1 /*
2 * Copyright (c) 1997, 2022, 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 "ci/ciFlatArrayKlass.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/c2/barrierSetC2.hpp"
32 #include "gc/shared/tlab_globals.hpp"
33 #include "memory/allocation.inline.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "oops/objArrayKlass.hpp"
36 #include "opto/addnode.hpp"
37 #include "opto/arraycopynode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/regalloc.hpp"
40 #include "opto/compile.hpp"
41 #include "opto/connode.hpp"
42 #include "opto/convertnode.hpp"
43 #include "opto/inlinetypenode.hpp"
44 #include "opto/loopnode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/matcher.hpp"
47 #include "opto/memnode.hpp"
48 #include "opto/mulnode.hpp"
49 #include "opto/narrowptrnode.hpp"
50 #include "opto/phaseX.hpp"
51 #include "opto/regmask.hpp"
52 #include "opto/rootnode.hpp"
53 #include "opto/vectornode.hpp"
54 #include "utilities/align.hpp"
55 #include "utilities/copy.hpp"
56 #include "utilities/macros.hpp"
57 #include "utilities/powerOfTwo.hpp"
58 #include "utilities/vmError.hpp"
59
60 // Portions of code courtesy of Clifford Click
61
62 // Optimization - Graph Style
63
64 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
65
66 //=============================================================================
67 uint MemNode::size_of() const { return sizeof(*this); }
68
69 const TypePtr *MemNode::adr_type() const {
70 Node* adr = in(Address);
71 if (adr == NULL) return NULL; // node is dead
72 const TypePtr* cross_check = NULL;
73 DEBUG_ONLY(cross_check = _adr_type);
74 return calculate_adr_type(adr->bottom_type(), cross_check);
75 }
76
77 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
78 if (adr != NULL) {
79 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
80 return true;
81 }
82 }
83 return false;
84 }
85
86 #ifndef PRODUCT
87 void MemNode::dump_spec(outputStream *st) const {
88 if (in(Address) == NULL) return; // node is dead
89 #ifndef ASSERT
90 // fake the missing field
91 const TypePtr* _adr_type = NULL;
92 if (in(Address) != NULL)
93 _adr_type = in(Address)->bottom_type()->isa_ptr();
94 #endif
95 dump_adr_type(this, _adr_type, st);
96
97 Compile* C = Compile::current();
98 if (C->alias_type(_adr_type)->is_volatile()) {
99 st->print(" Volatile!");
100 }
101 if (_unaligned_access) {
102 st->print(" unaligned");
103 }
104 if (_mismatched_access) {
105 st->print(" mismatched");
106 }
107 if (_unsafe_access) {
108 st->print(" unsafe");
109 }
110 }
111
112 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
113 st->print(" @");
114 if (adr_type == NULL) {
115 st->print("NULL");
116 } else {
117 adr_type->dump_on(st);
118 Compile* C = Compile::current();
119 Compile::AliasType* atp = NULL;
120 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
121 if (atp == NULL)
122 st->print(", idx=?\?;");
123 else if (atp->index() == Compile::AliasIdxBot)
124 st->print(", idx=Bot;");
125 else if (atp->index() == Compile::AliasIdxTop)
126 st->print(", idx=Top;");
127 else if (atp->index() == Compile::AliasIdxRaw)
128 st->print(", idx=Raw;");
129 else {
130 ciField* field = atp->field();
131 if (field) {
132 st->print(", name=");
133 field->print_name_on(st);
134 }
135 st->print(", idx=%d;", atp->index());
136 }
137 }
138 }
139
140 extern void print_alias_types();
141
142 #endif
143
144 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
145 assert((t_oop != NULL), "sanity");
146 bool is_instance = t_oop->is_known_instance_field();
147 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
148 (load != NULL) && load->is_Load() &&
149 (phase->is_IterGVN() != NULL);
150 if (!(is_instance || is_boxed_value_load))
151 return mchain; // don't try to optimize non-instance types
152 uint instance_id = t_oop->instance_id();
153 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
154 Node *prev = NULL;
155 Node *result = mchain;
156 while (prev != result) {
157 prev = result;
158 if (result == start_mem)
159 break; // hit one of our sentinels
160 // skip over a call which does not affect this memory slice
161 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
162 Node *proj_in = result->in(0);
163 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
164 break; // hit one of our sentinels
165 } else if (proj_in->is_Call()) {
166 // ArrayCopyNodes processed here as well
167 CallNode *call = proj_in->as_Call();
168 if (!call->may_modify(t_oop, phase)) { // returns false for instances
169 result = call->in(TypeFunc::Memory);
170 }
171 } else if (proj_in->is_Initialize()) {
172 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
173 // Stop if this is the initialization for the object instance which
174 // contains this memory slice, otherwise skip over it.
175 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
176 break;
177 }
178 if (is_instance) {
179 result = proj_in->in(TypeFunc::Memory);
180 } else if (is_boxed_value_load) {
181 Node* klass = alloc->in(AllocateNode::KlassNode);
182 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
183 if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) {
184 result = proj_in->in(TypeFunc::Memory); // not related allocation
185 }
186 }
187 } else if (proj_in->is_MemBar()) {
188 ArrayCopyNode* ac = NULL;
189 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
190 break;
191 }
192 result = proj_in->in(TypeFunc::Memory);
193 } else {
194 assert(false, "unexpected projection");
195 }
196 } else if (result->is_ClearArray()) {
197 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
198 // Can not bypass initialization of the instance
199 // we are looking for.
200 break;
201 }
202 // Otherwise skip it (the call updated 'result' value).
203 } else if (result->is_MergeMem()) {
204 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
205 }
206 }
207 return result;
208 }
209
210 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
211 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
212 if (t_oop == NULL)
213 return mchain; // don't try to optimize non-oop types
214 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
215 bool is_instance = t_oop->is_known_instance_field();
216 PhaseIterGVN *igvn = phase->is_IterGVN();
217 if (is_instance && igvn != NULL && result->is_Phi()) {
218 PhiNode *mphi = result->as_Phi();
219 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
220 const TypePtr *t = mphi->adr_type();
221 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
222 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
223 t->is_oopptr()->cast_to_exactness(true)
224 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
225 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
226 // clone the Phi with our address type
227 result = mphi->split_out_instance(t_adr, igvn);
228 } else {
229 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
230 }
231 }
232 return result;
233 }
234
235 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
236 uint alias_idx = phase->C->get_alias_index(tp);
237 Node *mem = mmem;
238 #ifdef ASSERT
239 {
240 // Check that current type is consistent with the alias index used during graph construction
241 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
242 bool consistent = adr_check == NULL || adr_check->empty() ||
243 phase->C->must_alias(adr_check, alias_idx );
244 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
245 if( !consistent && adr_check != NULL && !adr_check->empty() &&
246 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
247 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
248 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
249 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
250 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
251 // don't assert if it is dead code.
252 consistent = true;
253 }
254 if( !consistent ) {
255 st->print("alias_idx==%d, adr_check==", alias_idx);
256 if( adr_check == NULL ) {
257 st->print("NULL");
258 } else {
259 adr_check->dump();
260 }
261 st->cr();
262 print_alias_types();
263 assert(consistent, "adr_check must match alias idx");
264 }
265 }
266 #endif
267 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
268 // means an array I have not precisely typed yet. Do not do any
269 // alias stuff with it any time soon.
270 const TypeOopPtr *toop = tp->isa_oopptr();
271 if (tp->base() != Type::AnyPtr &&
272 !(toop &&
273 toop->isa_instptr() &&
274 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
275 toop->offset() == Type::OffsetBot)) {
276 // compress paths and change unreachable cycles to TOP
277 // If not, we can update the input infinitely along a MergeMem cycle
278 // Equivalent code in PhiNode::Ideal
279 Node* m = phase->transform(mmem);
280 // If transformed to a MergeMem, get the desired slice
281 // Otherwise the returned node represents memory for every slice
282 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
283 // Update input if it is progress over what we have now
284 }
285 return mem;
286 }
287
288 //--------------------------Ideal_common---------------------------------------
289 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
290 // Unhook non-raw memories from complete (macro-expanded) initializations.
291 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
292 // If our control input is a dead region, kill all below the region
293 Node *ctl = in(MemNode::Control);
294 if (ctl && remove_dead_region(phase, can_reshape))
295 return this;
296 ctl = in(MemNode::Control);
297 // Don't bother trying to transform a dead node
298 if (ctl && ctl->is_top()) return NodeSentinel;
299
300 PhaseIterGVN *igvn = phase->is_IterGVN();
301 // Wait if control on the worklist.
302 if (ctl && can_reshape && igvn != NULL) {
303 Node* bol = NULL;
304 Node* cmp = NULL;
305 if (ctl->in(0)->is_If()) {
306 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
307 bol = ctl->in(0)->in(1);
308 if (bol->is_Bool())
309 cmp = ctl->in(0)->in(1)->in(1);
310 }
311 if (igvn->_worklist.member(ctl) ||
312 (bol != NULL && igvn->_worklist.member(bol)) ||
313 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
314 // This control path may be dead.
315 // Delay this memory node transformation until the control is processed.
316 igvn->_worklist.push(this);
317 return NodeSentinel; // caller will return NULL
318 }
319 }
320 // Ignore if memory is dead, or self-loop
321 Node *mem = in(MemNode::Memory);
322 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
323 assert(mem != this, "dead loop in MemNode::Ideal");
324
325 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
326 // This memory slice may be dead.
327 // Delay this mem node transformation until the memory is processed.
328 igvn->_worklist.push(this);
329 return NodeSentinel; // caller will return NULL
330 }
331
332 Node *address = in(MemNode::Address);
333 const Type *t_adr = phase->type(address);
334 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
335
336 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
337 // Unsafe off-heap access with zero address. Remove access and other control users
338 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
339 assert(ctl != NULL, "unsafe accesses should be control dependent");
340 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
341 Node* u = ctl->fast_out(i);
342 if (u != ctl) {
343 igvn->rehash_node_delayed(u);
344 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
345 --i, imax -= nb;
346 }
347 }
348 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
349 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
350 phase->C->root()->add_req(halt);
351 return this;
352 }
353
354 if (can_reshape && igvn != NULL &&
355 (igvn->_worklist.member(address) ||
356 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
357 // The address's base and type may change when the address is processed.
358 // Delay this mem node transformation until the address is processed.
359 igvn->_worklist.push(this);
360 return NodeSentinel; // caller will return NULL
361 }
362
363 // Do NOT remove or optimize the next lines: ensure a new alias index
364 // is allocated for an oop pointer type before Escape Analysis.
365 // Note: C++ will not remove it since the call has side effect.
366 if (t_adr->isa_oopptr()) {
367 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
368 }
369
370 Node* base = NULL;
371 if (address->is_AddP()) {
372 base = address->in(AddPNode::Base);
373 }
374 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
375 !t_adr->isa_rawptr()) {
376 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
377 // Skip this node optimization if its address has TOP base.
378 return NodeSentinel; // caller will return NULL
379 }
380
381 // Avoid independent memory operations
382 Node* old_mem = mem;
383
384 // The code which unhooks non-raw memories from complete (macro-expanded)
385 // initializations was removed. After macro-expansion all stores caught
386 // by Initialize node became raw stores and there is no information
387 // which memory slices they modify. So it is unsafe to move any memory
388 // operation above these stores. Also in most cases hooked non-raw memories
389 // were already unhooked by using information from detect_ptr_independence()
390 // and find_previous_store().
391
392 if (mem->is_MergeMem()) {
393 MergeMemNode* mmem = mem->as_MergeMem();
394 const TypePtr *tp = t_adr->is_ptr();
395
396 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
397 }
398
399 if (mem != old_mem) {
400 set_req_X(MemNode::Memory, mem, phase);
401 if (phase->type(mem) == Type::TOP) return NodeSentinel;
402 return this;
403 }
404
405 // let the subclass continue analyzing...
406 return NULL;
407 }
408
409 // Helper function for proving some simple control dominations.
410 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
411 // Already assumes that 'dom' is available at 'sub', and that 'sub'
412 // is not a constant (dominated by the method's StartNode).
413 // Used by MemNode::find_previous_store to prove that the
414 // control input of a memory operation predates (dominates)
415 // an allocation it wants to look past.
416 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
417 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
418 return false; // Conservative answer for dead code
419
420 // Check 'dom'. Skip Proj and CatchProj nodes.
421 dom = dom->find_exact_control(dom);
422 if (dom == NULL || dom->is_top())
423 return false; // Conservative answer for dead code
424
425 if (dom == sub) {
426 // For the case when, for example, 'sub' is Initialize and the original
427 // 'dom' is Proj node of the 'sub'.
428 return false;
429 }
430
431 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
432 return true;
433
434 // 'dom' dominates 'sub' if its control edge and control edges
435 // of all its inputs dominate or equal to sub's control edge.
436
437 // Currently 'sub' is either Allocate, Initialize or Start nodes.
438 // Or Region for the check in LoadNode::Ideal();
439 // 'sub' should have sub->in(0) != NULL.
440 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
441 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
442
443 // Get control edge of 'sub'.
444 Node* orig_sub = sub;
445 sub = sub->find_exact_control(sub->in(0));
446 if (sub == NULL || sub->is_top())
447 return false; // Conservative answer for dead code
448
449 assert(sub->is_CFG(), "expecting control");
450
451 if (sub == dom)
452 return true;
453
454 if (sub->is_Start() || sub->is_Root())
455 return false;
456
457 {
458 // Check all control edges of 'dom'.
459
460 ResourceMark rm;
461 Node_List nlist;
462 Unique_Node_List dom_list;
463
464 dom_list.push(dom);
465 bool only_dominating_controls = false;
466
467 for (uint next = 0; next < dom_list.size(); next++) {
468 Node* n = dom_list.at(next);
469 if (n == orig_sub)
470 return false; // One of dom's inputs dominated by sub.
471 if (!n->is_CFG() && n->pinned()) {
472 // Check only own control edge for pinned non-control nodes.
473 n = n->find_exact_control(n->in(0));
474 if (n == NULL || n->is_top())
475 return false; // Conservative answer for dead code
476 assert(n->is_CFG(), "expecting control");
477 dom_list.push(n);
478 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
479 only_dominating_controls = true;
480 } else if (n->is_CFG()) {
481 if (n->dominates(sub, nlist))
482 only_dominating_controls = true;
483 else
484 return false;
485 } else {
486 // First, own control edge.
487 Node* m = n->find_exact_control(n->in(0));
488 if (m != NULL) {
489 if (m->is_top())
490 return false; // Conservative answer for dead code
491 dom_list.push(m);
492 }
493 // Now, the rest of edges.
494 uint cnt = n->req();
495 for (uint i = 1; i < cnt; i++) {
496 m = n->find_exact_control(n->in(i));
497 if (m == NULL || m->is_top())
498 continue;
499 dom_list.push(m);
500 }
501 }
502 }
503 return only_dominating_controls;
504 }
505 }
506
507 //---------------------detect_ptr_independence---------------------------------
508 // Used by MemNode::find_previous_store to prove that two base
509 // pointers are never equal.
510 // The pointers are accompanied by their associated allocations,
511 // if any, which have been previously discovered by the caller.
512 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
513 Node* p2, AllocateNode* a2,
514 PhaseTransform* phase) {
515 // Attempt to prove that these two pointers cannot be aliased.
516 // They may both manifestly be allocations, and they should differ.
517 // Or, if they are not both allocations, they can be distinct constants.
518 // Otherwise, one is an allocation and the other a pre-existing value.
519 if (a1 == NULL && a2 == NULL) { // neither an allocation
520 return (p1 != p2) && p1->is_Con() && p2->is_Con();
521 } else if (a1 != NULL && a2 != NULL) { // both allocations
522 return (a1 != a2);
523 } else if (a1 != NULL) { // one allocation a1
524 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
525 return all_controls_dominate(p2, a1);
526 } else { //(a2 != NULL) // one allocation a2
527 return all_controls_dominate(p1, a2);
528 }
529 return false;
530 }
531
532
533 // Find an arraycopy ac that produces the memory state represented by parameter mem.
534 // Return ac if
535 // (a) can_see_stored_value=true and ac must have set the value for this load or if
536 // (b) can_see_stored_value=false and ac could have set the value for this load or if
537 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
538 // In case (c) change the parameter mem to the memory input of ac to skip it
539 // when searching stored value.
540 // Otherwise return NULL.
541 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
542 ArrayCopyNode* ac = find_array_copy_clone(phase, ld_alloc, mem);
543 if (ac != NULL) {
544 Node* ld_addp = in(MemNode::Address);
545 Node* src = ac->in(ArrayCopyNode::Src);
546 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
547
548 // This is a load from a cloned array. The corresponding arraycopy ac must
549 // have set the value for the load and we can return ac but only if the load
550 // is known to be within bounds. This is checked below.
551 if (ary_t != NULL && ld_addp->is_AddP()) {
552 Node* ld_offs = ld_addp->in(AddPNode::Offset);
553 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
554 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
555 jlong elemsize = type2aelembytes(ary_elem);
556
557 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
558 const TypeInt* sizetype = ary_t->size();
559
560 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
561 // The load is known to be within bounds. It receives its value from ac.
562 return ac;
563 }
564 // The load is known to be out-of-bounds.
565 }
566 // The load could be out-of-bounds. It must not be hoisted but must remain
567 // dependent on the runtime range check. This is achieved by returning NULL.
568 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
569 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
570
571 if (ac->is_arraycopy_validated() ||
572 ac->is_copyof_validated() ||
573 ac->is_copyofrange_validated()) {
574 Node* ld_addp = in(MemNode::Address);
575 if (ld_addp->is_AddP()) {
576 Node* ld_base = ld_addp->in(AddPNode::Address);
577 Node* ld_offs = ld_addp->in(AddPNode::Offset);
578
579 Node* dest = ac->in(ArrayCopyNode::Dest);
580
581 if (dest == ld_base) {
582 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
583 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
584 return ac;
585 }
586 if (!can_see_stored_value) {
587 mem = ac->in(TypeFunc::Memory);
588 return ac;
589 }
590 }
591 }
592 }
593 }
594 return NULL;
595 }
596
597 ArrayCopyNode* MemNode::find_array_copy_clone(PhaseTransform* phase, Node* ld_alloc, Node* mem) const {
598 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
599 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
600 if (ld_alloc != NULL) {
601 // Check if there is an array copy for a clone
602 Node* mb = mem->in(0);
603 ArrayCopyNode* ac = NULL;
604 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
605 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
606 ac = mb->in(0)->in(0)->as_ArrayCopy();
607 } else {
608 // Step over GC barrier when ReduceInitialCardMarks is disabled
609 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
610 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
611
612 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
613 ac = control_proj_ac->in(0)->as_ArrayCopy();
614 }
615 }
616
617 if (ac != NULL && ac->is_clonebasic()) {
618 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase);
619 if (alloc != NULL && alloc == ld_alloc) {
620 return ac;
621 }
622 }
623 }
624 }
625 return NULL;
626 }
627
628 // The logic for reordering loads and stores uses four steps:
629 // (a) Walk carefully past stores and initializations which we
630 // can prove are independent of this load.
631 // (b) Observe that the next memory state makes an exact match
632 // with self (load or store), and locate the relevant store.
633 // (c) Ensure that, if we were to wire self directly to the store,
634 // the optimizer would fold it up somehow.
635 // (d) Do the rewiring, and return, depending on some other part of
636 // the optimizer to fold up the load.
637 // This routine handles steps (a) and (b). Steps (c) and (d) are
638 // specific to loads and stores, so they are handled by the callers.
639 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
640 //
641 Node* MemNode::find_previous_store(PhaseTransform* phase) {
642 Node* ctrl = in(MemNode::Control);
643 Node* adr = in(MemNode::Address);
644 intptr_t offset = 0;
645 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
646 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
647
648 if (offset == Type::OffsetBot)
649 return NULL; // cannot unalias unless there are precise offsets
650
651 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
652 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
653
654 intptr_t size_in_bytes = memory_size();
655
656 Node* mem = in(MemNode::Memory); // start searching here...
657
658 int cnt = 50; // Cycle limiter
659 for (;;) { // While we can dance past unrelated stores...
660 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
661
662 Node* prev = mem;
663 if (mem->is_Store()) {
664 Node* st_adr = mem->in(MemNode::Address);
665 intptr_t st_offset = 0;
666 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
667 if (st_base == NULL)
668 break; // inscrutable pointer
669
670 // For raw accesses it's not enough to prove that constant offsets don't intersect.
671 // We need the bases to be the equal in order for the offset check to make sense.
672 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
673 break;
674 }
675
676 if (st_offset != offset && st_offset != Type::OffsetBot) {
677 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
678 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
679 if (st_offset >= offset + size_in_bytes ||
680 st_offset <= offset - MAX_STORE ||
681 st_offset <= offset - mem->as_Store()->memory_size()) {
682 // Success: The offsets are provably independent.
683 // (You may ask, why not just test st_offset != offset and be done?
684 // The answer is that stores of different sizes can co-exist
685 // in the same sequence of RawMem effects. We sometimes initialize
686 // a whole 'tile' of array elements with a single jint or jlong.)
687 mem = mem->in(MemNode::Memory);
688 continue; // (a) advance through independent store memory
689 }
690 }
691 if (st_base != base &&
692 detect_ptr_independence(base, alloc,
693 st_base,
694 AllocateNode::Ideal_allocation(st_base, phase),
695 phase)) {
696 // Success: The bases are provably independent.
