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 bool 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->eqv_uncast(st_base, /*keep_deps=*/true)) 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 // Casted version may carry a dependency and it is respected.
1066 // Thus, we are able to replace L by V.
1067 }
1068 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1069 if (store_Opcode() != st->Opcode())
1070 return NULL;
1071 return st->in(MemNode::ValueIn);
1072 }
1073
1074 // A load from a freshly-created object always returns zero.
1075 // (This can happen after LoadNode::Ideal resets the load's memory input
1076 // to find_captured_store, which returned InitializeNode::zero_memory.)
1077 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1078 (st->in(0) == ld_alloc) &&
1079 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1080 // return a zero value for the load's basic type
1081 // (This is one of the few places where a generic PhaseTransform
1082 // can create new nodes. Think of it as lazily manifesting
1083 // virtually pre-existing constants.)
1084 return phase->zerocon(memory_type());
1085 }
1086
1087 // A load from an initialization barrier can match a captured store.
1088 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1089 InitializeNode* init = st->in(0)->as_Initialize();
1090 AllocateNode* alloc = init->allocation();
1091 if ((alloc != NULL) && (alloc == ld_alloc)) {
1092 // examine a captured store value
1093 st = init->find_captured_store(ld_off, memory_size(), phase);
1094 if (st != NULL) {
1095 continue; // take one more trip around
1096 }
1097 }
1098 }
1099
1100 // Load boxed value from result of valueOf() call is input parameter.
1101 if (this->is_Load() && ld_adr->is_AddP() &&
1102 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1103 intptr_t ignore = 0;
1104 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1105 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1106 base = bs->step_over_gc_barrier(base);
1107 if (base != NULL && base->is_Proj() &&
1108 base->as_Proj()->_con == TypeFunc::Parms &&
1109 base->in(0)->is_CallStaticJava() &&
1110 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1111 return base->in(0)->in(TypeFunc::Parms);
1112 }
1113 }
1114
1115 break;
1116 }
1117
1118 return NULL;
1119 }
1120
1121 //----------------------is_instance_field_load_with_local_phi------------------
1122 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1123 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1124 in(Address)->is_AddP() ) {
1125 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1126 // Only instances and boxed values.
1127 if( t_oop != NULL &&
1128 (t_oop->is_ptr_to_boxed_value() ||
1129 t_oop->is_known_instance_field()) &&
1130 t_oop->offset() != Type::OffsetBot &&
1131 t_oop->offset() != Type::OffsetTop) {
1132 return true;
1133 }
1134 }
1135 return false;
1136 }
1137
1138 //------------------------------Identity---------------------------------------
1139 // Loads are identity if previous store is to same address
1140 Node* LoadNode::Identity(PhaseGVN* phase) {
1141 // If the previous store-maker is the right kind of Store, and the store is
1142 // to the same address, then we are equal to the value stored.
1143 Node* mem = in(Memory);
1144 Node* value = can_see_stored_value(mem, phase);
1145 if( value ) {
1146 // byte, short & char stores truncate naturally.
1147 // A load has to load the truncated value which requires
1148 // some sort of masking operation and that requires an
1149 // Ideal call instead of an Identity call.
1150 if (memory_size() < BytesPerInt) {
1151 // If the input to the store does not fit with the load's result type,
1152 // it must be truncated via an Ideal call.
1153 if (!phase->type(value)->higher_equal(phase->type(this)))
1154 return this;
1155 }
1156 // (This works even when value is a Con, but LoadNode::Value
1157 // usually runs first, producing the singleton type of the Con.)
1158 return value;
1159 }
1160
1161 // Search for an existing data phi which was generated before for the same
1162 // instance's field to avoid infinite generation of phis in a loop.
1163 Node *region = mem->in(0);
1164 if (is_instance_field_load_with_local_phi(region)) {
1165 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1166 int this_index = phase->C->get_alias_index(addr_t);
1167 int this_offset = addr_t->offset();
1168 int this_iid = addr_t->instance_id();
1169 if (!addr_t->is_known_instance() &&
1170 addr_t->is_ptr_to_boxed_value()) {
1171 // Use _idx of address base (could be Phi node) for boxed values.
1172 intptr_t ignore = 0;
1173 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1174 if (base == NULL) {
1175 return this;
1176 }
1177 this_iid = base->_idx;
1178 }
1179 const Type* this_type = bottom_type();
1180 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1181 Node* phi = region->fast_out(i);
1182 if (phi->is_Phi() && phi != mem &&
1183 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1184 return phi;
1185 }
1186 }
1187 }
1188
1189 return this;
1190 }
1191
1192 // Construct an equivalent unsigned load.
1193 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1194 BasicType bt = T_ILLEGAL;
1195 const Type* rt = NULL;
1196 switch (Opcode()) {
1197 case Op_LoadUB: return this;
1198 case Op_LoadUS: return this;
1199 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1200 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1201 default:
1202 assert(false, "no unsigned variant: %s", Name());
1203 return NULL;
1204 }
1205 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1206 raw_adr_type(), rt, bt, _mo, _control_dependency,
1207 is_unaligned_access(), is_mismatched_access());
1208 }
1209
1210 // Construct an equivalent signed load.
1211 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1212 BasicType bt = T_ILLEGAL;
1213 const Type* rt = NULL;
1214 switch (Opcode()) {
1215 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1216 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1217 case Op_LoadB: // fall through
1218 case Op_LoadS: // fall through
1219 case Op_LoadI: // fall through
1220 case Op_LoadL: return this;
1221 default:
1222 assert(false, "no signed variant: %s", Name());
1223 return NULL;
1224 }
1225 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1226 raw_adr_type(), rt, bt, _mo, _control_dependency,
1227 is_unaligned_access(), is_mismatched_access());
1228 }
1229
1230 // We're loading from an object which has autobox behaviour.
1231 // If this object is result of a valueOf call we'll have a phi
1232 // merging a newly allocated object and a load from the cache.
1233 // We want to replace this load with the original incoming
1234 // argument to the valueOf call.
1235 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1236 assert(phase->C->eliminate_boxing(), "sanity");
1237 intptr_t ignore = 0;
1238 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1239 if ((base == NULL) || base->is_Phi()) {
1240 // Push the loads from the phi that comes from valueOf up
1241 // through it to allow elimination of the loads and the recovery
1242 // of the original value. It is done in split_through_phi().
1243 return NULL;
1244 } else if (base->is_Load() ||
1245 (base->is_DecodeN() && base->in(1)->is_Load())) {
1246 // Eliminate the load of boxed value for integer types from the cache
1247 // array by deriving the value from the index into the array.
1248 // Capture the offset of the load and then reverse the computation.
1249
1250 // Get LoadN node which loads a boxing object from 'cache' array.
1251 if (base->is_DecodeN()) {
1252 base = base->in(1);
1253 }
1254 if (!base->in(Address)->is_AddP()) {
1255 return NULL; // Complex address
1256 }
1257 AddPNode* address = base->in(Address)->as_AddP();
1258 Node* cache_base = address->in(AddPNode::Base);
1259 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1260 // Get ConP node which is static 'cache' field.
1261 cache_base = cache_base->in(1);
1262 }
1263 if ((cache_base != NULL) && cache_base->is_Con()) {
1264 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1265 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1266 Node* elements[4];
1267 int shift = exact_log2(type2aelembytes(T_OBJECT));
1268 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1269 if (count > 0 && elements[0]->is_Con() &&
1270 (count == 1 ||
1271 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1272 elements[1]->in(2) == phase->intcon(shift)))) {
1273 ciObjArray* array = base_type->const_oop()->as_obj_array();
1274 // Fetch the box object cache[0] at the base of the array and get its value
1275 ciInstance* box = array->obj_at(0)->as_instance();
1276 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1277 assert(ik->is_box_klass(), "sanity");
1278 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1279 if (ik->nof_nonstatic_fields() == 1) {
1280 // This should be true nonstatic_field_at requires calling
1281 // nof_nonstatic_fields so check it anyway
1282 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1283 BasicType bt = c.basic_type();
1284 // Only integer types have boxing cache.
1285 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1286 bt == T_BYTE || bt == T_SHORT ||
1287 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1288 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1289 if (cache_low != (int)cache_low) {
1290 return NULL; // should not happen since cache is array indexed by value
1291 }
1292 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1293 if (offset != (int)offset) {
1294 return NULL; // should not happen since cache is array indexed by value
1295 }
1296 // Add up all the offsets making of the address of the load
1297 Node* result = elements[0];
1298 for (int i = 1; i < count; i++) {
1299 result = phase->transform(new AddXNode(result, elements[i]));
1300 }
1301 // Remove the constant offset from the address and then
1302 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1303 // remove the scaling of the offset to recover the original index.
1304 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1305 // Peel the shift off directly but wrap it in a dummy node
1306 // since Ideal can't return existing nodes
1307 result = new RShiftXNode(result->in(1), phase->intcon(0));
1308 } else if (result->is_Add() && result->in(2)->is_Con() &&
1309 result->in(1)->Opcode() == Op_LShiftX &&
1310 result->in(1)->in(2) == phase->intcon(shift)) {
1311 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1312 // but for boxing cache access we know that X<<Z will not overflow
1313 // (there is range check) so we do this optimizatrion by hand here.
1314 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1315 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1316 } else {
1317 result = new RShiftXNode(result, phase->intcon(shift));
1318 }
1319 #ifdef _LP64
1320 if (bt != T_LONG) {
1321 result = new ConvL2INode(phase->transform(result));
1322 }
1323 #else
1324 if (bt == T_LONG) {
1325 result = new ConvI2LNode(phase->transform(result));
1326 }
1327 #endif
1328 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1329 // Need to preserve unboxing load type if it is unsigned.
1330 switch(this->Opcode()) {
1331 case Op_LoadUB:
1332 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1333 break;
1334 case Op_LoadUS:
1335 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1336 break;
1337 }
1338 return result;
1339 }
1340 }
1341 }
1342 }
1343 }
1344 return NULL;
1345 }
1346
1347 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1348 Node* region = phi->in(0);
1349 if (region == NULL) {
1350 return false; // Wait stable graph
1351 }
1352 uint cnt = phi->req();
1353 for (uint i = 1; i < cnt; i++) {
1354 Node* rc = region->in(i);
1355 if (rc == NULL || phase->type(rc) == Type::TOP)
1356 return false; // Wait stable graph
1357 Node* in = phi->in(i);
1358 if (in == NULL || phase->type(in) == Type::TOP)
1359 return false; // Wait stable graph
1360 }
1361 return true;
1362 }
1363 //------------------------------split_through_phi------------------------------
1364 // Split instance or boxed field load through Phi.
1365 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1366 Node* mem = in(Memory);
1367 Node* address = in(Address);
1368 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1369
1370 assert((t_oop != NULL) &&
1371 (t_oop->is_known_instance_field() ||
1372 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1373
1374 Compile* C = phase->C;
1375 intptr_t ignore = 0;
1376 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1377 bool base_is_phi = (base != NULL) && base->is_Phi();
1378 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1379 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1380 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1381
1382 if (!((mem->is_Phi() || base_is_phi) &&
1383 (load_boxed_values || t_oop->is_known_instance_field()))) {
1384 return NULL; // memory is not Phi
1385 }
1386
1387 if (mem->is_Phi()) {
1388 if (!stable_phi(mem->as_Phi(), phase)) {
1389 return NULL; // Wait stable graph
1390 }
1391 uint cnt = mem->req();
1392 // Check for loop invariant memory.
1393 if (cnt == 3) {
1394 for (uint i = 1; i < cnt; i++) {
1395 Node* in = mem->in(i);
1396 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1397 if (m == mem) {
1398 if (i == 1) {
1399 // if the first edge was a loop, check second edge too.
1400 // If both are replaceable - we are in an infinite loop
1401 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1402 if (n == mem) {
1403 break;
1404 }
1405 }
1406 set_req(Memory, mem->in(cnt - i));
1407 return this; // made change
1408 }
1409 }
1410 }
1411 }
1412 if (base_is_phi) {
1413 if (!stable_phi(base->as_Phi(), phase)) {
1414 return NULL; // Wait stable graph
1415 }
1416 uint cnt = base->req();
1417 // Check for loop invariant memory.
1418 if (cnt == 3) {
1419 for (uint i = 1; i < cnt; i++) {
1420 if (base->in(i) == base) {
1421 return NULL; // Wait stable graph
1422 }
1423 }
1424 }
1425 }
1426
1427 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1428
1429 // Split through Phi (see original code in loopopts.cpp).
1430 assert(C->have_alias_type(t_oop), "instance should have alias type");
1431
1432 // Do nothing here if Identity will find a value
1433 // (to avoid infinite chain of value phis generation).
1434 if (!phase->eqv(this, phase->apply_identity(this)))
1435 return NULL;
1436
1437 // Select Region to split through.
1438 Node* region;
1439 if (!base_is_phi) {
1440 assert(mem->is_Phi(), "sanity");
1441 region = mem->in(0);
1442 // Skip if the region dominates some control edge of the address.
1443 if (!MemNode::all_controls_dominate(address, region))
1444 return NULL;
1445 } else if (!mem->is_Phi()) {
1446 assert(base_is_phi, "sanity");
1447 region = base->in(0);
1448 // Skip if the region dominates some control edge of the memory.
1449 if (!MemNode::all_controls_dominate(mem, region))
1450 return NULL;
1451 } else if (base->in(0) != mem->in(0)) {
1452 assert(base_is_phi && mem->is_Phi(), "sanity");
1453 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1454 region = base->in(0);
1455 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1456 region = mem->in(0);
1457 } else {
1458 return NULL; // complex graph
1459 }
1460 } else {
1461 assert(base->in(0) == mem->in(0), "sanity");
1462 region = mem->in(0);
1463 }
1464
1465 const Type* this_type = this->bottom_type();
1466 int this_index = C->get_alias_index(t_oop);
1467 int this_offset = t_oop->offset();
1468 int this_iid = t_oop->instance_id();
1469 if (!t_oop->is_known_instance() && load_boxed_values) {
1470 // Use _idx of address base for boxed values.
1471 this_iid = base->_idx;
1472 }
1473 PhaseIterGVN* igvn = phase->is_IterGVN();
1474 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1475 for (uint i = 1; i < region->req(); i++) {
1476 Node* x;
1477 Node* the_clone = NULL;
1478 if (region->in(i) == C->top()) {
1479 x = C->top(); // Dead path? Use a dead data op
1480 } else {
1481 x = this->clone(); // Else clone up the data op
1482 the_clone = x; // Remember for possible deletion.
1483 // Alter data node to use pre-phi inputs
1484 if (this->in(0) == region) {
1485 x->set_req(0, region->in(i));
1486 } else {
1487 x->set_req(0, NULL);
1488 }
1489 if (mem->is_Phi() && (mem->in(0) == region)) {
1490 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1491 }
1492 if (address->is_Phi() && address->in(0) == region) {
1493 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1494 }
1495 if (base_is_phi && (base->in(0) == region)) {
1496 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1497 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1498 x->set_req(Address, adr_x);
1499 }
1500 }
1501 // Check for a 'win' on some paths
1502 const Type *t = x->Value(igvn);
1503
1504 bool singleton = t->singleton();
1505
1506 // See comments in PhaseIdealLoop::split_thru_phi().
1507 if (singleton && t == Type::TOP) {
1508 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1509 }
1510
1511 if (singleton) {
1512 x = igvn->makecon(t);
1513 } else {
1514 // We now call Identity to try to simplify the cloned node.
1515 // Note that some Identity methods call phase->type(this).
1516 // Make sure that the type array is big enough for
1517 // our new node, even though we may throw the node away.
1518 // (This tweaking with igvn only works because x is a new node.)
1519 igvn->set_type(x, t);
1520 // If x is a TypeNode, capture any more-precise type permanently into Node
1521 // otherwise it will be not updated during igvn->transform since
1522 // igvn->type(x) is set to x->Value() already.
1523 x->raise_bottom_type(t);
1524 Node *y = igvn->apply_identity(x);
1525 if (y != x) {
1526 x = y;
1527 } else {
1528 y = igvn->hash_find_insert(x);
1529 if (y) {
1530 x = y;
1531 } else {
1532 // Else x is a new node we are keeping
1533 // We do not need register_new_node_with_optimizer
1534 // because set_type has already been called.
