1 /*
2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shenandoah/brooksPointer.hpp"
29 #include "memory/allocation.inline.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/compile.hpp"
35 #include "opto/connode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/loopnode.hpp"
38 #include "opto/machnode.hpp"
39 #include "opto/matcher.hpp"
40 #include "opto/memnode.hpp"
41 #include "opto/mulnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/phaseX.hpp"
44 #include "opto/regmask.hpp"
45 #include "opto/shenandoahSupport.hpp"
46 #include "utilities/copy.hpp"
47
48 // Portions of code courtesy of Clifford Click
49
50 // Optimization - Graph Style
51
52 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
53
54 //=============================================================================
55 uint MemNode::size_of() const { return sizeof(*this); }
56
57 const TypePtr *MemNode::adr_type() const {
58 Node* adr = in(Address);
59 if (adr == NULL) return NULL; // node is dead
60 const TypePtr* cross_check = NULL;
61 DEBUG_ONLY(cross_check = _adr_type);
62 return calculate_adr_type(adr->bottom_type(), cross_check);
63 }
64
65 #ifndef PRODUCT
66 void MemNode::dump_spec(outputStream *st) const {
67 if (in(Address) == NULL) return; // node is dead
68 #ifndef ASSERT
69 // fake the missing field
70 const TypePtr* _adr_type = NULL;
71 if (in(Address) != NULL)
72 _adr_type = in(Address)->bottom_type()->isa_ptr();
73 #endif
74 dump_adr_type(this, _adr_type, st);
75
76 Compile* C = Compile::current();
77 if( C->alias_type(_adr_type)->is_volatile() )
78 st->print(" Volatile!");
79 }
80
81 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
82 st->print(" @");
83 if (adr_type == NULL) {
84 st->print("NULL");
85 } else {
86 adr_type->dump_on(st);
87 Compile* C = Compile::current();
88 Compile::AliasType* atp = NULL;
89 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
90 if (atp == NULL)
91 st->print(", idx=?\?;");
92 else if (atp->index() == Compile::AliasIdxBot)
93 st->print(", idx=Bot;");
94 else if (atp->index() == Compile::AliasIdxTop)
95 st->print(", idx=Top;");
96 else if (atp->index() == Compile::AliasIdxRaw)
97 st->print(", idx=Raw;");
98 else {
99 ciField* field = atp->field();
100 if (field) {
101 st->print(", name=");
102 field->print_name_on(st);
103 }
104 st->print(", idx=%d;", atp->index());
105 }
106 }
107 }
108
109 extern void print_alias_types();
110
111 #endif
112
113 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
114 assert((t_oop != NULL), "sanity");
115 bool is_instance = t_oop->is_known_instance_field();
116 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
117 (load != NULL) && load->is_Load() &&
118 (phase->is_IterGVN() != NULL);
119 if (!(is_instance || is_boxed_value_load))
120 return mchain; // don't try to optimize non-instance types
121 uint instance_id = t_oop->instance_id();
122 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
123 Node *prev = NULL;
124 Node *result = mchain;
125 while (prev != result) {
126 prev = result;
127 if (result == start_mem)
128 break; // hit one of our sentinels
129 // skip over a call which does not affect this memory slice
130 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
131 Node *proj_in = result->in(0);
132 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
133 break; // hit one of our sentinels
134 } else if (proj_in->is_Call()) {
135 // ArrayCopyNodes processed here as well
136 CallNode *call = proj_in->as_Call();
137 if (!call->may_modify(t_oop, phase)) { // returns false for instances
138 result = call->in(TypeFunc::Memory);
139 }
140 } else if (proj_in->is_Initialize()) {
141 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
142 // Stop if this is the initialization for the object instance which
143 // contains this memory slice, otherwise skip over it.
144 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
145 break;
146 }
147 if (is_instance) {
148 result = proj_in->in(TypeFunc::Memory);
149 } else if (is_boxed_value_load) {
150 Node* klass = alloc->in(AllocateNode::KlassNode);
151 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
152 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
153 result = proj_in->in(TypeFunc::Memory); // not related allocation
154 }
155 }
156 } else if (proj_in->is_MemBar()) {
157 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase)) {
158 break;
159 }
160 result = proj_in->in(TypeFunc::Memory);
161 } else {
162 assert(false, "unexpected projection");
163 }
164 } else if (result->is_ClearArray()) {
165 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
166 // Can not bypass initialization of the instance
167 // we are looking for.
168 break;
169 }
170 // Otherwise skip it (the call updated 'result' value).
171 } else if (result->is_MergeMem()) {
172 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
173 }
174 }
175 return result;
176 }
177
178 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
179 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
180 if (t_oop == NULL)
181 return mchain; // don't try to optimize non-oop types
182 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
183 bool is_instance = t_oop->is_known_instance_field();
184 PhaseIterGVN *igvn = phase->is_IterGVN();
185 if (is_instance && igvn != NULL && result->is_Phi()) {
186 PhiNode *mphi = result->as_Phi();
187 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
188 const TypePtr *t = mphi->adr_type();
189 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
190 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
191 t->is_oopptr()->cast_to_exactness(true)
192 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
193 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
194 // clone the Phi with our address type
195 result = mphi->split_out_instance(t_adr, igvn);
196 } else {
197 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
198 }
199 }
200 return result;
201 }
202
203 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
204 uint alias_idx = phase->C->get_alias_index(tp);
205 Node *mem = mmem;
206 #ifdef ASSERT
207 {
208 // Check that current type is consistent with the alias index used during graph construction
209 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
210 bool consistent = adr_check == NULL || adr_check->empty() ||
211 phase->C->must_alias(adr_check, alias_idx );
212 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
213 if( !consistent && adr_check != NULL && !adr_check->empty() &&
214 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
215 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
216 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
217 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
218 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
219 // don't assert if it is dead code.
220 consistent = true;
221 }
222 if( !consistent ) {
223 st->print("alias_idx==%d, adr_check==", alias_idx);
224 if( adr_check == NULL ) {
225 st->print("NULL");
226 } else {
227 adr_check->dump();
228 }
229 st->cr();
230 print_alias_types();
231 assert(consistent, "adr_check must match alias idx");
232 }
233 }
234 #endif
235 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
236 // means an array I have not precisely typed yet. Do not do any
237 // alias stuff with it any time soon.
238 const TypeOopPtr *toop = tp->isa_oopptr();
239 if( tp->base() != Type::AnyPtr &&
240 !(toop &&
241 toop->klass() != NULL &&
242 toop->klass()->is_java_lang_Object() &&
243 toop->offset() == Type::OffsetBot) ) {
244 // compress paths and change unreachable cycles to TOP
245 // If not, we can update the input infinitely along a MergeMem cycle
246 // Equivalent code in PhiNode::Ideal
247 Node* m = phase->transform(mmem);
248 // If transformed to a MergeMem, get the desired slice
249 // Otherwise the returned node represents memory for every slice
250 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
251 // Update input if it is progress over what we have now
252 }
253 return mem;
254 }
255
256 //--------------------------Ideal_common---------------------------------------
257 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
258 // Unhook non-raw memories from complete (macro-expanded) initializations.
259 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
260 // If our control input is a dead region, kill all below the region
261 Node *ctl = in(MemNode::Control);
262 if (ctl && remove_dead_region(phase, can_reshape))
263 return this;
264 ctl = in(MemNode::Control);
265 // Don't bother trying to transform a dead node
266 if (ctl && ctl->is_top()) return NodeSentinel;
267
268 PhaseIterGVN *igvn = phase->is_IterGVN();
269 // Wait if control on the worklist.
270 if (ctl && can_reshape && igvn != NULL) {
271 Node* bol = NULL;
272 Node* cmp = NULL;
273 if (ctl->in(0)->is_If()) {
274 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
275 bol = ctl->in(0)->in(1);
276 if (bol->is_Bool())
277 cmp = ctl->in(0)->in(1)->in(1);
278 }
279 if (igvn->_worklist.member(ctl) ||
280 (bol != NULL && igvn->_worklist.member(bol)) ||
281 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
282 // This control path may be dead.
283 // Delay this memory node transformation until the control is processed.
284 phase->is_IterGVN()->_worklist.push(this);
285 return NodeSentinel; // caller will return NULL
286 }
287 }
288 // Ignore if memory is dead, or self-loop
289 Node *mem = in(MemNode::Memory);
290 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
291 assert(mem != this, "dead loop in MemNode::Ideal");
292
293 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
294 // This memory slice may be dead.
295 // Delay this mem node transformation until the memory is processed.
296 phase->is_IterGVN()->_worklist.push(this);
297 return NodeSentinel; // caller will return NULL
298 }
299
300 Node *address = in(MemNode::Address);
301 const Type *t_adr = phase->type(address);
302 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
303
304 if (can_reshape && igvn != NULL &&
305 (igvn->_worklist.member(address) ||
306 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
307 // The address's base and type may change when the address is processed.
308 // Delay this mem node transformation until the address is processed.
309 phase->is_IterGVN()->_worklist.push(this);
310 return NodeSentinel; // caller will return NULL
311 }
312
313 // Do NOT remove or optimize the next lines: ensure a new alias index
314 // is allocated for an oop pointer type before Escape Analysis.
315 // Note: C++ will not remove it since the call has side effect.
316 if (t_adr->isa_oopptr()) {
317 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
318 }
319
320 Node* base = NULL;
321 if (address->is_AddP()) {
322 base = address->in(AddPNode::Base);
323 }
324 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
325 !t_adr->isa_rawptr()) {
326 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
327 // Skip this node optimization if its address has TOP base.
328 return NodeSentinel; // caller will return NULL
329 }
330
331 // Avoid independent memory operations
332 Node* old_mem = mem;
333
334 // The code which unhooks non-raw memories from complete (macro-expanded)
335 // initializations was removed. After macro-expansion all stores catched
336 // by Initialize node became raw stores and there is no information
337 // which memory slices they modify. So it is unsafe to move any memory
338 // operation above these stores. Also in most cases hooked non-raw memories
339 // were already unhooked by using information from detect_ptr_independence()
340 // and find_previous_store().
341
342 if (mem->is_MergeMem()) {
343 MergeMemNode* mmem = mem->as_MergeMem();
344 const TypePtr *tp = t_adr->is_ptr();
345
346 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
347 }
348
349 if (mem != old_mem) {
350 set_req(MemNode::Memory, mem);
351 if (can_reshape && old_mem->outcnt() == 0) {
352 igvn->_worklist.push(old_mem);
353 }
354 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
355 return this;
356 }
357
358 // let the subclass continue analyzing...
359 return NULL;
360 }
361
362 // Helper function for proving some simple control dominations.
363 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
364 // Already assumes that 'dom' is available at 'sub', and that 'sub'
365 // is not a constant (dominated by the method's StartNode).
366 // Used by MemNode::find_previous_store to prove that the
367 // control input of a memory operation predates (dominates)
368 // an allocation it wants to look past.
369 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
370 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
371 return false; // Conservative answer for dead code
372
373 // Check 'dom'. Skip Proj and CatchProj nodes.
374 dom = dom->find_exact_control(dom);
375 if (dom == NULL || dom->is_top())
376 return false; // Conservative answer for dead code
377
378 if (dom == sub) {
379 // For the case when, for example, 'sub' is Initialize and the original
380 // 'dom' is Proj node of the 'sub'.
381 return false;
382 }
383
384 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
385 return true;
386
387 // 'dom' dominates 'sub' if its control edge and control edges
388 // of all its inputs dominate or equal to sub's control edge.
389
390 // Currently 'sub' is either Allocate, Initialize or Start nodes.
391 // Or Region for the check in LoadNode::Ideal();
392 // 'sub' should have sub->in(0) != NULL.
393 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
394 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
395
396 // Get control edge of 'sub'.
397 Node* orig_sub = sub;
398 sub = sub->find_exact_control(sub->in(0));
399 if (sub == NULL || sub->is_top())
400 return false; // Conservative answer for dead code
401
402 assert(sub->is_CFG(), "expecting control");
403
404 if (sub == dom)
405 return true;
406
407 if (sub->is_Start() || sub->is_Root())
408 return false;
409
410 {
411 // Check all control edges of 'dom'.
412
413 ResourceMark rm;
414 Arena* arena = Thread::current()->resource_area();
415 Node_List nlist(arena);
416 Unique_Node_List dom_list(arena);
417
418 dom_list.push(dom);
419 bool only_dominating_controls = false;
420
421 for (uint next = 0; next < dom_list.size(); next++) {
422 Node* n = dom_list.at(next);
423 if (n == orig_sub)
424 return false; // One of dom's inputs dominated by sub.
425 if (!n->is_CFG() && n->pinned()) {
426 // Check only own control edge for pinned non-control nodes.
427 n = n->find_exact_control(n->in(0));
428 if (n == NULL || n->is_top())
429 return false; // Conservative answer for dead code
430 assert(n->is_CFG(), "expecting control");
431 dom_list.push(n);
432 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
433 only_dominating_controls = true;
434 } else if (n->is_CFG()) {
435 if (n->dominates(sub, nlist))
436 only_dominating_controls = true;
437 else
438 return false;
439 } else {
440 // First, own control edge.
441 Node* m = n->find_exact_control(n->in(0));
442 if (m != NULL) {
443 if (m->is_top())
444 return false; // Conservative answer for dead code
445 dom_list.push(m);
446 }
447 // Now, the rest of edges.
448 uint cnt = n->req();
449 for (uint i = 1; i < cnt; i++) {
450 m = n->find_exact_control(n->in(i));
451 if (m == NULL || m->is_top())
452 continue;
453 dom_list.push(m);
454 }
455 }
456 }
457 return only_dominating_controls;
458 }
459 }
460
461 //---------------------detect_ptr_independence---------------------------------
462 // Used by MemNode::find_previous_store to prove that two base
463 // pointers are never equal.
464 // The pointers are accompanied by their associated allocations,
465 // if any, which have been previously discovered by the caller.
466 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
467 Node* p2, AllocateNode* a2,
468 PhaseTransform* phase) {
469 // Attempt to prove that these two pointers cannot be aliased.
470 // They may both manifestly be allocations, and they should differ.
471 // Or, if they are not both allocations, they can be distinct constants.
472 // Otherwise, one is an allocation and the other a pre-existing value.
473 if (a1 == NULL && a2 == NULL) { // neither an allocation
474 return (p1 != p2) && p1->is_Con() && p2->is_Con();
475 } else if (a1 != NULL && a2 != NULL) { // both allocations
476 return (a1 != a2);
477 } else if (a1 != NULL) { // one allocation a1
478 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
479 return all_controls_dominate(p2, a1);
480 } else { //(a2 != NULL) // one allocation a2
481 return all_controls_dominate(p1, a2);
482 }
483 return false;
484 }
485
486
487 // Find an arraycopy that must have set (can_see_stored_value=true) or
488 // could have set (can_see_stored_value=false) the value for this load
489 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
490 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
491 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
492 Node* mb = mem->in(0);
493 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
494 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
495 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
496 if (ac->is_clonebasic()) {
497 intptr_t offset;
498 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
499 assert(alloc != NULL && alloc->initialization()->is_complete_with_arraycopy(), "broken allocation");
500 if (alloc == ld_alloc) {
501 return ac;
502 }
503 }
504 }
505 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
506 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
507
508 if (ac->is_arraycopy_validated() ||
509 ac->is_copyof_validated() ||
510 ac->is_copyofrange_validated()) {
511 Node* ld_addp = in(MemNode::Address);
512 if (ld_addp->is_AddP()) {
513 Node* ld_base = ld_addp->in(AddPNode::Address);
514 Node* ld_offs = ld_addp->in(AddPNode::Offset);
515
516 Node* dest = ac->in(ArrayCopyNode::Dest);
517
518 if (dest == ld_base) {
519 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
520 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
521 return ac;
522 }
523 if (!can_see_stored_value) {
524 mem = ac->in(TypeFunc::Memory);
525 }
526 }
527 }
528 }
529 }
530 return NULL;
531 }
532
533 // The logic for reordering loads and stores uses four steps:
534 // (a) Walk carefully past stores and initializations which we
535 // can prove are independent of this load.
536 // (b) Observe that the next memory state makes an exact match
537 // with self (load or store), and locate the relevant store.
538 // (c) Ensure that, if we were to wire self directly to the store,
539 // the optimizer would fold it up somehow.
540 // (d) Do the rewiring, and return, depending on some other part of
541 // the optimizer to fold up the load.
542 // This routine handles steps (a) and (b). Steps (c) and (d) are
543 // specific to loads and stores, so they are handled by the callers.
544 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
545 //
546 Node* MemNode::find_previous_store(PhaseTransform* phase) {
547 Node* ctrl = in(MemNode::Control);
548 Node* adr = in(MemNode::Address);
549 intptr_t offset = 0;
550 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
551 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
552
553 if (offset == Type::OffsetBot)
554 return NULL; // cannot unalias unless there are precise offsets
555
556 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
557
558 intptr_t size_in_bytes = memory_size();
559
560 Node* mem = in(MemNode::Memory); // start searching here...
561
562 int cnt = 50; // Cycle limiter
563 for (;;) { // While we can dance past unrelated stores...
564 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
565
566 Node* prev = mem;
567 if (mem->is_Store()) {
568 Node* st_adr = mem->in(MemNode::Address);
569 intptr_t st_offset = 0;
570 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
571 if (st_base == NULL)
572 break; // inscrutable pointer
573 if (st_offset != offset && st_offset != Type::OffsetBot) {
574 const int MAX_STORE = BytesPerLong;
575 if (st_offset >= offset + size_in_bytes ||
576 st_offset <= offset - MAX_STORE ||
577 st_offset <= offset - mem->as_Store()->memory_size()) {
578 // Success: The offsets are provably independent.
579 // (You may ask, why not just test st_offset != offset and be done?
580 // The answer is that stores of different sizes can co-exist
581 // in the same sequence of RawMem effects. We sometimes initialize
582 // a whole 'tile' of array elements with a single jint or jlong.)
583 mem = mem->in(MemNode::Memory);
584 continue; // (a) advance through independent store memory
585 }
586 }
587 if (st_base != base &&
588 detect_ptr_independence(base, alloc,
589 st_base,
590 AllocateNode::Ideal_allocation(st_base, phase),
591 phase)) {
592 // Success: The bases are provably independent.
593 mem = mem->in(MemNode::Memory);
594 continue; // (a) advance through independent store memory
595 }
596
597 // (b) At this point, if the bases or offsets do not agree, we lose,
598 // since we have not managed to prove 'this' and 'mem' independent.
599 if (st_base == base && st_offset == offset) {
600 return mem; // let caller handle steps (c), (d)
601 }
602
603 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
604 InitializeNode* st_init = mem->in(0)->as_Initialize();
605 AllocateNode* st_alloc = st_init->allocation();
606 if (st_alloc == NULL)
607 break; // something degenerated
608 bool known_identical = false;
609 bool known_independent = false;
610 if (alloc == st_alloc)
611 known_identical = true;
612 else if (alloc != NULL)
613 known_independent = true;
614 else if (all_controls_dominate(this, st_alloc))
615 known_independent = true;
616
617 if (known_independent) {
618 // The bases are provably independent: Either they are
619 // manifestly distinct allocations, or else the control
620 // of this load dominates the store's allocation.
