1 /*
2 * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/connode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/matcher.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/regmask.hpp"
47 #include "opto/rootnode.hpp"
48 #include "utilities/align.hpp"
49 #include "utilities/copy.hpp"
50 #include "utilities/macros.hpp"
51 #include "utilities/powerOfTwo.hpp"
52 #include "utilities/vmError.hpp"
53
54 // Portions of code courtesy of Clifford Click
55
56 // Optimization - Graph Style
57
58 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
59
60 //=============================================================================
61 uint MemNode::size_of() const { return sizeof(*this); }
62
63 const TypePtr *MemNode::adr_type() const {
64 Node* adr = in(Address);
65 if (adr == NULL) return NULL; // node is dead
66 const TypePtr* cross_check = NULL;
67 DEBUG_ONLY(cross_check = _adr_type);
68 return calculate_adr_type(adr->bottom_type(), cross_check);
69 }
70
71 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
72 if (adr != NULL) {
73 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
74 return true;
75 }
76 }
77 return false;
78 }
79
80 #ifndef PRODUCT
81 void MemNode::dump_spec(outputStream *st) const {
82 if (in(Address) == NULL) return; // node is dead
83 #ifndef ASSERT
84 // fake the missing field
85 const TypePtr* _adr_type = NULL;
86 if (in(Address) != NULL)
87 _adr_type = in(Address)->bottom_type()->isa_ptr();
88 #endif
89 dump_adr_type(this, _adr_type, st);
90
91 Compile* C = Compile::current();
92 if (C->alias_type(_adr_type)->is_volatile()) {
93 st->print(" Volatile!");
94 }
95 if (_unaligned_access) {
96 st->print(" unaligned");
97 }
98 if (_mismatched_access) {
99 st->print(" mismatched");
100 }
101 if (_unsafe_access) {
102 st->print(" unsafe");
103 }
104 }
105
106 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
107 st->print(" @");
108 if (adr_type == NULL) {
109 st->print("NULL");
110 } else {
111 adr_type->dump_on(st);
112 Compile* C = Compile::current();
113 Compile::AliasType* atp = NULL;
114 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
115 if (atp == NULL)
116 st->print(", idx=?\?;");
117 else if (atp->index() == Compile::AliasIdxBot)
118 st->print(", idx=Bot;");
119 else if (atp->index() == Compile::AliasIdxTop)
120 st->print(", idx=Top;");
121 else if (atp->index() == Compile::AliasIdxRaw)
122 st->print(", idx=Raw;");
123 else {
124 ciField* field = atp->field();
125 if (field) {
126 st->print(", name=");
127 field->print_name_on(st);
128 }
129 st->print(", idx=%d;", atp->index());
130 }
131 }
132 }
133
134 extern void print_alias_types();
135
136 #endif
137
138 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
139 assert((t_oop != NULL), "sanity");
140 bool is_instance = t_oop->is_known_instance_field();
141 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
142 (load != NULL) && load->is_Load() &&
143 (phase->is_IterGVN() != NULL);
144 if (!(is_instance || is_boxed_value_load))
145 return mchain; // don't try to optimize non-instance types
146 uint instance_id = t_oop->instance_id();
147 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
148 Node *prev = NULL;
149 Node *result = mchain;
150 while (prev != result) {
151 prev = result;
152 if (result == start_mem)
153 break; // hit one of our sentinels
154 // skip over a call which does not affect this memory slice
155 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
156 Node *proj_in = result->in(0);
157 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
158 break; // hit one of our sentinels
159 } else if (proj_in->is_Call()) {
160 // ArrayCopyNodes processed here as well
161 CallNode *call = proj_in->as_Call();
162 if (!call->may_modify(t_oop, phase)) { // returns false for instances
163 result = call->in(TypeFunc::Memory);
164 }
165 } else if (proj_in->is_Initialize()) {
166 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
167 // Stop if this is the initialization for the object instance which
168 // contains this memory slice, otherwise skip over it.
169 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
170 break;
171 }
172 if (is_instance) {
173 result = proj_in->in(TypeFunc::Memory);
174 } else if (is_boxed_value_load) {
175 Node* klass = alloc->in(AllocateNode::KlassNode);
176 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
177 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
178 result = proj_in->in(TypeFunc::Memory); // not related allocation
179 }
180 }
181 } else if (proj_in->is_MemBar()) {
182 ArrayCopyNode* ac = NULL;
183 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
184 break;
185 }
186 result = proj_in->in(TypeFunc::Memory);
187 } else {
188 assert(false, "unexpected projection");
189 }
190 } else if (result->is_ClearArray()) {
191 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
192 // Can not bypass initialization of the instance
193 // we are looking for.
194 break;
195 }
196 // Otherwise skip it (the call updated 'result' value).
197 } else if (result->is_MergeMem()) {
198 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
199 }
200 }
201 return result;
202 }
203
204 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
205 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
206 if (t_oop == NULL)
207 return mchain; // don't try to optimize non-oop types
208 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
209 bool is_instance = t_oop->is_known_instance_field();
210 PhaseIterGVN *igvn = phase->is_IterGVN();
211 if (is_instance && igvn != NULL && result->is_Phi()) {
212 PhiNode *mphi = result->as_Phi();
213 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
214 const TypePtr *t = mphi->adr_type();
215 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
216 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
217 t->is_oopptr()->cast_to_exactness(true)
218 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
219 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
220 // clone the Phi with our address type
221 result = mphi->split_out_instance(t_adr, igvn);
222 } else {
223 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
224 }
225 }
226 return result;
227 }
228
229 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
230 uint alias_idx = phase->C->get_alias_index(tp);
231 Node *mem = mmem;
232 #ifdef ASSERT
233 {
234 // Check that current type is consistent with the alias index used during graph construction
235 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
236 bool consistent = adr_check == NULL || adr_check->empty() ||
237 phase->C->must_alias(adr_check, alias_idx );
238 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
239 if( !consistent && adr_check != NULL && !adr_check->empty() &&
240 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
241 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
242 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
243 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
244 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
245 // don't assert if it is dead code.
246 consistent = true;
247 }
248 if( !consistent ) {
249 st->print("alias_idx==%d, adr_check==", alias_idx);
250 if( adr_check == NULL ) {
251 st->print("NULL");
252 } else {
253 adr_check->dump();
254 }
255 st->cr();
256 print_alias_types();
257 assert(consistent, "adr_check must match alias idx");
258 }
259 }
260 #endif
261 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
262 // means an array I have not precisely typed yet. Do not do any
263 // alias stuff with it any time soon.
264 const TypeOopPtr *toop = tp->isa_oopptr();
265 if( tp->base() != Type::AnyPtr &&
266 !(toop &&
267 toop->klass() != NULL &&
268 toop->klass()->is_java_lang_Object() &&
269 toop->offset() == Type::OffsetBot) ) {
270 // compress paths and change unreachable cycles to TOP
271 // If not, we can update the input infinitely along a MergeMem cycle
272 // Equivalent code in PhiNode::Ideal
273 Node* m = phase->transform(mmem);
274 // If transformed to a MergeMem, get the desired slice
275 // Otherwise the returned node represents memory for every slice
276 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
277 // Update input if it is progress over what we have now
278 }
279 return mem;
280 }
281
282 //--------------------------Ideal_common---------------------------------------
283 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
284 // Unhook non-raw memories from complete (macro-expanded) initializations.
285 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
286 // If our control input is a dead region, kill all below the region
287 Node *ctl = in(MemNode::Control);
288 if (ctl && remove_dead_region(phase, can_reshape))
289 return this;
290 ctl = in(MemNode::Control);
291 // Don't bother trying to transform a dead node
292 if (ctl && ctl->is_top()) return NodeSentinel;
293
294 PhaseIterGVN *igvn = phase->is_IterGVN();
295 // Wait if control on the worklist.
296 if (ctl && can_reshape && igvn != NULL) {
297 Node* bol = NULL;
298 Node* cmp = NULL;
299 if (ctl->in(0)->is_If()) {
300 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
301 bol = ctl->in(0)->in(1);
302 if (bol->is_Bool())
303 cmp = ctl->in(0)->in(1)->in(1);
304 }
305 if (igvn->_worklist.member(ctl) ||
306 (bol != NULL && igvn->_worklist.member(bol)) ||
307 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
308 // This control path may be dead.
309 // Delay this memory node transformation until the control is processed.
310 phase->is_IterGVN()->_worklist.push(this);
311 return NodeSentinel; // caller will return NULL
312 }
313 }
314 // Ignore if memory is dead, or self-loop
315 Node *mem = in(MemNode::Memory);
316 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
317 assert(mem != this, "dead loop in MemNode::Ideal");
318
319 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
320 // This memory slice may be dead.
321 // Delay this mem node transformation until the memory is processed.
322 phase->is_IterGVN()->_worklist.push(this);
323 return NodeSentinel; // caller will return NULL
324 }
325
326 Node *address = in(MemNode::Address);
327 const Type *t_adr = phase->type(address);
328 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
329
330 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
331 // Unsafe off-heap access with zero address. Remove access and other control users
332 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
333 assert(ctl != NULL, "unsafe accesses should be control dependent");
334 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
335 Node* u = ctl->fast_out(i);
336 if (u != ctl) {
337 igvn->rehash_node_delayed(u);
338 int nb = u->replace_edge(ctl, phase->C->top());
339 --i, imax -= nb;
340 }
341 }
342 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
343 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
344 phase->C->root()->add_req(halt);
345 return this;
346 }
347
348 if (can_reshape && igvn != NULL &&
349 (igvn->_worklist.member(address) ||
350 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
351 // The address's base and type may change when the address is processed.
352 // Delay this mem node transformation until the address is processed.
353 phase->is_IterGVN()->_worklist.push(this);
354 return NodeSentinel; // caller will return NULL
355 }
356
357 // Do NOT remove or optimize the next lines: ensure a new alias index
358 // is allocated for an oop pointer type before Escape Analysis.
359 // Note: C++ will not remove it since the call has side effect.
360 if (t_adr->isa_oopptr()) {
361 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
362 }
363
364 Node* base = NULL;
365 if (address->is_AddP()) {
366 base = address->in(AddPNode::Base);
367 }
368 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
369 !t_adr->isa_rawptr()) {
370 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
371 // Skip this node optimization if its address has TOP base.
372 return NodeSentinel; // caller will return NULL
373 }
374
375 // Avoid independent memory operations
376 Node* old_mem = mem;
377
378 // The code which unhooks non-raw memories from complete (macro-expanded)
379 // initializations was removed. After macro-expansion all stores catched
380 // by Initialize node became raw stores and there is no information
381 // which memory slices they modify. So it is unsafe to move any memory
382 // operation above these stores. Also in most cases hooked non-raw memories
383 // were already unhooked by using information from detect_ptr_independence()
384 // and find_previous_store().
385
386 if (mem->is_MergeMem()) {
387 MergeMemNode* mmem = mem->as_MergeMem();
388 const TypePtr *tp = t_adr->is_ptr();
389
390 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
391 }
392
393 if (mem != old_mem) {
394 set_req(MemNode::Memory, mem);
395 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
396 igvn->_worklist.push(old_mem);
397 }
398 if (phase->type(mem) == Type::TOP) return NodeSentinel;
399 return this;
400 }
401
402 // let the subclass continue analyzing...
403 return NULL;
404 }
405
406 // Helper function for proving some simple control dominations.
407 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
408 // Already assumes that 'dom' is available at 'sub', and that 'sub'
409 // is not a constant (dominated by the method's StartNode).
410 // Used by MemNode::find_previous_store to prove that the
411 // control input of a memory operation predates (dominates)
412 // an allocation it wants to look past.
413 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
414 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
415 return false; // Conservative answer for dead code
416
417 // Check 'dom'. Skip Proj and CatchProj nodes.
418 dom = dom->find_exact_control(dom);
419 if (dom == NULL || dom->is_top())
420 return false; // Conservative answer for dead code
421
422 if (dom == sub) {
423 // For the case when, for example, 'sub' is Initialize and the original
424 // 'dom' is Proj node of the 'sub'.
425 return false;
426 }
427
428 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
429 return true;
430
431 // 'dom' dominates 'sub' if its control edge and control edges
432 // of all its inputs dominate or equal to sub's control edge.
433
434 // Currently 'sub' is either Allocate, Initialize or Start nodes.
435 // Or Region for the check in LoadNode::Ideal();
436 // 'sub' should have sub->in(0) != NULL.
437 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
438 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
439
440 // Get control edge of 'sub'.
441 Node* orig_sub = sub;
442 sub = sub->find_exact_control(sub->in(0));
443 if (sub == NULL || sub->is_top())
444 return false; // Conservative answer for dead code
445
446 assert(sub->is_CFG(), "expecting control");
447
448 if (sub == dom)
449 return true;
450
451 if (sub->is_Start() || sub->is_Root())
452 return false;
453
454 {
455 // Check all control edges of 'dom'.
456
457 ResourceMark rm;
458 Arena* arena = Thread::current()->resource_area();
459 Node_List nlist(arena);
460 Unique_Node_List dom_list(arena);
461
462 dom_list.push(dom);
463 bool only_dominating_controls = false;
464
465 for (uint next = 0; next < dom_list.size(); next++) {
466 Node* n = dom_list.at(next);
467 if (n == orig_sub)
468 return false; // One of dom's inputs dominated by sub.
469 if (!n->is_CFG() && n->pinned()) {
470 // Check only own control edge for pinned non-control nodes.
471 n = n->find_exact_control(n->in(0));
472 if (n == NULL || n->is_top())
473 return false; // Conservative answer for dead code
474 assert(n->is_CFG(), "expecting control");
475 dom_list.push(n);
476 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
477 only_dominating_controls = true;
478 } else if (n->is_CFG()) {
479 if (n->dominates(sub, nlist))
480 only_dominating_controls = true;
481 else
482 return false;
483 } else {
484 // First, own control edge.
485 Node* m = n->find_exact_control(n->in(0));
486 if (m != NULL) {
487 if (m->is_top())
488 return false; // Conservative answer for dead code
489 dom_list.push(m);
490 }
491 // Now, the rest of edges.
492 uint cnt = n->req();
493 for (uint i = 1; i < cnt; i++) {
494 m = n->find_exact_control(n->in(i));
495 if (m == NULL || m->is_top())
496 continue;
497 dom_list.push(m);
498 }
499 }
500 }
501 return only_dominating_controls;
502 }
503 }
504
505 //---------------------detect_ptr_independence---------------------------------
506 // Used by MemNode::find_previous_store to prove that two base
507 // pointers are never equal.
508 // The pointers are accompanied by their associated allocations,
509 // if any, which have been previously discovered by the caller.
510 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
511 Node* p2, AllocateNode* a2,
512 PhaseTransform* phase) {
513 // Attempt to prove that these two pointers cannot be aliased.
514 // They may both manifestly be allocations, and they should differ.
515 // Or, if they are not both allocations, they can be distinct constants.
516 // Otherwise, one is an allocation and the other a pre-existing value.
517 if (a1 == NULL && a2 == NULL) { // neither an allocation
518 return (p1 != p2) && p1->is_Con() && p2->is_Con();
519 } else if (a1 != NULL && a2 != NULL) { // both allocations
520 return (a1 != a2);
521 } else if (a1 != NULL) { // one allocation a1
522 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
523 return all_controls_dominate(p2, a1);
524 } else { //(a2 != NULL) // one allocation a2
525 return all_controls_dominate(p1, a2);
526 }
527 return false;
528 }
529
530
531 // Find an arraycopy that must have set (can_see_stored_value=true) or
532 // could have set (can_see_stored_value=false) the value for this load
533 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
534 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
535 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
536 Node* mb = mem->in(0);
537 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
538 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
539 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
540 if (ac->is_clonebasic()) {
541 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase);
542 if (alloc != NULL && alloc == ld_alloc) {
543 return ac;
544 }
545 }
546 }
547 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
548 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
549
550 if (ac->is_arraycopy_validated() ||
551 ac->is_copyof_validated() ||
552 ac->is_copyofrange_validated()) {
553 Node* ld_addp = in(MemNode::Address);
554 if (ld_addp->is_AddP()) {
555 Node* ld_base = ld_addp->in(AddPNode::Address);
556 Node* ld_offs = ld_addp->in(AddPNode::Offset);
557
558 Node* dest = ac->in(ArrayCopyNode::Dest);
559
560 if (dest == ld_base) {
561 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
562 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
563 return ac;
564 }
565 if (!can_see_stored_value) {
566 mem = ac->in(TypeFunc::Memory);
567 }
568 }
569 }
570 }
571 }
572 return NULL;
573 }
574
575 // The logic for reordering loads and stores uses four steps:
576 // (a) Walk carefully past stores and initializations which we
577 // can prove are independent of this load.
578 // (b) Observe that the next memory state makes an exact match
579 // with self (load or store), and locate the relevant store.
580 // (c) Ensure that, if we were to wire self directly to the store,
581 // the optimizer would fold it up somehow.
582 // (d) Do the rewiring, and return, depending on some other part of
583 // the optimizer to fold up the load.
584 // This routine handles steps (a) and (b). Steps (c) and (d) are
585 // specific to loads and stores, so they are handled by the callers.
586 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
587 //
588 Node* MemNode::find_previous_store(PhaseTransform* phase) {
589 Node* ctrl = in(MemNode::Control);
590 Node* adr = in(MemNode::Address);
591 intptr_t offset = 0;
592 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
593 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
594
595 if (offset == Type::OffsetBot)
596 return NULL; // cannot unalias unless there are precise offsets
597
598 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
599 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
600
601 intptr_t size_in_bytes = memory_size();
602
603 Node* mem = in(MemNode::Memory); // start searching here...
604
605 int cnt = 50; // Cycle limiter
606 for (;;) { // While we can dance past unrelated stores...
607 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
608
609 Node* prev = mem;
610 if (mem->is_Store()) {
611 Node* st_adr = mem->in(MemNode::Address);
612 intptr_t st_offset = 0;
613 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
614 if (st_base == NULL)
615 break; // inscrutable pointer
616
617 // For raw accesses it's not enough to prove that constant offsets don't intersect.
618 // We need the bases to be the equal in order for the offset check to make sense.
619 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
620 break;
621 }
622
623 if (st_offset != offset && st_offset != Type::OffsetBot) {
624 const int MAX_STORE = BytesPerLong;
625 if (st_offset >= offset + size_in_bytes ||
626 st_offset <= offset - MAX_STORE ||
627 st_offset <= offset - mem->as_Store()->memory_size()) {
628 // Success: The offsets are provably independent.
629 // (You may ask, why not just test st_offset != offset and be done?
630 // The answer is that stores of different sizes can co-exist
631 // in the same sequence of RawMem effects. We sometimes initialize
632 // a whole 'tile' of array elements with a single jint or jlong.)
633 mem = mem->in(MemNode::Memory);
634 continue; // (a) advance through independent store memory
635 }
636 }
637 if (st_base != base &&
638 detect_ptr_independence(base, alloc,
639 st_base,
640 AllocateNode::Ideal_allocation(st_base, phase),
641 phase)) {
642 // Success: The bases are provably independent.
643 mem = mem->in(MemNode::Memory);
644 continue; // (a) advance through independent store memory
645 }
646
647 // (b) At this point, if the bases or offsets do not agree, we lose,
648 // since we have not managed to prove 'this' and 'mem' independent.