697 mem = mem->in(MemNode::Memory);
698 continue; // (a) advance through independent store memory
699 }
700
701 // (b) At this point, if the bases or offsets do not agree, we lose,
702 // since we have not managed to prove 'this' and 'mem' independent.
703 if (st_base == base && st_offset == offset) {
704 return mem; // let caller handle steps (c), (d)
705 }
706
707 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
708 InitializeNode* st_init = mem->in(0)->as_Initialize();
709 AllocateNode* st_alloc = st_init->allocation();
710 if (st_alloc == NULL)
711 break; // something degenerated
712 bool known_identical = false;
713 bool known_independent = false;
714 if (alloc == st_alloc)
715 known_identical = true;
716 else if (alloc != NULL)
717 known_independent = true;
718 else if (all_controls_dominate(this, st_alloc))
719 known_independent = true;
720
721 if (known_independent) {
722 // The bases are provably independent: Either they are
723 // manifestly distinct allocations, or else the control
724 // of this load dominates the store's allocation.
725 int alias_idx = phase->C->get_alias_index(adr_type());
726 if (alias_idx == Compile::AliasIdxRaw) {
727 mem = st_alloc->in(TypeFunc::Memory);
728 } else {
729 mem = st_init->memory(alias_idx);
730 }
731 continue; // (a) advance through independent store memory
732 }
733
734 // (b) at this point, if we are not looking at a store initializing
735 // the same allocation we are loading from, we lose.
736 if (known_identical) {
737 // From caller, can_see_stored_value will consult find_captured_store.
738 return mem; // let caller handle steps (c), (d)
739 }
740
741 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
742 if (prev != mem) {
743 // Found an arraycopy but it doesn't affect that load
744 continue;
745 }
746 // Found an arraycopy that may affect that load
747 return mem;
748 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
749 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
750 if (mem->is_Proj() && mem->in(0)->is_Call()) {
751 // ArrayCopyNodes processed here as well.
752 CallNode *call = mem->in(0)->as_Call();
753 if (!call->may_modify(addr_t, phase)) {
754 mem = call->in(TypeFunc::Memory);
755 continue; // (a) advance through independent call memory
756 }
757 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
758 ArrayCopyNode* ac = NULL;
759 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
760 break;
761 }
762 mem = mem->in(0)->in(TypeFunc::Memory);
763 continue; // (a) advance through independent MemBar memory
764 } else if (mem->is_ClearArray()) {
765 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
766 // (the call updated 'mem' value)
767 continue; // (a) advance through independent allocation memory
768 } else {
769 // Can not bypass initialization of the instance
770 // we are looking for.
771 return mem;
772 }
773 } else if (mem->is_MergeMem()) {
774 int alias_idx = phase->C->get_alias_index(adr_type());
775 mem = mem->as_MergeMem()->memory_at(alias_idx);
776 continue; // (a) advance through independent MergeMem memory
777 }
778 }
779
780 // Unless there is an explicit 'continue', we must bail out here,
781 // because 'mem' is an inscrutable memory state (e.g., a call).
782 break;
783 }
784
785 return NULL; // bail out
786 }
787
788 //----------------------calculate_adr_type-------------------------------------
789 // Helper function. Notices when the given type of address hits top or bottom.
790 // Also, asserts a cross-check of the type against the expected address type.
791 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
792 if (t == Type::TOP) return NULL; // does not touch memory any more?
793 #ifdef ASSERT
794 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
795 #endif
796 const TypePtr* tp = t->isa_ptr();
797 if (tp == NULL) {
798 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
799 return TypePtr::BOTTOM; // touches lots of memory
800 } else {
801 #ifdef ASSERT
802 // %%%% [phh] We don't check the alias index if cross_check is
803 // TypeRawPtr::BOTTOM. Needs to be investigated.
804 if (cross_check != NULL &&
805 cross_check != TypePtr::BOTTOM &&
806 cross_check != TypeRawPtr::BOTTOM) {
807 // Recheck the alias index, to see if it has changed (due to a bug).
808 Compile* C = Compile::current();
809 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
810 "must stay in the original alias category");
811 // The type of the address must be contained in the adr_type,
812 // disregarding "null"-ness.
813 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
814 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
815 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
816 "real address must not escape from expected memory type");
817 }
818 #endif
819 return tp;
820 }
821 }
822
823 //=============================================================================
824 // Should LoadNode::Ideal() attempt to remove control edges?
825 bool LoadNode::can_remove_control() const {
826 return true;
827 }
828 uint LoadNode::size_of() const { return sizeof(*this); }
829 bool LoadNode::cmp( const Node &n ) const
830 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
831 const Type *LoadNode::bottom_type() const { return _type; }
832 uint LoadNode::ideal_reg() const {
833 return _type->ideal_reg();
834 }
835
836 #ifndef PRODUCT
837 void LoadNode::dump_spec(outputStream *st) const {
838 MemNode::dump_spec(st);
839 if( !Verbose && !WizardMode ) {
840 // standard dump does this in Verbose and WizardMode
841 st->print(" #"); _type->dump_on(st);
842 }
843 if (!depends_only_on_test()) {
844 st->print(" (does not depend only on test)");
845 }
846 }
847 #endif
848
849 #ifdef ASSERT
850 //----------------------------is_immutable_value-------------------------------
851 // Helper function to allow a raw load without control edge for some cases
852 bool LoadNode::is_immutable_value(Node* adr) {
853 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
854 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
855
856 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
857 int offsets[] = {
858 in_bytes(JavaThread::osthread_offset()),
859 in_bytes(JavaThread::threadObj_offset()),
860 in_bytes(JavaThread::vthread_offset()),
861 in_bytes(JavaThread::extentLocalCache_offset()),
862 };
863
864 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
865 if (offset == offsets[i]) {
866 return true;
867 }
868 }
869 }
870
871 return false;
872 }
873 #endif
874
875 //----------------------------LoadNode::make-----------------------------------
876 // Polymorphic factory method:
877 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
878 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
879 Compile* C = gvn.C;
880
881 // sanity check the alias category against the created node type
882 assert(!(adr_type->isa_oopptr() &&
883 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
884 "use LoadKlassNode instead");
885 assert(!(adr_type->isa_aryptr() &&
886 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
887 "use LoadRangeNode instead");
888 // Check control edge of raw loads
889 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
890 // oop will be recorded in oop map if load crosses safepoint
891 rt->isa_oopptr() || is_immutable_value(adr),
892 "raw memory operations should have control edge");
893 LoadNode* load = NULL;
894 switch (bt) {
895 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
896 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
897 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
898 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
899 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
900 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
901 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
902 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
903 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
904 case T_PRIMITIVE_OBJECT:
905 case T_OBJECT:
906 #ifdef _LP64
907 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
908 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
909 } else
910 #endif
911 {
912 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
913 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
914 }
915 break;
916 default:
917 ShouldNotReachHere();
918 break;
919 }
920 assert(load != NULL, "LoadNode should have been created");
921 if (unaligned) {
922 load->set_unaligned_access();
923 }
924 if (mismatched) {
925 load->set_mismatched_access();
926 }
927 if (unsafe) {
928 load->set_unsafe_access();
929 }
930 load->set_barrier_data(barrier_data);
931 if (load->Opcode() == Op_LoadN) {
932 Node* ld = gvn.transform(load);
933 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
934 }
935
936 return load;
937 }
938
939 //------------------------------hash-------------------------------------------
940 uint LoadNode::hash() const {
941 // unroll addition of interesting fields
942 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
943 }
944
945 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
946 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
947 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
948 bool is_stable_ary = FoldStableValues &&
949 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
950 tp->isa_aryptr()->is_stable();
951
952 return (eliminate_boxing && non_volatile) || is_stable_ary;
953 }
954
955 return false;
956 }
957
958 // Is the value loaded previously stored by an arraycopy? If so return
959 // a load node that reads from the source array so we may be able to
960 // optimize out the ArrayCopy node later.
961 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
962 Node* ld_adr = in(MemNode::Address);
963 intptr_t ld_off = 0;
964 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
965 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
966 if (ac != NULL) {
967 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
968
969 Node* mem = ac->in(TypeFunc::Memory);
970 Node* ctl = ac->in(0);
971 Node* src = ac->in(ArrayCopyNode::Src);
972
973 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
974 return NULL;
975 }
976
977 LoadNode* ld = clone()->as_Load();
978 Node* addp = in(MemNode::Address)->clone();
979 if (ac->as_ArrayCopy()->is_clonebasic()) {
980 assert(ld_alloc != NULL, "need an alloc");
981 assert(addp->is_AddP(), "address must be addp");
982 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
983 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
984 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
985 addp->set_req(AddPNode::Base, src);
986 addp->set_req(AddPNode::Address, src);
987 } else {
988 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
989 ac->as_ArrayCopy()->is_copyof_validated() ||
990 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
991 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
992 addp->set_req(AddPNode::Base, src);
993 addp->set_req(AddPNode::Address, src);
994
995 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
996 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
997 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
998
999 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1000 uint shift = ary_t->is_flat() ? ary_t->flat_log_elem_size() : exact_log2(type2aelembytes(ary_elem));
1001
1002 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1003 #ifdef _LP64
1004 diff = phase->transform(new ConvI2LNode(diff));
1005 #endif
1006 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1007
1008 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1009 addp->set_req(AddPNode::Offset, offset);
1010 }
1011 addp = phase->transform(addp);
1012 #ifdef ASSERT
1013 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1014 ld->_adr_type = adr_type;
1015 #endif
1016 ld->set_req(MemNode::Address, addp);
1017 ld->set_req(0, ctl);
1018 ld->set_req(MemNode::Memory, mem);
1019 // load depends on the tests that validate the arraycopy
1020 ld->_control_dependency = UnknownControl;
1021 return ld;
1022 }
1023 return NULL;
1024 }
1025
1026
1027 //---------------------------can_see_stored_value------------------------------
1028 // This routine exists to make sure this set of tests is done the same
1029 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1030 // will change the graph shape in a way which makes memory alive twice at the
1031 // same time (uses the Oracle model of aliasing), then some
1032 // LoadXNode::Identity will fold things back to the equivalence-class model
1033 // of aliasing.
1034 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
1035 Node* ld_adr = in(MemNode::Address);
1036 intptr_t ld_off = 0;
1037 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1038 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase);
1039 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1040 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1041 // This is more general than load from boxing objects.
1042 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1043 uint alias_idx = atp->index();
1044 Node* result = NULL;
1045 Node* current = st;
1046 // Skip through chains of MemBarNodes checking the MergeMems for
1047 // new states for the slice of this load. Stop once any other
1048 // kind of node is encountered. Loads from final memory can skip
1049 // through any kind of MemBar but normal loads shouldn't skip
1050 // through MemBarAcquire since the could allow them to move out of
1051 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1052 // because alias info above them may be inaccurate (e.g., due to
1053 // mixed/mismatched unsafe accesses).
1054 bool is_final_mem = !atp->is_rewritable();
1055 while (current->is_Proj()) {
1056 int opc = current->in(0)->Opcode();
1057 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1058 opc == Op_MemBarAcquireLock ||
1059 opc == Op_LoadFence)) ||
1060 opc == Op_MemBarRelease ||
1061 opc == Op_StoreFence ||
1062 opc == Op_MemBarReleaseLock ||
1063 opc == Op_MemBarStoreStore ||
1064 opc == Op_StoreStoreFence) {
1065 Node* mem = current->in(0)->in(TypeFunc::Memory);
1066 if (mem->is_MergeMem()) {
1067 MergeMemNode* merge = mem->as_MergeMem();
1068 Node* new_st = merge->memory_at(alias_idx);
1069 if (new_st == merge->base_memory()) {
1070 // Keep searching
1071 current = new_st;
1072 continue;
1073 }
1074 // Save the new memory state for the slice and fall through
1075 // to exit.
1076 result = new_st;
1077 }
1078 }
1079 break;
1080 }
1081 if (result != NULL) {
1082 st = result;
1083 }
1084 }
1085
1086 // Loop around twice in the case Load -> Initialize -> Store.
1087 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1088 for (int trip = 0; trip <= 1; trip++) {
1089
1090 if (st->is_Store()) {
1091 Node* st_adr = st->in(MemNode::Address);
1092 if (st_adr != ld_adr) {
1093 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1094 intptr_t st_off = 0;
1095 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1096 if (ld_base == NULL) return NULL;
1097 if (st_base == NULL) return NULL;
1098 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL;
1099 if (ld_off != st_off) return NULL;
1100 if (ld_off == Type::OffsetBot) return NULL;
1101 // Same base, same offset.
1102 // Possible improvement for arrays: check index value instead of absolute offset.
1103
1104 // At this point we have proven something like this setup:
1105 // B = << base >>
1106 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1107 // S = StoreQ(AddP( B , #Off), V)
1108 // (Actually, we haven't yet proven the Q's are the same.)
1109 // In other words, we are loading from a casted version of
1110 // the same pointer-and-offset that we stored to.
1111 // Casted version may carry a dependency and it is respected.
1112 // Thus, we are able to replace L by V.
1113 }
1114 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1115 if (store_Opcode() != st->Opcode()) {
1116 return NULL;
1117 }
1118 // LoadVector/StoreVector needs additional check to ensure the types match.
1119 if (st->is_StoreVector()) {
1120 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1121 const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1122 if (in_vt != out_vt) {
1123 return NULL;
1124 }
1125 }
1126 return st->in(MemNode::ValueIn);
1127 }
1128
1129 // A load from a freshly-created object always returns zero.
1130 // (This can happen after LoadNode::Ideal resets the load's memory input
1131 // to find_captured_store, which returned InitializeNode::zero_memory.)
1132 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1133 (st->in(0) == ld_alloc) &&
1134 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1135 // return a zero value for the load's basic type
1136 // (This is one of the few places where a generic PhaseTransform
1137 // can create new nodes. Think of it as lazily manifesting
1138 // virtually pre-existing constants.)
1139 assert(memory_type() != T_PRIMITIVE_OBJECT, "should not be used for inline types");
1140 Node* default_value = ld_alloc->in(AllocateNode::DefaultValue);
1141 if (default_value != NULL) {
1142 return default_value;
1143 }
1144 assert(ld_alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
1145 if (memory_type() != T_VOID) {
1146 if (ReduceBulkZeroing || find_array_copy_clone(phase, ld_alloc, in(MemNode::Memory)) == NULL) {
1147 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1148 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1149 // by the ArrayCopyNode.
1150 return phase->zerocon(memory_type());
1151 }
1152 } else {
1153 // TODO: materialize all-zero vector constant
1154 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1155 }
1156 }
1157
1158 // A load from an initialization barrier can match a captured store.
1159 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1160 InitializeNode* init = st->in(0)->as_Initialize();
1161 AllocateNode* alloc = init->allocation();
1162 if ((alloc != NULL) && (alloc == ld_alloc)) {
1163 // examine a captured store value
1164 st = init->find_captured_store(ld_off, memory_size(), phase);
1165 if (st != NULL) {
1166 continue; // take one more trip around
1167 }
1168 }
1169 }
1170
1171 // Load boxed value from result of valueOf() call is input parameter.
1172 if (this->is_Load() && ld_adr->is_AddP() &&
1173 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1174 intptr_t ignore = 0;
1175 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1176 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1177 base = bs->step_over_gc_barrier(base);
1178 if (base != NULL && base->is_Proj() &&
1179 base->as_Proj()->_con == TypeFunc::Parms &&
1180 base->in(0)->is_CallStaticJava() &&
1181 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1182 return base->in(0)->in(TypeFunc::Parms);
1183 }
1184 }
1185
1186 break;
1187 }
1188
1189 return NULL;
1190 }
1191
1192 //----------------------is_instance_field_load_with_local_phi------------------
1193 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1194 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1195 in(Address)->is_AddP() ) {
1196 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1197 // Only instances and boxed values.
1198 if( t_oop != NULL &&
1199 (t_oop->is_ptr_to_boxed_value() ||
1200 t_oop->is_known_instance_field()) &&
1201 t_oop->offset() != Type::OffsetBot &&
1202 t_oop->offset() != Type::OffsetTop) {
1203 return true;
1204 }
1205 }
1206 return false;
1207 }
1208
1209 //------------------------------Identity---------------------------------------
1210 // Loads are identity if previous store is to same address
1211 Node* LoadNode::Identity(PhaseGVN* phase) {
1212 // Loading from an InlineTypePtr? The InlineTypePtr has the values of
1213 // all fields as input. Look for the field with matching offset.
1214 Node* addr = in(Address);
1215 intptr_t offset;
1216 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset);
1217 if (base != NULL && base->is_InlineTypePtr() && offset > oopDesc::klass_offset_in_bytes()) {
1218 Node* value = base->as_InlineTypePtr()->field_value_by_offset((int)offset, true);
1219 if (value->is_InlineType()) {
1220 // Non-flattened inline type field
1221 InlineTypeNode* vt = value->as_InlineType();
1222 if (vt->is_allocated(phase)) {
1223 value = vt->get_oop();
1224 } else {
1225 // Not yet allocated, bail out
1226 value = NULL;
1227 }
1228 }
1229 if (value != NULL) {
1230 if (Opcode() == Op_LoadN) {
1231 // Encode oop value if we are loading a narrow oop
1232 assert(!phase->type(value)->isa_narrowoop(), "should already be decoded");
1233 value = phase->transform(new EncodePNode(value, bottom_type()));
1234 }
1235 return value;
1236 }
1237 }
1238
1239 // If the previous store-maker is the right kind of Store, and the store is
1240 // to the same address, then we are equal to the value stored.
1241 Node* mem = in(Memory);
1242 Node* value = can_see_stored_value(mem, phase);
1243 if( value ) {
1244 // byte, short & char stores truncate naturally.
1245 // A load has to load the truncated value which requires
1246 // some sort of masking operation and that requires an
1247 // Ideal call instead of an Identity call.
1248 if (memory_size() < BytesPerInt) {
1249 // If the input to the store does not fit with the load's result type,
1250 // it must be truncated via an Ideal call.
1251 if (!phase->type(value)->higher_equal(phase->type(this)))
1252 return this;
1253 }
1254 // (This works even when value is a Con, but LoadNode::Value
1255 // usually runs first, producing the singleton type of the Con.)
1256 return value;
1257 }
1258
1259 // Search for an existing data phi which was generated before for the same
1260 // instance's field to avoid infinite generation of phis in a loop.
1261 Node *region = mem->in(0);
1262 if (is_instance_field_load_with_local_phi(region)) {
1263 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1264 int this_index = phase->C->get_alias_index(addr_t);
1265 int this_offset = addr_t->offset();
1266 int this_iid = addr_t->instance_id();
1267 if (!addr_t->is_known_instance() &&
1268 addr_t->is_ptr_to_boxed_value()) {
1269 // Use _idx of address base (could be Phi node) for boxed values.
1270 intptr_t ignore = 0;
1271 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1272 if (base == NULL) {
1273 return this;
1274 }
1275 this_iid = base->_idx;
1276 }
1277 const Type* this_type = bottom_type();
1278 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1279 Node* phi = region->fast_out(i);
1280 if (phi->is_Phi() && phi != mem &&
1281 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1282 return phi;
1283 }
1284 }
1285 }
1286
1287 return this;
1288 }
1289
1290 // Construct an equivalent unsigned load.
1291 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1292 BasicType bt = T_ILLEGAL;
1293 const Type* rt = NULL;
1294 switch (Opcode()) {
1295 case Op_LoadUB: return this;
1296 case Op_LoadUS: return this;
1297 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1298 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1299 default:
1300 assert(false, "no unsigned variant: %s", Name());
1301 return NULL;
1302 }
1303 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1304 raw_adr_type(), rt, bt, _mo, _control_dependency,
1305 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1306 }
1307
1308 // Construct an equivalent signed load.
1309 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1310 BasicType bt = T_ILLEGAL;
1311 const Type* rt = NULL;
1312 switch (Opcode()) {
1313 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1314 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1315 case Op_LoadB: // fall through
1316 case Op_LoadS: // fall through
1317 case Op_LoadI: // fall through
1318 case Op_LoadL: return this;
1319 default:
1320 assert(false, "no signed variant: %s", Name());
1321 return NULL;
1322 }
1323 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1324 raw_adr_type(), rt, bt, _mo, _control_dependency,
1325 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1326 }
1327
1328 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1329 BasicType bt = rt->basic_type();
1330 switch (Opcode()) {
1331 case Op_LoadI: return (bt == T_FLOAT);
1332 case Op_LoadL: return (bt == T_DOUBLE);
1333 case Op_LoadF: return (bt == T_INT);
1334 case Op_LoadD: return (bt == T_LONG);
1335
1336 default: return false;
1337 }
1338 }
1339
1340 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1341 BasicType bt = rt->basic_type();
1342 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1343 bool is_mismatched = is_mismatched_access();
1344 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1345 if (raw_type == NULL) {
1346 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1347 }
1348 const int op = Opcode();
1349 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1350 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1351 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1352 raw_adr_type(), rt, bt, _mo, _control_dependency,
1353 require_atomic_access, is_unaligned_access(), is_mismatched);
1354 }
1355
1356 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1357 BasicType bt = vt->basic_type();
1358 switch (Opcode()) {
1359 case Op_StoreI: return (bt == T_FLOAT);
1360 case Op_StoreL: return (bt == T_DOUBLE);
1361 case Op_StoreF: return (bt == T_INT);
1362 case Op_StoreD: return (bt == T_LONG);
1363
1364 default: return false;
1365 }
1366 }
1367
1368 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1369 BasicType bt = vt->basic_type();
1370 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1371 const int op = Opcode();
1372 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1373 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1374 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1375 raw_adr_type(), val, bt, _mo, require_atomic_access);
1376
1377 bool is_mismatched = is_mismatched_access();
1378 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1379 if (raw_type == NULL) {
1380 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1381 }
1382 if (is_mismatched) {
1383 st->set_mismatched_access();
1384 }
1385 return st;
1386 }
1387
1388 // We're loading from an object which has autobox behaviour.