1535 igvn->_worklist.push(x);
1536 }
1537 }
1538 }
1539 if (x != the_clone && the_clone != NULL) {
1540 igvn->remove_dead_node(the_clone);
1541 }
1542 phi->set_req(i, x);
1543 }
1544 // Record Phi
1545 igvn->register_new_node_with_optimizer(phi);
1546 return phi;
1547 }
1548
1549 //------------------------------Ideal------------------------------------------
1550 // If the load is from Field memory and the pointer is non-null, it might be possible to
1551 // zero out the control input.
1552 // If the offset is constant and the base is an object allocation,
1553 // try to hook me up to the exact initializing store.
1554 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1555 Node* p = MemNode::Ideal_common(phase, can_reshape);
1556 if (p) return (p == NodeSentinel) ? NULL : p;
1557
1558 Node* ctrl = in(MemNode::Control);
1559 Node* address = in(MemNode::Address);
1560 bool progress = false;
1561
1562 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1563 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1564
1565 // Skip up past a SafePoint control. Cannot do this for Stores because
1566 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1567 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1568 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1569 !addr_mark ) {
1570 ctrl = ctrl->in(0);
1571 set_req(MemNode::Control,ctrl);
1572 progress = true;
1573 }
1574
1575 intptr_t ignore = 0;
1576 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1577 if (base != NULL
1578 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1579 // Check for useless control edge in some common special cases
1580 if (in(MemNode::Control) != NULL
1581 && can_remove_control()
1582 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1583 && all_controls_dominate(base, phase->C->start())) {
1584 // A method-invariant, non-null address (constant or 'this' argument).
1585 set_req(MemNode::Control, NULL);
1586 progress = true;
1587 }
1588 }
1589
1590 Node* mem = in(MemNode::Memory);
1591 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1592
1593 if (can_reshape && (addr_t != NULL)) {
1594 // try to optimize our memory input
1595 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1596 if (opt_mem != mem) {
1597 set_req(MemNode::Memory, opt_mem);
1598 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1599 return this;
1600 }
1601 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1602 if ((t_oop != NULL) &&
1603 (t_oop->is_known_instance_field() ||
1604 t_oop->is_ptr_to_boxed_value())) {
1605 PhaseIterGVN *igvn = phase->is_IterGVN();
1606 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1607 // Delay this transformation until memory Phi is processed.
1608 phase->is_IterGVN()->_worklist.push(this);
1609 return NULL;
1610 }
1611 // Split instance field load through Phi.
1612 Node* result = split_through_phi(phase);
1613 if (result != NULL) return result;
1614
1615 if (t_oop->is_ptr_to_boxed_value()) {
1616 Node* result = eliminate_autobox(phase);
1617 if (result != NULL) return result;
1618 }
1619 }
1620 }
1621
1622 // Is there a dominating load that loads the same value? Leave
1623 // anything that is not a load of a field/array element (like
1624 // barriers etc.) alone
1625 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1626 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1627 Node *use = mem->fast_out(i);
1628 if (use != this &&
1629 use->Opcode() == Opcode() &&
1630 use->in(0) != NULL &&
1631 use->in(0) != in(0) &&
1632 use->in(Address) == in(Address)) {
1633 Node* ctl = in(0);
1634 for (int i = 0; i < 10 && ctl != NULL; i++) {
1635 ctl = IfNode::up_one_dom(ctl);
1636 if (ctl == use->in(0)) {
1637 set_req(0, use->in(0));
1638 return this;
1639 }
1640 }
1641 }
1642 }
1643 }
1644
1645 // Check for prior store with a different base or offset; make Load
1646 // independent. Skip through any number of them. Bail out if the stores
1647 // are in an endless dead cycle and report no progress. This is a key
1648 // transform for Reflection. However, if after skipping through the Stores
1649 // we can't then fold up against a prior store do NOT do the transform as
1650 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1651 // array memory alive twice: once for the hoisted Load and again after the
1652 // bypassed Store. This situation only works if EVERYBODY who does
1653 // anti-dependence work knows how to bypass. I.e. we need all
1654 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1655 // the alias index stuff. So instead, peek through Stores and IFF we can
1656 // fold up, do so.
1657 Node* prev_mem = find_previous_store(phase);
1658 if (prev_mem != NULL) {
1659 Node* value = can_see_arraycopy_value(prev_mem, phase);
1660 if (value != NULL) {
1661 return value;
1662 }
1663 }
1664 // Steps (a), (b): Walk past independent stores to find an exact match.
1665 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1666 // (c) See if we can fold up on the spot, but don't fold up here.
1667 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1668 // just return a prior value, which is done by Identity calls.
1669 if (can_see_stored_value(prev_mem, phase)) {
1670 // Make ready for step (d):
1671 set_req(MemNode::Memory, prev_mem);
1672 return this;
1673 }
1674 }
1675
1676 return progress ? this : NULL;
1677 }
1678
1679 // Helper to recognize certain Klass fields which are invariant across
1680 // some group of array types (e.g., int[] or all T[] where T < Object).
1681 const Type*
1682 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1683 ciKlass* klass) const {
1684 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1685 // The field is Klass::_modifier_flags. Return its (constant) value.
1686 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1687 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1688 return TypeInt::make(klass->modifier_flags());
1689 }
1690 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1691 // The field is Klass::_access_flags. Return its (constant) value.
1692 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1693 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1694 return TypeInt::make(klass->access_flags());
1695 }
1696 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1697 // The field is Klass::_layout_helper. Return its constant value if known.
1698 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1699 return TypeInt::make(klass->layout_helper());
1700 }
1701
1702 // No match.
1703 return NULL;
1704 }
1705
1706 //------------------------------Value-----------------------------------------
1707 const Type* LoadNode::Value(PhaseGVN* phase) const {
1708 // Either input is TOP ==> the result is TOP
1709 Node* mem = in(MemNode::Memory);
1710 const Type *t1 = phase->type(mem);
1711 if (t1 == Type::TOP) return Type::TOP;
1712 Node* adr = in(MemNode::Address);
1713 const TypePtr* tp = phase->type(adr)->isa_ptr();
1714 if (tp == NULL || tp->empty()) return Type::TOP;
1715 int off = tp->offset();
1716 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1717 Compile* C = phase->C;
1718
1719 // Try to guess loaded type from pointer type
1720 if (tp->isa_aryptr()) {
1721 const TypeAryPtr* ary = tp->is_aryptr();
1722 const Type* t = ary->elem();
1723
1724 // Determine whether the reference is beyond the header or not, by comparing
1725 // the offset against the offset of the start of the array's data.
1726 // Different array types begin at slightly different offsets (12 vs. 16).
1727 // We choose T_BYTE as an example base type that is least restrictive
1728 // as to alignment, which will therefore produce the smallest
1729 // possible base offset.
1730 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1731 const bool off_beyond_header = (off >= min_base_off);
1732
1733 // Try to constant-fold a stable array element.
1734 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1735 // Make sure the reference is not into the header and the offset is constant
1736 ciObject* aobj = ary->const_oop();
1737 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1738 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1739 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1740 stable_dimension,
1741 memory_type(), is_unsigned());
1742 if (con_type != NULL) {
1743 return con_type;
1744 }
1745 }
1746 }
1747
1748 // Don't do this for integer types. There is only potential profit if
1749 // the element type t is lower than _type; that is, for int types, if _type is
1750 // more restrictive than t. This only happens here if one is short and the other
1751 // char (both 16 bits), and in those cases we've made an intentional decision
1752 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1753 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1754 //
1755 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1756 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1757 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1758 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1759 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1760 // In fact, that could have been the original type of p1, and p1 could have
1761 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1762 // expression (LShiftL quux 3) independently optimized to the constant 8.
1763 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1764 && (_type->isa_vect() == NULL)
1765 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1766 // t might actually be lower than _type, if _type is a unique
1767 // concrete subclass of abstract class t.
1768 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
1769 const Type* jt = t->join_speculative(_type);
1770 // In any case, do not allow the join, per se, to empty out the type.
1771 if (jt->empty() && !t->empty()) {
1772 // This can happen if a interface-typed array narrows to a class type.
1773 jt = _type;
1774 }
1775 #ifdef ASSERT
1776 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1777 // The pointers in the autobox arrays are always non-null
1778 Node* base = adr->in(AddPNode::Base);
1779 if ((base != NULL) && base->is_DecodeN()) {
1780 // Get LoadN node which loads IntegerCache.cache field
1781 base = base->in(1);
1782 }
1783 if ((base != NULL) && base->is_Con()) {
1784 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1785 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1786 // It could be narrow oop
1787 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1788 }
1789 }
1790 }
1791 #endif
1792 return jt;
1793 }
1794 }
1795 } else if (tp->base() == Type::InstPtr) {
1796 assert( off != Type::OffsetBot ||
1797 // arrays can be cast to Objects
1798 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1799 // unsafe field access may not have a constant offset
1800 C->has_unsafe_access(),
1801 "Field accesses must be precise" );
1802 // For oop loads, we expect the _type to be precise.
1803
1804 // Optimize loads from constant fields.
1805 const TypeInstPtr* tinst = tp->is_instptr();
1806 ciObject* const_oop = tinst->const_oop();
1807 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1808 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1809 if (con_type != NULL) {
1810 return con_type;
1811 }
1812 }
1813 } else if (tp->base() == Type::KlassPtr) {
1814 assert( off != Type::OffsetBot ||
1815 // arrays can be cast to Objects
1816 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1817 // also allow array-loading from the primary supertype
1818 // array during subtype checks
1819 Opcode() == Op_LoadKlass,
1820 "Field accesses must be precise" );
1821 // For klass/static loads, we expect the _type to be precise
1822 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
1823 /* With mirrors being an indirect in the Klass*
1824 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1825 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1826 *
1827 * So check the type and klass of the node before the LoadP.
1828 */
1829 Node* adr2 = adr->in(MemNode::Address);
1830 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1831 if (tkls != NULL && !StressReflectiveCode) {
1832 ciKlass* klass = tkls->klass();
1833 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1834 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1835 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1836 return TypeInstPtr::make(klass->java_mirror());
1837 }
1838 }
1839 }
1840
1841 const TypeKlassPtr *tkls = tp->isa_klassptr();
1842 if (tkls != NULL && !StressReflectiveCode) {
1843 ciKlass* klass = tkls->klass();
1844 if (klass->is_loaded() && tkls->klass_is_exact()) {
1845 // We are loading a field from a Klass metaobject whose identity
1846 // is known at compile time (the type is "exact" or "precise").
1847 // Check for fields we know are maintained as constants by the VM.
1848 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1849 // The field is Klass::_super_check_offset. Return its (constant) value.
1850 // (Folds up type checking code.)
1851 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1852 return TypeInt::make(klass->super_check_offset());
1853 }
1854 // Compute index into primary_supers array
1855 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1856 // Check for overflowing; use unsigned compare to handle the negative case.
1857 if( depth < ciKlass::primary_super_limit() ) {
1858 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1859 // (Folds up type checking code.)
1860 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1861 ciKlass *ss = klass->super_of_depth(depth);
1862 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1863 }
1864 const Type* aift = load_array_final_field(tkls, klass);
1865 if (aift != NULL) return aift;
1866 }
1867
1868 // We can still check if we are loading from the primary_supers array at a
1869 // shallow enough depth. Even though the klass is not exact, entries less
1870 // than or equal to its super depth are correct.
1871 if (klass->is_loaded() ) {
1872 ciType *inner = klass;
1873 while( inner->is_obj_array_klass() )
1874 inner = inner->as_obj_array_klass()->base_element_type();
1875 if( inner->is_instance_klass() &&
1876 !inner->as_instance_klass()->flags().is_interface() ) {
1877 // Compute index into primary_supers array
1878 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1879 // Check for overflowing; use unsigned compare to handle the negative case.
1880 if( depth < ciKlass::primary_super_limit() &&
1881 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1882 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1883 // (Folds up type checking code.)
1884 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1885 ciKlass *ss = klass->super_of_depth(depth);
1886 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1887 }
1888 }
1889 }
1890
1891 // If the type is enough to determine that the thing is not an array,
1892 // we can give the layout_helper a positive interval type.
1893 // This will help short-circuit some reflective code.
1894 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1895 && !klass->is_array_klass() // not directly typed as an array
1896 && !klass->is_interface() // specifically not Serializable & Cloneable
1897 && !klass->is_java_lang_Object() // not the supertype of all T[]
1898 ) {
1899 // Note: When interfaces are reliable, we can narrow the interface
1900 // test to (klass != Serializable && klass != Cloneable).
1901 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1902 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1903 // The key property of this type is that it folds up tests
1904 // for array-ness, since it proves that the layout_helper is positive.
1905 // Thus, a generic value like the basic object layout helper works fine.
1906 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1907 }
1908 }
1909
1910 // If we are loading from a freshly-allocated object, produce a zero,
1911 // if the load is provably beyond the header of the object.
1912 // (Also allow a variable load from a fresh array to produce zero.)
1913 const TypeOopPtr *tinst = tp->isa_oopptr();
1914 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1915 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1916 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1917 Node* value = can_see_stored_value(mem,phase);
1918 if (value != NULL && value->is_Con()) {
1919 assert(value->bottom_type()->higher_equal(_type),"sanity");
1920 return value->bottom_type();
1921 }
1922 }
1923
1924 if (is_instance) {
1925 // If we have an instance type and our memory input is the
1926 // programs's initial memory state, there is no matching store,
1927 // so just return a zero of the appropriate type
1928 Node *mem = in(MemNode::Memory);
1929 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1930 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1931 return Type::get_zero_type(_type->basic_type());
1932 }
1933 }
1934 return _type;
1935 }
1936
1937 //------------------------------match_edge-------------------------------------
1938 // Do we Match on this edge index or not? Match only the address.
1939 uint LoadNode::match_edge(uint idx) const {
1940 return idx == MemNode::Address;
1941 }
1942
1943 //--------------------------LoadBNode::Ideal--------------------------------------
1944 //
1945 // If the previous store is to the same address as this load,
1946 // and the value stored was larger than a byte, replace this load
1947 // with the value stored truncated to a byte. If no truncation is
1948 // needed, the replacement is done in LoadNode::Identity().
1949 //
1950 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1951 Node* mem = in(MemNode::Memory);
1952 Node* value = can_see_stored_value(mem,phase);
1953 if( value && !phase->type(value)->higher_equal( _type ) ) {
1954 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1955 return new RShiftINode(result, phase->intcon(24));
1956 }
1957 // Identity call will handle the case where truncation is not needed.
1958 return LoadNode::Ideal(phase, can_reshape);
1959 }
1960
1961 const Type* LoadBNode::Value(PhaseGVN* phase) const {
1962 Node* mem = in(MemNode::Memory);
1963 Node* value = can_see_stored_value(mem,phase);
1964 if (value != NULL && value->is_Con() &&
1965 !value->bottom_type()->higher_equal(_type)) {
1966 // If the input to the store does not fit with the load's result type,
1967 // it must be truncated. We can't delay until Ideal call since
1968 // a singleton Value is needed for split_thru_phi optimization.
1969 int con = value->get_int();
1970 return TypeInt::make((con << 24) >> 24);
1971 }
1972 return LoadNode::Value(phase);
1973 }
1974
1975 //--------------------------LoadUBNode::Ideal-------------------------------------
1976 //
1977 // If the previous store is to the same address as this load,
1978 // and the value stored was larger than a byte, replace this load
1979 // with the value stored truncated to a byte. If no truncation is
1980 // needed, the replacement is done in LoadNode::Identity().
1981 //
1982 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1983 Node* mem = in(MemNode::Memory);
1984 Node* value = can_see_stored_value(mem, phase);
1985 if (value && !phase->type(value)->higher_equal(_type))
1986 return new AndINode(value, phase->intcon(0xFF));
1987 // Identity call will handle the case where truncation is not needed.
