed_value_load) {
176 Node* klass = alloc->in(AllocateNode::KlassNode);
177 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
178 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
179 result = proj_in->in(TypeFunc::Memory); // not related allocation
180 }
181 }
182 } else if (proj_in->is_MemBar()) {
183 ArrayCopyNode* ac = NULL;
184 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
185 break;
186 }
187 result = proj_in->in(TypeFunc::Memory);
188 } else {
189 assert(false, "unexpected projection");
190 }
191 } else if (result->is_ClearArray()) {
192 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
193 // Can not bypass initialization of the instance
194 // we are looking for.
195 break;
196 }
197 // Otherwise skip it (the call updated 'result' value).
198 } else if (result->is_MergeMem()) {
199 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
200 }
201 }
202 return result;
203 }
204
205 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
206 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
207 if (t_oop == NULL)
208 return mchain; // don't try to optimize non-oop types
209 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
210 bool is_instance = t_oop->is_known_instance_field();
211 PhaseIterGVN *igvn = phase->is_IterGVN();
212 if (is_instance && igvn != NULL && result->is_Phi()) {
213 PhiNode *mphi = result->as_Phi();
214 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
215 const TypePtr *t = mphi->adr_type();
216 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
217 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
218 t->is_oopptr()->cast_to_exactness(true)
219 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
220 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
221 // clone the Phi with our address type
222 result = mphi->split_out_instance(t_adr, igvn);
223 } else {
224 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
225 }
226 }
227 return result;
228 }
229
230 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
231 uint alias_idx = phase->C->get_alias_index(tp);
232 Node *mem = mmem;
233 #ifdef ASSERT
234 {
235 // Check that current type is consistent with the alias index used during graph construction
236 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
237 bool consistent = adr_check == NULL || adr_check->empty() ||
238 phase->C->must_alias(adr_check, alias_idx );
239 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
240 if( !consistent && adr_check != NULL && !adr_check->empty() &&
241 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
242 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
243 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
244 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
245 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
246 // don't assert if it is dead code.
247 consistent = true;
248 }
249 if( !consistent ) {
250 st->print("alias_idx==%d, adr_check==", alias_idx);
251 if( adr_check == NULL ) {
252 st->print("NULL");
253 } else {
254 adr_check->dump();
255 }
256 st->cr();
257 print_alias_types();
258 assert(consistent, "adr_check must match alias idx");
259 }
260 }
261 #endif
262 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
263 // means an array I have not precisely typed yet. Do not do any
264 // alias stuff with it any time soon.
265 const TypeOopPtr *toop = tp->isa_oopptr();
266 if( tp->base() != Type::AnyPtr &&
267 !(toop &&
268 toop->klass() != NULL &&
269 toop->klass()->is_java_lang_Object() &&
270 toop->offset() == Type::OffsetBot) ) {
271 // compress paths and change unreachable cycles to TOP
272 // If not, we can update the input infinitely along a MergeMem cycle
273 // Equivalent code in PhiNode::Ideal
274 Node* m = phase->transform(mmem);
275 // If transformed to a MergeMem, get the desired slice
276 // Otherwise the returned node represents memory for every slice
277 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
278 // Update input if it is progress over what we have now
279 }
280 return mem;
281 }
282
283 //--------------------------Ideal_common---------------------------------------
284 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
285 // Unhook non-raw memories from complete (macro-expanded) initializations.
286 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
287 // If our control input is a dead region, kill all below the region
288 Node *ctl = in(MemNode::Control);
289 if (ctl && remove_dead_region(phase, can_reshape))
290 return this;
291 ctl = in(MemNode::Control);
292 // Don't bother trying to transform a dead node
293 if (ctl && ctl->is_top()) return NodeSentinel;
294
295 PhaseIterGVN *igvn = phase->is_IterGVN();
296 // Wait if control on the worklist.
297 if (ctl && can_reshape && igvn != NULL) {
298 Node* bol = NULL;
299 Node* cmp = NULL;
300 if (ctl->in(0)->is_If()) {
301 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
302 bol = ctl->in(0)->in(1);
303 if (bol->is_Bool())
304 cmp = ctl->in(0)->in(1)->in(1);
305 }
306 if (igvn->_worklist.member(ctl) ||
307 (bol != NULL && igvn->_worklist.member(bol)) ||
308 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
309 // This control path may be dead.
310 // Delay this memory node transformation until the control is processed.
311 phase->is_IterGVN()->_worklist.push(this);
312 return NodeSentinel; // caller will return NULL
313 }
314 }
315 // Ignore if memory is dead, or self-loop
316 Node *mem = in(MemNode::Memory);
317 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
318 assert(mem != this, "dead loop in MemNode::Ideal");
319
320 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
321 // This memory slice may be dead.
322 // Delay this mem node transformation until the memory is processed.
323 phase->is_IterGVN()->_worklist.push(this);
324 return NodeSentinel; // caller will return NULL
325 }
326
327 Node *address = in(MemNode::Address);
328 const Type *t_adr = phase->type(address);
329 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
330
331 if (can_reshape && igvn != NULL &&
332 (igvn->_worklist.member(address) ||
333 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
334 // The address's base and type may change when the address is processed.
335 // Delay this mem node transformation until the address is processed.
336 phase->is_IterGVN()->_worklist.push(this);
337 return NodeSentinel; // caller will return NULL
338 }
339
340 // Do NOT remove or optimize the next lines: ensure a new alias index
341 // is allocated for an oop pointer type before Escape Analysis.
342 // Note: C++ will not remove it since the call has side effect.
343 if (t_adr->isa_oopptr()) {
344 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
345 }
346
347 Node* base = NULL;
348 if (address->is_AddP()) {
349 base = address->in(AddPNode::Base);
350 }
351 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
352 !t_adr->isa_rawptr()) {
353 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
354 // Skip this node optimization if its address has TOP base.
355 return NodeSentinel; // caller will return NULL
356 }
357
358 // Avoid independent memory operations
359 Node* old_mem = mem;
360
361 // The code which unhooks non-raw memories from complete (macro-expanded)
362 // initializations was removed. After macro-expansion all stores catched
363 // by Initialize node became raw stores and there is no information
364 // which memory slices they modify. So it is unsafe to move any memory
365 // operation above these stores. Also in most cases hooked non-raw memories
366 // were already unhooked by using information from detect_ptr_independence()
367 // and find_previous_store().
368
369 if (mem->is_MergeMem()) {
370 MergeMemNode* mmem = mem->as_MergeMem();
371 const TypePtr *tp = t_adr->is_ptr();
372
373 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
374 }
375
376 if (mem != old_mem) {
377 set_req(MemNode::Memory, mem);
378 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
379 igvn->_worklist.push(old_mem);
380 }
381 if (phase->type(mem) == Type::TOP) return NodeSentinel;
382 return this;
383 }
384
385 // let the subclass continue analyzing...
386 return NULL;
387 }
388
389 // Helper function for proving some simple control dominations.
390 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
391 // Already assumes that 'dom' is available at 'sub', and that 'sub'
392 // is not a constant (dominated by the method's StartNode).
393 // Used by MemNode::find_previous_store to prove that the
394 // control input of a memory operation predates (dominates)
395 // an allocation it wants to look past.
396 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
397 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
398 return false; // Conservative answer for dead code
399
400 // Check 'dom'. Skip Proj and CatchProj nodes.
401 dom = dom->find_exact_control(dom);
402 if (dom == NULL || dom->is_top())
403 return false; // Conservative answer for dead code
404
405 if (dom == sub) {
406 // For the case when, for example, 'sub' is Initialize and the original
407 // 'dom' is Proj node of the 'sub'.
408 return false;
409 }
410
411 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
412 return true;
413
414 // 'dom' dominates 'sub' if its control edge and control edges
415 // of all its inputs dominate or equal to sub's control edge.
416
417 // Currently 'sub' is either Allocate, Initialize or Start nodes.
418 // Or Region for the check in LoadNode::Ideal();
419 // 'sub' should have sub->in(0) != NULL.
420 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
421 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
422
423 // Get control edge of 'sub'.
424 Node* orig_sub = sub;
425 sub = sub->find_exact_control(sub->in(0));
426 if (sub == NULL || sub->is_top())
427 return false; // Conservative answer for dead code
428
429 assert(sub->is_CFG(), "expecting control");
430
431 if (sub == dom)
432 return true;
433
434 if (sub->is_Start() || sub->is_Root())
435 return false;
436
437 {
438 // Check all control edges of 'dom'.
439
440 ResourceMark rm;
441 Arena* arena = Thread::current()->resource_area();
442 Node_List nlist(arena);
443 Unique_Node_List dom_list(arena);
444
445 dom_list.push(dom);
446 bool only_dominating_controls = false;
447
448 for (uint next = 0; next < dom_list.size(); next++) {
449 Node* n = dom_list.at(next);
450 if (n == orig_sub)
451 return false; // One of dom's inputs dominated by sub.
452 if (!n->is_CFG() && n->pinned()) {
453 // Check only own control edge for pinned non-control nodes.
454 n = n->find_exact_control(n->in(0));
455 if (n == NULL || n->is_top())
456 return false; // Conservative answer for dead code
457 assert(n->is_CFG(), "expecting control");
458 dom_list.push(n);
459 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
460 only_dominating_controls = true;
461 } else if (n->is_CFG()) {
462 if (n->dominates(sub, nlist))
463 only_dominating_controls = true;
464 else
465 return false;
466 } else {
467 // First, own control edge.
468 Node* m = n->find_exact_control(n->in(0));
469 if (m != NULL) {
470 if (m->is_top())
471 return false; // Conservative answer for dead code
472 dom_list.push(m);
473 }
474 // Now, the rest of edges.
475 uint cnt = n->req();
476 for (uint i = 1; i < cnt; i++) {
477 m = n->find_exact_control(n->in(i));
478 if (m == NULL || m->is_top())
479 continue;
480 dom_list.push(m);
481 }
482 }
483 }
484 return only_dominating_controls;
485 }
486 }
487
488 //---------------------detect_ptr_independence---------------------------------
489 // Used by MemNode::find_previous_store to prove that two base
490 // pointers are never equal.
491 // The pointers are accompanied by their associated allocations,
492 // if any, which have been previously discovered by the caller.
493 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
494 Node* p2, AllocateNode* a2,
495 PhaseTransform* phase) {
496 // Attempt to prove that these two pointers cannot be aliased.
497 // They may both manifestly be allocations, and they should differ.
498 // Or, if they are not both allocations, they can be distinct constants.
499 // Otherwise, one is an allocation and the other a pre-existing value.
500 if (a1 == NULL && a2 == NULL) { // neither an allocation
501 return (p1 != p2) && p1->is_Con() && p2->is_Con();
502 } else if (a1 != NULL && a2 != NULL) { // both allocations
503 return (a1 != a2);
504 } else if (a1 != NULL) { // one allocation a1
505 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
506 return all_controls_dominate(p2, a1);
507 } else { //(a2 != NULL) // one allocation a2
508 return all_controls_dominate(p1, a2);
509 }
510 return false;
511 }
512
513
514 // Find an arraycopy that must have set (can_see_stored_value=true) or
515 // could have set (can_see_stored_value=false) the value for this load
516 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
517 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
518 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
519 Node* mb = mem->in(0);
520 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
521 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
522 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
523 if (ac->is_clonebasic()) {
524 intptr_t offset;
525 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
526 if (alloc != NULL && alloc == ld_alloc) {
527 return ac;
528 }
529 }
530 }
531 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
532 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
533
534 if (ac->is_arraycopy_validated() ||
535 ac->is_copyof_validated() ||
536 ac->is_copyofrange_validated()) {
537 Node* ld_addp = in(MemNode::Address);
538 if (ld_addp->is_AddP()) {
539 Node* ld_base = ld_addp->in(AddPNode::Address);
540 Node* ld_offs = ld_addp->in(AddPNode::Offset);
541
542 Node* dest = ac->in(ArrayCopyNode::Dest);
543
544 if (dest == ld_base) {
545 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
546 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
547 return ac;
548 }
549 if (!can_see_stored_value) {
550 mem = ac->in(TypeFunc::Memory);
551 }
552 }
553 }
554 }
555 }
556 return NULL;
557 }
558
559 // The logic for reordering loads and stores uses four steps:
560 // (a) Walk carefully past stores and initializations which we
561 // can prove are independent of this load.
562 // (b) Observe that the next memory state makes an exact match
563 // with self (load or store), and locate the relevant store.
564 // (c) Ensure that, if we were to wire self directly to the store,
565 // the optimizer would fold it up somehow.
566 // (d) Do the rewiring, and return, depending on some other part of
567 // the optimizer to fold up the load.
568 // This routine handles steps (a) and (b). Steps (c) and (d) are
569 // specific to loads and stores, so they are handled by the callers.
570 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
571 //
572 Node* MemNode::find_previous_store(PhaseTransform* phase) {
573 Node* ctrl = in(MemNode::Control);
574 Node* adr = in(MemNode::Address);
575 intptr_t offset = 0;
576 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
577 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
578
579 if (offset == Type::OffsetBot)
580 return NULL; // cannot unalias unless there are precise offsets
581
582 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
583 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
584
585 intptr_t size_in_bytes = memory_size();
586
587 Node* mem = in(MemNode::Memory); // start searching here...
588
589 int cnt = 50; // Cycle limiter
590 for (;;) { // While we can dance past unrelated stores...
591 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
592
593 Node* prev = mem;
594 if (mem->is_Store()) {
595 Node* st_adr = mem->in(MemNode::Address);
596 intptr_t st_offset = 0;
597 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
598 if (st_base == NULL)
599 break; // inscrutable pointer
600
601 // For raw accesses it's not enough to prove that constant offsets don't intersect.
602 // We need the bases to be the equal in order for the offset check to make sense.
603 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
604 break;
605 }
606
607 if (st_offset != offset && st_offset != Type::OffsetBot) {
608 const int MAX_STORE = BytesPerLong;
609 if (st_offset >= offset + size_in_bytes ||
610 st_offset <= offset - MAX_STORE ||
611 st_offset <= offset - mem->as_Store()->memory_size()) {
612 // Success: The offsets are provably independent.
613 // (You may ask, why not just test st_offset != offset and be done?
614 // The answer is that stores of different sizes can co-exist
615 // in the same sequence of RawMem effects. We sometimes initialize
616 // a whole 'tile' of array elements with a single jint or jlong.)
617 mem = mem->in(MemNode::Memory);
618 continue; // (a) advance through independent store memory
619 }
620 }
621 if (st_base != base &&
622 detect_ptr_independence(base, alloc,
623 st_base,
624 AllocateNode::Ideal_allocation(st_base, phase),
625 phase)) {
626 // Success: The bases are provably independent.
627 mem = mem->in(MemNode::Memory);
628 continue; // (a) advance through independent store memory
629 }
630
631 // (b) At this point, if the bases or offsets do not agree, we lose,
632 // since we have not managed to prove 'this' and 'mem' independent.
633 if (st_base == base && st_offset == offset) {
634 return mem; // let caller handle steps (c), (d)
635 }
636
637 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
638 InitializeNode* st_init = mem->in(0)->as_Initialize();
639 AllocateNode* st_alloc = st_init->allocation();
640 if (st_alloc == NULL)
641 break; // something degenerated
642 bool known_identical = false;
643 bool known_independent = false;
644 if (alloc == st_alloc)
645 known_identical = true;
646 else if (alloc != NULL)
647 known_independent = true;
648 else if (all_controls_dominate(this, st_alloc))
649 known_independent = true;
650
651 if (known_independent) {
652 // The bases are provably independent: Either they are
653 // manifestly distinct allocations, or else the control
654 // of this load dominates the store's allocation.
655 int alias_idx = phase->C->get_alias_index(adr_type());
656 if (alias_idx == Compile::AliasIdxRaw) {
657 mem = st_alloc->in(TypeFunc::Memory);
658 } else {
659 mem = st_init->memory(alias_idx);
660 }
661 continue; // (a) advance through independent store memory
662 }
663
664 // (b) at this point, if we are not looking at a store initializing
665 // the same allocation we are loading from, we lose.
666 if (known_identical) {
667 // From caller, can_see_stored_value will consult find_captured_store.
668 return mem; // let caller handle steps (c), (d)
669 }
670
671 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
672 if (prev != mem) {
673 // Found an arraycopy but it doesn't affect that load
674 continue;
675 }
676 // Found an arraycopy that may affect that load
677 return mem;
678 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
679 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
680 if (mem->is_Proj() && mem->in(0)->is_Call()) {
681 // ArrayCopyNodes processed here as well.
682 CallNode *call = mem->in(0)->as_Call();
683 if (!call->may_modify(addr_t, phase)) {
684 mem = call->in(TypeFunc::Memory);
685 continue; // (a) advance through independent call memory
686 }
687 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
688 ArrayCopyNode* ac = NULL;
689 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
690 break;
691 }
692 mem = mem->in(0)->in(TypeFunc::Memory);
693 continue; // (a) advance through independent MemBar memory
694 } else if (mem->is_ClearArray()) {
695 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
696 // (the call updated 'mem' value)
697 continue; // (a) advance through independent allocation memory
698 } else {
699 // Can not bypass initialization of the instance
700 // we are looking for.
701 return mem;
702 }
703 } else if (mem->is_MergeMem()) {
704 int alias_idx = phase->C->get_alias_index(adr_type());
705 mem = mem->as_MergeMem()->memory_at(alias_idx);
706 continue; // (a) advance through independent MergeMem memory
707 }
708 }
709
710 // Unless there is an explicit 'continue', we must bail out here,
711 // because 'mem' is an inscrutable memory state (e.g., a call).
712 break;
713 }
714
715 return NULL; // bail out
716 }
717
718 //----------------------calculate_adr_type-------------------------------------
719 // Helper function. Notices when the given type of address hits top or bottom.
720 // Also, asserts a cross-check of the type against the expected address type.
721 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
722 if (t == Type::TOP) return NULL; // does not touch memory any more?
723 #ifdef PRODUCT
724 cross_check = NULL;
725 #else
726 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
727 #endif
728 const TypePtr* tp = t->isa_ptr();
729 if (tp == NULL) {
730 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
731 return TypePtr::BOTTOM; // touches lots of memory
732 } else {
733 #ifdef ASSERT
734 // %%%% [phh] We don't check the alias index if cross_check is
735 // TypeRawPtr::BOTTOM. Needs to be investigated.
736 if (cross_check != NULL &&
737 cross_check != TypePtr::BOTTOM &&
738 cross_check != TypeRawPtr::BOTTOM) {
739 // Recheck the alias index, to see if it has changed (due to a bug).
740 Compile* C = Compile::current();
741 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
742 "must stay in the original alias category");
743 // The type of the address must be contained in the adr_type,
744 // disregarding "null"-ness.
745 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
746 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
747 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
748 "real address must not escape from expected memory type");
749 }
750 #endif
751 return tp;
752 }
753 }
754
755 //=============================================================================
756 // Should LoadNode::Ideal() attempt to remove control edges?
