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