1988 return LoadNode::Ideal(phase, can_reshape);
1989 }
1990
1991 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
1992 Node* mem = in(MemNode::Memory);
1993 Node* value = can_see_stored_value(mem,phase);
1994 if (value != NULL && value->is_Con() &&
1995 !value->bottom_type()->higher_equal(_type)) {
1996 // If the input to the store does not fit with the load's result type,
1997 // it must be truncated. We can't delay until Ideal call since
1998 // a singleton Value is needed for split_thru_phi optimization.
1999 int con = value->get_int();
2000 return TypeInt::make(con & 0xFF);
2001 }
2002 return LoadNode::Value(phase);
2003 }
2004
2005 //--------------------------LoadUSNode::Ideal-------------------------------------
2006 //
2007 // If the previous store is to the same address as this load,
2008 // and the value stored was larger than a char, replace this load
2009 // with the value stored truncated to a char. If no truncation is
2010 // needed, the replacement is done in LoadNode::Identity().
2011 //
2012 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2013 Node* mem = in(MemNode::Memory);
2014 Node* value = can_see_stored_value(mem,phase);
2015 if( value && !phase->type(value)->higher_equal( _type ) )
2016 return new AndINode(value,phase->intcon(0xFFFF));
2017 // Identity call will handle the case where truncation is not needed.
2018 return LoadNode::Ideal(phase, can_reshape);
2019 }
2020
2021 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2022 Node* mem = in(MemNode::Memory);
2023 Node* value = can_see_stored_value(mem,phase);
2024 if (value != NULL && value->is_Con() &&
2025 !value->bottom_type()->higher_equal(_type)) {
2026 // If the input to the store does not fit with the load's result type,
2027 // it must be truncated. We can't delay until Ideal call since
2028 // a singleton Value is needed for split_thru_phi optimization.
2029 int con = value->get_int();
2030 return TypeInt::make(con & 0xFFFF);
2031 }
2032 return LoadNode::Value(phase);
2033 }
2034
2035 //--------------------------LoadSNode::Ideal--------------------------------------
2036 //
2037 // If the previous store is to the same address as this load,
2038 // and the value stored was larger than a short, replace this load
2039 // with the value stored truncated to a short. If no truncation is
2040 // needed, the replacement is done in LoadNode::Identity().
2041 //
2042 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2043 Node* mem = in(MemNode::Memory);
2044 Node* value = can_see_stored_value(mem,phase);
2045 if( value && !phase->type(value)->higher_equal( _type ) ) {
2046 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2047 return new RShiftINode(result, phase->intcon(16));
2048 }
2049 // Identity call will handle the case where truncation is not needed.
2050 return LoadNode::Ideal(phase, can_reshape);
2051 }
2052
2053 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2054 Node* mem = in(MemNode::Memory);
2055 Node* value = can_see_stored_value(mem,phase);
2056 if (value != NULL && value->is_Con() &&
2057 !value->bottom_type()->higher_equal(_type)) {
2058 // If the input to the store does not fit with the load's result type,
2059 // it must be truncated. We can't delay until Ideal call since
2060 // a singleton Value is needed for split_thru_phi optimization.
2061 int con = value->get_int();
2062 return TypeInt::make((con << 16) >> 16);
2063 }
2064 return LoadNode::Value(phase);
2065 }
2066
2067 //=============================================================================
2068 //----------------------------LoadKlassNode::make------------------------------
2069 // Polymorphic factory method:
2070 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2071 // sanity check the alias category against the created node type
2072 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2073 assert(adr_type != NULL, "expecting TypeKlassPtr");
2074 #ifdef _LP64
2075 if (adr_type->is_ptr_to_narrowklass()) {
2076 assert(UseCompressedClassPointers, "no compressed klasses");
2077 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2078 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2079 }
2080 #endif
2081 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2082 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2083 }
2084
2085 //------------------------------Value------------------------------------------
2086 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2087 return klass_value_common(phase);
2088 }
2089
2090 // In most cases, LoadKlassNode does not have the control input set. If the control
2091 // input is set, it must not be removed (by LoadNode::Ideal()).
2092 bool LoadKlassNode::can_remove_control() const {
2093 return false;
2094 }
2095
2096 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2097 // Either input is TOP ==> the result is TOP
2098 const Type *t1 = phase->type( in(MemNode::Memory) );
2099 if (t1 == Type::TOP) return Type::TOP;
2100 Node *adr = in(MemNode::Address);
2101 const Type *t2 = phase->type( adr );
2102 if (t2 == Type::TOP) return Type::TOP;
2103 const TypePtr *tp = t2->is_ptr();
2104 if (TypePtr::above_centerline(tp->ptr()) ||
2105 tp->ptr() == TypePtr::Null) return Type::TOP;
2106
2107 // Return a more precise klass, if possible
2108 const TypeInstPtr *tinst = tp->isa_instptr();
2109 if (tinst != NULL) {
2110 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2111 int offset = tinst->offset();
2112 if (ik == phase->C->env()->Class_klass()
2113 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2114 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2115 // We are loading a special hidden field from a Class mirror object,
2116 // the field which points to the VM's Klass metaobject.
2117 ciType* t = tinst->java_mirror_type();
2118 // java_mirror_type returns non-null for compile-time Class constants.
2119 if (t != NULL) {
2120 // constant oop => constant klass
2121 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2122 if (t->is_void()) {
2123 // We cannot create a void array. Since void is a primitive type return null
2124 // klass. Users of this result need to do a null check on the returned klass.
2125 return TypePtr::NULL_PTR;
2126 }
2127 return TypeKlassPtr::make(ciArrayKlass::make(t));
2128 }
2129 if (!t->is_klass()) {
2130 // a primitive Class (e.g., int.class) has NULL for a klass field
2131 return TypePtr::NULL_PTR;
2132 }
2133 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2134 return TypeKlassPtr::make(t->as_klass());
2135 }
2136 // non-constant mirror, so we can't tell what's going on
2137 }
2138 if( !ik->is_loaded() )
2139 return _type; // Bail out if not loaded
2140 if (offset == oopDesc::klass_offset_in_bytes()) {
2141 if (tinst->klass_is_exact()) {
2142 return TypeKlassPtr::make(ik);
2143 }
2144 // See if we can become precise: no subklasses and no interface
2145 // (Note: We need to support verified interfaces.)
2146 if (!ik->is_interface() && !ik->has_subklass()) {
2147 //assert(!UseExactTypes, "this code should be useless with exact types");
2148 // Add a dependence; if any subclass added we need to recompile
2149 if (!ik->is_final()) {
2150 // %%% should use stronger assert_unique_concrete_subtype instead
2151 phase->C->dependencies()->assert_leaf_type(ik);
2152 }
2153 // Return precise klass
2154 return TypeKlassPtr::make(ik);
2155 }
2156
2157 // Return root of possible klass
2158 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2159 }
2160 }
2161
2162 // Check for loading klass from an array
2163 const TypeAryPtr *tary = tp->isa_aryptr();
2164 if( tary != NULL ) {
2165 ciKlass *tary_klass = tary->klass();
2166 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2167 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2168 if (tary->klass_is_exact()) {
2169 return TypeKlassPtr::make(tary_klass);
2170 }
2171 ciArrayKlass *ak = tary->klass()->as_array_klass();
2172 // If the klass is an object array, we defer the question to the
2173 // array component klass.
2174 if( ak->is_obj_array_klass() ) {
2175 assert( ak->is_loaded(), "" );
2176 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2177 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2178 ciInstanceKlass* ik = base_k->as_instance_klass();
2179 // See if we can become precise: no subklasses and no interface
2180 if (!ik->is_interface() && !ik->has_subklass()) {
2181 //assert(!UseExactTypes, "this code should be useless with exact types");
2182 // Add a dependence; if any subclass added we need to recompile
2183 if (!ik->is_final()) {
2184 phase->C->dependencies()->assert_leaf_type(ik);
2185 }
2186 // Return precise array klass
2187 return TypeKlassPtr::make(ak);
2188 }
2189 }
2190 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2191 } else { // Found a type-array?
2192 //assert(!UseExactTypes, "this code should be useless with exact types");
2193 assert( ak->is_type_array_klass(), "" );
2194 return TypeKlassPtr::make(ak); // These are always precise
2195 }
2196 }
2197 }
2198
2199 // Check for loading klass from an array klass
2200 const TypeKlassPtr *tkls = tp->isa_klassptr();
2201 if (tkls != NULL && !StressReflectiveCode) {
2202 ciKlass* klass = tkls->klass();
2203 if( !klass->is_loaded() )
2204 return _type; // Bail out if not loaded
2205 if( klass->is_obj_array_klass() &&
2206 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2207 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2208 // // Always returning precise element type is incorrect,
2209 // // e.g., element type could be object and array may contain strings
2210 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2211
2212 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2213 // according to the element type's subclassing.
2214 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2215 }
2216 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2217 tkls->offset() == in_bytes(Klass::super_offset())) {
2218 ciKlass* sup = klass->as_instance_klass()->super();
2219 // The field is Klass::_super. Return its (constant) value.
2220 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2221 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2222 }
2223 }
2224
2225 // Bailout case
2226 return LoadNode::Value(phase);
2227 }
2228
2229 //------------------------------Identity---------------------------------------
2230 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2231 // Also feed through the klass in Allocate(...klass...)._klass.
2232 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2233 return klass_identity_common(phase);
2234 }
2235
2236 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2237 Node* x = LoadNode::Identity(phase);
2238 if (x != this) return x;
2239
2240 // Take apart the address into an oop and and offset.
2241 // Return 'this' if we cannot.
2242 Node* adr = in(MemNode::Address);
2243 intptr_t offset = 0;
2244 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2245 if (base == NULL) return this;
2246 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2247 if (toop == NULL) return this;
2248
2249 // Step over potential GC barrier for OopHandle resolve
2250 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2251 if (bs->is_gc_barrier_node(base)) {
2252 base = bs->step_over_gc_barrier(base);
2253 }
2254
2255 // We can fetch the klass directly through an AllocateNode.
2256 // This works even if the klass is not constant (clone or newArray).
2257 if (offset == oopDesc::klass_offset_in_bytes()) {
2258 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2259 if (allocated_klass != NULL) {
2260 return allocated_klass;
2261 }
2262 }
2263
2264 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2265 // See inline_native_Class_query for occurrences of these patterns.
2266 // Java Example: x.getClass().isAssignableFrom(y)
2267 //
2268 // This improves reflective code, often making the Class
2269 // mirror go completely dead. (Current exception: Class
2270 // mirrors may appear in debug info, but we could clean them out by
2271 // introducing a new debug info operator for Klass.java_mirror).
2272
2273 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2274 && offset == java_lang_Class::klass_offset_in_bytes()) {
2275 if (base->is_Load()) {
2276 Node* base2 = base->in(MemNode::Address);
2277 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2278 Node* adr2 = base2->in(MemNode::Address);
2279 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2280 if (tkls != NULL && !tkls->empty()
2281 && (tkls->klass()->is_instance_klass() ||
2282 tkls->klass()->is_array_klass())
2283 && adr2->is_AddP()
2284 ) {
2285 int mirror_field = in_bytes(Klass::java_mirror_offset());
2286 if (tkls->offset() == mirror_field) {
2287 return adr2->in(AddPNode::Base);
2288 }
2289 }
2290 }
2291 }
2292 }
2293
2294 return this;
2295 }
2296
2297
2298 //------------------------------Value------------------------------------------
2299 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2300 const Type *t = klass_value_common(phase);
2301 if (t == Type::TOP)
2302 return t;
2303
2304 return t->make_narrowklass();
2305 }
2306
2307 //------------------------------Identity---------------------------------------
2308 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2309 // Also feed through the klass in Allocate(...klass...)._klass.
2310 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2311 Node *x = klass_identity_common(phase);
2312
2313 const Type *t = phase->type( x );
2314 if( t == Type::TOP ) return x;
2315 if( t->isa_narrowklass()) return x;
2316 assert (!t->isa_narrowoop(), "no narrow oop here");
2317
2318 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2319 }
2320
2321 //------------------------------Value-----------------------------------------
2322 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2323 // Either input is TOP ==> the result is TOP
2324 const Type *t1 = phase->type( in(MemNode::Memory) );
2325 if( t1 == Type::TOP ) return Type::TOP;
2326 Node *adr = in(MemNode::Address);
2327 const Type *t2 = phase->type( adr );
2328 if( t2 == Type::TOP ) return Type::TOP;
2329 const TypePtr *tp = t2->is_ptr();
2330 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2331 const TypeAryPtr *tap = tp->isa_aryptr();
2332 if( !tap ) return _type;
2333 return tap->size();
2334 }
2335
2336 //-------------------------------Ideal---------------------------------------
2337 // Feed through the length in AllocateArray(...length...)._length.
2338 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2339 Node* p = MemNode::Ideal_common(phase, can_reshape);
2340 if (p) return (p == NodeSentinel) ? NULL : p;
2341
2342 // Take apart the address into an oop and and offset.
2343 // Return 'this' if we cannot.
2344 Node* adr = in(MemNode::Address);
2345 intptr_t offset = 0;
2346 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2347 if (base == NULL) return NULL;
2348 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2349 if (tary == NULL) return NULL;
2350
2351 // We can fetch the length directly through an AllocateArrayNode.
2352 // This works even if the length is not constant (clone or newArray).
2353 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2354 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2355 if (alloc != NULL) {
2356 Node* allocated_length = alloc->Ideal_length();
2357 Node* len = alloc->make_ideal_length(tary, phase);
2358 if (allocated_length != len) {
2359 // New CastII improves on this.
2360 return len;
2361 }
2362 }
2363 }
2364
2365 return NULL;
2366 }
2367
2368 //------------------------------Identity---------------------------------------
2369 // Feed through the length in AllocateArray(...length...)._length.
2370 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2371 Node* x = LoadINode::Identity(phase);
2372 if (x != this) return x;
2373
2374 // Take apart the address into an oop and and offset.
2375 // Return 'this' if we cannot.
2376 Node* adr = in(MemNode::Address);
2377 intptr_t offset = 0;
2378 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2379 if (base == NULL) return this;
2380 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2381 if (tary == NULL) return this;
2382
2383 // We can fetch the length directly through an AllocateArrayNode.
2384 // This works even if the length is not constant (clone or newArray).
2385 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2386 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2387 if (alloc != NULL) {
2388 Node* allocated_length = alloc->Ideal_length();
2389 // Do not allow make_ideal_length to allocate a CastII node.
2390 Node* len = alloc->make_ideal_length(tary, phase, false);
2391 if (allocated_length == len) {
2392 // Return allocated_length only if it would not be improved by a CastII.
2393 return allocated_length;
2394 }
2395 }
2396 }
2397
2398 return this;
2399
2400 }
2401
2402 //=============================================================================
2403 //---------------------------StoreNode::make-----------------------------------
2404 // Polymorphic factory method:
2405 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2406 assert((mo == unordered || mo == release), "unexpected");
2407 Compile* C = gvn.C;
2408 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2409 ctl != NULL, "raw memory operations should have control edge");
2410
2411 switch (bt) {
2412 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2413 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2414 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2415 case T_CHAR:
2416 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2417 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2418 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2419 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2420 case T_METADATA:
2421 case T_ADDRESS:
2422 case T_OBJECT:
2423 #ifdef _LP64
2424 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2425 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2426 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2427 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2428 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2429 adr->bottom_type()->isa_rawptr())) {
2430 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2431 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2432 }
2433 #endif
2434 {
2435 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2436 }
2437 default:
2438 ShouldNotReachHere();
2439 return (StoreNode*)NULL;
2440 }
2441 }
2442
2443 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2444 bool require_atomic = true;
2445 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2446 }
2447
2448 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2449 bool require_atomic = true;
2450 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2451 }
2452
2453
2454 //--------------------------bottom_type----------------------------------------
2455 const Type *StoreNode::bottom_type() const {
2456 return Type::MEMORY;
2457 }
2458
2459 //------------------------------hash-------------------------------------------
2460 uint StoreNode::hash() const {
2461 // unroll addition of interesting fields
2462 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2463
2464 // Since they are not commoned, do not hash them:
2465 return NO_HASH;
2466 }
2467
2468 //------------------------------Ideal------------------------------------------
2469 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2470 // When a store immediately follows a relevant allocation/initialization,
2471 // try to capture it into the initialization, or hoist it above.