621 int alias_idx = phase->C->get_alias_index(adr_type());
622 if (alias_idx == Compile::AliasIdxRaw) {
623 mem = st_alloc->in(TypeFunc::Memory);
624 } else {
625 mem = st_init->memory(alias_idx);
626 }
627 continue; // (a) advance through independent store memory
628 }
629
630 // (b) at this point, if we are not looking at a store initializing
631 // the same allocation we are loading from, we lose.
632 if (known_identical) {
633 // From caller, can_see_stored_value will consult find_captured_store.
634 return mem; // let caller handle steps (c), (d)
635 }
636
637 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
638 if (prev != mem) {
639 // Found an arraycopy but it doesn't affect that load
640 continue;
641 }
642 // Found an arraycopy that may affect that load
643 return mem;
644 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
645 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
646 if (mem->is_Proj() && mem->in(0)->is_Call()) {
647 // ArrayCopyNodes processed here as well.
648 CallNode *call = mem->in(0)->as_Call();
649 if (!call->may_modify(addr_t, phase)) {
650 mem = call->in(TypeFunc::Memory);
651 continue; // (a) advance through independent call memory
652 }
653 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
654 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase)) {
655 break;
656 }
657 mem = mem->in(0)->in(TypeFunc::Memory);
658 continue; // (a) advance through independent MemBar memory
659 } else if (mem->is_ClearArray()) {
660 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
661 // (the call updated 'mem' value)
662 continue; // (a) advance through independent allocation memory
663 } else {
664 // Can not bypass initialization of the instance
665 // we are looking for.
666 return mem;
667 }
668 } else if (mem->is_MergeMem()) {
669 int alias_idx = phase->C->get_alias_index(adr_type());
670 mem = mem->as_MergeMem()->memory_at(alias_idx);
671 continue; // (a) advance through independent MergeMem memory
672 }
673 }
674
675 // Unless there is an explicit 'continue', we must bail out here,
676 // because 'mem' is an inscrutable memory state (e.g., a call).
677 break;
678 }
679
680 return NULL; // bail out
681 }
682
683 //----------------------calculate_adr_type-------------------------------------
684 // Helper function. Notices when the given type of address hits top or bottom.
685 // Also, asserts a cross-check of the type against the expected address type.
686 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
687 if (t == Type::TOP) return NULL; // does not touch memory any more?
688 #ifdef PRODUCT
689 cross_check = NULL;
690 #else
691 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
692 #endif
693 const TypePtr* tp = t->isa_ptr();
694 if (tp == NULL) {
695 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
696 return TypePtr::BOTTOM; // touches lots of memory
697 } else {
698 #ifdef ASSERT
699 // %%%% [phh] We don't check the alias index if cross_check is
700 // TypeRawPtr::BOTTOM. Needs to be investigated.
701 if (cross_check != NULL &&
702 cross_check != TypePtr::BOTTOM &&
703 cross_check != TypeRawPtr::BOTTOM) {
704 // Recheck the alias index, to see if it has changed (due to a bug).
705 Compile* C = Compile::current();
706 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
707 "must stay in the original alias category");
708 // The type of the address must be contained in the adr_type,
709 // disregarding "null"-ness.
710 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
711 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
712 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
713 "real address must not escape from expected memory type");
714 }
715 #endif
716 return tp;
717 }
718 }
719
720 //=============================================================================
721 // Should LoadNode::Ideal() attempt to remove control edges?
722 bool LoadNode::can_remove_control() const {
723 return true;
724 }
725 uint LoadNode::size_of() const { return sizeof(*this); }
726 uint LoadNode::cmp( const Node &n ) const
727 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
728 const Type *LoadNode::bottom_type() const { return _type; }
729 uint LoadNode::ideal_reg() const {
730 return _type->ideal_reg();
731 }
732
733 #ifndef PRODUCT
734 void LoadNode::dump_spec(outputStream *st) const {
735 MemNode::dump_spec(st);
736 if( !Verbose && !WizardMode ) {
737 // standard dump does this in Verbose and WizardMode
738 st->print(" #"); _type->dump_on(st);
739 }
740 if (!_depends_only_on_test) {
741 st->print(" (does not depend only on test)");
742 }
743 }
744 #endif
745
746 #ifdef ASSERT
747 //----------------------------is_immutable_value-------------------------------
748 // Helper function to allow a raw load without control edge for some cases
749 bool LoadNode::is_immutable_value(Node* adr) {
750 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
751 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
752 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
753 in_bytes(JavaThread::osthread_offset())));
754 }
755 #endif
756
757 //----------------------------LoadNode::make-----------------------------------
758 // Polymorphic factory method:
759 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo, ControlDependency control_dependency) {
760 Compile* C = gvn.C;
761
762 // sanity check the alias category against the created node type
763 assert(!(adr_type->isa_oopptr() &&
764 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
765 "use LoadKlassNode instead");
766 assert(!(adr_type->isa_aryptr() &&
767 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
768 "use LoadRangeNode instead");
769 // Check control edge of raw loads
770 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
771 // oop will be recorded in oop map if load crosses safepoint
772 rt->isa_oopptr() || is_immutable_value(adr),
773 "raw memory operations should have control edge");
774 switch (bt) {
775 case T_BOOLEAN: return new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency);
776 case T_BYTE: return new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency);
777 case T_INT: return new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency);
778 case T_CHAR: return new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency);
779 case T_SHORT: return new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency);
780 case T_LONG: return new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency);
781 case T_FLOAT: return new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency);
782 case T_DOUBLE: return new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency);
783 case T_ADDRESS: return new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
784 case T_OBJECT:
785 #ifdef _LP64
786 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
787 Node* load = gvn.transform(new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency));
788 return new DecodeNNode(load, load->bottom_type()->make_ptr());
789 } else
790 #endif
791 {
792 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
793 return new LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo, control_dependency);
794 }
795 }
796 ShouldNotReachHere();
797 return (LoadNode*)NULL;
798 }
799
800 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) {
801 bool require_atomic = true;
802 return new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
803 }
804
805 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo, ControlDependency control_dependency) {
806 bool require_atomic = true;
807 return new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
808 }
809
810
811
812 //------------------------------hash-------------------------------------------
813 uint LoadNode::hash() const {
814 // unroll addition of interesting fields
815 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
816 }
817
818 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
819 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
820 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
821 bool is_stable_ary = FoldStableValues &&
822 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
823 tp->isa_aryptr()->is_stable();
824
825 return (eliminate_boxing && non_volatile) || is_stable_ary;
826 }
827
828 return false;
829 }
830
831 // Is the value loaded previously stored by an arraycopy? If so return
832 // a load node that reads from the source array so we may be able to
833 // optimize out the ArrayCopy node later.
834 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseTransform* phase) const {
835 Node* ld_adr = in(MemNode::Address);
836 intptr_t ld_off = 0;
837 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
838 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
839 if (ac != NULL) {
840 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
841
842 Node* ld = clone();
843 if (ac->as_ArrayCopy()->is_clonebasic()) {
844 assert(ld_alloc != NULL, "need an alloc");
845 Node* addp = in(MemNode::Address)->clone();
846 assert(addp->is_AddP(), "address must be addp");
847 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern");
848 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern");
849 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)->in(AddPNode::Base));
850 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)->in(AddPNode::Address));
851 ld->set_req(MemNode::Address, phase->transform(addp));
852 if (in(0) != NULL) {
853 assert(ld_alloc->in(0) != NULL, "alloc must have control");
854 ld->set_req(0, ld_alloc->in(0));
855 }
856 } else {
857 Node* addp = in(MemNode::Address)->clone();
858 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
859 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src));
860 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src));
861
862 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
863 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
864 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
865 uint shift = exact_log2(type2aelembytes(ary_elem));
866
867 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
868 #ifdef _LP64
869 diff = phase->transform(new ConvI2LNode(diff));
870 #endif
871 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
872
873 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
874 addp->set_req(AddPNode::Offset, offset);
875 ld->set_req(MemNode::Address, phase->transform(addp));
876
877 if (in(0) != NULL) {
878 assert(ac->in(0) != NULL, "alloc must have control");
879 ld->set_req(0, ac->in(0));
880 }
881 }
882 // load depends on the tests that validate the arraycopy
883 ld->as_Load()->_depends_only_on_test = Pinned;
884 return ld;
885 }
886 return NULL;
887 }
888
889
890 //---------------------------can_see_stored_value------------------------------
891 // This routine exists to make sure this set of tests is done the same
892 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
893 // will change the graph shape in a way which makes memory alive twice at the
894 // same time (uses the Oracle model of aliasing), then some
895 // LoadXNode::Identity will fold things back to the equivalence-class model
896 // of aliasing.
897 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
898 Node* ld_adr = in(MemNode::Address);
899 intptr_t ld_off = 0;
900 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
901 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
902 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
903 // This is more general than load from boxing objects.
904 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
905 uint alias_idx = atp->index();
906 bool final = !atp->is_rewritable();
907 Node* result = NULL;
908 Node* current = st;
909 // Skip through chains of MemBarNodes checking the MergeMems for
910 // new states for the slice of this load. Stop once any other
911 // kind of node is encountered. Loads from final memory can skip
912 // through any kind of MemBar but normal loads shouldn't skip
913 // through MemBarAcquire since the could allow them to move out of
914 // a synchronized region.
915 while (current->is_Proj()) {
916 int opc = current->in(0)->Opcode();
917 if ((final && (opc == Op_MemBarAcquire ||
918 opc == Op_MemBarAcquireLock ||
919 opc == Op_LoadFence)) ||
920 opc == Op_MemBarRelease ||
921 opc == Op_StoreFence ||
922 opc == Op_MemBarReleaseLock ||
923 opc == Op_MemBarStoreStore ||
924 opc == Op_MemBarCPUOrder) {
925 Node* mem = current->in(0)->in(TypeFunc::Memory);
926 if (mem->is_MergeMem()) {
927 MergeMemNode* merge = mem->as_MergeMem();
928 Node* new_st = merge->memory_at(alias_idx);
929 if (new_st == merge->base_memory()) {
930 // Keep searching
931 current = new_st;
932 continue;
933 }
934 // Save the new memory state for the slice and fall through
935 // to exit.
936 result = new_st;
937 }
938 }
939 break;
940 }
941 if (result != NULL) {
942 st = result;
943 }
944 }
945
946 // Loop around twice in the case Load -> Initialize -> Store.
947 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
948 for (int trip = 0; trip <= 1; trip++) {
949
950 if (st->is_Store()) {
951 Node* st_adr = st->in(MemNode::Address);
952 if (!phase->eqv(st_adr, ld_adr)) {
953 // Try harder before giving up... Match raw and non-raw pointers.
954 intptr_t st_off = 0;
955 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
956 if (alloc == NULL) return NULL;
957 if (alloc != ld_alloc) return NULL;
958 if (ld_off != st_off) return NULL;
959 // At this point we have proven something like this setup:
960 // A = Allocate(...)
961 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
962 // S = StoreQ(, AddP(, A.Parm , #Off), V)
963 // (Actually, we haven't yet proven the Q's are the same.)
964 // In other words, we are loading from a casted version of
965 // the same pointer-and-offset that we stored to.
966 // Thus, we are able to replace L by V.
967 }
968 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
969 if (store_Opcode() != st->Opcode())
970 return NULL;
971 return st->in(MemNode::ValueIn);
972 }
973
974 // A load from a freshly-created object always returns zero.
975 // (This can happen after LoadNode::Ideal resets the load's memory input
976 // to find_captured_store, which returned InitializeNode::zero_memory.)
977 if (st->is_Proj() && st->in(0)->is_Allocate() &&
978 (st->in(0) == ld_alloc) &&
979 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
980 // return a zero value for the load's basic type
981 // (This is one of the few places where a generic PhaseTransform
982 // can create new nodes. Think of it as lazily manifesting
983 // virtually pre-existing constants.)
984 return phase->zerocon(memory_type());
985 }
986
987 // A load from an initialization barrier can match a captured store.
988 if (st->is_Proj() && st->in(0)->is_Initialize()) {
989 InitializeNode* init = st->in(0)->as_Initialize();
990 AllocateNode* alloc = init->allocation();
991 if ((alloc != NULL) && (alloc == ld_alloc)) {
992 // examine a captured store value
993 st = init->find_captured_store(ld_off, memory_size(), phase);
994 if (st != NULL) {
995 continue; // take one more trip around
996 }
997 }
998 }
999
1000 // Load boxed value from result of valueOf() call is input parameter.
1001 if (this->is_Load() && ld_adr->is_AddP() &&
1002 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1003 intptr_t ignore = 0;
1004 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1005 base = ShenandoahBarrierNode::skip_through_barrier(base);
1006 if (base != NULL && base->is_Proj() &&
1007 base->as_Proj()->_con == TypeFunc::Parms &&
1008 base->in(0)->is_CallStaticJava() &&
1009 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1010 return base->in(0)->in(TypeFunc::Parms);
1011 }
1012 }
1013
1014 break;
1015 }
1016
1017 return NULL;
1018 }
1019
1020 //----------------------is_instance_field_load_with_local_phi------------------
1021 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1022 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1023 in(Address)->is_AddP() ) {
1024 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1025 // Only instances and boxed values.
1026 if( t_oop != NULL &&
1027 (t_oop->is_ptr_to_boxed_value() ||
1028 t_oop->is_known_instance_field()) &&
1029 t_oop->offset() != Type::OffsetBot &&
1030 t_oop->offset() != Type::OffsetTop) {
1031 return true;
1032 }
1033 }
1034 return false;
1035 }
1036
1037 //------------------------------Identity---------------------------------------
1038 // Loads are identity if previous store is to same address
1039 Node *LoadNode::Identity( PhaseTransform *phase ) {
1040 // If the previous store-maker is the right kind of Store, and the store is
1041 // to the same address, then we are equal to the value stored.
1042 Node* mem = in(Memory);
1043 Node* value = can_see_stored_value(mem, phase);
1044 if( value ) {
1045 // byte, short & char stores truncate naturally.
1046 // A load has to load the truncated value which requires
1047 // some sort of masking operation and that requires an
1048 // Ideal call instead of an Identity call.
1049 if (memory_size() < BytesPerInt) {
1050 // If the input to the store does not fit with the load's result type,
1051 // it must be truncated via an Ideal call.
1052 if (!phase->type(value)->higher_equal(phase->type(this)))
1053 return this;
1054 }
1055 // (This works even when value is a Con, but LoadNode::Value
1056 // usually runs first, producing the singleton type of the Con.)
1057 return value;
1058 }
1059
1060 // Search for an existing data phi which was generated before for the same
1061 // instance's field to avoid infinite generation of phis in a loop.
1062 Node *region = mem->in(0);
1063 if (is_instance_field_load_with_local_phi(region)) {
1064 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1065 int this_index = phase->C->get_alias_index(addr_t);
1066 int this_offset = addr_t->offset();
1067 int this_iid = addr_t->instance_id();
1068 if (!addr_t->is_known_instance() &&
1069 addr_t->is_ptr_to_boxed_value()) {
1070 // Use _idx of address base (could be Phi node) for boxed values.
1071 intptr_t ignore = 0;
1072 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1073 this_iid = base->_idx;
1074 }
1075 const Type* this_type = bottom_type();
1076 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1077 Node* phi = region->fast_out(i);
1078 if (phi->is_Phi() && phi != mem &&
1079 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {
1080 return phi;
1081 }
1082 }
1083 }
1084
1085 return this;
1086 }
1087
1088 // We're loading from an object which has autobox behaviour.
1089 // If this object is result of a valueOf call we'll have a phi
1090 // merging a newly allocated object and a load from the cache.
1091 // We want to replace this load with the original incoming
1092 // argument to the valueOf call.
1093 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1094 assert(phase->C->eliminate_boxing(), "sanity");
1095 intptr_t ignore = 0;
1096 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1097 if ((base == NULL) || base->is_Phi()) {
1098 // Push the loads from the phi that comes from valueOf up
1099 // through it to allow elimination of the loads and the recovery
1100 // of the original value. It is done in split_through_phi().
1101 return NULL;
1102 } else if (base->is_Load() ||
1103 base->is_DecodeN() && base->in(1)->is_Load()) {
1104 // Eliminate the load of boxed value for integer types from the cache
1105 // array by deriving the value from the index into the array.
1106 // Capture the offset of the load and then reverse the computation.
1107
1108 // Get LoadN node which loads a boxing object from 'cache' array.
1109 if (base->is_DecodeN()) {
1110 base = base->in(1);
1111 }
1112 if (!base->in(Address)->is_AddP()) {
1113 return NULL; // Complex address
1114 }
1115 AddPNode* address = base->in(Address)->as_AddP();
1116 Node* cache_base = address->in(AddPNode::Base);
1117 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1118 // Get ConP node which is static 'cache' field.
1119 cache_base = cache_base->in(1);
1120 }
1121 if ((cache_base != NULL) && cache_base->is_Con()) {
1122 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1123 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1124 Node* elements[4];
1125 int shift = exact_log2(type2aelembytes(T_OBJECT));
1126 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1127 if ((count > 0) && elements[0]->is_Con() &&
1128 ((count == 1) ||
1129 (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1130 elements[1]->in(2) == phase->intcon(shift))) {
1131 ciObjArray* array = base_type->const_oop()->as_obj_array();
1132 // Fetch the box object cache[0] at the base of the array and get its value
1133 ciInstance* box = array->obj_at(0)->as_instance();
1134 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1135 assert(ik->is_box_klass(), "sanity");
1136 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1137 if (ik->nof_nonstatic_fields() == 1) {
1138 // This should be true nonstatic_field_at requires calling
1139 // nof_nonstatic_fields so check it anyway
1140 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1141 BasicType bt = c.basic_type();
1142 // Only integer types have boxing cache.
1143 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1144 bt == T_BYTE || bt == T_SHORT ||
1145 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt)));
1146 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1147 if (cache_low != (int)cache_low) {
1148 return NULL; // should not happen since cache is array indexed by value
1149 }
1150 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1151 if (offset != (int)offset) {
1152 return NULL; // should not happen since cache is array indexed by value
1153 }
1154 // Add up all the offsets making of the address of the load
1155 Node* result = elements[0];
1156 for (int i = 1; i < count; i++) {
1157 result = phase->transform(new AddXNode(result, elements[i]));
1158 }
1159 // Remove the constant offset from the address and then
1160 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1161 // remove the scaling of the offset to recover the original index.
1162 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1163 // Peel the shift off directly but wrap it in a dummy node
1164 // since Ideal can't return existing nodes
1165 result = new RShiftXNode(result->in(1), phase->intcon(0));
1166 } else if (result->is_Add() && result->in(2)->is_Con() &&
1167 result->in(1)->Opcode() == Op_LShiftX &&
1168 result->in(1)->in(2) == phase->intcon(shift)) {
1169 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1170 // but for boxing cache access we know that X<<Z will not overflow
1171 // (there is range check) so we do this optimizatrion by hand here.
1172 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1173 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1174 } else {
1175 result = new RShiftXNode(result, phase->intcon(shift));
1176 }
1177 #ifdef _LP64
1178 if (bt != T_LONG) {
1179 result = new ConvL2INode(phase->transform(result));
1180 }
1181 #else
1182 if (bt == T_LONG) {
1183 result = new ConvI2LNode(phase->transform(result));
1184 }
1185 #endif
1186 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1187 // Need to preserve unboxing load type if it is unsigned.