1389 // If this object is result of a valueOf call we'll have a phi
1390 // merging a newly allocated object and a load from the cache.
1391 // We want to replace this load with the original incoming
1392 // argument to the valueOf call.
1393 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1394 assert(igvn->C->eliminate_boxing(), "sanity");
1395 intptr_t ignore = 0;
1396 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1397 if ((base == NULL) || base->is_Phi()) {
1398 // Push the loads from the phi that comes from valueOf up
1399 // through it to allow elimination of the loads and the recovery
1400 // of the original value. It is done in split_through_phi().
1401 return NULL;
1402 } else if (base->is_Load() ||
1403 (base->is_DecodeN() && base->in(1)->is_Load())) {
1404 // Eliminate the load of boxed value for integer types from the cache
1405 // array by deriving the value from the index into the array.
1406 // Capture the offset of the load and then reverse the computation.
1407
1408 // Get LoadN node which loads a boxing object from 'cache' array.
1409 if (base->is_DecodeN()) {
1410 base = base->in(1);
1411 }
1412 if (!base->in(Address)->is_AddP()) {
1413 return NULL; // Complex address
1414 }
1415 AddPNode* address = base->in(Address)->as_AddP();
1416 Node* cache_base = address->in(AddPNode::Base);
1417 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1418 // Get ConP node which is static 'cache' field.
1419 cache_base = cache_base->in(1);
1420 }
1421 if ((cache_base != NULL) && cache_base->is_Con()) {
1422 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1423 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1424 Node* elements[4];
1425 int shift = exact_log2(type2aelembytes(T_OBJECT));
1426 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1427 if (count > 0 && elements[0]->is_Con() &&
1428 (count == 1 ||
1429 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1430 elements[1]->in(2) == igvn->intcon(shift)))) {
1431 ciObjArray* array = base_type->const_oop()->as_obj_array();
1432 // Fetch the box object cache[0] at the base of the array and get its value
1433 ciInstance* box = array->obj_at(0)->as_instance();
1434 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1435 assert(ik->is_box_klass(), "sanity");
1436 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1437 if (ik->nof_nonstatic_fields() == 1) {
1438 // This should be true nonstatic_field_at requires calling
1439 // nof_nonstatic_fields so check it anyway
1440 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1441 BasicType bt = c.basic_type();
1442 // Only integer types have boxing cache.
1443 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1444 bt == T_BYTE || bt == T_SHORT ||
1445 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1446 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1447 if (cache_low != (int)cache_low) {
1448 return NULL; // should not happen since cache is array indexed by value
1449 }
1450 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1451 if (offset != (int)offset) {
1452 return NULL; // should not happen since cache is array indexed by value
1453 }
1454 // Add up all the offsets making of the address of the load
1455 Node* result = elements[0];
1456 for (int i = 1; i < count; i++) {
1457 result = igvn->transform(new AddXNode(result, elements[i]));
1458 }
1459 // Remove the constant offset from the address and then
1460 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1461 // remove the scaling of the offset to recover the original index.
1462 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1463 // Peel the shift off directly but wrap it in a dummy node
1464 // since Ideal can't return existing nodes
1465 igvn->_worklist.push(result); // remove dead node later
1466 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1467 } else if (result->is_Add() && result->in(2)->is_Con() &&
1468 result->in(1)->Opcode() == Op_LShiftX &&
1469 result->in(1)->in(2) == igvn->intcon(shift)) {
1470 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1471 // but for boxing cache access we know that X<<Z will not overflow
1472 // (there is range check) so we do this optimizatrion by hand here.
1473 igvn->_worklist.push(result); // remove dead node later
1474 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1475 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1476 } else {
1477 result = new RShiftXNode(result, igvn->intcon(shift));
1478 }
1479 #ifdef _LP64
1480 if (bt != T_LONG) {
1481 result = new ConvL2INode(igvn->transform(result));
1482 }
1483 #else
1484 if (bt == T_LONG) {
1485 result = new ConvI2LNode(igvn->transform(result));
1486 }
1487 #endif
1488 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1489 // Need to preserve unboxing load type if it is unsigned.
1490 switch(this->Opcode()) {
1491 case Op_LoadUB:
1492 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1493 break;
1494 case Op_LoadUS:
1495 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1496 break;
1497 }
1498 return result;
1499 }
1500 }
1501 }
1502 }
1503 }
1504 return NULL;
1505 }
1506
1507 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1508 Node* region = phi->in(0);
1509 if (region == NULL) {
1510 return false; // Wait stable graph
1511 }
1512 uint cnt = phi->req();
1513 for (uint i = 1; i < cnt; i++) {
1514 Node* rc = region->in(i);
1515 if (rc == NULL || phase->type(rc) == Type::TOP)
1516 return false; // Wait stable graph
1517 Node* in = phi->in(i);
1518 if (in == NULL || phase->type(in) == Type::TOP)
1519 return false; // Wait stable graph
1520 }
1521 return true;
1522 }
1523 //------------------------------split_through_phi------------------------------
1524 // Split instance or boxed field load through Phi.
1525 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1526 Node* mem = in(Memory);
1527 Node* address = in(Address);
1528 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1529
1530 assert((t_oop != NULL) &&
1531 (t_oop->is_known_instance_field() ||
1532 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1533
1534 Compile* C = phase->C;
1535 intptr_t ignore = 0;
1536 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1537 bool base_is_phi = (base != NULL) && base->is_Phi();
1538 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1539 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1540 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1541
1542 if (!((mem->is_Phi() || base_is_phi) &&
1543 (load_boxed_values || t_oop->is_known_instance_field()))) {
1544 return NULL; // memory is not Phi
1545 }
1546
1547 if (mem->is_Phi()) {
1548 if (!stable_phi(mem->as_Phi(), phase)) {
1549 return NULL; // Wait stable graph
1550 }
1551 uint cnt = mem->req();
1552 // Check for loop invariant memory.
1553 if (cnt == 3) {
1554 for (uint i = 1; i < cnt; i++) {
1555 Node* in = mem->in(i);
1556 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1557 if (m == mem) {
1558 if (i == 1) {
1559 // if the first edge was a loop, check second edge too.
1560 // If both are replaceable - we are in an infinite loop
1561 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1562 if (n == mem) {
1563 break;
1564 }
1565 }
1566 set_req(Memory, mem->in(cnt - i));
1567 return this; // made change
1568 }
1569 }
1570 }
1571 }
1572 if (base_is_phi) {
1573 if (!stable_phi(base->as_Phi(), phase)) {
1574 return NULL; // Wait stable graph
1575 }
1576 uint cnt = base->req();
1577 // Check for loop invariant memory.
1578 if (cnt == 3) {
1579 for (uint i = 1; i < cnt; i++) {
1580 if (base->in(i) == base) {
1581 return NULL; // Wait stable graph
1582 }
1583 }
1584 }
1585 }
1586
1587 // Split through Phi (see original code in loopopts.cpp).
1588 assert(C->have_alias_type(t_oop), "instance should have alias type");
1589
1590 // Do nothing here if Identity will find a value
1591 // (to avoid infinite chain of value phis generation).
1592 if (this != Identity(phase)) {
1593 return NULL;
1594 }
1595
1596 // Select Region to split through.
1597 Node* region;
1598 if (!base_is_phi) {
1599 assert(mem->is_Phi(), "sanity");
1600 region = mem->in(0);
1601 // Skip if the region dominates some control edge of the address.
1602 if (!MemNode::all_controls_dominate(address, region))
1603 return NULL;
1604 } else if (!mem->is_Phi()) {
1605 assert(base_is_phi, "sanity");
1606 region = base->in(0);
1607 // Skip if the region dominates some control edge of the memory.
1608 if (!MemNode::all_controls_dominate(mem, region))
1609 return NULL;
1610 } else if (base->in(0) != mem->in(0)) {
1611 assert(base_is_phi && mem->is_Phi(), "sanity");
1612 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1613 region = base->in(0);
1614 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1615 region = mem->in(0);
1616 } else {
1617 return NULL; // complex graph
1618 }
1619 } else {
1620 assert(base->in(0) == mem->in(0), "sanity");
1621 region = mem->in(0);
1622 }
1623
1624 const Type* this_type = this->bottom_type();
1625 int this_index = C->get_alias_index(t_oop);
1626 int this_offset = t_oop->offset();
1627 int this_iid = t_oop->instance_id();
1628 if (!t_oop->is_known_instance() && load_boxed_values) {
1629 // Use _idx of address base for boxed values.
1630 this_iid = base->_idx;
1631 }
1632 PhaseIterGVN* igvn = phase->is_IterGVN();
1633 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1634 for (uint i = 1; i < region->req(); i++) {
1635 Node* x;
1636 Node* the_clone = NULL;
1637 Node* in = region->in(i);
1638 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1639 in != NULL && in->is_OuterStripMinedLoop()) {
1640 // No node should go in the outer strip mined loop
1641 in = in->in(LoopNode::EntryControl);
1642 }
1643 if (in == NULL || in == C->top()) {
1644 x = C->top(); // Dead path? Use a dead data op
1645 } else {
1646 x = this->clone(); // Else clone up the data op
1647 the_clone = x; // Remember for possible deletion.
1648 // Alter data node to use pre-phi inputs
1649 if (this->in(0) == region) {
1650 x->set_req(0, in);
1651 } else {
1652 x->set_req(0, NULL);
1653 }
1654 if (mem->is_Phi() && (mem->in(0) == region)) {
1655 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1656 }
1657 if (address->is_Phi() && address->in(0) == region) {
1658 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1659 }
1660 if (base_is_phi && (base->in(0) == region)) {
1661 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1662 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1663 x->set_req(Address, adr_x);
1664 }
1665 }
1666 // Check for a 'win' on some paths
1667 const Type *t = x->Value(igvn);
1668
1669 bool singleton = t->singleton();
1670
1671 // See comments in PhaseIdealLoop::split_thru_phi().
1672 if (singleton && t == Type::TOP) {
1673 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1674 }
1675
1676 if (singleton) {
1677 x = igvn->makecon(t);
1678 } else {
1679 // We now call Identity to try to simplify the cloned node.
1680 // Note that some Identity methods call phase->type(this).
1681 // Make sure that the type array is big enough for
1682 // our new node, even though we may throw the node away.
1683 // (This tweaking with igvn only works because x is a new node.)
1684 igvn->set_type(x, t);
1685 // If x is a TypeNode, capture any more-precise type permanently into Node
1686 // otherwise it will be not updated during igvn->transform since
1687 // igvn->type(x) is set to x->Value() already.
1688 x->raise_bottom_type(t);
1689 Node* y = x->Identity(igvn);
1690 if (y != x) {
1691 x = y;
1692 } else {
1693 y = igvn->hash_find_insert(x);
1694 if (y) {
1695 x = y;
1696 } else {
1697 // Else x is a new node we are keeping
1698 // We do not need register_new_node_with_optimizer
1699 // because set_type has already been called.
1700 igvn->_worklist.push(x);
1701 }
1702 }
1703 }
1704 if (x != the_clone && the_clone != NULL) {
1705 igvn->remove_dead_node(the_clone);
1706 }
1707 phi->set_req(i, x);
1708 }
1709 // Record Phi
1710 igvn->register_new_node_with_optimizer(phi);
1711 return phi;
1712 }
1713
1714 AllocateNode* LoadNode::is_new_object_mark_load(PhaseGVN *phase) const {
1715 if (Opcode() == Op_LoadX) {
1716 Node* address = in(MemNode::Address);
1717 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1718 Node* mem = in(MemNode::Memory);
1719 if (alloc != NULL && mem->is_Proj() &&
1720 mem->in(0) != NULL &&
1721 mem->in(0) == alloc->initialization() &&
1722 alloc->initialization()->proj_out_or_null(0) != NULL) {
1723 return alloc;
1724 }
1725 }
1726 return NULL;
1727 }
1728
1729
1730 //------------------------------Ideal------------------------------------------
1731 // If the load is from Field memory and the pointer is non-null, it might be possible to
1732 // zero out the control input.
1733 // If the offset is constant and the base is an object allocation,
1734 // try to hook me up to the exact initializing store.
1735 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1736 Node* p = MemNode::Ideal_common(phase, can_reshape);
1737 if (p) return (p == NodeSentinel) ? NULL : p;
1738
1739 Node* ctrl = in(MemNode::Control);
1740 Node* address = in(MemNode::Address);
1741 bool progress = false;
1742
1743 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1744 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1745
1746 // Skip up past a SafePoint control. Cannot do this for Stores because
1747 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1748 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1749 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1750 !addr_mark &&
1751 (depends_only_on_test() || has_unknown_control_dependency())) {
1752 ctrl = ctrl->in(0);
1753 set_req(MemNode::Control,ctrl);
1754 progress = true;
1755 }
1756
1757 intptr_t ignore = 0;
1758 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1759 if (base != NULL
1760 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1761 // Check for useless control edge in some common special cases
1762 if (in(MemNode::Control) != NULL
1763 && can_remove_control()
1764 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1765 && all_controls_dominate(base, phase->C->start())) {
1766 // A method-invariant, non-null address (constant or 'this' argument).
1767 set_req(MemNode::Control, NULL);
1768 progress = true;
1769 }
1770 }
1771
1772 Node* mem = in(MemNode::Memory);
1773 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1774
1775 if (can_reshape && (addr_t != NULL)) {
1776 // try to optimize our memory input
1777 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1778 if (opt_mem != mem) {
1779 set_req_X(MemNode::Memory, opt_mem, phase);
1780 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1781 return this;
1782 }
1783 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1784 if ((t_oop != NULL) &&
1785 (t_oop->is_known_instance_field() ||
1786 t_oop->is_ptr_to_boxed_value())) {
1787 PhaseIterGVN *igvn = phase->is_IterGVN();
1788 assert(igvn != NULL, "must be PhaseIterGVN when can_reshape is true");
1789 if (igvn->_worklist.member(opt_mem)) {
1790 // Delay this transformation until memory Phi is processed.
1791 igvn->_worklist.push(this);
1792 return NULL;
1793 }
1794 // Split instance field load through Phi.
1795 Node* result = split_through_phi(phase);
1796 if (result != NULL) return result;
1797
1798 if (t_oop->is_ptr_to_boxed_value()) {
1799 Node* result = eliminate_autobox(igvn);
1800 if (result != NULL) return result;
1801 }
1802 }
1803 }
1804
1805 // Is there a dominating load that loads the same value? Leave
1806 // anything that is not a load of a field/array element (like
1807 // barriers etc.) alone
1808 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1809 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1810 Node *use = mem->fast_out(i);
1811 if (use != this &&
1812 use->Opcode() == Opcode() &&
1813 use->in(0) != NULL &&
1814 use->in(0) != in(0) &&
1815 use->in(Address) == in(Address)) {
1816 Node* ctl = in(0);
1817 for (int i = 0; i < 10 && ctl != NULL; i++) {
1818 ctl = IfNode::up_one_dom(ctl);
1819 if (ctl == use->in(0)) {
1820 set_req(0, use->in(0));
1821 return this;
1822 }
1823 }
1824 }
1825 }
1826 }
1827
1828 // Check for prior store with a different base or offset; make Load
1829 // independent. Skip through any number of them. Bail out if the stores
1830 // are in an endless dead cycle and report no progress. This is a key
1831 // transform for Reflection. However, if after skipping through the Stores
1832 // we can't then fold up against a prior store do NOT do the transform as
1833 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1834 // array memory alive twice: once for the hoisted Load and again after the
1835 // bypassed Store. This situation only works if EVERYBODY who does
1836 // anti-dependence work knows how to bypass. I.e. we need all
1837 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1838 // the alias index stuff. So instead, peek through Stores and IFF we can
1839 // fold up, do so.
1840 Node* prev_mem = find_previous_store(phase);
1841 if (prev_mem != NULL) {
1842 Node* value = can_see_arraycopy_value(prev_mem, phase);
1843 if (value != NULL) {
1844 return value;
1845 }
1846 }
1847 // Steps (a), (b): Walk past independent stores to find an exact match.
1848 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1849 // (c) See if we can fold up on the spot, but don't fold up here.
1850 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1851 // just return a prior value, which is done by Identity calls.
1852 if (can_see_stored_value(prev_mem, phase)) {
1853 // Make ready for step (d):
1854 set_req_X(MemNode::Memory, prev_mem, phase);
1855 return this;
1856 }
1857 }
1858
1859 return progress ? this : NULL;
1860 }
1861
1862 // Helper to recognize certain Klass fields which are invariant across
1863 // some group of array types (e.g., int[] or all T[] where T < Object).
1864 const Type*
1865 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1866 ciKlass* klass) const {
1867 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1868 // The field is Klass::_modifier_flags. Return its (constant) value.
1869 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1870 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1871 return TypeInt::make(klass->modifier_flags());
1872 }
1873 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1874 // The field is Klass::_access_flags. Return its (constant) value.
1875 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1876 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1877 return TypeInt::make(klass->access_flags());
1878 }
1879 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1880 // The field is Klass::_layout_helper. Return its constant value if known.
1881 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1882 return TypeInt::make(klass->layout_helper());
1883 }
1884
1885 // No match.
1886 return NULL;
1887 }
1888
1889 //------------------------------Value-----------------------------------------
1890 const Type* LoadNode::Value(PhaseGVN* phase) const {
1891 // Either input is TOP ==> the result is TOP
1892 Node* mem = in(MemNode::Memory);
1893 const Type *t1 = phase->type(mem);
1894 if (t1 == Type::TOP) return Type::TOP;
1895 Node* adr = in(MemNode::Address);
1896 const TypePtr* tp = phase->type(adr)->isa_ptr();
1897 if (tp == NULL || tp->empty()) return Type::TOP;
1898 int off = tp->offset();
1899 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1900 Compile* C = phase->C;
1901
1902 // Try to guess loaded type from pointer type
1903 if (tp->isa_aryptr()) {
1904 const TypeAryPtr* ary = tp->is_aryptr();
1905 const Type* t = ary->elem();
1906
1907 // Determine whether the reference is beyond the header or not, by comparing
1908 // the offset against the offset of the start of the array's data.
1909 // Different array types begin at slightly different offsets (12 vs. 16).
1910 // We choose T_BYTE as an example base type that is least restrictive
1911 // as to alignment, which will therefore produce the smallest
1912 // possible base offset.
1913 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1914 const bool off_beyond_header = (off >= min_base_off);
1915
1916 // Try to constant-fold a stable array element.
1917 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1918 // Make sure the reference is not into the header and the offset is constant
1919 ciObject* aobj = ary->const_oop();
1920 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1921 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1922 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1923 stable_dimension,
1924 memory_type(), is_unsigned());
1925 if (con_type != NULL) {
1926 return con_type;
1927 }
1928 }
1929 }
1930
1931 // Don't do this for integer types. There is only potential profit if
1932 // the element type t is lower than _type; that is, for int types, if _type is
1933 // more restrictive than t. This only happens here if one is short and the other
1934 // char (both 16 bits), and in those cases we've made an intentional decision
1935 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1936 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1937 //
1938 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1939 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1940 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1941 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1942 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1943 // In fact, that could have been the original type of p1, and p1 could have
1944 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1945 // expression (LShiftL quux 3) independently optimized to the constant 8.
1946 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1947 && (_type->isa_vect() == NULL)
1948 && t->isa_inlinetype() == NULL
1949 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1950 // t might actually be lower than _type, if _type is a unique
1951 // concrete subclass of abstract class t.
1952 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
1953 const Type* jt = t->join_speculative(_type);
1954 // In any case, do not allow the join, per se, to empty out the type.
1955 if (jt->empty() && !t->empty()) {
1956 // This can happen if a interface-typed array narrows to a class type.
1957 jt = _type;
1958 }
1959 #ifdef ASSERT
1960 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1961 // The pointers in the autobox arrays are always non-null
1962 Node* base = adr->in(AddPNode::Base);
1963 if ((base != NULL) && base->is_DecodeN()) {
1964 // Get LoadN node which loads IntegerCache.cache field
1965 base = base->in(1);
1966 }
1967 if ((base != NULL) && base->is_Con()) {
1968 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1969 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1970 // It could be narrow oop
1971 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1972 }
1973 }
1974 }
1975 #endif
1976 return jt;
1977 }
1978 }
1979 } else if (tp->base() == Type::InstPtr) {
1980 assert( off != Type::OffsetBot ||
1981 // arrays can be cast to Objects
1982 !tp->isa_instptr() ||
1983 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
1984 // Default value load
1985 tp->is_instptr()->instance_klass() == ciEnv::current()->Class_klass() ||
1986 // unsafe field access may not have a constant offset
1987 C->has_unsafe_access(),
1988 "Field accesses must be precise" );
1989 // For oop loads, we expect the _type to be precise.