757 bool LoadNode::can_remove_control() const {
758 return true;
759 }
760 uint LoadNode::size_of() const { return sizeof(*this); }
761 uint LoadNode::cmp( const Node &n ) const
762 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
763 const Type *LoadNode::bottom_type() const { return _type; }
764 uint LoadNode::ideal_reg() const {
765 return _type->ideal_reg();
766 }
767
768 #ifndef PRODUCT
769 void LoadNode::dump_spec(outputStream *st) const {
770 MemNode::dump_spec(st);
771 if( !Verbose && !WizardMode ) {
772 // standard dump does this in Verbose and WizardMode
773 st->print(" #"); _type->dump_on(st);
774 }
775 if (!depends_only_on_test()) {
776 st->print(" (does not depend only on test)");
777 }
778 }
779 #endif
780
781 #ifdef ASSERT
782 //----------------------------is_immutable_value-------------------------------
783 // Helper function to allow a raw load without control edge for some cases
784 bool LoadNode::is_immutable_value(Node* adr) {
785 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
786 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
787 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
788 in_bytes(JavaThread::osthread_offset())));
789 }
790 #endif
791
792 //----------------------------LoadNode::make-----------------------------------
793 // Polymorphic factory method:
794 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
795 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
796 Compile* C = gvn.C;
797
798 // sanity check the alias category against the created node type
799 assert(!(adr_type->isa_oopptr() &&
800 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
801 "use LoadKlassNode instead");
802 assert(!(adr_type->isa_aryptr() &&
803 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
804 "use LoadRangeNode instead");
805 // Check control edge of raw loads
806 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
807 // oop will be recorded in oop map if load crosses safepoint
808 rt->isa_oopptr() || is_immutable_value(adr),
809 "raw memory operations should have control edge");
810 LoadNode* load = NULL;
811 switch (bt) {
812 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
813 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
814 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
815 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
816 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
817 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
818 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
819 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
820 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
821 case T_OBJECT:
822 #ifdef _LP64
823 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
824 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
825 } else
826 #endif
827 {
828 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
829 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
830 }
831 break;
832 default:
833 ShouldNotReachHere();
834 break;
835 }
836 assert(load != NULL, "LoadNode should have been created");
837 if (unaligned) {
838 load->set_unaligned_access();
839 }
840 if (mismatched) {
841 load->set_mismatched_access();
842 }
843 if (unsafe) {
844 load->set_unsafe_access();
845 }
846 if (load->Opcode() == Op_LoadN) {
847 Node* ld = gvn.transform(load);
848 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
849 }
850
851 return load;
852 }
853
854 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
855 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
856 bool require_atomic = true;
857 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
858 if (unaligned) {
859 load->set_unaligned_access();
860 }
861 if (mismatched) {
862 load->set_mismatched_access();
863 }
864 if (unsafe) {
865 load->set_unsafe_access();
866 }
867 return load;
868 }
869
870 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
871 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
872 bool require_atomic = true;
873 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
874 if (unaligned) {
875 load->set_unaligned_access();
876 }
877 if (mismatched) {
878 load->set_mismatched_access();
879 }
880 if (unsafe) {
881 load->set_unsafe_access();
882 }
883 return load;
884 }
885
886
887
888 //------------------------------hash-------------------------------------------
889 uint LoadNode::hash() const {
890 // unroll addition of interesting fields
891 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
892 }
893
894 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
895 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
896 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
897 bool is_stable_ary = FoldStableValues &&
898 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
899 tp->isa_aryptr()->is_stable();
900
901 return (eliminate_boxing && non_volatile) || is_stable_ary;
902 }
903
904 return false;
905 }
906
907 // Is the value loaded previously stored by an arraycopy? If so return
908 // a load node that reads from the source array so we may be able to
909 // optimize out the ArrayCopy node later.
910 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
911 #if INCLUDE_ZGC
912 if (UseZGC) {
913 if (bottom_type()->make_oopptr() != NULL) {
914 return NULL;
915 }
916 }
917 #endif
918
919 Node* ld_adr = in(MemNode::Address);
920 intptr_t ld_off = 0;
921 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
922 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
923 if (ac != NULL) {
924 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
925
926 Node* mem = ac->in(TypeFunc::Memory);
927 Node* ctl = ac->in(0);
928 Node* src = ac->in(ArrayCopyNode::Src);
929
930 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
931 return NULL;
932 }
933
934 LoadNode* ld = clone()->as_Load();
935 Node* addp = in(MemNode::Address)->clone();
936 if (ac->as_ArrayCopy()->is_clonebasic()) {
937 assert(ld_alloc != NULL, "need an alloc");
938 assert(addp->is_AddP(), "address must be addp");
939 assert(ac->in(ArrayCopyNode::Dest)->is_AddP(), "dest must be an address");
940 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
941 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base)), "strange pattern");
942 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address)), "strange pattern");
943 addp->set_req(AddPNode::Base, src->in(AddPNode::Base));
944 addp->set_req(AddPNode::Address, src->in(AddPNode::Address));
945 } else {
946 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
947 ac->as_ArrayCopy()->is_copyof_validated() ||
948 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
949 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
950 addp->set_req(AddPNode::Base, src);
951 addp->set_req(AddPNode::Address, src);
952
953 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
954 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
955 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
956 uint shift = exact_log2(type2aelembytes(ary_elem));
957
958 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
959 #ifdef _LP64
960 diff = phase->transform(new ConvI2LNode(diff));
961 #endif
962 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
963
964 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
965 addp->set_req(AddPNode::Offset, offset);
966 }
967 addp = phase->transform(addp);
968 #ifdef ASSERT
969 const TypePtr* adr_type = phase->type(addp)->is_ptr();
970 ld->_adr_type = adr_type;
971 #endif
972 ld->set_req(MemNode::Address, addp);
973 ld->set_req(0, ctl);
974 ld->set_req(MemNode::Memory, mem);
975 // load depends on the tests that validate the arraycopy
976 ld->_control_dependency = Pinned;
977 return ld;
978 }
979 return NULL;
980 }
981
982
983 //---------------------------can_see_stored_value------------------------------
984 // This routine exists to make sure this set of tests is done the same
985 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
986 // will change the graph shape in a way which makes memory alive twice at the
987 // same time (uses the Oracle model of aliasing), then some
988 // LoadXNode::Identity will fold things back to the equivalence-class model
989 // of aliasing.
990 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
991 Node* ld_adr = in(MemNode::Address);
992 intptr_t ld_off = 0;
993 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
994 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase);
995 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
996 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
997 // This is more general than load from boxing objects.
998 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
999 uint alias_idx = atp->index();
1000 bool final = !atp->is_rewritable();
1001 Node* result = NULL;
1002 Node* current = st;
1003 // Skip through chains of MemBarNodes checking the MergeMems for
1004 // new states for the slice of this load. Stop once any other
1005 // kind of node is encountered. Loads from final memory can skip
1006 // through any kind of MemBar but normal loads shouldn't skip
1007 // through MemBarAcquire since the could allow them to move out of
1008 // a synchronized region.
1009 while (current->is_Proj()) {
1010 int opc = current->in(0)->Opcode();
1011 if ((final && (opc == Op_MemBarAcquire ||
1012 opc == Op_MemBarAcquireLock ||
1013 opc == Op_LoadFence)) ||
1014 opc == Op_MemBarRelease ||
1015 opc == Op_StoreFence ||
1016 opc == Op_MemBarReleaseLock ||
1017 opc == Op_MemBarStoreStore ||
1018 opc == Op_MemBarCPUOrder) {
1019 Node* mem = current->in(0)->in(TypeFunc::Memory);
1020 if (mem->is_MergeMem()) {
1021 MergeMemNode* merge = mem->as_MergeMem();
1022 Node* new_st = merge->memory_at(alias_idx);
1023 if (new_st == merge->base_memory()) {
1024 // Keep searching
1025 current = new_st;
1026 continue;
1027 }
1028 // Save the new memory state for the slice and fall through
1029 // to exit.
1030 result = new_st;
1031 }
1032 }
1033 break;
1034 }
1035 if (result != NULL) {
1036 st = result;
1037 }
1038 }
1039
1040 // Loop around twice in the case Load -> Initialize -> Store.
1041 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1042 for (int trip = 0; trip <= 1; trip++) {
1043
1044 if (st->is_Store()) {
1045 Node* st_adr = st->in(MemNode::Address);
1046 if (!phase->eqv(st_adr, ld_adr)) {
1047 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1048 intptr_t st_off = 0;
1049 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1050 if (ld_base == NULL) return NULL;
1051 if (st_base == NULL) return NULL;
1052 if (ld_base->uncast() != st_base->uncast()) return NULL;
1053 if (ld_off != st_off) return NULL;
1054 if (ld_off == Type::OffsetBot) return NULL;
1055 // Same base, same offset.
1056 // Possible improvement for arrays: check index value instead of absolute offset.
1057
1058 // At this point we have proven something like this setup:
1059 // B = << base >>
1060 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1061 // S = StoreQ(AddP( B , #Off), V)
1062 // (Actually, we haven't yet proven the Q's are the same.)
1063 // In other words, we are loading from a casted version of
1064 // the same pointer-and-offset that we stored to.
1065 // Thus, we are able to replace L by V.
1066 }
1067 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1068 if (store_Opcode() != st->Opcode())
1069 return NULL;
1070 return st->in(MemNode::ValueIn);
1071 }
1072
1073 // A load from a freshly-created object always returns zero.
1074 // (This can happen after LoadNode::Ideal resets the load's memory input
1075 // to find_captured_store, which returned InitializeNode::zero_memory.)
1076 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1077 (st->in(0) == ld_alloc) &&
1078 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1079 // return a zero value for the load's basic type
1080 // (This is one of the few places where a generic PhaseTransform
1081 // can create new nodes. Think of it as lazily manifesting
1082 // virtually pre-existing constants.)
1083 return phase->zerocon(memory_type());
1084 }
1085
1086 // A load from an initialization barrier can match a captured store.
1087 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1088 InitializeNode* init = st->in(0)->as_Initialize();
1089 AllocateNode* alloc = init->allocation();
1090 if ((alloc != NULL) && (alloc == ld_alloc)) {
1091 // examine a captured store value
1092 st = init->find_captured_store(ld_off, memory_size(), phase);
1093 if (st != NULL) {
1094 continue; // take one more trip around
1095 }
1096 }
1097 }
1098
1099 // Load boxed value from result of valueOf() call is input parameter.
1100 if (this->is_Load() && ld_adr->is_AddP() &&
1101 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1102 intptr_t ignore = 0;
1103 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1104 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1105 base = bs->step_over_gc_barrier(base);
1106 if (base != NULL && base->is_Proj() &&
1107 base->as_Proj()->_con == TypeFunc::Parms &&
1108 base->in(0)->is_CallStaticJava() &&
1109 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1110 return base->in(0)->in(TypeFunc::Parms);
1111 }
1112 }
1113
1114 break;
1115 }
1116
1117 return NULL;
1118 }
1119
1120 //----------------------is_instance_field_load_with_local_phi------------------
1121 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1122 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1123 in(Address)->is_AddP() ) {
1124 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1125 // Only instances and boxed values.
1126 if( t_oop != NULL &&
1127 (t_oop->is_ptr_to_boxed_value() ||
1128 t_oop->is_known_instance_field()) &&
1129 t_oop->offset() != Type::OffsetBot &&
1130 t_oop->offset() != Type::OffsetTop) {
1131 return true;
1132 }
1133 }
1134 return false;
1135 }
1136
1137 //------------------------------Identity---------------------------------------
1138 // Loads are identity if previous store is to same address
1139 Node* LoadNode::Identity(PhaseGVN* phase) {
1140 // If the previous store-maker is the right kind of Store, and the store is
1141 // to the same address, then we are equal to the value stored.
1142 Node* mem = in(Memory);
1143 Node* value = can_see_stored_value(mem, phase);
1144 if( value ) {
1145 // byte, short & char stores truncate naturally.
1146 // A load has to load the truncated value which requires
1147 // some sort of masking operation and that requires an
1148 // Ideal call instead of an Identity call.
1149 if (memory_size() < BytesPerInt) {
1150 // If the input to the store does not fit with the load's result type,
1151 // it must be truncated via an Ideal call.
1152 if (!phase->type(value)->higher_equal(phase->type(this)))
1153 return this;
1154 }
1155 // (This works even when value is a Con, but LoadNode::Value
1156 // usually runs first, producing the singleton type of the Con.)
1157 return value;
1158 }
1159
1160 // Search for an existing data phi which was generated before for the same
1161 // instance's field to avoid infinite generation of phis in a loop.
1162 Node *region = mem->in(0);
1163 if (is_instance_field_load_with_local_phi(region)) {
1164 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1165 int this_index = phase->C->get_alias_index(addr_t);
1166 int this_offset = addr_t->offset();
1167 int this_iid = addr_t->instance_id();
1168 if (!addr_t->is_known_instance() &&
1169 addr_t->is_ptr_to_boxed_value()) {
1170 // Use _idx of address base (could be Phi node) for boxed values.
1171 intptr_t ignore = 0;
1172 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1173 if (base == NULL) {
1174 return this;
1175 }
1176 this_iid = base->_idx;
1177 }
1178 const Type* this_type = bottom_type();
1179 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1180 Node* phi = region->fast_out(i);
1181 if (phi->is_Phi() && phi != mem &&
1182 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1183 return phi;
1184 }
1185 }
1186 }
1187
1188 return this;
1189 }
1190
1191 // Construct an equivalent unsigned load.
1192 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1193 BasicType bt = T_ILLEGAL;
1194 const Type* rt = NULL;
1195 switch (Opcode()) {
1196 case Op_LoadUB: return this;
1197 case Op_LoadUS: return this;
1198 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1199 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1200 default:
1201 assert(false, "no unsigned variant: %s", Name());
1202 return NULL;
1203 }
1204 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1205 raw_adr_type(), rt, bt, _mo, _control_dependency,
1206 is_unaligned_access(), is_mismatched_access());
1207 }
1208
1209 // Construct an equivalent signed load.
1210 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1211 BasicType bt = T_ILLEGAL;
1212 const Type* rt = NULL;
1213 switch (Opcode()) {
1214 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1215 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1216 case Op_LoadB: // fall through
1217 case Op_LoadS: // fall through
1218 case Op_LoadI: // fall through
1219 case Op_LoadL: return this;
1220 default:
1221 assert(false, "no signed variant: %s", Name());
1222 return NULL;
1223 }
1224 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1225 raw_adr_type(), rt, bt, _mo, _control_dependency,
1226 is_unaligned_access(), is_mismatched_access());
1227 }
1228
1229 // We're loading from an object which has autobox behaviour.
1230 // If this object is result of a valueOf call we'll have a phi
1231 // merging a newly allocated object and a load from the cache.
1232 // We want to replace this load with the original incoming
1233 // argument to the valueOf call.
1234 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1235 assert(phase->C->eliminate_boxing(), "sanity");
1236 intptr_t ignore = 0;
1237 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1238 if ((base == NULL) || base->is_Phi()) {
1239 // Push the loads from the phi that comes from valueOf up
1240 // through it to allow elimination of the loads and the recovery
1241 // of the original value. It is done in split_through_phi().
1242 return NULL;
1243 } else if (base->is_Load() ||
1244 (base->is_DecodeN() && base->in(1)->is_Load())) {
1245 // Eliminate the load of boxed value for integer types from the cache
1246 // array by deriving the value from the index into the array.
1247 // Capture the offset of the load and then reverse the computation.
1248
1249 // Get LoadN node which loads a boxing object from 'cache' array.
1250 if (base->is_DecodeN()) {
1251 base = base->in(1);
1252 }
1253 if (!base->in(Address)->is_AddP()) {
1254 return NULL; // Complex address
1255 }
1256 AddPNode* address = base->in(Address)->as_AddP();
1257 Node* cache_base = address->in(AddPNode::Base);
1258 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1259 // Get ConP node which is static 'cache' field.
1260 cache_base = cache_base->in(1);
1261 }
1262 if ((cache_base != NULL) && cache_base->is_Con()) {
1263 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1264 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1265 Node* elements[4];
1266 int shift = exact_log2(type2aelembytes(T_OBJECT));
1267 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1268 if (count > 0 && elements[0]->is_Con() &&
1269 (count == 1 ||
1270 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1271 elements[1]->in(2) == phase->intcon(shift)))) {
1272 ciObjArray* array = base_type->const_oop()->as_obj_array();
1273 // Fetch the box object cache[0] at the base of the array and get its value
1274 ciInstance* box = array->obj_at(0)->as_instance();
1275 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1276 assert(ik->is_box_klass(), "sanity");
1277 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1278 if (ik->nof_nonstatic_fields() == 1) {
1279 // This should be true nonstatic_field_at requires calling
1280 // nof_nonstatic_fields so check it anyway
1281 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1282 BasicType bt = c.basic_type();
1283 // Only integer types have boxing cache.
1284 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1285 bt == T_BYTE || bt == T_SHORT ||
1286 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1287 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1288 if (cache_low != (int)cache_low) {
1289 return NULL; // should not happen since cache is array indexed by value
1290 }
1291 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1292 if (offset != (int)offset) {
1293 return NULL; // should not happen since cache is array indexed by value
1294 }
1295 // Add up all the offsets making of the address of the load
1296 Node* result = elements[0];
1297 for (int i = 1; i < count; i++) {
1298 result = phase->transform(new AddXNode(result, elements[i]));
1299 }
1300 // Remove the constant offset from the address and then
1301 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1302 // remove the scaling of the offset to recover the original index.
1303 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1304 // Peel the shift off directly but wrap it in a dummy node
1305 // since Ideal can't return existing nodes
1306 result = new RShiftXNode(result->in(1), phase->intcon(0));
1307 } else if (result->is_Add() && result->in(2)->is_Con() &&
1308 result->in(1)->Opcode() == Op_LShiftX &&
1309 result->in(1)->in(2) == phase->intcon(shift)) {
1310 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1311 // but for boxing cache access we know that X<<Z will not overflow
1312 // (there is range check) so we do this optimizatrion by hand here.
1313 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1314 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1315 } else {
1316 result = new RShiftXNode(result, phase->intcon(shift));
1317 }
1318 #ifdef _LP64
1319 if (bt != T_LONG) {
1320 result = new ConvL2INode(phase->transform(result));
1321 }
1322 #else
1323 if (bt == T_LONG) {
1324 result = new ConvI2LNode(phase->transform(result));
1325 }
1326 #endif
1327 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1328 // Need to preserve unboxing load type if it is unsigned.
1329 switch(this->Opcode()) {
1330 case Op_LoadUB:
1331 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1332 break;
1333 case Op_LoadUS:
1334 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1335 break;
1336 }
1337 return result;
1338 }
1339 }
1340 }
1341 }
1342 }
1343 return NULL;
1344 }
1345
1346 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1347 Node* region = phi->in(0);
1348 if (region == NULL) {
1349 return false; // Wait stable graph
1350 }
1351 uint cnt = phi->req();
1352 for (uint i = 1; i < cnt; i++) {
1353 Node* rc = region->in(i);
1354 if (rc == NULL || phase->type(rc) == Type::TOP)
1355 return false; // Wait stable graph
1356 Node* in = phi->in(i);
1357 if (in == NULL || phase->type(in) == Type::TOP)
1358 return false; // Wait stable graph
1359 }
1360 return true;
1361 }
1362 //------------------------------split_through_phi------------------------------
1363 // Split instance or boxed field load through Phi.
1364 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1365 Node* mem = in(Memory);
1366 Node* address = in(Address);
1367 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1368
1369 assert((t_oop != NULL) &&
1370 (t_oop->is_known_instance_field() ||
1371 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1372
1373 Compile* C = phase->C;
1374 intptr_t ignore = 0;
1375 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1376 bool base_is_phi = (base != NULL) && base->is_Phi();
1377 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1378 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1379 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1380
1381 if (!((mem->is_Phi() || base_is_phi) &&
1382 (load_boxed_values || t_oop->is_known_instance_field()))) {
1383 return NULL; // memory is not Phi
1384 }
1385
1386 if (mem->is_Phi()) {
1387 if (!stable_phi(mem->as_Phi(), phase)) {
1388 return NULL; // Wait stable graph
1389 }
1390 uint cnt = mem->req();
1391 // Check for loop invariant memory.