1188 switch(this->Opcode()) {
1189 case Op_LoadUB:
1190 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1191 break;
1192 case Op_LoadUS:
1193 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1194 break;
1195 }
1196 return result;
1197 }
1198 }
1199 }
1200 }
1201 }
1202 return NULL;
1203 }
1204
1205 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1206 Node* region = phi->in(0);
1207 if (region == NULL) {
1208 return false; // Wait stable graph
1209 }
1210 uint cnt = phi->req();
1211 for (uint i = 1; i < cnt; i++) {
1212 Node* rc = region->in(i);
1213 if (rc == NULL || phase->type(rc) == Type::TOP)
1214 return false; // Wait stable graph
1215 Node* in = phi->in(i);
1216 if (in == NULL || phase->type(in) == Type::TOP)
1217 return false; // Wait stable graph
1218 }
1219 return true;
1220 }
1221 //------------------------------split_through_phi------------------------------
1222 // Split instance or boxed field load through Phi.
1223 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1224 Node* mem = in(Memory);
1225 Node* address = in(Address);
1226 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1227
1228 assert((t_oop != NULL) &&
1229 (t_oop->is_known_instance_field() ||
1230 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1231
1232 Compile* C = phase->C;
1233 intptr_t ignore = 0;
1234 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1235 bool base_is_phi = (base != NULL) && base->is_Phi();
1236 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1237 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1238 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1239
1240 if (!((mem->is_Phi() || base_is_phi) &&
1241 (load_boxed_values || t_oop->is_known_instance_field()))) {
1242 return NULL; // memory is not Phi
1243 }
1244
1245 if (mem->is_Phi()) {
1246 if (!stable_phi(mem->as_Phi(), phase)) {
1247 return NULL; // Wait stable graph
1248 }
1249 uint cnt = mem->req();
1250 // Check for loop invariant memory.
1251 if (cnt == 3) {
1252 for (uint i = 1; i < cnt; i++) {
1253 Node* in = mem->in(i);
1254 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1255 if (m == mem) {
1256 set_req(Memory, mem->in(cnt - i));
1257 return this; // made change
1258 }
1259 }
1260 }
1261 }
1262 if (base_is_phi) {
1263 if (!stable_phi(base->as_Phi(), phase)) {
1264 return NULL; // Wait stable graph
1265 }
1266 uint cnt = base->req();
1267 // Check for loop invariant memory.
1268 if (cnt == 3) {
1269 for (uint i = 1; i < cnt; i++) {
1270 if (base->in(i) == base) {
1271 return NULL; // Wait stable graph
1272 }
1273 }
1274 }
1275 }
1276
1277 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1278
1279 // Split through Phi (see original code in loopopts.cpp).
1280 assert(C->have_alias_type(t_oop), "instance should have alias type");
1281
1282 // Do nothing here if Identity will find a value
1283 // (to avoid infinite chain of value phis generation).
1284 if (!phase->eqv(this, this->Identity(phase)))
1285 return NULL;
1286
1287 // Select Region to split through.
1288 Node* region;
1289 if (!base_is_phi) {
1290 assert(mem->is_Phi(), "sanity");
1291 region = mem->in(0);
1292 // Skip if the region dominates some control edge of the address.
1293 if (!MemNode::all_controls_dominate(address, region))
1294 return NULL;
1295 } else if (!mem->is_Phi()) {
1296 assert(base_is_phi, "sanity");
1297 region = base->in(0);
1298 // Skip if the region dominates some control edge of the memory.
1299 if (!MemNode::all_controls_dominate(mem, region))
1300 return NULL;
1301 } else if (base->in(0) != mem->in(0)) {
1302 assert(base_is_phi && mem->is_Phi(), "sanity");
1303 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1304 region = base->in(0);
1305 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1306 region = mem->in(0);
1307 } else {
1308 return NULL; // complex graph
1309 }
1310 } else {
1311 assert(base->in(0) == mem->in(0), "sanity");
1312 region = mem->in(0);
1313 }
1314
1315 const Type* this_type = this->bottom_type();
1316 int this_index = C->get_alias_index(t_oop);
1317 int this_offset = t_oop->offset();
1318 int this_iid = t_oop->instance_id();
1319 if (!t_oop->is_known_instance() && load_boxed_values) {
1320 // Use _idx of address base for boxed values.
1321 this_iid = base->_idx;
1322 }
1323 PhaseIterGVN* igvn = phase->is_IterGVN();
1324 Node* phi = new PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1325 for (uint i = 1; i < region->req(); i++) {
1326 Node* x;
1327 Node* the_clone = NULL;
1328 if (region->in(i) == C->top()) {
1329 x = C->top(); // Dead path? Use a dead data op
1330 } else {
1331 x = this->clone(); // Else clone up the data op
1332 the_clone = x; // Remember for possible deletion.
1333 // Alter data node to use pre-phi inputs
1334 if (this->in(0) == region) {
1335 x->set_req(0, region->in(i));
1336 } else {
1337 x->set_req(0, NULL);
1338 }
1339 if (mem->is_Phi() && (mem->in(0) == region)) {
1340 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1341 }
1342 if (address->is_Phi() && address->in(0) == region) {
1343 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1344 }
1345 if (base_is_phi && (base->in(0) == region)) {
1346 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1347 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1348 x->set_req(Address, adr_x);
1349 }
1350 }
1351 // Check for a 'win' on some paths
1352 const Type *t = x->Value(igvn);
1353
1354 bool singleton = t->singleton();
1355
1356 // See comments in PhaseIdealLoop::split_thru_phi().
1357 if (singleton && t == Type::TOP) {
1358 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1359 }
1360
1361 if (singleton) {
1362 x = igvn->makecon(t);
1363 } else {
1364 // We now call Identity to try to simplify the cloned node.
1365 // Note that some Identity methods call phase->type(this).
1366 // Make sure that the type array is big enough for
1367 // our new node, even though we may throw the node away.
1368 // (This tweaking with igvn only works because x is a new node.)
1369 igvn->set_type(x, t);
1370 // If x is a TypeNode, capture any more-precise type permanently into Node
1371 // otherwise it will be not updated during igvn->transform since
1372 // igvn->type(x) is set to x->Value() already.
1373 x->raise_bottom_type(t);
1374 Node *y = x->Identity(igvn);
1375 if (y != x) {
1376 x = y;
1377 } else {
1378 y = igvn->hash_find_insert(x);
1379 if (y) {
1380 x = y;
1381 } else {
1382 // Else x is a new node we are keeping
1383 // We do not need register_new_node_with_optimizer
1384 // because set_type has already been called.
1385 igvn->_worklist.push(x);
1386 }
1387 }
1388 }
1389 if (x != the_clone && the_clone != NULL) {
1390 igvn->remove_dead_node(the_clone);
1391 }
1392 phi->set_req(i, x);
1393 }
1394 // Record Phi
1395 igvn->register_new_node_with_optimizer(phi);
1396 return phi;
1397 }
1398
1399 //------------------------------Ideal------------------------------------------
1400 // If the load is from Field memory and the pointer is non-null, it might be possible to
1401 // zero out the control input.
1402 // If the offset is constant and the base is an object allocation,
1403 // try to hook me up to the exact initializing store.
1404 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1405 Node* p = MemNode::Ideal_common(phase, can_reshape);
1406 if (p) return (p == NodeSentinel) ? NULL : p;
1407
1408 Node* ctrl = in(MemNode::Control);
1409 Node* address = in(MemNode::Address);
1410 bool progress = false;
1411
1412 // Skip up past a SafePoint control. Cannot do this for Stores because
1413 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1414 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1415 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1416 ctrl = ctrl->in(0);
1417 set_req(MemNode::Control,ctrl);
1418 progress = true;
1419 }
1420
1421 intptr_t ignore = 0;
1422 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1423 if (base != NULL
1424 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1425 // Check for useless control edge in some common special cases
1426 if (in(MemNode::Control) != NULL
1427 && can_remove_control()
1428 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1429 && all_controls_dominate(base, phase->C->start())) {
1430 // A method-invariant, non-null address (constant or 'this' argument).
1431 set_req(MemNode::Control, NULL);
1432 progress = true;
1433 }
1434 }
1435
1436 Node* mem = in(MemNode::Memory);
1437 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1438
1439 if (can_reshape && (addr_t != NULL)) {
1440 // try to optimize our memory input
1441 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1442 if (opt_mem != mem) {
1443 set_req(MemNode::Memory, opt_mem);
1444 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1445 return this;
1446 }
1447 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1448 if ((t_oop != NULL) &&
1449 (t_oop->is_known_instance_field() ||
1450 t_oop->is_ptr_to_boxed_value())) {
1451 PhaseIterGVN *igvn = phase->is_IterGVN();
1452 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1453 // Delay this transformation until memory Phi is processed.
1454 phase->is_IterGVN()->_worklist.push(this);
1455 return NULL;
1456 }
1457 // Split instance field load through Phi.
1458 Node* result = split_through_phi(phase);
1459 if (result != NULL) return result;
1460
1461 if (t_oop->is_ptr_to_boxed_value()) {
1462 Node* result = eliminate_autobox(phase);
1463 if (result != NULL) return result;
1464 }
1465 }
1466 }
1467
1468 // Is there a dominating load that loads the same value? Leave
1469 // anything that is not a load of a field/array element (like
1470 // barriers etc.) alone
1471 if (in(0) != NULL && adr_type() != TypeRawPtr::BOTTOM && can_reshape) {
1472 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1473 Node *use = mem->fast_out(i);
1474 if (use != this &&
1475 use->Opcode() == Opcode() &&
1476 use->in(0) != NULL &&
1477 use->in(0) != in(0) &&
1478 use->in(Address) == in(Address)) {
1479 Node* ctl = in(0);
1480 for (int i = 0; i < 10 && ctl != NULL; i++) {
1481 ctl = IfNode::up_one_dom(ctl);
1482 if (ctl == use->in(0)) {
1483 set_req(0, use->in(0));
1484 return this;
1485 }
1486 }
1487 }
1488 }
1489 }
1490
1491 // Check for prior store with a different base or offset; make Load
1492 // independent. Skip through any number of them. Bail out if the stores
1493 // are in an endless dead cycle and report no progress. This is a key
1494 // transform for Reflection. However, if after skipping through the Stores
1495 // we can't then fold up against a prior store do NOT do the transform as
1496 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1497 // array memory alive twice: once for the hoisted Load and again after the
1498 // bypassed Store. This situation only works if EVERYBODY who does
1499 // anti-dependence work knows how to bypass. I.e. we need all
1500 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1501 // the alias index stuff. So instead, peek through Stores and IFF we can
1502 // fold up, do so.
1503 Node* prev_mem = find_previous_store(phase);
1504 if (prev_mem != NULL) {
1505 Node* value = can_see_arraycopy_value(prev_mem, phase);
1506 if (value != NULL) {
1507 return value;
1508 }
1509 }
1510 // Steps (a), (b): Walk past independent stores to find an exact match.
1511 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1512 // (c) See if we can fold up on the spot, but don't fold up here.
1513 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1514 // just return a prior value, which is done by Identity calls.
1515 if (can_see_stored_value(prev_mem, phase)) {
1516 // Make ready for step (d):
1517 set_req(MemNode::Memory, prev_mem);
1518 return this;
1519 }
1520 }
1521
1522 return progress ? this : NULL;
1523 }
1524
1525 // Helper to recognize certain Klass fields which are invariant across
1526 // some group of array types (e.g., int[] or all T[] where T < Object).
1527 const Type*
1528 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1529 ciKlass* klass) const {
1530 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1531 // The field is Klass::_modifier_flags. Return its (constant) value.
1532 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1533 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1534 return TypeInt::make(klass->modifier_flags());
1535 }
1536 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1537 // The field is Klass::_access_flags. Return its (constant) value.
1538 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1539 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1540 return TypeInt::make(klass->access_flags());
1541 }
1542 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1543 // The field is Klass::_layout_helper. Return its constant value if known.
1544 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1545 return TypeInt::make(klass->layout_helper());
1546 }
1547
1548 // No match.
1549 return NULL;
1550 }
1551
1552 // Try to constant-fold a stable array element.
1553 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) {
1554 assert(ary->const_oop(), "array should be constant");
1555 assert(ary->is_stable(), "array should be stable");
1556
1557 // Decode the results of GraphKit::array_element_address.
1558 ciArray* aobj = ary->const_oop()->as_array();
1559 ciConstant con = aobj->element_value_by_offset(off);
1560
1561 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) {
1562 const Type* con_type = Type::make_from_constant(con);
1563 if (con_type != NULL) {
1564 if (con_type->isa_aryptr()) {
1565 // Join with the array element type, in case it is also stable.
1566 int dim = ary->stable_dimension();
1567 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1);
1568 }
1569 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) {
1570 con_type = con_type->make_narrowoop();
1571 }
1572 #ifndef PRODUCT
1573 if (TraceIterativeGVN) {
1574 tty->print("FoldStableValues: array element [off=%d]: con_type=", off);
1575 con_type->dump(); tty->cr();
1576 }
1577 #endif //PRODUCT
1578 return con_type;
1579 }
1580 }
1581 return NULL;
1582 }
1583
1584 //------------------------------Value-----------------------------------------
1585 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1586 // Either input is TOP ==> the result is TOP
1587 Node* mem = in(MemNode::Memory);
1588 const Type *t1 = phase->type(mem);
1589 if (t1 == Type::TOP) return Type::TOP;
1590 Node* adr = in(MemNode::Address);
1591 const TypePtr* tp = phase->type(adr)->isa_ptr();
1592 if (tp == NULL || tp->empty()) return Type::TOP;
1593 int off = tp->offset();
1594 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1595 Compile* C = phase->C;
1596
1597 // Try to guess loaded type from pointer type
1598 if (tp->isa_aryptr()) {
1599 const TypeAryPtr* ary = tp->is_aryptr();
1600 const Type* t = ary->elem();
1601
1602 // Determine whether the reference is beyond the header or not, by comparing
1603 // the offset against the offset of the start of the array's data.
1604 // Different array types begin at slightly different offsets (12 vs. 16).
1605 // We choose T_BYTE as an example base type that is least restrictive
1606 // as to alignment, which will therefore produce the smallest
1607 // possible base offset.
1608 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1609 const bool off_beyond_header = (off != BrooksPointer::BYTE_OFFSET || !UseShenandoahGC) && ((uint)off >= (uint)min_base_off);
1610
1611 // Try to constant-fold a stable array element.
1612 if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) {
1613 // Make sure the reference is not into the header and the offset is constant
1614 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1615 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type());
1616 if (con_type != NULL) {
1617 return con_type;
1618 }
1619 }
1620 }
1621
1622 // Don't do this for integer types. There is only potential profit if
1623 // the element type t is lower than _type; that is, for int types, if _type is
1624 // more restrictive than t. This only happens here if one is short and the other
1625 // char (both 16 bits), and in those cases we've made an intentional decision
1626 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1627 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1628 //
1629 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1630 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1631 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1632 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1633 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1634 // In fact, that could have been the original type of p1, and p1 could have
1635 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1636 // expression (LShiftL quux 3) independently optimized to the constant 8.
1637 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1638 && (_type->isa_vect() == NULL)
1639 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1640 // t might actually be lower than _type, if _type is a unique
1641 // concrete subclass of abstract class t.
1642 if (off_beyond_header) { // is the offset beyond the header?
1643 const Type* jt = t->join_speculative(_type);
1644 // In any case, do not allow the join, per se, to empty out the type.
1645 if (jt->empty() && !t->empty()) {
1646 // This can happen if a interface-typed array narrows to a class type.
1647 jt = _type;
1648 }
1649 #ifdef ASSERT
1650 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1651 // The pointers in the autobox arrays are always non-null
1652 Node* base = adr->in(AddPNode::Base);
1653 if ((base != NULL) && base->is_DecodeN()) {
1654 // Get LoadN node which loads IntegerCache.cache field
1655 base = base->in(1);
1656 }
1657 if ((base != NULL) && base->is_Con()) {
1658 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1659 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1660 // It could be narrow oop
1661 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1662 }
1663 }
1664 }
1665 #endif
1666 return jt;
1667 }
1668 }
1669 } else if (tp->base() == Type::InstPtr) {
1670 ciEnv* env = C->env();
1671 const TypeInstPtr* tinst = tp->is_instptr();
1672 ciKlass* klass = tinst->klass();
1673 assert( off != Type::OffsetBot ||
1674 // arrays can be cast to Objects
1675 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1676 // unsafe field access may not have a constant offset
1677 C->has_unsafe_access(),
1678 "Field accesses must be precise" );
1679 // For oop loads, we expect the _type to be precise
1680 if (klass == env->String_klass() &&
1681 adr->is_AddP() && off != Type::OffsetBot) {
1682 // For constant Strings treat the final fields as compile time constants.
1683 // While we can list what field types java.lang.String has, it is more
1684 // future-proof to handle all possible field types, anticipating future
1685 // changes and experiments in String code.
1686 Node* base = adr->in(AddPNode::Base);
1687 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1688 if (t != NULL && t->singleton()) {
1689 ciField* field = env->String_klass()->get_field_by_offset(off, false);
1690 if (field != NULL && field->is_final()) {
1691 ciObject* string = t->const_oop();
1692 ciConstant constant = string->as_instance()->field_value(field);
1693 // Type::make_from_constant does not handle narrow oops, so handle it here.
1694 // Everything else can use the factory method.
1695 if ((constant.basic_type() == T_ARRAY || constant.basic_type() == T_OBJECT)
1696 && adr->bottom_type()->is_ptr_to_narrowoop()) {
1697 return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1698 } else {
1699 return Type::make_from_constant(constant, true);
1700 }
1701 }
1702 }
1703 }
1704 // Optimizations for constant objects
1705 ciObject* const_oop = tinst->const_oop();
1706 if (const_oop != NULL) {
1707 // For constant Boxed value treat the target field as a compile time constant.
1708 if (tinst->is_ptr_to_boxed_value()) {
1709 return tinst->get_const_boxed_value();
1710 } else
1711 // For constant CallSites treat the target field as a compile time constant.
1712 if (const_oop->is_call_site()) {
1713 ciCallSite* call_site = const_oop->as_call_site();
1714 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1715 if (field != NULL && field->is_call_site_target()) {
1716 ciMethodHandle* target = call_site->get_target();
1717 if (target != NULL) { // just in case
1718 ciConstant constant(T_OBJECT, target);
1719 const Type* t;
1720 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1721 t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1722 } else {
1723 t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1724 }
1725 // Add a dependence for invalidation of the optimization.
1726 if (!call_site->is_constant_call_site()) {
1727 C->dependencies()->assert_call_site_target_value(call_site, target);
1728 }
1729 return t;
1730 }
1731 }
1732 }
1733 }
1734 } else if (tp->base() == Type::KlassPtr) {
1735 assert( off != Type::OffsetBot ||
1736 // arrays can be cast to Objects
1737 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1738 // also allow array-loading from the primary supertype
1739 // array during subtype checks
1740 Opcode() == Op_LoadKlass,
1741 "Field accesses must be precise" );
1742 // For klass/static loads, we expect the _type to be precise
1743 }
1744
1745 const TypeKlassPtr *tkls = tp->isa_klassptr();
1746 if (tkls != NULL && !StressReflectiveCode) {
1747 ciKlass* klass = tkls->klass();
1748 if (klass->is_loaded() && tkls->klass_is_exact()) {
1749 // We are loading a field from a Klass metaobject whose identity
1750 // is known at compile time (the type is "exact" or "precise").