2472 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2473 Node* p = MemNode::Ideal_common(phase, can_reshape);
2474 if (p) return (p == NodeSentinel) ? NULL : p;
2475
2476 Node* mem = in(MemNode::Memory);
2477 Node* address = in(MemNode::Address);
2478 // Back-to-back stores to same address? Fold em up. Generally
2479 // unsafe if I have intervening uses... Also disallowed for StoreCM
2480 // since they must follow each StoreP operation. Redundant StoreCMs
2481 // are eliminated just before matching in final_graph_reshape.
2482 {
2483 Node* st = mem;
2484 // If Store 'st' has more than one use, we cannot fold 'st' away.
2485 // For example, 'st' might be the final state at a conditional
2486 // return. Or, 'st' might be used by some node which is live at
2487 // the same time 'st' is live, which might be unschedulable. So,
2488 // require exactly ONE user until such time as we clone 'mem' for
2489 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2490 // true).
2491 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2492 // Looking at a dead closed cycle of memory?
2493 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2494 assert(Opcode() == st->Opcode() ||
2495 st->Opcode() == Op_StoreVector ||
2496 Opcode() == Op_StoreVector ||
2497 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2498 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2499 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2500 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2501 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2502
2503 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2504 st->as_Store()->memory_size() <= this->memory_size()) {
2505 Node* use = st->raw_out(0);
2506 phase->igvn_rehash_node_delayed(use);
2507 if (can_reshape) {
2508 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2509 } else {
2510 // It's OK to do this in the parser, since DU info is always accurate,
2511 // and the parser always refers to nodes via SafePointNode maps.
2512 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2513 }
2514 return this;
2515 }
2516 st = st->in(MemNode::Memory);
2517 }
2518 }
2519
2520
2521 // Capture an unaliased, unconditional, simple store into an initializer.
2522 // Or, if it is independent of the allocation, hoist it above the allocation.
2523 if (ReduceFieldZeroing && /*can_reshape &&*/
2524 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2525 InitializeNode* init = mem->in(0)->as_Initialize();
2526 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2527 if (offset > 0) {
2528 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2529 // If the InitializeNode captured me, it made a raw copy of me,
2530 // and I need to disappear.
2531 if (moved != NULL) {
2532 // %%% hack to ensure that Ideal returns a new node:
2533 mem = MergeMemNode::make(mem);
2534 return mem; // fold me away
2535 }
2536 }
2537 }
2538
2539 return NULL; // No further progress
2540 }
2541
2542 //------------------------------Value-----------------------------------------
2543 const Type* StoreNode::Value(PhaseGVN* phase) const {
2544 // Either input is TOP ==> the result is TOP
2545 const Type *t1 = phase->type( in(MemNode::Memory) );
2546 if( t1 == Type::TOP ) return Type::TOP;
2547 const Type *t2 = phase->type( in(MemNode::Address) );
2548 if( t2 == Type::TOP ) return Type::TOP;
2549 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2550 if( t3 == Type::TOP ) return Type::TOP;
2551 return Type::MEMORY;
2552 }
2553
2554 //------------------------------Identity---------------------------------------
2555 // Remove redundant stores:
2556 // Store(m, p, Load(m, p)) changes to m.
2557 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2558 Node* StoreNode::Identity(PhaseGVN* phase) {
2559 Node* mem = in(MemNode::Memory);
2560 Node* adr = in(MemNode::Address);
2561 Node* val = in(MemNode::ValueIn);
2562
2563 Node* result = this;
2564
2565 // Load then Store? Then the Store is useless
2566 if (val->is_Load() &&
2567 val->in(MemNode::Address)->eqv_uncast(adr) &&
2568 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2569 val->as_Load()->store_Opcode() == Opcode()) {
2570 result = mem;
2571 }
2572
2573 // Two stores in a row of the same value?
2574 if (result == this &&
2575 mem->is_Store() &&
2576 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2577 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2578 mem->Opcode() == Opcode()) {
2579 result = mem;
2580 }
2581
2582 // Store of zero anywhere into a freshly-allocated object?
2583 // Then the store is useless.
2584 // (It must already have been captured by the InitializeNode.)
2585 if (result == this &&
2586 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2587 // a newly allocated object is already all-zeroes everywhere
2588 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2589 result = mem;
2590 }
2591
2592 if (result == this) {
2593 // the store may also apply to zero-bits in an earlier object
2594 Node* prev_mem = find_previous_store(phase);
2595 // Steps (a), (b): Walk past independent stores to find an exact match.
2596 if (prev_mem != NULL) {
2597 Node* prev_val = can_see_stored_value(prev_mem, phase);
2598 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2599 // prev_val and val might differ by a cast; it would be good
2600 // to keep the more informative of the two.
2601 result = mem;
2602 }
2603 }
2604 }
2605 }
2606
2607 if (result != this && phase->is_IterGVN() != NULL) {
2608 MemBarNode* trailing = trailing_membar();
2609 if (trailing != NULL) {
2610 #ifdef ASSERT
2611 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2612 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2613 #endif
2614 PhaseIterGVN* igvn = phase->is_IterGVN();
2615 trailing->remove(igvn);
2616 }
2617 }
2618
2619 return result;
2620 }
2621
2622 //------------------------------match_edge-------------------------------------
2623 // Do we Match on this edge index or not? Match only memory & value
2624 uint StoreNode::match_edge(uint idx) const {
2625 return idx == MemNode::Address || idx == MemNode::ValueIn;
2626 }
2627
2628 //------------------------------cmp--------------------------------------------
2629 // Do not common stores up together. They generally have to be split
2630 // back up anyways, so do not bother.
2631 bool StoreNode::cmp( const Node &n ) const {
2632 return (&n == this); // Always fail except on self
2633 }
2634
2635 //------------------------------Ideal_masked_input-----------------------------
2636 // Check for a useless mask before a partial-word store
2637 // (StoreB ... (AndI valIn conIa) )
2638 // If (conIa & mask == mask) this simplifies to
2639 // (StoreB ... (valIn) )
2640 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2641 Node *val = in(MemNode::ValueIn);
2642 if( val->Opcode() == Op_AndI ) {
2643 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2644 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2645 set_req(MemNode::ValueIn, val->in(1));
2646 return this;
2647 }
2648 }
2649 return NULL;
2650 }
2651
2652
2653 //------------------------------Ideal_sign_extended_input----------------------
2654 // Check for useless sign-extension before a partial-word store
2655 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2656 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2657 // (StoreB ... (valIn) )
2658 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2659 Node *val = in(MemNode::ValueIn);
2660 if( val->Opcode() == Op_RShiftI ) {
2661 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2662 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2663 Node *shl = val->in(1);
2664 if( shl->Opcode() == Op_LShiftI ) {
2665 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2666 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2667 set_req(MemNode::ValueIn, shl->in(1));
2668 return this;
2669 }
2670 }
2671 }
2672 }
2673 return NULL;
2674 }
2675
2676 //------------------------------value_never_loaded-----------------------------------
2677 // Determine whether there are any possible loads of the value stored.
2678 // For simplicity, we actually check if there are any loads from the
2679 // address stored to, not just for loads of the value stored by this node.
2680 //
2681 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2682 Node *adr = in(Address);
2683 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2684 if (adr_oop == NULL)
2685 return false;
2686 if (!adr_oop->is_known_instance_field())
2687 return false; // if not a distinct instance, there may be aliases of the address
2688 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2689 Node *use = adr->fast_out(i);
2690 if (use->is_Load() || use->is_LoadStore()) {
2691 return false;
2692 }
2693 }
2694 return true;
2695 }
2696
2697 MemBarNode* StoreNode::trailing_membar() const {
2698 if (is_release()) {
2699 MemBarNode* trailing_mb = NULL;
2700 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2701 Node* u = fast_out(i);
2702 if (u->is_MemBar()) {
2703 if (u->as_MemBar()->trailing_store()) {
2704 assert(u->Opcode() == Op_MemBarVolatile, "");
2705 assert(trailing_mb == NULL, "only one");
2706 trailing_mb = u->as_MemBar();
2707 #ifdef ASSERT
2708 Node* leading = u->as_MemBar()->leading_membar();
2709 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2710 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2711 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2712 #endif
2713 } else {
2714 assert(u->as_MemBar()->standalone(), "");
2715 }
2716 }
2717 }
2718 return trailing_mb;
2719 }
2720 return NULL;
2721 }
2722
2723
2724 //=============================================================================
2725 //------------------------------Ideal------------------------------------------
2726 // If the store is from an AND mask that leaves the low bits untouched, then
2727 // we can skip the AND operation. If the store is from a sign-extension
2728 // (a left shift, then right shift) we can skip both.
2729 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2730 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2731 if( progress != NULL ) return progress;
2732
2733 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2734 if( progress != NULL ) return progress;
2735
2736 // Finally check the default case
2737 return StoreNode::Ideal(phase, can_reshape);
2738 }
2739
2740 //=============================================================================
2741 //------------------------------Ideal------------------------------------------
2742 // If the store is from an AND mask that leaves the low bits untouched, then
2743 // we can skip the AND operation
2744 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2745 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2746 if( progress != NULL ) return progress;
2747
2748 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2749 if( progress != NULL ) return progress;
2750
2751 // Finally check the default case
2752 return StoreNode::Ideal(phase, can_reshape);
2753 }
2754
2755 //=============================================================================
2756 //------------------------------Identity---------------------------------------
2757 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2758 // No need to card mark when storing a null ptr
2759 Node* my_store = in(MemNode::OopStore);
2760 if (my_store->is_Store()) {
2761 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2762 if( t1 == TypePtr::NULL_PTR ) {
2763 return in(MemNode::Memory);
2764 }
2765 }
2766 return this;
2767 }
2768
2769 //=============================================================================
2770 //------------------------------Ideal---------------------------------------
2771 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2772 Node* progress = StoreNode::Ideal(phase, can_reshape);
2773 if (progress != NULL) return progress;
2774
2775 Node* my_store = in(MemNode::OopStore);
2776 if (my_store->is_MergeMem()) {
2777 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2778 set_req(MemNode::OopStore, mem);
2779 return this;
2780 }
2781
2782 return NULL;
2783 }
2784
2785 //------------------------------Value-----------------------------------------
2786 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2787 // Either input is TOP ==> the result is TOP
2788 const Type *t = phase->type( in(MemNode::Memory) );
2789 if( t == Type::TOP ) return Type::TOP;
2790 t = phase->type( in(MemNode::Address) );
2791 if( t == Type::TOP ) return Type::TOP;
2792 t = phase->type( in(MemNode::ValueIn) );
2793 if( t == Type::TOP ) return Type::TOP;
2794 // If extra input is TOP ==> the result is TOP
2795 t = phase->type( in(MemNode::OopStore) );
2796 if( t == Type::TOP ) return Type::TOP;
2797
2798 return StoreNode::Value( phase );
2799 }
2800
2801
2802 //=============================================================================
2803 //----------------------------------SCMemProjNode------------------------------
2804 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2805 {
2806 return bottom_type();
2807 }
2808
2809 //=============================================================================
2810 //----------------------------------LoadStoreNode------------------------------
2811 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2812 : Node(required),
2813 _type(rt),
2814 _adr_type(at)
2815 {
2816 init_req(MemNode::Control, c );
2817 init_req(MemNode::Memory , mem);
2818 init_req(MemNode::Address, adr);
2819 init_req(MemNode::ValueIn, val);
2820 init_class_id(Class_LoadStore);
2821 }
2822
2823 uint LoadStoreNode::ideal_reg() const {
2824 return _type->ideal_reg();
2825 }
2826
2827 bool LoadStoreNode::result_not_used() const {
2828 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2829 Node *x = fast_out(i);
2830 if (x->Opcode() == Op_SCMemProj) continue;
2831 return false;
2832 }
2833 return true;
2834 }
2835
2836 MemBarNode* LoadStoreNode::trailing_membar() const {
2837 MemBarNode* trailing = NULL;
2838 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2839 Node* u = fast_out(i);
2840 if (u->is_MemBar()) {
2841 if (u->as_MemBar()->trailing_load_store()) {
2842 assert(u->Opcode() == Op_MemBarAcquire, "");
2843 assert(trailing == NULL, "only one");
2844 trailing = u->as_MemBar();
2845 #ifdef ASSERT
2846 Node* leading = trailing->leading_membar();
2847 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2848 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
2849 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
2850 #endif
2851 } else {
2852 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
2853 }
2854 }
2855 }
2856
2857 return trailing;
2858 }
2859
2860 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2861
2862 //=============================================================================
2863 //----------------------------------LoadStoreConditionalNode--------------------
2864 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2865 init_req(ExpectedIn, ex );
2866 }
2867
2868 //=============================================================================
2869 //-------------------------------adr_type--------------------------------------
2870 const TypePtr* ClearArrayNode::adr_type() const {
2871 Node *adr = in(3);
2872 if (adr == NULL) return NULL; // node is dead
2873 return MemNode::calculate_adr_type(adr->bottom_type());
2874 }
2875
2876 //------------------------------match_edge-------------------------------------
2877 // Do we Match on this edge index or not? Do not match memory
2878 uint ClearArrayNode::match_edge(uint idx) const {
2879 return idx > 1;
2880 }
2881
2882 //------------------------------Identity---------------------------------------
2883 // Clearing a zero length array does nothing
2884 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2885 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2886 }
2887
2888 //------------------------------Idealize---------------------------------------
2889 // Clearing a short array is faster with stores
2890 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2891 // Already know this is a large node, do not try to ideal it
2892 if (!IdealizeClearArrayNode || _is_large) return NULL;
2893
2894 const int unit = BytesPerLong;
2895 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2896 if (!t) return NULL;
2897 if (!t->is_con()) return NULL;
2898 intptr_t raw_count = t->get_con();
2899 intptr_t size = raw_count;
2900 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2901 // Clearing nothing uses the Identity call.
2902 // Negative clears are possible on dead ClearArrays
2903 // (see jck test stmt114.stmt11402.val).
2904 if (size <= 0 || size % unit != 0) return NULL;
2905 intptr_t count = size / unit;
2906 // Length too long; communicate this to matchers and assemblers.
2907 // Assemblers are responsible to produce fast hardware clears for it.
2908 if (size > InitArrayShortSize) {
2909 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
2910 }
2911 Node *mem = in(1);
2912 if( phase->type(mem)==Type::TOP ) return NULL;
2913 Node *adr = in(3);
2914 const Type* at = phase->type(adr);
2915 if( at==Type::TOP ) return NULL;
2916 const TypePtr* atp = at->isa_ptr();
2917 // adjust atp to be the correct array element address type
2918 if (atp == NULL) atp = TypePtr::BOTTOM;
2919 else atp = atp->add_offset(Type::OffsetBot);
2920 // Get base for derived pointer purposes
2921 if( adr->Opcode() != Op_AddP ) Unimplemented();
2922 Node *base = adr->in(1);
2923
2924 Node *zero = phase->makecon(TypeLong::ZERO);
2925 Node *off = phase->MakeConX(BytesPerLong);
2926 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2927 count--;
2928 while( count-- ) {
2929 mem = phase->transform(mem);
2930 adr = phase->transform(new AddPNode(base,adr,off));
2931 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2932 }
2933 return mem;
2934 }
2935
2936 //----------------------------step_through----------------------------------
2937 // Return allocation input memory edge if it is different instance
2938 // or itself if it is the one we are looking for.
2939 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2940 Node* n = *np;
2941 assert(n->is_ClearArray(), "sanity");
2942 intptr_t offset;
2943 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2944 // This method is called only before Allocate nodes are expanded
2945 // during macro nodes expansion. Before that ClearArray nodes are
2946 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2947 // Allocate nodes are expanded) which follows allocations.
2948 assert(alloc != NULL, "should have allocation");
2949 if (alloc->_idx == instance_id) {
2950 // Can not bypass initialization of the instance we are looking for.
2951 return false;
2952 }
2953 // Otherwise skip it.
2954 InitializeNode* init = alloc->initialization();
2955 if (init != NULL)
2956 *np = init->in(TypeFunc::Memory);
2957 else
2958 *np = alloc->in(TypeFunc::Memory);
2959 return true;
2960 }
2961
2962 //----------------------------clear_memory-------------------------------------
2963 // Generate code to initialize object storage to zero.