649 if (st_base == base && st_offset == offset) {
650 return mem; // let caller handle steps (c), (d)
651 }
652
653 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
654 InitializeNode* st_init = mem->in(0)->as_Initialize();
655 AllocateNode* st_alloc = st_init->allocation();
656 if (st_alloc == NULL)
657 break; // something degenerated
658 bool known_identical = false;
659 bool known_independent = false;
660 if (alloc == st_alloc)
661 known_identical = true;
662 else if (alloc != NULL)
663 known_independent = true;
664 else if (all_controls_dominate(this, st_alloc))
665 known_independent = true;
666
667 if (known_independent) {
668 // The bases are provably independent: Either they are
669 // manifestly distinct allocations, or else the control
670 // of this load dominates the store's allocation.
671 int alias_idx = phase->C->get_alias_index(adr_type());
672 if (alias_idx == Compile::AliasIdxRaw) {
673 mem = st_alloc->in(TypeFunc::Memory);
674 } else {
675 mem = st_init->memory(alias_idx);
676 }
677 continue; // (a) advance through independent store memory
678 }
679
680 // (b) at this point, if we are not looking at a store initializing
681 // the same allocation we are loading from, we lose.
682 if (known_identical) {
683 // From caller, can_see_stored_value will consult find_captured_store.
684 return mem; // let caller handle steps (c), (d)
685 }
686
687 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
688 if (prev != mem) {
689 // Found an arraycopy but it doesn't affect that load
690 continue;
691 }
692 // Found an arraycopy that may affect that load
693 return mem;
694 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
695 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
696 if (mem->is_Proj() && mem->in(0)->is_Call()) {
697 // ArrayCopyNodes processed here as well.
698 CallNode *call = mem->in(0)->as_Call();
699 if (!call->may_modify(addr_t, phase)) {
700 mem = call->in(TypeFunc::Memory);
701 continue; // (a) advance through independent call memory
702 }
703 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
704 ArrayCopyNode* ac = NULL;
705 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
706 break;
707 }
708 mem = mem->in(0)->in(TypeFunc::Memory);
709 continue; // (a) advance through independent MemBar memory
710 } else if (mem->is_ClearArray()) {
711 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
712 // (the call updated 'mem' value)
713 continue; // (a) advance through independent allocation memory
714 } else {
715 // Can not bypass initialization of the instance
716 // we are looking for.
717 return mem;
718 }
719 } else if (mem->is_MergeMem()) {
720 int alias_idx = phase->C->get_alias_index(adr_type());
721 mem = mem->as_MergeMem()->memory_at(alias_idx);
722 continue; // (a) advance through independent MergeMem memory
723 }
724 }
725
726 // Unless there is an explicit 'continue', we must bail out here,
727 // because 'mem' is an inscrutable memory state (e.g., a call).
728 break;
729 }
730
731 return NULL; // bail out
732 }
733
734 //----------------------calculate_adr_type-------------------------------------
735 // Helper function. Notices when the given type of address hits top or bottom.
736 // Also, asserts a cross-check of the type against the expected address type.
737 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
738 if (t == Type::TOP) return NULL; // does not touch memory any more?
739 #ifdef ASSERT
740 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
741 #endif
742 const TypePtr* tp = t->isa_ptr();
743 if (tp == NULL) {
744 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
745 return TypePtr::BOTTOM; // touches lots of memory
746 } else {
747 #ifdef ASSERT
748 // %%%% [phh] We don't check the alias index if cross_check is
749 // TypeRawPtr::BOTTOM. Needs to be investigated.
750 if (cross_check != NULL &&
751 cross_check != TypePtr::BOTTOM &&
752 cross_check != TypeRawPtr::BOTTOM) {
753 // Recheck the alias index, to see if it has changed (due to a bug).
754 Compile* C = Compile::current();
755 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
756 "must stay in the original alias category");
757 // The type of the address must be contained in the adr_type,
758 // disregarding "null"-ness.
759 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
760 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
761 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
762 "real address must not escape from expected memory type");
763 }
764 #endif
765 return tp;
766 }
767 }
768
769 //=============================================================================
770 // Should LoadNode::Ideal() attempt to remove control edges?
771 bool LoadNode::can_remove_control() const {
772 return true;
773 }
774 uint LoadNode::size_of() const { return sizeof(*this); }
775 bool LoadNode::cmp( const Node &n ) const
776 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
777 const Type *LoadNode::bottom_type() const { return _type; }
778 uint LoadNode::ideal_reg() const {
779 return _type->ideal_reg();
780 }
781
782 #ifndef PRODUCT
783 void LoadNode::dump_spec(outputStream *st) const {
784 MemNode::dump_spec(st);
785 if( !Verbose && !WizardMode ) {
786 // standard dump does this in Verbose and WizardMode
787 st->print(" #"); _type->dump_on(st);
788 }
789 if (!depends_only_on_test()) {
790 st->print(" (does not depend only on test)");
791 }
792 }
793 #endif
794
795 #ifdef ASSERT
796 //----------------------------is_immutable_value-------------------------------
797 // Helper function to allow a raw load without control edge for some cases
798 bool LoadNode::is_immutable_value(Node* adr) {
799 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
800 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
801 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
802 in_bytes(JavaThread::osthread_offset())));
803 }
804 #endif
805
806 //----------------------------LoadNode::make-----------------------------------
807 // Polymorphic factory method:
808 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
809 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
810 Compile* C = gvn.C;
811
812 // sanity check the alias category against the created node type
813 assert(!(adr_type->isa_oopptr() &&
814 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
815 "use LoadKlassNode instead");
816 assert(!(adr_type->isa_aryptr() &&
817 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
818 "use LoadRangeNode instead");
819 // Check control edge of raw loads
820 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
821 // oop will be recorded in oop map if load crosses safepoint
822 rt->isa_oopptr() || is_immutable_value(adr),
823 "raw memory operations should have control edge");
824 LoadNode* load = NULL;
825 switch (bt) {
826 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
827 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
828 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
829 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
830 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
831 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
832 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
833 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
834 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
835 case T_OBJECT:
836 #ifdef _LP64
837 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
838 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
839 } else
840 #endif
841 {
842 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
843 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
844 }
845 break;
846 default:
847 ShouldNotReachHere();
848 break;
849 }
850 assert(load != NULL, "LoadNode should have been created");
851 if (unaligned) {
852 load->set_unaligned_access();
853 }
854 if (mismatched) {
855 load->set_mismatched_access();
856 }
857 if (unsafe) {
858 load->set_unsafe_access();
859 }
860 load->set_barrier_data(barrier_data);
861 if (load->Opcode() == Op_LoadN) {
862 Node* ld = gvn.transform(load);
863 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
864 }
865
866 return load;
867 }
868
869 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
870 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
871 bool require_atomic = true;
872 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
873 if (unaligned) {
874 load->set_unaligned_access();
875 }
876 if (mismatched) {
877 load->set_mismatched_access();
878 }
879 if (unsafe) {
880 load->set_unsafe_access();
881 }
882 load->set_barrier_data(barrier_data);
883 return load;
884 }
885
886 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
887 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
888 bool require_atomic = true;
889 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
890 if (unaligned) {
891 load->set_unaligned_access();
892 }
893 if (mismatched) {
894 load->set_mismatched_access();
895 }
896 if (unsafe) {
897 load->set_unsafe_access();
898 }
899 load->set_barrier_data(barrier_data);
900 return load;
901 }
902
903
904
905 //------------------------------hash-------------------------------------------
906 uint LoadNode::hash() const {
907 // unroll addition of interesting fields
908 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
909 }
910
911 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
912 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
913 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
914 bool is_stable_ary = FoldStableValues &&
915 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
916 tp->isa_aryptr()->is_stable();
917
918 return (eliminate_boxing && non_volatile) || is_stable_ary;
919 }
920
921 return false;
922 }
923
924 // Is the value loaded previously stored by an arraycopy? If so return
925 // a load node that reads from the source array so we may be able to
926 // optimize out the ArrayCopy node later.
927 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
928 Node* ld_adr = in(MemNode::Address);
929 intptr_t ld_off = 0;
930 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
931 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
932 if (ac != NULL) {
933 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
934
935 Node* mem = ac->in(TypeFunc::Memory);
936 Node* ctl = ac->in(0);
937 Node* src = ac->in(ArrayCopyNode::Src);
938
939 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
940 return NULL;
941 }
942
943 LoadNode* ld = clone()->as_Load();
944 Node* addp = in(MemNode::Address)->clone();
945 if (ac->as_ArrayCopy()->is_clonebasic()) {
946 assert(ld_alloc != NULL, "need an alloc");
947 assert(addp->is_AddP(), "address must be addp");
948 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
949 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
950 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
951 addp->set_req(AddPNode::Base, src);
952 addp->set_req(AddPNode::Address, src);
953 } else {
954 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
955 ac->as_ArrayCopy()->is_copyof_validated() ||
956 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
957 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
958 addp->set_req(AddPNode::Base, src);
959 addp->set_req(AddPNode::Address, src);
960
961 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
962 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
963 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
964 uint shift = exact_log2(type2aelembytes(ary_elem));
965
966 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
967 #ifdef _LP64
968 diff = phase->transform(new ConvI2LNode(diff));
969 #endif
970 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
971
972 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
973 addp->set_req(AddPNode::Offset, offset);
974 }
975 addp = phase->transform(addp);
976 #ifdef ASSERT
977 const TypePtr* adr_type = phase->type(addp)->is_ptr();
978 ld->_adr_type = adr_type;
979 #endif
980 ld->set_req(MemNode::Address, addp);
981 ld->set_req(0, ctl);
982 ld->set_req(MemNode::Memory, mem);
983 // load depends on the tests that validate the arraycopy
984 ld->_control_dependency = UnknownControl;
985 return ld;
986 }
987 return NULL;
988 }
989
990
991 //---------------------------can_see_stored_value------------------------------
992 // This routine exists to make sure this set of tests is done the same
993 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
994 // will change the graph shape in a way which makes memory alive twice at the
995 // same time (uses the Oracle model of aliasing), then some
996 // LoadXNode::Identity will fold things back to the equivalence-class model
997 // of aliasing.
998 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
999 Node* ld_adr = in(MemNode::Address);
1000 intptr_t ld_off = 0;
1001 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1002 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase);
1003 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1004 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1005 // This is more general than load from boxing objects.
1006 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1007 uint alias_idx = atp->index();
1008 bool final = !atp->is_rewritable();
1009 Node* result = NULL;
1010 Node* current = st;
1011 // Skip through chains of MemBarNodes checking the MergeMems for
1012 // new states for the slice of this load. Stop once any other
1013 // kind of node is encountered. Loads from final memory can skip
1014 // through any kind of MemBar but normal loads shouldn't skip
1015 // through MemBarAcquire since the could allow them to move out of
1016 // a synchronized region.
1017 while (current->is_Proj()) {
1018 int opc = current->in(0)->Opcode();
1019 if ((final && (opc == Op_MemBarAcquire ||
1020 opc == Op_MemBarAcquireLock ||
1021 opc == Op_LoadFence)) ||
1022 opc == Op_MemBarRelease ||
1023 opc == Op_StoreFence ||
1024 opc == Op_MemBarReleaseLock ||
1025 opc == Op_MemBarStoreStore ||
1026 opc == Op_MemBarCPUOrder) {
1027 Node* mem = current->in(0)->in(TypeFunc::Memory);
1028 if (mem->is_MergeMem()) {
1029 MergeMemNode* merge = mem->as_MergeMem();
1030 Node* new_st = merge->memory_at(alias_idx);
1031 if (new_st == merge->base_memory()) {
1032 // Keep searching
1033 current = new_st;
1034 continue;
1035 }
1036 // Save the new memory state for the slice and fall through
1037 // to exit.
1038 result = new_st;
1039 }
1040 }
1041 break;
1042 }
1043 if (result != NULL) {
1044 st = result;
1045 }
1046 }
1047
1048 // Loop around twice in the case Load -> Initialize -> Store.
1049 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1050 for (int trip = 0; trip <= 1; trip++) {
1051
1052 if (st->is_Store()) {
1053 Node* st_adr = st->in(MemNode::Address);
1054 if (!phase->eqv(st_adr, ld_adr)) {
1055 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1056 intptr_t st_off = 0;
1057 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1058 if (ld_base == NULL) return NULL;
1059 if (st_base == NULL) return NULL;
1060 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL;
1061 if (ld_off != st_off) return NULL;
1062 if (ld_off == Type::OffsetBot) return NULL;
1063 // Same base, same offset.
1064 // Possible improvement for arrays: check index value instead of absolute offset.
1065
1066 // At this point we have proven something like this setup:
1067 // B = << base >>
1068 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1069 // S = StoreQ(AddP( B , #Off), V)
1070 // (Actually, we haven't yet proven the Q's are the same.)
1071 // In other words, we are loading from a casted version of
1072 // the same pointer-and-offset that we stored to.
1073 // Casted version may carry a dependency and it is respected.
1074 // Thus, we are able to replace L by V.
1075 }
1076 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1077 if (store_Opcode() != st->Opcode())
1078 return NULL;
1079 return st->in(MemNode::ValueIn);
1080 }
1081
1082 // A load from a freshly-created object always returns zero.
1083 // (This can happen after LoadNode::Ideal resets the load's memory input
1084 // to find_captured_store, which returned InitializeNode::zero_memory.)
1085 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1086 (st->in(0) == ld_alloc) &&
1087 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1088 // return a zero value for the load's basic type
1089 // (This is one of the few places where a generic PhaseTransform
1090 // can create new nodes. Think of it as lazily manifesting
1091 // virtually pre-existing constants.)
1092 return phase->zerocon(memory_type());
1093 }
1094
1095 // A load from an initialization barrier can match a captured store.
1096 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1097 InitializeNode* init = st->in(0)->as_Initialize();
1098 AllocateNode* alloc = init->allocation();
1099 if ((alloc != NULL) && (alloc == ld_alloc)) {
1100 // examine a captured store value
1101 st = init->find_captured_store(ld_off, memory_size(), phase);
1102 if (st != NULL) {
1103 continue; // take one more trip around
1104 }
1105 }
1106 }
1107
1108 // Load boxed value from result of valueOf() call is input parameter.
1109 if (this->is_Load() && ld_adr->is_AddP() &&
1110 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1111 intptr_t ignore = 0;
1112 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1113 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1114 base = bs->step_over_gc_barrier(base);
1115 if (base != NULL && base->is_Proj() &&
1116 base->as_Proj()->_con == TypeFunc::Parms &&
1117 base->in(0)->is_CallStaticJava() &&
1118 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1119 return base->in(0)->in(TypeFunc::Parms);
1120 }
1121 }
1122
1123 break;
1124 }
1125
1126 return NULL;
1127 }
1128
1129 //----------------------is_instance_field_load_with_local_phi------------------
1130 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1131 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1132 in(Address)->is_AddP() ) {
1133 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1134 // Only instances and boxed values.
1135 if( t_oop != NULL &&
1136 (t_oop->is_ptr_to_boxed_value() ||
1137 t_oop->is_known_instance_field()) &&
1138 t_oop->offset() != Type::OffsetBot &&
1139 t_oop->offset() != Type::OffsetTop) {
1140 return true;
1141 }
1142 }
1143 return false;
1144 }
1145
1146 //------------------------------Identity---------------------------------------
1147 // Loads are identity if previous store is to same address
1148 Node* LoadNode::Identity(PhaseGVN* phase) {
1149 // If the previous store-maker is the right kind of Store, and the store is
1150 // to the same address, then we are equal to the value stored.
1151 Node* mem = in(Memory);
1152 Node* value = can_see_stored_value(mem, phase);
1153 if( value ) {
1154 // byte, short & char stores truncate naturally.
1155 // A load has to load the truncated value which requires
1156 // some sort of masking operation and that requires an
1157 // Ideal call instead of an Identity call.
1158 if (memory_size() < BytesPerInt) {
1159 // If the input to the store does not fit with the load's result type,
1160 // it must be truncated via an Ideal call.
1161 if (!phase->type(value)->higher_equal(phase->type(this)))
1162 return this;
1163 }
1164 // (This works even when value is a Con, but LoadNode::Value
1165 // usually runs first, producing the singleton type of the Con.)
1166 return value;
1167 }
1168
1169 // Search for an existing data phi which was generated before for the same
1170 // instance's field to avoid infinite generation of phis in a loop.
1171 Node *region = mem->in(0);
1172 if (is_instance_field_load_with_local_phi(region)) {
1173 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1174 int this_index = phase->C->get_alias_index(addr_t);
1175 int this_offset = addr_t->offset();
1176 int this_iid = addr_t->instance_id();
1177 if (!addr_t->is_known_instance() &&
1178 addr_t->is_ptr_to_boxed_value()) {
1179 // Use _idx of address base (could be Phi node) for boxed values.
1180 intptr_t ignore = 0;
1181 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1182 if (base == NULL) {
1183 return this;
1184 }
1185 this_iid = base->_idx;
1186 }
1187 const Type* this_type = bottom_type();
1188 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1189 Node* phi = region->fast_out(i);
1190 if (phi->is_Phi() && phi != mem &&
1191 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1192 return phi;
1193 }
1194 }
1195 }
1196
1197 return this;
1198 }
1199
1200 // Construct an equivalent unsigned load.
1201 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1202 BasicType bt = T_ILLEGAL;
1203 const Type* rt = NULL;
1204 switch (Opcode()) {
1205 case Op_LoadUB: return this;
1206 case Op_LoadUS: return this;
1207 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1208 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1209 default:
1210 assert(false, "no unsigned variant: %s", Name());
1211 return NULL;
1212 }
1213 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1214 raw_adr_type(), rt, bt, _mo, _control_dependency,
1215 is_unaligned_access(), is_mismatched_access());
1216 }
1217
1218 // Construct an equivalent signed load.
1219 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1220 BasicType bt = T_ILLEGAL;
1221 const Type* rt = NULL;
1222 switch (Opcode()) {
1223 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1224 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1225 case Op_LoadB: // fall through
1226 case Op_LoadS: // fall through
1227 case Op_LoadI: // fall through
1228 case Op_LoadL: return this;
1229 default:
1230 assert(false, "no signed variant: %s", Name());
1231 return NULL;
1232 }
1233 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1234 raw_adr_type(), rt, bt, _mo, _control_dependency,
1235 is_unaligned_access(), is_mismatched_access());
1236 }
1237
1238 // We're loading from an object which has autobox behaviour.
1239 // If this object is result of a valueOf call we'll have a phi
1240 // merging a newly allocated object and a load from the cache.
1241 // We want to replace this load with the original incoming
1242 // argument to the valueOf call.
1243 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1244 assert(phase->C->eliminate_boxing(), "sanity");
1245 intptr_t ignore = 0;
1246 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1247 if ((base == NULL) || base->is_Phi()) {
1248 // Push the loads from the phi that comes from valueOf up
1249 // through it to allow elimination of the loads and the recovery
1250 // of the original value. It is done in split_through_phi().
1251 return NULL;
1252 } else if (base->is_Load() ||
1253 (base->is_DecodeN() && base->in(1)->is_Load())) {
1254 // Eliminate the load of boxed value for integer types from the cache
1255 // array by deriving the value from the index into the array.
1256 // Capture the offset of the load and then reverse the computation.
1257
1258 // Get LoadN node which loads a boxing object from 'cache' array.