1751 // Check for fields we know are maintained as constants by the VM.
1752 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1753 // The field is Klass::_super_check_offset. Return its (constant) value.
1754 // (Folds up type checking code.)
1755 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1756 return TypeInt::make(klass->super_check_offset());
1757 }
1758 // Compute index into primary_supers array
1759 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1760 // Check for overflowing; use unsigned compare to handle the negative case.
1761 if( depth < ciKlass::primary_super_limit() ) {
1762 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1763 // (Folds up type checking code.)
1764 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1765 ciKlass *ss = klass->super_of_depth(depth);
1766 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1767 }
1768 const Type* aift = load_array_final_field(tkls, klass);
1769 if (aift != NULL) return aift;
1770 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1771 // The field is Klass::_java_mirror. Return its (constant) value.
1772 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1773 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1774 return TypeInstPtr::make(klass->java_mirror());
1775 }
1776 }
1777
1778 // We can still check if we are loading from the primary_supers array at a
1779 // shallow enough depth. Even though the klass is not exact, entries less
1780 // than or equal to its super depth are correct.
1781 if (klass->is_loaded() ) {
1782 ciType *inner = klass;
1783 while( inner->is_obj_array_klass() )
1784 inner = inner->as_obj_array_klass()->base_element_type();
1785 if( inner->is_instance_klass() &&
1786 !inner->as_instance_klass()->flags().is_interface() ) {
1787 // Compute index into primary_supers array
1788 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1789 // Check for overflowing; use unsigned compare to handle the negative case.
1790 if( depth < ciKlass::primary_super_limit() &&
1791 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1792 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1793 // (Folds up type checking code.)
1794 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1795 ciKlass *ss = klass->super_of_depth(depth);
1796 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1797 }
1798 }
1799 }
1800
1801 // If the type is enough to determine that the thing is not an array,
1802 // we can give the layout_helper a positive interval type.
1803 // This will help short-circuit some reflective code.
1804 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1805 && !klass->is_array_klass() // not directly typed as an array
1806 && !klass->is_interface() // specifically not Serializable & Cloneable
1807 && !klass->is_java_lang_Object() // not the supertype of all T[]
1808 ) {
1809 // Note: When interfaces are reliable, we can narrow the interface
1810 // test to (klass != Serializable && klass != Cloneable).
1811 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1812 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1813 // The key property of this type is that it folds up tests
1814 // for array-ness, since it proves that the layout_helper is positive.
1815 // Thus, a generic value like the basic object layout helper works fine.
1816 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1817 }
1818 }
1819
1820 // If we are loading from a freshly-allocated object, produce a zero,
1821 // if the load is provably beyond the header of the object.
1822 // (Also allow a variable load from a fresh array to produce zero.)
1823 const TypeOopPtr *tinst = tp->isa_oopptr();
1824 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1825 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1826 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1827 Node* value = can_see_stored_value(mem,phase);
1828 if (value != NULL && value->is_Con()) {
1829 assert(value->bottom_type()->higher_equal(_type),"sanity");
1830 return value->bottom_type();
1831 }
1832 }
1833
1834 if (is_instance) {
1835 // If we have an instance type and our memory input is the
1836 // programs's initial memory state, there is no matching store,
1837 // so just return a zero of the appropriate type
1838 Node *mem = in(MemNode::Memory);
1839 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1840 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1841 return Type::get_zero_type(_type->basic_type());
1842 }
1843 }
1844 return _type;
1845 }
1846
1847 //------------------------------match_edge-------------------------------------
1848 // Do we Match on this edge index or not? Match only the address.
1849 uint LoadNode::match_edge(uint idx) const {
1850 return idx == MemNode::Address;
1851 }
1852
1853 //--------------------------LoadBNode::Ideal--------------------------------------
1854 //
1855 // If the previous store is to the same address as this load,
1856 // and the value stored was larger than a byte, replace this load
1857 // with the value stored truncated to a byte. If no truncation is
1858 // needed, the replacement is done in LoadNode::Identity().
1859 //
1860 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1861 Node* mem = in(MemNode::Memory);
1862 Node* value = can_see_stored_value(mem,phase);
1863 if( value && !phase->type(value)->higher_equal( _type ) ) {
1864 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1865 return new RShiftINode(result, phase->intcon(24));
1866 }
1867 // Identity call will handle the case where truncation is not needed.
1868 return LoadNode::Ideal(phase, can_reshape);
1869 }
1870
1871 const Type* LoadBNode::Value(PhaseTransform *phase) const {
1872 Node* mem = in(MemNode::Memory);
1873 Node* value = can_see_stored_value(mem,phase);
1874 if (value != NULL && value->is_Con() &&
1875 !value->bottom_type()->higher_equal(_type)) {
1876 // If the input to the store does not fit with the load's result type,
1877 // it must be truncated. We can't delay until Ideal call since
1878 // a singleton Value is needed for split_thru_phi optimization.
1879 int con = value->get_int();
1880 return TypeInt::make((con << 24) >> 24);
1881 }
1882 return LoadNode::Value(phase);
1883 }
1884
1885 //--------------------------LoadUBNode::Ideal-------------------------------------
1886 //
1887 // If the previous store is to the same address as this load,
1888 // and the value stored was larger than a byte, replace this load
1889 // with the value stored truncated to a byte. If no truncation is
1890 // needed, the replacement is done in LoadNode::Identity().
1891 //
1892 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1893 Node* mem = in(MemNode::Memory);
1894 Node* value = can_see_stored_value(mem, phase);
1895 if (value && !phase->type(value)->higher_equal(_type))
1896 return new AndINode(value, phase->intcon(0xFF));
1897 // Identity call will handle the case where truncation is not needed.
1898 return LoadNode::Ideal(phase, can_reshape);
1899 }
1900
1901 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1902 Node* mem = in(MemNode::Memory);
1903 Node* value = can_see_stored_value(mem,phase);
1904 if (value != NULL && value->is_Con() &&
1905 !value->bottom_type()->higher_equal(_type)) {
1906 // If the input to the store does not fit with the load's result type,
1907 // it must be truncated. We can't delay until Ideal call since
1908 // a singleton Value is needed for split_thru_phi optimization.
1909 int con = value->get_int();
1910 return TypeInt::make(con & 0xFF);
1911 }
1912 return LoadNode::Value(phase);
1913 }
1914
1915 //--------------------------LoadUSNode::Ideal-------------------------------------
1916 //
1917 // If the previous store is to the same address as this load,
1918 // and the value stored was larger than a char, replace this load
1919 // with the value stored truncated to a char. If no truncation is
1920 // needed, the replacement is done in LoadNode::Identity().
1921 //
1922 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1923 Node* mem = in(MemNode::Memory);
1924 Node* value = can_see_stored_value(mem,phase);
1925 if( value && !phase->type(value)->higher_equal( _type ) )
1926 return new AndINode(value,phase->intcon(0xFFFF));
1927 // Identity call will handle the case where truncation is not needed.
1928 return LoadNode::Ideal(phase, can_reshape);
1929 }
1930
1931 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1932 Node* mem = in(MemNode::Memory);
1933 Node* value = can_see_stored_value(mem,phase);
1934 if (value != NULL && value->is_Con() &&
1935 !value->bottom_type()->higher_equal(_type)) {
1936 // If the input to the store does not fit with the load's result type,
1937 // it must be truncated. We can't delay until Ideal call since
1938 // a singleton Value is needed for split_thru_phi optimization.
1939 int con = value->get_int();
1940 return TypeInt::make(con & 0xFFFF);
1941 }
1942 return LoadNode::Value(phase);
1943 }
1944
1945 //--------------------------LoadSNode::Ideal--------------------------------------
1946 //
1947 // If the previous store is to the same address as this load,
1948 // and the value stored was larger than a short, replace this load
1949 // with the value stored truncated to a short. If no truncation is
1950 // needed, the replacement is done in LoadNode::Identity().
1951 //
1952 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1953 Node* mem = in(MemNode::Memory);
1954 Node* value = can_see_stored_value(mem,phase);
1955 if( value && !phase->type(value)->higher_equal( _type ) ) {
1956 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
1957 return new RShiftINode(result, phase->intcon(16));
1958 }
1959 // Identity call will handle the case where truncation is not needed.
1960 return LoadNode::Ideal(phase, can_reshape);
1961 }
1962
1963 const Type* LoadSNode::Value(PhaseTransform *phase) const {
1964 Node* mem = in(MemNode::Memory);
1965 Node* value = can_see_stored_value(mem,phase);
1966 if (value != NULL && value->is_Con() &&
1967 !value->bottom_type()->higher_equal(_type)) {
1968 // If the input to the store does not fit with the load's result type,
1969 // it must be truncated. We can't delay until Ideal call since
1970 // a singleton Value is needed for split_thru_phi optimization.
1971 int con = value->get_int();
1972 return TypeInt::make((con << 16) >> 16);
1973 }
1974 return LoadNode::Value(phase);
1975 }
1976
1977 //=============================================================================
1978 //----------------------------LoadKlassNode::make------------------------------
1979 // Polymorphic factory method:
1980 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
1981 // sanity check the alias category against the created node type
1982 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
1983 assert(adr_type != NULL, "expecting TypeKlassPtr");
1984 #ifdef _LP64
1985 if (adr_type->is_ptr_to_narrowklass()) {
1986 assert(UseCompressedClassPointers, "no compressed klasses");
1987 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
1988 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
1989 }
1990 #endif
1991 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1992 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
1993 }
1994
1995 //------------------------------Value------------------------------------------
1996 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1997 return klass_value_common(phase);
1998 }
1999
2000 // In most cases, LoadKlassNode does not have the control input set. If the control
2001 // input is set, it must not be removed (by LoadNode::Ideal()).
2002 bool LoadKlassNode::can_remove_control() const {
2003 return false;
2004 }
2005
2006 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
2007 // Either input is TOP ==> the result is TOP
2008 const Type *t1 = phase->type( in(MemNode::Memory) );
2009 if (t1 == Type::TOP) return Type::TOP;
2010 Node *adr = in(MemNode::Address);
2011 const Type *t2 = phase->type( adr );
2012 if (t2 == Type::TOP) return Type::TOP;
2013 const TypePtr *tp = t2->is_ptr();
2014 if (TypePtr::above_centerline(tp->ptr()) ||
2015 tp->ptr() == TypePtr::Null) return Type::TOP;
2016
2017 // Return a more precise klass, if possible
2018 const TypeInstPtr *tinst = tp->isa_instptr();
2019 if (tinst != NULL) {
2020 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2021 int offset = tinst->offset();
2022 if (ik == phase->C->env()->Class_klass()
2023 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2024 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2025 // We are loading a special hidden field from a Class mirror object,
2026 // the field which points to the VM's Klass metaobject.
2027 ciType* t = tinst->java_mirror_type();
2028 // java_mirror_type returns non-null for compile-time Class constants.
2029 if (t != NULL) {
2030 // constant oop => constant klass
2031 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2032 if (t->is_void()) {
2033 // We cannot create a void array. Since void is a primitive type return null
2034 // klass. Users of this result need to do a null check on the returned klass.
2035 return TypePtr::NULL_PTR;
2036 }
2037 return TypeKlassPtr::make(ciArrayKlass::make(t));
2038 }
2039 if (!t->is_klass()) {
2040 // a primitive Class (e.g., int.class) has NULL for a klass field
2041 return TypePtr::NULL_PTR;
2042 }
2043 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2044 return TypeKlassPtr::make(t->as_klass());
2045 }
2046 // non-constant mirror, so we can't tell what's going on
2047 }
2048 if( !ik->is_loaded() )
2049 return _type; // Bail out if not loaded
2050 if (offset == oopDesc::klass_offset_in_bytes()) {
2051 if (tinst->klass_is_exact()) {
2052 return TypeKlassPtr::make(ik);
2053 }
2054 // See if we can become precise: no subklasses and no interface
2055 // (Note: We need to support verified interfaces.)
2056 if (!ik->is_interface() && !ik->has_subklass()) {
2057 //assert(!UseExactTypes, "this code should be useless with exact types");
2058 // Add a dependence; if any subclass added we need to recompile
2059 if (!ik->is_final()) {
2060 // %%% should use stronger assert_unique_concrete_subtype instead
2061 phase->C->dependencies()->assert_leaf_type(ik);
2062 }
2063 // Return precise klass
2064 return TypeKlassPtr::make(ik);
2065 }
2066
2067 // Return root of possible klass
2068 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2069 }
2070 }
2071
2072 // Check for loading klass from an array
2073 const TypeAryPtr *tary = tp->isa_aryptr();
2074 if( tary != NULL ) {
2075 ciKlass *tary_klass = tary->klass();
2076 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2077 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2078 if (tary->klass_is_exact()) {
2079 return TypeKlassPtr::make(tary_klass);
2080 }
2081 ciArrayKlass *ak = tary->klass()->as_array_klass();
2082 // If the klass is an object array, we defer the question to the
2083 // array component klass.
2084 if( ak->is_obj_array_klass() ) {
2085 assert( ak->is_loaded(), "" );
2086 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2087 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2088 ciInstanceKlass* ik = base_k->as_instance_klass();
2089 // See if we can become precise: no subklasses and no interface
2090 if (!ik->is_interface() && !ik->has_subklass()) {
2091 //assert(!UseExactTypes, "this code should be useless with exact types");
2092 // Add a dependence; if any subclass added we need to recompile
2093 if (!ik->is_final()) {
2094 phase->C->dependencies()->assert_leaf_type(ik);
2095 }
2096 // Return precise array klass
2097 return TypeKlassPtr::make(ak);
2098 }
2099 }
2100 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2101 } else { // Found a type-array?
2102 //assert(!UseExactTypes, "this code should be useless with exact types");
2103 assert( ak->is_type_array_klass(), "" );
2104 return TypeKlassPtr::make(ak); // These are always precise
2105 }
2106 }
2107 }
2108
2109 // Check for loading klass from an array klass
2110 const TypeKlassPtr *tkls = tp->isa_klassptr();
2111 if (tkls != NULL && !StressReflectiveCode) {
2112 ciKlass* klass = tkls->klass();
2113 if( !klass->is_loaded() )
2114 return _type; // Bail out if not loaded
2115 if( klass->is_obj_array_klass() &&
2116 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2117 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2118 // // Always returning precise element type is incorrect,
2119 // // e.g., element type could be object and array may contain strings
2120 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2121
2122 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2123 // according to the element type's subclassing.
2124 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2125 }
2126 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2127 tkls->offset() == in_bytes(Klass::super_offset())) {
2128 ciKlass* sup = klass->as_instance_klass()->super();
2129 // The field is Klass::_super. Return its (constant) value.
2130 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2131 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2132 }
2133 }
2134
2135 // Bailout case
2136 return LoadNode::Value(phase);
2137 }
2138
2139 //------------------------------Identity---------------------------------------
2140 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2141 // Also feed through the klass in Allocate(...klass...)._klass.
2142 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
2143 return klass_identity_common(phase);
2144 }
2145
2146 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
2147 Node* x = LoadNode::Identity(phase);
2148 if (x != this) return x;
2149
2150 // Take apart the address into an oop and and offset.
2151 // Return 'this' if we cannot.
2152 Node* adr = in(MemNode::Address);
2153 intptr_t offset = 0;
2154 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2155 if (base == NULL) return this;
2156 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2157 if (toop == NULL) return this;
2158
2159 // We can fetch the klass directly through an AllocateNode.
2160 // This works even if the klass is not constant (clone or newArray).
2161 if (offset == oopDesc::klass_offset_in_bytes()) {
2162 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2163 if (allocated_klass != NULL) {
2164 return allocated_klass;
2165 }
2166 }
2167
2168 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2169 // See inline_native_Class_query for occurrences of these patterns.
2170 // Java Example: x.getClass().isAssignableFrom(y)
2171 //
2172 // This improves reflective code, often making the Class
2173 // mirror go completely dead. (Current exception: Class
2174 // mirrors may appear in debug info, but we could clean them out by
2175 // introducing a new debug info operator for Klass*.java_mirror).
2176 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2177 && offset == java_lang_Class::klass_offset_in_bytes()) {
2178 // We are loading a special hidden field from a Class mirror,
2179 // the field which points to its Klass or ArrayKlass metaobject.
2180 if (base->is_Load()) {
2181 Node* adr2 = base->in(MemNode::Address);
2182 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2183 if (tkls != NULL && !tkls->empty()
2184 && (tkls->klass()->is_instance_klass() ||
2185 tkls->klass()->is_array_klass())
2186 && adr2->is_AddP()
2187 ) {
2188 int mirror_field = in_bytes(Klass::java_mirror_offset());
2189 if (tkls->offset() == mirror_field) {
2190 return adr2->in(AddPNode::Base);
2191 }
2192 }
2193 }
2194 }
2195
2196 return this;
2197 }
2198
2199
2200 //------------------------------Value------------------------------------------
2201 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
2202 const Type *t = klass_value_common(phase);
2203 if (t == Type::TOP)
2204 return t;
2205
2206 return t->make_narrowklass();
2207 }
2208
2209 //------------------------------Identity---------------------------------------
2210 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2211 // Also feed through the klass in Allocate(...klass...)._klass.
2212 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2213 Node *x = klass_identity_common(phase);
2214
2215 const Type *t = phase->type( x );
2216 if( t == Type::TOP ) return x;
2217 if( t->isa_narrowklass()) return x;
2218 assert (!t->isa_narrowoop(), "no narrow oop here");
2219
2220 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2221 }
2222
2223 //------------------------------Value-----------------------------------------
2224 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2225 // Either input is TOP ==> the result is TOP
2226 const Type *t1 = phase->type( in(MemNode::Memory) );
2227 if( t1 == Type::TOP ) return Type::TOP;
2228 Node *adr = in(MemNode::Address);
2229 const Type *t2 = phase->type( adr );
2230 if( t2 == Type::TOP ) return Type::TOP;
2231 const TypePtr *tp = t2->is_ptr();
2232 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2233 const TypeAryPtr *tap = tp->isa_aryptr();
2234 if( !tap ) return _type;
2235 return tap->size();
2236 }
2237
2238 //-------------------------------Ideal---------------------------------------
2239 // Feed through the length in AllocateArray(...length...)._length.
2240 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2241 Node* p = MemNode::Ideal_common(phase, can_reshape);
2242 if (p) return (p == NodeSentinel) ? NULL : p;
2243
2244 // Take apart the address into an oop and and offset.
2245 // Return 'this' if we cannot.
2246 Node* adr = in(MemNode::Address);
2247 intptr_t offset = 0;
2248 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2249 if (base == NULL) return NULL;
2250 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2251 if (tary == NULL) return NULL;
2252
2253 // We can fetch the length directly through an AllocateArrayNode.
2254 // This works even if the length is not constant (clone or newArray).