1392 if (cnt == 3) {
1393 for (uint i = 1; i < cnt; i++) {
1394 Node* in = mem->in(i);
1395 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1396 if (m == mem) {
1397 set_req(Memory, mem->in(cnt - i));
1398 return this; // made change
1399 }
1400 }
1401 }
1402 }
1403 if (base_is_phi) {
1404 if (!stable_phi(base->as_Phi(), phase)) {
1405 return NULL; // Wait stable graph
1406 }
1407 uint cnt = base->req();
1408 // Check for loop invariant memory.
1409 if (cnt == 3) {
1410 for (uint i = 1; i < cnt; i++) {
1411 if (base->in(i) == base) {
1412 return NULL; // Wait stable graph
1413 }
1414 }
1415 }
1416 }
1417
1418 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1419
1420 // Split through Phi (see original code in loopopts.cpp).
1421 assert(C->have_alias_type(t_oop), "instance should have alias type");
1422
1423 // Do nothing here if Identity will find a value
1424 // (to avoid infinite chain of value phis generation).
1425 if (!phase->eqv(this, phase->apply_identity(this)))
1426 return NULL;
1427
1428 // Select Region to split through.
1429 Node* region;
1430 if (!base_is_phi) {
1431 assert(mem->is_Phi(), "sanity");
1432 region = mem->in(0);
1433 // Skip if the region dominates some control edge of the address.
1434 if (!MemNode::all_controls_dominate(address, region))
1435 return NULL;
1436 } else if (!mem->is_Phi()) {
1437 assert(base_is_phi, "sanity");
1438 region = base->in(0);
1439 // Skip if the region dominates some control edge of the memory.
1440 if (!MemNode::all_controls_dominate(mem, region))
1441 return NULL;
1442 } else if (base->in(0) != mem->in(0)) {
1443 assert(base_is_phi && mem->is_Phi(), "sanity");
1444 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1445 region = base->in(0);
1446 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1447 region = mem->in(0);
1448 } else {
1449 return NULL; // complex graph
1450 }
1451 } else {
1452 assert(base->in(0) == mem->in(0), "sanity");
1453 region = mem->in(0);
1454 }
1455
1456 const Type* this_type = this->bottom_type();
1457 int this_index = C->get_alias_index(t_oop);
1458 int this_offset = t_oop->offset();
1459 int this_iid = t_oop->instance_id();
1460 if (!t_oop->is_known_instance() && load_boxed_values) {
1461 // Use _idx of address base for boxed values.
1462 this_iid = base->_idx;
1463 }
1464 PhaseIterGVN* igvn = phase->is_IterGVN();
1465 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1466 for (uint i = 1; i < region->req(); i++) {
1467 Node* x;
1468 Node* the_clone = NULL;
1469 if (region->in(i) == C->top()) {
1470 x = C->top(); // Dead path? Use a dead data op
1471 } else {
1472 x = this->clone(); // Else clone up the data op
1473 the_clone = x; // Remember for possible deletion.
1474 // Alter data node to use pre-phi inputs
1475 if (this->in(0) == region) {
1476 x->set_req(0, region->in(i));
1477 } else {
1478 x->set_req(0, NULL);
1479 }
1480 if (mem->is_Phi() && (mem->in(0) == region)) {
1481 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1482 }
1483 if (address->is_Phi() && address->in(0) == region) {
1484 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1485 }
1486 if (base_is_phi && (base->in(0) == region)) {
1487 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1488 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1489 x->set_req(Address, adr_x);
1490 }
1491 }
1492 // Check for a 'win' on some paths
1493 const Type *t = x->Value(igvn);
1494
1495 bool singleton = t->singleton();
1496
1497 // See comments in PhaseIdealLoop::split_thru_phi().
1498 if (singleton && t == Type::TOP) {
1499 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1500 }
1501
1502 if (singleton) {
1503 x = igvn->makecon(t);
1504 } else {
1505 // We now call Identity to try to simplify the cloned node.
1506 // Note that some Identity methods call phase->type(this).
1507 // Make sure that the type array is big enough for
1508 // our new node, even though we may throw the node away.
1509 // (This tweaking with igvn only works because x is a new node.)
1510 igvn->set_type(x, t);
1511 // If x is a TypeNode, capture any more-precise type permanently into Node
1512 // otherwise it will be not updated during igvn->transform since
1513 // igvn->type(x) is set to x->Value() already.
1514 x->raise_bottom_type(t);
1515 Node *y = igvn->apply_identity(x);
1516 if (y != x) {
1517 x = y;
1518 } else {
1519 y = igvn->hash_find_insert(x);
1520 if (y) {
1521 x = y;
1522 } else {
1523 // Else x is a new node we are keeping
1524 // We do not need register_new_node_with_optimizer
1525 // because set_type has already been called.
1526 igvn->_worklist.push(x);
1527 }
1528 }
1529 }
1530 if (x != the_clone && the_clone != NULL) {
1531 igvn->remove_dead_node(the_clone);
1532 }
1533 phi->set_req(i, x);
1534 }
1535 // Record Phi
1536 igvn->register_new_node_with_optimizer(phi);
1537 return phi;
1538 }
1539
1540 //------------------------------Ideal------------------------------------------
1541 // If the load is from Field memory and the pointer is non-null, it might be possible to
1542 // zero out the control input.
1543 // If the offset is constant and the base is an object allocation,
1544 // try to hook me up to the exact initializing store.
1545 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1546 Node* p = MemNode::Ideal_common(phase, can_reshape);
1547 if (p) return (p == NodeSentinel) ? NULL : p;
1548
1549 Node* ctrl = in(MemNode::Control);
1550 Node* address = in(MemNode::Address);
1551 bool progress = false;
1552
1553 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1554 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1555
1556 // Skip up past a SafePoint control. Cannot do this for Stores because
1557 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1558 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1559 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1560 !addr_mark ) {
1561 ctrl = ctrl->in(0);
1562 set_req(MemNode::Control,ctrl);
1563 progress = true;
1564 }
1565
1566 intptr_t ignore = 0;
1567 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1568 if (base != NULL
1569 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1570 // Check for useless control edge in some common special cases
1571 if (in(MemNode::Control) != NULL
1572 && can_remove_control()
1573 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1574 && all_controls_dominate(base, phase->C->start())) {
1575 // A method-invariant, non-null address (constant or 'this' argument).
1576 set_req(MemNode::Control, NULL);
1577 progress = true;
1578 }
1579 }
1580
1581 Node* mem = in(MemNode::Memory);
1582 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1583
1584 if (can_reshape && (addr_t != NULL)) {
1585 // try to optimize our memory input
1586 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1587 if (opt_mem != mem) {
1588 set_req(MemNode::Memory, opt_mem);
1589 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1590 return this;
1591 }
1592 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1593 if ((t_oop != NULL) &&
1594 (t_oop->is_known_instance_field() ||
1595 t_oop->is_ptr_to_boxed_value())) {
1596 PhaseIterGVN *igvn = phase->is_IterGVN();
1597 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1598 // Delay this transformation until memory Phi is processed.
1599 phase->is_IterGVN()->_worklist.push(this);
1600 return NULL;
1601 }
1602 // Split instance field load through Phi.
1603 Node* result = split_through_phi(phase);
1604 if (result != NULL) return result;
1605
1606 if (t_oop->is_ptr_to_boxed_value()) {
1607 Node* result = eliminate_autobox(phase);
1608 if (result != NULL) return result;
1609 }
1610 }
1611 }
1612
1613 // Is there a dominating load that loads the same value? Leave
1614 // anything that is not a load of a field/array element (like
1615 // barriers etc.) alone
1616 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1617 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1618 Node *use = mem->fast_out(i);
1619 if (use != this &&
1620 use->Opcode() == Opcode() &&
1621 use->in(0) != NULL &&
1622 use->in(0) != in(0) &&
1623 use->in(Address) == in(Address)) {
1624 Node* ctl = in(0);
1625 for (int i = 0; i < 10 && ctl != NULL; i++) {
1626 ctl = IfNode::up_one_dom(ctl);
1627 if (ctl == use->in(0)) {
1628 set_req(0, use->in(0));
1629 return this;
1630 }
1631 }
1632 }
1633 }
1634 }
1635
1636 // Check for prior store with a different base or offset; make Load
1637 // independent. Skip through any number of them. Bail out if the stores
1638 // are in an endless dead cycle and report no progress. This is a key
1639 // transform for Reflection. However, if after skipping through the Stores
1640 // we can't then fold up against a prior store do NOT do the transform as
1641 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1642 // array memory alive twice: once for the hoisted Load and again after the
1643 // bypassed Store. This situation only works if EVERYBODY who does
1644 // anti-dependence work knows how to bypass. I.e. we need all
1645 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1646 // the alias index stuff. So instead, peek through Stores and IFF we can
1647 // fold up, do so.
1648 Node* prev_mem = find_previous_store(phase);
1649 if (prev_mem != NULL) {
1650 Node* value = can_see_arraycopy_value(prev_mem, phase);
1651 if (value != NULL) {
1652 return value;
1653 }
1654 }
1655 // Steps (a), (b): Walk past independent stores to find an exact match.
1656 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1657 // (c) See if we can fold up on the spot, but don't fold up here.
1658 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1659 // just return a prior value, which is done by Identity calls.
1660 if (can_see_stored_value(prev_mem, phase)) {
1661 // Make ready for step (d):
1662 set_req(MemNode::Memory, prev_mem);
1663 return this;
1664 }
1665 }
1666
1667 return progress ? this : NULL;
1668 }
1669
1670 // Helper to recognize certain Klass fields which are invariant across
1671 // some group of array types (e.g., int[] or all T[] where T < Object).
1672 const Type*
1673 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1674 ciKlass* klass) const {
1675 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1676 // The field is Klass::_modifier_flags. Return its (constant) value.
1677 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1678 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1679 return TypeInt::make(klass->modifier_flags());
1680 }
1681 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1682 // The field is Klass::_access_flags. Return its (constant) value.
1683 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1684 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1685 return TypeInt::make(klass->access_flags());
1686 }
1687 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1688 // The field is Klass::_layout_helper. Return its constant value if known.
1689 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1690 return TypeInt::make(klass->layout_helper());
1691 }
1692
1693 // No match.
1694 return NULL;
1695 }
1696
1697 //------------------------------Value-----------------------------------------
1698 const Type* LoadNode::Value(PhaseGVN* phase) const {
1699 // Either input is TOP ==> the result is TOP
1700 Node* mem = in(MemNode::Memory);
1701 const Type *t1 = phase->type(mem);
1702 if (t1 == Type::TOP) return Type::TOP;
1703 Node* adr = in(MemNode::Address);
1704 const TypePtr* tp = phase->type(adr)->isa_ptr();
1705 if (tp == NULL || tp->empty()) return Type::TOP;
1706 int off = tp->offset();
1707 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1708 Compile* C = phase->C;
1709
1710 // Try to guess loaded type from pointer type
1711 if (tp->isa_aryptr()) {
1712 const TypeAryPtr* ary = tp->is_aryptr();
1713 const Type* t = ary->elem();
1714
1715 // Determine whether the reference is beyond the header or not, by comparing
1716 // the offset against the offset of the start of the array's data.
1717 // Different array types begin at slightly different offsets (12 vs. 16).
1718 // We choose T_BYTE as an example base type that is least restrictive
1719 // as to alignment, which will therefore produce the smallest
1720 // possible base offset.
1721 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1722 const bool off_beyond_header = (off >= min_base_off);
1723
1724 // Try to constant-fold a stable array element.
1725 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1726 // Make sure the reference is not into the header and the offset is constant
1727 ciObject* aobj = ary->const_oop();
1728 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1729 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1730 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1731 stable_dimension,
1732 memory_type(), is_unsigned());
1733 if (con_type != NULL) {
1734 return con_type;
1735 }
1736 }
1737 }
1738
1739 // Don't do this for integer types. There is only potential profit if
1740 // the element type t is lower than _type; that is, for int types, if _type is
1741 // more restrictive than t. This only happens here if one is short and the other
1742 // char (both 16 bits), and in those cases we've made an intentional decision
1743 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1744 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1745 //
1746 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1747 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1748 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1749 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1750 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1751 // In fact, that could have been the original type of p1, and p1 could have
1752 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1753 // expression (LShiftL quux 3) independently optimized to the constant 8.
1754 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1755 && (_type->isa_vect() == NULL)
1756 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1757 // t might actually be lower than _type, if _type is a unique
1758 // concrete subclass of abstract class t.
1759 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
1760 const Type* jt = t->join_speculative(_type);
1761 // In any case, do not allow the join, per se, to empty out the type.
1762 if (jt->empty() && !t->empty()) {
1763 // This can happen if a interface-typed array narrows to a class type.
1764 jt = _type;
1765 }
1766 #ifdef ASSERT
1767 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1768 // The pointers in the autobox arrays are always non-null
1769 Node* base = adr->in(AddPNode::Base);
1770 if ((base != NULL) && base->is_DecodeN()) {
1771 // Get LoadN node which loads IntegerCache.cache field
1772 base = base->in(1);
1773 }
1774 if ((base != NULL) && base->is_Con()) {
1775 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1776 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1777 // It could be narrow oop
1778 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1779 }
1780 }
1781 }
1782 #endif
1783 return jt;
1784 }
1785 }
1786 } else if (tp->base() == Type::InstPtr) {
1787 assert( off != Type::OffsetBot ||
1788 // arrays can be cast to Objects
1789 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1790 // unsafe field access may not have a constant offset
1791 C->has_unsafe_access(),
1792 "Field accesses must be precise" );
1793 // For oop loads, we expect the _type to be precise.
1794
1795 // Optimize loads from constant fields.
1796 const TypeInstPtr* tinst = tp->is_instptr();
1797 ciObject* const_oop = tinst->const_oop();
1798 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1799 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1800 if (con_type != NULL) {
1801 return con_type;
1802 }
1803 }
1804 } else if (tp->base() == Type::KlassPtr) {
1805 assert( off != Type::OffsetBot ||
1806 // arrays can be cast to Objects
1807 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1808 // also allow array-loading from the primary supertype
1809 // array during subtype checks
1810 Opcode() == Op_LoadKlass,
1811 "Field accesses must be precise" );
1812 // For klass/static loads, we expect the _type to be precise
1813 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
1814 /* With mirrors being an indirect in the Klass*
1815 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1816 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1817 *
1818 * So check the type and klass of the node before the LoadP.
1819 */
1820 Node* adr2 = adr->in(MemNode::Address);
1821 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1822 if (tkls != NULL && !StressReflectiveCode) {
1823 ciKlass* klass = tkls->klass();
1824 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1825 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1826 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1827 return TypeInstPtr::make(klass->java_mirror());
1828 }
1829 }
1830 }
1831
1832 const TypeKlassPtr *tkls = tp->isa_klassptr();
1833 if (tkls != NULL && !StressReflectiveCode) {
1834 ciKlass* klass = tkls->klass();
1835 if (klass->is_loaded() && tkls->klass_is_exact()) {
1836 // We are loading a field from a Klass metaobject whose identity
1837 // is known at compile time (the type is "exact" or "precise").
1838 // Check for fields we know are maintained as constants by the VM.
1839 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1840 // The field is Klass::_super_check_offset. Return its (constant) value.
1841 // (Folds up type checking code.)
1842 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1843 return TypeInt::make(klass->super_check_offset());
1844 }
1845 // Compute index into primary_supers array
1846 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1847 // Check for overflowing; use unsigned compare to handle the negative case.
1848 if( depth < ciKlass::primary_super_limit() ) {
1849 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1850 // (Folds up type checking code.)
1851 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1852 ciKlass *ss = klass->super_of_depth(depth);
1853 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1854 }
1855 const Type* aift = load_array_final_field(tkls, klass);
1856 if (aift != NULL) return aift;
1857 }
1858
1859 // We can still check if we are loading from the primary_supers array at a
1860 // shallow enough depth. Even though the klass is not exact, entries less
1861 // than or equal to its super depth are correct.
1862 if (klass->is_loaded() ) {
1863 ciType *inner = klass;
1864 while( inner->is_obj_array_klass() )
1865 inner = inner->as_obj_array_klass()->base_element_type();
1866 if( inner->is_instance_klass() &&
1867 !inner->as_instance_klass()->flags().is_interface() ) {
1868 // Compute index into primary_supers array
1869 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1870 // Check for overflowing; use unsigned compare to handle the negative case.
1871 if( depth < ciKlass::primary_super_limit() &&
1872 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1873 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1874 // (Folds up type checking code.)
1875 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1876 ciKlass *ss = klass->super_of_depth(depth);
1877 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1878 }
1879 }
1880 }
1881
1882 // If the type is enough to determine that the thing is not an array,
1883 // we can give the layout_helper a positive interval type.
1884 // This will help short-circuit some reflective code.
1885 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1886 && !klass->is_array_klass() // not directly typed as an array
1887 && !klass->is_interface() // specifically not Serializable & Cloneable
1888 && !klass->is_java_lang_Object() // not the supertype of all T[]
1889 ) {
1890 // Note: When interfaces are reliable, we can narrow the interface
1891 // test to (klass != Serializable && klass != Cloneable).
1892 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1893 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1894 // The key property of this type is that it folds up tests
1895 // for array-ness, since it proves that the layout_helper is positive.
1896 // Thus, a generic value like the basic object layout helper works fine.
1897 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1898 }
1899 }
1900
1901 // If we are loading from a freshly-allocated object, produce a zero,
1902 // if the load is provably beyond the header of the object.
1903 // (Also allow a variable load from a fresh array to produce zero.)
1904 const TypeOopPtr *tinst = tp->isa_oopptr();
1905 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1906 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1907 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1908 Node* value = can_see_stored_value(mem,phase);
1909 if (value != NULL && value->is_Con()) {
1910 assert(value->bottom_type()->higher_equal(_type),"sanity");
1911 return value->bottom_type();
1912 }
1913 }
1914
1915 if (is_instance) {
1916 // If we have an instance type and our memory input is the
1917 // programs's initial memory state, there is no matching store,
1918 // so just return a zero of the appropriate type
1919 Node *mem = in(MemNode::Memory);
1920 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1921 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1922 return Type::get_zero_type(_type->basic_type());
1923 }
1924 }
1925 return _type;
1926 }
1927
1928 //------------------------------match_edge-------------------------------------
1929 // Do we Match on this edge index or not? Match only the address.
1930 uint LoadNode::match_edge(uint idx) const {
1931 return idx == MemNode::Address;
1932 }
1933
1934 //--------------------------LoadBNode::Ideal--------------------------------------
1935 //
1936 // If the previous store is to the same address as this load,
1937 // and the value stored was larger than a byte, replace this load
1938 // with the value stored truncated to a byte. If no truncation is
1939 // needed, the replacement is done in LoadNode::Identity().
1940 //
1941 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1942 Node* mem = in(MemNode::Memory);
1943 Node* value = can_see_stored_value(mem,phase);
1944 if( value && !phase->type(value)->higher_equal( _type ) ) {
1945 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1946 return new RShiftINode(result, phase->intcon(24));
1947 }
1948 // Identity call will handle the case where truncation is not needed.
1949 return LoadNode::Ideal(phase, can_reshape);
1950 }
1951
1952 const Type* LoadBNode::Value(PhaseGVN* phase) const {
1953 Node* mem = in(MemNode::Memory);
1954 Node* value = can_see_stored_value(mem,phase);
1955 if (value != NULL && value->is_Con() &&
1956 !value->bottom_type()->higher_equal(_type)) {
1957 // If the input to the store does not fit with the load's result type,
1958 // it must be truncated. We can't delay until Ideal call since
1959 // a singleton Value is needed for split_thru_phi optimization.