1259 if (base->is_DecodeN()) {
1260 base = base->in(1);
1261 }
1262 if (!base->in(Address)->is_AddP()) {
1263 return NULL; // Complex address
1264 }
1265 AddPNode* address = base->in(Address)->as_AddP();
1266 Node* cache_base = address->in(AddPNode::Base);
1267 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1268 // Get ConP node which is static 'cache' field.
1269 cache_base = cache_base->in(1);
1270 }
1271 if ((cache_base != NULL) && cache_base->is_Con()) {
1272 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1273 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1274 Node* elements[4];
1275 int shift = exact_log2(type2aelembytes(T_OBJECT));
1276 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1277 if (count > 0 && elements[0]->is_Con() &&
1278 (count == 1 ||
1279 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1280 elements[1]->in(2) == phase->intcon(shift)))) {
1281 ciObjArray* array = base_type->const_oop()->as_obj_array();
1282 // Fetch the box object cache[0] at the base of the array and get its value
1283 ciInstance* box = array->obj_at(0)->as_instance();
1284 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1285 assert(ik->is_box_klass(), "sanity");
1286 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1287 if (ik->nof_nonstatic_fields() == 1) {
1288 // This should be true nonstatic_field_at requires calling
1289 // nof_nonstatic_fields so check it anyway
1290 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1291 BasicType bt = c.basic_type();
1292 // Only integer types have boxing cache.
1293 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1294 bt == T_BYTE || bt == T_SHORT ||
1295 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1296 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1297 if (cache_low != (int)cache_low) {
1298 return NULL; // should not happen since cache is array indexed by value
1299 }
1300 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1301 if (offset != (int)offset) {
1302 return NULL; // should not happen since cache is array indexed by value
1303 }
1304 // Add up all the offsets making of the address of the load
1305 Node* result = elements[0];
1306 for (int i = 1; i < count; i++) {
1307 result = phase->transform(new AddXNode(result, elements[i]));
1308 }
1309 // Remove the constant offset from the address and then
1310 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1311 // remove the scaling of the offset to recover the original index.
1312 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1313 // Peel the shift off directly but wrap it in a dummy node
1314 // since Ideal can't return existing nodes
1315 result = new RShiftXNode(result->in(1), phase->intcon(0));
1316 } else if (result->is_Add() && result->in(2)->is_Con() &&
1317 result->in(1)->Opcode() == Op_LShiftX &&
1318 result->in(1)->in(2) == phase->intcon(shift)) {
1319 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1320 // but for boxing cache access we know that X<<Z will not overflow
1321 // (there is range check) so we do this optimizatrion by hand here.
1322 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1323 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1324 } else {
1325 result = new RShiftXNode(result, phase->intcon(shift));
1326 }
1327 #ifdef _LP64
1328 if (bt != T_LONG) {
1329 result = new ConvL2INode(phase->transform(result));
1330 }
1331 #else
1332 if (bt == T_LONG) {
1333 result = new ConvI2LNode(phase->transform(result));
1334 }
1335 #endif
1336 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1337 // Need to preserve unboxing load type if it is unsigned.
1338 switch(this->Opcode()) {
1339 case Op_LoadUB:
1340 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1341 break;
1342 case Op_LoadUS:
1343 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1344 break;
1345 }
1346 return result;
1347 }
1348 }
1349 }
1350 }
1351 }
1352 return NULL;
1353 }
1354
1355 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1356 Node* region = phi->in(0);
1357 if (region == NULL) {
1358 return false; // Wait stable graph
1359 }
1360 uint cnt = phi->req();
1361 for (uint i = 1; i < cnt; i++) {
1362 Node* rc = region->in(i);
1363 if (rc == NULL || phase->type(rc) == Type::TOP)
1364 return false; // Wait stable graph
1365 Node* in = phi->in(i);
1366 if (in == NULL || phase->type(in) == Type::TOP)
1367 return false; // Wait stable graph
1368 }
1369 return true;
1370 }
1371 //------------------------------split_through_phi------------------------------
1372 // Split instance or boxed field load through Phi.
1373 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1374 Node* mem = in(Memory);
1375 Node* address = in(Address);
1376 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1377
1378 assert((t_oop != NULL) &&
1379 (t_oop->is_known_instance_field() ||
1380 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1381
1382 Compile* C = phase->C;
1383 intptr_t ignore = 0;
1384 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1385 bool base_is_phi = (base != NULL) && base->is_Phi();
1386 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1387 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1388 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1389
1390 if (!((mem->is_Phi() || base_is_phi) &&
1391 (load_boxed_values || t_oop->is_known_instance_field()))) {
1392 return NULL; // memory is not Phi
1393 }
1394
1395 if (mem->is_Phi()) {
1396 if (!stable_phi(mem->as_Phi(), phase)) {
1397 return NULL; // Wait stable graph
1398 }
1399 uint cnt = mem->req();
1400 // Check for loop invariant memory.
1401 if (cnt == 3) {
1402 for (uint i = 1; i < cnt; i++) {
1403 Node* in = mem->in(i);
1404 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1405 if (m == mem) {
1406 if (i == 1) {
1407 // if the first edge was a loop, check second edge too.
1408 // If both are replaceable - we are in an infinite loop
1409 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1410 if (n == mem) {
1411 break;
1412 }
1413 }
1414 set_req(Memory, mem->in(cnt - i));
1415 return this; // made change
1416 }
1417 }
1418 }
1419 }
1420 if (base_is_phi) {
1421 if (!stable_phi(base->as_Phi(), phase)) {
1422 return NULL; // Wait stable graph
1423 }
1424 uint cnt = base->req();
1425 // Check for loop invariant memory.
1426 if (cnt == 3) {
1427 for (uint i = 1; i < cnt; i++) {
1428 if (base->in(i) == base) {
1429 return NULL; // Wait stable graph
1430 }
1431 }
1432 }
1433 }
1434
1435 // Split through Phi (see original code in loopopts.cpp).
1436 assert(C->have_alias_type(t_oop), "instance should have alias type");
1437
1438 // Do nothing here if Identity will find a value
1439 // (to avoid infinite chain of value phis generation).
1440 if (!phase->eqv(this, this->Identity(phase))) {
1441 return NULL;
1442 }
1443
1444 // Select Region to split through.
1445 Node* region;
1446 if (!base_is_phi) {
1447 assert(mem->is_Phi(), "sanity");
1448 region = mem->in(0);
1449 // Skip if the region dominates some control edge of the address.
1450 if (!MemNode::all_controls_dominate(address, region))
1451 return NULL;
1452 } else if (!mem->is_Phi()) {
1453 assert(base_is_phi, "sanity");
1454 region = base->in(0);
1455 // Skip if the region dominates some control edge of the memory.
1456 if (!MemNode::all_controls_dominate(mem, region))
1457 return NULL;
1458 } else if (base->in(0) != mem->in(0)) {
1459 assert(base_is_phi && mem->is_Phi(), "sanity");
1460 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1461 region = base->in(0);
1462 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1463 region = mem->in(0);
1464 } else {
1465 return NULL; // complex graph
1466 }
1467 } else {
1468 assert(base->in(0) == mem->in(0), "sanity");
1469 region = mem->in(0);
1470 }
1471
1472 const Type* this_type = this->bottom_type();
1473 int this_index = C->get_alias_index(t_oop);
1474 int this_offset = t_oop->offset();
1475 int this_iid = t_oop->instance_id();
1476 if (!t_oop->is_known_instance() && load_boxed_values) {
1477 // Use _idx of address base for boxed values.
1478 this_iid = base->_idx;
1479 }
1480 PhaseIterGVN* igvn = phase->is_IterGVN();
1481 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1482 for (uint i = 1; i < region->req(); i++) {
1483 Node* x;
1484 Node* the_clone = NULL;
1485 Node* in = region->in(i);
1486 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1487 in != NULL && in->is_OuterStripMinedLoop()) {
1488 // No node should go in the outer strip mined loop
1489 in = in->in(LoopNode::EntryControl);
1490 }
1491 if (in == NULL || in == C->top()) {
1492 x = C->top(); // Dead path? Use a dead data op
1493 } else {
1494 x = this->clone(); // Else clone up the data op
1495 the_clone = x; // Remember for possible deletion.
1496 // Alter data node to use pre-phi inputs
1497 if (this->in(0) == region) {
1498 x->set_req(0, in);
1499 } else {
1500 x->set_req(0, NULL);
1501 }
1502 if (mem->is_Phi() && (mem->in(0) == region)) {
1503 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1504 }
1505 if (address->is_Phi() && address->in(0) == region) {
1506 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1507 }
1508 if (base_is_phi && (base->in(0) == region)) {
1509 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1510 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1511 x->set_req(Address, adr_x);
1512 }
1513 }
1514 // Check for a 'win' on some paths
1515 const Type *t = x->Value(igvn);
1516
1517 bool singleton = t->singleton();
1518
1519 // See comments in PhaseIdealLoop::split_thru_phi().
1520 if (singleton && t == Type::TOP) {
1521 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1522 }
1523
1524 if (singleton) {
1525 x = igvn->makecon(t);
1526 } else {
1527 // We now call Identity to try to simplify the cloned node.
1528 // Note that some Identity methods call phase->type(this).
1529 // Make sure that the type array is big enough for
1530 // our new node, even though we may throw the node away.
1531 // (This tweaking with igvn only works because x is a new node.)
1532 igvn->set_type(x, t);
1533 // If x is a TypeNode, capture any more-precise type permanently into Node
1534 // otherwise it will be not updated during igvn->transform since
1535 // igvn->type(x) is set to x->Value() already.
1536 x->raise_bottom_type(t);
1537 Node* y = x->Identity(igvn);
1538 if (y != x) {
1539 x = y;
1540 } else {
1541 y = igvn->hash_find_insert(x);
1542 if (y) {
1543 x = y;
1544 } else {
1545 // Else x is a new node we are keeping
1546 // We do not need register_new_node_with_optimizer
1547 // because set_type has already been called.
1548 igvn->_worklist.push(x);
1549 }
1550 }
1551 }
1552 if (x != the_clone && the_clone != NULL) {
1553 igvn->remove_dead_node(the_clone);
1554 }
1555 phi->set_req(i, x);
1556 }
1557 // Record Phi
1558 igvn->register_new_node_with_optimizer(phi);
1559 return phi;
1560 }
1561
1562 AllocateNode* LoadNode::is_new_object_mark_load(PhaseGVN *phase) const {
1563 if (Opcode() == Op_LoadX) {
1564 Node* address = in(MemNode::Address);
1565 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1566 Node* mem = in(MemNode::Memory);
1567 if (alloc != NULL && mem->is_Proj() &&
1568 mem->in(0) != NULL &&
1569 mem->in(0) == alloc->initialization() &&
1570 alloc->initialization()->proj_out_or_null(0) != NULL) {
1571 return alloc;
1572 }
1573 }
1574 return NULL;
1575 }
1576
1577
1578 //------------------------------Ideal------------------------------------------
1579 // If the load is from Field memory and the pointer is non-null, it might be possible to
1580 // zero out the control input.
1581 // If the offset is constant and the base is an object allocation,
1582 // try to hook me up to the exact initializing store.
1583 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1584 Node* p = MemNode::Ideal_common(phase, can_reshape);
1585 if (p) return (p == NodeSentinel) ? NULL : p;
1586
1587 Node* ctrl = in(MemNode::Control);
1588 Node* address = in(MemNode::Address);
1589 bool progress = false;
1590
1591 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1592 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1593
1594 // Skip up past a SafePoint control. Cannot do this for Stores because
1595 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1596 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1597 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1598 !addr_mark &&
1599 (depends_only_on_test() || has_unknown_control_dependency())) {
1600 ctrl = ctrl->in(0);
1601 set_req(MemNode::Control,ctrl);
1602 progress = true;
1603 }
1604
1605 intptr_t ignore = 0;
1606 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1607 if (base != NULL
1608 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1609 // Check for useless control edge in some common special cases
1610 if (in(MemNode::Control) != NULL
1611 && can_remove_control()
1612 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1613 && all_controls_dominate(base, phase->C->start())) {
1614 // A method-invariant, non-null address (constant or 'this' argument).
1615 set_req(MemNode::Control, NULL);
1616 progress = true;
1617 }
1618 }
1619
1620 Node* mem = in(MemNode::Memory);
1621 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1622
1623 if (can_reshape && (addr_t != NULL)) {
1624 // try to optimize our memory input
1625 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1626 if (opt_mem != mem) {
1627 set_req(MemNode::Memory, opt_mem);
1628 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1629 return this;
1630 }
1631 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1632 if ((t_oop != NULL) &&
1633 (t_oop->is_known_instance_field() ||
1634 t_oop->is_ptr_to_boxed_value())) {
1635 PhaseIterGVN *igvn = phase->is_IterGVN();
1636 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1637 // Delay this transformation until memory Phi is processed.
1638 phase->is_IterGVN()->_worklist.push(this);
1639 return NULL;
1640 }
1641 // Split instance field load through Phi.
1642 Node* result = split_through_phi(phase);
1643 if (result != NULL) return result;
1644
1645 if (t_oop->is_ptr_to_boxed_value()) {
1646 Node* result = eliminate_autobox(phase);
1647 if (result != NULL) return result;
1648 }
1649 }
1650 }
1651
1652 // Is there a dominating load that loads the same value? Leave
1653 // anything that is not a load of a field/array element (like
1654 // barriers etc.) alone
1655 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1656 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1657 Node *use = mem->fast_out(i);
1658 if (use != this &&
1659 use->Opcode() == Opcode() &&
1660 use->in(0) != NULL &&
1661 use->in(0) != in(0) &&
1662 use->in(Address) == in(Address)) {
1663 Node* ctl = in(0);
1664 for (int i = 0; i < 10 && ctl != NULL; i++) {
1665 ctl = IfNode::up_one_dom(ctl);
1666 if (ctl == use->in(0)) {
1667 set_req(0, use->in(0));
1668 return this;
1669 }
1670 }
1671 }
1672 }
1673 }
1674
1675 // Check for prior store with a different base or offset; make Load
1676 // independent. Skip through any number of them. Bail out if the stores
1677 // are in an endless dead cycle and report no progress. This is a key
1678 // transform for Reflection. However, if after skipping through the Stores
1679 // we can't then fold up against a prior store do NOT do the transform as
1680 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1681 // array memory alive twice: once for the hoisted Load and again after the
1682 // bypassed Store. This situation only works if EVERYBODY who does
1683 // anti-dependence work knows how to bypass. I.e. we need all
1684 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1685 // the alias index stuff. So instead, peek through Stores and IFF we can
1686 // fold up, do so.
1687 Node* prev_mem = find_previous_store(phase);
1688 if (prev_mem != NULL) {
1689 Node* value = can_see_arraycopy_value(prev_mem, phase);
1690 if (value != NULL) {
1691 return value;
1692 }
1693 }
1694 // Steps (a), (b): Walk past independent stores to find an exact match.
1695 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1696 // (c) See if we can fold up on the spot, but don't fold up here.
1697 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1698 // just return a prior value, which is done by Identity calls.
1699 if (can_see_stored_value(prev_mem, phase)) {
1700 // Make ready for step (d):
1701 set_req(MemNode::Memory, prev_mem);
1702 return this;
1703 }
1704 }
1705
1706 AllocateNode* alloc = is_new_object_mark_load(phase);
1707 if (alloc != NULL && alloc->Opcode() == Op_Allocate && UseBiasedLocking) {
1708 InitializeNode* init = alloc->initialization();
1709 Node* control = init->proj_out(0);
1710 return alloc->make_ideal_mark(phase, address, control, mem);
1711 }
1712
1713 return progress ? this : NULL;
1714 }
1715
1716 // Helper to recognize certain Klass fields which are invariant across
1717 // some group of array types (e.g., int[] or all T[] where T < Object).
1718 const Type*
1719 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1720 ciKlass* klass) const {
1721 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1722 // The field is Klass::_modifier_flags. Return its (constant) value.
1723 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1724 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1725 return TypeInt::make(klass->modifier_flags());
1726 }
1727 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1728 // The field is Klass::_access_flags. Return its (constant) value.
1729 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1730 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1731 return TypeInt::make(klass->access_flags());
1732 }
1733 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1734 // The field is Klass::_layout_helper. Return its constant value if known.
1735 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1736 return TypeInt::make(klass->layout_helper());
1737 }
1738
1739 // No match.
1740 return NULL;
1741 }
1742
1743 //------------------------------Value-----------------------------------------
1744 const Type* LoadNode::Value(PhaseGVN* phase) const {
1745 // Either input is TOP ==> the result is TOP
1746 Node* mem = in(MemNode::Memory);
1747 const Type *t1 = phase->type(mem);
1748 if (t1 == Type::TOP) return Type::TOP;
1749 Node* adr = in(MemNode::Address);
1750 const TypePtr* tp = phase->type(adr)->isa_ptr();
1751 if (tp == NULL || tp->empty()) return Type::TOP;
1752 int off = tp->offset();
1753 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1754 Compile* C = phase->C;
1755
1756 // Try to guess loaded type from pointer type
1757 if (tp->isa_aryptr()) {
1758 const TypeAryPtr* ary = tp->is_aryptr();
1759 const Type* t = ary->elem();
1760
1761 // Determine whether the reference is beyond the header or not, by comparing
1762 // the offset against the offset of the start of the array's data.
1763 // Different array types begin at slightly different offsets (12 vs. 16).
1764 // We choose T_BYTE as an example base type that is least restrictive
1765 // as to alignment, which will therefore produce the smallest
1766 // possible base offset.
1767 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1768 const bool off_beyond_header = (off >= min_base_off);
1769
1770 // Try to constant-fold a stable array element.
1771 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1772 // Make sure the reference is not into the header and the offset is constant
1773 ciObject* aobj = ary->const_oop();
1774 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1775 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1776 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1777 stable_dimension,
1778 memory_type(), is_unsigned());
1779 if (con_type != NULL) {
1780 return con_type;
1781 }
1782 }
1783 }
1784
1785 // Don't do this for integer types. There is only potential profit if
1786 // the element type t is lower than _type; that is, for int types, if _type is
1787 // more restrictive than t. This only happens here if one is short and the other
1788 // char (both 16 bits), and in those cases we've made an intentional decision
1789 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1790 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1791 //
1792 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1793 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1794 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1795 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1796 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1797 // In fact, that could have been the original type of p1, and p1 could have
1798 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1799 // expression (LShiftL quux 3) independently optimized to the constant 8.
1800 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1801 && (_type->isa_vect() == NULL)
1802 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1803 // t might actually be lower than _type, if _type is a unique
1804 // concrete subclass of abstract class t.
1805 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
1806 const Type* jt = t->join_speculative(_type);
1807 // In any case, do not allow the join, per se, to empty out the type.
1808 if (jt->empty() && !t->empty()) {
1809 // This can happen if a interface-typed array narrows to a class type.
1810 jt = _type;
1811 }
1812 #ifdef ASSERT
1813 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1814 // The pointers in the autobox arrays are always non-null
1815 Node* base = adr->in(AddPNode::Base);
1816 if ((base != NULL) && base->is_DecodeN()) {
1817 // Get LoadN node which loads IntegerCache.cache field
1818 base = base->in(1);
1819 }
1820 if ((base != NULL) && base->is_Con()) {
1821 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1822 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1823 // It could be narrow oop
1824 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1825 }
1826 }
1827 }
1828 #endif
1829 return jt;
1830 }
1831 }
1832 } else if (tp->base() == Type::InstPtr) {
1833 assert( off != Type::OffsetBot ||
1834 // arrays can be cast to Objects
1835 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1836 // unsafe field access may not have a constant offset
1837 C->has_unsafe_access(),
1838 "Field accesses must be precise" );
1839 // For oop loads, we expect the _type to be precise.