2255 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2256 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2257 if (alloc != NULL) {
2258 Node* allocated_length = alloc->Ideal_length();
2259 Node* len = alloc->make_ideal_length(tary, phase);
2260 if (allocated_length != len) {
2261 // New CastII improves on this.
2262 return len;
2263 }
2264 }
2265 }
2266
2267 return NULL;
2268 }
2269
2270 //------------------------------Identity---------------------------------------
2271 // Feed through the length in AllocateArray(...length...)._length.
2272 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2273 Node* x = LoadINode::Identity(phase);
2274 if (x != this) return x;
2275
2276 // Take apart the address into an oop and and offset.
2277 // Return 'this' if we cannot.
2278 Node* adr = in(MemNode::Address);
2279 intptr_t offset = 0;
2280 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2281 if (base == NULL) return this;
2282 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2283 if (tary == NULL) return this;
2284
2285 // We can fetch the length directly through an AllocateArrayNode.
2286 // This works even if the length is not constant (clone or newArray).
2287 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2288 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2289 if (alloc != NULL) {
2290 Node* allocated_length = alloc->Ideal_length();
2291 // Do not allow make_ideal_length to allocate a CastII node.
2292 Node* len = alloc->make_ideal_length(tary, phase, false);
2293 if (allocated_length == len) {
2294 // Return allocated_length only if it would not be improved by a CastII.
2295 return allocated_length;
2296 }
2297 }
2298 }
2299
2300 return this;
2301
2302 }
2303
2304 //=============================================================================
2305 //---------------------------StoreNode::make-----------------------------------
2306 // Polymorphic factory method:
2307 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2308 assert((mo == unordered || mo == release), "unexpected");
2309 Compile* C = gvn.C;
2310 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2311 ctl != NULL, "raw memory operations should have control edge");
2312
2313 switch (bt) {
2314 case T_BOOLEAN:
2315 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2316 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2317 case T_CHAR:
2318 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2319 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2320 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2321 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2322 case T_METADATA:
2323 case T_ADDRESS:
2324 case T_OBJECT:
2325 #ifdef _LP64
2326 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2327 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2328 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2329 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2330 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2331 adr->bottom_type()->isa_rawptr())) {
2332 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2333 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2334 }
2335 #endif
2336 {
2337 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2338 }
2339 }
2340 ShouldNotReachHere();
2341 return (StoreNode*)NULL;
2342 }
2343
2344 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2345 bool require_atomic = true;
2346 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2347 }
2348
2349 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2350 bool require_atomic = true;
2351 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2352 }
2353
2354
2355 //--------------------------bottom_type----------------------------------------
2356 const Type *StoreNode::bottom_type() const {
2357 return Type::MEMORY;
2358 }
2359
2360 //------------------------------hash-------------------------------------------
2361 uint StoreNode::hash() const {
2362 // unroll addition of interesting fields
2363 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2364
2365 // Since they are not commoned, do not hash them:
2366 return NO_HASH;
2367 }
2368
2369 //------------------------------Ideal------------------------------------------
2370 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2371 // When a store immediately follows a relevant allocation/initialization,
2372 // try to capture it into the initialization, or hoist it above.
2373 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2374 Node* p = MemNode::Ideal_common(phase, can_reshape);
2375 if (p) return (p == NodeSentinel) ? NULL : p;
2376
2377 Node* mem = in(MemNode::Memory);
2378 Node* address = in(MemNode::Address);
2379 // Back-to-back stores to same address? Fold em up. Generally
2380 // unsafe if I have intervening uses... Also disallowed for StoreCM
2381 // since they must follow each StoreP operation. Redundant StoreCMs
2382 // are eliminated just before matching in final_graph_reshape.
2383 {
2384 Node* st = mem;
2385 // If Store 'st' has more than one use, we cannot fold 'st' away.
2386 // For example, 'st' might be the final state at a conditional
2387 // return. Or, 'st' might be used by some node which is live at
2388 // the same time 'st' is live, which might be unschedulable. So,
2389 // require exactly ONE user until such time as we clone 'mem' for
2390 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2391 // true).
2392 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2393 // Looking at a dead closed cycle of memory?
2394 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2395 assert(Opcode() == st->Opcode() ||
2396 st->Opcode() == Op_StoreVector ||
2397 Opcode() == Op_StoreVector ||
2398 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2399 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI), // expanded ClearArrayNode
2400 err_msg_res("no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]));
2401
2402 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2403 st->as_Store()->memory_size() <= this->memory_size()) {
2404 Node* use = st->raw_out(0);
2405 phase->igvn_rehash_node_delayed(use);
2406 if (can_reshape) {
2407 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2408 } else {
2409 // It's OK to do this in the parser, since DU info is always accurate,
2410 // and the parser always refers to nodes via SafePointNode maps.
2411 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2412 }
2413 return this;
2414 }
2415 st = st->in(MemNode::Memory);
2416 }
2417 }
2418
2419
2420 // Capture an unaliased, unconditional, simple store into an initializer.
2421 // Or, if it is independent of the allocation, hoist it above the allocation.
2422 if (ReduceFieldZeroing && /*can_reshape &&*/
2423 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2424 InitializeNode* init = mem->in(0)->as_Initialize();
2425 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2426 if (offset > 0) {
2427 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2428 // If the InitializeNode captured me, it made a raw copy of me,
2429 // and I need to disappear.
2430 if (moved != NULL) {
2431 // %%% hack to ensure that Ideal returns a new node:
2432 mem = MergeMemNode::make(mem);
2433 return mem; // fold me away
2434 }
2435 }
2436 }
2437
2438 return NULL; // No further progress
2439 }
2440
2441 //------------------------------Value-----------------------------------------
2442 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2443 // Either input is TOP ==> the result is TOP
2444 const Type *t1 = phase->type( in(MemNode::Memory) );
2445 if( t1 == Type::TOP ) return Type::TOP;
2446 const Type *t2 = phase->type( in(MemNode::Address) );
2447 if( t2 == Type::TOP ) return Type::TOP;
2448 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2449 if( t3 == Type::TOP ) return Type::TOP;
2450 return Type::MEMORY;
2451 }
2452
2453 //------------------------------Identity---------------------------------------
2454 // Remove redundant stores:
2455 // Store(m, p, Load(m, p)) changes to m.
2456 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2457 Node *StoreNode::Identity( PhaseTransform *phase ) {
2458 Node* mem = in(MemNode::Memory);
2459 Node* adr = in(MemNode::Address);
2460 Node* val = in(MemNode::ValueIn);
2461
2462 // Load then Store? Then the Store is useless
2463 if (val->is_Load() &&
2464 val->in(MemNode::Address)->eqv_uncast(adr) &&
2465 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2466 val->as_Load()->store_Opcode() == Opcode()) {
2467 return mem;
2468 }
2469
2470 // Two stores in a row of the same value?
2471 if (mem->is_Store() &&
2472 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2473 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2474 mem->Opcode() == Opcode()) {
2475 return mem;
2476 }
2477
2478 // Store of zero anywhere into a freshly-allocated object?
2479 // Then the store is useless.
2480 // (It must already have been captured by the InitializeNode.)
2481 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2482 // a newly allocated object is already all-zeroes everywhere
2483 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2484 return mem;
2485 }
2486
2487 // the store may also apply to zero-bits in an earlier object
2488 Node* prev_mem = find_previous_store(phase);
2489 // Steps (a), (b): Walk past independent stores to find an exact match.
2490 if (prev_mem != NULL) {
2491 Node* prev_val = can_see_stored_value(prev_mem, phase);
2492 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2493 // prev_val and val might differ by a cast; it would be good
2494 // to keep the more informative of the two.
2495 return mem;
2496 }
2497 }
2498 }
2499
2500 return this;
2501 }
2502
2503 //------------------------------match_edge-------------------------------------
2504 // Do we Match on this edge index or not? Match only memory & value
2505 uint StoreNode::match_edge(uint idx) const {
2506 return idx == MemNode::Address || idx == MemNode::ValueIn;
2507 }
2508
2509 //------------------------------cmp--------------------------------------------
2510 // Do not common stores up together. They generally have to be split
2511 // back up anyways, so do not bother.
2512 uint StoreNode::cmp( const Node &n ) const {
2513 return (&n == this); // Always fail except on self
2514 }
2515
2516 //------------------------------Ideal_masked_input-----------------------------
2517 // Check for a useless mask before a partial-word store
2518 // (StoreB ... (AndI valIn conIa) )
2519 // If (conIa & mask == mask) this simplifies to
2520 // (StoreB ... (valIn) )
2521 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2522 Node *val = in(MemNode::ValueIn);
2523 if( val->Opcode() == Op_AndI ) {
2524 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2525 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2526 set_req(MemNode::ValueIn, val->in(1));
2527 return this;
2528 }
2529 }
2530 return NULL;
2531 }
2532
2533
2534 //------------------------------Ideal_sign_extended_input----------------------
2535 // Check for useless sign-extension before a partial-word store
2536 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2537 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2538 // (StoreB ... (valIn) )
2539 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2540 Node *val = in(MemNode::ValueIn);
2541 if( val->Opcode() == Op_RShiftI ) {
2542 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2543 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2544 Node *shl = val->in(1);
2545 if( shl->Opcode() == Op_LShiftI ) {
2546 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2547 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2548 set_req(MemNode::ValueIn, shl->in(1));
2549 return this;
2550 }
2551 }
2552 }
2553 }
2554 return NULL;
2555 }
2556
2557 //------------------------------value_never_loaded-----------------------------------
2558 // Determine whether there are any possible loads of the value stored.
2559 // For simplicity, we actually check if there are any loads from the
2560 // address stored to, not just for loads of the value stored by this node.
2561 //
2562 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2563 Node *adr = in(Address);
2564 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2565 if (adr_oop == NULL)
2566 return false;
2567 if (!adr_oop->is_known_instance_field())
2568 return false; // if not a distinct instance, there may be aliases of the address
2569 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2570 Node *use = adr->fast_out(i);
2571 if (use->is_Load() || use->is_LoadStore()) {
2572 return false;
2573 }
2574 }
2575 return true;
2576 }
2577
2578 //=============================================================================
2579 //------------------------------Ideal------------------------------------------
2580 // If the store is from an AND mask that leaves the low bits untouched, then
2581 // we can skip the AND operation. If the store is from a sign-extension
2582 // (a left shift, then right shift) we can skip both.
2583 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2584 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2585 if( progress != NULL ) return progress;
2586
2587 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2588 if( progress != NULL ) return progress;
2589
2590 // Finally check the default case
2591 return StoreNode::Ideal(phase, can_reshape);
2592 }
2593
2594 //=============================================================================
2595 //------------------------------Ideal------------------------------------------
2596 // If the store is from an AND mask that leaves the low bits untouched, then
2597 // we can skip the AND operation
2598 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2599 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2600 if( progress != NULL ) return progress;
2601
2602 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2603 if( progress != NULL ) return progress;
2604
2605 // Finally check the default case
2606 return StoreNode::Ideal(phase, can_reshape);
2607 }
2608
2609 //=============================================================================
2610 //------------------------------Identity---------------------------------------
2611 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2612 // No need to card mark when storing a null ptr
2613 Node* my_store = in(MemNode::OopStore);
2614 if (my_store->is_Store()) {
2615 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2616 if( t1 == TypePtr::NULL_PTR ) {
2617 return in(MemNode::Memory);
2618 }
2619 }
2620 return this;
2621 }
2622
2623 //=============================================================================
2624 //------------------------------Ideal---------------------------------------
2625 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2626 Node* progress = StoreNode::Ideal(phase, can_reshape);
2627 if (progress != NULL) return progress;
2628
2629 Node* my_store = in(MemNode::OopStore);
2630 if (my_store->is_MergeMem()) {
2631 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2632 set_req(MemNode::OopStore, mem);
2633 return this;
2634 }
2635
2636 return NULL;
2637 }
2638
2639 //------------------------------Value-----------------------------------------
2640 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2641 // Either input is TOP ==> the result is TOP
2642 const Type *t = phase->type( in(MemNode::Memory) );
2643 if( t == Type::TOP ) return Type::TOP;
2644 t = phase->type( in(MemNode::Address) );
2645 if( t == Type::TOP ) return Type::TOP;
2646 t = phase->type( in(MemNode::ValueIn) );
2647 if( t == Type::TOP ) return Type::TOP;
2648 // If extra input is TOP ==> the result is TOP
2649 t = phase->type( in(MemNode::OopStore) );
2650 if( t == Type::TOP ) return Type::TOP;
2651
2652 return StoreNode::Value( phase );
2653 }
2654
2655
2656 //=============================================================================
2657 //----------------------------------SCMemProjNode------------------------------
2658 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2659 {
2660 return bottom_type();
2661 }
2662
2663 //=============================================================================
2664 //----------------------------------LoadStoreNode------------------------------
2665 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2666 : Node(required),
2667 _type(rt),
2668 _adr_type(at)
2669 {
2670 init_req(MemNode::Control, c );
2671 init_req(MemNode::Memory , mem);
2672 init_req(MemNode::Address, adr);
2673 init_req(MemNode::ValueIn, val);
2674 init_class_id(Class_LoadStore);
2675 }
2676
2677 uint LoadStoreNode::ideal_reg() const {
2678 return _type->ideal_reg();
2679 }
2680
2681 bool LoadStoreNode::result_not_used() const {
2682 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2683 Node *x = fast_out(i);
2684 if (x->Opcode() == Op_SCMemProj) continue;
2685 return false;
2686 }
2687 return true;
2688 }
2689
2690 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2691
2692 //=============================================================================
2693 //----------------------------------LoadStoreConditionalNode--------------------
2694 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2695 init_req(ExpectedIn, ex );
2696 }
2697
2698 //=============================================================================
2699 //-------------------------------adr_type--------------------------------------
2700 const TypePtr* ClearArrayNode::adr_type() const {
2701 Node *adr = in(3);
2702 if (adr == NULL) return NULL; // node is dead
2703 return MemNode::calculate_adr_type(adr->bottom_type());
2704 }
2705
2706 //------------------------------match_edge-------------------------------------
2707 // Do we Match on this edge index or not? Do not match memory
2708 uint ClearArrayNode::match_edge(uint idx) const {
2709 return idx > 1;
2710 }
2711
2712 //------------------------------Identity---------------------------------------
2713 // Clearing a zero length array does nothing
2714 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2715 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2716 }
2717
2718 //------------------------------Idealize---------------------------------------
2719 // Clearing a short array is faster with stores
2720 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2721 const int unit = BytesPerLong;
2722 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2723 if (!t) return NULL;
2724 if (!t->is_con()) return NULL;
2725 intptr_t raw_count = t->get_con();
2726 intptr_t size = raw_count;
2727 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2728 // Clearing nothing uses the Identity call.
2729 // Negative clears are possible on dead ClearArrays
2730 // (see jck test stmt114.stmt11402.val).
2731 if (size <= 0 || size % unit != 0) return NULL;
2732 intptr_t count = size / unit;
2733 // Length too long; use fast hardware clear
2734 if (size > Matcher::init_array_short_size) return NULL;
2735 Node *mem = in(1);
2736 if( phase->type(mem)==Type::TOP ) return NULL;
2737 Node *adr = in(3);
2738 const Type* at = phase->type(adr);
2739 if( at==Type::TOP ) return NULL;
2740 const TypePtr* atp = at->isa_ptr();
2741 // adjust atp to be the correct array element address type
2742 if (atp == NULL) atp = TypePtr::BOTTOM;
2743 else atp = atp->add_offset(Type::OffsetBot);
2744 // Get base for derived pointer purposes
2745 if( adr->Opcode() != Op_AddP ) Unimplemented();
2746 Node *base = adr->in(1);
2747
2748 Node *zero = phase->makecon(TypeLong::ZERO);
2749 Node *off = phase->MakeConX(BytesPerLong);
2750 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2751 count--;
2752 while( count-- ) {
2753 mem = phase->transform(mem);
2754 adr = phase->transform(new AddPNode(base,adr,off));
2755 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2756 }
2757 return mem;
2758 }
2759
2760 //----------------------------step_through----------------------------------
2761 // Return allocation input memory edge if it is different instance
2762 // or itself if it is the one we are looking for.
2763 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2764 Node* n = *np;
2765 assert(n->is_ClearArray(), "sanity");
2766 intptr_t offset;
2767 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2768 // This method is called only before Allocate nodes are expanded
2769 // during macro nodes expansion. Before that ClearArray nodes are
2770 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2771 // Allocate nodes are expanded) which follows allocations.
2772 assert(alloc != NULL, "should have allocation");
2773 if (alloc->_idx == instance_id) {
2774 // Can not bypass initialization of the instance we are looking for.
2775 return false;
2776 }
2777 // Otherwise skip it.
2778 InitializeNode* init = alloc->initialization();
2779 if (init != NULL)
2780 *np = init->in(TypeFunc::Memory);
2781 else
2782 *np = alloc->in(TypeFunc::Memory);
2783 return true;
2784 }
2785
2786 //----------------------------clear_memory-------------------------------------
2787 // Generate code to initialize object storage to zero.
2788 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2789 intptr_t start_offset,
2790 Node* end_offset,
2791 PhaseGVN* phase) {
2792 intptr_t offset = start_offset;
2793
2794 int unit = BytesPerLong;
2795 if ((offset % unit) != 0) {
2796 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
2797 adr = phase->transform(adr);
2798 const TypePtr* atp = TypeRawPtr::BOTTOM;
2799 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2800 mem = phase->transform(mem);
2801 offset += BytesPerInt;
2802 }
2803 assert((offset % unit) == 0, "");
2804
2805 // Initialize the remaining stuff, if any, with a ClearArray.