2964 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2965 intptr_t start_offset,
2966 Node* end_offset,
2967 PhaseGVN* phase) {
2968 intptr_t offset = start_offset;
2969
2970 int unit = BytesPerLong;
2971 if ((offset % unit) != 0) {
2972 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
2973 adr = phase->transform(adr);
2974 const TypePtr* atp = TypeRawPtr::BOTTOM;
2975 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2976 mem = phase->transform(mem);
2977 offset += BytesPerInt;
2978 }
2979 assert((offset % unit) == 0, "");
2980
2981 // Initialize the remaining stuff, if any, with a ClearArray.
2982 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2983 }
2984
2985 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2986 Node* start_offset,
2987 Node* end_offset,
2988 PhaseGVN* phase) {
2989 if (start_offset == end_offset) {
2990 // nothing to do
2991 return mem;
2992 }
2993
2994 int unit = BytesPerLong;
2995 Node* zbase = start_offset;
2996 Node* zend = end_offset;
2997
2998 // Scale to the unit required by the CPU:
2999 if (!Matcher::init_array_count_is_in_bytes) {
3000 Node* shift = phase->intcon(exact_log2(unit));
3001 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3002 zend = phase->transform(new URShiftXNode(zend, shift) );
3003 }
3004
3005 // Bulk clear double-words
3006 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3007 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3008 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3009 return phase->transform(mem);
3010 }
3011
3012 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3013 intptr_t start_offset,
3014 intptr_t end_offset,
3015 PhaseGVN* phase) {
3016 if (start_offset == end_offset) {
3017 // nothing to do
3018 return mem;
3019 }
3020
3021 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3022 intptr_t done_offset = end_offset;
3023 if ((done_offset % BytesPerLong) != 0) {
3024 done_offset -= BytesPerInt;
3025 }
3026 if (done_offset > start_offset) {
3027 mem = clear_memory(ctl, mem, dest,
3028 start_offset, phase->MakeConX(done_offset), phase);
3029 }
3030 if (done_offset < end_offset) { // emit the final 32-bit store
3031 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3032 adr = phase->transform(adr);
3033 const TypePtr* atp = TypeRawPtr::BOTTOM;
3034 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3035 mem = phase->transform(mem);
3036 done_offset += BytesPerInt;
3037 }
3038 assert(done_offset == end_offset, "");
3039 return mem;
3040 }
3041
3042 //=============================================================================
3043 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3044 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3045 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3046 #ifdef ASSERT
3047 , _pair_idx(0)
3048 #endif
3049 {
3050 init_class_id(Class_MemBar);
3051 Node* top = C->top();
3052 init_req(TypeFunc::I_O,top);
3053 init_req(TypeFunc::FramePtr,top);
3054 init_req(TypeFunc::ReturnAdr,top);
3055 if (precedent != NULL)
3056 init_req(TypeFunc::Parms, precedent);
3057 }
3058
3059 //------------------------------cmp--------------------------------------------
3060 uint MemBarNode::hash() const { return NO_HASH; }
3061 bool MemBarNode::cmp( const Node &n ) const {
3062 return (&n == this); // Always fail except on self
3063 }
3064
3065 //------------------------------make-------------------------------------------
3066 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3067 switch (opcode) {
3068 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3069 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3070 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3071 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3072 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3073 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3074 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3075 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3076 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3077 case Op_Initialize: return new InitializeNode(C, atp, pn);
3078 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3079 default: ShouldNotReachHere(); return NULL;
3080 }
3081 }
3082
3083 void MemBarNode::remove(PhaseIterGVN *igvn) {
3084 if (outcnt() != 2) {
3085 return;
3086 }
3087 if (trailing_store() || trailing_load_store()) {
3088 MemBarNode* leading = leading_membar();
3089 if (leading != NULL) {
3090 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3091 leading->remove(igvn);
3092 }
3093 }
3094 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3095 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3096 }
3097
3098 //------------------------------Ideal------------------------------------------
3099 // Return a node which is more "ideal" than the current node. Strip out
3100 // control copies
3101 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3102 if (remove_dead_region(phase, can_reshape)) return this;
3103 // Don't bother trying to transform a dead node
3104 if (in(0) && in(0)->is_top()) {
3105 return NULL;
3106 }
3107
3108 #if INCLUDE_ZGC
3109 if (UseZGC) {
3110 if (req() == (Precedent+1) && in(MemBarNode::Precedent)->in(0) != NULL && in(MemBarNode::Precedent)->in(0)->is_LoadBarrier()) {
3111 Node* load_node = in(MemBarNode::Precedent)->in(0)->in(LoadBarrierNode::Oop);
3112 set_req(MemBarNode::Precedent, load_node);
3113 return this;
3114 }
3115 }
3116 #endif
3117
3118 bool progress = false;
3119 // Eliminate volatile MemBars for scalar replaced objects.
3120 if (can_reshape && req() == (Precedent+1)) {
3121 bool eliminate = false;
3122 int opc = Opcode();
3123 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3124 // Volatile field loads and stores.
3125 Node* my_mem = in(MemBarNode::Precedent);
3126 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3127 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3128 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3129 // replace this Precedent (decodeN) with the Load instead.
3130 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3131 Node* load_node = my_mem->in(1);
3132 set_req(MemBarNode::Precedent, load_node);
3133 phase->is_IterGVN()->_worklist.push(my_mem);
3134 my_mem = load_node;
3135 } else {
3136 assert(my_mem->unique_out() == this, "sanity");
3137 del_req(Precedent);
3138 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3139 my_mem = NULL;
3140 }
3141 progress = true;
3142 }
3143 if (my_mem != NULL && my_mem->is_Mem()) {
3144 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3145 // Check for scalar replaced object reference.
3146 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3147 t_oop->offset() != Type::OffsetBot &&
3148 t_oop->offset() != Type::OffsetTop) {
3149 eliminate = true;
3150 }
3151 }
3152 } else if (opc == Op_MemBarRelease) {
3153 // Final field stores.
3154 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3155 if ((alloc != NULL) && alloc->is_Allocate() &&
3156 alloc->as_Allocate()->does_not_escape_thread()) {
3157 // The allocated object does not escape.
3158 eliminate = true;
3159 }
3160 }
3161 if (eliminate) {
3162 // Replace MemBar projections by its inputs.
3163 PhaseIterGVN* igvn = phase->is_IterGVN();
3164 remove(igvn);
3165 // Must return either the original node (now dead) or a new node
3166 // (Do not return a top here, since that would break the uniqueness of top.)
3167 return new ConINode(TypeInt::ZERO);
3168 }
3169 }
3170 return progress ? this : NULL;
3171 }
3172
3173 //------------------------------Value------------------------------------------
3174 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3175 if( !in(0) ) return Type::TOP;
3176 if( phase->type(in(0)) == Type::TOP )
3177 return Type::TOP;
3178 return TypeTuple::MEMBAR;
3179 }
3180
3181 //------------------------------match------------------------------------------
3182 // Construct projections for memory.
3183 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3184 switch (proj->_con) {
3185 case TypeFunc::Control:
3186 case TypeFunc::Memory:
3187 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3188 }
3189 ShouldNotReachHere();
3190 return NULL;
3191 }
3192
3193 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3194 trailing->_kind = TrailingStore;
3195 leading->_kind = LeadingStore;
3196 #ifdef ASSERT
3197 trailing->_pair_idx = leading->_idx;
3198 leading->_pair_idx = leading->_idx;
3199 #endif
3200 }
3201
3202 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3203 trailing->_kind = TrailingLoadStore;
3204 leading->_kind = LeadingLoadStore;
3205 #ifdef ASSERT
3206 trailing->_pair_idx = leading->_idx;
3207 leading->_pair_idx = leading->_idx;
3208 #endif
3209 }
3210
3211 MemBarNode* MemBarNode::trailing_membar() const {
3212 ResourceMark rm;
3213 Node* trailing = (Node*)this;
3214 VectorSet seen(Thread::current()->resource_area());
3215 Node_Stack multis(0);
3216 do {
3217 Node* c = trailing;
3218 uint i = 0;
3219 do {
3220 trailing = NULL;
3221 for (; i < c->outcnt(); i++) {
3222 Node* next = c->raw_out(i);
3223 if (next != c && next->is_CFG()) {
3224 if (c->is_MultiBranch()) {
3225 if (multis.node() == c) {
3226 multis.set_index(i+1);
3227 } else {
3228 multis.push(c, i+1);
3229 }
3230 }
3231 trailing = next;
3232 break;
3233 }
3234 }
3235 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3236 break;
3237 }
3238 while (multis.size() > 0) {
3239 c = multis.node();
3240 i = multis.index();
3241 if (i < c->req()) {
3242 break;
3243 }
3244 multis.pop();
3245 }
3246 } while (multis.size() > 0);
3247 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3248
3249 MemBarNode* mb = trailing->as_MemBar();
3250 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3251 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3252 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3253 return mb;
3254 }
3255
3256 MemBarNode* MemBarNode::leading_membar() const {
3257 ResourceMark rm;
3258 VectorSet seen(Thread::current()->resource_area());
3259 Node_Stack regions(0);
3260 Node* leading = in(0);
3261 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3262 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3263 leading = NULL;
3264 while (regions.size() > 0 && leading == NULL) {
3265 Node* r = regions.node();
3266 uint i = regions.index();
3267 if (i < r->req()) {
3268 leading = r->in(i);
3269 regions.set_index(i+1);
3270 } else {
3271 regions.pop();
3272 }
3273 }
3274 if (leading == NULL) {
3275 assert(regions.size() == 0, "all paths should have been tried");
3276 return NULL;
3277 }
3278 }
3279 if (leading->is_Region()) {
3280 regions.push(leading, 2);
3281 leading = leading->in(1);
3282 } else {
3283 leading = leading->in(0);
3284 }
3285 }
3286 #ifdef ASSERT
3287 Unique_Node_List wq;
3288 wq.push((Node*)this);
3289 uint found = 0;
3290 for (uint i = 0; i < wq.size(); i++) {
3291 Node* n = wq.at(i);
3292 if (n->is_Region()) {
3293 for (uint j = 1; j < n->req(); j++) {
3294 Node* in = n->in(j);
3295 if (in != NULL && !in->is_top()) {
3296 wq.push(in);
3297 }
3298 }
3299 } else {
3300 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3301 assert(n == leading, "consistency check failed");
3302 found++;
3303 } else {
3304 Node* in = n->in(0);
3305 if (in != NULL && !in->is_top()) {
3306 wq.push(in);
3307 }
3308 }
3309 }
3310 }
3311 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3312 #endif
3313 if (leading == NULL) {
3314 return NULL;
3315 }
3316 MemBarNode* mb = leading->as_MemBar();
3317 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3318 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3319 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3320 return mb;
3321 }
3322
3323 //===========================InitializeNode====================================
3324 // SUMMARY:
3325 // This node acts as a memory barrier on raw memory, after some raw stores.
3326 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3327 // The Initialize can 'capture' suitably constrained stores as raw inits.
3328 // It can coalesce related raw stores into larger units (called 'tiles').
3329 // It can avoid zeroing new storage for memory units which have raw inits.
3330 // At macro-expansion, it is marked 'complete', and does not optimize further.
3331 //
3332 // EXAMPLE:
3333 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3334 // ctl = incoming control; mem* = incoming memory
3335 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3336 // First allocate uninitialized memory and fill in the header:
3337 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3338 // ctl := alloc.Control; mem* := alloc.Memory*
3339 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3340 // Then initialize to zero the non-header parts of the raw memory block:
3341 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3342 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3343 // After the initialize node executes, the object is ready for service:
3344 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3345 // Suppose its body is immediately initialized as {1,2}:
3346 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3347 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3348 // mem.SLICE(#short[*]) := store2
3349 //
3350 // DETAILS:
3351 // An InitializeNode collects and isolates object initialization after
3352 // an AllocateNode and before the next possible safepoint. As a
3353 // memory barrier (MemBarNode), it keeps critical stores from drifting
3354 // down past any safepoint or any publication of the allocation.
3355 // Before this barrier, a newly-allocated object may have uninitialized bits.
3356 // After this barrier, it may be treated as a real oop, and GC is allowed.
3357 //
3358 // The semantics of the InitializeNode include an implicit zeroing of
3359 // the new object from object header to the end of the object.
3360 // (The object header and end are determined by the AllocateNode.)
3361 //
3362 // Certain stores may be added as direct inputs to the InitializeNode.
3363 // These stores must update raw memory, and they must be to addresses
3364 // derived from the raw address produced by AllocateNode, and with
3365 // a constant offset. They must be ordered by increasing offset.
3366 // The first one is at in(RawStores), the last at in(req()-1).
3367 // Unlike most memory operations, they are not linked in a chain,
3368 // but are displayed in parallel as users of the rawmem output of
3369 // the allocation.
3370 //
3371 // (See comments in InitializeNode::capture_store, which continue
3372 // the example given above.)
3373 //
3374 // When the associated Allocate is macro-expanded, the InitializeNode
3375 // may be rewritten to optimize collected stores. A ClearArrayNode
3376 // may also be created at that point to represent any required zeroing.
3377 // The InitializeNode is then marked 'complete', prohibiting further
3378 // capturing of nearby memory operations.
3379 //
3380 // During macro-expansion, all captured initializations which store
3381 // constant values of 32 bits or smaller are coalesced (if advantageous)
3382 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3383 // initialized in fewer memory operations. Memory words which are
3384 // covered by neither tiles nor non-constant stores are pre-zeroed
3385 // by explicit stores of zero. (The code shape happens to do all
3386 // zeroing first, then all other stores, with both sequences occurring
3387 // in order of ascending offsets.)
3388 //
3389 // Alternatively, code may be inserted between an AllocateNode and its
3390 // InitializeNode, to perform arbitrary initialization of the new object.
3391 // E.g., the object copying intrinsics insert complex data transfers here.
3392 // The initialization must then be marked as 'complete' disable the
3393 // built-in zeroing semantics and the collection of initializing stores.
3394 //
3395 // While an InitializeNode is incomplete, reads from the memory state
3396 // produced by it are optimizable if they match the control edge and
3397 // new oop address associated with the allocation/initialization.
3398 // They return a stored value (if the offset matches) or else zero.
3399 // A write to the memory state, if it matches control and address,
3400 // and if it is to a constant offset, may be 'captured' by the
3401 // InitializeNode. It is cloned as a raw memory operation and rewired
3402 // inside the initialization, to the raw oop produced by the allocation.
3403 // Operations on addresses which are provably distinct (e.g., to
3404 // other AllocateNodes) are allowed to bypass the initialization.
3405 //
3406 // The effect of all this is to consolidate object initialization
3407 // (both arrays and non-arrays, both piecewise and bulk) into a
3408 // single location, where it can be optimized as a unit.
3409 //
3410 // Only stores with an offset less than TrackedInitializationLimit words
3411 // will be considered for capture by an InitializeNode. This puts a
3412 // reasonable limit on the complexity of optimized initializations.
3413
3414 //---------------------------InitializeNode------------------------------------
3415 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3416 : MemBarNode(C, adr_type, rawoop),
3417 _is_complete(Incomplete), _does_not_escape(false)
3418 {
3419 init_class_id(Class_Initialize);
3420
3421 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3422 assert(in(RawAddress) == rawoop, "proper init");
3423 // Note: allocation() can be NULL, for secondary initialization barriers
3424 }
3425
3426 // Since this node is not matched, it will be processed by the
3427 // register allocator. Declare that there are no constraints
3428 // on the allocation of the RawAddress edge.
3429 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3430 // This edge should be set to top, by the set_complete. But be conservative.