1840
1841 // Optimize loads from constant fields.
1842 const TypeInstPtr* tinst = tp->is_instptr();
1843 ciObject* const_oop = tinst->const_oop();
1844 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1845 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1846 if (con_type != NULL) {
1847 return con_type;
1848 }
1849 }
1850 } else if (tp->base() == Type::KlassPtr) {
1851 assert( off != Type::OffsetBot ||
1852 // arrays can be cast to Objects
1853 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1854 // also allow array-loading from the primary supertype
1855 // array during subtype checks
1856 Opcode() == Op_LoadKlass,
1857 "Field accesses must be precise" );
1858 // For klass/static loads, we expect the _type to be precise
1859 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
1860 /* With mirrors being an indirect in the Klass*
1861 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1862 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1863 *
1864 * So check the type and klass of the node before the LoadP.
1865 */
1866 Node* adr2 = adr->in(MemNode::Address);
1867 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1868 if (tkls != NULL && !StressReflectiveCode) {
1869 ciKlass* klass = tkls->klass();
1870 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1871 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1872 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1873 return TypeInstPtr::make(klass->java_mirror());
1874 }
1875 }
1876 }
1877
1878 const TypeKlassPtr *tkls = tp->isa_klassptr();
1879 if (tkls != NULL && !StressReflectiveCode) {
1880 ciKlass* klass = tkls->klass();
1881 if (klass->is_loaded() && tkls->klass_is_exact()) {
1882 // We are loading a field from a Klass metaobject whose identity
1883 // is known at compile time (the type is "exact" or "precise").
1884 // Check for fields we know are maintained as constants by the VM.
1885 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1886 // The field is Klass::_super_check_offset. Return its (constant) value.
1887 // (Folds up type checking code.)
1888 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1889 return TypeInt::make(klass->super_check_offset());
1890 }
1891 // Compute index into primary_supers array
1892 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1893 // Check for overflowing; use unsigned compare to handle the negative case.
1894 if( depth < ciKlass::primary_super_limit() ) {
1895 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1896 // (Folds up type checking code.)
1897 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1898 ciKlass *ss = klass->super_of_depth(depth);
1899 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1900 }
1901 const Type* aift = load_array_final_field(tkls, klass);
1902 if (aift != NULL) return aift;
1903 }
1904
1905 // We can still check if we are loading from the primary_supers array at a
1906 // shallow enough depth. Even though the klass is not exact, entries less
1907 // than or equal to its super depth are correct.
1908 if (klass->is_loaded() ) {
1909 ciType *inner = klass;
1910 while( inner->is_obj_array_klass() )
1911 inner = inner->as_obj_array_klass()->base_element_type();
1912 if( inner->is_instance_klass() &&
1913 !inner->as_instance_klass()->flags().is_interface() ) {
1914 // Compute index into primary_supers array
1915 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1916 // Check for overflowing; use unsigned compare to handle the negative case.
1917 if( depth < ciKlass::primary_super_limit() &&
1918 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1919 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1920 // (Folds up type checking code.)
1921 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1922 ciKlass *ss = klass->super_of_depth(depth);
1923 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1924 }
1925 }
1926 }
1927
1928 // If the type is enough to determine that the thing is not an array,
1929 // we can give the layout_helper a positive interval type.
1930 // This will help short-circuit some reflective code.
1931 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1932 && !klass->is_array_klass() // not directly typed as an array
1933 && !klass->is_interface() // specifically not Serializable & Cloneable
1934 && !klass->is_java_lang_Object() // not the supertype of all T[]
1935 ) {
1936 // Note: When interfaces are reliable, we can narrow the interface
1937 // test to (klass != Serializable && klass != Cloneable).
1938 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1939 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1940 // The key property of this type is that it folds up tests
1941 // for array-ness, since it proves that the layout_helper is positive.
1942 // Thus, a generic value like the basic object layout helper works fine.
1943 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1944 }
1945 }
1946
1947 // If we are loading from a freshly-allocated object, produce a zero,
1948 // if the load is provably beyond the header of the object.
1949 // (Also allow a variable load from a fresh array to produce zero.)
1950 const TypeOopPtr *tinst = tp->isa_oopptr();
1951 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1952 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1953 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1954 Node* value = can_see_stored_value(mem,phase);
1955 if (value != NULL && value->is_Con()) {
1956 assert(value->bottom_type()->higher_equal(_type),"sanity");
1957 return value->bottom_type();
1958 }
1959 }
1960
1961 if (is_instance) {
1962 // If we have an instance type and our memory input is the
1963 // programs's initial memory state, there is no matching store,
1964 // so just return a zero of the appropriate type
1965 Node *mem = in(MemNode::Memory);
1966 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1967 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1968 return Type::get_zero_type(_type->basic_type());
1969 }
1970 }
1971
1972 Node* alloc = is_new_object_mark_load(phase);
1973 if (alloc != NULL && !(alloc->Opcode() == Op_Allocate && UseBiasedLocking)) {
1974 return TypeX::make(markWord::prototype().value());
1975 }
1976
1977 return _type;
1978 }
1979
1980 //------------------------------match_edge-------------------------------------
1981 // Do we Match on this edge index or not? Match only the address.
1982 uint LoadNode::match_edge(uint idx) const {
1983 return idx == MemNode::Address;
1984 }
1985
1986 //--------------------------LoadBNode::Ideal--------------------------------------
1987 //
1988 // If the previous store is to the same address as this load,
1989 // and the value stored was larger than a byte, replace this load
1990 // with the value stored truncated to a byte. If no truncation is
1991 // needed, the replacement is done in LoadNode::Identity().
1992 //
1993 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1994 Node* mem = in(MemNode::Memory);
1995 Node* value = can_see_stored_value(mem,phase);
1996 if( value && !phase->type(value)->higher_equal( _type ) ) {
1997 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1998 return new RShiftINode(result, phase->intcon(24));
1999 }
2000 // Identity call will handle the case where truncation is not needed.
2001 return LoadNode::Ideal(phase, can_reshape);
2002 }
2003
2004 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2005 Node* mem = in(MemNode::Memory);
2006 Node* value = can_see_stored_value(mem,phase);
2007 if (value != NULL && value->is_Con() &&
2008 !value->bottom_type()->higher_equal(_type)) {
2009 // If the input to the store does not fit with the load's result type,
2010 // it must be truncated. We can't delay until Ideal call since
2011 // a singleton Value is needed for split_thru_phi optimization.
2012 int con = value->get_int();
2013 return TypeInt::make((con << 24) >> 24);
2014 }
2015 return LoadNode::Value(phase);
2016 }
2017
2018 //--------------------------LoadUBNode::Ideal-------------------------------------
2019 //
2020 // If the previous store is to the same address as this load,
2021 // and the value stored was larger than a byte, replace this load
2022 // with the value stored truncated to a byte. If no truncation is
2023 // needed, the replacement is done in LoadNode::Identity().
2024 //
2025 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2026 Node* mem = in(MemNode::Memory);
2027 Node* value = can_see_stored_value(mem, phase);
2028 if (value && !phase->type(value)->higher_equal(_type))
2029 return new AndINode(value, phase->intcon(0xFF));
2030 // Identity call will handle the case where truncation is not needed.
2031 return LoadNode::Ideal(phase, can_reshape);
2032 }
2033
2034 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2035 Node* mem = in(MemNode::Memory);
2036 Node* value = can_see_stored_value(mem,phase);
2037 if (value != NULL && value->is_Con() &&
2038 !value->bottom_type()->higher_equal(_type)) {
2039 // If the input to the store does not fit with the load's result type,
2040 // it must be truncated. We can't delay until Ideal call since
2041 // a singleton Value is needed for split_thru_phi optimization.
2042 int con = value->get_int();
2043 return TypeInt::make(con & 0xFF);
2044 }
2045 return LoadNode::Value(phase);
2046 }
2047
2048 //--------------------------LoadUSNode::Ideal-------------------------------------
2049 //
2050 // If the previous store is to the same address as this load,
2051 // and the value stored was larger than a char, replace this load
2052 // with the value stored truncated to a char. If no truncation is
2053 // needed, the replacement is done in LoadNode::Identity().
2054 //
2055 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2056 Node* mem = in(MemNode::Memory);
2057 Node* value = can_see_stored_value(mem,phase);
2058 if( value && !phase->type(value)->higher_equal( _type ) )
2059 return new AndINode(value,phase->intcon(0xFFFF));
2060 // Identity call will handle the case where truncation is not needed.
2061 return LoadNode::Ideal(phase, can_reshape);
2062 }
2063
2064 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2065 Node* mem = in(MemNode::Memory);
2066 Node* value = can_see_stored_value(mem,phase);
2067 if (value != NULL && value->is_Con() &&
2068 !value->bottom_type()->higher_equal(_type)) {
2069 // If the input to the store does not fit with the load's result type,
2070 // it must be truncated. We can't delay until Ideal call since
2071 // a singleton Value is needed for split_thru_phi optimization.
2072 int con = value->get_int();
2073 return TypeInt::make(con & 0xFFFF);
2074 }
2075 return LoadNode::Value(phase);
2076 }
2077
2078 //--------------------------LoadSNode::Ideal--------------------------------------
2079 //
2080 // If the previous store is to the same address as this load,
2081 // and the value stored was larger than a short, replace this load
2082 // with the value stored truncated to a short. If no truncation is
2083 // needed, the replacement is done in LoadNode::Identity().
2084 //
2085 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2086 Node* mem = in(MemNode::Memory);
2087 Node* value = can_see_stored_value(mem,phase);
2088 if( value && !phase->type(value)->higher_equal( _type ) ) {
2089 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2090 return new RShiftINode(result, phase->intcon(16));
2091 }
2092 // Identity call will handle the case where truncation is not needed.
2093 return LoadNode::Ideal(phase, can_reshape);
2094 }
2095
2096 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2097 Node* mem = in(MemNode::Memory);
2098 Node* value = can_see_stored_value(mem,phase);
2099 if (value != NULL && value->is_Con() &&
2100 !value->bottom_type()->higher_equal(_type)) {
2101 // If the input to the store does not fit with the load's result type,
2102 // it must be truncated. We can't delay until Ideal call since
2103 // a singleton Value is needed for split_thru_phi optimization.
2104 int con = value->get_int();
2105 return TypeInt::make((con << 16) >> 16);
2106 }
2107 return LoadNode::Value(phase);
2108 }
2109
2110 //=============================================================================
2111 //----------------------------LoadKlassNode::make------------------------------
2112 // Polymorphic factory method:
2113 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2114 // sanity check the alias category against the created node type
2115 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2116 assert(adr_type != NULL, "expecting TypeKlassPtr");
2117 #ifdef _LP64
2118 if (adr_type->is_ptr_to_narrowklass()) {
2119 assert(UseCompressedClassPointers, "no compressed klasses");
2120 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2121 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2122 }
2123 #endif
2124 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2125 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2126 }
2127
2128 //------------------------------Value------------------------------------------
2129 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2130 return klass_value_common(phase);
2131 }
2132
2133 // In most cases, LoadKlassNode does not have the control input set. If the control
2134 // input is set, it must not be removed (by LoadNode::Ideal()).
2135 bool LoadKlassNode::can_remove_control() const {
2136 return false;
2137 }
2138
2139 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2140 // Either input is TOP ==> the result is TOP
2141 const Type *t1 = phase->type( in(MemNode::Memory) );
2142 if (t1 == Type::TOP) return Type::TOP;
2143 Node *adr = in(MemNode::Address);
2144 const Type *t2 = phase->type( adr );
2145 if (t2 == Type::TOP) return Type::TOP;
2146 const TypePtr *tp = t2->is_ptr();
2147 if (TypePtr::above_centerline(tp->ptr()) ||
2148 tp->ptr() == TypePtr::Null) return Type::TOP;
2149
2150 // Return a more precise klass, if possible
2151 const TypeInstPtr *tinst = tp->isa_instptr();
2152 if (tinst != NULL) {
2153 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2154 int offset = tinst->offset();
2155 if (ik == phase->C->env()->Class_klass()
2156 && (offset == java_lang_Class::klass_offset() ||
2157 offset == java_lang_Class::array_klass_offset())) {
2158 // We are loading a special hidden field from a Class mirror object,
2159 // the field which points to the VM's Klass metaobject.
2160 ciType* t = tinst->java_mirror_type();
2161 // java_mirror_type returns non-null for compile-time Class constants.
2162 if (t != NULL) {
2163 // constant oop => constant klass
2164 if (offset == java_lang_Class::array_klass_offset()) {
2165 if (t->is_void()) {
2166 // We cannot create a void array. Since void is a primitive type return null
2167 // klass. Users of this result need to do a null check on the returned klass.
2168 return TypePtr::NULL_PTR;
2169 }
2170 return TypeKlassPtr::make(ciArrayKlass::make(t));
2171 }
2172 if (!t->is_klass()) {
2173 // a primitive Class (e.g., int.class) has NULL for a klass field
2174 return TypePtr::NULL_PTR;
2175 }
2176 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2177 return TypeKlassPtr::make(t->as_klass());
2178 }
2179 // non-constant mirror, so we can't tell what's going on
2180 }
2181 if( !ik->is_loaded() )
2182 return _type; // Bail out if not loaded
2183 if (offset == oopDesc::klass_offset_in_bytes()) {
2184 if (tinst->klass_is_exact()) {
2185 return TypeKlassPtr::make(ik);
2186 }
2187 // See if we can become precise: no subklasses and no interface
2188 // (Note: We need to support verified interfaces.)
2189 if (!ik->is_interface() && !ik->has_subklass()) {
2190 // Add a dependence; if any subclass added we need to recompile
2191 if (!ik->is_final()) {
2192 // %%% should use stronger assert_unique_concrete_subtype instead
2193 phase->C->dependencies()->assert_leaf_type(ik);
2194 }
2195 // Return precise klass
2196 return TypeKlassPtr::make(ik);
2197 }
2198
2199 // Return root of possible klass
2200 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2201 }
2202 }
2203
2204 // Check for loading klass from an array
2205 const TypeAryPtr *tary = tp->isa_aryptr();
2206 if( tary != NULL ) {
2207 ciKlass *tary_klass = tary->klass();
2208 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2209 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2210 if (tary->klass_is_exact()) {
2211 return TypeKlassPtr::make(tary_klass);
2212 }
2213 ciArrayKlass *ak = tary->klass()->as_array_klass();
2214 // If the klass is an object array, we defer the question to the
2215 // array component klass.
2216 if( ak->is_obj_array_klass() ) {
2217 assert( ak->is_loaded(), "" );
2218 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2219 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2220 ciInstanceKlass* ik = base_k->as_instance_klass();
2221 // See if we can become precise: no subklasses and no interface
2222 if (!ik->is_interface() && !ik->has_subklass()) {
2223 // Add a dependence; if any subclass added we need to recompile
2224 if (!ik->is_final()) {
2225 phase->C->dependencies()->assert_leaf_type(ik);
2226 }
2227 // Return precise array klass
2228 return TypeKlassPtr::make(ak);
2229 }
2230 }
2231 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2232 } else { // Found a type-array?
2233 assert( ak->is_type_array_klass(), "" );
2234 return TypeKlassPtr::make(ak); // These are always precise
2235 }
2236 }
2237 }
2238
2239 // Check for loading klass from an array klass
2240 const TypeKlassPtr *tkls = tp->isa_klassptr();
2241 if (tkls != NULL && !StressReflectiveCode) {
2242 ciKlass* klass = tkls->klass();
2243 if( !klass->is_loaded() )
2244 return _type; // Bail out if not loaded
2245 if( klass->is_obj_array_klass() &&
2246 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2247 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2248 // // Always returning precise element type is incorrect,
2249 // // e.g., element type could be object and array may contain strings
2250 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2251
2252 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2253 // according to the element type's subclassing.
2254 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2255 }
2256 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2257 tkls->offset() == in_bytes(Klass::super_offset())) {
2258 ciKlass* sup = klass->as_instance_klass()->super();
2259 // The field is Klass::_super. Return its (constant) value.
2260 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2261 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2262 }
2263 }
2264
2265 // Bailout case
2266 return LoadNode::Value(phase);
2267 }
2268
2269 //------------------------------Identity---------------------------------------
2270 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2271 // Also feed through the klass in Allocate(...klass...)._klass.
2272 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2273 return klass_identity_common(phase);
2274 }
2275
2276 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2277 Node* x = LoadNode::Identity(phase);
2278 if (x != this) return x;
2279
2280 // Take apart the address into an oop and and offset.
2281 // Return 'this' if we cannot.
2282 Node* adr = in(MemNode::Address);
2283 intptr_t offset = 0;
2284 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2285 if (base == NULL) return this;
2286 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2287 if (toop == NULL) return this;
2288
2289 // Step over potential GC barrier for OopHandle resolve
2290 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2291 if (bs->is_gc_barrier_node(base)) {
2292 base = bs->step_over_gc_barrier(base);
2293 }
2294
2295 // We can fetch the klass directly through an AllocateNode.
2296 // This works even if the klass is not constant (clone or newArray).
2297 if (offset == oopDesc::klass_offset_in_bytes()) {
2298 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2299 if (allocated_klass != NULL) {
2300 return allocated_klass;
2301 }
2302 }
2303
2304 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2305 // See inline_native_Class_query for occurrences of these patterns.
2306 // Java Example: x.getClass().isAssignableFrom(y)
2307 //
2308 // This improves reflective code, often making the Class
2309 // mirror go completely dead. (Current exception: Class
2310 // mirrors may appear in debug info, but we could clean them out by
2311 // introducing a new debug info operator for Klass.java_mirror).
2312
2313 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2314 && offset == java_lang_Class::klass_offset()) {
2315 if (base->is_Load()) {
2316 Node* base2 = base->in(MemNode::Address);
2317 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2318 Node* adr2 = base2->in(MemNode::Address);
2319 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2320 if (tkls != NULL && !tkls->empty()
2321 && (tkls->klass()->is_instance_klass() ||
2322 tkls->klass()->is_array_klass())
2323 && adr2->is_AddP()
2324 ) {
2325 int mirror_field = in_bytes(Klass::java_mirror_offset());
2326 if (tkls->offset() == mirror_field) {
2327 return adr2->in(AddPNode::Base);
2328 }
2329 }
2330 }
2331 }
2332 }
2333
2334 return this;
2335 }
2336
2337
2338 //------------------------------Value------------------------------------------
2339 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2340 const Type *t = klass_value_common(phase);
2341 if (t == Type::TOP)
2342 return t;
2343
2344 return t->make_narrowklass();
2345 }
2346
2347 //------------------------------Identity---------------------------------------
2348 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2349 // Also feed through the klass in Allocate(...klass...)._klass.