2806 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2807 }
2808
2809 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2810 Node* start_offset,
2811 Node* end_offset,
2812 PhaseGVN* phase) {
2813 if (start_offset == end_offset) {
2814 // nothing to do
2815 return mem;
2816 }
2817
2818 int unit = BytesPerLong;
2819 Node* zbase = start_offset;
2820 Node* zend = end_offset;
2821
2822 // Scale to the unit required by the CPU:
2823 if (!Matcher::init_array_count_is_in_bytes) {
2824 Node* shift = phase->intcon(exact_log2(unit));
2825 zbase = phase->transform(new URShiftXNode(zbase, shift) );
2826 zend = phase->transform(new URShiftXNode(zend, shift) );
2827 }
2828
2829 // Bulk clear double-words
2830 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
2831 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
2832 mem = new ClearArrayNode(ctl, mem, zsize, adr);
2833 return phase->transform(mem);
2834 }
2835
2836 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2837 intptr_t start_offset,
2838 intptr_t end_offset,
2839 PhaseGVN* phase) {
2840 if (start_offset == end_offset) {
2841 // nothing to do
2842 return mem;
2843 }
2844
2845 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2846 intptr_t done_offset = end_offset;
2847 if ((done_offset % BytesPerLong) != 0) {
2848 done_offset -= BytesPerInt;
2849 }
2850 if (done_offset > start_offset) {
2851 mem = clear_memory(ctl, mem, dest,
2852 start_offset, phase->MakeConX(done_offset), phase);
2853 }
2854 if (done_offset < end_offset) { // emit the final 32-bit store
2855 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
2856 adr = phase->transform(adr);
2857 const TypePtr* atp = TypeRawPtr::BOTTOM;
2858 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2859 mem = phase->transform(mem);
2860 done_offset += BytesPerInt;
2861 }
2862 assert(done_offset == end_offset, "");
2863 return mem;
2864 }
2865
2866 //=============================================================================
2867 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2868 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2869 _adr_type(C->get_adr_type(alias_idx))
2870 {
2871 init_class_id(Class_MemBar);
2872 Node* top = C->top();
2873 init_req(TypeFunc::I_O,top);
2874 init_req(TypeFunc::FramePtr,top);
2875 init_req(TypeFunc::ReturnAdr,top);
2876 if (precedent != NULL)
2877 init_req(TypeFunc::Parms, precedent);
2878 }
2879
2880 //------------------------------cmp--------------------------------------------
2881 uint MemBarNode::hash() const { return NO_HASH; }
2882 uint MemBarNode::cmp( const Node &n ) const {
2883 return (&n == this); // Always fail except on self
2884 }
2885
2886 //------------------------------make-------------------------------------------
2887 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2888 switch (opcode) {
2889 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
2890 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
2891 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
2892 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
2893 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
2894 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
2895 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
2896 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
2897 case Op_Initialize: return new InitializeNode(C, atp, pn);
2898 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
2899 default: ShouldNotReachHere(); return NULL;
2900 }
2901 }
2902
2903 //------------------------------Ideal------------------------------------------
2904 // Return a node which is more "ideal" than the current node. Strip out
2905 // control copies
2906 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2907 if (remove_dead_region(phase, can_reshape)) return this;
2908 // Don't bother trying to transform a dead node
2909 if (in(0) && in(0)->is_top()) {
2910 return NULL;
2911 }
2912
2913 bool progress = false;
2914 // Eliminate volatile MemBars for scalar replaced objects.
2915 if (can_reshape && req() == (Precedent+1)) {
2916 bool eliminate = false;
2917 int opc = Opcode();
2918 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
2919 // Volatile field loads and stores.
2920 Node* my_mem = in(MemBarNode::Precedent);
2921 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
2922 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
2923 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
2924 // replace this Precedent (decodeN) with the Load instead.
2925 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
2926 Node* load_node = my_mem->in(1);
2927 set_req(MemBarNode::Precedent, load_node);
2928 phase->is_IterGVN()->_worklist.push(my_mem);
2929 my_mem = load_node;
2930 } else {
2931 assert(my_mem->unique_out() == this, "sanity");
2932 del_req(Precedent);
2933 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
2934 my_mem = NULL;
2935 }
2936 progress = true;
2937 }
2938 if (my_mem != NULL && my_mem->is_Mem()) {
2939 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2940 // Check for scalar replaced object reference.
2941 if( t_oop != NULL && t_oop->is_known_instance_field() &&
2942 t_oop->offset() != Type::OffsetBot &&
2943 t_oop->offset() != Type::OffsetTop) {
2944 eliminate = true;
2945 }
2946 }
2947 } else if (opc == Op_MemBarRelease) {
2948 // Final field stores.
2949 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
2950 if ((alloc != NULL) && alloc->is_Allocate() &&
2951 alloc->as_Allocate()->_is_non_escaping) {
2952 // The allocated object does not escape.
2953 eliminate = true;
2954 }
2955 }
2956 if (eliminate) {
2957 // Replace MemBar projections by its inputs.
2958 PhaseIterGVN* igvn = phase->is_IterGVN();
2959 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2960 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2961 // Must return either the original node (now dead) or a new node
2962 // (Do not return a top here, since that would break the uniqueness of top.)
2963 return new ConINode(TypeInt::ZERO);
2964 }
2965 }
2966 return progress ? this : NULL;
2967 }
2968
2969 //------------------------------Value------------------------------------------
2970 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2971 if( !in(0) ) return Type::TOP;
2972 if( phase->type(in(0)) == Type::TOP )
2973 return Type::TOP;
2974 return TypeTuple::MEMBAR;
2975 }
2976
2977 //------------------------------match------------------------------------------
2978 // Construct projections for memory.
2979 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2980 switch (proj->_con) {
2981 case TypeFunc::Control:
2982 case TypeFunc::Memory:
2983 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2984 }
2985 ShouldNotReachHere();
2986 return NULL;
2987 }
2988
2989 //===========================InitializeNode====================================
2990 // SUMMARY:
2991 // This node acts as a memory barrier on raw memory, after some raw stores.
2992 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2993 // The Initialize can 'capture' suitably constrained stores as raw inits.
2994 // It can coalesce related raw stores into larger units (called 'tiles').
2995 // It can avoid zeroing new storage for memory units which have raw inits.
2996 // At macro-expansion, it is marked 'complete', and does not optimize further.
2997 //
2998 // EXAMPLE:
2999 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3000 // ctl = incoming control; mem* = incoming memory
3001 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3002 // First allocate uninitialized memory and fill in the header:
3003 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3004 // ctl := alloc.Control; mem* := alloc.Memory*
3005 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3006 // Then initialize to zero the non-header parts of the raw memory block:
3007 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3008 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3009 // After the initialize node executes, the object is ready for service:
3010 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3011 // Suppose its body is immediately initialized as {1,2}:
3012 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3013 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3014 // mem.SLICE(#short[*]) := store2
3015 //
3016 // DETAILS:
3017 // An InitializeNode collects and isolates object initialization after
3018 // an AllocateNode and before the next possible safepoint. As a
3019 // memory barrier (MemBarNode), it keeps critical stores from drifting
3020 // down past any safepoint or any publication of the allocation.
3021 // Before this barrier, a newly-allocated object may have uninitialized bits.
3022 // After this barrier, it may be treated as a real oop, and GC is allowed.
3023 //
3024 // The semantics of the InitializeNode include an implicit zeroing of
3025 // the new object from object header to the end of the object.
3026 // (The object header and end are determined by the AllocateNode.)
3027 //
3028 // Certain stores may be added as direct inputs to the InitializeNode.
3029 // These stores must update raw memory, and they must be to addresses
3030 // derived from the raw address produced by AllocateNode, and with
3031 // a constant offset. They must be ordered by increasing offset.
3032 // The first one is at in(RawStores), the last at in(req()-1).
3033 // Unlike most memory operations, they are not linked in a chain,
3034 // but are displayed in parallel as users of the rawmem output of
3035 // the allocation.
3036 //
3037 // (See comments in InitializeNode::capture_store, which continue
3038 // the example given above.)
3039 //
3040 // When the associated Allocate is macro-expanded, the InitializeNode
3041 // may be rewritten to optimize collected stores. A ClearArrayNode
3042 // may also be created at that point to represent any required zeroing.
3043 // The InitializeNode is then marked 'complete', prohibiting further
3044 // capturing of nearby memory operations.
3045 //
3046 // During macro-expansion, all captured initializations which store
3047 // constant values of 32 bits or smaller are coalesced (if advantageous)
3048 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3049 // initialized in fewer memory operations. Memory words which are
3050 // covered by neither tiles nor non-constant stores are pre-zeroed
3051 // by explicit stores of zero. (The code shape happens to do all
3052 // zeroing first, then all other stores, with both sequences occurring
3053 // in order of ascending offsets.)
3054 //
3055 // Alternatively, code may be inserted between an AllocateNode and its
3056 // InitializeNode, to perform arbitrary initialization of the new object.
3057 // E.g., the object copying intrinsics insert complex data transfers here.
3058 // The initialization must then be marked as 'complete' disable the
3059 // built-in zeroing semantics and the collection of initializing stores.
3060 //
3061 // While an InitializeNode is incomplete, reads from the memory state
3062 // produced by it are optimizable if they match the control edge and
3063 // new oop address associated with the allocation/initialization.
3064 // They return a stored value (if the offset matches) or else zero.
3065 // A write to the memory state, if it matches control and address,
3066 // and if it is to a constant offset, may be 'captured' by the
3067 // InitializeNode. It is cloned as a raw memory operation and rewired
3068 // inside the initialization, to the raw oop produced by the allocation.
3069 // Operations on addresses which are provably distinct (e.g., to
3070 // other AllocateNodes) are allowed to bypass the initialization.
3071 //
3072 // The effect of all this is to consolidate object initialization
3073 // (both arrays and non-arrays, both piecewise and bulk) into a
3074 // single location, where it can be optimized as a unit.
3075 //
3076 // Only stores with an offset less than TrackedInitializationLimit words
3077 // will be considered for capture by an InitializeNode. This puts a
3078 // reasonable limit on the complexity of optimized initializations.
3079
3080 //---------------------------InitializeNode------------------------------------
3081 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3082 : _is_complete(Incomplete), _does_not_escape(false),
3083 MemBarNode(C, adr_type, rawoop)
3084 {
3085 init_class_id(Class_Initialize);
3086
3087 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3088 assert(in(RawAddress) == rawoop, "proper init");
3089 // Note: allocation() can be NULL, for secondary initialization barriers
3090 }
3091
3092 // Since this node is not matched, it will be processed by the
3093 // register allocator. Declare that there are no constraints
3094 // on the allocation of the RawAddress edge.
3095 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3096 // This edge should be set to top, by the set_complete. But be conservative.
3097 if (idx == InitializeNode::RawAddress)
3098 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3099 return RegMask::Empty;
3100 }
3101
3102 Node* InitializeNode::memory(uint alias_idx) {
3103 Node* mem = in(Memory);
3104 if (mem->is_MergeMem()) {
3105 return mem->as_MergeMem()->memory_at(alias_idx);
3106 } else {
3107 // incoming raw memory is not split
3108 return mem;
3109 }
3110 }
3111
3112 bool InitializeNode::is_non_zero() {
3113 if (is_complete()) return false;
3114 remove_extra_zeroes();
3115 return (req() > RawStores);
3116 }
3117
3118 void InitializeNode::set_complete(PhaseGVN* phase) {
3119 assert(!is_complete(), "caller responsibility");
3120 _is_complete = Complete;
3121
3122 // After this node is complete, it contains a bunch of
3123 // raw-memory initializations. There is no need for
3124 // it to have anything to do with non-raw memory effects.
3125 // Therefore, tell all non-raw users to re-optimize themselves,
3126 // after skipping the memory effects of this initialization.
3127 PhaseIterGVN* igvn = phase->is_IterGVN();
3128 if (igvn) igvn->add_users_to_worklist(this);
3129 }
3130
3131 // convenience function
3132 // return false if the init contains any stores already
3133 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3134 InitializeNode* init = initialization();
3135 if (init == NULL || init->is_complete()) return false;
3136 init->remove_extra_zeroes();
3137 // for now, if this allocation has already collected any inits, bail:
3138 if (init->is_non_zero()) return false;
3139 init->set_complete(phase);
3140 return true;
3141 }
3142
3143 void InitializeNode::remove_extra_zeroes() {
3144 if (req() == RawStores) return;
3145 Node* zmem = zero_memory();
3146 uint fill = RawStores;
3147 for (uint i = fill; i < req(); i++) {
3148 Node* n = in(i);
3149 if (n->is_top() || n == zmem) continue; // skip
3150 if (fill < i) set_req(fill, n); // compact
3151 ++fill;
3152 }
3153 // delete any empty spaces created:
3154 while (fill < req()) {
3155 del_req(fill);
3156 }
3157 }
3158
3159 // Helper for remembering which stores go with which offsets.
3160 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3161 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3162 intptr_t offset = -1;
3163 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3164 phase, offset);
3165 if (base == NULL) return -1; // something is dead,
3166 if (offset < 0) return -1; // dead, dead
3167 return offset;
3168 }
3169
3170 // Helper for proving that an initialization expression is
3171 // "simple enough" to be folded into an object initialization.
3172 // Attempts to prove that a store's initial value 'n' can be captured
3173 // within the initialization without creating a vicious cycle, such as:
3174 // { Foo p = new Foo(); p.next = p; }
3175 // True for constants and parameters and small combinations thereof.
3176 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3177 if (n == NULL) return true; // (can this really happen?)
3178 if (n->is_Proj()) n = n->in(0);
3179 if (n == this) return false; // found a cycle
3180 if (n->is_Con()) return true;
3181 if (n->is_Start()) return true; // params, etc., are OK
3182 if (n->is_Root()) return true; // even better
3183
3184 Node* ctl = n->in(0);
3185 if (ctl != NULL && !ctl->is_top()) {
3186 if (ctl->is_Proj()) ctl = ctl->in(0);
3187 if (ctl == this) return false;
3188
3189 // If we already know that the enclosing memory op is pinned right after
3190 // the init, then any control flow that the store has picked up
3191 // must have preceded the init, or else be equal to the init.
3192 // Even after loop optimizations (which might change control edges)
3193 // a store is never pinned *before* the availability of its inputs.
3194 if (!MemNode::all_controls_dominate(n, this))
3195 return false; // failed to prove a good control
3196 }
3197
3198 // Check data edges for possible dependencies on 'this'.
3199 if ((count += 1) > 20) return false; // complexity limit
3200 for (uint i = 1; i < n->req(); i++) {
3201 Node* m = n->in(i);
3202 if (m == NULL || m == n || m->is_top()) continue;
3203 uint first_i = n->find_edge(m);
3204 if (i != first_i) continue; // process duplicate edge just once
3205 if (!detect_init_independence(m, count)) {
3206 return false;
3207 }
3208 }
3209
3210 return true;
3211 }
3212
3213 // Here are all the checks a Store must pass before it can be moved into
3214 // an initialization. Returns zero if a check fails.
3215 // On success, returns the (constant) offset to which the store applies,
3216 // within the initialized memory.
3217 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3218 const int FAIL = 0;
3219 if (st->req() != MemNode::ValueIn + 1)
3220 return FAIL; // an inscrutable StoreNode (card mark?)
3221 Node* ctl = st->in(MemNode::Control);
3222 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3223 return FAIL; // must be unconditional after the initialization
3224 Node* mem = st->in(MemNode::Memory);
3225 if (!(mem->is_Proj() && mem->in(0) == this))
3226 return FAIL; // must not be preceded by other stores
3227 Node* adr = st->in(MemNode::Address);
3228 intptr_t offset;
3229 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3230 if (alloc == NULL)
3231 return FAIL; // inscrutable address
3232 if (alloc != allocation())
3233 return FAIL; // wrong allocation! (store needs to float up)
3234 Node* val = st->in(MemNode::ValueIn);
3235 int complexity_count = 0;
3236 if (!detect_init_independence(val, complexity_count))
3237 return FAIL; // stored value must be 'simple enough'
3238
3239 // The Store can be captured only if nothing after the allocation
3240 // and before the Store is using the memory location that the store
3241 // overwrites.
3242 bool failed = false;
3243 // If is_complete_with_arraycopy() is true the shape of the graph is
3244 // well defined and is safe so no need for extra checks.
3245 if (!is_complete_with_arraycopy()) {
3246 // We are going to look at each use of the memory state following
3247 // the allocation to make sure nothing reads the memory that the
3248 // Store writes.
3249 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3250 int alias_idx = phase->C->get_alias_index(t_adr);
3251 ResourceMark rm;
3252 Unique_Node_List mems;
3253 mems.push(mem);
3254 Node* unique_merge = NULL;
3255 for (uint next = 0; next < mems.size(); ++next) {
3256 Node *m = mems.at(next);
3257 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3258 Node *n = m->fast_out(j);
3259 if (n->outcnt() == 0) {
3260 continue;
3261 }
3262 if (n == st) {
3263 continue;
3264 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3265 // If the control of this use is different from the control
3266 // of the Store which is right after the InitializeNode then
3267 // this node cannot be between the InitializeNode and the
3268 // Store.
3269 continue;
3270 } else if (n->is_MergeMem()) {
3271 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3272 // We can hit a MergeMemNode (that will likely go away
3273 // later) that is a direct use of the memory state
3274 // following the InitializeNode on the same slice as the
3275 // store node that we'd like to capture. We need to check
3276 // the uses of the MergeMemNode.
3277 mems.push(n);
3278 }
3279 } else if (n->is_Mem()) {
3280 Node* other_adr = n->in(MemNode::Address);
3281 if (other_adr == adr) {
3282 failed = true;
3283 break;
3284 } else {
3285 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3286 if (other_t_adr != NULL) {
3287 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3288 if (other_alias_idx == alias_idx) {
3289 // A load from the same memory slice as the store right
3290 // after the InitializeNode. We check the control of the
3291 // object/array that is loaded from. If it's the same as
3292 // the store control then we cannot capture the store.
3293 assert(!n->is_Store(), "2 stores to same slice on same control?");
3294 Node* base = other_adr;
3295 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
3296 base = base->in(AddPNode::Base);
3297 if (base != NULL) {
3298 base = base->uncast();
3299 if (base->is_Proj() && base->in(0) == alloc) {
3300 failed = true;
3301 break;
3302 }
3303 }
3304 }
3305 }
3306 }
3307 } else {
3308 failed = true;
3309 break;
3310 }
3311 }
3312 }
3313 }
3314 if (failed) {
3315 if (!can_reshape) {
3316 // We decided we couldn't capture the store during parsing. We
3317 // should try again during the next IGVN once the graph is
3318 // cleaner.
3319 phase->C->record_for_igvn(st);
3320 }
3321 return FAIL;
3322 }
3323
3324 return offset; // success
3325 }
3326
3327 // Find the captured store in(i) which corresponds to the range
3328 // [start..start+size) in the initialized object.
3329 // If there is one, return its index i. If there isn't, return the
3330 // negative of the index where it should be inserted.
3331 // Return 0 if the queried range overlaps an initialization boundary
3332 // or if dead code is encountered.
3333 // If size_in_bytes is zero, do not bother with overlap checks.
3334 int InitializeNode::captured_store_insertion_point(intptr_t start,
3335 int size_in_bytes,
3336 PhaseTransform* phase) {
3337 const int FAIL = 0, MAX_STORE = BytesPerLong;
3338
3339 if (is_complete())
3340 return FAIL; // arraycopy got here first; punt
3341
3342 assert(allocation() != NULL, "must be present");
3343
3344 // no negatives, no header fields:
3345 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3346
3347 // after a certain size, we bail out on tracking all the stores:
3348 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3349 if (start >= ti_limit) return FAIL;
3350
3351 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3352 if (i >= limit) return -(int)i; // not found; here is where to put it
3353
3354 Node* st = in(i);
3355 intptr_t st_off = get_store_offset(st, phase);
3356 if (st_off < 0) {
3357 if (st != zero_memory()) {
3358 return FAIL; // bail out if there is dead garbage
3359 }
3360 } else if (st_off > start) {
3361 // ...we are done, since stores are ordered
3362 if (st_off < start + size_in_bytes) {
3363 return FAIL; // the next store overlaps
3364 }
3365 return -(int)i; // not found; here is where to put it
3366 } else if (st_off < start) {
3367 if (size_in_bytes != 0 &&
3368 start < st_off + MAX_STORE &&
3369 start < st_off + st->as_Store()->memory_size()) {
3370 return FAIL; // the previous store overlaps
3371 }
3372 } else {
3373 if (size_in_bytes != 0 &&
3374 st->as_Store()->memory_size() != size_in_bytes) {
3375 return FAIL; // mismatched store size
3376 }
3377 return i;
3378 }
3379
3380 ++i;
3381 }
3382 }
3383
3384 // Look for a captured store which initializes at the offset 'start'
3385 // with the given size. If there is no such store, and no other
3386 // initialization interferes, then return zero_memory (the memory
3387 // projection of the AllocateNode).