3431 if (idx == InitializeNode::RawAddress)
3432 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3433 return RegMask::Empty;
3434 }
3435
3436 Node* InitializeNode::memory(uint alias_idx) {
3437 Node* mem = in(Memory);
3438 if (mem->is_MergeMem()) {
3439 return mem->as_MergeMem()->memory_at(alias_idx);
3440 } else {
3441 // incoming raw memory is not split
3442 return mem;
3443 }
3444 }
3445
3446 bool InitializeNode::is_non_zero() {
3447 if (is_complete()) return false;
3448 remove_extra_zeroes();
3449 return (req() > RawStores);
3450 }
3451
3452 void InitializeNode::set_complete(PhaseGVN* phase) {
3453 assert(!is_complete(), "caller responsibility");
3454 _is_complete = Complete;
3455
3456 // After this node is complete, it contains a bunch of
3457 // raw-memory initializations. There is no need for
3458 // it to have anything to do with non-raw memory effects.
3459 // Therefore, tell all non-raw users to re-optimize themselves,
3460 // after skipping the memory effects of this initialization.
3461 PhaseIterGVN* igvn = phase->is_IterGVN();
3462 if (igvn) igvn->add_users_to_worklist(this);
3463 }
3464
3465 // convenience function
3466 // return false if the init contains any stores already
3467 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3468 InitializeNode* init = initialization();
3469 if (init == NULL || init->is_complete()) return false;
3470 init->remove_extra_zeroes();
3471 // for now, if this allocation has already collected any inits, bail:
3472 if (init->is_non_zero()) return false;
3473 init->set_complete(phase);
3474 return true;
3475 }
3476
3477 void InitializeNode::remove_extra_zeroes() {
3478 if (req() == RawStores) return;
3479 Node* zmem = zero_memory();
3480 uint fill = RawStores;
3481 for (uint i = fill; i < req(); i++) {
3482 Node* n = in(i);
3483 if (n->is_top() || n == zmem) continue; // skip
3484 if (fill < i) set_req(fill, n); // compact
3485 ++fill;
3486 }
3487 // delete any empty spaces created:
3488 while (fill < req()) {
3489 del_req(fill);
3490 }
3491 }
3492
3493 // Helper for remembering which stores go with which offsets.
3494 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3495 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3496 intptr_t offset = -1;
3497 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3498 phase, offset);
3499 if (base == NULL) return -1; // something is dead,
3500 if (offset < 0) return -1; // dead, dead
3501 return offset;
3502 }
3503
3504 // Helper for proving that an initialization expression is
3505 // "simple enough" to be folded into an object initialization.
3506 // Attempts to prove that a store's initial value 'n' can be captured
3507 // within the initialization without creating a vicious cycle, such as:
3508 // { Foo p = new Foo(); p.next = p; }
3509 // True for constants and parameters and small combinations thereof.
3510 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3511 if (n == NULL) return true; // (can this really happen?)
3512 if (n->is_Proj()) n = n->in(0);
3513 if (n == this) return false; // found a cycle
3514 if (n->is_Con()) return true;
3515 if (n->is_Start()) return true; // params, etc., are OK
3516 if (n->is_Root()) return true; // even better
3517
3518 Node* ctl = n->in(0);
3519 if (ctl != NULL && !ctl->is_top()) {
3520 if (ctl->is_Proj()) ctl = ctl->in(0);
3521 if (ctl == this) return false;
3522
3523 // If we already know that the enclosing memory op is pinned right after
3524 // the init, then any control flow that the store has picked up
3525 // must have preceded the init, or else be equal to the init.
3526 // Even after loop optimizations (which might change control edges)
3527 // a store is never pinned *before* the availability of its inputs.
3528 if (!MemNode::all_controls_dominate(n, this))
3529 return false; // failed to prove a good control
3530 }
3531
3532 // Check data edges for possible dependencies on 'this'.
3533 if ((count += 1) > 20) return false; // complexity limit
3534 for (uint i = 1; i < n->req(); i++) {
3535 Node* m = n->in(i);
3536 if (m == NULL || m == n || m->is_top()) continue;
3537 uint first_i = n->find_edge(m);
3538 if (i != first_i) continue; // process duplicate edge just once
3539 if (!detect_init_independence(m, count)) {
3540 return false;
3541 }
3542 }
3543
3544 return true;
3545 }
3546
3547 // Here are all the checks a Store must pass before it can be moved into
3548 // an initialization. Returns zero if a check fails.
3549 // On success, returns the (constant) offset to which the store applies,
3550 // within the initialized memory.
3551 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3552 const int FAIL = 0;
3553 if (st->req() != MemNode::ValueIn + 1)
3554 return FAIL; // an inscrutable StoreNode (card mark?)
3555 Node* ctl = st->in(MemNode::Control);
3556 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3557 return FAIL; // must be unconditional after the initialization
3558 Node* mem = st->in(MemNode::Memory);
3559 if (!(mem->is_Proj() && mem->in(0) == this))
3560 return FAIL; // must not be preceded by other stores
3561 Node* adr = st->in(MemNode::Address);
3562 intptr_t offset;
3563 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3564 if (alloc == NULL)
3565 return FAIL; // inscrutable address
3566 if (alloc != allocation())
3567 return FAIL; // wrong allocation! (store needs to float up)
3568 int size_in_bytes = st->memory_size();
3569 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3570 return FAIL; // mismatched access
3571 }
3572 Node* val = st->in(MemNode::ValueIn);
3573 int complexity_count = 0;
3574 if (!detect_init_independence(val, complexity_count))
3575 return FAIL; // stored value must be 'simple enough'
3576
3577 // The Store can be captured only if nothing after the allocation
3578 // and before the Store is using the memory location that the store
3579 // overwrites.
3580 bool failed = false;
3581 // If is_complete_with_arraycopy() is true the shape of the graph is
3582 // well defined and is safe so no need for extra checks.
3583 if (!is_complete_with_arraycopy()) {
3584 // We are going to look at each use of the memory state following
3585 // the allocation to make sure nothing reads the memory that the
3586 // Store writes.
3587 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3588 int alias_idx = phase->C->get_alias_index(t_adr);
3589 ResourceMark rm;
3590 Unique_Node_List mems;
3591 mems.push(mem);
3592 Node* unique_merge = NULL;
3593 for (uint next = 0; next < mems.size(); ++next) {
3594 Node *m = mems.at(next);
3595 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3596 Node *n = m->fast_out(j);
3597 if (n->outcnt() == 0) {
3598 continue;
3599 }
3600 if (n == st) {
3601 continue;
3602 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3603 // If the control of this use is different from the control
3604 // of the Store which is right after the InitializeNode then
3605 // this node cannot be between the InitializeNode and the
3606 // Store.
3607 continue;
3608 } else if (n->is_MergeMem()) {
3609 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3610 // We can hit a MergeMemNode (that will likely go away
3611 // later) that is a direct use of the memory state
3612 // following the InitializeNode on the same slice as the
3613 // store node that we'd like to capture. We need to check
3614 // the uses of the MergeMemNode.
3615 mems.push(n);
3616 }
3617 } else if (n->is_Mem()) {
3618 Node* other_adr = n->in(MemNode::Address);
3619 if (other_adr == adr) {
3620 failed = true;
3621 break;
3622 } else {
3623 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3624 if (other_t_adr != NULL) {
3625 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3626 if (other_alias_idx == alias_idx) {
3627 // A load from the same memory slice as the store right
3628 // after the InitializeNode. We check the control of the
3629 // object/array that is loaded from. If it's the same as
3630 // the store control then we cannot capture the store.
3631 assert(!n->is_Store(), "2 stores to same slice on same control?");
3632 Node* base = other_adr;
3633 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3634 base = base->in(AddPNode::Base);
3635 if (base != NULL) {
3636 base = base->uncast();
3637 if (base->is_Proj() && base->in(0) == alloc) {
3638 failed = true;
3639 break;
3640 }
3641 }
3642 }
3643 }
3644 }
3645 } else {
3646 failed = true;
3647 break;
3648 }
3649 }
3650 }
3651 }
3652 if (failed) {
3653 if (!can_reshape) {
3654 // We decided we couldn't capture the store during parsing. We
3655 // should try again during the next IGVN once the graph is
3656 // cleaner.
3657 phase->C->record_for_igvn(st);
3658 }
3659 return FAIL;
3660 }
3661
3662 return offset; // success
3663 }
3664
3665 // Find the captured store in(i) which corresponds to the range
3666 // [start..start+size) in the initialized object.
3667 // If there is one, return its index i. If there isn't, return the
3668 // negative of the index where it should be inserted.
3669 // Return 0 if the queried range overlaps an initialization boundary
3670 // or if dead code is encountered.
3671 // If size_in_bytes is zero, do not bother with overlap checks.
3672 int InitializeNode::captured_store_insertion_point(intptr_t start,
3673 int size_in_bytes,
3674 PhaseTransform* phase) {
3675 const int FAIL = 0, MAX_STORE = BytesPerLong;
3676
3677 if (is_complete())
3678 return FAIL; // arraycopy got here first; punt
3679
3680 assert(allocation() != NULL, "must be present");
3681
3682 // no negatives, no header fields:
3683 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3684
3685 // after a certain size, we bail out on tracking all the stores:
3686 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3687 if (start >= ti_limit) return FAIL;
3688
3689 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3690 if (i >= limit) return -(int)i; // not found; here is where to put it
3691
3692 Node* st = in(i);
3693 intptr_t st_off = get_store_offset(st, phase);
3694 if (st_off < 0) {
3695 if (st != zero_memory()) {
3696 return FAIL; // bail out if there is dead garbage
3697 }
3698 } else if (st_off > start) {
3699 // ...we are done, since stores are ordered
3700 if (st_off < start + size_in_bytes) {
3701 return FAIL; // the next store overlaps
3702 }
3703 return -(int)i; // not found; here is where to put it
3704 } else if (st_off < start) {
3705 if (size_in_bytes != 0 &&
3706 start < st_off + MAX_STORE &&
3707 start < st_off + st->as_Store()->memory_size()) {
3708 return FAIL; // the previous store overlaps
3709 }
3710 } else {
3711 if (size_in_bytes != 0 &&
3712 st->as_Store()->memory_size() != size_in_bytes) {
3713 return FAIL; // mismatched store size
3714 }
3715 return i;
3716 }
3717
3718 ++i;
3719 }
3720 }
3721
3722 // Look for a captured store which initializes at the offset 'start'
3723 // with the given size. If there is no such store, and no other
3724 // initialization interferes, then return zero_memory (the memory
3725 // projection of the AllocateNode).
3726 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3727 PhaseTransform* phase) {
3728 assert(stores_are_sane(phase), "");
3729 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3730 if (i == 0) {
3731 return NULL; // something is dead
3732 } else if (i < 0) {
3733 return zero_memory(); // just primordial zero bits here
3734 } else {
3735 Node* st = in(i); // here is the store at this position
3736 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3737 return st;
3738 }
3739 }
3740
3741 // Create, as a raw pointer, an address within my new object at 'offset'.
3742 Node* InitializeNode::make_raw_address(intptr_t offset,
3743 PhaseTransform* phase) {
3744 Node* addr = in(RawAddress);
3745 if (offset != 0) {
3746 Compile* C = phase->C;
3747 addr = phase->transform( new AddPNode(C->top(), addr,
3748 phase->MakeConX(offset)) );
3749 }
3750 return addr;
3751 }
3752
3753 // Clone the given store, converting it into a raw store
3754 // initializing a field or element of my new object.
3755 // Caller is responsible for retiring the original store,
3756 // with subsume_node or the like.
3757 //
3758 // From the example above InitializeNode::InitializeNode,
3759 // here are the old stores to be captured:
3760 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3761 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3762 //
3763 // Here is the changed code; note the extra edges on init:
3764 // alloc = (Allocate ...)
3765 // rawoop = alloc.RawAddress
3766 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3767 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3768 // init = (Initialize alloc.Control alloc.Memory rawoop
3769 // rawstore1 rawstore2)
3770 //
3771 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3772 PhaseTransform* phase, bool can_reshape) {
3773 assert(stores_are_sane(phase), "");
3774
3775 if (start < 0) return NULL;
3776 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3777
3778 Compile* C = phase->C;
3779 int size_in_bytes = st->memory_size();
3780 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3781 if (i == 0) return NULL; // bail out
3782 Node* prev_mem = NULL; // raw memory for the captured store
3783 if (i > 0) {
3784 prev_mem = in(i); // there is a pre-existing store under this one
3785 set_req(i, C->top()); // temporarily disconnect it
3786 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3787 } else {
3788 i = -i; // no pre-existing store
3789 prev_mem = zero_memory(); // a slice of the newly allocated object
3790 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3791 set_req(--i, C->top()); // reuse this edge; it has been folded away
3792 else
3793 ins_req(i, C->top()); // build a new edge
3794 }
3795 Node* new_st = st->clone();
3796 new_st->set_req(MemNode::Control, in(Control));
3797 new_st->set_req(MemNode::Memory, prev_mem);
3798 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3799 new_st = phase->transform(new_st);
3800
3801 // At this point, new_st might have swallowed a pre-existing store
3802 // at the same offset, or perhaps new_st might have disappeared,
3803 // if it redundantly stored the same value (or zero to fresh memory).
3804
3805 // In any case, wire it in:
3806 phase->igvn_rehash_node_delayed(this);
3807 set_req(i, new_st);
3808
3809 // The caller may now kill the old guy.
3810 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3811 assert(check_st == new_st || check_st == NULL, "must be findable");
3812 assert(!is_complete(), "");
3813 return new_st;
3814 }
3815
3816 static bool store_constant(jlong* tiles, int num_tiles,
3817 intptr_t st_off, int st_size,
3818 jlong con) {
3819 if ((st_off & (st_size-1)) != 0)
3820 return false; // strange store offset (assume size==2**N)
3821 address addr = (address)tiles + st_off;
3822 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3823 switch (st_size) {
3824 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3825 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3826 case sizeof(jint): *(jint*) addr = (jint) con; break;
3827 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3828 default: return false; // strange store size (detect size!=2**N here)
3829 }
3830 return true; // return success to caller
3831 }
3832
3833 // Coalesce subword constants into int constants and possibly
3834 // into long constants. The goal, if the CPU permits,
3835 // is to initialize the object with a small number of 64-bit tiles.
3836 // Also, convert floating-point constants to bit patterns.
3837 // Non-constants are not relevant to this pass.
3838 //
3839 // In terms of the running example on InitializeNode::InitializeNode
3840 // and InitializeNode::capture_store, here is the transformation
3841 // of rawstore1 and rawstore2 into rawstore12:
3842 // alloc = (Allocate ...)
3843 // rawoop = alloc.RawAddress
3844 // tile12 = 0x00010002
3845 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3846 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3847 //
3848 void
3849 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3850 Node* size_in_bytes,
3851 PhaseGVN* phase) {
3852 Compile* C = phase->C;
3853
3854 assert(stores_are_sane(phase), "");
3855 // Note: After this pass, they are not completely sane,
3856 // since there may be some overlaps.
3857
3858 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3859
3860 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3861 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3862 size_limit = MIN2(size_limit, ti_limit);
3863 size_limit = align_up(size_limit, BytesPerLong);
3864 int num_tiles = size_limit / BytesPerLong;
3865
3866 // allocate space for the tile map:
3867 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3868 jlong tiles_buf[small_len];
3869 Node* nodes_buf[small_len];
3870 jlong inits_buf[small_len];
3871 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3872 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3873 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3874 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3875 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3876 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3877 // tiles: exact bitwise model of all primitive constants
3878 // nodes: last constant-storing node subsumed into the tiles model
3879 // inits: which bytes (in each tile) are touched by any initializations
3880
3881 //// Pass A: Fill in the tile model with any relevant stores.
3882
3883 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3884 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3885 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3886 Node* zmem = zero_memory(); // initially zero memory state
3887 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3888 Node* st = in(i);
3889 intptr_t st_off = get_store_offset(st, phase);
3890
3891 // Figure out the store's offset and constant value:
3892 if (st_off < header_size) continue; //skip (ignore header)
3893 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3894 int st_size = st->as_Store()->memory_size();
3895 if (st_off + st_size > size_limit) break;
3896
3897 // Record which bytes are touched, whether by constant or not.