1990
1991 const TypeInstPtr* tinst = tp->is_instptr();
1992 BasicType bt = memory_type();
1993
1994 // Optimize loads from constant fields.
1995 ciObject* const_oop = tinst->const_oop();
1996 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1997 ciType* mirror_type = const_oop->as_instance()->java_mirror_type();
1998 if (mirror_type != NULL) {
1999 const Type* const_oop = NULL;
2000 ciInlineKlass* vk = mirror_type->is_inlinetype() ? mirror_type->as_inline_klass() : NULL;
2001 // Fold default value loads
2002 if (vk != NULL && off == vk->default_value_offset()) {
2003 const_oop = TypeInstPtr::make(vk->default_instance());
2004 }
2005 // Fold class mirror loads
2006 if (off == java_lang_Class::primary_mirror_offset()) {
2007 const_oop = (vk == NULL) ? TypePtr::NULL_PTR : TypeInstPtr::make(vk->ref_instance());
2008 } else if (off == java_lang_Class::secondary_mirror_offset()) {
2009 const_oop = (vk == NULL) ? TypePtr::NULL_PTR : TypeInstPtr::make(vk->val_instance());
2010 }
2011 if (const_oop != NULL) {
2012 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop;
2013 }
2014 }
2015 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
2016 if (con_type != NULL) {
2017 return con_type;
2018 }
2019 }
2020 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2021 assert(off != Type::OffsetBot ||
2022 !tp->isa_instklassptr() ||
2023 // arrays can be cast to Objects
2024 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2025 // also allow array-loading from the primary supertype
2026 // array during subtype checks
2027 Opcode() == Op_LoadKlass,
2028 "Field accesses must be precise");
2029 // For klass/static loads, we expect the _type to be precise
2030 } else if (tp->base() == Type::RawPtr && !StressReflectiveCode) {
2031 if (adr->is_Load() && off == 0) {
2032 /* With mirrors being an indirect in the Klass*
2033 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2034 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2035 *
2036 * So check the type and klass of the node before the LoadP.
2037 */
2038 Node* adr2 = adr->in(MemNode::Address);
2039 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2040 if (tkls != NULL) {
2041 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2042 ciKlass* klass = tkls->exact_klass();
2043 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2044 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2045 return TypeInstPtr::make(klass->java_mirror());
2046 }
2047 }
2048 } else {
2049 // Check for a load of the default value offset from the InlineKlassFixedBlock:
2050 // LoadI(LoadP(inline_klass, adr_inlineklass_fixed_block_offset), default_value_offset_offset)
2051 intptr_t offset = 0;
2052 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2053 if (base != NULL && base->is_Load() && offset == in_bytes(InlineKlass::default_value_offset_offset())) {
2054 const TypeKlassPtr* tkls = phase->type(base->in(MemNode::Address))->isa_klassptr();
2055 if (tkls != NULL && tkls->is_loaded() && tkls->klass_is_exact() && tkls->exact_klass()->is_inlinetype() &&
2056 tkls->offset() == in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset())) {
2057 assert(base->Opcode() == Op_LoadP, "must load an oop from klass");
2058 assert(Opcode() == Op_LoadI, "must load an int from fixed block");
2059 return TypeInt::make(tkls->exact_klass()->as_inline_klass()->default_value_offset());
2060 }
2061 }
2062 }
2063 }
2064
2065 const TypeKlassPtr *tkls = tp->isa_klassptr();
2066 if (tkls != NULL && !StressReflectiveCode) {
2067 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2068 ciKlass* klass = tkls->exact_klass();
2069 // We are loading a field from a Klass metaobject whose identity
2070 // is known at compile time (the type is "exact" or "precise").
2071 // Check for fields we know are maintained as constants by the VM.
2072 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2073 // The field is Klass::_super_check_offset. Return its (constant) value.
2074 // (Folds up type checking code.)
2075 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2076 return TypeInt::make(klass->super_check_offset());
2077 }
2078 // Compute index into primary_supers array
2079 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2080 // Check for overflowing; use unsigned compare to handle the negative case.
2081 if( depth < ciKlass::primary_super_limit() ) {
2082 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2083 // (Folds up type checking code.)
2084 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2085 ciKlass *ss = klass->super_of_depth(depth);
2086 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2087 }
2088 const Type* aift = load_array_final_field(tkls, klass);
2089 if (aift != NULL) return aift;
2090 }
2091
2092 // We can still check if we are loading from the primary_supers array at a
2093 // shallow enough depth. Even though the klass is not exact, entries less
2094 // than or equal to its super depth are correct.
2095 if (tkls->is_loaded()) {
2096 ciKlass* klass = NULL;
2097 if (tkls->isa_instklassptr()) {
2098 klass = tkls->is_instklassptr()->instance_klass();
2099 } else {
2100 int dims;
2101 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2102 if (inner->isa_instklassptr()) {
2103 klass = inner->is_instklassptr()->instance_klass();
2104 klass = ciObjArrayKlass::make(klass, dims);
2105 }
2106 }
2107 if (klass != NULL) {
2108 // Compute index into primary_supers array
2109 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2110 // Check for overflowing; use unsigned compare to handle the negative case.
2111 if (depth < ciKlass::primary_super_limit() &&
2112 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2113 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2114 // (Folds up type checking code.)
2115 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2116 ciKlass *ss = klass->super_of_depth(depth);
2117 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2118 }
2119 }
2120 }
2121
2122 // If the type is enough to determine that the thing is not an array,
2123 // we can give the layout_helper a positive interval type.
2124 // This will help short-circuit some reflective code.
2125 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2126 tkls->isa_instklassptr() && // not directly typed as an array
2127 !tkls->is_instklassptr()->instance_klass()->is_java_lang_Object() // not the supertype of all T[] and specifically not Serializable & Cloneable
2128 ) {
2129 // Note: When interfaces are reliable, we can narrow the interface
2130 // test to (klass != Serializable && klass != Cloneable).
2131 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2132 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2133 // The key property of this type is that it folds up tests
2134 // for array-ness, since it proves that the layout_helper is positive.
2135 // Thus, a generic value like the basic object layout helper works fine.
2136 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2137 }
2138 }
2139
2140 // If we are loading from a freshly-allocated object, produce a zero,
2141 // if the load is provably beyond the header of the object.
2142 // (Also allow a variable load from a fresh array to produce zero.)
2143 const TypeOopPtr *tinst = tp->isa_oopptr();
2144 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
2145 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
2146 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2147 Node* value = can_see_stored_value(mem,phase);
2148 if (value != NULL && value->is_Con()) {
2149 assert(value->bottom_type()->higher_equal(_type),"sanity");
2150 return value->bottom_type();
2151 }
2152 }
2153
2154 bool is_vect = (_type->isa_vect() != NULL);
2155 if (is_instance && !is_vect) {
2156 // If we have an instance type and our memory input is the
2157 // programs's initial memory state, there is no matching store,
2158 // so just return a zero of the appropriate type -
2159 // except if it is vectorized - then we have no zero constant.
2160 Node *mem = in(MemNode::Memory);
2161 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2162 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2163 return Type::get_zero_type(_type->basic_type());
2164 }
2165 }
2166 Node* alloc = is_new_object_mark_load(phase);
2167 if (alloc != NULL) {
2168 if (EnableValhalla) {
2169 // The mark word may contain property bits (inline, flat, null-free)
2170 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2171 const TypeKlassPtr* tkls = phase->type(klass_node)->is_klassptr();
2172 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2173 return TypeX::make(tkls->exact_klass()->prototype_header().value());
2174 }
2175 } else {
2176 return TypeX::make(markWord::prototype().value());
2177 }
2178 }
2179
2180 return _type;
2181 }
2182
2183 //------------------------------match_edge-------------------------------------
2184 // Do we Match on this edge index or not? Match only the address.
2185 uint LoadNode::match_edge(uint idx) const {
2186 return idx == MemNode::Address;
2187 }
2188
2189 //--------------------------LoadBNode::Ideal--------------------------------------
2190 //
2191 // If the previous store is to the same address as this load,
2192 // and the value stored was larger than a byte, replace this load
2193 // with the value stored truncated to a byte. If no truncation is
2194 // needed, the replacement is done in LoadNode::Identity().
2195 //
2196 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2197 Node* mem = in(MemNode::Memory);
2198 Node* value = can_see_stored_value(mem,phase);
2199 if (value != NULL) {
2200 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2201 if (narrow != value) {
2202 return narrow;
2203 }
2204 }
2205 // Identity call will handle the case where truncation is not needed.
2206 return LoadNode::Ideal(phase, can_reshape);
2207 }
2208
2209 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2210 Node* mem = in(MemNode::Memory);
2211 Node* value = can_see_stored_value(mem,phase);
2212 if (value != NULL && value->is_Con() &&
2213 !value->bottom_type()->higher_equal(_type)) {
2214 // If the input to the store does not fit with the load's result type,
2215 // it must be truncated. We can't delay until Ideal call since
2216 // a singleton Value is needed for split_thru_phi optimization.
2217 int con = value->get_int();
2218 return TypeInt::make((con << 24) >> 24);
2219 }
2220 return LoadNode::Value(phase);
2221 }
2222
2223 //--------------------------LoadUBNode::Ideal-------------------------------------
2224 //
2225 // If the previous store is to the same address as this load,
2226 // and the value stored was larger than a byte, replace this load
2227 // with the value stored truncated to a byte. If no truncation is
2228 // needed, the replacement is done in LoadNode::Identity().
2229 //
2230 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2231 Node* mem = in(MemNode::Memory);
2232 Node* value = can_see_stored_value(mem, phase);
2233 if (value != NULL) {
2234 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2235 if (narrow != value) {
2236 return narrow;
2237 }
2238 }
2239 // Identity call will handle the case where truncation is not needed.
2240 return LoadNode::Ideal(phase, can_reshape);
2241 }
2242
2243 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2244 Node* mem = in(MemNode::Memory);
2245 Node* value = can_see_stored_value(mem,phase);
2246 if (value != NULL && value->is_Con() &&
2247 !value->bottom_type()->higher_equal(_type)) {
2248 // If the input to the store does not fit with the load's result type,
2249 // it must be truncated. We can't delay until Ideal call since
2250 // a singleton Value is needed for split_thru_phi optimization.
2251 int con = value->get_int();
2252 return TypeInt::make(con & 0xFF);
2253 }
2254 return LoadNode::Value(phase);
2255 }
2256
2257 //--------------------------LoadUSNode::Ideal-------------------------------------
2258 //
2259 // If the previous store is to the same address as this load,
2260 // and the value stored was larger than a char, replace this load
2261 // with the value stored truncated to a char. If no truncation is
2262 // needed, the replacement is done in LoadNode::Identity().
2263 //
2264 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2265 Node* mem = in(MemNode::Memory);
2266 Node* value = can_see_stored_value(mem,phase);
2267 if (value != NULL) {
2268 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2269 if (narrow != value) {
2270 return narrow;
2271 }
2272 }
2273 // Identity call will handle the case where truncation is not needed.
2274 return LoadNode::Ideal(phase, can_reshape);
2275 }
2276
2277 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2278 Node* mem = in(MemNode::Memory);
2279 Node* value = can_see_stored_value(mem,phase);
2280 if (value != NULL && value->is_Con() &&
2281 !value->bottom_type()->higher_equal(_type)) {
2282 // If the input to the store does not fit with the load's result type,
2283 // it must be truncated. We can't delay until Ideal call since
2284 // a singleton Value is needed for split_thru_phi optimization.
2285 int con = value->get_int();
2286 return TypeInt::make(con & 0xFFFF);
2287 }
2288 return LoadNode::Value(phase);
2289 }
2290
2291 //--------------------------LoadSNode::Ideal--------------------------------------
2292 //
2293 // If the previous store is to the same address as this load,
2294 // and the value stored was larger than a short, replace this load
2295 // with the value stored truncated to a short. If no truncation is
2296 // needed, the replacement is done in LoadNode::Identity().
2297 //
2298 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2299 Node* mem = in(MemNode::Memory);
2300 Node* value = can_see_stored_value(mem,phase);
2301 if (value != NULL) {
2302 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2303 if (narrow != value) {
2304 return narrow;
2305 }
2306 }
2307 // Identity call will handle the case where truncation is not needed.
2308 return LoadNode::Ideal(phase, can_reshape);
2309 }
2310
2311 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2312 Node* mem = in(MemNode::Memory);
2313 Node* value = can_see_stored_value(mem,phase);
2314 if (value != NULL && value->is_Con() &&
2315 !value->bottom_type()->higher_equal(_type)) {
2316 // If the input to the store does not fit with the load's result type,
2317 // it must be truncated. We can't delay until Ideal call since
2318 // a singleton Value is needed for split_thru_phi optimization.
2319 int con = value->get_int();
2320 return TypeInt::make((con << 16) >> 16);
2321 }
2322 return LoadNode::Value(phase);
2323 }
2324
2325 //=============================================================================
2326 //----------------------------LoadKlassNode::make------------------------------
2327 // Polymorphic factory method:
2328 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at,
2329 const TypeKlassPtr* tk) {
2330 // sanity check the alias category against the created node type
2331 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2332 assert(adr_type != NULL, "expecting TypeKlassPtr");
2333 #ifdef _LP64
2334 if (adr_type->is_ptr_to_narrowklass()) {
2335 assert(UseCompressedClassPointers, "no compressed klasses");
2336 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2337 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2338 }
2339 #endif
2340 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2341 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2342 }
2343
2344 //------------------------------Value------------------------------------------
2345 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2346 return klass_value_common(phase);
2347 }
2348
2349 // In most cases, LoadKlassNode does not have the control input set. If the control
2350 // input is set, it must not be removed (by LoadNode::Ideal()).
2351 bool LoadKlassNode::can_remove_control() const {
2352 return false;
2353 }
2354
2355 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2356 // Either input is TOP ==> the result is TOP
2357 const Type *t1 = phase->type( in(MemNode::Memory) );
2358 if (t1 == Type::TOP) return Type::TOP;
2359 Node *adr = in(MemNode::Address);
2360 const Type *t2 = phase->type( adr );
2361 if (t2 == Type::TOP) return Type::TOP;
2362 const TypePtr *tp = t2->is_ptr();
2363 if (TypePtr::above_centerline(tp->ptr()) ||
2364 tp->ptr() == TypePtr::Null) return Type::TOP;
2365
2366 // Return a more precise klass, if possible
2367 const TypeInstPtr *tinst = tp->isa_instptr();
2368 if (tinst != NULL) {
2369 ciInstanceKlass* ik = tinst->instance_klass();
2370 int offset = tinst->offset();
2371 if (ik == phase->C->env()->Class_klass()
2372 && (offset == java_lang_Class::klass_offset() ||
2373 offset == java_lang_Class::array_klass_offset())) {
2374 // We are loading a special hidden field from a Class mirror object,
2375 // the field which points to the VM's Klass metaobject.
2376 bool null_free = false;
2377 ciType* t = tinst->java_mirror_type(&null_free);
2378 // java_mirror_type returns non-null for compile-time Class constants.
2379 if (t != NULL) {
2380 // constant oop => constant klass
2381 if (offset == java_lang_Class::array_klass_offset()) {
2382 if (t->is_void()) {
2383 // We cannot create a void array. Since void is a primitive type return null
2384 // klass. Users of this result need to do a null check on the returned klass.
2385 return TypePtr::NULL_PTR;
2386 }
2387 return TypeKlassPtr::make(ciArrayKlass::make(t, null_free));
2388 }
2389 if (!t->is_klass()) {
2390 // a primitive Class (e.g., int.class) has NULL for a klass field
2391 return TypePtr::NULL_PTR;
2392 }
2393 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2394 return TypeKlassPtr::make(t->as_klass());
2395 }
2396 // non-constant mirror, so we can't tell what's going on
2397 }
2398 if (!tinst->is_loaded())
2399 return _type; // Bail out if not loaded
2400 if (offset == oopDesc::klass_offset_in_bytes()) {
2401 return tinst->as_klass_type(true);
2402 }
2403 }
2404
2405 // Check for loading klass from an array
2406 const TypeAryPtr* tary = tp->isa_aryptr();
2407 if (tary != NULL && tary->elem() != Type::BOTTOM &&
2408 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2409 return tary->as_klass_type(true);
2410 }
2411
2412 // Check for loading klass from an array klass
2413 const TypeKlassPtr *tkls = tp->isa_klassptr();
2414 if (tkls != NULL && !StressReflectiveCode) {
2415 if (!tkls->is_loaded())
2416 return _type; // Bail out if not loaded
2417 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2418 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2419 // // Always returning precise element type is incorrect,
2420 // // e.g., element type could be object and array may contain strings
2421 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2422
2423 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2424 // according to the element type's subclassing.
2425 return tkls->is_aryklassptr()->elem();
2426 }
2427 if (tkls->isa_instklassptr() != NULL && tkls->klass_is_exact() &&
2428 tkls->offset() == in_bytes(Klass::super_offset())) {
2429 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2430 // The field is Klass::_super. Return its (constant) value.
2431 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2432 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2433 }
2434 }
2435
2436 // Bailout case
2437 return LoadNode::Value(phase);
2438 }
2439
2440 //------------------------------Identity---------------------------------------
2441 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2442 // Also feed through the klass in Allocate(...klass...)._klass.
2443 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2444 return klass_identity_common(phase);
2445 }
2446
2447 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2448 Node* x = LoadNode::Identity(phase);
2449 if (x != this) return x;
2450
2451 // Take apart the address into an oop and and offset.
2452 // Return 'this' if we cannot.
2453 Node* adr = in(MemNode::Address);
2454 intptr_t offset = 0;
2455 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2456 if (base == NULL) return this;
2457 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2458 if (toop == NULL) return this;
2459
2460 // Step over potential GC barrier for OopHandle resolve
2461 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2462 if (bs->is_gc_barrier_node(base)) {
2463 base = bs->step_over_gc_barrier(base);
2464 }
2465
2466 // We can fetch the klass directly through an AllocateNode.
2467 // This works even if the klass is not constant (clone or newArray).
2468 if (offset == oopDesc::klass_offset_in_bytes()) {
2469 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2470 if (allocated_klass != NULL) {
2471 return allocated_klass;
2472 }
2473 }
2474
2475 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2476 // See inline_native_Class_query for occurrences of these patterns.
2477 // Java Example: x.getClass().isAssignableFrom(y)
2478 //
2479 // This improves reflective code, often making the Class
2480 // mirror go completely dead. (Current exception: Class
2481 // mirrors may appear in debug info, but we could clean them out by
2482 // introducing a new debug info operator for Klass.java_mirror).
2483
2484 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2485 && offset == java_lang_Class::klass_offset()) {
2486 if (base->is_Load()) {
2487 Node* base2 = base->in(MemNode::Address);
2488 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2489 Node* adr2 = base2->in(MemNode::Address);
2490 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2491 if (tkls != NULL && !tkls->empty()
2492 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2493 && adr2->is_AddP()
2494 ) {
2495 int mirror_field = in_bytes(Klass::java_mirror_offset());
2496 if (tkls->offset() == mirror_field) {
2497 return adr2->in(AddPNode::Base);
2498 }
2499 }
2500 }
2501 }
2502 }
2503
2504 return this;
2505 }
2506
2507
2508 //------------------------------Value------------------------------------------
2509 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2510 const Type *t = klass_value_common(phase);
2511 if (t == Type::TOP)
2512 return t;
2513
2514 return t->make_narrowklass();
2515 }
2516
2517 //------------------------------Identity---------------------------------------
2518 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2519 // Also feed through the klass in Allocate(...klass...)._klass.
2520 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2521 Node *x = klass_identity_common(phase);
2522
2523 const Type *t = phase->type( x );
2524 if( t == Type::TOP ) return x;
2525 if( t->isa_narrowklass()) return x;
2526 assert (!t->isa_narrowoop(), "no narrow oop here");
2527
2528 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2529 }
2530
2531 //------------------------------Value-----------------------------------------
2532 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2533 // Either input is TOP ==> the result is TOP
2534 const Type *t1 = phase->type( in(MemNode::Memory) );
2535 if( t1 == Type::TOP ) return Type::TOP;
2536 Node *adr = in(MemNode::Address);
2537 const Type *t2 = phase->type( adr );
2538 if( t2 == Type::TOP ) return Type::TOP;
2539 const TypePtr *tp = t2->is_ptr();
2540 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2541 const TypeAryPtr *tap = tp->isa_aryptr();
2542 if( !tap ) return _type;
2543 return tap->size();
2544 }
2545
2546 //-------------------------------Ideal---------------------------------------
2547 // Feed through the length in AllocateArray(...length...)._length.
2548 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2549 Node* p = MemNode::Ideal_common(phase, can_reshape);
2550 if (p) return (p == NodeSentinel) ? NULL : p;
2551
2552 // Take apart the address into an oop and and offset.
2553 // Return 'this' if we cannot.
2554 Node* adr = in(MemNode::Address);
2555 intptr_t offset = 0;
2556 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2557 if (base == NULL) return NULL;
2558 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2559 if (tary == NULL) return NULL;
2560
2561 // We can fetch the length directly through an AllocateArrayNode.
2562 // This works even if the length is not constant (clone or newArray).
2563 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2564 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2565 if (alloc != NULL) {
2566 Node* allocated_length = alloc->Ideal_length();
2567 Node* len = alloc->make_ideal_length(tary, phase);
2568 if (allocated_length != len) {
2569 // New CastII improves on this.