1960 int con = value->get_int();
1961 return TypeInt::make((con << 24) >> 24);
1962 }
1963 return LoadNode::Value(phase);
1964 }
1965
1966 //--------------------------LoadUBNode::Ideal-------------------------------------
1967 //
1968 // If the previous store is to the same address as this load,
1969 // and the value stored was larger than a byte, replace this load
1970 // with the value stored truncated to a byte. If no truncation is
1971 // needed, the replacement is done in LoadNode::Identity().
1972 //
1973 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1974 Node* mem = in(MemNode::Memory);
1975 Node* value = can_see_stored_value(mem, phase);
1976 if (value && !phase->type(value)->higher_equal(_type))
1977 return new AndINode(value, phase->intcon(0xFF));
1978 // Identity call will handle the case where truncation is not needed.
1979 return LoadNode::Ideal(phase, can_reshape);
1980 }
1981
1982 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
1983 Node* mem = in(MemNode::Memory);
1984 Node* value = can_see_stored_value(mem,phase);
1985 if (value != NULL && value->is_Con() &&
1986 !value->bottom_type()->higher_equal(_type)) {
1987 // If the input to the store does not fit with the load's result type,
1988 // it must be truncated. We can't delay until Ideal call since
1989 // a singleton Value is needed for split_thru_phi optimization.
1990 int con = value->get_int();
1991 return TypeInt::make(con & 0xFF);
1992 }
1993 return LoadNode::Value(phase);
1994 }
1995
1996 //--------------------------LoadUSNode::Ideal-------------------------------------
1997 //
1998 // If the previous store is to the same address as this load,
1999 // and the value stored was larger than a char, replace this load
2000 // with the value stored truncated to a char. If no truncation is
2001 // needed, the replacement is done in LoadNode::Identity().
2002 //
2003 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2004 Node* mem = in(MemNode::Memory);
2005 Node* value = can_see_stored_value(mem,phase);
2006 if( value && !phase->type(value)->higher_equal( _type ) )
2007 return new AndINode(value,phase->intcon(0xFFFF));
2008 // Identity call will handle the case where truncation is not needed.
2009 return LoadNode::Ideal(phase, can_reshape);
2010 }
2011
2012 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2013 Node* mem = in(MemNode::Memory);
2014 Node* value = can_see_stored_value(mem,phase);
2015 if (value != NULL && value->is_Con() &&
2016 !value->bottom_type()->higher_equal(_type)) {
2017 // If the input to the store does not fit with the load's result type,
2018 // it must be truncated. We can't delay until Ideal call since
2019 // a singleton Value is needed for split_thru_phi optimization.
2020 int con = value->get_int();
2021 return TypeInt::make(con & 0xFFFF);
2022 }
2023 return LoadNode::Value(phase);
2024 }
2025
2026 //--------------------------LoadSNode::Ideal--------------------------------------
2027 //
2028 // If the previous store is to the same address as this load,
2029 // and the value stored was larger than a short, replace this load
2030 // with the value stored truncated to a short. If no truncation is
2031 // needed, the replacement is done in LoadNode::Identity().
2032 //
2033 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2034 Node* mem = in(MemNode::Memory);
2035 Node* value = can_see_stored_value(mem,phase);
2036 if( value && !phase->type(value)->higher_equal( _type ) ) {
2037 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2038 return new RShiftINode(result, phase->intcon(16));
2039 }
2040 // Identity call will handle the case where truncation is not needed.
2041 return LoadNode::Ideal(phase, can_reshape);
2042 }
2043
2044 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2045 Node* mem = in(MemNode::Memory);
2046 Node* value = can_see_stored_value(mem,phase);
2047 if (value != NULL && value->is_Con() &&
2048 !value->bottom_type()->higher_equal(_type)) {
2049 // If the input to the store does not fit with the load's result type,
2050 // it must be truncated. We can't delay until Ideal call since
2051 // a singleton Value is needed for split_thru_phi optimization.
2052 int con = value->get_int();
2053 return TypeInt::make((con << 16) >> 16);
2054 }
2055 return LoadNode::Value(phase);
2056 }
2057
2058 //=============================================================================
2059 //----------------------------LoadKlassNode::make------------------------------
2060 // Polymorphic factory method:
2061 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2062 // sanity check the alias category against the created node type
2063 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2064 assert(adr_type != NULL, "expecting TypeKlassPtr");
2065 #ifdef _LP64
2066 if (adr_type->is_ptr_to_narrowklass()) {
2067 assert(UseCompressedClassPointers, "no compressed klasses");
2068 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2069 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2070 }
2071 #endif
2072 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2073 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2074 }
2075
2076 //------------------------------Value------------------------------------------
2077 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2078 return klass_value_common(phase);
2079 }
2080
2081 // In most cases, LoadKlassNode does not have the control input set. If the control
2082 // input is set, it must not be removed (by LoadNode::Ideal()).
2083 bool LoadKlassNode::can_remove_control() const {
2084 return false;
2085 }
2086
2087 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2088 // Either input is TOP ==> the result is TOP
2089 const Type *t1 = phase->type( in(MemNode::Memory) );
2090 if (t1 == Type::TOP) return Type::TOP;
2091 Node *adr = in(MemNode::Address);
2092 const Type *t2 = phase->type( adr );
2093 if (t2 == Type::TOP) return Type::TOP;
2094 const TypePtr *tp = t2->is_ptr();
2095 if (TypePtr::above_centerline(tp->ptr()) ||
2096 tp->ptr() == TypePtr::Null) return Type::TOP;
2097
2098 // Return a more precise klass, if possible
2099 const TypeInstPtr *tinst = tp->isa_instptr();
2100 if (tinst != NULL) {
2101 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2102 int offset = tinst->offset();
2103 if (ik == phase->C->env()->Class_klass()
2104 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2105 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2106 // We are loading a special hidden field from a Class mirror object,
2107 // the field which points to the VM's Klass metaobject.
2108 ciType* t = tinst->java_mirror_type();
2109 // java_mirror_type returns non-null for compile-time Class constants.
2110 if (t != NULL) {
2111 // constant oop => constant klass
2112 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2113 if (t->is_void()) {
2114 // We cannot create a void array. Since void is a primitive type return null
2115 // klass. Users of this result need to do a null check on the returned klass.
2116 return TypePtr::NULL_PTR;
2117 }
2118 return TypeKlassPtr::make(ciArrayKlass::make(t));
2119 }
2120 if (!t->is_klass()) {
2121 // a primitive Class (e.g., int.class) has NULL for a klass field
2122 return TypePtr::NULL_PTR;
2123 }
2124 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2125 return TypeKlassPtr::make(t->as_klass());
2126 }
2127 // non-constant mirror, so we can't tell what's going on
2128 }
2129 if( !ik->is_loaded() )
2130 return _type; // Bail out if not loaded
2131 if (offset == oopDesc::klass_offset_in_bytes()) {
2132 if (tinst->klass_is_exact()) {
2133 return TypeKlassPtr::make(ik);
2134 }
2135 // See if we can become precise: no subklasses and no interface
2136 // (Note: We need to support verified interfaces.)
2137 if (!ik->is_interface() && !ik->has_subklass()) {
2138 //assert(!UseExactTypes, "this code should be useless with exact types");
2139 // Add a dependence; if any subclass added we need to recompile
2140 if (!ik->is_final()) {
2141 // %%% should use stronger assert_unique_concrete_subtype instead
2142 phase->C->dependencies()->assert_leaf_type(ik);
2143 }
2144 // Return precise klass
2145 return TypeKlassPtr::make(ik);
2146 }
2147
2148 // Return root of possible klass
2149 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2150 }
2151 }
2152
2153 // Check for loading klass from an array
2154 const TypeAryPtr *tary = tp->isa_aryptr();
2155 if( tary != NULL ) {
2156 ciKlass *tary_klass = tary->klass();
2157 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2158 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2159 if (tary->klass_is_exact()) {
2160 return TypeKlassPtr::make(tary_klass);
2161 }
2162 ciArrayKlass *ak = tary->klass()->as_array_klass();
2163 // If the klass is an object array, we defer the question to the
2164 // array component klass.
2165 if( ak->is_obj_array_klass() ) {
2166 assert( ak->is_loaded(), "" );
2167 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2168 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2169 ciInstanceKlass* ik = base_k->as_instance_klass();
2170 // See if we can become precise: no subklasses and no interface
2171 if (!ik->is_interface() && !ik->has_subklass()) {
2172 //assert(!UseExactTypes, "this code should be useless with exact types");
2173 // Add a dependence; if any subclass added we need to recompile
2174 if (!ik->is_final()) {
2175 phase->C->dependencies()->assert_leaf_type(ik);
2176 }
2177 // Return precise array klass
2178 return TypeKlassPtr::make(ak);
2179 }
2180 }
2181 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2182 } else { // Found a type-array?
2183 //assert(!UseExactTypes, "this code should be useless with exact types");
2184 assert( ak->is_type_array_klass(), "" );
2185 return TypeKlassPtr::make(ak); // These are always precise
2186 }
2187 }
2188 }
2189
2190 // Check for loading klass from an array klass
2191 const TypeKlassPtr *tkls = tp->isa_klassptr();
2192 if (tkls != NULL && !StressReflectiveCode) {
2193 ciKlass* klass = tkls->klass();
2194 if( !klass->is_loaded() )
2195 return _type; // Bail out if not loaded
2196 if( klass->is_obj_array_klass() &&
2197 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2198 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2199 // // Always returning precise element type is incorrect,
2200 // // e.g., element type could be object and array may contain strings
2201 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2202
2203 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2204 // according to the element type's subclassing.
2205 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2206 }
2207 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2208 tkls->offset() == in_bytes(Klass::super_offset())) {
2209 ciKlass* sup = klass->as_instance_klass()->super();
2210 // The field is Klass::_super. Return its (constant) value.
2211 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2212 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2213 }
2214 }
2215
2216 // Bailout case
2217 return LoadNode::Value(phase);
2218 }
2219
2220 //------------------------------Identity---------------------------------------
2221 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2222 // Also feed through the klass in Allocate(...klass...)._klass.
2223 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2224 return klass_identity_common(phase);
2225 }
2226
2227 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2228 Node* x = LoadNode::Identity(phase);
2229 if (x != this) return x;
2230
2231 // Take apart the address into an oop and and offset.
2232 // Return 'this' if we cannot.
2233 Node* adr = in(MemNode::Address);
2234 intptr_t offset = 0;
2235 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2236 if (base == NULL) return this;
2237 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2238 if (toop == NULL) return this;
2239
2240 // Step over potential GC barrier for OopHandle resolve
2241 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2242 if (bs->is_gc_barrier_node(base)) {
2243 base = bs->step_over_gc_barrier(base);
2244 }
2245
2246 // We can fetch the klass directly through an AllocateNode.
2247 // This works even if the klass is not constant (clone or newArray).
2248 if (offset == oopDesc::klass_offset_in_bytes()) {
2249 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2250 if (allocated_klass != NULL) {
2251 return allocated_klass;
2252 }
2253 }
2254
2255 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2256 // See inline_native_Class_query for occurrences of these patterns.
2257 // Java Example: x.getClass().isAssignableFrom(y)
2258 //
2259 // This improves reflective code, often making the Class
2260 // mirror go completely dead. (Current exception: Class
2261 // mirrors may appear in debug info, but we could clean them out by
2262 // introducing a new debug info operator for Klass.java_mirror).
2263
2264 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2265 && offset == java_lang_Class::klass_offset_in_bytes()) {
2266 if (base->is_Load()) {
2267 Node* base2 = base->in(MemNode::Address);
2268 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2269 Node* adr2 = base2->in(MemNode::Address);
2270 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2271 if (tkls != NULL && !tkls->empty()
2272 && (tkls->klass()->is_instance_klass() ||
2273 tkls->klass()->is_array_klass())
2274 && adr2->is_AddP()
2275 ) {
2276 int mirror_field = in_bytes(Klass::java_mirror_offset());
2277 if (tkls->offset() == mirror_field) {
2278 return adr2->in(AddPNode::Base);
2279 }
2280 }
2281 }
2282 }
2283 }
2284
2285 return this;
2286 }
2287
2288
2289 //------------------------------Value------------------------------------------
2290 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2291 const Type *t = klass_value_common(phase);
2292 if (t == Type::TOP)
2293 return t;
2294
2295 return t->make_narrowklass();
2296 }
2297
2298 //------------------------------Identity---------------------------------------
2299 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2300 // Also feed through the klass in Allocate(...klass...)._klass.
2301 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2302 Node *x = klass_identity_common(phase);
2303
2304 const Type *t = phase->type( x );
2305 if( t == Type::TOP ) return x;
2306 if( t->isa_narrowklass()) return x;
2307 assert (!t->isa_narrowoop(), "no narrow oop here");
2308
2309 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2310 }
2311
2312 //------------------------------Value-----------------------------------------
2313 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2314 // Either input is TOP ==> the result is TOP
2315 const Type *t1 = phase->type( in(MemNode::Memory) );
2316 if( t1 == Type::TOP ) return Type::TOP;
2317 Node *adr = in(MemNode::Address);
2318 const Type *t2 = phase->type( adr );
2319 if( t2 == Type::TOP ) return Type::TOP;
2320 const TypePtr *tp = t2->is_ptr();
2321 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2322 const TypeAryPtr *tap = tp->isa_aryptr();
2323 if( !tap ) return _type;
2324 return tap->size();
2325 }
2326
2327 //-------------------------------Ideal---------------------------------------
2328 // Feed through the length in AllocateArray(...length...)._length.
2329 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2330 Node* p = MemNode::Ideal_common(phase, can_reshape);
2331 if (p) return (p == NodeSentinel) ? NULL : p;
2332
2333 // Take apart the address into an oop and and offset.
2334 // Return 'this' if we cannot.
2335 Node* adr = in(MemNode::Address);
2336 intptr_t offset = 0;
2337 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2338 if (base == NULL) return NULL;
2339 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2340 if (tary == NULL) return NULL;
2341
2342 // We can fetch the length directly through an AllocateArrayNode.
2343 // This works even if the length is not constant (clone or newArray).
2344 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2345 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2346 if (alloc != NULL) {
2347 Node* allocated_length = alloc->Ideal_length();
2348 Node* len = alloc->make_ideal_length(tary, phase);
2349 if (allocated_length != len) {
2350 // New CastII improves on this.
2351 return len;
2352 }
2353 }
2354 }
2355
2356 return NULL;
2357 }
2358
2359 //------------------------------Identity---------------------------------------
2360 // Feed through the length in AllocateArray(...length...)._length.
2361 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2362 Node* x = LoadINode::Identity(phase);
2363 if (x != this) return x;
2364
2365 // Take apart the address into an oop and and offset.
2366 // Return 'this' if we cannot.
2367 Node* adr = in(MemNode::Address);
2368 intptr_t offset = 0;
2369 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2370 if (base == NULL) return this;
2371 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2372 if (tary == NULL) return this;
2373
2374 // We can fetch the length directly through an AllocateArrayNode.
2375 // This works even if the length is not constant (clone or newArray).
2376 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2377 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2378 if (alloc != NULL) {
2379 Node* allocated_length = alloc->Ideal_length();
2380 // Do not allow make_ideal_length to allocate a CastII node.
2381 Node* len = alloc->make_ideal_length(tary, phase, false);
2382 if (allocated_length == len) {
2383 // Return allocated_length only if it would not be improved by a CastII.
2384 return allocated_length;
2385 }
2386 }
2387 }
2388
2389 return this;
2390
2391 }
2392
2393 //=============================================================================
2394 //---------------------------StoreNode::make-----------------------------------
2395 // Polymorphic factory method:
2396 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2397 assert((mo == unordered || mo == release), "unexpected");
2398 Compile* C = gvn.C;
2399 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2400 ctl != NULL, "raw memory operations should have control edge");
2401
2402 switch (bt) {
2403 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2404 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2405 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2406 case T_CHAR:
2407 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2408 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2409 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2410 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2411 case T_METADATA:
2412 case T_ADDRESS:
2413 case T_OBJECT:
2414 #ifdef _LP64
2415 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2416 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2417 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2418 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2419 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2420 adr->bottom_type()->isa_rawptr())) {
2421 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2422 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2423 }
2424 #endif
2425 {
2426 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2427 }
2428 default:
2429 ShouldNotReachHere();
2430 return (StoreNode*)NULL;
2431 }
2432 }
2433
2434 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2435 bool require_atomic = true;
2436 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2437 }
2438
2439 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2440 bool require_atomic = true;
2441 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2442 }
2443
2444
2445 //--------------------------bottom_type----------------------------------------
2446 const Type *StoreNode::bottom_type() const {
2447 return Type::MEMORY;
2448 }
2449
2450 //------------------------------hash-------------------------------------------
2451 uint StoreNode::hash() const {
2452 // unroll addition of interesting fields
2453 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2454
2455 // Since they are not commoned, do not hash them:
2456 return NO_HASH;
2457 }
2458
2459 //------------------------------Ideal------------------------------------------
2460 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2461 // When a store immediately follows a relevant allocation/initialization,
2462 // try to capture it into the initialization, or hoist it above.
2463 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2464 Node* p = MemNode::Ideal_common(phase, can_reshape);
2465 if (p) return (p == NodeSentinel) ? NULL : p;
2466
2467 Node* mem = in(MemNode::Memory);
2468 Node* address = in(MemNode::Address);
2469 // Back-to-back stores to same address? Fold em up. Generally
2470 // unsafe if I have intervening uses... Also disallowed for StoreCM
2471 // since they must follow each StoreP operation. Redundant StoreCMs
2472 // are eliminated just before matching in final_graph_reshape.
2473 {
2474 Node* st = mem;
2475 // If Store 'st' has more than one use, we cannot fold 'st' away.
2476 // For example, 'st' might be the final state at a conditional
2477 // return. Or, 'st' might be used by some node which is live at
2478 // the same time 'st' is live, which might be unschedulable. So,
2479 // require exactly ONE user until such time as we clone 'mem' for
2480 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2481 // true).
2482 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2483 // Looking at a dead closed cycle of memory?
2484 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2485 assert(Opcode() == st->Opcode() ||
2486 st->Opcode() == Op_StoreVector ||
2487 Opcode() == Op_StoreVector ||
2488 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2489 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2490 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2491 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2492 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2493
2494 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2495 st->as_Store()->memory_size() <= this->memory_size()) {
2496 Node* use = st->raw_out(0);
2497 phase->igvn_rehash_node_delayed(use);
2498 if (can_reshape) {
2499 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2500 } else {
2501 // It's OK to do this in the parser, since DU info is always accurate,
2502 // and the parser always refers to nodes via SafePointNode maps.
2503 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2504 }
2505 return this;
2506 }
2507 st = st->in(MemNode::Memory);
2508 }
2509 }
2510
2511
2512 // Capture an unaliased, unconditional, simple store into an initializer.
2513 // Or, if it is independent of the allocation, hoist it above the allocation.
2514 if (ReduceFieldZeroing && /*can_reshape &&*/
2515 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2516 InitializeNode* init = mem->in(0)->as_Initialize();
2517 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2518 if (offset > 0) {
2519 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2520 // If the InitializeNode captured me, it made a raw copy of me,
2521 // and I need to disappear.