2350 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2351 Node *x = klass_identity_common(phase);
2352
2353 const Type *t = phase->type( x );
2354 if( t == Type::TOP ) return x;
2355 if( t->isa_narrowklass()) return x;
2356 assert (!t->isa_narrowoop(), "no narrow oop here");
2357
2358 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2359 }
2360
2361 //------------------------------Value-----------------------------------------
2362 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2363 // Either input is TOP ==> the result is TOP
2364 const Type *t1 = phase->type( in(MemNode::Memory) );
2365 if( t1 == Type::TOP ) return Type::TOP;
2366 Node *adr = in(MemNode::Address);
2367 const Type *t2 = phase->type( adr );
2368 if( t2 == Type::TOP ) return Type::TOP;
2369 const TypePtr *tp = t2->is_ptr();
2370 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2371 const TypeAryPtr *tap = tp->isa_aryptr();
2372 if( !tap ) return _type;
2373 return tap->size();
2374 }
2375
2376 //-------------------------------Ideal---------------------------------------
2377 // Feed through the length in AllocateArray(...length...)._length.
2378 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2379 Node* p = MemNode::Ideal_common(phase, can_reshape);
2380 if (p) return (p == NodeSentinel) ? NULL : p;
2381
2382 // Take apart the address into an oop and and offset.
2383 // Return 'this' if we cannot.
2384 Node* adr = in(MemNode::Address);
2385 intptr_t offset = 0;
2386 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2387 if (base == NULL) return NULL;
2388 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2389 if (tary == NULL) return NULL;
2390
2391 // We can fetch the length directly through an AllocateArrayNode.
2392 // This works even if the length is not constant (clone or newArray).
2393 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2394 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2395 if (alloc != NULL) {
2396 Node* allocated_length = alloc->Ideal_length();
2397 Node* len = alloc->make_ideal_length(tary, phase);
2398 if (allocated_length != len) {
2399 // New CastII improves on this.
2400 return len;
2401 }
2402 }
2403 }
2404
2405 return NULL;
2406 }
2407
2408 //------------------------------Identity---------------------------------------
2409 // Feed through the length in AllocateArray(...length...)._length.
2410 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2411 Node* x = LoadINode::Identity(phase);
2412 if (x != this) return x;
2413
2414 // Take apart the address into an oop and and offset.
2415 // Return 'this' if we cannot.
2416 Node* adr = in(MemNode::Address);
2417 intptr_t offset = 0;
2418 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2419 if (base == NULL) return this;
2420 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2421 if (tary == NULL) return this;
2422
2423 // We can fetch the length directly through an AllocateArrayNode.
2424 // This works even if the length is not constant (clone or newArray).
2425 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2426 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2427 if (alloc != NULL) {
2428 Node* allocated_length = alloc->Ideal_length();
2429 // Do not allow make_ideal_length to allocate a CastII node.
2430 Node* len = alloc->make_ideal_length(tary, phase, false);
2431 if (allocated_length == len) {
2432 // Return allocated_length only if it would not be improved by a CastII.
2433 return allocated_length;
2434 }
2435 }
2436 }
2437
2438 return this;
2439
2440 }
2441
2442 //=============================================================================
2443 //---------------------------StoreNode::make-----------------------------------
2444 // Polymorphic factory method:
2445 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2446 assert((mo == unordered || mo == release), "unexpected");
2447 Compile* C = gvn.C;
2448 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2449 ctl != NULL, "raw memory operations should have control edge");
2450
2451 switch (bt) {
2452 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2453 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2454 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2455 case T_CHAR:
2456 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2457 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2458 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2459 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2460 case T_METADATA:
2461 case T_ADDRESS:
2462 case T_OBJECT:
2463 #ifdef _LP64
2464 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2465 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2466 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2467 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2468 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2469 adr->bottom_type()->isa_rawptr())) {
2470 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2471 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2472 }
2473 #endif
2474 {
2475 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2476 }
2477 default:
2478 ShouldNotReachHere();
2479 return (StoreNode*)NULL;
2480 }
2481 }
2482
2483 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2484 bool require_atomic = true;
2485 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2486 }
2487
2488 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2489 bool require_atomic = true;
2490 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2491 }
2492
2493
2494 //--------------------------bottom_type----------------------------------------
2495 const Type *StoreNode::bottom_type() const {
2496 return Type::MEMORY;
2497 }
2498
2499 //------------------------------hash-------------------------------------------
2500 uint StoreNode::hash() const {
2501 // unroll addition of interesting fields
2502 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2503
2504 // Since they are not commoned, do not hash them:
2505 return NO_HASH;
2506 }
2507
2508 //------------------------------Ideal------------------------------------------
2509 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2510 // When a store immediately follows a relevant allocation/initialization,
2511 // try to capture it into the initialization, or hoist it above.
2512 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2513 Node* p = MemNode::Ideal_common(phase, can_reshape);
2514 if (p) return (p == NodeSentinel) ? NULL : p;
2515
2516 Node* mem = in(MemNode::Memory);
2517 Node* address = in(MemNode::Address);
2518 // Back-to-back stores to same address? Fold em up. Generally
2519 // unsafe if I have intervening uses... Also disallowed for StoreCM
2520 // since they must follow each StoreP operation. Redundant StoreCMs
2521 // are eliminated just before matching in final_graph_reshape.
2522 {
2523 Node* st = mem;
2524 // If Store 'st' has more than one use, we cannot fold 'st' away.
2525 // For example, 'st' might be the final state at a conditional
2526 // return. Or, 'st' might be used by some node which is live at
2527 // the same time 'st' is live, which might be unschedulable. So,
2528 // require exactly ONE user until such time as we clone 'mem' for
2529 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2530 // true).
2531 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2532 // Looking at a dead closed cycle of memory?
2533 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2534 assert(Opcode() == st->Opcode() ||
2535 st->Opcode() == Op_StoreVector ||
2536 Opcode() == Op_StoreVector ||
2537 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2538 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2539 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2540 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2541 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2542
2543 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2544 st->as_Store()->memory_size() <= this->memory_size()) {
2545 Node* use = st->raw_out(0);
2546 phase->igvn_rehash_node_delayed(use);
2547 if (can_reshape) {
2548 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2549 } else {
2550 // It's OK to do this in the parser, since DU info is always accurate,
2551 // and the parser always refers to nodes via SafePointNode maps.
2552 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2553 }
2554 return this;
2555 }
2556 st = st->in(MemNode::Memory);
2557 }
2558 }
2559
2560
2561 // Capture an unaliased, unconditional, simple store into an initializer.
2562 // Or, if it is independent of the allocation, hoist it above the allocation.
2563 if (ReduceFieldZeroing && /*can_reshape &&*/
2564 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2565 InitializeNode* init = mem->in(0)->as_Initialize();
2566 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2567 if (offset > 0) {
2568 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2569 // If the InitializeNode captured me, it made a raw copy of me,
2570 // and I need to disappear.
2571 if (moved != NULL) {
2572 // %%% hack to ensure that Ideal returns a new node:
2573 mem = MergeMemNode::make(mem);
2574 return mem; // fold me away
2575 }
2576 }
2577 }
2578
2579 return NULL; // No further progress
2580 }
2581
2582 //------------------------------Value-----------------------------------------
2583 const Type* StoreNode::Value(PhaseGVN* phase) const {
2584 // Either input is TOP ==> the result is TOP
2585 const Type *t1 = phase->type( in(MemNode::Memory) );
2586 if( t1 == Type::TOP ) return Type::TOP;
2587 const Type *t2 = phase->type( in(MemNode::Address) );
2588 if( t2 == Type::TOP ) return Type::TOP;
2589 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2590 if( t3 == Type::TOP ) return Type::TOP;
2591 return Type::MEMORY;
2592 }
2593
2594 //------------------------------Identity---------------------------------------
2595 // Remove redundant stores:
2596 // Store(m, p, Load(m, p)) changes to m.
2597 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2598 Node* StoreNode::Identity(PhaseGVN* phase) {
2599 Node* mem = in(MemNode::Memory);
2600 Node* adr = in(MemNode::Address);
2601 Node* val = in(MemNode::ValueIn);
2602
2603 Node* result = this;
2604
2605 // Load then Store? Then the Store is useless
2606 if (val->is_Load() &&
2607 val->in(MemNode::Address)->eqv_uncast(adr) &&
2608 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2609 val->as_Load()->store_Opcode() == Opcode()) {
2610 result = mem;
2611 }
2612
2613 // Two stores in a row of the same value?
2614 if (result == this &&
2615 mem->is_Store() &&
2616 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2617 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2618 mem->Opcode() == Opcode()) {
2619 result = mem;
2620 }
2621
2622 // Store of zero anywhere into a freshly-allocated object?
2623 // Then the store is useless.
2624 // (It must already have been captured by the InitializeNode.)
2625 if (result == this &&
2626 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2627 // a newly allocated object is already all-zeroes everywhere
2628 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2629 result = mem;
2630 }
2631
2632 if (result == this) {
2633 // the store may also apply to zero-bits in an earlier object
2634 Node* prev_mem = find_previous_store(phase);
2635 // Steps (a), (b): Walk past independent stores to find an exact match.
2636 if (prev_mem != NULL) {
2637 Node* prev_val = can_see_stored_value(prev_mem, phase);
2638 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2639 // prev_val and val might differ by a cast; it would be good
2640 // to keep the more informative of the two.
2641 result = mem;
2642 }
2643 }
2644 }
2645 }
2646
2647 if (result != this && phase->is_IterGVN() != NULL) {
2648 MemBarNode* trailing = trailing_membar();
2649 if (trailing != NULL) {
2650 #ifdef ASSERT
2651 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2652 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2653 #endif
2654 PhaseIterGVN* igvn = phase->is_IterGVN();
2655 trailing->remove(igvn);
2656 }
2657 }
2658
2659 return result;
2660 }
2661
2662 //------------------------------match_edge-------------------------------------
2663 // Do we Match on this edge index or not? Match only memory & value
2664 uint StoreNode::match_edge(uint idx) const {
2665 return idx == MemNode::Address || idx == MemNode::ValueIn;
2666 }
2667
2668 //------------------------------cmp--------------------------------------------
2669 // Do not common stores up together. They generally have to be split
2670 // back up anyways, so do not bother.
2671 bool StoreNode::cmp( const Node &n ) const {
2672 return (&n == this); // Always fail except on self
2673 }
2674
2675 //------------------------------Ideal_masked_input-----------------------------
2676 // Check for a useless mask before a partial-word store
2677 // (StoreB ... (AndI valIn conIa) )
2678 // If (conIa & mask == mask) this simplifies to
2679 // (StoreB ... (valIn) )
2680 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2681 Node *val = in(MemNode::ValueIn);
2682 if( val->Opcode() == Op_AndI ) {
2683 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2684 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2685 set_req(MemNode::ValueIn, val->in(1));
2686 return this;
2687 }
2688 }
2689 return NULL;
2690 }
2691
2692
2693 //------------------------------Ideal_sign_extended_input----------------------
2694 // Check for useless sign-extension before a partial-word store
2695 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2696 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2697 // (StoreB ... (valIn) )
2698 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2699 Node *val = in(MemNode::ValueIn);
2700 if( val->Opcode() == Op_RShiftI ) {
2701 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2702 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2703 Node *shl = val->in(1);
2704 if( shl->Opcode() == Op_LShiftI ) {
2705 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2706 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2707 set_req(MemNode::ValueIn, shl->in(1));
2708 return this;
2709 }
2710 }
2711 }
2712 }
2713 return NULL;
2714 }
2715
2716 //------------------------------value_never_loaded-----------------------------------
2717 // Determine whether there are any possible loads of the value stored.
2718 // For simplicity, we actually check if there are any loads from the
2719 // address stored to, not just for loads of the value stored by this node.
2720 //
2721 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2722 Node *adr = in(Address);
2723 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2724 if (adr_oop == NULL)
2725 return false;
2726 if (!adr_oop->is_known_instance_field())
2727 return false; // if not a distinct instance, there may be aliases of the address
2728 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2729 Node *use = adr->fast_out(i);
2730 if (use->is_Load() || use->is_LoadStore()) {
2731 return false;
2732 }
2733 }
2734 return true;
2735 }
2736
2737 MemBarNode* StoreNode::trailing_membar() const {
2738 if (is_release()) {
2739 MemBarNode* trailing_mb = NULL;
2740 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2741 Node* u = fast_out(i);
2742 if (u->is_MemBar()) {
2743 if (u->as_MemBar()->trailing_store()) {
2744 assert(u->Opcode() == Op_MemBarVolatile, "");
2745 assert(trailing_mb == NULL, "only one");
2746 trailing_mb = u->as_MemBar();
2747 #ifdef ASSERT
2748 Node* leading = u->as_MemBar()->leading_membar();
2749 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2750 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2751 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2752 #endif
2753 } else {
2754 assert(u->as_MemBar()->standalone(), "");
2755 }
2756 }
2757 }
2758 return trailing_mb;
2759 }
2760 return NULL;
2761 }
2762
2763
2764 //=============================================================================
2765 //------------------------------Ideal------------------------------------------
2766 // If the store is from an AND mask that leaves the low bits untouched, then
2767 // we can skip the AND operation. If the store is from a sign-extension
2768 // (a left shift, then right shift) we can skip both.
2769 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2770 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2771 if( progress != NULL ) return progress;
2772
2773 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2774 if( progress != NULL ) return progress;
2775
2776 // Finally check the default case
2777 return StoreNode::Ideal(phase, can_reshape);
2778 }
2779
2780 //=============================================================================
2781 //------------------------------Ideal------------------------------------------
2782 // If the store is from an AND mask that leaves the low bits untouched, then
2783 // we can skip the AND operation
2784 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2785 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2786 if( progress != NULL ) return progress;
2787
2788 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2789 if( progress != NULL ) return progress;
2790
2791 // Finally check the default case
2792 return StoreNode::Ideal(phase, can_reshape);
2793 }
2794
2795 //=============================================================================
2796 //------------------------------Identity---------------------------------------
2797 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2798 // No need to card mark when storing a null ptr
2799 Node* my_store = in(MemNode::OopStore);
2800 if (my_store->is_Store()) {
2801 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2802 if( t1 == TypePtr::NULL_PTR ) {
2803 return in(MemNode::Memory);
2804 }
2805 }
2806 return this;
2807 }
2808
2809 //=============================================================================
2810 //------------------------------Ideal---------------------------------------
2811 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2812 Node* progress = StoreNode::Ideal(phase, can_reshape);
2813 if (progress != NULL) return progress;
2814
2815 Node* my_store = in(MemNode::OopStore);
2816 if (my_store->is_MergeMem()) {
2817 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2818 set_req(MemNode::OopStore, mem);
2819 return this;
2820 }
2821
2822 return NULL;
2823 }
2824
2825 //------------------------------Value-----------------------------------------
2826 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2827 // Either input is TOP ==> the result is TOP
2828 const Type *t = phase->type( in(MemNode::Memory) );
2829 if( t == Type::TOP ) return Type::TOP;
2830 t = phase->type( in(MemNode::Address) );
2831 if( t == Type::TOP ) return Type::TOP;
2832 t = phase->type( in(MemNode::ValueIn) );
2833 if( t == Type::TOP ) return Type::TOP;
2834 // If extra input is TOP ==> the result is TOP
2835 t = phase->type( in(MemNode::OopStore) );
2836 if( t == Type::TOP ) return Type::TOP;
2837
2838 return StoreNode::Value( phase );
2839 }
2840
2841
2842 //=============================================================================
2843 //----------------------------------SCMemProjNode------------------------------
2844 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2845 {
2846 return bottom_type();
2847 }
2848
2849 //=============================================================================
2850 //----------------------------------LoadStoreNode------------------------------
2851 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2852 : Node(required),
2853 _type(rt),
2854 _adr_type(at),
2855 _barrier(0)
2856 {
2857 init_req(MemNode::Control, c );
2858 init_req(MemNode::Memory , mem);
2859 init_req(MemNode::Address, adr);
2860 init_req(MemNode::ValueIn, val);
2861 init_class_id(Class_LoadStore);
2862 }
2863
2864 uint LoadStoreNode::ideal_reg() const {
2865 return _type->ideal_reg();
2866 }
2867
2868 bool LoadStoreNode::result_not_used() const {
2869 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2870 Node *x = fast_out(i);
2871 if (x->Opcode() == Op_SCMemProj) continue;
2872 return false;
2873 }
2874 return true;
2875 }
2876
2877 MemBarNode* LoadStoreNode::trailing_membar() const {
2878 MemBarNode* trailing = NULL;
2879 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2880 Node* u = fast_out(i);
2881 if (u->is_MemBar()) {
2882 if (u->as_MemBar()->trailing_load_store()) {
2883 assert(u->Opcode() == Op_MemBarAcquire, "");
2884 assert(trailing == NULL, "only one");
2885 trailing = u->as_MemBar();
2886 #ifdef ASSERT
2887 Node* leading = trailing->leading_membar();
2888 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2889 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
2890 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
2891 #endif
2892 } else {
2893 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
2894 }
2895 }
2896 }
2897
2898 return trailing;
2899 }
2900
2901 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2902
2903 //=============================================================================
2904 //----------------------------------LoadStoreConditionalNode--------------------
2905 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2906 init_req(ExpectedIn, ex );
2907 }
2908
2909 //=============================================================================
2910 //-------------------------------adr_type--------------------------------------
2911 const TypePtr* ClearArrayNode::adr_type() const {
2912 Node *adr = in(3);
2913 if (adr == NULL) return NULL; // node is dead
2914 return MemNode::calculate_adr_type(adr->bottom_type());
2915 }
2916
2917 //------------------------------match_edge-------------------------------------
2918 // Do we Match on this edge index or not? Do not match memory
2919 uint ClearArrayNode::match_edge(uint idx) const {
2920 return idx > 1;
2921 }
2922
2923 //------------------------------Identity---------------------------------------
2924 // Clearing a zero length array does nothing
2925 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2926 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2927 }
2928
2929 //------------------------------Idealize---------------------------------------
2930 // Clearing a short array is faster with stores
2931 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2932 // Already know this is a large node, do not try to ideal it
2933 if (!IdealizeClearArrayNode || _is_large) return NULL;
2934
2935 const int unit = BytesPerLong;
2936 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2937 if (!t) return NULL;
2938 if (!t->is_con()) return NULL;
2939 intptr_t raw_count = t->get_con();
2940 intptr_t size = raw_count;
2941 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2942 // Clearing nothing uses the Identity call.
2943 // Negative clears are possible on dead ClearArrays
2944 // (see jck test stmt114.stmt11402.val).
2945 if (size <= 0 || size % unit != 0) return NULL;
2946 intptr_t count = size / unit;
2947 // Length too long; communicate this to matchers and assemblers.
2948 // Assemblers are responsible to produce fast hardware clears for it.
2949 if (size > InitArrayShortSize) {
2950 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
2951 }
2952 Node *mem = in(1);
2953 if( phase->type(mem)==Type::TOP ) return NULL;
2954 Node *adr = in(3);
2955 const Type* at = phase->type(adr);
2956 if( at==Type::TOP ) return NULL;
2957 const TypePtr* atp = at->isa_ptr();
2958 // adjust atp to be the correct array element address type
2959 if (atp == NULL) atp = TypePtr::BOTTOM;
2960 else atp = atp->add_offset(Type::OffsetBot);
2961 // Get base for derived pointer purposes
2962 if( adr->Opcode() != Op_AddP ) Unimplemented();
2963 Node *base = adr->in(1);
2964
2965 Node *zero = phase->makecon(TypeLong::ZERO);
2966 Node *off = phase->MakeConX(BytesPerLong);
2967 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2968 count--;
2969 while( count-- ) {
2970 mem = phase->transform(mem);
2971 adr = phase->transform(new AddPNode(base,adr,off));
2972 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2973 }
2974 return mem;
2975 }
2976
2977 //----------------------------step_through----------------------------------
2978 // Return allocation input memory edge if it is different instance
2979 // or itself if it is the one we are looking for.