2570 return len;
2571 }
2572 }
2573 }
2574
2575 return NULL;
2576 }
2577
2578 //------------------------------Identity---------------------------------------
2579 // Feed through the length in AllocateArray(...length...)._length.
2580 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2581 Node* x = LoadINode::Identity(phase);
2582 if (x != this) return x;
2583
2584 // Take apart the address into an oop and and offset.
2585 // Return 'this' if we cannot.
2586 Node* adr = in(MemNode::Address);
2587 intptr_t offset = 0;
2588 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2589 if (base == NULL) return this;
2590 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2591 if (tary == NULL) return this;
2592
2593 // We can fetch the length directly through an AllocateArrayNode.
2594 // This works even if the length is not constant (clone or newArray).
2595 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2596 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2597 if (alloc != NULL) {
2598 Node* allocated_length = alloc->Ideal_length();
2599 // Do not allow make_ideal_length to allocate a CastII node.
2600 Node* len = alloc->make_ideal_length(tary, phase, false);
2601 if (allocated_length == len) {
2602 // Return allocated_length only if it would not be improved by a CastII.
2603 return allocated_length;
2604 }
2605 }
2606 }
2607
2608 return this;
2609
2610 }
2611
2612 //=============================================================================
2613 //---------------------------StoreNode::make-----------------------------------
2614 // Polymorphic factory method:
2615 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo, bool require_atomic_access) {
2616 assert((mo == unordered || mo == release), "unexpected");
2617 Compile* C = gvn.C;
2618 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2619 ctl != NULL, "raw memory operations should have control edge");
2620
2621 switch (bt) {
2622 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2623 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2624 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2625 case T_CHAR:
2626 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2627 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2628 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2629 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2630 case T_METADATA:
2631 case T_ADDRESS:
2632 case T_PRIMITIVE_OBJECT:
2633 case T_OBJECT:
2634 #ifdef _LP64
2635 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2636 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2637 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2638 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2639 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2640 adr->bottom_type()->isa_rawptr())) {
2641 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2642 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2643 }
2644 #endif
2645 {
2646 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2647 }
2648 default:
2649 ShouldNotReachHere();
2650 return (StoreNode*)NULL;
2651 }
2652 }
2653
2654 //--------------------------bottom_type----------------------------------------
2655 const Type *StoreNode::bottom_type() const {
2656 return Type::MEMORY;
2657 }
2658
2659 //------------------------------hash-------------------------------------------
2660 uint StoreNode::hash() const {
2661 // unroll addition of interesting fields
2662 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2663
2664 // Since they are not commoned, do not hash them:
2665 return NO_HASH;
2666 }
2667
2668 //------------------------------Ideal------------------------------------------
2669 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2670 // When a store immediately follows a relevant allocation/initialization,
2671 // try to capture it into the initialization, or hoist it above.
2672 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2673 Node* p = MemNode::Ideal_common(phase, can_reshape);
2674 if (p) return (p == NodeSentinel) ? NULL : p;
2675
2676 Node* mem = in(MemNode::Memory);
2677 Node* address = in(MemNode::Address);
2678 Node* value = in(MemNode::ValueIn);
2679 // Back-to-back stores to same address? Fold em up. Generally
2680 // unsafe if I have intervening uses... Also disallowed for StoreCM
2681 // since they must follow each StoreP operation. Redundant StoreCMs
2682 // are eliminated just before matching in final_graph_reshape.
2683 if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::INLINES) {
2684 Node* st = mem;
2685 // If Store 'st' has more than one use, we cannot fold 'st' away.
2686 // For example, 'st' might be the final state at a conditional
2687 // return. Or, 'st' might be used by some node which is live at
2688 // the same time 'st' is live, which might be unschedulable. So,
2689 // require exactly ONE user until such time as we clone 'mem' for
2690 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2691 // true).
2692 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2693 // Looking at a dead closed cycle of memory?
2694 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2695 assert(Opcode() == st->Opcode() ||
2696 st->Opcode() == Op_StoreVector ||
2697 Opcode() == Op_StoreVector ||
2698 st->Opcode() == Op_StoreVectorScatter ||
2699 Opcode() == Op_StoreVectorScatter ||
2700 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2701 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2702 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2703 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
2704 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2705 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2706
2707 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2708 st->as_Store()->memory_size() <= this->memory_size()) {
2709 Node* use = st->raw_out(0);
2710 if (phase->is_IterGVN()) {
2711 phase->is_IterGVN()->rehash_node_delayed(use);
2712 }
2713 // It's OK to do this in the parser, since DU info is always accurate,
2714 // and the parser always refers to nodes via SafePointNode maps.
2715 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
2716 return this;
2717 }
2718 st = st->in(MemNode::Memory);
2719 }
2720 }
2721
2722
2723 // Capture an unaliased, unconditional, simple store into an initializer.
2724 // Or, if it is independent of the allocation, hoist it above the allocation.
2725 if (ReduceFieldZeroing && /*can_reshape &&*/
2726 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2727 InitializeNode* init = mem->in(0)->as_Initialize();
2728 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2729 if (offset > 0) {
2730 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2731 // If the InitializeNode captured me, it made a raw copy of me,
2732 // and I need to disappear.
2733 if (moved != NULL) {
2734 // %%% hack to ensure that Ideal returns a new node:
2735 mem = MergeMemNode::make(mem);
2736 return mem; // fold me away
2737 }
2738 }
2739 }
2740
2741 // Fold reinterpret cast into memory operation:
2742 // StoreX mem (MoveY2X v) => StoreY mem v
2743 if (value->is_Move()) {
2744 const Type* vt = value->in(1)->bottom_type();
2745 if (has_reinterpret_variant(vt)) {
2746 if (phase->C->post_loop_opts_phase()) {
2747 return convert_to_reinterpret_store(*phase, value->in(1), vt);
2748 } else {
2749 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
2750 }
2751 }
2752 }
2753
2754 return NULL; // No further progress
2755 }
2756
2757 //------------------------------Value-----------------------------------------
2758 const Type* StoreNode::Value(PhaseGVN* phase) const {
2759 // Either input is TOP ==> the result is TOP
2760 const Type *t1 = phase->type( in(MemNode::Memory) );
2761 if( t1 == Type::TOP ) return Type::TOP;
2762 const Type *t2 = phase->type( in(MemNode::Address) );
2763 if( t2 == Type::TOP ) return Type::TOP;
2764 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2765 if( t3 == Type::TOP ) return Type::TOP;
2766 return Type::MEMORY;
2767 }
2768
2769 //------------------------------Identity---------------------------------------
2770 // Remove redundant stores:
2771 // Store(m, p, Load(m, p)) changes to m.
2772 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2773 Node* StoreNode::Identity(PhaseGVN* phase) {
2774 Node* mem = in(MemNode::Memory);
2775 Node* adr = in(MemNode::Address);
2776 Node* val = in(MemNode::ValueIn);
2777
2778 Node* result = this;
2779
2780 // Load then Store? Then the Store is useless
2781 if (val->is_Load() &&
2782 val->in(MemNode::Address)->eqv_uncast(adr) &&
2783 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2784 val->as_Load()->store_Opcode() == Opcode()) {
2785 result = mem;
2786 }
2787
2788 // Two stores in a row of the same value?
2789 if (result == this &&
2790 mem->is_Store() &&
2791 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2792 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2793 mem->Opcode() == Opcode()) {
2794 result = mem;
2795 }
2796
2797 // Store of zero anywhere into a freshly-allocated object?
2798 // Then the store is useless.
2799 // (It must already have been captured by the InitializeNode.)
2800 if (result == this && ReduceFieldZeroing) {
2801 // a newly allocated object is already all-zeroes everywhere
2802 if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
2803 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) {
2804 assert(!phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == NULL, "storing null to inline type array is forbidden");
2805 result = mem;
2806 }
2807
2808 if (result == this && phase->type(val)->is_zero_type()) {
2809 // the store may also apply to zero-bits in an earlier object
2810 Node* prev_mem = find_previous_store(phase);
2811 // Steps (a), (b): Walk past independent stores to find an exact match.
2812 if (prev_mem != NULL) {
2813 Node* prev_val = can_see_stored_value(prev_mem, phase);
2814 if (prev_val != NULL && prev_val == val) {
2815 // prev_val and val might differ by a cast; it would be good
2816 // to keep the more informative of the two.
2817 result = mem;
2818 }
2819 }
2820 }
2821 }
2822
2823 PhaseIterGVN* igvn = phase->is_IterGVN();
2824 if (result != this && igvn != NULL) {
2825 MemBarNode* trailing = trailing_membar();
2826 if (trailing != NULL) {
2827 #ifdef ASSERT
2828 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2829 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2830 #endif
2831 trailing->remove(igvn);
2832 }
2833 }
2834
2835 return result;
2836 }
2837
2838 //------------------------------match_edge-------------------------------------
2839 // Do we Match on this edge index or not? Match only memory & value
2840 uint StoreNode::match_edge(uint idx) const {
2841 return idx == MemNode::Address || idx == MemNode::ValueIn;
2842 }
2843
2844 //------------------------------cmp--------------------------------------------
2845 // Do not common stores up together. They generally have to be split
2846 // back up anyways, so do not bother.
2847 bool StoreNode::cmp( const Node &n ) const {
2848 return (&n == this); // Always fail except on self
2849 }
2850
2851 //------------------------------Ideal_masked_input-----------------------------
2852 // Check for a useless mask before a partial-word store
2853 // (StoreB ... (AndI valIn conIa) )
2854 // If (conIa & mask == mask) this simplifies to
2855 // (StoreB ... (valIn) )
2856 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2857 Node *val = in(MemNode::ValueIn);
2858 if( val->Opcode() == Op_AndI ) {
2859 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2860 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2861 set_req_X(MemNode::ValueIn, val->in(1), phase);
2862 return this;
2863 }
2864 }
2865 return NULL;
2866 }
2867
2868
2869 //------------------------------Ideal_sign_extended_input----------------------
2870 // Check for useless sign-extension before a partial-word store
2871 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2872 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2873 // (StoreB ... (valIn) )
2874 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2875 Node *val = in(MemNode::ValueIn);
2876 if( val->Opcode() == Op_RShiftI ) {
2877 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2878 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2879 Node *shl = val->in(1);
2880 if( shl->Opcode() == Op_LShiftI ) {
2881 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2882 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2883 set_req_X(MemNode::ValueIn, shl->in(1), phase);
2884 return this;
2885 }
2886 }
2887 }
2888 }
2889 return NULL;
2890 }
2891
2892 //------------------------------value_never_loaded-----------------------------------
2893 // Determine whether there are any possible loads of the value stored.
2894 // For simplicity, we actually check if there are any loads from the
2895 // address stored to, not just for loads of the value stored by this node.
2896 //
2897 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2898 Node *adr = in(Address);
2899 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2900 if (adr_oop == NULL)
2901 return false;
2902 if (!adr_oop->is_known_instance_field())
2903 return false; // if not a distinct instance, there may be aliases of the address
2904 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2905 Node *use = adr->fast_out(i);
2906 if (use->is_Load() || use->is_LoadStore()) {
2907 return false;
2908 }
2909 }
2910 return true;
2911 }
2912
2913 MemBarNode* StoreNode::trailing_membar() const {
2914 if (is_release()) {
2915 MemBarNode* trailing_mb = NULL;
2916 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2917 Node* u = fast_out(i);
2918 if (u->is_MemBar()) {
2919 if (u->as_MemBar()->trailing_store()) {
2920 assert(u->Opcode() == Op_MemBarVolatile, "");
2921 assert(trailing_mb == NULL, "only one");
2922 trailing_mb = u->as_MemBar();
2923 #ifdef ASSERT
2924 Node* leading = u->as_MemBar()->leading_membar();
2925 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2926 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2927 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2928 #endif
2929 } else {
2930 assert(u->as_MemBar()->standalone(), "");
2931 }
2932 }
2933 }
2934 return trailing_mb;
2935 }
2936 return NULL;
2937 }
2938
2939
2940 //=============================================================================
2941 //------------------------------Ideal------------------------------------------
2942 // If the store is from an AND mask that leaves the low bits untouched, then
2943 // we can skip the AND operation. If the store is from a sign-extension
2944 // (a left shift, then right shift) we can skip both.
2945 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2946 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2947 if( progress != NULL ) return progress;
2948
2949 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2950 if( progress != NULL ) return progress;
2951
2952 // Finally check the default case
2953 return StoreNode::Ideal(phase, can_reshape);
2954 }
2955
2956 //=============================================================================
2957 //------------------------------Ideal------------------------------------------
2958 // If the store is from an AND mask that leaves the low bits untouched, then
2959 // we can skip the AND operation
2960 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2961 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2962 if( progress != NULL ) return progress;
2963
2964 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2965 if( progress != NULL ) return progress;
2966
2967 // Finally check the default case
2968 return StoreNode::Ideal(phase, can_reshape);
2969 }
2970
2971 //=============================================================================
2972 //------------------------------Identity---------------------------------------
2973 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2974 // No need to card mark when storing a null ptr
2975 Node* my_store = in(MemNode::OopStore);
2976 if (my_store->is_Store()) {
2977 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2978 if( t1 == TypePtr::NULL_PTR ) {
2979 return in(MemNode::Memory);
2980 }
2981 }
2982 return this;
2983 }
2984
2985 //=============================================================================
2986 //------------------------------Ideal---------------------------------------
2987 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2988 Node* progress = StoreNode::Ideal(phase, can_reshape);
2989 if (progress != NULL) return progress;
2990
2991 Node* my_store = in(MemNode::OopStore);
2992 if (my_store->is_MergeMem()) {
2993 if (oop_alias_idx() != phase->C->get_alias_index(TypeAryPtr::INLINES) ||
2994 phase->C->flattened_accesses_share_alias()) {
2995 // The alias that was recorded is no longer accurate enough.
2996 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2997 set_req_X(MemNode::OopStore, mem, phase);
2998 return this;
2999 }
3000 }
3001
3002 return NULL;
3003 }
3004
3005 //------------------------------Value-----------------------------------------
3006 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
3007 // Either input is TOP ==> the result is TOP (checked in StoreNode::Value).
3008 // If extra input is TOP ==> the result is TOP
3009 const Type* t = phase->type(in(MemNode::OopStore));
3010 if (t == Type::TOP) {
3011 return Type::TOP;
3012 }
3013 return StoreNode::Value(phase);
3014 }
3015
3016
3017 //=============================================================================
3018 //----------------------------------SCMemProjNode------------------------------
3019 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3020 {
3021 if (in(0) == NULL || phase->type(in(0)) == Type::TOP) {
3022 return Type::TOP;
3023 }
3024 return bottom_type();
3025 }
3026
3027 //=============================================================================
3028 //----------------------------------LoadStoreNode------------------------------
3029 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3030 : Node(required),
3031 _type(rt),
3032 _adr_type(at),
3033 _barrier_data(0)
3034 {
3035 init_req(MemNode::Control, c );
3036 init_req(MemNode::Memory , mem);
3037 init_req(MemNode::Address, adr);
3038 init_req(MemNode::ValueIn, val);
3039 init_class_id(Class_LoadStore);
3040 }
3041
3042 //------------------------------Value-----------------------------------------
3043 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
3044 // Either input is TOP ==> the result is TOP
3045 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
3046 return Type::TOP;
3047 }
3048 const Type* t = phase->type(in(MemNode::Memory));
3049 if (t == Type::TOP) {
3050 return Type::TOP;
3051 }
3052 t = phase->type(in(MemNode::Address));
3053 if (t == Type::TOP) {
3054 return Type::TOP;
3055 }
3056 t = phase->type(in(MemNode::ValueIn));
3057 if (t == Type::TOP) {
3058 return Type::TOP;
3059 }
3060 return bottom_type();
3061 }
3062
3063 uint LoadStoreNode::ideal_reg() const {
3064 return _type->ideal_reg();
3065 }
3066
3067 bool LoadStoreNode::result_not_used() const {
3068 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
3069 Node *x = fast_out(i);
3070 if (x->Opcode() == Op_SCMemProj) continue;
3071 return false;
3072 }
3073 return true;
3074 }
3075
3076 MemBarNode* LoadStoreNode::trailing_membar() const {
3077 MemBarNode* trailing = NULL;
3078 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3079 Node* u = fast_out(i);
3080 if (u->is_MemBar()) {
3081 if (u->as_MemBar()->trailing_load_store()) {
3082 assert(u->Opcode() == Op_MemBarAcquire, "");
3083 assert(trailing == NULL, "only one");
3084 trailing = u->as_MemBar();
3085 #ifdef ASSERT
3086 Node* leading = trailing->leading_membar();
3087 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3088 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3089 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3090 #endif
3091 } else {
3092 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3093 }
3094 }
3095 }
3096
3097 return trailing;
3098 }
3099
3100 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3101
3102 //=============================================================================
3103 //----------------------------------LoadStoreConditionalNode--------------------
3104 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
3105 init_req(ExpectedIn, ex );
3106 }
3107
3108 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
3109 // Either input is TOP ==> the result is TOP
3110 const Type* t = phase->type(in(ExpectedIn));
3111 if (t == Type::TOP) {
3112 return Type::TOP;
3113 }
3114 return LoadStoreNode::Value(phase);
3115 }
3116
3117 //=============================================================================
3118 //-------------------------------adr_type--------------------------------------
3119 const TypePtr* ClearArrayNode::adr_type() const {
3120 Node *adr = in(3);
3121 if (adr == NULL) return NULL; // node is dead
3122 return MemNode::calculate_adr_type(adr->bottom_type());
3123 }
3124
3125 //------------------------------match_edge-------------------------------------
3126 // Do we Match on this edge index or not? Do not match memory
3127 uint ClearArrayNode::match_edge(uint idx) const {
3128 return idx > 1;
3129 }
3130
3131 //------------------------------Identity---------------------------------------
3132 // Clearing a zero length array does nothing
3133 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3134 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
3135 }
3136
3137 //------------------------------Idealize---------------------------------------
3138 // Clearing a short array is faster with stores
3139 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3140 // Already know this is a large node, do not try to ideal it
3141 if (_is_large) return NULL;
3142
3143 const int unit = BytesPerLong;
3144 const TypeX* t = phase->type(in(2))->isa_intptr_t();
3145 if (!t) return NULL;
3146 if (!t->is_con()) return NULL;
3147 intptr_t raw_count = t->get_con();
3148 intptr_t size = raw_count;
3149 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3150 // Clearing nothing uses the Identity call.
3151 // Negative clears are possible on dead ClearArrays
3152 // (see jck test stmt114.stmt11402.val).
3153 if (size <= 0 || size % unit != 0) return NULL;
3154 intptr_t count = size / unit;
3155 // Length too long; communicate this to matchers and assemblers.
3156 // Assemblers are responsible to produce fast hardware clears for it.
3157 if (size > InitArrayShortSize) {
3158 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
3159 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
3160 return NULL;
3161 }
3162 if (!IdealizeClearArrayNode) return NULL;
3163 Node *mem = in(1);
3164 if( phase->type(mem)==Type::TOP ) return NULL;
3165 Node *adr = in(3);
3166 const Type* at = phase->type(adr);
3167 if( at==Type::TOP ) return NULL;
3168 const TypePtr* atp = at->isa_ptr();
3169 // adjust atp to be the correct array element address type
3170 if (atp == NULL) atp = TypePtr::BOTTOM;
3171 else atp = atp->add_offset(Type::OffsetBot);
3172 // Get base for derived pointer purposes
3173 if( adr->Opcode() != Op_AddP ) Unimplemented();
3174 Node *base = adr->in(1);
3175
3176 Node *val = in(4);
3177 Node *off = phase->MakeConX(BytesPerLong);
3178 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
3179 count--;
3180 while( count-- ) {
3181 mem = phase->transform(mem);
3182 adr = phase->transform(new AddPNode(base,adr,off));
3183 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
3184 }
3185 return mem;
3186 }
3187
3188 //----------------------------step_through----------------------------------
3189 // Return allocation input memory edge if it is different instance
3190 // or itself if it is the one we are looking for.
3191 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
3192 Node* n = *np;
3193 assert(n->is_ClearArray(), "sanity");
3194 intptr_t offset;
3195 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3196 // This method is called only before Allocate nodes are expanded
3197 // during macro nodes expansion. Before that ClearArray nodes are
3198 // only generated in PhaseMacroExpand::generate_arraycopy() (before
3199 // Allocate nodes are expanded) which follows allocations.
3200 assert(alloc != NULL, "should have allocation");
3201 if (alloc->_idx == instance_id) {
3202 // Can not bypass initialization of the instance we are looking for.
3203 return false;
3204 }
3205 // Otherwise skip it.
3206 InitializeNode* init = alloc->initialization();
3207 if (init != NULL)
3208 *np = init->in(TypeFunc::Memory);
3209 else
3210 *np = alloc->in(TypeFunc::Memory);
3211 return true;
3212 }
3213
3214 //----------------------------clear_memory-------------------------------------
3215 // Generate code to initialize object storage to zero.
3216 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3217 Node* val,
3218 Node* raw_val,
3219 intptr_t start_offset,
3220 Node* end_offset,
3221 PhaseGVN* phase) {
3222 intptr_t offset = start_offset;
3223
3224 int unit = BytesPerLong;
3225 if ((offset % unit) != 0) {
3226 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3227 adr = phase->transform(adr);
3228 const TypePtr* atp = TypeRawPtr::BOTTOM;
3229 if (val != NULL) {
3230 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3231 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3232 } else {
3233 assert(raw_val == NULL, "val may not be null");
3234 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3235 }
3236 mem = phase->transform(mem);
3237 offset += BytesPerInt;
3238 }
3239 assert((offset % unit) == 0, "");
3240
3241 // Initialize the remaining stuff, if any, with a ClearArray.