3898 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3899 continue; // skip (strange store size)
3900
3901 const Type* val = phase->type(st->in(MemNode::ValueIn));
3902 if (!val->singleton()) continue; //skip (non-con store)
3903 BasicType type = val->basic_type();
3904
3905 jlong con = 0;
3906 switch (type) {
3907 case T_INT: con = val->is_int()->get_con(); break;
3908 case T_LONG: con = val->is_long()->get_con(); break;
3909 case T_FLOAT: con = jint_cast(val->getf()); break;
3910 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3911 default: continue; //skip (odd store type)
3912 }
3913
3914 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3915 st->Opcode() == Op_StoreL) {
3916 continue; // This StoreL is already optimal.
3917 }
3918
3919 // Store down the constant.
3920 store_constant(tiles, num_tiles, st_off, st_size, con);
3921
3922 intptr_t j = st_off >> LogBytesPerLong;
3923
3924 if (type == T_INT && st_size == BytesPerInt
3925 && (st_off & BytesPerInt) == BytesPerInt) {
3926 jlong lcon = tiles[j];
3927 if (!Matcher::isSimpleConstant64(lcon) &&
3928 st->Opcode() == Op_StoreI) {
3929 // This StoreI is already optimal by itself.
3930 jint* intcon = (jint*) &tiles[j];
3931 intcon[1] = 0; // undo the store_constant()
3932
3933 // If the previous store is also optimal by itself, back up and
3934 // undo the action of the previous loop iteration... if we can.
3935 // But if we can't, just let the previous half take care of itself.
3936 st = nodes[j];
3937 st_off -= BytesPerInt;
3938 con = intcon[0];
3939 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3940 assert(st_off >= header_size, "still ignoring header");
3941 assert(get_store_offset(st, phase) == st_off, "must be");
3942 assert(in(i-1) == zmem, "must be");
3943 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3944 assert(con == tcon->is_int()->get_con(), "must be");
3945 // Undo the effects of the previous loop trip, which swallowed st:
3946 intcon[0] = 0; // undo store_constant()
3947 set_req(i-1, st); // undo set_req(i, zmem)
3948 nodes[j] = NULL; // undo nodes[j] = st
3949 --old_subword; // undo ++old_subword
3950 }
3951 continue; // This StoreI is already optimal.
3952 }
3953 }
3954
3955 // This store is not needed.
3956 set_req(i, zmem);
3957 nodes[j] = st; // record for the moment
3958 if (st_size < BytesPerLong) // something has changed
3959 ++old_subword; // includes int/float, but who's counting...
3960 else ++old_long;
3961 }
3962
3963 if ((old_subword + old_long) == 0)
3964 return; // nothing more to do
3965
3966 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3967 // Be sure to insert them before overlapping non-constant stores.
3968 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3969 for (int j = 0; j < num_tiles; j++) {
3970 jlong con = tiles[j];
3971 jlong init = inits[j];
3972 if (con == 0) continue;
3973 jint con0, con1; // split the constant, address-wise
3974 jint init0, init1; // split the init map, address-wise
3975 { union { jlong con; jint intcon[2]; } u;
3976 u.con = con;
3977 con0 = u.intcon[0];
3978 con1 = u.intcon[1];
3979 u.con = init;
3980 init0 = u.intcon[0];
3981 init1 = u.intcon[1];
3982 }
3983
3984 Node* old = nodes[j];
3985 assert(old != NULL, "need the prior store");
3986 intptr_t offset = (j * BytesPerLong);
3987
3988 bool split = !Matcher::isSimpleConstant64(con);
3989
3990 if (offset < header_size) {
3991 assert(offset + BytesPerInt >= header_size, "second int counts");
3992 assert(*(jint*)&tiles[j] == 0, "junk in header");
3993 split = true; // only the second word counts
3994 // Example: int a[] = { 42 ... }
3995 } else if (con0 == 0 && init0 == -1) {
3996 split = true; // first word is covered by full inits
3997 // Example: int a[] = { ... foo(), 42 ... }
3998 } else if (con1 == 0 && init1 == -1) {
3999 split = true; // second word is covered by full inits
4000 // Example: int a[] = { ... 42, foo() ... }
4001 }
4002
4003 // Here's a case where init0 is neither 0 nor -1:
4004 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4005 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4006 // In this case the tile is not split; it is (jlong)42.
4007 // The big tile is stored down, and then the foo() value is inserted.
4008 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4009
4010 Node* ctl = old->in(MemNode::Control);
4011 Node* adr = make_raw_address(offset, phase);
4012 const TypePtr* atp = TypeRawPtr::BOTTOM;
4013
4014 // One or two coalesced stores to plop down.
4015 Node* st[2];
4016 intptr_t off[2];
4017 int nst = 0;
4018 if (!split) {
4019 ++new_long;
4020 off[nst] = offset;
4021 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4022 phase->longcon(con), T_LONG, MemNode::unordered);
4023 } else {
4024 // Omit either if it is a zero.
4025 if (con0 != 0) {
4026 ++new_int;
4027 off[nst] = offset;
4028 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4029 phase->intcon(con0), T_INT, MemNode::unordered);
4030 }
4031 if (con1 != 0) {
4032 ++new_int;
4033 offset += BytesPerInt;
4034 adr = make_raw_address(offset, phase);
4035 off[nst] = offset;
4036 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4037 phase->intcon(con1), T_INT, MemNode::unordered);
4038 }
4039 }
4040
4041 // Insert second store first, then the first before the second.
4042 // Insert each one just before any overlapping non-constant stores.
4043 while (nst > 0) {
4044 Node* st1 = st[--nst];
4045 C->copy_node_notes_to(st1, old);
4046 st1 = phase->transform(st1);
4047 offset = off[nst];
4048 assert(offset >= header_size, "do not smash header");
4049 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4050 guarantee(ins_idx != 0, "must re-insert constant store");
4051 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4052 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4053 set_req(--ins_idx, st1);
4054 else
4055 ins_req(ins_idx, st1);
4056 }
4057 }
4058
4059 if (PrintCompilation && WizardMode)
4060 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4061 old_subword, old_long, new_int, new_long);
4062 if (C->log() != NULL)
4063 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4064 old_subword, old_long, new_int, new_long);
4065
4066 // Clean up any remaining occurrences of zmem:
4067 remove_extra_zeroes();
4068 }
4069
4070 // Explore forward from in(start) to find the first fully initialized
4071 // word, and return its offset. Skip groups of subword stores which
4072 // together initialize full words. If in(start) is itself part of a
4073 // fully initialized word, return the offset of in(start). If there
4074 // are no following full-word stores, or if something is fishy, return
4075 // a negative value.
4076 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4077 int int_map = 0;
4078 intptr_t int_map_off = 0;
4079 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4080
4081 for (uint i = start, limit = req(); i < limit; i++) {
4082 Node* st = in(i);
4083
4084 intptr_t st_off = get_store_offset(st, phase);
4085 if (st_off < 0) break; // return conservative answer
4086
4087 int st_size = st->as_Store()->memory_size();
4088 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4089 return st_off; // we found a complete word init
4090 }
4091
4092 // update the map:
4093
4094 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4095 if (this_int_off != int_map_off) {
4096 // reset the map:
4097 int_map = 0;
4098 int_map_off = this_int_off;
4099 }
4100
4101 int subword_off = st_off - this_int_off;
4102 int_map |= right_n_bits(st_size) << subword_off;
4103 if ((int_map & FULL_MAP) == FULL_MAP) {
4104 return this_int_off; // we found a complete word init
4105 }
4106
4107 // Did this store hit or cross the word boundary?
4108 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4109 if (next_int_off == this_int_off + BytesPerInt) {
4110 // We passed the current int, without fully initializing it.
4111 int_map_off = next_int_off;
4112 int_map >>= BytesPerInt;
4113 } else if (next_int_off > this_int_off + BytesPerInt) {
4114 // We passed the current and next int.
4115 return this_int_off + BytesPerInt;
4116 }
4117 }
4118
4119 return -1;
4120 }
4121
4122
4123 // Called when the associated AllocateNode is expanded into CFG.
4124 // At this point, we may perform additional optimizations.
4125 // Linearize the stores by ascending offset, to make memory
4126 // activity as coherent as possible.
4127 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4128 intptr_t header_size,
4129 Node* size_in_bytes,
4130 PhaseGVN* phase) {
4131 assert(!is_complete(), "not already complete");
4132 assert(stores_are_sane(phase), "");
4133 assert(allocation() != NULL, "must be present");
4134
4135 remove_extra_zeroes();
4136
4137 if (ReduceFieldZeroing || ReduceBulkZeroing)
4138 // reduce instruction count for common initialization patterns
4139 coalesce_subword_stores(header_size, size_in_bytes, phase);
4140
4141 Node* zmem = zero_memory(); // initially zero memory state
4142 Node* inits = zmem; // accumulating a linearized chain of inits
4143 #ifdef ASSERT
4144 intptr_t first_offset = allocation()->minimum_header_size();
4145 intptr_t last_init_off = first_offset; // previous init offset
4146 intptr_t last_init_end = first_offset; // previous init offset+size
4147 intptr_t last_tile_end = first_offset; // previous tile offset+size
4148 #endif
4149 intptr_t zeroes_done = header_size;
4150
4151 bool do_zeroing = true; // we might give up if inits are very sparse
4152 int big_init_gaps = 0; // how many large gaps have we seen?
4153
4154 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4155 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4156
4157 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4158 Node* st = in(i);
4159 intptr_t st_off = get_store_offset(st, phase);
4160 if (st_off < 0)
4161 break; // unknown junk in the inits
4162 if (st->in(MemNode::Memory) != zmem)
4163 break; // complicated store chains somehow in list
4164
4165 int st_size = st->as_Store()->memory_size();
4166 intptr_t next_init_off = st_off + st_size;
4167
4168 if (do_zeroing && zeroes_done < next_init_off) {
4169 // See if this store needs a zero before it or under it.
4170 intptr_t zeroes_needed = st_off;
4171
4172 if (st_size < BytesPerInt) {
4173 // Look for subword stores which only partially initialize words.
4174 // If we find some, we must lay down some word-level zeroes first,
4175 // underneath the subword stores.
4176 //
4177 // Examples:
4178 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4179 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4180 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4181 //
4182 // Note: coalesce_subword_stores may have already done this,
4183 // if it was prompted by constant non-zero subword initializers.
4184 // But this case can still arise with non-constant stores.
4185
4186 intptr_t next_full_store = find_next_fullword_store(i, phase);
4187
4188 // In the examples above:
4189 // in(i) p q r s x y z
4190 // st_off 12 13 14 15 12 13 14
4191 // st_size 1 1 1 1 1 1 1
4192 // next_full_s. 12 16 16 16 16 16 16
4193 // z's_done 12 16 16 16 12 16 12
4194 // z's_needed 12 16 16 16 16 16 16
4195 // zsize 0 0 0 0 4 0 4
4196 if (next_full_store < 0) {
4197 // Conservative tack: Zero to end of current word.
4198 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4199 } else {
4200 // Zero to beginning of next fully initialized word.
4201 // Or, don't zero at all, if we are already in that word.
4202 assert(next_full_store >= zeroes_needed, "must go forward");
4203 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4204 zeroes_needed = next_full_store;
4205 }
4206 }
4207
4208 if (zeroes_needed > zeroes_done) {
4209 intptr_t zsize = zeroes_needed - zeroes_done;
4210 // Do some incremental zeroing on rawmem, in parallel with inits.
4211 zeroes_done = align_down(zeroes_done, BytesPerInt);
4212 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4213 zeroes_done, zeroes_needed,
4214 phase);
4215 zeroes_done = zeroes_needed;
4216 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4217 do_zeroing = false; // leave the hole, next time
4218 }
4219 }
4220
4221 // Collect the store and move on:
4222 st->set_req(MemNode::Memory, inits);
4223 inits = st; // put it on the linearized chain
4224 set_req(i, zmem); // unhook from previous position
4225
4226 if (zeroes_done == st_off)
4227 zeroes_done = next_init_off;
4228
4229 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4230
4231 #ifdef ASSERT
4232 // Various order invariants. Weaker than stores_are_sane because
4233 // a large constant tile can be filled in by smaller non-constant stores.
4234 assert(st_off >= last_init_off, "inits do not reverse");
4235 last_init_off = st_off;
4236 const Type* val = NULL;
4237 if (st_size >= BytesPerInt &&
4238 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4239 (int)val->basic_type() < (int)T_OBJECT) {
4240 assert(st_off >= last_tile_end, "tiles do not overlap");
4241 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4242 last_tile_end = MAX2(last_tile_end, next_init_off);
4243 } else {
4244 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4245 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4246 assert(st_off >= last_init_end, "inits do not overlap");
4247 last_init_end = next_init_off; // it's a non-tile
4248 }
4249 #endif //ASSERT
4250 }
4251
4252 remove_extra_zeroes(); // clear out all the zmems left over
4253 add_req(inits);
4254
4255 if (!(UseTLAB && ZeroTLAB)) {
4256 // If anything remains to be zeroed, zero it all now.
4257 zeroes_done = align_down(zeroes_done, BytesPerInt);
4258 // if it is the last unused 4 bytes of an instance, forget about it
4259 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4260 if (zeroes_done + BytesPerLong >= size_limit) {
4261 AllocateNode* alloc = allocation();
4262 assert(alloc != NULL, "must be present");
4263 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4264 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4265 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4266 if (zeroes_done == k->layout_helper())
4267 zeroes_done = size_limit;
4268 }
4269 }
4270 if (zeroes_done < size_limit) {
4271 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4272 zeroes_done, size_in_bytes, phase);
4273 }
4274 }
4275
4276 set_complete(phase);
4277 return rawmem;
4278 }
4279
4280
4281 #ifdef ASSERT
4282 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4283 if (is_complete())
4284 return true; // stores could be anything at this point
4285 assert(allocation() != NULL, "must be present");
4286 intptr_t last_off = allocation()->minimum_header_size();
4287 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4288 Node* st = in(i);
4289 intptr_t st_off = get_store_offset(st, phase);
4290 if (st_off < 0) continue; // ignore dead garbage
4291 if (last_off > st_off) {
4292 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4293 this->dump(2);
4294 assert(false, "ascending store offsets");
4295 return false;
4296 }
4297 last_off = st_off + st->as_Store()->memory_size();
4298 }
4299 return true;
4300 }
4301 #endif //ASSERT
4302
4303
4304
4305
4306 //============================MergeMemNode=====================================
4307 //
4308 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4309 // contributing store or call operations. Each contributor provides the memory
4310 // state for a particular "alias type" (see Compile::alias_type). For example,
4311 // if a MergeMem has an input X for alias category #6, then any memory reference
4312 // to alias category #6 may use X as its memory state input, as an exact equivalent
4313 // to using the MergeMem as a whole.
4314 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4315 //
4316 // (Here, the <N> notation gives the index of the relevant adr_type.)
4317 //
4318 // In one special case (and more cases in the future), alias categories overlap.
4319 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4320 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4321 // it is exactly equivalent to that state W:
4322 // MergeMem(<Bot>: W) <==> W
4323 //
4324 // Usually, the merge has more than one input. In that case, where inputs
4325 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4326 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4327 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4328 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4329 //
4330 // A merge can take a "wide" memory state as one of its narrow inputs.
4331 // This simply means that the merge observes out only the relevant parts of
4332 // the wide input. That is, wide memory states arriving at narrow merge inputs
4333 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4334 //
4335 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4336 // and that memory slices "leak through":
4337 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4338 //
4339 // But, in such a cascade, repeated memory slices can "block the leak":
4340 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4341 //
4342 // In the last example, Y is not part of the combined memory state of the
4343 // outermost MergeMem. The system must, of course, prevent unschedulable
4344 // memory states from arising, so you can be sure that the state Y is somehow
4345 // a precursor to state Y'.
4346 //
4347 //
4348 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4349 // of each MergeMemNode array are exactly the numerical alias indexes, including
4350 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4351 // Compile::alias_type (and kin) produce and manage these indexes.
4352 //
4353 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4354 // (Note that this provides quick access to the top node inside MergeMem methods,
4355 // without the need to reach out via TLS to Compile::current.)
4356 //
4357 // As a consequence of what was just described, a MergeMem that represents a full
4358 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4359 // containing all alias categories.
4360 //
4361 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4362 //
4363 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4364 // a memory state for the alias type <N>, or else the top node, meaning that
4365 // there is no particular input for that alias type. Note that the length of
4366 // a MergeMem is variable, and may be extended at any time to accommodate new
4367 // memory states at larger alias indexes. When merges grow, they are of course
4368 // filled with "top" in the unused in() positions.