2980 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2981 Node* n = *np;
2982 assert(n->is_ClearArray(), "sanity");
2983 intptr_t offset;
2984 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2985 // This method is called only before Allocate nodes are expanded
2986 // during macro nodes expansion. Before that ClearArray nodes are
2987 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2988 // Allocate nodes are expanded) which follows allocations.
2989 assert(alloc != NULL, "should have allocation");
2990 if (alloc->_idx == instance_id) {
2991 // Can not bypass initialization of the instance we are looking for.
2992 return false;
2993 }
2994 // Otherwise skip it.
2995 InitializeNode* init = alloc->initialization();
2996 if (init != NULL)
2997 *np = init->in(TypeFunc::Memory);
2998 else
2999 *np = alloc->in(TypeFunc::Memory);
3000 return true;
3001 }
3002
3003 //----------------------------clear_memory-------------------------------------
3004 // Generate code to initialize object storage to zero.
3005 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3006 intptr_t start_offset,
3007 Node* end_offset,
3008 PhaseGVN* phase) {
3009 intptr_t offset = start_offset;
3010
3011 int unit = BytesPerLong;
3012 if ((offset % unit) != 0) {
3013 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3014 adr = phase->transform(adr);
3015 const TypePtr* atp = TypeRawPtr::BOTTOM;
3016 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3017 mem = phase->transform(mem);
3018 offset += BytesPerInt;
3019 }
3020 assert((offset % unit) == 0, "");
3021
3022 // Initialize the remaining stuff, if any, with a ClearArray.
3023 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
3024 }
3025
3026 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3027 Node* start_offset,
3028 Node* end_offset,
3029 PhaseGVN* phase) {
3030 if (start_offset == end_offset) {
3031 // nothing to do
3032 return mem;
3033 }
3034
3035 int unit = BytesPerLong;
3036 Node* zbase = start_offset;
3037 Node* zend = end_offset;
3038
3039 // Scale to the unit required by the CPU:
3040 if (!Matcher::init_array_count_is_in_bytes) {
3041 Node* shift = phase->intcon(exact_log2(unit));
3042 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3043 zend = phase->transform(new URShiftXNode(zend, shift) );
3044 }
3045
3046 // Bulk clear double-words
3047 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3048 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3049 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3050 return phase->transform(mem);
3051 }
3052
3053 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3054 intptr_t start_offset,
3055 intptr_t end_offset,
3056 PhaseGVN* phase) {
3057 if (start_offset == end_offset) {
3058 // nothing to do
3059 return mem;
3060 }
3061
3062 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3063 intptr_t done_offset = end_offset;
3064 if ((done_offset % BytesPerLong) != 0) {
3065 done_offset -= BytesPerInt;
3066 }
3067 if (done_offset > start_offset) {
3068 mem = clear_memory(ctl, mem, dest,
3069 start_offset, phase->MakeConX(done_offset), phase);
3070 }
3071 if (done_offset < end_offset) { // emit the final 32-bit store
3072 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3073 adr = phase->transform(adr);
3074 const TypePtr* atp = TypeRawPtr::BOTTOM;
3075 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3076 mem = phase->transform(mem);
3077 done_offset += BytesPerInt;
3078 }
3079 assert(done_offset == end_offset, "");
3080 return mem;
3081 }
3082
3083 //=============================================================================
3084 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3085 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3086 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3087 #ifdef ASSERT
3088 , _pair_idx(0)
3089 #endif
3090 {
3091 init_class_id(Class_MemBar);
3092 Node* top = C->top();
3093 init_req(TypeFunc::I_O,top);
3094 init_req(TypeFunc::FramePtr,top);
3095 init_req(TypeFunc::ReturnAdr,top);
3096 if (precedent != NULL)
3097 init_req(TypeFunc::Parms, precedent);
3098 }
3099
3100 //------------------------------cmp--------------------------------------------
3101 uint MemBarNode::hash() const { return NO_HASH; }
3102 bool MemBarNode::cmp( const Node &n ) const {
3103 return (&n == this); // Always fail except on self
3104 }
3105
3106 //------------------------------make-------------------------------------------
3107 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3108 switch (opcode) {
3109 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3110 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3111 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3112 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3113 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3114 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3115 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3116 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3117 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3118 case Op_Initialize: return new InitializeNode(C, atp, pn);
3119 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3120 default: ShouldNotReachHere(); return NULL;
3121 }
3122 }
3123
3124 void MemBarNode::remove(PhaseIterGVN *igvn) {
3125 if (outcnt() != 2) {
3126 return;
3127 }
3128 if (trailing_store() || trailing_load_store()) {
3129 MemBarNode* leading = leading_membar();
3130 if (leading != NULL) {
3131 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3132 leading->remove(igvn);
3133 }
3134 }
3135 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3136 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3137 }
3138
3139 //------------------------------Ideal------------------------------------------
3140 // Return a node which is more "ideal" than the current node. Strip out
3141 // control copies
3142 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3143 if (remove_dead_region(phase, can_reshape)) return this;
3144 // Don't bother trying to transform a dead node
3145 if (in(0) && in(0)->is_top()) {
3146 return NULL;
3147 }
3148
3149 bool progress = false;
3150 // Eliminate volatile MemBars for scalar replaced objects.
3151 if (can_reshape && req() == (Precedent+1)) {
3152 bool eliminate = false;
3153 int opc = Opcode();
3154 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3155 // Volatile field loads and stores.
3156 Node* my_mem = in(MemBarNode::Precedent);
3157 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3158 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3159 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3160 // replace this Precedent (decodeN) with the Load instead.
3161 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3162 Node* load_node = my_mem->in(1);
3163 set_req(MemBarNode::Precedent, load_node);
3164 phase->is_IterGVN()->_worklist.push(my_mem);
3165 my_mem = load_node;
3166 } else {
3167 assert(my_mem->unique_out() == this, "sanity");
3168 del_req(Precedent);
3169 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3170 my_mem = NULL;
3171 }
3172 progress = true;
3173 }
3174 if (my_mem != NULL && my_mem->is_Mem()) {
3175 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3176 // Check for scalar replaced object reference.
3177 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3178 t_oop->offset() != Type::OffsetBot &&
3179 t_oop->offset() != Type::OffsetTop) {
3180 eliminate = true;
3181 }
3182 }
3183 } else if (opc == Op_MemBarRelease) {
3184 // Final field stores.
3185 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3186 if ((alloc != NULL) && alloc->is_Allocate() &&
3187 alloc->as_Allocate()->does_not_escape_thread()) {
3188 // The allocated object does not escape.
3189 eliminate = true;
3190 }
3191 }
3192 if (eliminate) {
3193 // Replace MemBar projections by its inputs.
3194 PhaseIterGVN* igvn = phase->is_IterGVN();
3195 remove(igvn);
3196 // Must return either the original node (now dead) or a new node
3197 // (Do not return a top here, since that would break the uniqueness of top.)
3198 return new ConINode(TypeInt::ZERO);
3199 }
3200 }
3201 return progress ? this : NULL;
3202 }
3203
3204 //------------------------------Value------------------------------------------
3205 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3206 if( !in(0) ) return Type::TOP;
3207 if( phase->type(in(0)) == Type::TOP )
3208 return Type::TOP;
3209 return TypeTuple::MEMBAR;
3210 }
3211
3212 //------------------------------match------------------------------------------
3213 // Construct projections for memory.
3214 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3215 switch (proj->_con) {
3216 case TypeFunc::Control:
3217 case TypeFunc::Memory:
3218 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3219 }
3220 ShouldNotReachHere();
3221 return NULL;
3222 }
3223
3224 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3225 trailing->_kind = TrailingStore;
3226 leading->_kind = LeadingStore;
3227 #ifdef ASSERT
3228 trailing->_pair_idx = leading->_idx;
3229 leading->_pair_idx = leading->_idx;
3230 #endif
3231 }
3232
3233 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3234 trailing->_kind = TrailingLoadStore;
3235 leading->_kind = LeadingLoadStore;
3236 #ifdef ASSERT
3237 trailing->_pair_idx = leading->_idx;
3238 leading->_pair_idx = leading->_idx;
3239 #endif
3240 }
3241
3242 MemBarNode* MemBarNode::trailing_membar() const {
3243 ResourceMark rm;
3244 Node* trailing = (Node*)this;
3245 VectorSet seen(Thread::current()->resource_area());
3246 Node_Stack multis(0);
3247 do {
3248 Node* c = trailing;
3249 uint i = 0;
3250 do {
3251 trailing = NULL;
3252 for (; i < c->outcnt(); i++) {
3253 Node* next = c->raw_out(i);
3254 if (next != c && next->is_CFG()) {
3255 if (c->is_MultiBranch()) {
3256 if (multis.node() == c) {
3257 multis.set_index(i+1);
3258 } else {
3259 multis.push(c, i+1);
3260 }
3261 }
3262 trailing = next;
3263 break;
3264 }
3265 }
3266 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3267 break;
3268 }
3269 while (multis.size() > 0) {
3270 c = multis.node();
3271 i = multis.index();
3272 if (i < c->req()) {
3273 break;
3274 }
3275 multis.pop();
3276 }
3277 } while (multis.size() > 0);
3278 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3279
3280 MemBarNode* mb = trailing->as_MemBar();
3281 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3282 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3283 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3284 return mb;
3285 }
3286
3287 MemBarNode* MemBarNode::leading_membar() const {
3288 ResourceMark rm;
3289 VectorSet seen(Thread::current()->resource_area());
3290 Node_Stack regions(0);
3291 Node* leading = in(0);
3292 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3293 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3294 leading = NULL;
3295 while (regions.size() > 0 && leading == NULL) {
3296 Node* r = regions.node();
3297 uint i = regions.index();
3298 if (i < r->req()) {
3299 leading = r->in(i);
3300 regions.set_index(i+1);
3301 } else {
3302 regions.pop();
3303 }
3304 }
3305 if (leading == NULL) {
3306 assert(regions.size() == 0, "all paths should have been tried");
3307 return NULL;
3308 }
3309 }
3310 if (leading->is_Region()) {
3311 regions.push(leading, 2);
3312 leading = leading->in(1);
3313 } else {
3314 leading = leading->in(0);
3315 }
3316 }
3317 #ifdef ASSERT
3318 Unique_Node_List wq;
3319 wq.push((Node*)this);
3320 uint found = 0;
3321 for (uint i = 0; i < wq.size(); i++) {
3322 Node* n = wq.at(i);
3323 if (n->is_Region()) {
3324 for (uint j = 1; j < n->req(); j++) {
3325 Node* in = n->in(j);
3326 if (in != NULL && !in->is_top()) {
3327 wq.push(in);
3328 }
3329 }
3330 } else {
3331 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3332 assert(n == leading, "consistency check failed");
3333 found++;
3334 } else {
3335 Node* in = n->in(0);
3336 if (in != NULL && !in->is_top()) {
3337 wq.push(in);
3338 }
3339 }
3340 }
3341 }
3342 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3343 #endif
3344 if (leading == NULL) {
3345 return NULL;
3346 }
3347 MemBarNode* mb = leading->as_MemBar();
3348 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3349 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3350 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3351 return mb;
3352 }
3353
3354 //===========================InitializeNode====================================
3355 // SUMMARY:
3356 // This node acts as a memory barrier on raw memory, after some raw stores.
3357 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3358 // The Initialize can 'capture' suitably constrained stores as raw inits.
3359 // It can coalesce related raw stores into larger units (called 'tiles').
3360 // It can avoid zeroing new storage for memory units which have raw inits.
3361 // At macro-expansion, it is marked 'complete', and does not optimize further.
3362 //
3363 // EXAMPLE:
3364 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3365 // ctl = incoming control; mem* = incoming memory
3366 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3367 // First allocate uninitialized memory and fill in the header:
3368 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3369 // ctl := alloc.Control; mem* := alloc.Memory*
3370 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3371 // Then initialize to zero the non-header parts of the raw memory block:
3372 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3373 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3374 // After the initialize node executes, the object is ready for service:
3375 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3376 // Suppose its body is immediately initialized as {1,2}:
3377 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3378 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3379 // mem.SLICE(#short[*]) := store2
3380 //
3381 // DETAILS:
3382 // An InitializeNode collects and isolates object initialization after
3383 // an AllocateNode and before the next possible safepoint. As a
3384 // memory barrier (MemBarNode), it keeps critical stores from drifting
3385 // down past any safepoint or any publication of the allocation.
3386 // Before this barrier, a newly-allocated object may have uninitialized bits.
3387 // After this barrier, it may be treated as a real oop, and GC is allowed.
3388 //
3389 // The semantics of the InitializeNode include an implicit zeroing of
3390 // the new object from object header to the end of the object.
3391 // (The object header and end are determined by the AllocateNode.)
3392 //
3393 // Certain stores may be added as direct inputs to the InitializeNode.
3394 // These stores must update raw memory, and they must be to addresses
3395 // derived from the raw address produced by AllocateNode, and with
3396 // a constant offset. They must be ordered by increasing offset.
3397 // The first one is at in(RawStores), the last at in(req()-1).
3398 // Unlike most memory operations, they are not linked in a chain,
3399 // but are displayed in parallel as users of the rawmem output of
3400 // the allocation.
3401 //
3402 // (See comments in InitializeNode::capture_store, which continue
3403 // the example given above.)
3404 //
3405 // When the associated Allocate is macro-expanded, the InitializeNode
3406 // may be rewritten to optimize collected stores. A ClearArrayNode
3407 // may also be created at that point to represent any required zeroing.
3408 // The InitializeNode is then marked 'complete', prohibiting further
3409 // capturing of nearby memory operations.
3410 //
3411 // During macro-expansion, all captured initializations which store
3412 // constant values of 32 bits or smaller are coalesced (if advantageous)
3413 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3414 // initialized in fewer memory operations. Memory words which are
3415 // covered by neither tiles nor non-constant stores are pre-zeroed
3416 // by explicit stores of zero. (The code shape happens to do all
3417 // zeroing first, then all other stores, with both sequences occurring
3418 // in order of ascending offsets.)
3419 //
3420 // Alternatively, code may be inserted between an AllocateNode and its
3421 // InitializeNode, to perform arbitrary initialization of the new object.
3422 // E.g., the object copying intrinsics insert complex data transfers here.
3423 // The initialization must then be marked as 'complete' disable the
3424 // built-in zeroing semantics and the collection of initializing stores.
3425 //
3426 // While an InitializeNode is incomplete, reads from the memory state
3427 // produced by it are optimizable if they match the control edge and
3428 // new oop address associated with the allocation/initialization.
3429 // They return a stored value (if the offset matches) or else zero.
3430 // A write to the memory state, if it matches control and address,
3431 // and if it is to a constant offset, may be 'captured' by the
3432 // InitializeNode. It is cloned as a raw memory operation and rewired
3433 // inside the initialization, to the raw oop produced by the allocation.
3434 // Operations on addresses which are provably distinct (e.g., to
3435 // other AllocateNodes) are allowed to bypass the initialization.
3436 //
3437 // The effect of all this is to consolidate object initialization
3438 // (both arrays and non-arrays, both piecewise and bulk) into a
3439 // single location, where it can be optimized as a unit.
3440 //
3441 // Only stores with an offset less than TrackedInitializationLimit words
3442 // will be considered for capture by an InitializeNode. This puts a
3443 // reasonable limit on the complexity of optimized initializations.
3444
3445 //---------------------------InitializeNode------------------------------------
3446 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3447 : MemBarNode(C, adr_type, rawoop),
3448 _is_complete(Incomplete), _does_not_escape(false)
3449 {
3450 init_class_id(Class_Initialize);
3451
3452 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3453 assert(in(RawAddress) == rawoop, "proper init");
3454 // Note: allocation() can be NULL, for secondary initialization barriers
3455 }
3456
3457 // Since this node is not matched, it will be processed by the
3458 // register allocator. Declare that there are no constraints
3459 // on the allocation of the RawAddress edge.
3460 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3461 // This edge should be set to top, by the set_complete. But be conservative.
3462 if (idx == InitializeNode::RawAddress)
3463 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3464 return RegMask::Empty;
3465 }
3466
3467 Node* InitializeNode::memory(uint alias_idx) {
3468 Node* mem = in(Memory);
3469 if (mem->is_MergeMem()) {
3470 return mem->as_MergeMem()->memory_at(alias_idx);
3471 } else {
3472 // incoming raw memory is not split
3473 return mem;
3474 }
3475 }
3476
3477 bool InitializeNode::is_non_zero() {
3478 if (is_complete()) return false;
3479 remove_extra_zeroes();
3480 return (req() > RawStores);
3481 }
3482
3483 void InitializeNode::set_complete(PhaseGVN* phase) {
3484 assert(!is_complete(), "caller responsibility");
3485 _is_complete = Complete;
3486
3487 // After this node is complete, it contains a bunch of
3488 // raw-memory initializations. There is no need for
3489 // it to have anything to do with non-raw memory effects.
3490 // Therefore, tell all non-raw users to re-optimize themselves,
3491 // after skipping the memory effects of this initialization.
3492 PhaseIterGVN* igvn = phase->is_IterGVN();
3493 if (igvn) igvn->add_users_to_worklist(this);
3494 }
3495
3496 // convenience function
3497 // return false if the init contains any stores already
3498 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3499 InitializeNode* init = initialization();
3500 if (init == NULL || init->is_complete()) return false;
3501 init->remove_extra_zeroes();
3502 // for now, if this allocation has already collected any inits, bail:
3503 if (init->is_non_zero()) return false;
3504 init->set_complete(phase);
3505 return true;
3506 }
3507
3508 void InitializeNode::remove_extra_zeroes() {
3509 if (req() == RawStores) return;
3510 Node* zmem = zero_memory();
3511 uint fill = RawStores;
3512 for (uint i = fill; i < req(); i++) {
3513 Node* n = in(i);
3514 if (n->is_top() || n == zmem) continue; // skip
3515 if (fill < i) set_req(fill, n); // compact
3516 ++fill;
3517 }
3518 // delete any empty spaces created:
3519 while (fill < req()) {
3520 del_req(fill);
3521 }
3522 }
3523
3524 // Helper for remembering which stores go with which offsets.
3525 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3526 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3527 intptr_t offset = -1;
3528 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3529 phase, offset);
3530 if (base == NULL) return -1; // something is dead,
3531 if (offset < 0) return -1; // dead, dead
3532 return offset;
3533 }
3534
3535 // Helper for proving that an initialization expression is
3536 // "simple enough" to be folded into an object initialization.
3537 // Attempts to prove that a store's initial value 'n' can be captured
3538 // within the initialization without creating a vicious cycle, such as:
3539 // { Foo p = new Foo(); p.next = p; }
3540 // True for constants and parameters and small combinations thereof.
3541 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
3542 ResourceMark rm;
3543 Unique_Node_List worklist;
3544 worklist.push(value);
3545
3546 uint complexity_limit = 20;
3547 for (uint j = 0; j < worklist.size(); j++) {
3548 if (j >= complexity_limit) {
3549 return false; // Bail out if processed too many nodes
3550 }
3551
3552 Node* n = worklist.at(j);
3553 if (n == NULL) continue; // (can this really happen?)