2522 if (moved != NULL) {
2523 // %%% hack to ensure that Ideal returns a new node:
2524 mem = MergeMemNode::make(mem);
2525 return mem; // fold me away
2526 }
2527 }
2528 }
2529
2530 return NULL; // No further progress
2531 }
2532
2533 //------------------------------Value-----------------------------------------
2534 const Type* StoreNode::Value(PhaseGVN* phase) const {
2535 // Either input is TOP ==> the result is TOP
2536 const Type *t1 = phase->type( in(MemNode::Memory) );
2537 if( t1 == Type::TOP ) return Type::TOP;
2538 const Type *t2 = phase->type( in(MemNode::Address) );
2539 if( t2 == Type::TOP ) return Type::TOP;
2540 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2541 if( t3 == Type::TOP ) return Type::TOP;
2542 return Type::MEMORY;
2543 }
2544
2545 //------------------------------Identity---------------------------------------
2546 // Remove redundant stores:
2547 // Store(m, p, Load(m, p)) changes to m.
2548 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2549 Node* StoreNode::Identity(PhaseGVN* phase) {
2550 Node* mem = in(MemNode::Memory);
2551 Node* adr = in(MemNode::Address);
2552 Node* val = in(MemNode::ValueIn);
2553
2554 Node* result = this;
2555
2556 // Load then Store? Then the Store is useless
2557 if (val->is_Load() &&
2558 val->in(MemNode::Address)->eqv_uncast(adr) &&
2559 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2560 val->as_Load()->store_Opcode() == Opcode()) {
2561 result = mem;
2562 }
2563
2564 // Two stores in a row of the same value?
2565 if (result == this &&
2566 mem->is_Store() &&
2567 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2568 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2569 mem->Opcode() == Opcode()) {
2570 result = mem;
2571 }
2572
2573 // Store of zero anywhere into a freshly-allocated object?
2574 // Then the store is useless.
2575 // (It must already have been captured by the InitializeNode.)
2576 if (result == this &&
2577 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2578 // a newly allocated object is already all-zeroes everywhere
2579 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2580 result = mem;
2581 }
2582
2583 if (result == this) {
2584 // the store may also apply to zero-bits in an earlier object
2585 Node* prev_mem = find_previous_store(phase);
2586 // Steps (a), (b): Walk past independent stores to find an exact match.
2587 if (prev_mem != NULL) {
2588 Node* prev_val = can_see_stored_value(prev_mem, phase);
2589 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2590 // prev_val and val might differ by a cast; it would be good
2591 // to keep the more informative of the two.
2592 result = mem;
2593 }
2594 }
2595 }
2596 }
2597
2598 if (result != this && phase->is_IterGVN() != NULL) {
2599 MemBarNode* trailing = trailing_membar();
2600 if (trailing != NULL) {
2601 #ifdef ASSERT
2602 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2603 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2604 #endif
2605 PhaseIterGVN* igvn = phase->is_IterGVN();
2606 trailing->remove(igvn);
2607 }
2608 }
2609
2610 return result;
2611 }
2612
2613 //------------------------------match_edge-------------------------------------
2614 // Do we Match on this edge index or not? Match only memory & value
2615 uint StoreNode::match_edge(uint idx) const {
2616 return idx == MemNode::Address || idx == MemNode::ValueIn;
2617 }
2618
2619 //------------------------------cmp--------------------------------------------
2620 // Do not common stores up together. They generally have to be split
2621 // back up anyways, so do not bother.
2622 uint StoreNode::cmp( const Node &n ) const {
2623 return (&n == this); // Always fail except on self
2624 }
2625
2626 //------------------------------Ideal_masked_input-----------------------------
2627 // Check for a useless mask before a partial-word store
2628 // (StoreB ... (AndI valIn conIa) )
2629 // If (conIa & mask == mask) this simplifies to
2630 // (StoreB ... (valIn) )
2631 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2632 Node *val = in(MemNode::ValueIn);
2633 if( val->Opcode() == Op_AndI ) {
2634 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2635 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2636 set_req(MemNode::ValueIn, val->in(1));
2637 return this;
2638 }
2639 }
2640 return NULL;
2641 }
2642
2643
2644 //------------------------------Ideal_sign_extended_input----------------------
2645 // Check for useless sign-extension before a partial-word store
2646 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2647 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2648 // (StoreB ... (valIn) )
2649 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2650 Node *val = in(MemNode::ValueIn);
2651 if( val->Opcode() == Op_RShiftI ) {
2652 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2653 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2654 Node *shl = val->in(1);
2655 if( shl->Opcode() == Op_LShiftI ) {
2656 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2657 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2658 set_req(MemNode::ValueIn, shl->in(1));
2659 return this;
2660 }
2661 }
2662 }
2663 }
2664 return NULL;
2665 }
2666
2667 //------------------------------value_never_loaded-----------------------------------
2668 // Determine whether there are any possible loads of the value stored.
2669 // For simplicity, we actually check if there are any loads from the
2670 // address stored to, not just for loads of the value stored by this node.
2671 //
2672 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2673 Node *adr = in(Address);
2674 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2675 if (adr_oop == NULL)
2676 return false;
2677 if (!adr_oop->is_known_instance_field())
2678 return false; // if not a distinct instance, there may be aliases of the address
2679 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2680 Node *use = adr->fast_out(i);
2681 if (use->is_Load() || use->is_LoadStore()) {
2682 return false;
2683 }
2684 }
2685 return true;
2686 }
2687
2688 MemBarNode* StoreNode::trailing_membar() const {
2689 if (is_release()) {
2690 MemBarNode* trailing_mb = NULL;
2691 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2692 Node* u = fast_out(i);
2693 if (u->is_MemBar()) {
2694 if (u->as_MemBar()->trailing_store()) {
2695 assert(u->Opcode() == Op_MemBarVolatile, "");
2696 assert(trailing_mb == NULL, "only one");
2697 trailing_mb = u->as_MemBar();
2698 #ifdef ASSERT
2699 Node* leading = u->as_MemBar()->leading_membar();
2700 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2701 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2702 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2703 #endif
2704 } else {
2705 assert(u->as_MemBar()->standalone(), "");
2706 }
2707 }
2708 }
2709 return trailing_mb;
2710 }
2711 return NULL;
2712 }
2713
2714
2715 //=============================================================================
2716 //------------------------------Ideal------------------------------------------
2717 // If the store is from an AND mask that leaves the low bits untouched, then
2718 // we can skip the AND operation. If the store is from a sign-extension
2719 // (a left shift, then right shift) we can skip both.
2720 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2721 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2722 if( progress != NULL ) return progress;
2723
2724 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2725 if( progress != NULL ) return progress;
2726
2727 // Finally check the default case
2728 return StoreNode::Ideal(phase, can_reshape);
2729 }
2730
2731 //=============================================================================
2732 //------------------------------Ideal------------------------------------------
2733 // If the store is from an AND mask that leaves the low bits untouched, then
2734 // we can skip the AND operation
2735 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2736 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2737 if( progress != NULL ) return progress;
2738
2739 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2740 if( progress != NULL ) return progress;
2741
2742 // Finally check the default case
2743 return StoreNode::Ideal(phase, can_reshape);
2744 }
2745
2746 //=============================================================================
2747 //------------------------------Identity---------------------------------------
2748 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2749 // No need to card mark when storing a null ptr
2750 Node* my_store = in(MemNode::OopStore);
2751 if (my_store->is_Store()) {
2752 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2753 if( t1 == TypePtr::NULL_PTR ) {
2754 return in(MemNode::Memory);
2755 }
2756 }
2757 return this;
2758 }
2759
2760 //=============================================================================
2761 //------------------------------Ideal---------------------------------------
2762 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2763 Node* progress = StoreNode::Ideal(phase, can_reshape);
2764 if (progress != NULL) return progress;
2765
2766 Node* my_store = in(MemNode::OopStore);
2767 if (my_store->is_MergeMem()) {
2768 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2769 set_req(MemNode::OopStore, mem);
2770 return this;
2771 }
2772
2773 return NULL;
2774 }
2775
2776 //------------------------------Value-----------------------------------------
2777 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2778 // Either input is TOP ==> the result is TOP
2779 const Type *t = phase->type( in(MemNode::Memory) );
2780 if( t == Type::TOP ) return Type::TOP;
2781 t = phase->type( in(MemNode::Address) );
2782 if( t == Type::TOP ) return Type::TOP;
2783 t = phase->type( in(MemNode::ValueIn) );
2784 if( t == Type::TOP ) return Type::TOP;
2785 // If extra input is TOP ==> the result is TOP
2786 t = phase->type( in(MemNode::OopStore) );
2787 if( t == Type::TOP ) return Type::TOP;
2788
2789 return StoreNode::Value( phase );
2790 }
2791
2792
2793 //=============================================================================
2794 //----------------------------------SCMemProjNode------------------------------
2795 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2796 {
2797 return bottom_type();
2798 }
2799
2800 //=============================================================================
2801 //----------------------------------LoadStoreNode------------------------------
2802 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2803 : Node(required),
2804 _type(rt),
2805 _adr_type(at)
2806 {
2807 init_req(MemNode::Control, c );
2808 init_req(MemNode::Memory , mem);
2809 init_req(MemNode::Address, adr);
2810 init_req(MemNode::ValueIn, val);
2811 init_class_id(Class_LoadStore);
2812 }
2813
2814 uint LoadStoreNode::ideal_reg() const {
2815 return _type->ideal_reg();
2816 }
2817
2818 bool LoadStoreNode::result_not_used() const {
2819 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2820 Node *x = fast_out(i);
2821 if (x->Opcode() == Op_SCMemProj) continue;
2822 return false;
2823 }
2824 return true;
2825 }
2826
2827 MemBarNode* LoadStoreNode::trailing_membar() const {
2828 MemBarNode* trailing = NULL;
2829 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2830 Node* u = fast_out(i);
2831 if (u->is_MemBar()) {
2832 if (u->as_MemBar()->trailing_load_store()) {
2833 assert(u->Opcode() == Op_MemBarAcquire, "");
2834 assert(trailing == NULL, "only one");
2835 trailing = u->as_MemBar();
2836 #ifdef ASSERT
2837 Node* leading = trailing->leading_membar();
2838 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2839 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
2840 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
2841 #endif
2842 } else {
2843 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
2844 }
2845 }
2846 }
2847
2848 return trailing;
2849 }
2850
2851 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2852
2853 //=============================================================================
2854 //----------------------------------LoadStoreConditionalNode--------------------
2855 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2856 init_req(ExpectedIn, ex );
2857 }
2858
2859 //=============================================================================
2860 //-------------------------------adr_type--------------------------------------
2861 const TypePtr* ClearArrayNode::adr_type() const {
2862 Node *adr = in(3);
2863 if (adr == NULL) return NULL; // node is dead
2864 return MemNode::calculate_adr_type(adr->bottom_type());
2865 }
2866
2867 //------------------------------match_edge-------------------------------------
2868 // Do we Match on this edge index or not? Do not match memory
2869 uint ClearArrayNode::match_edge(uint idx) const {
2870 return idx > 1;
2871 }
2872
2873 //------------------------------Identity---------------------------------------
2874 // Clearing a zero length array does nothing
2875 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2876 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2877 }
2878
2879 //------------------------------Idealize---------------------------------------
2880 // Clearing a short array is faster with stores
2881 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2882 // Already know this is a large node, do not try to ideal it
2883 if (!IdealizeClearArrayNode || _is_large) return NULL;
2884
2885 const int unit = BytesPerLong;
2886 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2887 if (!t) return NULL;
2888 if (!t->is_con()) return NULL;
2889 intptr_t raw_count = t->get_con();
2890 intptr_t size = raw_count;
2891 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2892 // Clearing nothing uses the Identity call.
2893 // Negative clears are possible on dead ClearArrays
2894 // (see jck test stmt114.stmt11402.val).
2895 if (size <= 0 || size % unit != 0) return NULL;
2896 intptr_t count = size / unit;
2897 // Length too long; communicate this to matchers and assemblers.
2898 // Assemblers are responsible to produce fast hardware clears for it.
2899 if (size > InitArrayShortSize) {
2900 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
2901 }
2902 Node *mem = in(1);
2903 if( phase->type(mem)==Type::TOP ) return NULL;
2904 Node *adr = in(3);
2905 const Type* at = phase->type(adr);
2906 if( at==Type::TOP ) return NULL;
2907 const TypePtr* atp = at->isa_ptr();
2908 // adjust atp to be the correct array element address type
2909 if (atp == NULL) atp = TypePtr::BOTTOM;
2910 else atp = atp->add_offset(Type::OffsetBot);
2911 // Get base for derived pointer purposes
2912 if( adr->Opcode() != Op_AddP ) Unimplemented();
2913 Node *base = adr->in(1);
2914
2915 Node *zero = phase->makecon(TypeLong::ZERO);
2916 Node *off = phase->MakeConX(BytesPerLong);
2917 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2918 count--;
2919 while( count-- ) {
2920 mem = phase->transform(mem);
2921 adr = phase->transform(new AddPNode(base,adr,off));
2922 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2923 }
2924 return mem;
2925 }
2926
2927 //----------------------------step_through----------------------------------
2928 // Return allocation input memory edge if it is different instance
2929 // or itself if it is the one we are looking for.
2930 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2931 Node* n = *np;
2932 assert(n->is_ClearArray(), "sanity");
2933 intptr_t offset;
2934 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2935 // This method is called only before Allocate nodes are expanded
2936 // during macro nodes expansion. Before that ClearArray nodes are
2937 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2938 // Allocate nodes are expanded) which follows allocations.
2939 assert(alloc != NULL, "should have allocation");
2940 if (alloc->_idx == instance_id) {
2941 // Can not bypass initialization of the instance we are looking for.
2942 return false;
2943 }
2944 // Otherwise skip it.
2945 InitializeNode* init = alloc->initialization();
2946 if (init != NULL)
2947 *np = init->in(TypeFunc::Memory);
2948 else
2949 *np = alloc->in(TypeFunc::Memory);
2950 return true;
2951 }
2952
2953 //----------------------------clear_memory-------------------------------------
2954 // Generate code to initialize object storage to zero.
2955 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2956 intptr_t start_offset,
2957 Node* end_offset,
2958 PhaseGVN* phase) {
2959 intptr_t offset = start_offset;
2960
2961 int unit = BytesPerLong;
2962 if ((offset % unit) != 0) {
2963 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
2964 adr = phase->transform(adr);
2965 const TypePtr* atp = TypeRawPtr::BOTTOM;
2966 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2967 mem = phase->transform(mem);
2968 offset += BytesPerInt;
2969 }
2970 assert((offset % unit) == 0, "");
2971
2972 // Initialize the remaining stuff, if any, with a ClearArray.
2973 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2974 }
2975
2976 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2977 Node* start_offset,
2978 Node* end_offset,
2979 PhaseGVN* phase) {
2980 if (start_offset == end_offset) {
2981 // nothing to do
2982 return mem;
2983 }
2984
2985 int unit = BytesPerLong;
2986 Node* zbase = start_offset;
2987 Node* zend = end_offset;
2988
2989 // Scale to the unit required by the CPU:
2990 if (!Matcher::init_array_count_is_in_bytes) {
2991 Node* shift = phase->intcon(exact_log2(unit));
2992 zbase = phase->transform(new URShiftXNode(zbase, shift) );
2993 zend = phase->transform(new URShiftXNode(zend, shift) );
2994 }
2995
2996 // Bulk clear double-words
2997 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
2998 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
2999 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3000 return phase->transform(mem);
3001 }
3002
3003 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3004 intptr_t start_offset,
3005 intptr_t end_offset,
3006 PhaseGVN* phase) {
3007 if (start_offset == end_offset) {
3008 // nothing to do
3009 return mem;
3010 }
3011
3012 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3013 intptr_t done_offset = end_offset;
3014 if ((done_offset % BytesPerLong) != 0) {
3015 done_offset -= BytesPerInt;
3016 }
3017 if (done_offset > start_offset) {
3018 mem = clear_memory(ctl, mem, dest,
3019 start_offset, phase->MakeConX(done_offset), phase);
3020 }
3021 if (done_offset < end_offset) { // emit the final 32-bit store
3022 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3023 adr = phase->transform(adr);
3024 const TypePtr* atp = TypeRawPtr::BOTTOM;
3025 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3026 mem = phase->transform(mem);
3027 done_offset += BytesPerInt;
3028 }
3029 assert(done_offset == end_offset, "");
3030 return mem;
3031 }
3032
3033 //=============================================================================
3034 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3035 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3036 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3037 #ifdef ASSERT
3038 , _pair_idx(0)
3039 #endif
3040 {
3041 init_class_id(Class_MemBar);
3042 Node* top = C->top();
3043 init_req(TypeFunc::I_O,top);
3044 init_req(TypeFunc::FramePtr,top);
3045 init_req(TypeFunc::ReturnAdr,top);
3046 if (precedent != NULL)
3047 init_req(TypeFunc::Parms, precedent);
3048 }
3049
3050 //------------------------------cmp--------------------------------------------
3051 uint MemBarNode::hash() const { return NO_HASH; }
3052 uint MemBarNode::cmp( const Node &n ) const {
3053 return (&n == this); // Always fail except on self
3054 }
3055
3056 //------------------------------make-------------------------------------------
3057 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3058 switch (opcode) {
3059 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3060 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3061 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3062 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3063 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3064 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3065 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3066 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3067 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3068 case Op_Initialize: return new InitializeNode(C, atp, pn);
3069 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3070 default: ShouldNotReachHere(); return NULL;
3071 }
3072 }
3073
3074 void MemBarNode::remove(PhaseIterGVN *igvn) {
3075 if (outcnt() != 2) {
3076 return;
3077 }
3078 if (trailing_store() || trailing_load_store()) {
3079 MemBarNode* leading = leading_membar();
3080 if (leading != NULL) {
3081 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3082 leading->remove(igvn);
3083 }
3084 }
3085 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3086 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3087 }
3088
3089 //------------------------------Ideal------------------------------------------
3090 // Return a node which is more "ideal" than the current node. Strip out
3091 // control copies
3092 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3093 if (remove_dead_region(phase, can_reshape)) return this;
3094 // Don't bother trying to transform a dead node
3095 if (in(0) && in(0)->is_top()) {
3096 return NULL;
3097 }
3098
3099 #if INCLUDE_ZGC
3100 if (UseZGC) {
3101 if (req() == (Precedent+1) && in(MemBarNode::Precedent)->in(0) != NULL && in(MemBarNode::Precedent)->in(0)->is_LoadBarrier()) {
3102 Node* load_node = in(MemBarNode::Precedent)->in(0)->in(LoadBarrierNode::Oop);
3103 set_req(MemBarNode::Precedent, load_node);
3104 return this;
3105 }
3106 }
3107 #endif
3108
3109 bool progress = false;
3110 // Eliminate volatile MemBars for scalar replaced objects.
3111 if (can_reshape && req() == (Precedent+1)) {
3112 bool eliminate = false;
3113 int opc = Opcode();
3114 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3115 // Volatile field loads and stores.
3116 Node* my_mem = in(MemBarNode::Precedent);
3117 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3118 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3119 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3120 // replace this Precedent (decodeN) with the Load instead.
3121 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3122 Node* load_node = my_mem->in(1);
3123 set_req(MemBarNode::Precedent, load_node);
3124 phase->is_IterGVN()->_worklist.push(my_mem);
3125 my_mem = load_node;
3126 } else {
3127 assert(my_mem->unique_out() == this, "sanity");
3128 del_req(Precedent);
3129 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3130 my_mem = NULL;
3131 }
3132 progress = true;
3133 }
3134 if (my_mem != NULL && my_mem->is_Mem()) {
3135 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3136 // Check for scalar replaced object reference.