3242 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
3243 }
3244
3245 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3246 Node* raw_val,
3247 Node* start_offset,
3248 Node* end_offset,
3249 PhaseGVN* phase) {
3250 if (start_offset == end_offset) {
3251 // nothing to do
3252 return mem;
3253 }
3254
3255 int unit = BytesPerLong;
3256 Node* zbase = start_offset;
3257 Node* zend = end_offset;
3258
3259 // Scale to the unit required by the CPU:
3260 if (!Matcher::init_array_count_is_in_bytes) {
3261 Node* shift = phase->intcon(exact_log2(unit));
3262 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3263 zend = phase->transform(new URShiftXNode(zend, shift) );
3264 }
3265
3266 // Bulk clear double-words
3267 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3268 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3269 if (raw_val == NULL) {
3270 raw_val = phase->MakeConX(0);
3271 }
3272 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
3273 return phase->transform(mem);
3274 }
3275
3276 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3277 Node* val,
3278 Node* raw_val,
3279 intptr_t start_offset,
3280 intptr_t end_offset,
3281 PhaseGVN* phase) {
3282 if (start_offset == end_offset) {
3283 // nothing to do
3284 return mem;
3285 }
3286
3287 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3288 intptr_t done_offset = end_offset;
3289 if ((done_offset % BytesPerLong) != 0) {
3290 done_offset -= BytesPerInt;
3291 }
3292 if (done_offset > start_offset) {
3293 mem = clear_memory(ctl, mem, dest, val, raw_val,
3294 start_offset, phase->MakeConX(done_offset), phase);
3295 }
3296 if (done_offset < end_offset) { // emit the final 32-bit store
3297 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3298 adr = phase->transform(adr);
3299 const TypePtr* atp = TypeRawPtr::BOTTOM;
3300 if (val != NULL) {
3301 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3302 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3303 } else {
3304 assert(raw_val == NULL, "val may not be null");
3305 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3306 }
3307 mem = phase->transform(mem);
3308 done_offset += BytesPerInt;
3309 }
3310 assert(done_offset == end_offset, "");
3311 return mem;
3312 }
3313
3314 //=============================================================================
3315 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3316 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3317 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3318 #ifdef ASSERT
3319 , _pair_idx(0)
3320 #endif
3321 {
3322 init_class_id(Class_MemBar);
3323 Node* top = C->top();
3324 init_req(TypeFunc::I_O,top);
3325 init_req(TypeFunc::FramePtr,top);
3326 init_req(TypeFunc::ReturnAdr,top);
3327 if (precedent != NULL)
3328 init_req(TypeFunc::Parms, precedent);
3329 }
3330
3331 //------------------------------cmp--------------------------------------------
3332 uint MemBarNode::hash() const { return NO_HASH; }
3333 bool MemBarNode::cmp( const Node &n ) const {
3334 return (&n == this); // Always fail except on self
3335 }
3336
3337 //------------------------------make-------------------------------------------
3338 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3339 switch (opcode) {
3340 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3341 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3342 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3343 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3344 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3345 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
3346 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3347 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3348 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3349 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3350 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3351 case Op_Initialize: return new InitializeNode(C, atp, pn);
3352 case Op_Blackhole: return new BlackholeNode(C, atp, pn);
3353 default: ShouldNotReachHere(); return NULL;
3354 }
3355 }
3356
3357 void MemBarNode::remove(PhaseIterGVN *igvn) {
3358 if (outcnt() != 2) {
3359 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
3360 assert(outcnt() == 1, "Only control then");
3361 }
3362 if (trailing_store() || trailing_load_store()) {
3363 MemBarNode* leading = leading_membar();
3364 if (leading != NULL) {
3365 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3366 leading->remove(igvn);
3367 }
3368 }
3369 if (proj_out_or_null(TypeFunc::Memory) != NULL) {
3370 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3371 }
3372 if (proj_out_or_null(TypeFunc::Control) != NULL) {
3373 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3374 }
3375 }
3376
3377 //------------------------------Ideal------------------------------------------
3378 // Return a node which is more "ideal" than the current node. Strip out
3379 // control copies
3380 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3381 if (remove_dead_region(phase, can_reshape)) return this;
3382 // Don't bother trying to transform a dead node
3383 if (in(0) && in(0)->is_top()) {
3384 return NULL;
3385 }
3386
3387 bool progress = false;
3388 // Eliminate volatile MemBars for scalar replaced objects.
3389 if (can_reshape && req() == (Precedent+1)) {
3390 bool eliminate = false;
3391 int opc = Opcode();
3392 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3393 // Volatile field loads and stores.
3394 Node* my_mem = in(MemBarNode::Precedent);
3395 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3396 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3397 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3398 // replace this Precedent (decodeN) with the Load instead.
3399 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3400 Node* load_node = my_mem->in(1);
3401 set_req(MemBarNode::Precedent, load_node);
3402 phase->is_IterGVN()->_worklist.push(my_mem);
3403 my_mem = load_node;
3404 } else {
3405 assert(my_mem->unique_out() == this, "sanity");
3406 del_req(Precedent);
3407 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3408 my_mem = NULL;
3409 }
3410 progress = true;
3411 }
3412 if (my_mem != NULL && my_mem->is_Mem()) {
3413 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3414 // Check for scalar replaced object reference.
3415 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3416 t_oop->offset() != Type::OffsetBot &&
3417 t_oop->offset() != Type::OffsetTop) {
3418 eliminate = true;
3419 }
3420 }
3421 } else if (opc == Op_MemBarRelease) {
3422 // Final field stores.
3423 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3424 if ((alloc != NULL) && alloc->is_Allocate() &&
3425 alloc->as_Allocate()->does_not_escape_thread()) {
3426 // The allocated object does not escape.
3427 eliminate = true;
3428 }
3429 }
3430 if (eliminate) {
3431 // Replace MemBar projections by its inputs.
3432 PhaseIterGVN* igvn = phase->is_IterGVN();
3433 remove(igvn);
3434 // Must return either the original node (now dead) or a new node
3435 // (Do not return a top here, since that would break the uniqueness of top.)
3436 return new ConINode(TypeInt::ZERO);
3437 }
3438 }
3439 return progress ? this : NULL;
3440 }
3441
3442 //------------------------------Value------------------------------------------
3443 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3444 if( !in(0) ) return Type::TOP;
3445 if( phase->type(in(0)) == Type::TOP )
3446 return Type::TOP;
3447 return TypeTuple::MEMBAR;
3448 }
3449
3450 //------------------------------match------------------------------------------
3451 // Construct projections for memory.
3452 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
3453 switch (proj->_con) {
3454 case TypeFunc::Control:
3455 case TypeFunc::Memory:
3456 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3457 }
3458 ShouldNotReachHere();
3459 return NULL;
3460 }
3461
3462 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3463 trailing->_kind = TrailingStore;
3464 leading->_kind = LeadingStore;
3465 #ifdef ASSERT
3466 trailing->_pair_idx = leading->_idx;
3467 leading->_pair_idx = leading->_idx;
3468 #endif
3469 }
3470
3471 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3472 trailing->_kind = TrailingLoadStore;
3473 leading->_kind = LeadingLoadStore;
3474 #ifdef ASSERT
3475 trailing->_pair_idx = leading->_idx;
3476 leading->_pair_idx = leading->_idx;
3477 #endif
3478 }
3479
3480 MemBarNode* MemBarNode::trailing_membar() const {
3481 ResourceMark rm;
3482 Node* trailing = (Node*)this;
3483 VectorSet seen;
3484 Node_Stack multis(0);
3485 do {
3486 Node* c = trailing;
3487 uint i = 0;
3488 do {
3489 trailing = NULL;
3490 for (; i < c->outcnt(); i++) {
3491 Node* next = c->raw_out(i);
3492 if (next != c && next->is_CFG()) {
3493 if (c->is_MultiBranch()) {
3494 if (multis.node() == c) {
3495 multis.set_index(i+1);
3496 } else {
3497 multis.push(c, i+1);
3498 }
3499 }
3500 trailing = next;
3501 break;
3502 }
3503 }
3504 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3505 break;
3506 }
3507 while (multis.size() > 0) {
3508 c = multis.node();
3509 i = multis.index();
3510 if (i < c->req()) {
3511 break;
3512 }
3513 multis.pop();
3514 }
3515 } while (multis.size() > 0);
3516 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3517
3518 MemBarNode* mb = trailing->as_MemBar();
3519 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3520 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3521 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3522 return mb;
3523 }
3524
3525 MemBarNode* MemBarNode::leading_membar() const {
3526 ResourceMark rm;
3527 VectorSet seen;
3528 Node_Stack regions(0);
3529 Node* leading = in(0);
3530 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3531 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3532 leading = NULL;
3533 while (regions.size() > 0 && leading == NULL) {
3534 Node* r = regions.node();
3535 uint i = regions.index();
3536 if (i < r->req()) {
3537 leading = r->in(i);
3538 regions.set_index(i+1);
3539 } else {
3540 regions.pop();
3541 }
3542 }
3543 if (leading == NULL) {
3544 assert(regions.size() == 0, "all paths should have been tried");
3545 return NULL;
3546 }
3547 }
3548 if (leading->is_Region()) {
3549 regions.push(leading, 2);
3550 leading = leading->in(1);
3551 } else {
3552 leading = leading->in(0);
3553 }
3554 }
3555 #ifdef ASSERT
3556 Unique_Node_List wq;
3557 wq.push((Node*)this);
3558 uint found = 0;
3559 for (uint i = 0; i < wq.size(); i++) {
3560 Node* n = wq.at(i);
3561 if (n->is_Region()) {
3562 for (uint j = 1; j < n->req(); j++) {
3563 Node* in = n->in(j);
3564 if (in != NULL && !in->is_top()) {
3565 wq.push(in);
3566 }
3567 }
3568 } else {
3569 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3570 assert(n == leading, "consistency check failed");
3571 found++;
3572 } else {
3573 Node* in = n->in(0);
3574 if (in != NULL && !in->is_top()) {
3575 wq.push(in);
3576 }
3577 }
3578 }
3579 }
3580 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3581 #endif
3582 if (leading == NULL) {
3583 return NULL;
3584 }
3585 MemBarNode* mb = leading->as_MemBar();
3586 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3587 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3588 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3589 return mb;
3590 }
3591
3592 #ifndef PRODUCT
3593 void BlackholeNode::format(PhaseRegAlloc* ra, outputStream* st) const {
3594 st->print("blackhole ");
3595 bool first = true;
3596 for (uint i = 0; i < req(); i++) {
3597 Node* n = in(i);
3598 if (n != NULL && OptoReg::is_valid(ra->get_reg_first(n))) {
3599 if (first) {
3600 first = false;
3601 } else {
3602 st->print(", ");
3603 }
3604 char buf[128];
3605 ra->dump_register(n, buf);
3606 st->print("%s", buf);
3607 }
3608 }
3609 st->cr();
3610 }
3611 #endif
3612
3613 //===========================InitializeNode====================================
3614 // SUMMARY:
3615 // This node acts as a memory barrier on raw memory, after some raw stores.
3616 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3617 // The Initialize can 'capture' suitably constrained stores as raw inits.
3618 // It can coalesce related raw stores into larger units (called 'tiles').
3619 // It can avoid zeroing new storage for memory units which have raw inits.
3620 // At macro-expansion, it is marked 'complete', and does not optimize further.
3621 //
3622 // EXAMPLE:
3623 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3624 // ctl = incoming control; mem* = incoming memory
3625 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3626 // First allocate uninitialized memory and fill in the header:
3627 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3628 // ctl := alloc.Control; mem* := alloc.Memory*
3629 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3630 // Then initialize to zero the non-header parts of the raw memory block:
3631 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3632 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3633 // After the initialize node executes, the object is ready for service:
3634 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3635 // Suppose its body is immediately initialized as {1,2}:
3636 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3637 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3638 // mem.SLICE(#short[*]) := store2
3639 //
3640 // DETAILS:
3641 // An InitializeNode collects and isolates object initialization after
3642 // an AllocateNode and before the next possible safepoint. As a
3643 // memory barrier (MemBarNode), it keeps critical stores from drifting
3644 // down past any safepoint or any publication of the allocation.
3645 // Before this barrier, a newly-allocated object may have uninitialized bits.
3646 // After this barrier, it may be treated as a real oop, and GC is allowed.
3647 //
3648 // The semantics of the InitializeNode include an implicit zeroing of
3649 // the new object from object header to the end of the object.
3650 // (The object header and end are determined by the AllocateNode.)
3651 //
3652 // Certain stores may be added as direct inputs to the InitializeNode.
3653 // These stores must update raw memory, and they must be to addresses
3654 // derived from the raw address produced by AllocateNode, and with
3655 // a constant offset. They must be ordered by increasing offset.
3656 // The first one is at in(RawStores), the last at in(req()-1).
3657 // Unlike most memory operations, they are not linked in a chain,
3658 // but are displayed in parallel as users of the rawmem output of
3659 // the allocation.
3660 //
3661 // (See comments in InitializeNode::capture_store, which continue
3662 // the example given above.)
3663 //
3664 // When the associated Allocate is macro-expanded, the InitializeNode
3665 // may be rewritten to optimize collected stores. A ClearArrayNode
3666 // may also be created at that point to represent any required zeroing.
3667 // The InitializeNode is then marked 'complete', prohibiting further
3668 // capturing of nearby memory operations.
3669 //
3670 // During macro-expansion, all captured initializations which store
3671 // constant values of 32 bits or smaller are coalesced (if advantageous)
3672 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3673 // initialized in fewer memory operations. Memory words which are
3674 // covered by neither tiles nor non-constant stores are pre-zeroed
3675 // by explicit stores of zero. (The code shape happens to do all
3676 // zeroing first, then all other stores, with both sequences occurring
3677 // in order of ascending offsets.)
3678 //
3679 // Alternatively, code may be inserted between an AllocateNode and its
3680 // InitializeNode, to perform arbitrary initialization of the new object.
3681 // E.g., the object copying intrinsics insert complex data transfers here.
3682 // The initialization must then be marked as 'complete' disable the
3683 // built-in zeroing semantics and the collection of initializing stores.
3684 //
3685 // While an InitializeNode is incomplete, reads from the memory state
3686 // produced by it are optimizable if they match the control edge and
3687 // new oop address associated with the allocation/initialization.
3688 // They return a stored value (if the offset matches) or else zero.
3689 // A write to the memory state, if it matches control and address,
3690 // and if it is to a constant offset, may be 'captured' by the
3691 // InitializeNode. It is cloned as a raw memory operation and rewired
3692 // inside the initialization, to the raw oop produced by the allocation.
3693 // Operations on addresses which are provably distinct (e.g., to
3694 // other AllocateNodes) are allowed to bypass the initialization.
3695 //
3696 // The effect of all this is to consolidate object initialization
3697 // (both arrays and non-arrays, both piecewise and bulk) into a
3698 // single location, where it can be optimized as a unit.
3699 //
3700 // Only stores with an offset less than TrackedInitializationLimit words
3701 // will be considered for capture by an InitializeNode. This puts a
3702 // reasonable limit on the complexity of optimized initializations.
3703
3704 //---------------------------InitializeNode------------------------------------
3705 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3706 : MemBarNode(C, adr_type, rawoop),
3707 _is_complete(Incomplete), _does_not_escape(false)
3708 {
3709 init_class_id(Class_Initialize);
3710
3711 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3712 assert(in(RawAddress) == rawoop, "proper init");
3713 // Note: allocation() can be NULL, for secondary initialization barriers
3714 }
3715
3716 // Since this node is not matched, it will be processed by the
3717 // register allocator. Declare that there are no constraints
3718 // on the allocation of the RawAddress edge.
3719 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3720 // This edge should be set to top, by the set_complete. But be conservative.
3721 if (idx == InitializeNode::RawAddress)
3722 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3723 return RegMask::Empty;
3724 }
3725
3726 Node* InitializeNode::memory(uint alias_idx) {
3727 Node* mem = in(Memory);
3728 if (mem->is_MergeMem()) {
3729 return mem->as_MergeMem()->memory_at(alias_idx);
3730 } else {
3731 // incoming raw memory is not split
3732 return mem;
3733 }
3734 }
3735
3736 bool InitializeNode::is_non_zero() {
3737 if (is_complete()) return false;
3738 remove_extra_zeroes();
3739 return (req() > RawStores);
3740 }
3741
3742 void InitializeNode::set_complete(PhaseGVN* phase) {
3743 assert(!is_complete(), "caller responsibility");
3744 _is_complete = Complete;
3745
3746 // After this node is complete, it contains a bunch of
3747 // raw-memory initializations. There is no need for
3748 // it to have anything to do with non-raw memory effects.
3749 // Therefore, tell all non-raw users to re-optimize themselves,
3750 // after skipping the memory effects of this initialization.
3751 PhaseIterGVN* igvn = phase->is_IterGVN();
3752 if (igvn) igvn->add_users_to_worklist(this);
3753 }
3754
3755 // convenience function
3756 // return false if the init contains any stores already
3757 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3758 InitializeNode* init = initialization();
3759 if (init == NULL || init->is_complete()) {
3760 return false;
3761 }
3762 init->remove_extra_zeroes();
3763 // for now, if this allocation has already collected any inits, bail:
3764 if (init->is_non_zero()) return false;
3765 init->set_complete(phase);
3766 return true;
3767 }
3768
3769 void InitializeNode::remove_extra_zeroes() {
3770 if (req() == RawStores) return;
3771 Node* zmem = zero_memory();
3772 uint fill = RawStores;
3773 for (uint i = fill; i < req(); i++) {
3774 Node* n = in(i);
3775 if (n->is_top() || n == zmem) continue; // skip
3776 if (fill < i) set_req(fill, n); // compact
3777 ++fill;
3778 }
3779 // delete any empty spaces created:
3780 while (fill < req()) {
3781 del_req(fill);
3782 }
3783 }
3784
3785 // Helper for remembering which stores go with which offsets.
3786 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3787 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3788 intptr_t offset = -1;
3789 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3790 phase, offset);
3791 if (base == NULL) return -1; // something is dead,
3792 if (offset < 0) return -1; // dead, dead
3793 return offset;
3794 }
3795
3796 // Helper for proving that an initialization expression is
3797 // "simple enough" to be folded into an object initialization.
3798 // Attempts to prove that a store's initial value 'n' can be captured
3799 // within the initialization without creating a vicious cycle, such as:
3800 // { Foo p = new Foo(); p.next = p; }
3801 // True for constants and parameters and small combinations thereof.
3802 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
3803 ResourceMark rm;
3804 Unique_Node_List worklist;
3805 worklist.push(value);
3806
3807 uint complexity_limit = 20;
3808 for (uint j = 0; j < worklist.size(); j++) {
3809 if (j >= complexity_limit) {
3810 return false; // Bail out if processed too many nodes
3811 }
3812
3813 Node* n = worklist.at(j);
3814 if (n == NULL) continue; // (can this really happen?)
3815 if (n->is_Proj()) n = n->in(0);
3816 if (n == this) return false; // found a cycle
3817 if (n->is_Con()) continue;
3818 if (n->is_Start()) continue; // params, etc., are OK
3819 if (n->is_Root()) continue; // even better
3820
3821 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
3822 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
3823 continue;
3824 }
3825
3826 Node* ctl = n->in(0);
3827 if (ctl != NULL && !ctl->is_top()) {
3828 if (ctl->is_Proj()) ctl = ctl->in(0);
3829 if (ctl == this) return false;
3830
3831 // If we already know that the enclosing memory op is pinned right after
3832 // the init, then any control flow that the store has picked up
3833 // must have preceded the init, or else be equal to the init.
3834 // Even after loop optimizations (which might change control edges)
3835 // a store is never pinned *before* the availability of its inputs.
3836 if (!MemNode::all_controls_dominate(n, this))
3837 return false; // failed to prove a good control
3838 }
3839
3840 // Check data edges for possible dependencies on 'this'.
3841 for (uint i = 1; i < n->req(); i++) {
3842 Node* m = n->in(i);
3843 if (m == NULL || m == n || m->is_top()) continue;
3844
3845 // Only process data inputs once
3846 worklist.push(m);
3847 }
3848 }
3849
3850 return true;
3851 }
3852
3853 // Here are all the checks a Store must pass before it can be moved into
3854 // an initialization. Returns zero if a check fails.
3855 // On success, returns the (constant) offset to which the store applies,
3856 // within the initialized memory.
3857 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
3858 const int FAIL = 0;
3859 if (st->req() != MemNode::ValueIn + 1)
3860 return FAIL; // an inscrutable StoreNode (card mark?)