3554 if (n->is_Proj()) n = n->in(0);
3555 if (n == this) return false; // found a cycle
3556 if (n->is_Con()) continue;
3557 if (n->is_Start()) continue; // params, etc., are OK
3558 if (n->is_Root()) continue; // even better
3559
3560 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
3561 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
3562 continue;
3563 }
3564
3565 Node* ctl = n->in(0);
3566 if (ctl != NULL && !ctl->is_top()) {
3567 if (ctl->is_Proj()) ctl = ctl->in(0);
3568 if (ctl == this) return false;
3569
3570 // If we already know that the enclosing memory op is pinned right after
3571 // the init, then any control flow that the store has picked up
3572 // must have preceded the init, or else be equal to the init.
3573 // Even after loop optimizations (which might change control edges)
3574 // a store is never pinned *before* the availability of its inputs.
3575 if (!MemNode::all_controls_dominate(n, this))
3576 return false; // failed to prove a good control
3577 }
3578
3579 // Check data edges for possible dependencies on 'this'.
3580 for (uint i = 1; i < n->req(); i++) {
3581 Node* m = n->in(i);
3582 if (m == NULL || m == n || m->is_top()) continue;
3583
3584 // Only process data inputs once
3585 worklist.push(m);
3586 }
3587 }
3588
3589 return true;
3590 }
3591
3592 // Here are all the checks a Store must pass before it can be moved into
3593 // an initialization. Returns zero if a check fails.
3594 // On success, returns the (constant) offset to which the store applies,
3595 // within the initialized memory.
3596 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
3597 const int FAIL = 0;
3598 if (st->req() != MemNode::ValueIn + 1)
3599 return FAIL; // an inscrutable StoreNode (card mark?)
3600 Node* ctl = st->in(MemNode::Control);
3601 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3602 return FAIL; // must be unconditional after the initialization
3603 Node* mem = st->in(MemNode::Memory);
3604 if (!(mem->is_Proj() && mem->in(0) == this))
3605 return FAIL; // must not be preceded by other stores
3606 Node* adr = st->in(MemNode::Address);
3607 intptr_t offset;
3608 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3609 if (alloc == NULL)
3610 return FAIL; // inscrutable address
3611 if (alloc != allocation())
3612 return FAIL; // wrong allocation! (store needs to float up)
3613 int size_in_bytes = st->memory_size();
3614 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3615 return FAIL; // mismatched access
3616 }
3617 Node* val = st->in(MemNode::ValueIn);
3618
3619 if (!detect_init_independence(val, phase))
3620 return FAIL; // stored value must be 'simple enough'
3621
3622 // The Store can be captured only if nothing after the allocation
3623 // and before the Store is using the memory location that the store
3624 // overwrites.
3625 bool failed = false;
3626 // If is_complete_with_arraycopy() is true the shape of the graph is
3627 // well defined and is safe so no need for extra checks.
3628 if (!is_complete_with_arraycopy()) {
3629 // We are going to look at each use of the memory state following
3630 // the allocation to make sure nothing reads the memory that the
3631 // Store writes.
3632 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3633 int alias_idx = phase->C->get_alias_index(t_adr);
3634 ResourceMark rm;
3635 Unique_Node_List mems;
3636 mems.push(mem);
3637 Node* unique_merge = NULL;
3638 for (uint next = 0; next < mems.size(); ++next) {
3639 Node *m = mems.at(next);
3640 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3641 Node *n = m->fast_out(j);
3642 if (n->outcnt() == 0) {
3643 continue;
3644 }
3645 if (n == st) {
3646 continue;
3647 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3648 // If the control of this use is different from the control
3649 // of the Store which is right after the InitializeNode then
3650 // this node cannot be between the InitializeNode and the
3651 // Store.
3652 continue;
3653 } else if (n->is_MergeMem()) {
3654 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3655 // We can hit a MergeMemNode (that will likely go away
3656 // later) that is a direct use of the memory state
3657 // following the InitializeNode on the same slice as the
3658 // store node that we'd like to capture. We need to check
3659 // the uses of the MergeMemNode.
3660 mems.push(n);
3661 }
3662 } else if (n->is_Mem()) {
3663 Node* other_adr = n->in(MemNode::Address);
3664 if (other_adr == adr) {
3665 failed = true;
3666 break;
3667 } else {
3668 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3669 if (other_t_adr != NULL) {
3670 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3671 if (other_alias_idx == alias_idx) {
3672 // A load from the same memory slice as the store right
3673 // after the InitializeNode. We check the control of the
3674 // object/array that is loaded from. If it's the same as
3675 // the store control then we cannot capture the store.
3676 assert(!n->is_Store(), "2 stores to same slice on same control?");
3677 Node* base = other_adr;
3678 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3679 base = base->in(AddPNode::Base);
3680 if (base != NULL) {
3681 base = base->uncast();
3682 if (base->is_Proj() && base->in(0) == alloc) {
3683 failed = true;
3684 break;
3685 }
3686 }
3687 }
3688 }
3689 }
3690 } else {
3691 failed = true;
3692 break;
3693 }
3694 }
3695 }
3696 }
3697 if (failed) {
3698 if (!can_reshape) {
3699 // We decided we couldn't capture the store during parsing. We
3700 // should try again during the next IGVN once the graph is
3701 // cleaner.
3702 phase->C->record_for_igvn(st);
3703 }
3704 return FAIL;
3705 }
3706
3707 return offset; // success
3708 }
3709
3710 // Find the captured store in(i) which corresponds to the range
3711 // [start..start+size) in the initialized object.
3712 // If there is one, return its index i. If there isn't, return the
3713 // negative of the index where it should be inserted.
3714 // Return 0 if the queried range overlaps an initialization boundary
3715 // or if dead code is encountered.
3716 // If size_in_bytes is zero, do not bother with overlap checks.
3717 int InitializeNode::captured_store_insertion_point(intptr_t start,
3718 int size_in_bytes,
3719 PhaseTransform* phase) {
3720 const int FAIL = 0, MAX_STORE = BytesPerLong;
3721
3722 if (is_complete())
3723 return FAIL; // arraycopy got here first; punt
3724
3725 assert(allocation() != NULL, "must be present");
3726
3727 // no negatives, no header fields:
3728 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3729
3730 // after a certain size, we bail out on tracking all the stores:
3731 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3732 if (start >= ti_limit) return FAIL;
3733
3734 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3735 if (i >= limit) return -(int)i; // not found; here is where to put it
3736
3737 Node* st = in(i);
3738 intptr_t st_off = get_store_offset(st, phase);
3739 if (st_off < 0) {
3740 if (st != zero_memory()) {
3741 return FAIL; // bail out if there is dead garbage
3742 }
3743 } else if (st_off > start) {
3744 // ...we are done, since stores are ordered
3745 if (st_off < start + size_in_bytes) {
3746 return FAIL; // the next store overlaps
3747 }
3748 return -(int)i; // not found; here is where to put it
3749 } else if (st_off < start) {
3750 if (size_in_bytes != 0 &&
3751 start < st_off + MAX_STORE &&
3752 start < st_off + st->as_Store()->memory_size()) {
3753 return FAIL; // the previous store overlaps
3754 }
3755 } else {
3756 if (size_in_bytes != 0 &&
3757 st->as_Store()->memory_size() != size_in_bytes) {
3758 return FAIL; // mismatched store size
3759 }
3760 return i;
3761 }
3762
3763 ++i;
3764 }
3765 }
3766
3767 // Look for a captured store which initializes at the offset 'start'
3768 // with the given size. If there is no such store, and no other
3769 // initialization interferes, then return zero_memory (the memory
3770 // projection of the AllocateNode).
3771 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3772 PhaseTransform* phase) {
3773 assert(stores_are_sane(phase), "");
3774 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3775 if (i == 0) {
3776 return NULL; // something is dead
3777 } else if (i < 0) {
3778 return zero_memory(); // just primordial zero bits here
3779 } else {
3780 Node* st = in(i); // here is the store at this position
3781 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3782 return st;
3783 }
3784 }
3785
3786 // Create, as a raw pointer, an address within my new object at 'offset'.
3787 Node* InitializeNode::make_raw_address(intptr_t offset,
3788 PhaseTransform* phase) {
3789 Node* addr = in(RawAddress);
3790 if (offset != 0) {
3791 Compile* C = phase->C;
3792 addr = phase->transform( new AddPNode(C->top(), addr,
3793 phase->MakeConX(offset)) );
3794 }
3795 return addr;
3796 }
3797
3798 // Clone the given store, converting it into a raw store
3799 // initializing a field or element of my new object.
3800 // Caller is responsible for retiring the original store,
3801 // with subsume_node or the like.
3802 //
3803 // From the example above InitializeNode::InitializeNode,
3804 // here are the old stores to be captured:
3805 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3806 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3807 //
3808 // Here is the changed code; note the extra edges on init:
3809 // alloc = (Allocate ...)
3810 // rawoop = alloc.RawAddress
3811 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3812 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3813 // init = (Initialize alloc.Control alloc.Memory rawoop
3814 // rawstore1 rawstore2)
3815 //
3816 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3817 PhaseGVN* phase, bool can_reshape) {
3818 assert(stores_are_sane(phase), "");
3819
3820 if (start < 0) return NULL;
3821 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3822
3823 Compile* C = phase->C;
3824 int size_in_bytes = st->memory_size();
3825 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3826 if (i == 0) return NULL; // bail out
3827 Node* prev_mem = NULL; // raw memory for the captured store
3828 if (i > 0) {
3829 prev_mem = in(i); // there is a pre-existing store under this one
3830 set_req(i, C->top()); // temporarily disconnect it
3831 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3832 } else {
3833 i = -i; // no pre-existing store
3834 prev_mem = zero_memory(); // a slice of the newly allocated object
3835 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3836 set_req(--i, C->top()); // reuse this edge; it has been folded away
3837 else
3838 ins_req(i, C->top()); // build a new edge
3839 }
3840 Node* new_st = st->clone();
3841 new_st->set_req(MemNode::Control, in(Control));
3842 new_st->set_req(MemNode::Memory, prev_mem);
3843 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3844 new_st = phase->transform(new_st);
3845
3846 // At this point, new_st might have swallowed a pre-existing store
3847 // at the same offset, or perhaps new_st might have disappeared,
3848 // if it redundantly stored the same value (or zero to fresh memory).
3849
3850 // In any case, wire it in:
3851 phase->igvn_rehash_node_delayed(this);
3852 set_req(i, new_st);
3853
3854 // The caller may now kill the old guy.
3855 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3856 assert(check_st == new_st || check_st == NULL, "must be findable");
3857 assert(!is_complete(), "");
3858 return new_st;
3859 }
3860
3861 static bool store_constant(jlong* tiles, int num_tiles,
3862 intptr_t st_off, int st_size,
3863 jlong con) {
3864 if ((st_off & (st_size-1)) != 0)
3865 return false; // strange store offset (assume size==2**N)
3866 address addr = (address)tiles + st_off;
3867 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3868 switch (st_size) {
3869 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3870 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3871 case sizeof(jint): *(jint*) addr = (jint) con; break;
3872 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3873 default: return false; // strange store size (detect size!=2**N here)
3874 }
3875 return true; // return success to caller
3876 }
3877
3878 // Coalesce subword constants into int constants and possibly
3879 // into long constants. The goal, if the CPU permits,
3880 // is to initialize the object with a small number of 64-bit tiles.
3881 // Also, convert floating-point constants to bit patterns.
3882 // Non-constants are not relevant to this pass.
3883 //
3884 // In terms of the running example on InitializeNode::InitializeNode
3885 // and InitializeNode::capture_store, here is the transformation
3886 // of rawstore1 and rawstore2 into rawstore12:
3887 // alloc = (Allocate ...)
3888 // rawoop = alloc.RawAddress
3889 // tile12 = 0x00010002
3890 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3891 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3892 //
3893 void
3894 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3895 Node* size_in_bytes,
3896 PhaseGVN* phase) {
3897 Compile* C = phase->C;
3898
3899 assert(stores_are_sane(phase), "");
3900 // Note: After this pass, they are not completely sane,
3901 // since there may be some overlaps.
3902
3903 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3904
3905 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3906 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3907 size_limit = MIN2(size_limit, ti_limit);
3908 size_limit = align_up(size_limit, BytesPerLong);
3909 int num_tiles = size_limit / BytesPerLong;
3910
3911 // allocate space for the tile map:
3912 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3913 jlong tiles_buf[small_len];
3914 Node* nodes_buf[small_len];
3915 jlong inits_buf[small_len];
3916 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3917 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3918 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3919 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3920 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3921 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3922 // tiles: exact bitwise model of all primitive constants
3923 // nodes: last constant-storing node subsumed into the tiles model
3924 // inits: which bytes (in each tile) are touched by any initializations
3925
3926 //// Pass A: Fill in the tile model with any relevant stores.
3927
3928 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3929 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3930 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3931 Node* zmem = zero_memory(); // initially zero memory state
3932 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3933 Node* st = in(i);
3934 intptr_t st_off = get_store_offset(st, phase);
3935
3936 // Figure out the store's offset and constant value:
3937 if (st_off < header_size) continue; //skip (ignore header)
3938 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3939 int st_size = st->as_Store()->memory_size();
3940 if (st_off + st_size > size_limit) break;
3941
3942 // Record which bytes are touched, whether by constant or not.
3943 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3944 continue; // skip (strange store size)
3945
3946 const Type* val = phase->type(st->in(MemNode::ValueIn));
3947 if (!val->singleton()) continue; //skip (non-con store)
3948 BasicType type = val->basic_type();
3949
3950 jlong con = 0;
3951 switch (type) {
3952 case T_INT: con = val->is_int()->get_con(); break;
3953 case T_LONG: con = val->is_long()->get_con(); break;
3954 case T_FLOAT: con = jint_cast(val->getf()); break;
3955 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3956 default: continue; //skip (odd store type)
3957 }
3958
3959 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3960 st->Opcode() == Op_StoreL) {
3961 continue; // This StoreL is already optimal.
3962 }
3963
3964 // Store down the constant.
3965 store_constant(tiles, num_tiles, st_off, st_size, con);
3966
3967 intptr_t j = st_off >> LogBytesPerLong;
3968
3969 if (type == T_INT && st_size == BytesPerInt
3970 && (st_off & BytesPerInt) == BytesPerInt) {
3971 jlong lcon = tiles[j];
3972 if (!Matcher::isSimpleConstant64(lcon) &&
3973 st->Opcode() == Op_StoreI) {
3974 // This StoreI is already optimal by itself.
3975 jint* intcon = (jint*) &tiles[j];
3976 intcon[1] = 0; // undo the store_constant()
3977
3978 // If the previous store is also optimal by itself, back up and
3979 // undo the action of the previous loop iteration... if we can.
3980 // But if we can't, just let the previous half take care of itself.
3981 st = nodes[j];
3982 st_off -= BytesPerInt;
3983 con = intcon[0];
3984 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3985 assert(st_off >= header_size, "still ignoring header");
3986 assert(get_store_offset(st, phase) == st_off, "must be");
3987 assert(in(i-1) == zmem, "must be");
3988 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3989 assert(con == tcon->is_int()->get_con(), "must be");
3990 // Undo the effects of the previous loop trip, which swallowed st:
3991 intcon[0] = 0; // undo store_constant()
3992 set_req(i-1, st); // undo set_req(i, zmem)
3993 nodes[j] = NULL; // undo nodes[j] = st
3994 --old_subword; // undo ++old_subword
3995 }
3996 continue; // This StoreI is already optimal.
3997 }
3998 }
3999
4000 // This store is not needed.
4001 set_req(i, zmem);
4002 nodes[j] = st; // record for the moment
4003 if (st_size < BytesPerLong) // something has changed
4004 ++old_subword; // includes int/float, but who's counting...
4005 else ++old_long;
4006 }
4007
4008 if ((old_subword + old_long) == 0)
4009 return; // nothing more to do
4010
4011 //// Pass B: Convert any non-zero tiles into optimal constant stores.
4012 // Be sure to insert them before overlapping non-constant stores.
4013 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
4014 for (int j = 0; j < num_tiles; j++) {
4015 jlong con = tiles[j];
4016 jlong init = inits[j];
4017 if (con == 0) continue;
4018 jint con0, con1; // split the constant, address-wise
4019 jint init0, init1; // split the init map, address-wise
4020 { union { jlong con; jint intcon[2]; } u;
4021 u.con = con;
4022 con0 = u.intcon[0];
4023 con1 = u.intcon[1];
4024 u.con = init;
4025 init0 = u.intcon[0];
4026 init1 = u.intcon[1];
4027 }
4028
4029 Node* old = nodes[j];
4030 assert(old != NULL, "need the prior store");
4031 intptr_t offset = (j * BytesPerLong);
4032
4033 bool split = !Matcher::isSimpleConstant64(con);
4034
4035 if (offset < header_size) {
4036 assert(offset + BytesPerInt >= header_size, "second int counts");
4037 assert(*(jint*)&tiles[j] == 0, "junk in header");
4038 split = true; // only the second word counts
4039 // Example: int a[] = { 42 ... }
4040 } else if (con0 == 0 && init0 == -1) {
4041 split = true; // first word is covered by full inits
4042 // Example: int a[] = { ... foo(), 42 ... }
4043 } else if (con1 == 0 && init1 == -1) {
4044 split = true; // second word is covered by full inits
4045 // Example: int a[] = { ... 42, foo() ... }
4046 }
4047
4048 // Here's a case where init0 is neither 0 nor -1:
4049 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4050 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4051 // In this case the tile is not split; it is (jlong)42.
4052 // The big tile is stored down, and then the foo() value is inserted.
4053 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4054
4055 Node* ctl = old->in(MemNode::Control);
4056 Node* adr = make_raw_address(offset, phase);
4057 const TypePtr* atp = TypeRawPtr::BOTTOM;
4058
4059 // One or two coalesced stores to plop down.
4060 Node* st[2];
4061 intptr_t off[2];
4062 int nst = 0;
4063 if (!split) {
4064 ++new_long;
4065 off[nst] = offset;
4066 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4067 phase->longcon(con), T_LONG, MemNode::unordered);
4068 } else {
4069 // Omit either if it is a zero.
4070 if (con0 != 0) {
4071 ++new_int;
4072 off[nst] = offset;
4073 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4074 phase->intcon(con0), T_INT, MemNode::unordered);
4075 }
4076 if (con1 != 0) {
4077 ++new_int;
4078 offset += BytesPerInt;
4079 adr = make_raw_address(offset, phase);
4080 off[nst] = offset;
4081 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4082 phase->intcon(con1), T_INT, MemNode::unordered);
4083 }
4084 }
4085
4086 // Insert second store first, then the first before the second.
4087 // Insert each one just before any overlapping non-constant stores.
4088 while (nst > 0) {
4089 Node* st1 = st[--nst];
4090 C->copy_node_notes_to(st1, old);
4091 st1 = phase->transform(st1);
4092 offset = off[nst];
4093 assert(offset >= header_size, "do not smash header");
4094 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4095 guarantee(ins_idx != 0, "must re-insert constant store");
4096 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4097 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4098 set_req(--ins_idx, st1);
4099 else
4100 ins_req(ins_idx, st1);
4101 }
4102 }
4103
4104 if (PrintCompilation && WizardMode)
4105 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4106 old_subword, old_long, new_int, new_long);
4107 if (C->log() != NULL)
4108 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4109 old_subword, old_long, new_int, new_long);
4110
4111 // Clean up any remaining occurrences of zmem:
4112 remove_extra_zeroes();
4113 }
4114
4115 // Explore forward from in(start) to find the first fully initialized
4116 // word, and return its offset. Skip groups of subword stores which
4117 // together initialize full words. If in(start) is itself part of a
4118 // fully initialized word, return the offset of in(start). If there
4119 // are no following full-word stores, or if something is fishy, return
4120 // a negative value.