3388 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3389 PhaseTransform* phase) {
3390 assert(stores_are_sane(phase), "");
3391 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3392 if (i == 0) {
3393 return NULL; // something is dead
3394 } else if (i < 0) {
3395 return zero_memory(); // just primordial zero bits here
3396 } else {
3397 Node* st = in(i); // here is the store at this position
3398 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3399 return st;
3400 }
3401 }
3402
3403 // Create, as a raw pointer, an address within my new object at 'offset'.
3404 Node* InitializeNode::make_raw_address(intptr_t offset,
3405 PhaseTransform* phase) {
3406 Node* addr = in(RawAddress);
3407 if (offset != 0) {
3408 Compile* C = phase->C;
3409 addr = phase->transform( new AddPNode(C->top(), addr,
3410 phase->MakeConX(offset)) );
3411 }
3412 return addr;
3413 }
3414
3415 // Clone the given store, converting it into a raw store
3416 // initializing a field or element of my new object.
3417 // Caller is responsible for retiring the original store,
3418 // with subsume_node or the like.
3419 //
3420 // From the example above InitializeNode::InitializeNode,
3421 // here are the old stores to be captured:
3422 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3423 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3424 //
3425 // Here is the changed code; note the extra edges on init:
3426 // alloc = (Allocate ...)
3427 // rawoop = alloc.RawAddress
3428 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3429 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3430 // init = (Initialize alloc.Control alloc.Memory rawoop
3431 // rawstore1 rawstore2)
3432 //
3433 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3434 PhaseTransform* phase, bool can_reshape) {
3435 assert(stores_are_sane(phase), "");
3436
3437 if (start < 0) return NULL;
3438 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3439
3440 Compile* C = phase->C;
3441 int size_in_bytes = st->memory_size();
3442 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3443 if (i == 0) return NULL; // bail out
3444 Node* prev_mem = NULL; // raw memory for the captured store
3445 if (i > 0) {
3446 prev_mem = in(i); // there is a pre-existing store under this one
3447 set_req(i, C->top()); // temporarily disconnect it
3448 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3449 } else {
3450 i = -i; // no pre-existing store
3451 prev_mem = zero_memory(); // a slice of the newly allocated object
3452 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3453 set_req(--i, C->top()); // reuse this edge; it has been folded away
3454 else
3455 ins_req(i, C->top()); // build a new edge
3456 }
3457 Node* new_st = st->clone();
3458 new_st->set_req(MemNode::Control, in(Control));
3459 new_st->set_req(MemNode::Memory, prev_mem);
3460 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3461 new_st = phase->transform(new_st);
3462
3463 // At this point, new_st might have swallowed a pre-existing store
3464 // at the same offset, or perhaps new_st might have disappeared,
3465 // if it redundantly stored the same value (or zero to fresh memory).
3466
3467 // In any case, wire it in:
3468 phase->igvn_rehash_node_delayed(this);
3469 set_req(i, new_st);
3470
3471 // The caller may now kill the old guy.
3472 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3473 assert(check_st == new_st || check_st == NULL, "must be findable");
3474 assert(!is_complete(), "");
3475 return new_st;
3476 }
3477
3478 static bool store_constant(jlong* tiles, int num_tiles,
3479 intptr_t st_off, int st_size,
3480 jlong con) {
3481 if ((st_off & (st_size-1)) != 0)
3482 return false; // strange store offset (assume size==2**N)
3483 address addr = (address)tiles + st_off;
3484 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3485 switch (st_size) {
3486 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3487 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3488 case sizeof(jint): *(jint*) addr = (jint) con; break;
3489 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3490 default: return false; // strange store size (detect size!=2**N here)
3491 }
3492 return true; // return success to caller
3493 }
3494
3495 // Coalesce subword constants into int constants and possibly
3496 // into long constants. The goal, if the CPU permits,
3497 // is to initialize the object with a small number of 64-bit tiles.
3498 // Also, convert floating-point constants to bit patterns.
3499 // Non-constants are not relevant to this pass.
3500 //
3501 // In terms of the running example on InitializeNode::InitializeNode
3502 // and InitializeNode::capture_store, here is the transformation
3503 // of rawstore1 and rawstore2 into rawstore12:
3504 // alloc = (Allocate ...)
3505 // rawoop = alloc.RawAddress
3506 // tile12 = 0x00010002
3507 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3508 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3509 //
3510 void
3511 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3512 Node* size_in_bytes,
3513 PhaseGVN* phase) {
3514 Compile* C = phase->C;
3515
3516 assert(stores_are_sane(phase), "");
3517 // Note: After this pass, they are not completely sane,
3518 // since there may be some overlaps.
3519
3520 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3521
3522 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3523 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3524 size_limit = MIN2(size_limit, ti_limit);
3525 size_limit = align_size_up(size_limit, BytesPerLong);
3526 int num_tiles = size_limit / BytesPerLong;
3527
3528 // allocate space for the tile map:
3529 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3530 jlong tiles_buf[small_len];
3531 Node* nodes_buf[small_len];
3532 jlong inits_buf[small_len];
3533 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3534 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3535 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3536 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3537 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3538 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3539 // tiles: exact bitwise model of all primitive constants
3540 // nodes: last constant-storing node subsumed into the tiles model
3541 // inits: which bytes (in each tile) are touched by any initializations
3542
3543 //// Pass A: Fill in the tile model with any relevant stores.
3544
3545 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3546 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3547 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3548 Node* zmem = zero_memory(); // initially zero memory state
3549 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3550 Node* st = in(i);
3551 intptr_t st_off = get_store_offset(st, phase);
3552
3553 // Figure out the store's offset and constant value:
3554 if (st_off < header_size) continue; //skip (ignore header)
3555 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3556 int st_size = st->as_Store()->memory_size();
3557 if (st_off + st_size > size_limit) break;
3558
3559 // Record which bytes are touched, whether by constant or not.
3560 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3561 continue; // skip (strange store size)
3562
3563 const Type* val = phase->type(st->in(MemNode::ValueIn));
3564 if (!val->singleton()) continue; //skip (non-con store)
3565 BasicType type = val->basic_type();
3566
3567 jlong con = 0;
3568 switch (type) {
3569 case T_INT: con = val->is_int()->get_con(); break;
3570 case T_LONG: con = val->is_long()->get_con(); break;
3571 case T_FLOAT: con = jint_cast(val->getf()); break;
3572 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3573 default: continue; //skip (odd store type)
3574 }
3575
3576 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3577 st->Opcode() == Op_StoreL) {
3578 continue; // This StoreL is already optimal.
3579 }
3580
3581 // Store down the constant.
3582 store_constant(tiles, num_tiles, st_off, st_size, con);
3583
3584 intptr_t j = st_off >> LogBytesPerLong;
3585
3586 if (type == T_INT && st_size == BytesPerInt
3587 && (st_off & BytesPerInt) == BytesPerInt) {
3588 jlong lcon = tiles[j];
3589 if (!Matcher::isSimpleConstant64(lcon) &&
3590 st->Opcode() == Op_StoreI) {
3591 // This StoreI is already optimal by itself.
3592 jint* intcon = (jint*) &tiles[j];
3593 intcon[1] = 0; // undo the store_constant()
3594
3595 // If the previous store is also optimal by itself, back up and
3596 // undo the action of the previous loop iteration... if we can.
3597 // But if we can't, just let the previous half take care of itself.
3598 st = nodes[j];
3599 st_off -= BytesPerInt;
3600 con = intcon[0];
3601 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3602 assert(st_off >= header_size, "still ignoring header");
3603 assert(get_store_offset(st, phase) == st_off, "must be");
3604 assert(in(i-1) == zmem, "must be");
3605 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3606 assert(con == tcon->is_int()->get_con(), "must be");
3607 // Undo the effects of the previous loop trip, which swallowed st:
3608 intcon[0] = 0; // undo store_constant()
3609 set_req(i-1, st); // undo set_req(i, zmem)
3610 nodes[j] = NULL; // undo nodes[j] = st
3611 --old_subword; // undo ++old_subword
3612 }
3613 continue; // This StoreI is already optimal.
3614 }
3615 }
3616
3617 // This store is not needed.
3618 set_req(i, zmem);
3619 nodes[j] = st; // record for the moment
3620 if (st_size < BytesPerLong) // something has changed
3621 ++old_subword; // includes int/float, but who's counting...
3622 else ++old_long;
3623 }
3624
3625 if ((old_subword + old_long) == 0)
3626 return; // nothing more to do
3627
3628 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3629 // Be sure to insert them before overlapping non-constant stores.
3630 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3631 for (int j = 0; j < num_tiles; j++) {
3632 jlong con = tiles[j];
3633 jlong init = inits[j];
3634 if (con == 0) continue;
3635 jint con0, con1; // split the constant, address-wise
3636 jint init0, init1; // split the init map, address-wise
3637 { union { jlong con; jint intcon[2]; } u;
3638 u.con = con;
3639 con0 = u.intcon[0];
3640 con1 = u.intcon[1];
3641 u.con = init;
3642 init0 = u.intcon[0];
3643 init1 = u.intcon[1];
3644 }
3645
3646 Node* old = nodes[j];
3647 assert(old != NULL, "need the prior store");
3648 intptr_t offset = (j * BytesPerLong);
3649
3650 bool split = !Matcher::isSimpleConstant64(con);
3651
3652 if (offset < header_size) {
3653 assert(offset + BytesPerInt >= header_size, "second int counts");
3654 assert(*(jint*)&tiles[j] == 0, "junk in header");
3655 split = true; // only the second word counts
3656 // Example: int a[] = { 42 ... }
3657 } else if (con0 == 0 && init0 == -1) {
3658 split = true; // first word is covered by full inits
3659 // Example: int a[] = { ... foo(), 42 ... }
3660 } else if (con1 == 0 && init1 == -1) {
3661 split = true; // second word is covered by full inits
3662 // Example: int a[] = { ... 42, foo() ... }
3663 }
3664
3665 // Here's a case where init0 is neither 0 nor -1:
3666 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3667 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3668 // In this case the tile is not split; it is (jlong)42.
3669 // The big tile is stored down, and then the foo() value is inserted.
3670 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3671
3672 Node* ctl = old->in(MemNode::Control);
3673 Node* adr = make_raw_address(offset, phase);
3674 const TypePtr* atp = TypeRawPtr::BOTTOM;
3675
3676 // One or two coalesced stores to plop down.
3677 Node* st[2];
3678 intptr_t off[2];
3679 int nst = 0;
3680 if (!split) {
3681 ++new_long;
3682 off[nst] = offset;
3683 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3684 phase->longcon(con), T_LONG, MemNode::unordered);
3685 } else {
3686 // Omit either if it is a zero.
3687 if (con0 != 0) {
3688 ++new_int;
3689 off[nst] = offset;
3690 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3691 phase->intcon(con0), T_INT, MemNode::unordered);
3692 }
3693 if (con1 != 0) {
3694 ++new_int;
3695 offset += BytesPerInt;
3696 adr = make_raw_address(offset, phase);
3697 off[nst] = offset;
3698 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3699 phase->intcon(con1), T_INT, MemNode::unordered);
3700 }
3701 }
3702
3703 // Insert second store first, then the first before the second.
3704 // Insert each one just before any overlapping non-constant stores.
3705 while (nst > 0) {
3706 Node* st1 = st[--nst];
3707 C->copy_node_notes_to(st1, old);
3708 st1 = phase->transform(st1);
3709 offset = off[nst];
3710 assert(offset >= header_size, "do not smash header");
3711 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3712 guarantee(ins_idx != 0, "must re-insert constant store");
3713 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3714 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3715 set_req(--ins_idx, st1);
3716 else
3717 ins_req(ins_idx, st1);
3718 }
3719 }
3720
3721 if (PrintCompilation && WizardMode)
3722 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3723 old_subword, old_long, new_int, new_long);
3724 if (C->log() != NULL)
3725 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3726 old_subword, old_long, new_int, new_long);
3727
3728 // Clean up any remaining occurrences of zmem:
3729 remove_extra_zeroes();
3730 }
3731
3732 // Explore forward from in(start) to find the first fully initialized
3733 // word, and return its offset. Skip groups of subword stores which
3734 // together initialize full words. If in(start) is itself part of a
3735 // fully initialized word, return the offset of in(start). If there
3736 // are no following full-word stores, or if something is fishy, return
3737 // a negative value.
3738 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3739 int int_map = 0;
3740 intptr_t int_map_off = 0;
3741 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3742
3743 for (uint i = start, limit = req(); i < limit; i++) {
3744 Node* st = in(i);
3745
3746 intptr_t st_off = get_store_offset(st, phase);
3747 if (st_off < 0) break; // return conservative answer
3748
3749 int st_size = st->as_Store()->memory_size();
3750 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3751 return st_off; // we found a complete word init
3752 }
3753
3754 // update the map:
3755
3756 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3757 if (this_int_off != int_map_off) {
3758 // reset the map:
3759 int_map = 0;
3760 int_map_off = this_int_off;
3761 }
3762
3763 int subword_off = st_off - this_int_off;
3764 int_map |= right_n_bits(st_size) << subword_off;
3765 if ((int_map & FULL_MAP) == FULL_MAP) {
3766 return this_int_off; // we found a complete word init
3767 }
3768
3769 // Did this store hit or cross the word boundary?
3770 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3771 if (next_int_off == this_int_off + BytesPerInt) {
3772 // We passed the current int, without fully initializing it.
3773 int_map_off = next_int_off;
3774 int_map >>= BytesPerInt;
3775 } else if (next_int_off > this_int_off + BytesPerInt) {
3776 // We passed the current and next int.
3777 return this_int_off + BytesPerInt;
3778 }
3779 }
3780
3781 return -1;
3782 }
3783
3784
3785 // Called when the associated AllocateNode is expanded into CFG.
3786 // At this point, we may perform additional optimizations.
3787 // Linearize the stores by ascending offset, to make memory
3788 // activity as coherent as possible.
3789 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3790 intptr_t header_size,
3791 Node* size_in_bytes,
3792 PhaseGVN* phase) {
3793 assert(!is_complete(), "not already complete");
3794 assert(stores_are_sane(phase), "");
3795 assert(allocation() != NULL, "must be present");
3796
3797 remove_extra_zeroes();
3798
3799 if (ReduceFieldZeroing || ReduceBulkZeroing)
3800 // reduce instruction count for common initialization patterns
3801 coalesce_subword_stores(header_size, size_in_bytes, phase);
3802
3803 Node* zmem = zero_memory(); // initially zero memory state
3804 Node* inits = zmem; // accumulating a linearized chain of inits
3805 #ifdef ASSERT
3806 intptr_t first_offset = allocation()->minimum_header_size();
3807 intptr_t last_init_off = first_offset; // previous init offset
3808 intptr_t last_init_end = first_offset; // previous init offset+size
3809 intptr_t last_tile_end = first_offset; // previous tile offset+size
3810 #endif
3811 intptr_t zeroes_done = header_size;
3812
3813 bool do_zeroing = true; // we might give up if inits are very sparse
3814 int big_init_gaps = 0; // how many large gaps have we seen?
3815
3816 if (ZeroTLAB) do_zeroing = false;
3817 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3818
3819 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3820 Node* st = in(i);
3821 intptr_t st_off = get_store_offset(st, phase);
3822 if (st_off < 0)
3823 break; // unknown junk in the inits
3824 if (st->in(MemNode::Memory) != zmem)
3825 break; // complicated store chains somehow in list
3826
3827 int st_size = st->as_Store()->memory_size();
3828 intptr_t next_init_off = st_off + st_size;
3829
3830 if (do_zeroing && zeroes_done < next_init_off) {
3831 // See if this store needs a zero before it or under it.
3832 intptr_t zeroes_needed = st_off;
3833
3834 if (st_size < BytesPerInt) {
3835 // Look for subword stores which only partially initialize words.
3836 // If we find some, we must lay down some word-level zeroes first,
3837 // underneath the subword stores.
3838 //
3839 // Examples:
3840 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3841 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3842 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3843 //
3844 // Note: coalesce_subword_stores may have already done this,
3845 // if it was prompted by constant non-zero subword initializers.
3846 // But this case can still arise with non-constant stores.
3847
3848 intptr_t next_full_store = find_next_fullword_store(i, phase);
3849
3850 // In the examples above:
3851 // in(i) p q r s x y z
3852 // st_off 12 13 14 15 12 13 14
3853 // st_size 1 1 1 1 1 1 1
3854 // next_full_s. 12 16 16 16 16 16 16
3855 // z's_done 12 16 16 16 12 16 12
3856 // z's_needed 12 16 16 16 16 16 16
3857 // zsize 0 0 0 0 4 0 4
3858 if (next_full_store < 0) {
3859 // Conservative tack: Zero to end of current word.
3860 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3861 } else {
3862 // Zero to beginning of next fully initialized word.
3863 // Or, don't zero at all, if we are already in that word.
3864 assert(next_full_store >= zeroes_needed, "must go forward");
3865 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3866 zeroes_needed = next_full_store;
3867 }
3868 }
3869
3870 if (zeroes_needed > zeroes_done) {
3871 intptr_t zsize = zeroes_needed - zeroes_done;
3872 // Do some incremental zeroing on rawmem, in parallel with inits.
3873 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3874 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3875 zeroes_done, zeroes_needed,
3876 phase);
3877 zeroes_done = zeroes_needed;
3878 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3879 do_zeroing = false; // leave the hole, next time
3880 }
3881 }
3882
3883 // Collect the store and move on:
3884 st->set_req(MemNode::Memory, inits);
3885 inits = st; // put it on the linearized chain
3886 set_req(i, zmem); // unhook from previous position
3887
3888 if (zeroes_done == st_off)
3889 zeroes_done = next_init_off;
3890
3891 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3892
3893 #ifdef ASSERT
3894 // Various order invariants. Weaker than stores_are_sane because
3895 // a large constant tile can be filled in by smaller non-constant stores.
3896 assert(st_off >= last_init_off, "inits do not reverse");
3897 last_init_off = st_off;
3898 const Type* val = NULL;
3899 if (st_size >= BytesPerInt &&
3900 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3901 (int)val->basic_type() < (int)T_OBJECT) {
3902 assert(st_off >= last_tile_end, "tiles do not overlap");
3903 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3904 last_tile_end = MAX2(last_tile_end, next_init_off);
3905 } else {
3906 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3907 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3908 assert(st_off >= last_init_end, "inits do not overlap");
3909 last_init_end = next_init_off; // it's a non-tile
3910 }
3911 #endif //ASSERT
3912 }
3913
3914 remove_extra_zeroes(); // clear out all the zmems left over
3915 add_req(inits);
3916
3917 if (!ZeroTLAB) {
3918 // If anything remains to be zeroed, zero it all now.