3861 Node* ctl = st->in(MemNode::Control);
3862 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3863 return FAIL; // must be unconditional after the initialization
3864 Node* mem = st->in(MemNode::Memory);
3865 if (!(mem->is_Proj() && mem->in(0) == this))
3866 return FAIL; // must not be preceded by other stores
3867 Node* adr = st->in(MemNode::Address);
3868 intptr_t offset;
3869 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3870 if (alloc == NULL)
3871 return FAIL; // inscrutable address
3872 if (alloc != allocation())
3873 return FAIL; // wrong allocation! (store needs to float up)
3874 int size_in_bytes = st->memory_size();
3875 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3876 return FAIL; // mismatched access
3877 }
3878 Node* val = st->in(MemNode::ValueIn);
3879
3880 if (!detect_init_independence(val, phase))
3881 return FAIL; // stored value must be 'simple enough'
3882
3883 // The Store can be captured only if nothing after the allocation
3884 // and before the Store is using the memory location that the store
3885 // overwrites.
3886 bool failed = false;
3887 // If is_complete_with_arraycopy() is true the shape of the graph is
3888 // well defined and is safe so no need for extra checks.
3889 if (!is_complete_with_arraycopy()) {
3890 // We are going to look at each use of the memory state following
3891 // the allocation to make sure nothing reads the memory that the
3892 // Store writes.
3893 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3894 int alias_idx = phase->C->get_alias_index(t_adr);
3895 ResourceMark rm;
3896 Unique_Node_List mems;
3897 mems.push(mem);
3898 Node* unique_merge = NULL;
3899 for (uint next = 0; next < mems.size(); ++next) {
3900 Node *m = mems.at(next);
3901 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3902 Node *n = m->fast_out(j);
3903 if (n->outcnt() == 0) {
3904 continue;
3905 }
3906 if (n == st) {
3907 continue;
3908 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3909 // If the control of this use is different from the control
3910 // of the Store which is right after the InitializeNode then
3911 // this node cannot be between the InitializeNode and the
3912 // Store.
3913 continue;
3914 } else if (n->is_MergeMem()) {
3915 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3916 // We can hit a MergeMemNode (that will likely go away
3917 // later) that is a direct use of the memory state
3918 // following the InitializeNode on the same slice as the
3919 // store node that we'd like to capture. We need to check
3920 // the uses of the MergeMemNode.
3921 mems.push(n);
3922 }
3923 } else if (n->is_Mem()) {
3924 Node* other_adr = n->in(MemNode::Address);
3925 if (other_adr == adr) {
3926 failed = true;
3927 break;
3928 } else {
3929 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3930 if (other_t_adr != NULL) {
3931 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3932 if (other_alias_idx == alias_idx) {
3933 // A load from the same memory slice as the store right
3934 // after the InitializeNode. We check the control of the
3935 // object/array that is loaded from. If it's the same as
3936 // the store control then we cannot capture the store.
3937 assert(!n->is_Store(), "2 stores to same slice on same control?");
3938 Node* base = other_adr;
3939 if (base->is_Phi()) {
3940 // In rare case, base may be a PhiNode and it may read
3941 // the same memory slice between InitializeNode and store.
3942 failed = true;
3943 break;
3944 }
3945 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3946 base = base->in(AddPNode::Base);
3947 if (base != NULL) {
3948 base = base->uncast();
3949 if (base->is_Proj() && base->in(0) == alloc) {
3950 failed = true;
3951 break;
3952 }
3953 }
3954 }
3955 }
3956 }
3957 } else {
3958 failed = true;
3959 break;
3960 }
3961 }
3962 }
3963 }
3964 if (failed) {
3965 if (!can_reshape) {
3966 // We decided we couldn't capture the store during parsing. We
3967 // should try again during the next IGVN once the graph is
3968 // cleaner.
3969 phase->C->record_for_igvn(st);
3970 }
3971 return FAIL;
3972 }
3973
3974 return offset; // success
3975 }
3976
3977 // Find the captured store in(i) which corresponds to the range
3978 // [start..start+size) in the initialized object.
3979 // If there is one, return its index i. If there isn't, return the
3980 // negative of the index where it should be inserted.
3981 // Return 0 if the queried range overlaps an initialization boundary
3982 // or if dead code is encountered.
3983 // If size_in_bytes is zero, do not bother with overlap checks.
3984 int InitializeNode::captured_store_insertion_point(intptr_t start,
3985 int size_in_bytes,
3986 PhaseTransform* phase) {
3987 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
3988
3989 if (is_complete())
3990 return FAIL; // arraycopy got here first; punt
3991
3992 assert(allocation() != NULL, "must be present");
3993
3994 // no negatives, no header fields:
3995 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3996
3997 // after a certain size, we bail out on tracking all the stores:
3998 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3999 if (start >= ti_limit) return FAIL;
4000
4001 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4002 if (i >= limit) return -(int)i; // not found; here is where to put it
4003
4004 Node* st = in(i);
4005 intptr_t st_off = get_store_offset(st, phase);
4006 if (st_off < 0) {
4007 if (st != zero_memory()) {
4008 return FAIL; // bail out if there is dead garbage
4009 }
4010 } else if (st_off > start) {
4011 // ...we are done, since stores are ordered
4012 if (st_off < start + size_in_bytes) {
4013 return FAIL; // the next store overlaps
4014 }
4015 return -(int)i; // not found; here is where to put it
4016 } else if (st_off < start) {
4017 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
4018 if (size_in_bytes != 0 &&
4019 start < st_off + MAX_STORE &&
4020 start < st_off + st->as_Store()->memory_size()) {
4021 return FAIL; // the previous store overlaps
4022 }
4023 } else {
4024 if (size_in_bytes != 0 &&
4025 st->as_Store()->memory_size() != size_in_bytes) {
4026 return FAIL; // mismatched store size
4027 }
4028 return i;
4029 }
4030
4031 ++i;
4032 }
4033 }
4034
4035 // Look for a captured store which initializes at the offset 'start'
4036 // with the given size. If there is no such store, and no other
4037 // initialization interferes, then return zero_memory (the memory
4038 // projection of the AllocateNode).
4039 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
4040 PhaseTransform* phase) {
4041 assert(stores_are_sane(phase), "");
4042 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4043 if (i == 0) {
4044 return NULL; // something is dead
4045 } else if (i < 0) {
4046 return zero_memory(); // just primordial zero bits here
4047 } else {
4048 Node* st = in(i); // here is the store at this position
4049 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
4050 return st;
4051 }
4052 }
4053
4054 // Create, as a raw pointer, an address within my new object at 'offset'.
4055 Node* InitializeNode::make_raw_address(intptr_t offset,
4056 PhaseTransform* phase) {
4057 Node* addr = in(RawAddress);
4058 if (offset != 0) {
4059 Compile* C = phase->C;
4060 addr = phase->transform( new AddPNode(C->top(), addr,
4061 phase->MakeConX(offset)) );
4062 }
4063 return addr;
4064 }
4065
4066 // Clone the given store, converting it into a raw store
4067 // initializing a field or element of my new object.
4068 // Caller is responsible for retiring the original store,
4069 // with subsume_node or the like.
4070 //
4071 // From the example above InitializeNode::InitializeNode,
4072 // here are the old stores to be captured:
4073 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4074 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4075 //
4076 // Here is the changed code; note the extra edges on init:
4077 // alloc = (Allocate ...)
4078 // rawoop = alloc.RawAddress
4079 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
4080 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
4081 // init = (Initialize alloc.Control alloc.Memory rawoop
4082 // rawstore1 rawstore2)
4083 //
4084 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
4085 PhaseGVN* phase, bool can_reshape) {
4086 assert(stores_are_sane(phase), "");
4087
4088 if (start < 0) return NULL;
4089 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
4090
4091 Compile* C = phase->C;
4092 int size_in_bytes = st->memory_size();
4093 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4094 if (i == 0) return NULL; // bail out
4095 Node* prev_mem = NULL; // raw memory for the captured store
4096 if (i > 0) {
4097 prev_mem = in(i); // there is a pre-existing store under this one
4098 set_req(i, C->top()); // temporarily disconnect it
4099 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
4100 } else {
4101 i = -i; // no pre-existing store
4102 prev_mem = zero_memory(); // a slice of the newly allocated object
4103 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
4104 set_req(--i, C->top()); // reuse this edge; it has been folded away
4105 else
4106 ins_req(i, C->top()); // build a new edge
4107 }
4108 Node* new_st = st->clone();
4109 new_st->set_req(MemNode::Control, in(Control));
4110 new_st->set_req(MemNode::Memory, prev_mem);
4111 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
4112 new_st = phase->transform(new_st);
4113
4114 // At this point, new_st might have swallowed a pre-existing store
4115 // at the same offset, or perhaps new_st might have disappeared,
4116 // if it redundantly stored the same value (or zero to fresh memory).
4117
4118 // In any case, wire it in:
4119 PhaseIterGVN* igvn = phase->is_IterGVN();
4120 if (igvn) {
4121 igvn->rehash_node_delayed(this);
4122 }
4123 set_req(i, new_st);
4124
4125 // The caller may now kill the old guy.
4126 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
4127 assert(check_st == new_st || check_st == NULL, "must be findable");
4128 assert(!is_complete(), "");
4129 return new_st;
4130 }
4131
4132 static bool store_constant(jlong* tiles, int num_tiles,
4133 intptr_t st_off, int st_size,
4134 jlong con) {
4135 if ((st_off & (st_size-1)) != 0)
4136 return false; // strange store offset (assume size==2**N)
4137 address addr = (address)tiles + st_off;
4138 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4139 switch (st_size) {
4140 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
4141 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
4142 case sizeof(jint): *(jint*) addr = (jint) con; break;
4143 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
4144 default: return false; // strange store size (detect size!=2**N here)
4145 }
4146 return true; // return success to caller
4147 }
4148
4149 // Coalesce subword constants into int constants and possibly
4150 // into long constants. The goal, if the CPU permits,
4151 // is to initialize the object with a small number of 64-bit tiles.
4152 // Also, convert floating-point constants to bit patterns.
4153 // Non-constants are not relevant to this pass.
4154 //
4155 // In terms of the running example on InitializeNode::InitializeNode
4156 // and InitializeNode::capture_store, here is the transformation
4157 // of rawstore1 and rawstore2 into rawstore12:
4158 // alloc = (Allocate ...)
4159 // rawoop = alloc.RawAddress
4160 // tile12 = 0x00010002
4161 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4162 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4163 //
4164 void
4165 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4166 Node* size_in_bytes,
4167 PhaseGVN* phase) {
4168 Compile* C = phase->C;
4169
4170 assert(stores_are_sane(phase), "");
4171 // Note: After this pass, they are not completely sane,
4172 // since there may be some overlaps.
4173
4174 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4175
4176 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4177 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4178 size_limit = MIN2(size_limit, ti_limit);
4179 size_limit = align_up(size_limit, BytesPerLong);
4180 int num_tiles = size_limit / BytesPerLong;
4181
4182 // allocate space for the tile map:
4183 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4184 jlong tiles_buf[small_len];
4185 Node* nodes_buf[small_len];
4186 jlong inits_buf[small_len];
4187 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4188 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4189 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4190 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4191 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4192 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4193 // tiles: exact bitwise model of all primitive constants
4194 // nodes: last constant-storing node subsumed into the tiles model
4195 // inits: which bytes (in each tile) are touched by any initializations
4196
4197 //// Pass A: Fill in the tile model with any relevant stores.
4198
4199 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4200 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4201 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4202 Node* zmem = zero_memory(); // initially zero memory state
4203 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4204 Node* st = in(i);
4205 intptr_t st_off = get_store_offset(st, phase);
4206
4207 // Figure out the store's offset and constant value:
4208 if (st_off < header_size) continue; //skip (ignore header)
4209 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
4210 int st_size = st->as_Store()->memory_size();
4211 if (st_off + st_size > size_limit) break;
4212
4213 // Record which bytes are touched, whether by constant or not.
4214 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4215 continue; // skip (strange store size)
4216
4217 const Type* val = phase->type(st->in(MemNode::ValueIn));
4218 if (!val->singleton()) continue; //skip (non-con store)
4219 BasicType type = val->basic_type();
4220
4221 jlong con = 0;
4222 switch (type) {
4223 case T_INT: con = val->is_int()->get_con(); break;
4224 case T_LONG: con = val->is_long()->get_con(); break;
4225 case T_FLOAT: con = jint_cast(val->getf()); break;
4226 case T_DOUBLE: con = jlong_cast(val->getd()); break;
4227 default: continue; //skip (odd store type)
4228 }
4229
4230 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4231 st->Opcode() == Op_StoreL) {
4232 continue; // This StoreL is already optimal.
4233 }
4234
4235 // Store down the constant.
4236 store_constant(tiles, num_tiles, st_off, st_size, con);
4237
4238 intptr_t j = st_off >> LogBytesPerLong;
4239
4240 if (type == T_INT && st_size == BytesPerInt
4241 && (st_off & BytesPerInt) == BytesPerInt) {
4242 jlong lcon = tiles[j];
4243 if (!Matcher::isSimpleConstant64(lcon) &&
4244 st->Opcode() == Op_StoreI) {
4245 // This StoreI is already optimal by itself.
4246 jint* intcon = (jint*) &tiles[j];
4247 intcon[1] = 0; // undo the store_constant()
4248
4249 // If the previous store is also optimal by itself, back up and
4250 // undo the action of the previous loop iteration... if we can.
4251 // But if we can't, just let the previous half take care of itself.
4252 st = nodes[j];
4253 st_off -= BytesPerInt;
4254 con = intcon[0];
4255 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
4256 assert(st_off >= header_size, "still ignoring header");
4257 assert(get_store_offset(st, phase) == st_off, "must be");
4258 assert(in(i-1) == zmem, "must be");
4259 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4260 assert(con == tcon->is_int()->get_con(), "must be");
4261 // Undo the effects of the previous loop trip, which swallowed st:
4262 intcon[0] = 0; // undo store_constant()
4263 set_req(i-1, st); // undo set_req(i, zmem)
4264 nodes[j] = NULL; // undo nodes[j] = st
4265 --old_subword; // undo ++old_subword
4266 }
4267 continue; // This StoreI is already optimal.
4268 }
4269 }
4270
4271 // This store is not needed.
4272 set_req(i, zmem);
4273 nodes[j] = st; // record for the moment
4274 if (st_size < BytesPerLong) // something has changed
4275 ++old_subword; // includes int/float, but who's counting...
4276 else ++old_long;
4277 }
4278
4279 if ((old_subword + old_long) == 0)
4280 return; // nothing more to do
4281
4282 //// Pass B: Convert any non-zero tiles into optimal constant stores.
4283 // Be sure to insert them before overlapping non-constant stores.
4284 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
4285 for (int j = 0; j < num_tiles; j++) {
4286 jlong con = tiles[j];
4287 jlong init = inits[j];
4288 if (con == 0) continue;
4289 jint con0, con1; // split the constant, address-wise
4290 jint init0, init1; // split the init map, address-wise
4291 { union { jlong con; jint intcon[2]; } u;
4292 u.con = con;
4293 con0 = u.intcon[0];
4294 con1 = u.intcon[1];
4295 u.con = init;
4296 init0 = u.intcon[0];
4297 init1 = u.intcon[1];
4298 }
4299
4300 Node* old = nodes[j];
4301 assert(old != NULL, "need the prior store");
4302 intptr_t offset = (j * BytesPerLong);
4303
4304 bool split = !Matcher::isSimpleConstant64(con);
4305
4306 if (offset < header_size) {
4307 assert(offset + BytesPerInt >= header_size, "second int counts");
4308 assert(*(jint*)&tiles[j] == 0, "junk in header");
4309 split = true; // only the second word counts
4310 // Example: int a[] = { 42 ... }
4311 } else if (con0 == 0 && init0 == -1) {
4312 split = true; // first word is covered by full inits
4313 // Example: int a[] = { ... foo(), 42 ... }
4314 } else if (con1 == 0 && init1 == -1) {
4315 split = true; // second word is covered by full inits
4316 // Example: int a[] = { ... 42, foo() ... }
4317 }
4318
4319 // Here's a case where init0 is neither 0 nor -1:
4320 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4321 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4322 // In this case the tile is not split; it is (jlong)42.
4323 // The big tile is stored down, and then the foo() value is inserted.
4324 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4325
4326 Node* ctl = old->in(MemNode::Control);
4327 Node* adr = make_raw_address(offset, phase);
4328 const TypePtr* atp = TypeRawPtr::BOTTOM;
4329
4330 // One or two coalesced stores to plop down.
4331 Node* st[2];
4332 intptr_t off[2];
4333 int nst = 0;
4334 if (!split) {
4335 ++new_long;
4336 off[nst] = offset;
4337 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4338 phase->longcon(con), T_LONG, MemNode::unordered);
4339 } else {
4340 // Omit either if it is a zero.
4341 if (con0 != 0) {
4342 ++new_int;
4343 off[nst] = offset;
4344 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4345 phase->intcon(con0), T_INT, MemNode::unordered);
4346 }
4347 if (con1 != 0) {
4348 ++new_int;
4349 offset += BytesPerInt;
4350 adr = make_raw_address(offset, phase);
4351 off[nst] = offset;
4352 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4353 phase->intcon(con1), T_INT, MemNode::unordered);
4354 }
4355 }
4356
4357 // Insert second store first, then the first before the second.
4358 // Insert each one just before any overlapping non-constant stores.
4359 while (nst > 0) {
4360 Node* st1 = st[--nst];
4361 C->copy_node_notes_to(st1, old);
4362 st1 = phase->transform(st1);
4363 offset = off[nst];
4364 assert(offset >= header_size, "do not smash header");
4365 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4366 guarantee(ins_idx != 0, "must re-insert constant store");
4367 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4368 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4369 set_req(--ins_idx, st1);
4370 else
4371 ins_req(ins_idx, st1);
4372 }
4373 }
4374
4375 if (PrintCompilation && WizardMode)
4376 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4377 old_subword, old_long, new_int, new_long);
4378 if (C->log() != NULL)
4379 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4380 old_subword, old_long, new_int, new_long);
4381
4382 // Clean up any remaining occurrences of zmem:
4383 remove_extra_zeroes();
4384 }
4385
4386 // Explore forward from in(start) to find the first fully initialized
4387 // word, and return its offset. Skip groups of subword stores which
4388 // together initialize full words. If in(start) is itself part of a
4389 // fully initialized word, return the offset of in(start). If there
4390 // are no following full-word stores, or if something is fishy, return
4391 // a negative value.
4392 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4393 int int_map = 0;
4394 intptr_t int_map_off = 0;
4395 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4396
4397 for (uint i = start, limit = req(); i < limit; i++) {
4398 Node* st = in(i);
4399
4400 intptr_t st_off = get_store_offset(st, phase);
4401 if (st_off < 0) break; // return conservative answer
4402
4403 int st_size = st->as_Store()->memory_size();
4404 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4405 return st_off; // we found a complete word init
4406 }
4407
4408 // update the map:
4409
4410 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4411 if (this_int_off != int_map_off) {
4412 // reset the map:
4413 int_map = 0;
4414 int_map_off = this_int_off;
4415 }
4416
4417 int subword_off = st_off - this_int_off;
4418 int_map |= right_n_bits(st_size) << subword_off;
4419 if ((int_map & FULL_MAP) == FULL_MAP) {
4420 return this_int_off; // we found a complete word init
4421 }
4422
4423 // Did this store hit or cross the word boundary?
4424 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4425 if (next_int_off == this_int_off + BytesPerInt) {
4426 // We passed the current int, without fully initializing it.
4427 int_map_off = next_int_off;
4428 int_map >>= BytesPerInt;
4429 } else if (next_int_off > this_int_off + BytesPerInt) {
4430 // We passed the current and next int.
4431 return this_int_off + BytesPerInt;
4432 }
4433 }
4434
4435 return -1;
4436 }
4437
4438
4439 // Called when the associated AllocateNode is expanded into CFG.
4440 // At this point, we may perform additional optimizations.
4441 // Linearize the stores by ascending offset, to make memory
4442 // activity as coherent as possible.
4443 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4444 intptr_t header_size,
4445 Node* size_in_bytes,
4446 PhaseIterGVN* phase) {
4447 assert(!is_complete(), "not already complete");
4448 assert(stores_are_sane(phase), "");
4449 assert(allocation() != NULL, "must be present");
4450
4451 remove_extra_zeroes();
4452
4453 if (ReduceFieldZeroing || ReduceBulkZeroing)
4454 // reduce instruction count for common initialization patterns
4455 coalesce_subword_stores(header_size, size_in_bytes, phase);
4456
4457 Node* zmem = zero_memory(); // initially zero memory state
4458 Node* inits = zmem; // accumulating a linearized chain of inits
4459 #ifdef ASSERT
4460 intptr_t first_offset = allocation()->minimum_header_size();
4461 intptr_t last_init_off = first_offset; // previous init offset
4462 intptr_t last_init_end = first_offset; // previous init offset+size
4463 intptr_t last_tile_end = first_offset; // previous tile offset+size
4464 #endif
4465 intptr_t zeroes_done = header_size;
4466
4467 bool do_zeroing = true; // we might give up if inits are very sparse
4468 int big_init_gaps = 0; // how many large gaps have we seen?
4469
4470 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4471 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4472
4473 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4474 Node* st = in(i);
4475 intptr_t st_off = get_store_offset(st, phase);
4476 if (st_off < 0)
4477 break; // unknown junk in the inits
4478 if (st->in(MemNode::Memory) != zmem)
4479 break; // complicated store chains somehow in list
4480
4481 int st_size = st->as_Store()->memory_size();
4482 intptr_t next_init_off = st_off + st_size;
4483
4484 if (do_zeroing && zeroes_done < next_init_off) {
4485 // See if this store needs a zero before it or under it.