3137 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3138 t_oop->offset() != Type::OffsetBot &&
3139 t_oop->offset() != Type::OffsetTop) {
3140 eliminate = true;
3141 }
3142 }
3143 } else if (opc == Op_MemBarRelease) {
3144 // Final field stores.
3145 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3146 if ((alloc != NULL) && alloc->is_Allocate() &&
3147 alloc->as_Allocate()->does_not_escape_thread()) {
3148 // The allocated object does not escape.
3149 eliminate = true;
3150 }
3151 }
3152 if (eliminate) {
3153 // Replace MemBar projections by its inputs.
3154 PhaseIterGVN* igvn = phase->is_IterGVN();
3155 remove(igvn);
3156 // Must return either the original node (now dead) or a new node
3157 // (Do not return a top here, since that would break the uniqueness of top.)
3158 return new ConINode(TypeInt::ZERO);
3159 }
3160 }
3161 return progress ? this : NULL;
3162 }
3163
3164 //------------------------------Value------------------------------------------
3165 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3166 if( !in(0) ) return Type::TOP;
3167 if( phase->type(in(0)) == Type::TOP )
3168 return Type::TOP;
3169 return TypeTuple::MEMBAR;
3170 }
3171
3172 //------------------------------match------------------------------------------
3173 // Construct projections for memory.
3174 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3175 switch (proj->_con) {
3176 case TypeFunc::Control:
3177 case TypeFunc::Memory:
3178 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3179 }
3180 ShouldNotReachHere();
3181 return NULL;
3182 }
3183
3184 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3185 trailing->_kind = TrailingStore;
3186 leading->_kind = LeadingStore;
3187 #ifdef ASSERT
3188 trailing->_pair_idx = leading->_idx;
3189 leading->_pair_idx = leading->_idx;
3190 #endif
3191 }
3192
3193 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3194 trailing->_kind = TrailingLoadStore;
3195 leading->_kind = LeadingLoadStore;
3196 #ifdef ASSERT
3197 trailing->_pair_idx = leading->_idx;
3198 leading->_pair_idx = leading->_idx;
3199 #endif
3200 }
3201
3202 MemBarNode* MemBarNode::trailing_membar() const {
3203 ResourceMark rm;
3204 Node* trailing = (Node*)this;
3205 VectorSet seen(Thread::current()->resource_area());
3206 Node_Stack multis(0);
3207 do {
3208 Node* c = trailing;
3209 uint i = 0;
3210 do {
3211 trailing = NULL;
3212 for (; i < c->outcnt(); i++) {
3213 Node* next = c->raw_out(i);
3214 if (next != c && next->is_CFG()) {
3215 if (c->is_MultiBranch()) {
3216 if (multis.node() == c) {
3217 multis.set_index(i+1);
3218 } else {
3219 multis.push(c, i+1);
3220 }
3221 }
3222 trailing = next;
3223 break;
3224 }
3225 }
3226 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3227 break;
3228 }
3229 while (multis.size() > 0) {
3230 c = multis.node();
3231 i = multis.index();
3232 if (i < c->req()) {
3233 break;
3234 }
3235 multis.pop();
3236 }
3237 } while (multis.size() > 0);
3238 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3239
3240 MemBarNode* mb = trailing->as_MemBar();
3241 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3242 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3243 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3244 return mb;
3245 }
3246
3247 MemBarNode* MemBarNode::leading_membar() const {
3248 ResourceMark rm;
3249 VectorSet seen(Thread::current()->resource_area());
3250 Node_Stack regions(0);
3251 Node* leading = in(0);
3252 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3253 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3254 leading = NULL;
3255 while (regions.size() > 0 && leading == NULL) {
3256 Node* r = regions.node();
3257 uint i = regions.index();
3258 if (i < r->req()) {
3259 leading = r->in(i);
3260 regions.set_index(i+1);
3261 } else {
3262 regions.pop();
3263 }
3264 }
3265 if (leading == NULL) {
3266 assert(regions.size() == 0, "all paths should have been tried");
3267 return NULL;
3268 }
3269 }
3270 if (leading->is_Region()) {
3271 regions.push(leading, 2);
3272 leading = leading->in(1);
3273 } else {
3274 leading = leading->in(0);
3275 }
3276 }
3277 #ifdef ASSERT
3278 Unique_Node_List wq;
3279 wq.push((Node*)this);
3280 uint found = 0;
3281 for (uint i = 0; i < wq.size(); i++) {
3282 Node* n = wq.at(i);
3283 if (n->is_Region()) {
3284 for (uint j = 1; j < n->req(); j++) {
3285 Node* in = n->in(j);
3286 if (in != NULL && !in->is_top()) {
3287 wq.push(in);
3288 }
3289 }
3290 } else {
3291 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3292 assert(n == leading, "consistency check failed");
3293 found++;
3294 } else {
3295 Node* in = n->in(0);
3296 if (in != NULL && !in->is_top()) {
3297 wq.push(in);
3298 }
3299 }
3300 }
3301 }
3302 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3303 #endif
3304 if (leading == NULL) {
3305 return NULL;
3306 }
3307 MemBarNode* mb = leading->as_MemBar();
3308 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3309 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3310 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3311 return mb;
3312 }
3313
3314 //===========================InitializeNode====================================
3315 // SUMMARY:
3316 // This node acts as a memory barrier on raw memory, after some raw stores.
3317 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3318 // The Initialize can 'capture' suitably constrained stores as raw inits.
3319 // It can coalesce related raw stores into larger units (called 'tiles').
3320 // It can avoid zeroing new storage for memory units which have raw inits.
3321 // At macro-expansion, it is marked 'complete', and does not optimize further.
3322 //
3323 // EXAMPLE:
3324 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3325 // ctl = incoming control; mem* = incoming memory
3326 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3327 // First allocate uninitialized memory and fill in the header:
3328 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3329 // ctl := alloc.Control; mem* := alloc.Memory*
3330 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3331 // Then initialize to zero the non-header parts of the raw memory block:
3332 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3333 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3334 // After the initialize node executes, the object is ready for service:
3335 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3336 // Suppose its body is immediately initialized as {1,2}:
3337 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3338 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3339 // mem.SLICE(#short[*]) := store2
3340 //
3341 // DETAILS:
3342 // An InitializeNode collects and isolates object initialization after
3343 // an AllocateNode and before the next possible safepoint. As a
3344 // memory barrier (MemBarNode), it keeps critical stores from drifting
3345 // down past any safepoint or any publication of the allocation.
3346 // Before this barrier, a newly-allocated object may have uninitialized bits.
3347 // After this barrier, it may be treated as a real oop, and GC is allowed.
3348 //
3349 // The semantics of the InitializeNode include an implicit zeroing of
3350 // the new object from object header to the end of the object.
3351 // (The object header and end are determined by the AllocateNode.)
3352 //
3353 // Certain stores may be added as direct inputs to the InitializeNode.
3354 // These stores must update raw memory, and they must be to addresses
3355 // derived from the raw address produced by AllocateNode, and with
3356 // a constant offset. They must be ordered by increasing offset.
3357 // The first one is at in(RawStores), the last at in(req()-1).
3358 // Unlike most memory operations, they are not linked in a chain,
3359 // but are displayed in parallel as users of the rawmem output of
3360 // the allocation.
3361 //
3362 // (See comments in InitializeNode::capture_store, which continue
3363 // the example given above.)
3364 //
3365 // When the associated Allocate is macro-expanded, the InitializeNode
3366 // may be rewritten to optimize collected stores. A ClearArrayNode
3367 // may also be created at that point to represent any required zeroing.
3368 // The InitializeNode is then marked 'complete', prohibiting further
3369 // capturing of nearby memory operations.
3370 //
3371 // During macro-expansion, all captured initializations which store
3372 // constant values of 32 bits or smaller are coalesced (if advantageous)
3373 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3374 // initialized in fewer memory operations. Memory words which are
3375 // covered by neither tiles nor non-constant stores are pre-zeroed
3376 // by explicit stores of zero. (The code shape happens to do all
3377 // zeroing first, then all other stores, with both sequences occurring
3378 // in order of ascending offsets.)
3379 //
3380 // Alternatively, code may be inserted between an AllocateNode and its
3381 // InitializeNode, to perform arbitrary initialization of the new object.
3382 // E.g., the object copying intrinsics insert complex data transfers here.
3383 // The initialization must then be marked as 'complete' disable the
3384 // built-in zeroing semantics and the collection of initializing stores.
3385 //
3386 // While an InitializeNode is incomplete, reads from the memory state
3387 // produced by it are optimizable if they match the control edge and
3388 // new oop address associated with the allocation/initialization.
3389 // They return a stored value (if the offset matches) or else zero.
3390 // A write to the memory state, if it matches control and address,
3391 // and if it is to a constant offset, may be 'captured' by the
3392 // InitializeNode. It is cloned as a raw memory operation and rewired
3393 // inside the initialization, to the raw oop produced by the allocation.
3394 // Operations on addresses which are provably distinct (e.g., to
3395 // other AllocateNodes) are allowed to bypass the initialization.
3396 //
3397 // The effect of all this is to consolidate object initialization
3398 // (both arrays and non-arrays, both piecewise and bulk) into a
3399 // single location, where it can be optimized as a unit.
3400 //
3401 // Only stores with an offset less than TrackedInitializationLimit words
3402 // will be considered for capture by an InitializeNode. This puts a
3403 // reasonable limit on the complexity of optimized initializations.
3404
3405 //---------------------------InitializeNode------------------------------------
3406 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3407 : MemBarNode(C, adr_type, rawoop),
3408 _is_complete(Incomplete), _does_not_escape(false)
3409 {
3410 init_class_id(Class_Initialize);
3411
3412 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3413 assert(in(RawAddress) == rawoop, "proper init");
3414 // Note: allocation() can be NULL, for secondary initialization barriers
3415 }
3416
3417 // Since this node is not matched, it will be processed by the
3418 // register allocator. Declare that there are no constraints
3419 // on the allocation of the RawAddress edge.
3420 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3421 // This edge should be set to top, by the set_complete. But be conservative.
3422 if (idx == InitializeNode::RawAddress)
3423 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3424 return RegMask::Empty;
3425 }
3426
3427 Node* InitializeNode::memory(uint alias_idx) {
3428 Node* mem = in(Memory);
3429 if (mem->is_MergeMem()) {
3430 return mem->as_MergeMem()->memory_at(alias_idx);
3431 } else {
3432 // incoming raw memory is not split
3433 return mem;
3434 }
3435 }
3436
3437 bool InitializeNode::is_non_zero() {
3438 if (is_complete()) return false;
3439 remove_extra_zeroes();
3440 return (req() > RawStores);
3441 }
3442
3443 void InitializeNode::set_complete(PhaseGVN* phase) {
3444 assert(!is_complete(), "caller responsibility");
3445 _is_complete = Complete;
3446
3447 // After this node is complete, it contains a bunch of
3448 // raw-memory initializations. There is no need for
3449 // it to have anything to do with non-raw memory effects.
3450 // Therefore, tell all non-raw users to re-optimize themselves,
3451 // after skipping the memory effects of this initialization.
3452 PhaseIterGVN* igvn = phase->is_IterGVN();
3453 if (igvn) igvn->add_users_to_worklist(this);
3454 }
3455
3456 // convenience function
3457 // return false if the init contains any stores already
3458 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3459 InitializeNode* init = initialization();
3460 if (init == NULL || init->is_complete()) return false;
3461 init->remove_extra_zeroes();
3462 // for now, if this allocation has already collected any inits, bail:
3463 if (init->is_non_zero()) return false;
3464 init->set_complete(phase);
3465 return true;
3466 }
3467
3468 void InitializeNode::remove_extra_zeroes() {
3469 if (req() == RawStores) return;
3470 Node* zmem = zero_memory();
3471 uint fill = RawStores;
3472 for (uint i = fill; i < req(); i++) {
3473 Node* n = in(i);
3474 if (n->is_top() || n == zmem) continue; // skip
3475 if (fill < i) set_req(fill, n); // compact
3476 ++fill;
3477 }
3478 // delete any empty spaces created:
3479 while (fill < req()) {
3480 del_req(fill);
3481 }
3482 }
3483
3484 // Helper for remembering which stores go with which offsets.
3485 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3486 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3487 intptr_t offset = -1;
3488 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3489 phase, offset);
3490 if (base == NULL) return -1; // something is dead,
3491 if (offset < 0) return -1; // dead, dead
3492 return offset;
3493 }
3494
3495 // Helper for proving that an initialization expression is
3496 // "simple enough" to be folded into an object initialization.
3497 // Attempts to prove that a store's initial value 'n' can be captured
3498 // within the initialization without creating a vicious cycle, such as:
3499 // { Foo p = new Foo(); p.next = p; }
3500 // True for constants and parameters and small combinations thereof.
3501 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3502 if (n == NULL) return true; // (can this really happen?)
3503 if (n->is_Proj()) n = n->in(0);
3504 if (n == this) return false; // found a cycle
3505 if (n->is_Con()) return true;
3506 if (n->is_Start()) return true; // params, etc., are OK
3507 if (n->is_Root()) return true; // even better
3508
3509 Node* ctl = n->in(0);
3510 if (ctl != NULL && !ctl->is_top()) {
3511 if (ctl->is_Proj()) ctl = ctl->in(0);
3512 if (ctl == this) return false;
3513
3514 // If we already know that the enclosing memory op is pinned right after
3515 // the init, then any control flow that the store has picked up
3516 // must have preceded the init, or else be equal to the init.
3517 // Even after loop optimizations (which might change control edges)
3518 // a store is never pinned *before* the availability of its inputs.
3519 if (!MemNode::all_controls_dominate(n, this))
3520 return false; // failed to prove a good control
3521 }
3522
3523 // Check data edges for possible dependencies on 'this'.
3524 if ((count += 1) > 20) return false; // complexity limit
3525 for (uint i = 1; i < n->req(); i++) {
3526 Node* m = n->in(i);
3527 if (m == NULL || m == n || m->is_top()) continue;
3528 uint first_i = n->find_edge(m);
3529 if (i != first_i) continue; // process duplicate edge just once
3530 if (!detect_init_independence(m, count)) {
3531 return false;
3532 }
3533 }
3534
3535 return true;
3536 }
3537
3538 // Here are all the checks a Store must pass before it can be moved into
3539 // an initialization. Returns zero if a check fails.
3540 // On success, returns the (constant) offset to which the store applies,
3541 // within the initialized memory.
3542 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3543 const int FAIL = 0;
3544 if (st->is_unaligned_access() || ((get_store_offset(st, phase) % BytesPerInt) != 0)) {
3545 return FAIL;
3546 }
3547 if (st->req() != MemNode::ValueIn + 1)
3548 return FAIL; // an inscrutable StoreNode (card mark?)
3549 Node* ctl = st->in(MemNode::Control);
3550 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3551 return FAIL; // must be unconditional after the initialization
3552 Node* mem = st->in(MemNode::Memory);
3553 if (!(mem->is_Proj() && mem->in(0) == this))
3554 return FAIL; // must not be preceded by other stores
3555 Node* adr = st->in(MemNode::Address);
3556 intptr_t offset;
3557 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3558 if (alloc == NULL)
3559 return FAIL; // inscrutable address
3560 if (alloc != allocation())
3561 return FAIL; // wrong allocation! (store needs to float up)
3562 Node* val = st->in(MemNode::ValueIn);
3563 int complexity_count = 0;
3564 if (!detect_init_independence(val, complexity_count))
3565 return FAIL; // stored value must be 'simple enough'
3566
3567 // The Store can be captured only if nothing after the allocation
3568 // and before the Store is using the memory location that the store
3569 // overwrites.
3570 bool failed = false;
3571 // If is_complete_with_arraycopy() is true the shape of the graph is
3572 // well defined and is safe so no need for extra checks.
3573 if (!is_complete_with_arraycopy()) {
3574 // We are going to look at each use of the memory state following
3575 // the allocation to make sure nothing reads the memory that the
3576 // Store writes.
3577 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3578 int alias_idx = phase->C->get_alias_index(t_adr);
3579 ResourceMark rm;
3580 Unique_Node_List mems;
3581 mems.push(mem);
3582 Node* unique_merge = NULL;
3583 for (uint next = 0; next < mems.size(); ++next) {
3584 Node *m = mems.at(next);
3585 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3586 Node *n = m->fast_out(j);
3587 if (n->outcnt() == 0) {
3588 continue;
3589 }
3590 if (n == st) {
3591 continue;
3592 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3593 // If the control of this use is different from the control
3594 // of the Store which is right after the InitializeNode then
3595 // this node cannot be between the InitializeNode and the
3596 // Store.
3597 continue;
3598 } else if (n->is_MergeMem()) {
3599 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3600 // We can hit a MergeMemNode (that will likely go away
3601 // later) that is a direct use of the memory state
3602 // following the InitializeNode on the same slice as the
3603 // store node that we'd like to capture. We need to check
3604 // the uses of the MergeMemNode.
3605 mems.push(n);
3606 }
3607 } else if (n->is_Mem()) {
3608 Node* other_adr = n->in(MemNode::Address);
3609 if (other_adr == adr) {
3610 failed = true;
3611 break;
3612 } else {
3613 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3614 if (other_t_adr != NULL) {
3615 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3616 if (other_alias_idx == alias_idx) {
3617 // A load from the same memory slice as the store right
3618 // after the InitializeNode. We check the control of the
3619 // object/array that is loaded from. If it's the same as
3620 // the store control then we cannot capture the store.
3621 assert(!n->is_Store(), "2 stores to same slice on same control?");
3622 Node* base = other_adr;
3623 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3624 base = base->in(AddPNode::Base);
3625 if (base != NULL) {
3626 base = base->uncast();
3627 if (base->is_Proj() && base->in(0) == alloc) {
3628 failed = true;
3629 break;
3630 }
3631 }
3632 }
3633 }
3634 }
3635 } else {
3636 failed = true;
3637 break;
3638 }
3639 }
3640 }
3641 }
3642 if (failed) {
3643 if (!can_reshape) {
3644 // We decided we couldn't capture the store during parsing. We
3645 // should try again during the next IGVN once the graph is
3646 // cleaner.
3647 phase->C->record_for_igvn(st);
3648 }
3649 return FAIL;
3650 }
3651
3652 return offset; // success
3653 }
3654
3655 // Find the captured store in(i) which corresponds to the range
3656 // [start..start+size) in the initialized object.
3657 // If there is one, return its index i. If there isn't, return the
3658 // negative of the index where it should be inserted.
3659 // Return 0 if the queried range overlaps an initialization boundary
3660 // or if dead code is encountered.
3661 // If size_in_bytes is zero, do not bother with overlap checks.
3662 int InitializeNode::captured_store_insertion_point(intptr_t start,
3663 int size_in_bytes,
3664 PhaseTransform* phase) {
3665 const int FAIL = 0, MAX_STORE = BytesPerLong;
3666
3667 if (is_complete())
3668 return FAIL; // arraycopy got here first; punt
3669
3670 assert(allocation() != NULL, "must be present");
3671
3672 // no negatives, no header fields:
3673 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3674
3675 // after a certain size, we bail out on tracking all the stores:
3676 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3677 if (start >= ti_limit) return FAIL;
3678
3679 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3680 if (i >= limit) return -(int)i; // not found; here is where to put it
3681
3682 Node* st = in(i);
3683 intptr_t st_off = get_store_offset(st, phase);
3684 if (st_off < 0) {
3685 if (st != zero_memory()) {
3686 return FAIL; // bail out if there is dead garbage
3687 }
3688 } else if (st_off > start) {
3689 // ...we are done, since stores are ordered
3690 if (st_off < start + size_in_bytes) {
3691 return FAIL; // the next store overlaps
3692 }
3693 return -(int)i; // not found; here is where to put it
3694 } else if (st_off < start) {
3695 if (size_in_bytes != 0 &&
3696 start < st_off + MAX_STORE &&
3697 start < st_off + st->as_Store()->memory_size()) {
3698 return FAIL; // the previous store overlaps
3699 }
3700 } else {
3701 if (size_in_bytes != 0 &&
3702 st->as_Store()->memory_size() != size_in_bytes) {
3703 return FAIL; // mismatched store size
3704 }
3705 return i;
3706 }
3707
3708 ++i;
3709 }
3710 }
3711
3712 // Look for a captured store which initializes at the offset 'start'
3713 // with the given size. If there is no such store, and no other
3714 // initialization interferes, then return zero_memory (the memory
3715 // projection of the AllocateNode).