3919 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3920 // if it is the last unused 4 bytes of an instance, forget about it
3921 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3922 if (zeroes_done + BytesPerLong >= size_limit) {
3923 assert(allocation() != NULL, "");
3924 if (allocation()->Opcode() == Op_Allocate) {
3925 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3926 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3927 if (zeroes_done == k->layout_helper())
3928 zeroes_done = size_limit;
3929 }
3930 }
3931 if (zeroes_done < size_limit) {
3932 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3933 zeroes_done, size_in_bytes, phase);
3934 }
3935 }
3936
3937 set_complete(phase);
3938 return rawmem;
3939 }
3940
3941
3942 #ifdef ASSERT
3943 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3944 if (is_complete())
3945 return true; // stores could be anything at this point
3946 assert(allocation() != NULL, "must be present");
3947 intptr_t last_off = allocation()->minimum_header_size();
3948 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3949 Node* st = in(i);
3950 intptr_t st_off = get_store_offset(st, phase);
3951 if (st_off < 0) continue; // ignore dead garbage
3952 if (last_off > st_off) {
3953 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
3954 this->dump(2);
3955 assert(false, "ascending store offsets");
3956 return false;
3957 }
3958 last_off = st_off + st->as_Store()->memory_size();
3959 }
3960 return true;
3961 }
3962 #endif //ASSERT
3963
3964
3965
3966
3967 //============================MergeMemNode=====================================
3968 //
3969 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3970 // contributing store or call operations. Each contributor provides the memory
3971 // state for a particular "alias type" (see Compile::alias_type). For example,
3972 // if a MergeMem has an input X for alias category #6, then any memory reference
3973 // to alias category #6 may use X as its memory state input, as an exact equivalent
3974 // to using the MergeMem as a whole.
3975 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3976 //
3977 // (Here, the <N> notation gives the index of the relevant adr_type.)
3978 //
3979 // In one special case (and more cases in the future), alias categories overlap.
3980 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3981 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3982 // it is exactly equivalent to that state W:
3983 // MergeMem(<Bot>: W) <==> W
3984 //
3985 // Usually, the merge has more than one input. In that case, where inputs
3986 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3987 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3988 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3989 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3990 //
3991 // A merge can take a "wide" memory state as one of its narrow inputs.
3992 // This simply means that the merge observes out only the relevant parts of
3993 // the wide input. That is, wide memory states arriving at narrow merge inputs
3994 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3995 //
3996 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3997 // and that memory slices "leak through":
3998 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3999 //
4000 // But, in such a cascade, repeated memory slices can "block the leak":
4001 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4002 //
4003 // In the last example, Y is not part of the combined memory state of the
4004 // outermost MergeMem. The system must, of course, prevent unschedulable
4005 // memory states from arising, so you can be sure that the state Y is somehow
4006 // a precursor to state Y'.
4007 //
4008 //
4009 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4010 // of each MergeMemNode array are exactly the numerical alias indexes, including
4011 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4012 // Compile::alias_type (and kin) produce and manage these indexes.
4013 //
4014 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4015 // (Note that this provides quick access to the top node inside MergeMem methods,
4016 // without the need to reach out via TLS to Compile::current.)
4017 //
4018 // As a consequence of what was just described, a MergeMem that represents a full
4019 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4020 // containing all alias categories.
4021 //
4022 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4023 //
4024 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4025 // a memory state for the alias type <N>, or else the top node, meaning that
4026 // there is no particular input for that alias type. Note that the length of
4027 // a MergeMem is variable, and may be extended at any time to accommodate new
4028 // memory states at larger alias indexes. When merges grow, they are of course
4029 // filled with "top" in the unused in() positions.
4030 //
4031 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4032 // (Top was chosen because it works smoothly with passes like GCM.)
4033 //
4034 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4035 // the type of random VM bits like TLS references.) Since it is always the
4036 // first non-Bot memory slice, some low-level loops use it to initialize an
4037 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4038 //
4039 //
4040 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4041 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4042 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4043 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4044 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4045 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4046 //
4047 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4048 // really that different from the other memory inputs. An abbreviation called
4049 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4050 //
4051 //
4052 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4053 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4054 // that "emerges though" the base memory will be marked as excluding the alias types
4055 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4056 //
4057 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4058 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4059 //
4060 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4061 // (It is currently unimplemented.) As you can see, the resulting merge is
4062 // actually a disjoint union of memory states, rather than an overlay.
4063 //
4064
4065 //------------------------------MergeMemNode-----------------------------------
4066 Node* MergeMemNode::make_empty_memory() {
4067 Node* empty_memory = (Node*) Compile::current()->top();
4068 assert(empty_memory->is_top(), "correct sentinel identity");
4069 return empty_memory;
4070 }
4071
4072 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4073 init_class_id(Class_MergeMem);
4074 // all inputs are nullified in Node::Node(int)
4075 // set_input(0, NULL); // no control input
4076
4077 // Initialize the edges uniformly to top, for starters.
4078 Node* empty_mem = make_empty_memory();
4079 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4080 init_req(i,empty_mem);
4081 }
4082 assert(empty_memory() == empty_mem, "");
4083
4084 if( new_base != NULL && new_base->is_MergeMem() ) {
4085 MergeMemNode* mdef = new_base->as_MergeMem();
4086 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4087 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4088 mms.set_memory(mms.memory2());
4089 }
4090 assert(base_memory() == mdef->base_memory(), "");
4091 } else {
4092 set_base_memory(new_base);
4093 }
4094 }
4095
4096 // Make a new, untransformed MergeMem with the same base as 'mem'.
4097 // If mem is itself a MergeMem, populate the result with the same edges.
4098 MergeMemNode* MergeMemNode::make(Node* mem) {
4099 return new MergeMemNode(mem);
4100 }
4101
4102 //------------------------------cmp--------------------------------------------
4103 uint MergeMemNode::hash() const { return NO_HASH; }
4104 uint MergeMemNode::cmp( const Node &n ) const {
4105 return (&n == this); // Always fail except on self
4106 }
4107
4108 //------------------------------Identity---------------------------------------
4109 Node* MergeMemNode::Identity(PhaseTransform *phase) {
4110 // Identity if this merge point does not record any interesting memory
4111 // disambiguations.
4112 Node* base_mem = base_memory();
4113 Node* empty_mem = empty_memory();
4114 if (base_mem != empty_mem) { // Memory path is not dead?
4115 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4116 Node* mem = in(i);
4117 if (mem != empty_mem && mem != base_mem) {
4118 return this; // Many memory splits; no change
4119 }
4120 }
4121 }
4122 return base_mem; // No memory splits; ID on the one true input
4123 }
4124
4125 //------------------------------Ideal------------------------------------------
4126 // This method is invoked recursively on chains of MergeMem nodes
4127 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4128 // Remove chain'd MergeMems
4129 //
4130 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4131 // relative to the "in(Bot)". Since we are patching both at the same time,
4132 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4133 // but rewrite each "in(i)" relative to the new "in(Bot)".
4134 Node *progress = NULL;
4135
4136
4137 Node* old_base = base_memory();
4138 Node* empty_mem = empty_memory();
4139 if (old_base == empty_mem)
4140 return NULL; // Dead memory path.
4141
4142 MergeMemNode* old_mbase;
4143 if (old_base != NULL && old_base->is_MergeMem())
4144 old_mbase = old_base->as_MergeMem();
4145 else
4146 old_mbase = NULL;
4147 Node* new_base = old_base;
4148
4149 // simplify stacked MergeMems in base memory
4150 if (old_mbase) new_base = old_mbase->base_memory();
4151
4152 // the base memory might contribute new slices beyond my req()
4153 if (old_mbase) grow_to_match(old_mbase);
4154
4155 // Look carefully at the base node if it is a phi.
4156 PhiNode* phi_base;
4157 if (new_base != NULL && new_base->is_Phi())
4158 phi_base = new_base->as_Phi();
4159 else
4160 phi_base = NULL;
4161
4162 Node* phi_reg = NULL;
4163 uint phi_len = (uint)-1;
4164 if (phi_base != NULL && !phi_base->is_copy()) {
4165 // do not examine phi if degraded to a copy
4166 phi_reg = phi_base->region();
4167 phi_len = phi_base->req();
4168 // see if the phi is unfinished
4169 for (uint i = 1; i < phi_len; i++) {
4170 if (phi_base->in(i) == NULL) {
4171 // incomplete phi; do not look at it yet!
4172 phi_reg = NULL;
4173 phi_len = (uint)-1;
4174 break;
4175 }
4176 }
4177 }
4178
4179 // Note: We do not call verify_sparse on entry, because inputs
4180 // can normalize to the base_memory via subsume_node or similar
4181 // mechanisms. This method repairs that damage.
4182
4183 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4184
4185 // Look at each slice.
4186 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4187 Node* old_in = in(i);
4188 // calculate the old memory value
4189 Node* old_mem = old_in;
4190 if (old_mem == empty_mem) old_mem = old_base;
4191 assert(old_mem == memory_at(i), "");
4192
4193 // maybe update (reslice) the old memory value
4194
4195 // simplify stacked MergeMems
4196 Node* new_mem = old_mem;
4197 MergeMemNode* old_mmem;
4198 if (old_mem != NULL && old_mem->is_MergeMem())
4199 old_mmem = old_mem->as_MergeMem();
4200 else
4201 old_mmem = NULL;
4202 if (old_mmem == this) {
4203 // This can happen if loops break up and safepoints disappear.
4204 // A merge of BotPtr (default) with a RawPtr memory derived from a
4205 // safepoint can be rewritten to a merge of the same BotPtr with
4206 // the BotPtr phi coming into the loop. If that phi disappears
4207 // also, we can end up with a self-loop of the mergemem.
4208 // In general, if loops degenerate and memory effects disappear,
4209 // a mergemem can be left looking at itself. This simply means
4210 // that the mergemem's default should be used, since there is
4211 // no longer any apparent effect on this slice.
4212 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4213 // from start. Update the input to TOP.
4214 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4215 }
4216 else if (old_mmem != NULL) {
4217 new_mem = old_mmem->memory_at(i);
4218 }
4219 // else preceding memory was not a MergeMem
4220
4221 // replace equivalent phis (unfortunately, they do not GVN together)
4222 if (new_mem != NULL && new_mem != new_base &&
4223 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4224 if (new_mem->is_Phi()) {
4225 PhiNode* phi_mem = new_mem->as_Phi();
4226 for (uint i = 1; i < phi_len; i++) {
4227 if (phi_base->in(i) != phi_mem->in(i)) {
4228 phi_mem = NULL;
4229 break;
4230 }
4231 }
4232 if (phi_mem != NULL) {
4233 // equivalent phi nodes; revert to the def
4234 new_mem = new_base;
4235 }
4236 }
4237 }
4238
4239 // maybe store down a new value
4240 Node* new_in = new_mem;
4241 if (new_in == new_base) new_in = empty_mem;
4242
4243 if (new_in != old_in) {
4244 // Warning: Do not combine this "if" with the previous "if"
4245 // A memory slice might have be be rewritten even if it is semantically
4246 // unchanged, if the base_memory value has changed.
4247 set_req(i, new_in);
4248 progress = this; // Report progress
4249 }
4250 }
4251
4252 if (new_base != old_base) {
4253 set_req(Compile::AliasIdxBot, new_base);
4254 // Don't use set_base_memory(new_base), because we need to update du.
4255 assert(base_memory() == new_base, "");
4256 progress = this;
4257 }
4258
4259 if( base_memory() == this ) {
4260 // a self cycle indicates this memory path is dead
4261 set_req(Compile::AliasIdxBot, empty_mem);
4262 }
4263
4264 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4265 // Recursion must occur after the self cycle check above
4266 if( base_memory()->is_MergeMem() ) {
4267 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4268 Node *m = phase->transform(new_mbase); // Rollup any cycles
4269 if( m != NULL && (m->is_top() ||
4270 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4271 // propagate rollup of dead cycle to self
4272 set_req(Compile::AliasIdxBot, empty_mem);
4273 }
4274 }
4275
4276 if( base_memory() == empty_mem ) {
4277 progress = this;
4278 // Cut inputs during Parse phase only.
4279 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4280 if( !can_reshape ) {
4281 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4282 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4283 }
4284 }
4285 }
4286
4287 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4288 // Check if PhiNode::Ideal's "Split phis through memory merges"
4289 // transform should be attempted. Look for this->phi->this cycle.
4290 uint merge_width = req();
4291 if (merge_width > Compile::AliasIdxRaw) {
4292 PhiNode* phi = base_memory()->as_Phi();
4293 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4294 if (phi->in(i) == this) {
4295 phase->is_IterGVN()->_worklist.push(phi);
4296 break;
4297 }
4298 }
4299 }
4300 }
4301
4302 assert(progress || verify_sparse(), "please, no dups of base");
4303 return progress;
4304 }
4305
4306 //-------------------------set_base_memory-------------------------------------
4307 void MergeMemNode::set_base_memory(Node *new_base) {
4308 Node* empty_mem = empty_memory();
4309 set_req(Compile::AliasIdxBot, new_base);
4310 assert(memory_at(req()) == new_base, "must set default memory");
4311 // Clear out other occurrences of new_base:
4312 if (new_base != empty_mem) {
4313 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4314 if (in(i) == new_base) set_req(i, empty_mem);
4315 }
4316 }
4317 }
4318
4319 //------------------------------out_RegMask------------------------------------
4320 const RegMask &MergeMemNode::out_RegMask() const {
4321 return RegMask::Empty;
4322 }
4323
4324 //------------------------------dump_spec--------------------------------------
4325 #ifndef PRODUCT
4326 void MergeMemNode::dump_spec(outputStream *st) const {
4327 st->print(" {");
4328 Node* base_mem = base_memory();
4329 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4330 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4331 if (mem == base_mem) { st->print(" -"); continue; }
4332 st->print( " N%d:", mem->_idx );
4333 Compile::current()->get_adr_type(i)->dump_on(st);
4334 }
4335 st->print(" }");
4336 }
4337 #endif // !PRODUCT
4338
4339
4340 #ifdef ASSERT
4341 static bool might_be_same(Node* a, Node* b) {
4342 if (a == b) return true;
4343 if (!(a->is_Phi() || b->is_Phi())) return false;
4344 // phis shift around during optimization
4345 return true; // pretty stupid...
4346 }
4347
4348 // verify a narrow slice (either incoming or outgoing)
4349 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4350 if (!VerifyAliases) return; // don't bother to verify unless requested
4351 if (is_error_reported()) return; // muzzle asserts when debugging an error
4352 if (Node::in_dump()) return; // muzzle asserts when printing
4353 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4354 assert(n != NULL, "");
4355 // Elide intervening MergeMem's
4356 while (n->is_MergeMem()) {
4357 n = n->as_MergeMem()->memory_at(alias_idx);
4358 }
4359 Compile* C = Compile::current();
4360 const TypePtr* n_adr_type = n->adr_type();
4361 if (n == m->empty_memory()) {
4362 // Implicit copy of base_memory()
4363 } else if (n_adr_type != TypePtr::BOTTOM) {
4364 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4365 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4366 } else {
4367 // A few places like make_runtime_call "know" that VM calls are narrow,
4368 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4369 bool expected_wide_mem = false;
4370 if (n == m->base_memory()) {
4371 expected_wide_mem = true;
4372 } else if (alias_idx == Compile::AliasIdxRaw ||
4373 n == m->memory_at(Compile::AliasIdxRaw)) {
4374 expected_wide_mem = true;
4375 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4376 // memory can "leak through" calls on channels that
4377 // are write-once. Allow this also.
4378 expected_wide_mem = true;
4379 }
4380 assert(expected_wide_mem, "expected narrow slice replacement");
4381 }
4382 }
4383 #else // !ASSERT
4384 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4385 #endif
4386
4387
4388 //-----------------------------memory_at---------------------------------------
4389 Node* MergeMemNode::memory_at(uint alias_idx) const {
4390 assert(alias_idx >= Compile::AliasIdxRaw ||
4391 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4392 "must avoid base_memory and AliasIdxTop");
4393
4394 // Otherwise, it is a narrow slice.
4395 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4396 Compile *C = Compile::current();
4397 if (is_empty_memory(n)) {
4398 // the array is sparse; empty slots are the "top" node
4399 n = base_memory();
4400 assert(Node::in_dump()
4401 || n == NULL || n->bottom_type() == Type::TOP
4402 || n->adr_type() == NULL // address is TOP
4403 || n->adr_type() == TypePtr::BOTTOM
4404 || n->adr_type() == TypeRawPtr::BOTTOM
4405 || Compile::current()->AliasLevel() == 0,
4406 "must be a wide memory");
4407 // AliasLevel == 0 if we are organizing the memory states manually.
4408 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4409 } else {
4410 // make sure the stored slice is sane
4411 #ifdef ASSERT
4412 if (is_error_reported() || Node::in_dump()) {
4413 } else if (might_be_same(n, base_memory())) {
4414 // Give it a pass: It is a mostly harmless repetition of the base.
4415 // This can arise normally from node subsumption during optimization.
4416 } else {
4417 verify_memory_slice(this, alias_idx, n);
4418 }
4419 #endif
4420 }
4421 return n;
4422 }
4423
4424 //---------------------------set_memory_at-------------------------------------
4425 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4426 verify_memory_slice(this, alias_idx, n);
4427 Node* empty_mem = empty_memory();
4428 if (n == base_memory()) n = empty_mem; // collapse default
4429 uint need_req = alias_idx+1;
4430 if (req() < need_req) {
4431 if (n == empty_mem) return; // already the default, so do not grow me
4432 // grow the sparse array
4433 do {
4434 add_req(empty_mem);
4435 } while (req() < need_req);
4436 }
4437 set_req( alias_idx, n );
4438 }
4439
4440
4441
4442 //--------------------------iteration_setup------------------------------------
4443 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4444 if (other != NULL) {
4445 grow_to_match(other);
4446 // invariant: the finite support of mm2 is within mm->req()
4447 #ifdef ASSERT
4448 for (uint i = req(); i < other->req(); i++) {
4449 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4450 }
4451 #endif
4452 }
4453 // Replace spurious copies of base_memory by top.
4454 Node* base_mem = base_memory();
4455 if (base_mem != NULL && !base_mem->is_top()) {
4456 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4457 if (in(i) == base_mem)
4458 set_req(i, empty_memory());
4459 }
4460 }
4461 }
4462
4463 //---------------------------grow_to_match-------------------------------------
4464 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4465 Node* empty_mem = empty_memory();
4466 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4467 // look for the finite support of the other memory
4468 for (uint i = other->req(); --i >= req(); ) {
4469 if (other->in(i) != empty_mem) {
4470 uint new_len = i+1;
4471 while (req() < new_len) add_req(empty_mem);
4472 break;
4473 }
4474 }
4475 }
4476
4477 //---------------------------verify_sparse-------------------------------------
4478 #ifndef PRODUCT
4479 bool MergeMemNode::verify_sparse() const {
4480 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4481 Node* base_mem = base_memory();
4482 // The following can happen in degenerate cases, since empty==top.
4483 if (is_empty_memory(base_mem)) return true;
4484 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4485 assert(in(i) != NULL, "sane slice");
4486 if (in(i) == base_mem) return false; // should have been the sentinel value!
4487 }
4488 return true;
4489 }
4490
4491 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4492 Node* n;
4493 n = mm->in(idx);
4494 if (mem == n) return true; // might be empty_memory()
4495 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4496 if (mem == n) return true;
4497 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4498 if (mem == n) return true;
4499 if (n == NULL) break;
4500 }
4501 return false;
4502 }
4503 #endif // !PRODUCT