4369 //
4370 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4371 // (Top was chosen because it works smoothly with passes like GCM.)
4372 //
4373 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4374 // the type of random VM bits like TLS references.) Since it is always the
4375 // first non-Bot memory slice, some low-level loops use it to initialize an
4376 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4377 //
4378 //
4379 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4380 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4381 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4382 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4383 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4384 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4385 //
4386 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4387 // really that different from the other memory inputs. An abbreviation called
4388 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4389 //
4390 //
4391 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4392 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4393 // that "emerges though" the base memory will be marked as excluding the alias types
4394 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4395 //
4396 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4397 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4398 //
4399 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4400 // (It is currently unimplemented.) As you can see, the resulting merge is
4401 // actually a disjoint union of memory states, rather than an overlay.
4402 //
4403
4404 //------------------------------MergeMemNode-----------------------------------
4405 Node* MergeMemNode::make_empty_memory() {
4406 Node* empty_memory = (Node*) Compile::current()->top();
4407 assert(empty_memory->is_top(), "correct sentinel identity");
4408 return empty_memory;
4409 }
4410
4411 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4412 init_class_id(Class_MergeMem);
4413 // all inputs are nullified in Node::Node(int)
4414 // set_input(0, NULL); // no control input
4415
4416 // Initialize the edges uniformly to top, for starters.
4417 Node* empty_mem = make_empty_memory();
4418 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4419 init_req(i,empty_mem);
4420 }
4421 assert(empty_memory() == empty_mem, "");
4422
4423 if( new_base != NULL && new_base->is_MergeMem() ) {
4424 MergeMemNode* mdef = new_base->as_MergeMem();
4425 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4426 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4427 mms.set_memory(mms.memory2());
4428 }
4429 assert(base_memory() == mdef->base_memory(), "");
4430 } else {
4431 set_base_memory(new_base);
4432 }
4433 }
4434
4435 // Make a new, untransformed MergeMem with the same base as 'mem'.
4436 // If mem is itself a MergeMem, populate the result with the same edges.
4437 MergeMemNode* MergeMemNode::make(Node* mem) {
4438 return new MergeMemNode(mem);
4439 }
4440
4441 //------------------------------cmp--------------------------------------------
4442 uint MergeMemNode::hash() const { return NO_HASH; }
4443 bool MergeMemNode::cmp( const Node &n ) const {
4444 return (&n == this); // Always fail except on self
4445 }
4446
4447 //------------------------------Identity---------------------------------------
4448 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4449 // Identity if this merge point does not record any interesting memory
4450 // disambiguations.
4451 Node* base_mem = base_memory();
4452 Node* empty_mem = empty_memory();
4453 if (base_mem != empty_mem) { // Memory path is not dead?
4454 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4455 Node* mem = in(i);
4456 if (mem != empty_mem && mem != base_mem) {
4457 return this; // Many memory splits; no change
4458 }
4459 }
4460 }
4461 return base_mem; // No memory splits; ID on the one true input
4462 }
4463
4464 //------------------------------Ideal------------------------------------------
4465 // This method is invoked recursively on chains of MergeMem nodes
4466 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4467 // Remove chain'd MergeMems
4468 //
4469 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4470 // relative to the "in(Bot)". Since we are patching both at the same time,
4471 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4472 // but rewrite each "in(i)" relative to the new "in(Bot)".
4473 Node *progress = NULL;
4474
4475
4476 Node* old_base = base_memory();
4477 Node* empty_mem = empty_memory();
4478 if (old_base == empty_mem)
4479 return NULL; // Dead memory path.
4480
4481 MergeMemNode* old_mbase;
4482 if (old_base != NULL && old_base->is_MergeMem())
4483 old_mbase = old_base->as_MergeMem();
4484 else
4485 old_mbase = NULL;
4486 Node* new_base = old_base;
4487
4488 // simplify stacked MergeMems in base memory
4489 if (old_mbase) new_base = old_mbase->base_memory();
4490
4491 // the base memory might contribute new slices beyond my req()
4492 if (old_mbase) grow_to_match(old_mbase);
4493
4494 // Look carefully at the base node if it is a phi.
4495 PhiNode* phi_base;
4496 if (new_base != NULL && new_base->is_Phi())
4497 phi_base = new_base->as_Phi();
4498 else
4499 phi_base = NULL;
4500
4501 Node* phi_reg = NULL;
4502 uint phi_len = (uint)-1;
4503 if (phi_base != NULL && !phi_base->is_copy()) {
4504 // do not examine phi if degraded to a copy
4505 phi_reg = phi_base->region();
4506 phi_len = phi_base->req();
4507 // see if the phi is unfinished
4508 for (uint i = 1; i < phi_len; i++) {
4509 if (phi_base->in(i) == NULL) {
4510 // incomplete phi; do not look at it yet!
4511 phi_reg = NULL;
4512 phi_len = (uint)-1;
4513 break;
4514 }
4515 }
4516 }
4517
4518 // Note: We do not call verify_sparse on entry, because inputs
4519 // can normalize to the base_memory via subsume_node or similar
4520 // mechanisms. This method repairs that damage.
4521
4522 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4523
4524 // Look at each slice.
4525 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4526 Node* old_in = in(i);
4527 // calculate the old memory value
4528 Node* old_mem = old_in;
4529 if (old_mem == empty_mem) old_mem = old_base;
4530 assert(old_mem == memory_at(i), "");
4531
4532 // maybe update (reslice) the old memory value
4533
4534 // simplify stacked MergeMems
4535 Node* new_mem = old_mem;
4536 MergeMemNode* old_mmem;
4537 if (old_mem != NULL && old_mem->is_MergeMem())
4538 old_mmem = old_mem->as_MergeMem();
4539 else
4540 old_mmem = NULL;
4541 if (old_mmem == this) {
4542 // This can happen if loops break up and safepoints disappear.
4543 // A merge of BotPtr (default) with a RawPtr memory derived from a
4544 // safepoint can be rewritten to a merge of the same BotPtr with
4545 // the BotPtr phi coming into the loop. If that phi disappears
4546 // also, we can end up with a self-loop of the mergemem.
4547 // In general, if loops degenerate and memory effects disappear,
4548 // a mergemem can be left looking at itself. This simply means
4549 // that the mergemem's default should be used, since there is
4550 // no longer any apparent effect on this slice.
4551 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4552 // from start. Update the input to TOP.
4553 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4554 }
4555 else if (old_mmem != NULL) {
4556 new_mem = old_mmem->memory_at(i);
4557 }
4558 // else preceding memory was not a MergeMem
4559
4560 // replace equivalent phis (unfortunately, they do not GVN together)
4561 if (new_mem != NULL && new_mem != new_base &&
4562 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4563 if (new_mem->is_Phi()) {
4564 PhiNode* phi_mem = new_mem->as_Phi();
4565 for (uint i = 1; i < phi_len; i++) {
4566 if (phi_base->in(i) != phi_mem->in(i)) {
4567 phi_mem = NULL;
4568 break;
4569 }
4570 }
4571 if (phi_mem != NULL) {
4572 // equivalent phi nodes; revert to the def
4573 new_mem = new_base;
4574 }
4575 }
4576 }
4577
4578 // maybe store down a new value
4579 Node* new_in = new_mem;
4580 if (new_in == new_base) new_in = empty_mem;
4581
4582 if (new_in != old_in) {
4583 // Warning: Do not combine this "if" with the previous "if"
4584 // A memory slice might have be be rewritten even if it is semantically
4585 // unchanged, if the base_memory value has changed.
4586 set_req(i, new_in);
4587 progress = this; // Report progress
4588 }
4589 }
4590
4591 if (new_base != old_base) {
4592 set_req(Compile::AliasIdxBot, new_base);
4593 // Don't use set_base_memory(new_base), because we need to update du.
4594 assert(base_memory() == new_base, "");
4595 progress = this;
4596 }
4597
4598 if( base_memory() == this ) {
4599 // a self cycle indicates this memory path is dead
4600 set_req(Compile::AliasIdxBot, empty_mem);
4601 }
4602
4603 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4604 // Recursion must occur after the self cycle check above
4605 if( base_memory()->is_MergeMem() ) {
4606 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4607 Node *m = phase->transform(new_mbase); // Rollup any cycles
4608 if( m != NULL &&
4609 (m->is_top() ||
4610 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4611 // propagate rollup of dead cycle to self
4612 set_req(Compile::AliasIdxBot, empty_mem);
4613 }
4614 }
4615
4616 if( base_memory() == empty_mem ) {
4617 progress = this;
4618 // Cut inputs during Parse phase only.
4619 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4620 if( !can_reshape ) {
4621 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4622 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4623 }
4624 }
4625 }
4626
4627 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4628 // Check if PhiNode::Ideal's "Split phis through memory merges"
4629 // transform should be attempted. Look for this->phi->this cycle.
4630 uint merge_width = req();
4631 if (merge_width > Compile::AliasIdxRaw) {
4632 PhiNode* phi = base_memory()->as_Phi();
4633 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4634 if (phi->in(i) == this) {
4635 phase->is_IterGVN()->_worklist.push(phi);
4636 break;
4637 }
4638 }
4639 }
4640 }
4641
4642 assert(progress || verify_sparse(), "please, no dups of base");
4643 return progress;
4644 }
4645
4646 //-------------------------set_base_memory-------------------------------------
4647 void MergeMemNode::set_base_memory(Node *new_base) {
4648 Node* empty_mem = empty_memory();
4649 set_req(Compile::AliasIdxBot, new_base);
4650 assert(memory_at(req()) == new_base, "must set default memory");
4651 // Clear out other occurrences of new_base:
4652 if (new_base != empty_mem) {
4653 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4654 if (in(i) == new_base) set_req(i, empty_mem);
4655 }
4656 }
4657 }
4658
4659 //------------------------------out_RegMask------------------------------------
4660 const RegMask &MergeMemNode::out_RegMask() const {
4661 return RegMask::Empty;
4662 }
4663
4664 //------------------------------dump_spec--------------------------------------
4665 #ifndef PRODUCT
4666 void MergeMemNode::dump_spec(outputStream *st) const {
4667 st->print(" {");
4668 Node* base_mem = base_memory();
4669 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4670 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4671 if (mem == base_mem) { st->print(" -"); continue; }
4672 st->print( " N%d:", mem->_idx );
4673 Compile::current()->get_adr_type(i)->dump_on(st);
4674 }
4675 st->print(" }");
4676 }
4677 #endif // !PRODUCT
4678
4679
4680 #ifdef ASSERT
4681 static bool might_be_same(Node* a, Node* b) {
4682 if (a == b) return true;
4683 if (!(a->is_Phi() || b->is_Phi())) return false;
4684 // phis shift around during optimization
4685 return true; // pretty stupid...
4686 }
4687
4688 // verify a narrow slice (either incoming or outgoing)
4689 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4690 if (!VerifyAliases) return; // don't bother to verify unless requested
4691 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4692 if (Node::in_dump()) return; // muzzle asserts when printing
4693 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4694 assert(n != NULL, "");
4695 // Elide intervening MergeMem's
4696 while (n->is_MergeMem()) {
4697 n = n->as_MergeMem()->memory_at(alias_idx);
4698 }
4699 Compile* C = Compile::current();
4700 const TypePtr* n_adr_type = n->adr_type();
4701 if (n == m->empty_memory()) {
4702 // Implicit copy of base_memory()
4703 } else if (n_adr_type != TypePtr::BOTTOM) {
4704 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4705 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4706 } else {
4707 // A few places like make_runtime_call "know" that VM calls are narrow,
4708 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4709 bool expected_wide_mem = false;
4710 if (n == m->base_memory()) {
4711 expected_wide_mem = true;
4712 } else if (alias_idx == Compile::AliasIdxRaw ||
4713 n == m->memory_at(Compile::AliasIdxRaw)) {
4714 expected_wide_mem = true;
4715 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4716 // memory can "leak through" calls on channels that
4717 // are write-once. Allow this also.
4718 expected_wide_mem = true;
4719 }
4720 assert(expected_wide_mem, "expected narrow slice replacement");
4721 }
4722 }
4723 #else // !ASSERT
4724 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4725 #endif
4726
4727
4728 //-----------------------------memory_at---------------------------------------
4729 Node* MergeMemNode::memory_at(uint alias_idx) const {
4730 assert(alias_idx >= Compile::AliasIdxRaw ||
4731 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4732 "must avoid base_memory and AliasIdxTop");
4733
4734 // Otherwise, it is a narrow slice.
4735 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4736 Compile *C = Compile::current();
4737 if (is_empty_memory(n)) {
4738 // the array is sparse; empty slots are the "top" node
4739 n = base_memory();
4740 assert(Node::in_dump()
4741 || n == NULL || n->bottom_type() == Type::TOP
4742 || n->adr_type() == NULL // address is TOP
4743 || n->adr_type() == TypePtr::BOTTOM
4744 || n->adr_type() == TypeRawPtr::BOTTOM
4745 || Compile::current()->AliasLevel() == 0,
4746 "must be a wide memory");
4747 // AliasLevel == 0 if we are organizing the memory states manually.
4748 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4749 } else {
4750 // make sure the stored slice is sane
4751 #ifdef ASSERT
4752 if (VMError::is_error_reported() || Node::in_dump()) {
4753 } else if (might_be_same(n, base_memory())) {
4754 // Give it a pass: It is a mostly harmless repetition of the base.
4755 // This can arise normally from node subsumption during optimization.
4756 } else {
4757 verify_memory_slice(this, alias_idx, n);
4758 }
4759 #endif
4760 }
4761 return n;
4762 }
4763
4764 //---------------------------set_memory_at-------------------------------------
4765 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4766 verify_memory_slice(this, alias_idx, n);
4767 Node* empty_mem = empty_memory();
4768 if (n == base_memory()) n = empty_mem; // collapse default
4769 uint need_req = alias_idx+1;
4770 if (req() < need_req) {
4771 if (n == empty_mem) return; // already the default, so do not grow me
4772 // grow the sparse array
4773 do {
4774 add_req(empty_mem);
4775 } while (req() < need_req);
4776 }
4777 set_req( alias_idx, n );
4778 }
4779
4780
4781
4782 //--------------------------iteration_setup------------------------------------
4783 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4784 if (other != NULL) {
4785 grow_to_match(other);
4786 // invariant: the finite support of mm2 is within mm->req()
4787 #ifdef ASSERT
4788 for (uint i = req(); i < other->req(); i++) {
4789 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4790 }
4791 #endif
4792 }
4793 // Replace spurious copies of base_memory by top.
4794 Node* base_mem = base_memory();
4795 if (base_mem != NULL && !base_mem->is_top()) {
4796 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4797 if (in(i) == base_mem)
4798 set_req(i, empty_memory());
4799 }
4800 }
4801 }
4802
4803 //---------------------------grow_to_match-------------------------------------
4804 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4805 Node* empty_mem = empty_memory();
4806 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4807 // look for the finite support of the other memory
4808 for (uint i = other->req(); --i >= req(); ) {
4809 if (other->in(i) != empty_mem) {
4810 uint new_len = i+1;
4811 while (req() < new_len) add_req(empty_mem);
4812 break;
4813 }
4814 }
4815 }
4816
4817 //---------------------------verify_sparse-------------------------------------
4818 #ifndef PRODUCT
4819 bool MergeMemNode::verify_sparse() const {
4820 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4821 Node* base_mem = base_memory();
4822 // The following can happen in degenerate cases, since empty==top.
4823 if (is_empty_memory(base_mem)) return true;
4824 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4825 assert(in(i) != NULL, "sane slice");
4826 if (in(i) == base_mem) return false; // should have been the sentinel value!
4827 }
4828 return true;
4829 }
4830
4831 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4832 Node* n;
4833 n = mm->in(idx);
4834 if (mem == n) return true; // might be empty_memory()
4835 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4836 if (mem == n) return true;
4837 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4838 if (mem == n) return true;
4839 if (n == NULL) break;
4840 }
4841 return false;
4842 }
4843 #endif // !PRODUCT