3716 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3717 PhaseTransform* phase) {
3718 assert(stores_are_sane(phase), "");
3719 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3720 if (i == 0) {
3721 return NULL; // something is dead
3722 } else if (i < 0) {
3723 return zero_memory(); // just primordial zero bits here
3724 } else {
3725 Node* st = in(i); // here is the store at this position
3726 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3727 return st;
3728 }
3729 }
3730
3731 // Create, as a raw pointer, an address within my new object at 'offset'.
3732 Node* InitializeNode::make_raw_address(intptr_t offset,
3733 PhaseTransform* phase) {
3734 Node* addr = in(RawAddress);
3735 if (offset != 0) {
3736 Compile* C = phase->C;
3737 addr = phase->transform( new AddPNode(C->top(), addr,
3738 phase->MakeConX(offset)) );
3739 }
3740 return addr;
3741 }
3742
3743 // Clone the given store, converting it into a raw store
3744 // initializing a field or element of my new object.
3745 // Caller is responsible for retiring the original store,
3746 // with subsume_node or the like.
3747 //
3748 // From the example above InitializeNode::InitializeNode,
3749 // here are the old stores to be captured:
3750 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3751 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3752 //
3753 // Here is the changed code; note the extra edges on init:
3754 // alloc = (Allocate ...)
3755 // rawoop = alloc.RawAddress
3756 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3757 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3758 // init = (Initialize alloc.Control alloc.Memory rawoop
3759 // rawstore1 rawstore2)
3760 //
3761 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3762 PhaseTransform* phase, bool can_reshape) {
3763 assert(stores_are_sane(phase), "");
3764
3765 if (start < 0) return NULL;
3766 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3767
3768 Compile* C = phase->C;
3769 int size_in_bytes = st->memory_size();
3770 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3771 if (i == 0) return NULL; // bail out
3772 Node* prev_mem = NULL; // raw memory for the captured store
3773 if (i > 0) {
3774 prev_mem = in(i); // there is a pre-existing store under this one
3775 set_req(i, C->top()); // temporarily disconnect it
3776 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3777 } else {
3778 i = -i; // no pre-existing store
3779 prev_mem = zero_memory(); // a slice of the newly allocated object
3780 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3781 set_req(--i, C->top()); // reuse this edge; it has been folded away
3782 else
3783 ins_req(i, C->top()); // build a new edge
3784 }
3785 Node* new_st = st->clone();
3786 new_st->set_req(MemNode::Control, in(Control));
3787 new_st->set_req(MemNode::Memory, prev_mem);
3788 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3789 new_st = phase->transform(new_st);
3790
3791 // At this point, new_st might have swallowed a pre-existing store
3792 // at the same offset, or perhaps new_st might have disappeared,
3793 // if it redundantly stored the same value (or zero to fresh memory).
3794
3795 // In any case, wire it in:
3796 phase->igvn_rehash_node_delayed(this);
3797 set_req(i, new_st);
3798
3799 // The caller may now kill the old guy.
3800 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3801 assert(check_st == new_st || check_st == NULL, "must be findable");
3802 assert(!is_complete(), "");
3803 return new_st;
3804 }
3805
3806 static bool store_constant(jlong* tiles, int num_tiles,
3807 intptr_t st_off, int st_size,
3808 jlong con) {
3809 if ((st_off & (st_size-1)) != 0)
3810 return false; // strange store offset (assume size==2**N)
3811 address addr = (address)tiles + st_off;
3812 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3813 switch (st_size) {
3814 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3815 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3816 case sizeof(jint): *(jint*) addr = (jint) con; break;
3817 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3818 default: return false; // strange store size (detect size!=2**N here)
3819 }
3820 return true; // return success to caller
3821 }
3822
3823 // Coalesce subword constants into int constants and possibly
3824 // into long constants. The goal, if the CPU permits,
3825 // is to initialize the object with a small number of 64-bit tiles.
3826 // Also, convert floating-point constants to bit patterns.
3827 // Non-constants are not relevant to this pass.
3828 //
3829 // In terms of the running example on InitializeNode::InitializeNode
3830 // and InitializeNode::capture_store, here is the transformation
3831 // of rawstore1 and rawstore2 into rawstore12:
3832 // alloc = (Allocate ...)
3833 // rawoop = alloc.RawAddress
3834 // tile12 = 0x00010002
3835 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3836 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3837 //
3838 void
3839 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3840 Node* size_in_bytes,
3841 PhaseGVN* phase) {
3842 Compile* C = phase->C;
3843
3844 assert(stores_are_sane(phase), "");
3845 // Note: After this pass, they are not completely sane,
3846 // since there may be some overlaps.
3847
3848 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3849
3850 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3851 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3852 size_limit = MIN2(size_limit, ti_limit);
3853 size_limit = align_up(size_limit, BytesPerLong);
3854 int num_tiles = size_limit / BytesPerLong;
3855
3856 // allocate space for the tile map:
3857 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3858 jlong tiles_buf[small_len];
3859 Node* nodes_buf[small_len];
3860 jlong inits_buf[small_len];
3861 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3862 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3863 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3864 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3865 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3866 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3867 // tiles: exact bitwise model of all primitive constants
3868 // nodes: last constant-storing node subsumed into the tiles model
3869 // inits: which bytes (in each tile) are touched by any initializations
3870
3871 //// Pass A: Fill in the tile model with any relevant stores.
3872
3873 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3874 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3875 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3876 Node* zmem = zero_memory(); // initially zero memory state
3877 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3878 Node* st = in(i);
3879 intptr_t st_off = get_store_offset(st, phase);
3880
3881 // Figure out the store's offset and constant value:
3882 if (st_off < header_size) continue; //skip (ignore header)
3883 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3884 int st_size = st->as_Store()->memory_size();
3885 if (st_off + st_size > size_limit) break;
3886
3887 // Record which bytes are touched, whether by constant or not.
3888 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3889 continue; // skip (strange store size)
3890
3891 const Type* val = phase->type(st->in(MemNode::ValueIn));
3892 if (!val->singleton()) continue; //skip (non-con store)
3893 BasicType type = val->basic_type();
3894
3895 jlong con = 0;
3896 switch (type) {
3897 case T_INT: con = val->is_int()->get_con(); break;
3898 case T_LONG: con = val->is_long()->get_con(); break;
3899 case T_FLOAT: con = jint_cast(val->getf()); break;
3900 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3901 default: continue; //skip (odd store type)
3902 }
3903
3904 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3905 st->Opcode() == Op_StoreL) {
3906 continue; // This StoreL is already optimal.
3907 }
3908
3909 // Store down the constant.
3910 store_constant(tiles, num_tiles, st_off, st_size, con);
3911
3912 intptr_t j = st_off >> LogBytesPerLong;
3913
3914 if (type == T_INT && st_size == BytesPerInt
3915 && (st_off & BytesPerInt) == BytesPerInt) {
3916 jlong lcon = tiles[j];
3917 if (!Matcher::isSimpleConstant64(lcon) &&
3918 st->Opcode() == Op_StoreI) {
3919 // This StoreI is already optimal by itself.
3920 jint* intcon = (jint*) &tiles[j];
3921 intcon[1] = 0; // undo the store_constant()
3922
3923 // If the previous store is also optimal by itself, back up and
3924 // undo the action of the previous loop iteration... if we can.
3925 // But if we can't, just let the previous half take care of itself.
3926 st = nodes[j];
3927 st_off -= BytesPerInt;
3928 con = intcon[0];
3929 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3930 assert(st_off >= header_size, "still ignoring header");
3931 assert(get_store_offset(st, phase) == st_off, "must be");
3932 assert(in(i-1) == zmem, "must be");
3933 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3934 assert(con == tcon->is_int()->get_con(), "must be");
3935 // Undo the effects of the previous loop trip, which swallowed st:
3936 intcon[0] = 0; // undo store_constant()
3937 set_req(i-1, st); // undo set_req(i, zmem)
3938 nodes[j] = NULL; // undo nodes[j] = st
3939 --old_subword; // undo ++old_subword
3940 }
3941 continue; // This StoreI is already optimal.
3942 }
3943 }
3944
3945 // This store is not needed.
3946 set_req(i, zmem);
3947 nodes[j] = st; // record for the moment
3948 if (st_size < BytesPerLong) // something has changed
3949 ++old_subword; // includes int/float, but who's counting...
3950 else ++old_long;
3951 }
3952
3953 if ((old_subword + old_long) == 0)
3954 return; // nothing more to do
3955
3956 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3957 // Be sure to insert them before overlapping non-constant stores.
3958 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3959 for (int j = 0; j < num_tiles; j++) {
3960 jlong con = tiles[j];
3961 jlong init = inits[j];
3962 if (con == 0) continue;
3963 jint con0, con1; // split the constant, address-wise
3964 jint init0, init1; // split the init map, address-wise
3965 { union { jlong con; jint intcon[2]; } u;
3966 u.con = con;
3967 con0 = u.intcon[0];
3968 con1 = u.intcon[1];
3969 u.con = init;
3970 init0 = u.intcon[0];
3971 init1 = u.intcon[1];
3972 }
3973
3974 Node* old = nodes[j];
3975 assert(old != NULL, "need the prior store");
3976 intptr_t offset = (j * BytesPerLong);
3977
3978 bool split = !Matcher::isSimpleConstant64(con);
3979
3980 if (offset < header_size) {
3981 assert(offset + BytesPerInt >= header_size, "second int counts");
3982 assert(*(jint*)&tiles[j] == 0, "junk in header");
3983 split = true; // only the second word counts
3984 // Example: int a[] = { 42 ... }
3985 } else if (con0 == 0 && init0 == -1) {
3986 split = true; // first word is covered by full inits
3987 // Example: int a[] = { ... foo(), 42 ... }
3988 } else if (con1 == 0 && init1 == -1) {
3989 split = true; // second word is covered by full inits
3990 // Example: int a[] = { ... 42, foo() ... }
3991 }
3992
3993 // Here's a case where init0 is neither 0 nor -1:
3994 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3995 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3996 // In this case the tile is not split; it is (jlong)42.
3997 // The big tile is stored down, and then the foo() value is inserted.
3998 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3999
4000 Node* ctl = old->in(MemNode::Control);
4001 Node* adr = make_raw_address(offset, phase);
4002 const TypePtr* atp = TypeRawPtr::BOTTOM;
4003
4004 // One or two coalesced stores to plop down.
4005 Node* st[2];
4006 intptr_t off[2];
4007 int nst = 0;
4008 if (!split) {
4009 ++new_long;
4010 off[nst] = offset;
4011 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4012 phase->longcon(con), T_LONG, MemNode::unordered);
4013 } else {
4014 // Omit either if it is a zero.
4015 if (con0 != 0) {
4016 ++new_int;
4017 off[nst] = offset;
4018 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4019 phase->intcon(con0), T_INT, MemNode::unordered);
4020 }
4021 if (con1 != 0) {
4022 ++new_int;
4023 offset += BytesPerInt;
4024 adr = make_raw_address(offset, phase);
4025 off[nst] = offset;
4026 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4027 phase->intcon(con1), T_INT, MemNode::unordered);
4028 }
4029 }
4030
4031 // Insert second store first, then the first before the second.
4032 // Insert each one just before any overlapping non-constant stores.
4033 while (nst > 0) {
4034 Node* st1 = st[--nst];
4035 C->copy_node_notes_to(st1, old);
4036 st1 = phase->transform(st1);
4037 offset = off[nst];
4038 assert(offset >= header_size, "do not smash header");
4039 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4040 guarantee(ins_idx != 0, "must re-insert constant store");
4041 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4042 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4043 set_req(--ins_idx, st1);
4044 else
4045 ins_req(ins_idx, st1);
4046 }
4047 }
4048
4049 if (PrintCompilation && WizardMode)
4050 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4051 old_subword, old_long, new_int, new_long);
4052 if (C->log() != NULL)
4053 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4054 old_subword, old_long, new_int, new_long);
4055
4056 // Clean up any remaining occurrences of zmem:
4057 remove_extra_zeroes();
4058 }
4059
4060 // Explore forward from in(start) to find the first fully initialized
4061 // word, and return its offset. Skip groups of subword stores which
4062 // together initialize full words. If in(start) is itself part of a
4063 // fully initialized word, return the offset of in(start). If there
4064 // are no following full-word stores, or if something is fishy, return
4065 // a negative value.
4066 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4067 int int_map = 0;
4068 intptr_t int_map_off = 0;
4069 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4070
4071 for (uint i = start, limit = req(); i < limit; i++) {
4072 Node* st = in(i);
4073
4074 intptr_t st_off = get_store_offset(st, phase);
4075 if (st_off < 0) break; // return conservative answer
4076
4077 int st_size = st->as_Store()->memory_size();
4078 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4079 return st_off; // we found a complete word init
4080 }
4081
4082 // update the map:
4083
4084 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4085 if (this_int_off != int_map_off) {
4086 // reset the map:
4087 int_map = 0;
4088 int_map_off = this_int_off;
4089 }
4090
4091 int subword_off = st_off - this_int_off;
4092 int_map |= right_n_bits(st_size) << subword_off;
4093 if ((int_map & FULL_MAP) == FULL_MAP) {
4094 return this_int_off; // we found a complete word init
4095 }
4096
4097 // Did this store hit or cross the word boundary?
4098 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4099 if (next_int_off == this_int_off + BytesPerInt) {
4100 // We passed the current int, without fully initializing it.
4101 int_map_off = next_int_off;
4102 int_map >>= BytesPerInt;
4103 } else if (next_int_off > this_int_off + BytesPerInt) {
4104 // We passed the current and next int.
4105 return this_int_off + BytesPerInt;
4106 }
4107 }
4108
4109 return -1;
4110 }
4111
4112
4113 // Called when the associated AllocateNode is expanded into CFG.
4114 // At this point, we may perform additional optimizations.
4115 // Linearize the stores by ascending offset, to make memory
4116 // activity as coherent as possible.
4117 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4118 intptr_t header_size,
4119 Node* size_in_bytes,
4120 PhaseGVN* phase) {
4121 assert(!is_complete(), "not already complete");
4122 assert(stores_are_sane(phase), "");
4123 assert(allocation() != NULL, "must be present");
4124
4125 remove_extra_zeroes();
4126
4127 if (ReduceFieldZeroing || ReduceBulkZeroing)
4128 // reduce instruction count for common initialization patterns
4129 coalesce_subword_stores(header_size, size_in_bytes, phase);
4130
4131 Node* zmem = zero_memory(); // initially zero memory state
4132 Node* inits = zmem; // accumulating a linearized chain of inits
4133 #ifdef ASSERT
4134 intptr_t first_offset = allocation()->minimum_header_size();
4135 intptr_t last_init_off = first_offset; // previous init offset
4136 intptr_t last_init_end = first_offset; // previous init offset+size
4137 intptr_t last_tile_end = first_offset; // previous tile offset+size
4138 #endif
4139 intptr_t zeroes_done = header_size;
4140
4141 bool do_zeroing = true; // we might give up if inits are very sparse
4142 int big_init_gaps = 0; // how many large gaps have we seen?
4143
4144 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4145 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4146
4147 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4148 Node* st = in(i);
4149 intptr_t st_off = get_store_offset(st, phase);
4150 if (st_off < 0)
4151 break; // unknown junk in the inits
4152 if (st->in(MemNode::Memory) != zmem)
4153 break; // complicated store chains somehow in list
4154
4155 int st_size = st->as_Store()->memory_size();
4156 intptr_t next_init_off = st_off + st_size;
4157
4158 if (do_zeroing && zeroes_done < next_init_off) {
4159 // See if this store needs a zero before it or under it.
4160 intptr_t zeroes_needed = st_off;
4161
4162 if (st_size < BytesPerInt) {
4163 // Look for subword stores which only partially initialize words.
4164 // If we find some, we must lay down some word-level zeroes first,
4165 // underneath the subword stores.
4166 //
4167 // Examples:
4168 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4169 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4170 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4171 //
4172 // Note: coalesce_subword_stores may have already done this,
4173 // if it was prompted by constant non-zero subword initializers.
4174 // But this case can still arise with non-constant stores.
4175
4176 intptr_t next_full_store = find_next_fullword_store(i, phase);
4177
4178 // In the examples above:
4179 // in(i) p q r s x y z
4180 // st_off 12 13 14 15 12 13 14
4181 // st_size 1 1 1 1 1 1 1
4182 // next_full_s. 12 16 16 16 16 16 16
4183 // z's_done 12 16 16 16 12 16 12
4184 // z's_needed 12 16 16 16 16 16 16
4185 // zsize 0 0 0 0 4 0 4
4186 if (next_full_store < 0) {
4187 // Conservative tack: Zero to end of current word.
4188 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4189 } else {
4190 // Zero to beginning of next fully initialized word.
4191 // Or, don't zero at all, if we are already in that word.
4192 assert(next_full_store >= zeroes_needed, "must go forward");
4193 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4194 zeroes_needed = next_full_store;
4195 }
4196 }
4197
4198 if (zeroes_needed > zeroes_done) {
4199 intptr_t zsize = zeroes_needed - zeroes_done;
4200 // Do some incremental zeroing on rawmem, in parallel with inits.
4201 zeroes_done = align_down(zeroes_done, BytesPerInt);
4202 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4203 zeroes_done, zeroes_needed,
4204 phase);
4205 zeroes_done = zeroes_needed;
4206 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4207 do_zeroing = false; // leave the hole, next time
4208 }
4209 }
4210
4211 // Collect the store and move on:
4212 st->set_req(MemNode::Memory, inits);
4213 inits = st; // put it on the linearized chain
4214 set_req(i, zmem); // unhook from previous position
4215
4216 if (zeroes_done == st_off)
4217 zeroes_done = next_init_off;
4218
4219 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4220
4221 #ifdef ASSERT
4222 // Various order invariants. Weaker than stores_are_sane because
4223 // a large constant tile can be filled in by smaller non-constant stores.
4224 assert(st_off >= last_init_off, "inits do not reverse");
4225 last_init_off = st_off;
4226 const Type* val = NULL;
4227 if (st_size >= BytesPerInt &&
4228 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4229 (int)val->basic_type() < (int)T_OBJECT) {
4230 assert(st_off >= last_tile_end, "tiles do not overlap");
4231 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4232 last_tile_end = MAX2(last_tile_end, next_init_off);
4233 } else {
4234 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4235 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4236 assert(st_off >= last_init_end, "inits do not overlap");
4237 last_init_end = next_init_off; // it's a non-tile
4238 }
4239 #endif //ASSERT
4240 }
4241
4242 remove_extra_zeroes(); // clear out all the zmems left over
4243 add_req(inits);
4244
4245 if (!(UseTLAB && ZeroTLAB)) {
4246 // If anything remains to be zeroed, zero it all now.