4486 intptr_t zeroes_needed = st_off;
4487
4488 if (st_size < BytesPerInt) {
4489 // Look for subword stores which only partially initialize words.
4490 // If we find some, we must lay down some word-level zeroes first,
4491 // underneath the subword stores.
4492 //
4493 // Examples:
4494 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4495 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4496 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4497 //
4498 // Note: coalesce_subword_stores may have already done this,
4499 // if it was prompted by constant non-zero subword initializers.
4500 // But this case can still arise with non-constant stores.
4501
4502 intptr_t next_full_store = find_next_fullword_store(i, phase);
4503
4504 // In the examples above:
4505 // in(i) p q r s x y z
4506 // st_off 12 13 14 15 12 13 14
4507 // st_size 1 1 1 1 1 1 1
4508 // next_full_s. 12 16 16 16 16 16 16
4509 // z's_done 12 16 16 16 12 16 12
4510 // z's_needed 12 16 16 16 16 16 16
4511 // zsize 0 0 0 0 4 0 4
4512 if (next_full_store < 0) {
4513 // Conservative tack: Zero to end of current word.
4514 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4515 } else {
4516 // Zero to beginning of next fully initialized word.
4517 // Or, don't zero at all, if we are already in that word.
4518 assert(next_full_store >= zeroes_needed, "must go forward");
4519 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4520 zeroes_needed = next_full_store;
4521 }
4522 }
4523
4524 if (zeroes_needed > zeroes_done) {
4525 intptr_t zsize = zeroes_needed - zeroes_done;
4526 // Do some incremental zeroing on rawmem, in parallel with inits.
4527 zeroes_done = align_down(zeroes_done, BytesPerInt);
4528 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4529 allocation()->in(AllocateNode::DefaultValue),
4530 allocation()->in(AllocateNode::RawDefaultValue),
4531 zeroes_done, zeroes_needed,
4532 phase);
4533 zeroes_done = zeroes_needed;
4534 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4535 do_zeroing = false; // leave the hole, next time
4536 }
4537 }
4538
4539 // Collect the store and move on:
4540 phase->replace_input_of(st, MemNode::Memory, inits);
4541 inits = st; // put it on the linearized chain
4542 set_req(i, zmem); // unhook from previous position
4543
4544 if (zeroes_done == st_off)
4545 zeroes_done = next_init_off;
4546
4547 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4548
4549 #ifdef ASSERT
4550 // Various order invariants. Weaker than stores_are_sane because
4551 // a large constant tile can be filled in by smaller non-constant stores.
4552 assert(st_off >= last_init_off, "inits do not reverse");
4553 last_init_off = st_off;
4554 const Type* val = NULL;
4555 if (st_size >= BytesPerInt &&
4556 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4557 (int)val->basic_type() < (int)T_OBJECT) {
4558 assert(st_off >= last_tile_end, "tiles do not overlap");
4559 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4560 last_tile_end = MAX2(last_tile_end, next_init_off);
4561 } else {
4562 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4563 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4564 assert(st_off >= last_init_end, "inits do not overlap");
4565 last_init_end = next_init_off; // it's a non-tile
4566 }
4567 #endif //ASSERT
4568 }
4569
4570 remove_extra_zeroes(); // clear out all the zmems left over
4571 add_req(inits);
4572
4573 if (!(UseTLAB && ZeroTLAB)) {
4574 // If anything remains to be zeroed, zero it all now.
4575 zeroes_done = align_down(zeroes_done, BytesPerInt);
4576 // if it is the last unused 4 bytes of an instance, forget about it
4577 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4578 if (zeroes_done + BytesPerLong >= size_limit) {
4579 AllocateNode* alloc = allocation();
4580 assert(alloc != NULL, "must be present");
4581 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4582 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4583 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
4584 if (zeroes_done == k->layout_helper())
4585 zeroes_done = size_limit;
4586 }
4587 }
4588 if (zeroes_done < size_limit) {
4589 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4590 allocation()->in(AllocateNode::DefaultValue),
4591 allocation()->in(AllocateNode::RawDefaultValue),
4592 zeroes_done, size_in_bytes, phase);
4593 }
4594 }
4595
4596 set_complete(phase);
4597 return rawmem;
4598 }
4599
4600
4601 #ifdef ASSERT
4602 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4603 if (is_complete())
4604 return true; // stores could be anything at this point
4605 assert(allocation() != NULL, "must be present");
4606 intptr_t last_off = allocation()->minimum_header_size();
4607 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4608 Node* st = in(i);
4609 intptr_t st_off = get_store_offset(st, phase);
4610 if (st_off < 0) continue; // ignore dead garbage
4611 if (last_off > st_off) {
4612 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4613 this->dump(2);
4614 assert(false, "ascending store offsets");
4615 return false;
4616 }
4617 last_off = st_off + st->as_Store()->memory_size();
4618 }
4619 return true;
4620 }
4621 #endif //ASSERT
4622
4623
4624
4625
4626 //============================MergeMemNode=====================================
4627 //
4628 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4629 // contributing store or call operations. Each contributor provides the memory
4630 // state for a particular "alias type" (see Compile::alias_type). For example,
4631 // if a MergeMem has an input X for alias category #6, then any memory reference
4632 // to alias category #6 may use X as its memory state input, as an exact equivalent
4633 // to using the MergeMem as a whole.
4634 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4635 //
4636 // (Here, the <N> notation gives the index of the relevant adr_type.)
4637 //
4638 // In one special case (and more cases in the future), alias categories overlap.
4639 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4640 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4641 // it is exactly equivalent to that state W:
4642 // MergeMem(<Bot>: W) <==> W
4643 //
4644 // Usually, the merge has more than one input. In that case, where inputs
4645 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4646 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4647 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4648 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4649 //
4650 // A merge can take a "wide" memory state as one of its narrow inputs.
4651 // This simply means that the merge observes out only the relevant parts of
4652 // the wide input. That is, wide memory states arriving at narrow merge inputs
4653 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4654 //
4655 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4656 // and that memory slices "leak through":
4657 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4658 //
4659 // But, in such a cascade, repeated memory slices can "block the leak":
4660 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4661 //
4662 // In the last example, Y is not part of the combined memory state of the
4663 // outermost MergeMem. The system must, of course, prevent unschedulable
4664 // memory states from arising, so you can be sure that the state Y is somehow
4665 // a precursor to state Y'.
4666 //
4667 //
4668 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4669 // of each MergeMemNode array are exactly the numerical alias indexes, including
4670 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4671 // Compile::alias_type (and kin) produce and manage these indexes.
4672 //
4673 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4674 // (Note that this provides quick access to the top node inside MergeMem methods,
4675 // without the need to reach out via TLS to Compile::current.)
4676 //
4677 // As a consequence of what was just described, a MergeMem that represents a full
4678 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4679 // containing all alias categories.
4680 //
4681 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4682 //
4683 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4684 // a memory state for the alias type <N>, or else the top node, meaning that
4685 // there is no particular input for that alias type. Note that the length of
4686 // a MergeMem is variable, and may be extended at any time to accommodate new
4687 // memory states at larger alias indexes. When merges grow, they are of course
4688 // filled with "top" in the unused in() positions.
4689 //
4690 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4691 // (Top was chosen because it works smoothly with passes like GCM.)
4692 //
4693 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4694 // the type of random VM bits like TLS references.) Since it is always the
4695 // first non-Bot memory slice, some low-level loops use it to initialize an
4696 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4697 //
4698 //
4699 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4700 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4701 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4702 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4703 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4704 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4705 //
4706 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4707 // really that different from the other memory inputs. An abbreviation called
4708 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4709 //
4710 //
4711 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4712 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4713 // that "emerges though" the base memory will be marked as excluding the alias types
4714 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4715 //
4716 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4717 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4718 //
4719 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4720 // (It is currently unimplemented.) As you can see, the resulting merge is
4721 // actually a disjoint union of memory states, rather than an overlay.
4722 //
4723
4724 //------------------------------MergeMemNode-----------------------------------
4725 Node* MergeMemNode::make_empty_memory() {
4726 Node* empty_memory = (Node*) Compile::current()->top();
4727 assert(empty_memory->is_top(), "correct sentinel identity");
4728 return empty_memory;
4729 }
4730
4731 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4732 init_class_id(Class_MergeMem);
4733 // all inputs are nullified in Node::Node(int)
4734 // set_input(0, NULL); // no control input
4735
4736 // Initialize the edges uniformly to top, for starters.
4737 Node* empty_mem = make_empty_memory();
4738 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4739 init_req(i,empty_mem);
4740 }
4741 assert(empty_memory() == empty_mem, "");
4742
4743 if( new_base != NULL && new_base->is_MergeMem() ) {
4744 MergeMemNode* mdef = new_base->as_MergeMem();
4745 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4746 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4747 mms.set_memory(mms.memory2());
4748 }
4749 assert(base_memory() == mdef->base_memory(), "");
4750 } else {
4751 set_base_memory(new_base);
4752 }
4753 }
4754
4755 // Make a new, untransformed MergeMem with the same base as 'mem'.
4756 // If mem is itself a MergeMem, populate the result with the same edges.
4757 MergeMemNode* MergeMemNode::make(Node* mem) {
4758 return new MergeMemNode(mem);
4759 }
4760
4761 //------------------------------cmp--------------------------------------------
4762 uint MergeMemNode::hash() const { return NO_HASH; }
4763 bool MergeMemNode::cmp( const Node &n ) const {
4764 return (&n == this); // Always fail except on self
4765 }
4766
4767 //------------------------------Identity---------------------------------------
4768 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4769 // Identity if this merge point does not record any interesting memory
4770 // disambiguations.
4771 Node* base_mem = base_memory();
4772 Node* empty_mem = empty_memory();
4773 if (base_mem != empty_mem) { // Memory path is not dead?
4774 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4775 Node* mem = in(i);
4776 if (mem != empty_mem && mem != base_mem) {
4777 return this; // Many memory splits; no change
4778 }
4779 }
4780 }
4781 return base_mem; // No memory splits; ID on the one true input
4782 }
4783
4784 //------------------------------Ideal------------------------------------------
4785 // This method is invoked recursively on chains of MergeMem nodes
4786 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4787 // Remove chain'd MergeMems
4788 //
4789 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4790 // relative to the "in(Bot)". Since we are patching both at the same time,
4791 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4792 // but rewrite each "in(i)" relative to the new "in(Bot)".
4793 Node *progress = NULL;
4794
4795
4796 Node* old_base = base_memory();
4797 Node* empty_mem = empty_memory();
4798 if (old_base == empty_mem)
4799 return NULL; // Dead memory path.
4800
4801 MergeMemNode* old_mbase;
4802 if (old_base != NULL && old_base->is_MergeMem())
4803 old_mbase = old_base->as_MergeMem();
4804 else
4805 old_mbase = NULL;
4806 Node* new_base = old_base;
4807
4808 // simplify stacked MergeMems in base memory
4809 if (old_mbase) new_base = old_mbase->base_memory();
4810
4811 // the base memory might contribute new slices beyond my req()
4812 if (old_mbase) grow_to_match(old_mbase);
4813
4814 // Note: We do not call verify_sparse on entry, because inputs
4815 // can normalize to the base_memory via subsume_node or similar
4816 // mechanisms. This method repairs that damage.
4817
4818 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4819
4820 // Look at each slice.
4821 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4822 Node* old_in = in(i);
4823 // calculate the old memory value
4824 Node* old_mem = old_in;
4825 if (old_mem == empty_mem) old_mem = old_base;
4826 assert(old_mem == memory_at(i), "");
4827
4828 // maybe update (reslice) the old memory value
4829
4830 // simplify stacked MergeMems
4831 Node* new_mem = old_mem;
4832 MergeMemNode* old_mmem;
4833 if (old_mem != NULL && old_mem->is_MergeMem())
4834 old_mmem = old_mem->as_MergeMem();
4835 else
4836 old_mmem = NULL;
4837 if (old_mmem == this) {
4838 // This can happen if loops break up and safepoints disappear.
4839 // A merge of BotPtr (default) with a RawPtr memory derived from a
4840 // safepoint can be rewritten to a merge of the same BotPtr with
4841 // the BotPtr phi coming into the loop. If that phi disappears
4842 // also, we can end up with a self-loop of the mergemem.
4843 // In general, if loops degenerate and memory effects disappear,
4844 // a mergemem can be left looking at itself. This simply means
4845 // that the mergemem's default should be used, since there is
4846 // no longer any apparent effect on this slice.
4847 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4848 // from start. Update the input to TOP.
4849 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4850 }
4851 else if (old_mmem != NULL) {
4852 new_mem = old_mmem->memory_at(i);
4853 }
4854 // else preceding memory was not a MergeMem
4855
4856 // maybe store down a new value
4857 Node* new_in = new_mem;
4858 if (new_in == new_base) new_in = empty_mem;
4859
4860 if (new_in != old_in) {
4861 // Warning: Do not combine this "if" with the previous "if"
4862 // A memory slice might have be be rewritten even if it is semantically
4863 // unchanged, if the base_memory value has changed.
4864 set_req_X(i, new_in, phase);
4865 progress = this; // Report progress
4866 }
4867 }
4868
4869 if (new_base != old_base) {
4870 set_req_X(Compile::AliasIdxBot, new_base, phase);
4871 // Don't use set_base_memory(new_base), because we need to update du.
4872 assert(base_memory() == new_base, "");
4873 progress = this;
4874 }
4875
4876 if( base_memory() == this ) {
4877 // a self cycle indicates this memory path is dead
4878 set_req(Compile::AliasIdxBot, empty_mem);
4879 }
4880
4881 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4882 // Recursion must occur after the self cycle check above
4883 if( base_memory()->is_MergeMem() ) {
4884 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4885 Node *m = phase->transform(new_mbase); // Rollup any cycles
4886 if( m != NULL &&
4887 (m->is_top() ||
4888 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4889 // propagate rollup of dead cycle to self
4890 set_req(Compile::AliasIdxBot, empty_mem);
4891 }
4892 }
4893
4894 if( base_memory() == empty_mem ) {
4895 progress = this;
4896 // Cut inputs during Parse phase only.
4897 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4898 if( !can_reshape ) {
4899 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4900 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4901 }
4902 }
4903 }
4904
4905 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4906 // Check if PhiNode::Ideal's "Split phis through memory merges"
4907 // transform should be attempted. Look for this->phi->this cycle.
4908 uint merge_width = req();
4909 if (merge_width > Compile::AliasIdxRaw) {
4910 PhiNode* phi = base_memory()->as_Phi();
4911 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4912 if (phi->in(i) == this) {
4913 phase->is_IterGVN()->_worklist.push(phi);
4914 break;
4915 }
4916 }
4917 }
4918 }
4919
4920 assert(progress || verify_sparse(), "please, no dups of base");
4921 return progress;
4922 }
4923
4924 //-------------------------set_base_memory-------------------------------------
4925 void MergeMemNode::set_base_memory(Node *new_base) {
4926 Node* empty_mem = empty_memory();
4927 set_req(Compile::AliasIdxBot, new_base);
4928 assert(memory_at(req()) == new_base, "must set default memory");
4929 // Clear out other occurrences of new_base:
4930 if (new_base != empty_mem) {
4931 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4932 if (in(i) == new_base) set_req(i, empty_mem);
4933 }
4934 }
4935 }
4936
4937 //------------------------------out_RegMask------------------------------------
4938 const RegMask &MergeMemNode::out_RegMask() const {
4939 return RegMask::Empty;
4940 }
4941
4942 //------------------------------dump_spec--------------------------------------
4943 #ifndef PRODUCT
4944 void MergeMemNode::dump_spec(outputStream *st) const {
4945 st->print(" {");
4946 Node* base_mem = base_memory();
4947 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4948 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4949 if (mem == base_mem) { st->print(" -"); continue; }
4950 st->print( " N%d:", mem->_idx );
4951 Compile::current()->get_adr_type(i)->dump_on(st);
4952 }
4953 st->print(" }");
4954 }
4955 #endif // !PRODUCT
4956
4957
4958 #ifdef ASSERT
4959 static bool might_be_same(Node* a, Node* b) {
4960 if (a == b) return true;
4961 if (!(a->is_Phi() || b->is_Phi())) return false;
4962 // phis shift around during optimization
4963 return true; // pretty stupid...
4964 }
4965
4966 // verify a narrow slice (either incoming or outgoing)
4967 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4968 if (!VerifyAliases) return; // don't bother to verify unless requested
4969 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4970 if (Node::in_dump()) return; // muzzle asserts when printing
4971 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4972 assert(n != NULL, "");
4973 // Elide intervening MergeMem's
4974 while (n->is_MergeMem()) {
4975 n = n->as_MergeMem()->memory_at(alias_idx);
4976 }
4977 Compile* C = Compile::current();
4978 const TypePtr* n_adr_type = n->adr_type();
4979 if (n == m->empty_memory()) {
4980 // Implicit copy of base_memory()
4981 } else if (n_adr_type != TypePtr::BOTTOM) {
4982 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4983 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4984 } else {
4985 // A few places like make_runtime_call "know" that VM calls are narrow,
4986 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4987 bool expected_wide_mem = false;
4988 if (n == m->base_memory()) {
4989 expected_wide_mem = true;
4990 } else if (alias_idx == Compile::AliasIdxRaw ||
4991 n == m->memory_at(Compile::AliasIdxRaw)) {
4992 expected_wide_mem = true;
4993 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4994 // memory can "leak through" calls on channels that
4995 // are write-once. Allow this also.
4996 expected_wide_mem = true;
4997 }
4998 assert(expected_wide_mem, "expected narrow slice replacement");
4999 }
5000 }
5001 #else // !ASSERT
5002 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
5003 #endif
5004
5005
5006 //-----------------------------memory_at---------------------------------------
5007 Node* MergeMemNode::memory_at(uint alias_idx) const {
5008 assert(alias_idx >= Compile::AliasIdxRaw ||
5009 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
5010 "must avoid base_memory and AliasIdxTop");
5011
5012 // Otherwise, it is a narrow slice.
5013 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
5014 if (is_empty_memory(n)) {
5015 // the array is sparse; empty slots are the "top" node
5016 n = base_memory();
5017 assert(Node::in_dump()
5018 || n == NULL || n->bottom_type() == Type::TOP
5019 || n->adr_type() == NULL // address is TOP
5020 || n->adr_type() == TypePtr::BOTTOM
5021 || n->adr_type() == TypeRawPtr::BOTTOM
5022 || Compile::current()->AliasLevel() == 0,
5023 "must be a wide memory");
5024 // AliasLevel == 0 if we are organizing the memory states manually.
5025 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
5026 } else {
5027 // make sure the stored slice is sane
5028 #ifdef ASSERT
5029 if (VMError::is_error_reported() || Node::in_dump()) {
5030 } else if (might_be_same(n, base_memory())) {
5031 // Give it a pass: It is a mostly harmless repetition of the base.
5032 // This can arise normally from node subsumption during optimization.
5033 } else {
5034 verify_memory_slice(this, alias_idx, n);
5035 }
5036 #endif
5037 }
5038 return n;
5039 }
5040
5041 //---------------------------set_memory_at-------------------------------------
5042 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
5043 verify_memory_slice(this, alias_idx, n);
5044 Node* empty_mem = empty_memory();
5045 if (n == base_memory()) n = empty_mem; // collapse default
5046 uint need_req = alias_idx+1;
5047 if (req() < need_req) {
5048 if (n == empty_mem) return; // already the default, so do not grow me
5049 // grow the sparse array
5050 do {
5051 add_req(empty_mem);
5052 } while (req() < need_req);
5053 }
5054 set_req( alias_idx, n );
5055 }
5056
5057
5058
5059 //--------------------------iteration_setup------------------------------------
5060 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
5061 if (other != NULL) {
5062 grow_to_match(other);
5063 // invariant: the finite support of mm2 is within mm->req()
5064 #ifdef ASSERT
5065 for (uint i = req(); i < other->req(); i++) {
5066 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
5067 }
5068 #endif
5069 }
5070 // Replace spurious copies of base_memory by top.
5071 Node* base_mem = base_memory();
5072 if (base_mem != NULL && !base_mem->is_top()) {
5073 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
5074 if (in(i) == base_mem)
5075 set_req(i, empty_memory());
5076 }
5077 }
5078 }
5079
5080 //---------------------------grow_to_match-------------------------------------
5081 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
5082 Node* empty_mem = empty_memory();
5083 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
5084 // look for the finite support of the other memory
5085 for (uint i = other->req(); --i >= req(); ) {
5086 if (other->in(i) != empty_mem) {
5087 uint new_len = i+1;
5088 while (req() < new_len) add_req(empty_mem);
5089 break;
5090 }
5091 }
5092 }
5093
5094 //---------------------------verify_sparse-------------------------------------
5095 #ifndef PRODUCT
5096 bool MergeMemNode::verify_sparse() const {
5097 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
5098 Node* base_mem = base_memory();
5099 // The following can happen in degenerate cases, since empty==top.
5100 if (is_empty_memory(base_mem)) return true;
5101 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5102 assert(in(i) != NULL, "sane slice");
5103 if (in(i) == base_mem) return false; // should have been the sentinel value!
5104 }
5105 return true;
5106 }
5107
5108 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5109 Node* n;
5110 n = mm->in(idx);
5111 if (mem == n) return true; // might be empty_memory()
5112 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5113 if (mem == n) return true;
5114 return false;
5115 }
5116 #endif // !PRODUCT