4247 zeroes_done = align_down(zeroes_done, BytesPerInt);
4248 // if it is the last unused 4 bytes of an instance, forget about it
4249 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4250 if (zeroes_done + BytesPerLong >= size_limit) {
4251 AllocateNode* alloc = allocation();
4252 assert(alloc != NULL, "must be present");
4253 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4254 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4255 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4256 if (zeroes_done == k->layout_helper())
4257 zeroes_done = size_limit;
4258 }
4259 }
4260 if (zeroes_done < size_limit) {
4261 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4262 zeroes_done, size_in_bytes, phase);
4263 }
4264 }
4265
4266 set_complete(phase);
4267 return rawmem;
4268 }
4269
4270
4271 #ifdef ASSERT
4272 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4273 if (is_complete())
4274 return true; // stores could be anything at this point
4275 assert(allocation() != NULL, "must be present");
4276 intptr_t last_off = allocation()->minimum_header_size();
4277 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4278 Node* st = in(i);
4279 intptr_t st_off = get_store_offset(st, phase);
4280 if (st_off < 0) continue; // ignore dead garbage
4281 if (last_off > st_off) {
4282 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4283 this->dump(2);
4284 assert(false, "ascending store offsets");
4285 return false;
4286 }
4287 last_off = st_off + st->as_Store()->memory_size();
4288 }
4289 return true;
4290 }
4291 #endif //ASSERT
4292
4293
4294
4295
4296 //============================MergeMemNode=====================================
4297 //
4298 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4299 // contributing store or call operations. Each contributor provides the memory
4300 // state for a particular "alias type" (see Compile::alias_type). For example,
4301 // if a MergeMem has an input X for alias category #6, then any memory reference
4302 // to alias category #6 may use X as its memory state input, as an exact equivalent
4303 // to using the MergeMem as a whole.
4304 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4305 //
4306 // (Here, the <N> notation gives the index of the relevant adr_type.)
4307 //
4308 // In one special case (and more cases in the future), alias categories overlap.
4309 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4310 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4311 // it is exactly equivalent to that state W:
4312 // MergeMem(<Bot>: W) <==> W
4313 //
4314 // Usually, the merge has more than one input. In that case, where inputs
4315 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4316 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4317 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4318 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4319 //
4320 // A merge can take a "wide" memory state as one of its narrow inputs.
4321 // This simply means that the merge observes out only the relevant parts of
4322 // the wide input. That is, wide memory states arriving at narrow merge inputs
4323 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4324 //
4325 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4326 // and that memory slices "leak through":
4327 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4328 //
4329 // But, in such a cascade, repeated memory slices can "block the leak":
4330 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4331 //
4332 // In the last example, Y is not part of the combined memory state of the
4333 // outermost MergeMem. The system must, of course, prevent unschedulable
4334 // memory states from arising, so you can be sure that the state Y is somehow
4335 // a precursor to state Y'.
4336 //
4337 //
4338 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4339 // of each MergeMemNode array are exactly the numerical alias indexes, including
4340 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4341 // Compile::alias_type (and kin) produce and manage these indexes.
4342 //
4343 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4344 // (Note that this provides quick access to the top node inside MergeMem methods,
4345 // without the need to reach out via TLS to Compile::current.)
4346 //
4347 // As a consequence of what was just described, a MergeMem that represents a full
4348 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4349 // containing all alias categories.
4350 //
4351 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4352 //
4353 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4354 // a memory state for the alias type <N>, or else the top node, meaning that
4355 // there is no particular input for that alias type. Note that the length of
4356 // a MergeMem is variable, and may be extended at any time to accommodate new
4357 // memory states at larger alias indexes. When merges grow, they are of course
4358 // filled with "top" in the unused in() positions.
4359 //
4360 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4361 // (Top was chosen because it works smoothly with passes like GCM.)
4362 //
4363 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4364 // the type of random VM bits like TLS references.) Since it is always the
4365 // first non-Bot memory slice, some low-level loops use it to initialize an
4366 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4367 //
4368 //
4369 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4370 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4371 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4372 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4373 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4374 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4375 //
4376 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4377 // really that different from the other memory inputs. An abbreviation called
4378 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4379 //
4380 //
4381 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4382 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4383 // that "emerges though" the base memory will be marked as excluding the alias types
4384 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4385 //
4386 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4387 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4388 //
4389 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4390 // (It is currently unimplemented.) As you can see, the resulting merge is
4391 // actually a disjoint union of memory states, rather than an overlay.
4392 //
4393
4394 //------------------------------MergeMemNode-----------------------------------
4395 Node* MergeMemNode::make_empty_memory() {
4396 Node* empty_memory = (Node*) Compile::current()->top();
4397 assert(empty_memory->is_top(), "correct sentinel identity");
4398 return empty_memory;
4399 }
4400
4401 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4402 init_class_id(Class_MergeMem);
4403 // all inputs are nullified in Node::Node(int)
4404 // set_input(0, NULL); // no control input
4405
4406 // Initialize the edges uniformly to top, for starters.
4407 Node* empty_mem = make_empty_memory();
4408 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4409 init_req(i,empty_mem);
4410 }
4411 assert(empty_memory() == empty_mem, "");
4412
4413 if( new_base != NULL && new_base->is_MergeMem() ) {
4414 MergeMemNode* mdef = new_base->as_MergeMem();
4415 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4416 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4417 mms.set_memory(mms.memory2());
4418 }
4419 assert(base_memory() == mdef->base_memory(), "");
4420 } else {
4421 set_base_memory(new_base);
4422 }
4423 }
4424
4425 // Make a new, untransformed MergeMem with the same base as 'mem'.
4426 // If mem is itself a MergeMem, populate the result with the same edges.
4427 MergeMemNode* MergeMemNode::make(Node* mem) {
4428 return new MergeMemNode(mem);
4429 }
4430
4431 //------------------------------cmp--------------------------------------------
4432 uint MergeMemNode::hash() const { return NO_HASH; }
4433 uint MergeMemNode::cmp( const Node &n ) const {
4434 return (&n == this); // Always fail except on self
4435 }
4436
4437 //------------------------------Identity---------------------------------------
4438 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4439 // Identity if this merge point does not record any interesting memory
4440 // disambiguations.
4441 Node* base_mem = base_memory();
4442 Node* empty_mem = empty_memory();
4443 if (base_mem != empty_mem) { // Memory path is not dead?
4444 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4445 Node* mem = in(i);
4446 if (mem != empty_mem && mem != base_mem) {
4447 return this; // Many memory splits; no change
4448 }
4449 }
4450 }
4451 return base_mem; // No memory splits; ID on the one true input
4452 }
4453
4454 //------------------------------Ideal------------------------------------------
4455 // This method is invoked recursively on chains of MergeMem nodes
4456 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4457 // Remove chain'd MergeMems
4458 //
4459 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4460 // relative to the "in(Bot)". Since we are patching both at the same time,
4461 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4462 // but rewrite each "in(i)" relative to the new "in(Bot)".
4463 Node *progress = NULL;
4464
4465
4466 Node* old_base = base_memory();
4467 Node* empty_mem = empty_memory();
4468 if (old_base == empty_mem)
4469 return NULL; // Dead memory path.
4470
4471 MergeMemNode* old_mbase;
4472 if (old_base != NULL && old_base->is_MergeMem())
4473 old_mbase = old_base->as_MergeMem();
4474 else
4475 old_mbase = NULL;
4476 Node* new_base = old_base;
4477
4478 // simplify stacked MergeMems in base memory
4479 if (old_mbase) new_base = old_mbase->base_memory();
4480
4481 // the base memory might contribute new slices beyond my req()
4482 if (old_mbase) grow_to_match(old_mbase);
4483
4484 // Look carefully at the base node if it is a phi.
4485 PhiNode* phi_base;
4486 if (new_base != NULL && new_base->is_Phi())
4487 phi_base = new_base->as_Phi();
4488 else
4489 phi_base = NULL;
4490
4491 Node* phi_reg = NULL;
4492 uint phi_len = (uint)-1;
4493 if (phi_base != NULL && !phi_base->is_copy()) {
4494 // do not examine phi if degraded to a copy
4495 phi_reg = phi_base->region();
4496 phi_len = phi_base->req();
4497 // see if the phi is unfinished
4498 for (uint i = 1; i < phi_len; i++) {
4499 if (phi_base->in(i) == NULL) {
4500 // incomplete phi; do not look at it yet!
4501 phi_reg = NULL;
4502 phi_len = (uint)-1;
4503 break;
4504 }
4505 }
4506 }
4507
4508 // Note: We do not call verify_sparse on entry, because inputs
4509 // can normalize to the base_memory via subsume_node or similar
4510 // mechanisms. This method repairs that damage.
4511
4512 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4513
4514 // Look at each slice.
4515 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4516 Node* old_in = in(i);
4517 // calculate the old memory value
4518 Node* old_mem = old_in;
4519 if (old_mem == empty_mem) old_mem = old_base;
4520 assert(old_mem == memory_at(i), "");
4521
4522 // maybe update (reslice) the old memory value
4523
4524 // simplify stacked MergeMems
4525 Node* new_mem = old_mem;
4526 MergeMemNode* old_mmem;
4527 if (old_mem != NULL && old_mem->is_MergeMem())
4528 old_mmem = old_mem->as_MergeMem();
4529 else
4530 old_mmem = NULL;
4531 if (old_mmem == this) {
4532 // This can happen if loops break up and safepoints disappear.
4533 // A merge of BotPtr (default) with a RawPtr memory derived from a
4534 // safepoint can be rewritten to a merge of the same BotPtr with
4535 // the BotPtr phi coming into the loop. If that phi disappears
4536 // also, we can end up with a self-loop of the mergemem.
4537 // In general, if loops degenerate and memory effects disappear,
4538 // a mergemem can be left looking at itself. This simply means
4539 // that the mergemem's default should be used, since there is
4540 // no longer any apparent effect on this slice.
4541 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4542 // from start. Update the input to TOP.
4543 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4544 }
4545 else if (old_mmem != NULL) {
4546 new_mem = old_mmem->memory_at(i);
4547 }
4548 // else preceding memory was not a MergeMem
4549
4550 // replace equivalent phis (unfortunately, they do not GVN together)
4551 if (new_mem != NULL && new_mem != new_base &&
4552 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4553 if (new_mem->is_Phi()) {
4554 PhiNode* phi_mem = new_mem->as_Phi();
4555 for (uint i = 1; i < phi_len; i++) {
4556 if (phi_base->in(i) != phi_mem->in(i)) {
4557 phi_mem = NULL;
4558 break;
4559 }
4560 }
4561 if (phi_mem != NULL) {
4562 // equivalent phi nodes; revert to the def
4563 new_mem = new_base;
4564 }
4565 }
4566 }
4567
4568 // maybe store down a new value
4569 Node* new_in = new_mem;
4570 if (new_in == new_base) new_in = empty_mem;
4571
4572 if (new_in != old_in) {
4573 // Warning: Do not combine this "if" with the previous "if"
4574 // A memory slice might have be be rewritten even if it is semantically
4575 // unchanged, if the base_memory value has changed.
4576 set_req(i, new_in);
4577 progress = this; // Report progress
4578 }
4579 }
4580
4581 if (new_base != old_base) {
4582 set_req(Compile::AliasIdxBot, new_base);
4583 // Don't use set_base_memory(new_base), because we need to update du.
4584 assert(base_memory() == new_base, "");
4585 progress = this;
4586 }
4587
4588 if( base_memory() == this ) {
4589 // a self cycle indicates this memory path is dead
4590 set_req(Compile::AliasIdxBot, empty_mem);
4591 }
4592
4593 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4594 // Recursion must occur after the self cycle check above
4595 if( base_memory()->is_MergeMem() ) {
4596 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4597 Node *m = phase->transform(new_mbase); // Rollup any cycles
4598 if( m != NULL &&
4599 (m->is_top() ||
4600 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4601 // propagate rollup of dead cycle to self
4602 set_req(Compile::AliasIdxBot, empty_mem);
4603 }
4604 }
4605
4606 if( base_memory() == empty_mem ) {
4607 progress = this;
4608 // Cut inputs during Parse phase only.
4609 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4610 if( !can_reshape ) {
4611 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4612 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4613 }
4614 }
4615 }
4616
4617 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4618 // Check if PhiNode::Ideal's "Split phis through memory merges"
4619 // transform should be attempted. Look for this->phi->this cycle.
4620 uint merge_width = req();
4621 if (merge_width > Compile::AliasIdxRaw) {
4622 PhiNode* phi = base_memory()->as_Phi();
4623 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4624 if (phi->in(i) == this) {
4625 phase->is_IterGVN()->_worklist.push(phi);
4626 break;
4627 }
4628 }
4629 }
4630 }
4631
4632 assert(progress || verify_sparse(), "please, no dups of base");
4633 return progress;
4634 }
4635
4636 //-------------------------set_base_memory-------------------------------------
4637 void MergeMemNode::set_base_memory(Node *new_base) {
4638 Node* empty_mem = empty_memory();
4639 set_req(Compile::AliasIdxBot, new_base);
4640 assert(memory_at(req()) == new_base, "must set default memory");
4641 // Clear out other occurrences of new_base:
4642 if (new_base != empty_mem) {
4643 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4644 if (in(i) == new_base) set_req(i, empty_mem);
4645 }
4646 }
4647 }
4648
4649 //------------------------------out_RegMask------------------------------------
4650 const RegMask &MergeMemNode::out_RegMask() const {
4651 return RegMask::Empty;
4652 }
4653
4654 //------------------------------dump_spec--------------------------------------
4655 #ifndef PRODUCT
4656 void MergeMemNode::dump_spec(outputStream *st) const {
4657 st->print(" {");
4658 Node* base_mem = base_memory();
4659 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4660 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4661 if (mem == base_mem) { st->print(" -"); continue; }
4662 st->print( " N%d:", mem->_idx );
4663 Compile::current()->get_adr_type(i)->dump_on(st);
4664 }
4665 st->print(" }");
4666 }
4667 #endif // !PRODUCT
4668
4669
4670 #ifdef ASSERT
4671 static bool might_be_same(Node* a, Node* b) {
4672 if (a == b) return true;
4673 if (!(a->is_Phi() || b->is_Phi())) return false;
4674 // phis shift around during optimization
4675 return true; // pretty stupid...
4676 }
4677
4678 // verify a narrow slice (either incoming or outgoing)
4679 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4680 if (!VerifyAliases) return; // don't bother to verify unless requested
4681 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4682 if (Node::in_dump()) return; // muzzle asserts when printing
4683 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4684 assert(n != NULL, "");
4685 // Elide intervening MergeMem's
4686 while (n->is_MergeMem()) {
4687 n = n->as_MergeMem()->memory_at(alias_idx);
4688 }
4689 Compile* C = Compile::current();
4690 const TypePtr* n_adr_type = n->adr_type();
4691 if (n == m->empty_memory()) {
4692 // Implicit copy of base_memory()
4693 } else if (n_adr_type != TypePtr::BOTTOM) {
4694 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4695 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4696 } else {
4697 // A few places like make_runtime_call "know" that VM calls are narrow,
4698 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4699 bool expected_wide_mem = false;
4700 if (n == m->base_memory()) {
4701 expected_wide_mem = true;
4702 } else if (alias_idx == Compile::AliasIdxRaw ||
4703 n == m->memory_at(Compile::AliasIdxRaw)) {
4704 expected_wide_mem = true;
4705 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4706 // memory can "leak through" calls on channels that
4707 // are write-once. Allow this also.
4708 expected_wide_mem = true;
4709 }
4710 assert(expected_wide_mem, "expected narrow slice replacement");
4711 }
4712 }
4713 #else // !ASSERT
4714 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4715 #endif
4716
4717
4718 //-----------------------------memory_at---------------------------------------
4719 Node* MergeMemNode::memory_at(uint alias_idx) const {
4720 assert(alias_idx >= Compile::AliasIdxRaw ||
4721 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4722 "must avoid base_memory and AliasIdxTop");
4723
4724 // Otherwise, it is a narrow slice.
4725 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4726 Compile *C = Compile::current();
4727 if (is_empty_memory(n)) {
4728 // the array is sparse; empty slots are the "top" node
4729 n = base_memory();
4730 assert(Node::in_dump()
4731 || n == NULL || n->bottom_type() == Type::TOP
4732 || n->adr_type() == NULL // address is TOP
4733 || n->adr_type() == TypePtr::BOTTOM
4734 || n->adr_type() == TypeRawPtr::BOTTOM
4735 || Compile::current()->AliasLevel() == 0,
4736 "must be a wide memory");
4737 // AliasLevel == 0 if we are organizing the memory states manually.
4738 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4739 } else {
4740 // make sure the stored slice is sane
4741 #ifdef ASSERT
4742 if (VMError::is_error_reported() || Node::in_dump()) {
4743 } else if (might_be_same(n, base_memory())) {
4744 // Give it a pass: It is a mostly harmless repetition of the base.
4745 // This can arise normally from node subsumption during optimization.
4746 } else {
4747 verify_memory_slice(this, alias_idx, n);
4748 }
4749 #endif
4750 }
4751 return n;
4752 }
4753
4754 //---------------------------set_memory_at-------------------------------------
4755 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4756 verify_memory_slice(this, alias_idx, n);
4757 Node* empty_mem = empty_memory();
4758 if (n == base_memory()) n = empty_mem; // collapse default
4759 uint need_req = alias_idx+1;
4760 if (req() < need_req) {
4761 if (n == empty_mem) return; // already the default, so do not grow me
4762 // grow the sparse array
4763 do {
4764 add_req(empty_mem);
4765 } while (req() < need_req);
4766 }
4767 set_req( alias_idx, n );
4768 }
4769
4770
4771
4772 //--------------------------iteration_setup------------------------------------
4773 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4774 if (other != NULL) {
4775 grow_to_match(other);
4776 // invariant: the finite support of mm2 is within mm->req()
4777 #ifdef ASSERT
4778 for (uint i = req(); i < other->req(); i++) {
4779 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4780 }
4781 #endif
4782 }
4783 // Replace spurious copies of base_memory by top.
4784 Node* base_mem = base_memory();
4785 if (base_mem != NULL && !base_mem->is_top()) {
4786 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4787 if (in(i) == base_mem)
4788 set_req(i, empty_memory());
4789 }
4790 }
4791 }
4792
4793 //---------------------------grow_to_match-------------------------------------
4794 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4795 Node* empty_mem = empty_memory();
4796 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4797 // look for the finite support of the other memory
4798 for (uint i = other->req(); --i >= req(); ) {
4799 if (other->in(i) != empty_mem) {
4800 uint new_len = i+1;
4801 while (req() < new_len) add_req(empty_mem);
4802 break;
4803 }
4804 }
4805 }
4806
4807 //---------------------------verify_sparse-------------------------------------
4808 #ifndef PRODUCT
4809 bool MergeMemNode::verify_sparse() const {
4810 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4811 Node* base_mem = base_memory();
4812 // The following can happen in degenerate cases, since empty==top.
4813 if (is_empty_memory(base_mem)) return true;
4814 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4815 assert(in(i) != NULL, "sane slice");
4816 if (in(i) == base_mem) return false; // should have been the sentinel value!
4817 }
4818 return true;
4819 }
4820
4821 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4822 Node* n;
4823 n = mm->in(idx);
4824 if (mem == n) return true; // might be empty_memory()
4825 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4826 if (mem == n) return true;
4827 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4828 if (mem == n) return true;
4829 if (n == NULL) break;
4830 }
4831 return false;
4832 }
4833 #endif // !PRODUCT