4121 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4122 int int_map = 0;
4123 intptr_t int_map_off = 0;
4124 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4125
4126 for (uint i = start, limit = req(); i < limit; i++) {
4127 Node* st = in(i);
4128
4129 intptr_t st_off = get_store_offset(st, phase);
4130 if (st_off < 0) break; // return conservative answer
4131
4132 int st_size = st->as_Store()->memory_size();
4133 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4134 return st_off; // we found a complete word init
4135 }
4136
4137 // update the map:
4138
4139 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4140 if (this_int_off != int_map_off) {
4141 // reset the map:
4142 int_map = 0;
4143 int_map_off = this_int_off;
4144 }
4145
4146 int subword_off = st_off - this_int_off;
4147 int_map |= right_n_bits(st_size) << subword_off;
4148 if ((int_map & FULL_MAP) == FULL_MAP) {
4149 return this_int_off; // we found a complete word init
4150 }
4151
4152 // Did this store hit or cross the word boundary?
4153 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4154 if (next_int_off == this_int_off + BytesPerInt) {
4155 // We passed the current int, without fully initializing it.
4156 int_map_off = next_int_off;
4157 int_map >>= BytesPerInt;
4158 } else if (next_int_off > this_int_off + BytesPerInt) {
4159 // We passed the current and next int.
4160 return this_int_off + BytesPerInt;
4161 }
4162 }
4163
4164 return -1;
4165 }
4166
4167
4168 // Called when the associated AllocateNode is expanded into CFG.
4169 // At this point, we may perform additional optimizations.
4170 // Linearize the stores by ascending offset, to make memory
4171 // activity as coherent as possible.
4172 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4173 intptr_t header_size,
4174 Node* size_in_bytes,
4175 PhaseIterGVN* phase) {
4176 assert(!is_complete(), "not already complete");
4177 assert(stores_are_sane(phase), "");
4178 assert(allocation() != NULL, "must be present");
4179
4180 remove_extra_zeroes();
4181
4182 if (ReduceFieldZeroing || ReduceBulkZeroing)
4183 // reduce instruction count for common initialization patterns
4184 coalesce_subword_stores(header_size, size_in_bytes, phase);
4185
4186 Node* zmem = zero_memory(); // initially zero memory state
4187 Node* inits = zmem; // accumulating a linearized chain of inits
4188 #ifdef ASSERT
4189 intptr_t first_offset = allocation()->minimum_header_size();
4190 intptr_t last_init_off = first_offset; // previous init offset
4191 intptr_t last_init_end = first_offset; // previous init offset+size
4192 intptr_t last_tile_end = first_offset; // previous tile offset+size
4193 #endif
4194 intptr_t zeroes_done = header_size;
4195
4196 bool do_zeroing = true; // we might give up if inits are very sparse
4197 int big_init_gaps = 0; // how many large gaps have we seen?
4198
4199 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4200 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4201
4202 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4203 Node* st = in(i);
4204 intptr_t st_off = get_store_offset(st, phase);
4205 if (st_off < 0)
4206 break; // unknown junk in the inits
4207 if (st->in(MemNode::Memory) != zmem)
4208 break; // complicated store chains somehow in list
4209
4210 int st_size = st->as_Store()->memory_size();
4211 intptr_t next_init_off = st_off + st_size;
4212
4213 if (do_zeroing && zeroes_done < next_init_off) {
4214 // See if this store needs a zero before it or under it.
4215 intptr_t zeroes_needed = st_off;
4216
4217 if (st_size < BytesPerInt) {
4218 // Look for subword stores which only partially initialize words.
4219 // If we find some, we must lay down some word-level zeroes first,
4220 // underneath the subword stores.
4221 //
4222 // Examples:
4223 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4224 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4225 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4226 //
4227 // Note: coalesce_subword_stores may have already done this,
4228 // if it was prompted by constant non-zero subword initializers.
4229 // But this case can still arise with non-constant stores.
4230
4231 intptr_t next_full_store = find_next_fullword_store(i, phase);
4232
4233 // In the examples above:
4234 // in(i) p q r s x y z
4235 // st_off 12 13 14 15 12 13 14
4236 // st_size 1 1 1 1 1 1 1
4237 // next_full_s. 12 16 16 16 16 16 16
4238 // z's_done 12 16 16 16 12 16 12
4239 // z's_needed 12 16 16 16 16 16 16
4240 // zsize 0 0 0 0 4 0 4
4241 if (next_full_store < 0) {
4242 // Conservative tack: Zero to end of current word.
4243 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4244 } else {
4245 // Zero to beginning of next fully initialized word.
4246 // Or, don't zero at all, if we are already in that word.
4247 assert(next_full_store >= zeroes_needed, "must go forward");
4248 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4249 zeroes_needed = next_full_store;
4250 }
4251 }
4252
4253 if (zeroes_needed > zeroes_done) {
4254 intptr_t zsize = zeroes_needed - zeroes_done;
4255 // Do some incremental zeroing on rawmem, in parallel with inits.
4256 zeroes_done = align_down(zeroes_done, BytesPerInt);
4257 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4258 zeroes_done, zeroes_needed,
4259 phase);
4260 zeroes_done = zeroes_needed;
4261 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4262 do_zeroing = false; // leave the hole, next time
4263 }
4264 }
4265
4266 // Collect the store and move on:
4267 phase->replace_input_of(st, MemNode::Memory, inits);
4268 inits = st; // put it on the linearized chain
4269 set_req(i, zmem); // unhook from previous position
4270
4271 if (zeroes_done == st_off)
4272 zeroes_done = next_init_off;
4273
4274 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4275
4276 #ifdef ASSERT
4277 // Various order invariants. Weaker than stores_are_sane because
4278 // a large constant tile can be filled in by smaller non-constant stores.
4279 assert(st_off >= last_init_off, "inits do not reverse");
4280 last_init_off = st_off;
4281 const Type* val = NULL;
4282 if (st_size >= BytesPerInt &&
4283 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4284 (int)val->basic_type() < (int)T_OBJECT) {
4285 assert(st_off >= last_tile_end, "tiles do not overlap");
4286 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4287 last_tile_end = MAX2(last_tile_end, next_init_off);
4288 } else {
4289 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4290 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4291 assert(st_off >= last_init_end, "inits do not overlap");
4292 last_init_end = next_init_off; // it's a non-tile
4293 }
4294 #endif //ASSERT
4295 }
4296
4297 remove_extra_zeroes(); // clear out all the zmems left over
4298 add_req(inits);
4299
4300 if (!(UseTLAB && ZeroTLAB)) {
4301 // If anything remains to be zeroed, zero it all now.
4302 zeroes_done = align_down(zeroes_done, BytesPerInt);
4303 // if it is the last unused 4 bytes of an instance, forget about it
4304 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4305 if (zeroes_done + BytesPerLong >= size_limit) {
4306 AllocateNode* alloc = allocation();
4307 assert(alloc != NULL, "must be present");
4308 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4309 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4310 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4311 if (zeroes_done == k->layout_helper())
4312 zeroes_done = size_limit;
4313 }
4314 }
4315 if (zeroes_done < size_limit) {
4316 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4317 zeroes_done, size_in_bytes, phase);
4318 }
4319 }
4320
4321 set_complete(phase);
4322 return rawmem;
4323 }
4324
4325
4326 #ifdef ASSERT
4327 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4328 if (is_complete())
4329 return true; // stores could be anything at this point
4330 assert(allocation() != NULL, "must be present");
4331 intptr_t last_off = allocation()->minimum_header_size();
4332 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4333 Node* st = in(i);
4334 intptr_t st_off = get_store_offset(st, phase);
4335 if (st_off < 0) continue; // ignore dead garbage
4336 if (last_off > st_off) {
4337 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4338 this->dump(2);
4339 assert(false, "ascending store offsets");
4340 return false;
4341 }
4342 last_off = st_off + st->as_Store()->memory_size();
4343 }
4344 return true;
4345 }
4346 #endif //ASSERT
4347
4348
4349
4350
4351 //============================MergeMemNode=====================================
4352 //
4353 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4354 // contributing store or call operations. Each contributor provides the memory
4355 // state for a particular "alias type" (see Compile::alias_type). For example,
4356 // if a MergeMem has an input X for alias category #6, then any memory reference
4357 // to alias category #6 may use X as its memory state input, as an exact equivalent
4358 // to using the MergeMem as a whole.
4359 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4360 //
4361 // (Here, the <N> notation gives the index of the relevant adr_type.)
4362 //
4363 // In one special case (and more cases in the future), alias categories overlap.
4364 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4365 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4366 // it is exactly equivalent to that state W:
4367 // MergeMem(<Bot>: W) <==> W
4368 //
4369 // Usually, the merge has more than one input. In that case, where inputs
4370 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4371 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4372 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4373 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4374 //
4375 // A merge can take a "wide" memory state as one of its narrow inputs.
4376 // This simply means that the merge observes out only the relevant parts of
4377 // the wide input. That is, wide memory states arriving at narrow merge inputs
4378 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4379 //
4380 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4381 // and that memory slices "leak through":
4382 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4383 //
4384 // But, in such a cascade, repeated memory slices can "block the leak":
4385 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4386 //
4387 // In the last example, Y is not part of the combined memory state of the
4388 // outermost MergeMem. The system must, of course, prevent unschedulable
4389 // memory states from arising, so you can be sure that the state Y is somehow
4390 // a precursor to state Y'.
4391 //
4392 //
4393 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4394 // of each MergeMemNode array are exactly the numerical alias indexes, including
4395 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4396 // Compile::alias_type (and kin) produce and manage these indexes.
4397 //
4398 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4399 // (Note that this provides quick access to the top node inside MergeMem methods,
4400 // without the need to reach out via TLS to Compile::current.)
4401 //
4402 // As a consequence of what was just described, a MergeMem that represents a full
4403 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4404 // containing all alias categories.
4405 //
4406 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4407 //
4408 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4409 // a memory state for the alias type <N>, or else the top node, meaning that
4410 // there is no particular input for that alias type. Note that the length of
4411 // a MergeMem is variable, and may be extended at any time to accommodate new
4412 // memory states at larger alias indexes. When merges grow, they are of course
4413 // filled with "top" in the unused in() positions.
4414 //
4415 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4416 // (Top was chosen because it works smoothly with passes like GCM.)
4417 //
4418 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4419 // the type of random VM bits like TLS references.) Since it is always the
4420 // first non-Bot memory slice, some low-level loops use it to initialize an
4421 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4422 //
4423 //
4424 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4425 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4426 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4427 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4428 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4429 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4430 //
4431 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4432 // really that different from the other memory inputs. An abbreviation called
4433 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4434 //
4435 //
4436 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4437 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4438 // that "emerges though" the base memory will be marked as excluding the alias types
4439 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4440 //
4441 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4442 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4443 //
4444 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4445 // (It is currently unimplemented.) As you can see, the resulting merge is
4446 // actually a disjoint union of memory states, rather than an overlay.
4447 //
4448
4449 //------------------------------MergeMemNode-----------------------------------
4450 Node* MergeMemNode::make_empty_memory() {
4451 Node* empty_memory = (Node*) Compile::current()->top();
4452 assert(empty_memory->is_top(), "correct sentinel identity");
4453 return empty_memory;
4454 }
4455
4456 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4457 init_class_id(Class_MergeMem);
4458 // all inputs are nullified in Node::Node(int)
4459 // set_input(0, NULL); // no control input
4460
4461 // Initialize the edges uniformly to top, for starters.
4462 Node* empty_mem = make_empty_memory();
4463 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4464 init_req(i,empty_mem);
4465 }
4466 assert(empty_memory() == empty_mem, "");
4467
4468 if( new_base != NULL && new_base->is_MergeMem() ) {
4469 MergeMemNode* mdef = new_base->as_MergeMem();
4470 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4471 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4472 mms.set_memory(mms.memory2());
4473 }
4474 assert(base_memory() == mdef->base_memory(), "");
4475 } else {
4476 set_base_memory(new_base);
4477 }
4478 }
4479
4480 // Make a new, untransformed MergeMem with the same base as 'mem'.
4481 // If mem is itself a MergeMem, populate the result with the same edges.
4482 MergeMemNode* MergeMemNode::make(Node* mem) {
4483 return new MergeMemNode(mem);
4484 }
4485
4486 //------------------------------cmp--------------------------------------------
4487 uint MergeMemNode::hash() const { return NO_HASH; }
4488 bool MergeMemNode::cmp( const Node &n ) const {
4489 return (&n == this); // Always fail except on self
4490 }
4491
4492 //------------------------------Identity---------------------------------------
4493 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4494 // Identity if this merge point does not record any interesting memory
4495 // disambiguations.
4496 Node* base_mem = base_memory();
4497 Node* empty_mem = empty_memory();
4498 if (base_mem != empty_mem) { // Memory path is not dead?
4499 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4500 Node* mem = in(i);
4501 if (mem != empty_mem && mem != base_mem) {
4502 return this; // Many memory splits; no change
4503 }
4504 }
4505 }
4506 return base_mem; // No memory splits; ID on the one true input
4507 }
4508
4509 //------------------------------Ideal------------------------------------------
4510 // This method is invoked recursively on chains of MergeMem nodes
4511 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4512 // Remove chain'd MergeMems
4513 //
4514 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4515 // relative to the "in(Bot)". Since we are patching both at the same time,
4516 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4517 // but rewrite each "in(i)" relative to the new "in(Bot)".
4518 Node *progress = NULL;
4519
4520
4521 Node* old_base = base_memory();
4522 Node* empty_mem = empty_memory();
4523 if (old_base == empty_mem)
4524 return NULL; // Dead memory path.
4525
4526 MergeMemNode* old_mbase;
4527 if (old_base != NULL && old_base->is_MergeMem())
4528 old_mbase = old_base->as_MergeMem();
4529 else
4530 old_mbase = NULL;
4531 Node* new_base = old_base;
4532
4533 // simplify stacked MergeMems in base memory
4534 if (old_mbase) new_base = old_mbase->base_memory();
4535
4536 // the base memory might contribute new slices beyond my req()
4537 if (old_mbase) grow_to_match(old_mbase);
4538
4539 // Look carefully at the base node if it is a phi.
4540 PhiNode* phi_base;
4541 if (new_base != NULL && new_base->is_Phi())
4542 phi_base = new_base->as_Phi();
4543 else
4544 phi_base = NULL;
4545
4546 Node* phi_reg = NULL;
4547 uint phi_len = (uint)-1;
4548 if (phi_base != NULL && !phi_base->is_copy()) {
4549 // do not examine phi if degraded to a copy
4550 phi_reg = phi_base->region();
4551 phi_len = phi_base->req();
4552 // see if the phi is unfinished
4553 for (uint i = 1; i < phi_len; i++) {
4554 if (phi_base->in(i) == NULL) {
4555 // incomplete phi; do not look at it yet!
4556 phi_reg = NULL;
4557 phi_len = (uint)-1;
4558 break;
4559 }
4560 }
4561 }
4562
4563 // Note: We do not call verify_sparse on entry, because inputs
4564 // can normalize to the base_memory via subsume_node or similar
4565 // mechanisms. This method repairs that damage.
4566
4567 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4568
4569 // Look at each slice.
4570 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4571 Node* old_in = in(i);
4572 // calculate the old memory value
4573 Node* old_mem = old_in;
4574 if (old_mem == empty_mem) old_mem = old_base;
4575 assert(old_mem == memory_at(i), "");
4576
4577 // maybe update (reslice) the old memory value
4578
4579 // simplify stacked MergeMems
4580 Node* new_mem = old_mem;
4581 MergeMemNode* old_mmem;
4582 if (old_mem != NULL && old_mem->is_MergeMem())
4583 old_mmem = old_mem->as_MergeMem();
4584 else
4585 old_mmem = NULL;
4586 if (old_mmem == this) {
4587 // This can happen if loops break up and safepoints disappear.
4588 // A merge of BotPtr (default) with a RawPtr memory derived from a
4589 // safepoint can be rewritten to a merge of the same BotPtr with
4590 // the BotPtr phi coming into the loop. If that phi disappears
4591 // also, we can end up with a self-loop of the mergemem.
4592 // In general, if loops degenerate and memory effects disappear,
4593 // a mergemem can be left looking at itself. This simply means
4594 // that the mergemem's default should be used, since there is
4595 // no longer any apparent effect on this slice.
4596 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4597 // from start. Update the input to TOP.
4598 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4599 }
4600 else if (old_mmem != NULL) {
4601 new_mem = old_mmem->memory_at(i);
4602 }
4603 // else preceding memory was not a MergeMem
4604
4605 // replace equivalent phis (unfortunately, they do not GVN together)
4606 if (new_mem != NULL && new_mem != new_base &&
4607 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4608 if (new_mem->is_Phi()) {
4609 PhiNode* phi_mem = new_mem->as_Phi();
4610 for (uint i = 1; i < phi_len; i++) {
4611 if (phi_base->in(i) != phi_mem->in(i)) {
4612 phi_mem = NULL;
4613 break;
4614 }
4615 }
4616 if (phi_mem != NULL) {
4617 // equivalent phi nodes; revert to the def
4618 new_mem = new_base;
4619 }
4620 }
4621 }
4622
4623 // maybe store down a new value
4624 Node* new_in = new_mem;
4625 if (new_in == new_base) new_in = empty_mem;
4626
4627 if (new_in != old_in) {
4628 // Warning: Do not combine this "if" with the previous "if"
4629 // A memory slice might have be be rewritten even if it is semantically
4630 // unchanged, if the base_memory value has changed.
4631 set_req(i, new_in);
4632 progress = this; // Report progress
4633 }
4634 }
4635
4636 if (new_base != old_base) {
4637 set_req(Compile::AliasIdxBot, new_base);
4638 // Don't use set_base_memory(new_base), because we need to update du.
4639 assert(base_memory() == new_base, "");
4640 progress = this;
4641 }
4642
4643 if( base_memory() == this ) {
4644 // a self cycle indicates this memory path is dead
4645 set_req(Compile::AliasIdxBot, empty_mem);
4646 }
4647
4648 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4649 // Recursion must occur after the self cycle check above
4650 if( base_memory()->is_MergeMem() ) {
4651 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4652 Node *m = phase->transform(new_mbase); // Rollup any cycles
4653 if( m != NULL &&
4654 (m->is_top() ||
4655 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4656 // propagate rollup of dead cycle to self
4657 set_req(Compile::AliasIdxBot, empty_mem);
4658 }
4659 }
4660
4661 if( base_memory() == empty_mem ) {
4662 progress = this;
4663 // Cut inputs during Parse phase only.
4664 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4665 if( !can_reshape ) {
4666 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4667 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4668 }
4669 }
4670 }
4671
4672 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4673 // Check if PhiNode::Ideal's "Split phis through memory merges"
4674 // transform should be attempted. Look for this->phi->this cycle.
4675 uint merge_width = req();
4676 if (merge_width > Compile::AliasIdxRaw) {
4677 PhiNode* phi = base_memory()->as_Phi();
4678 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4679 if (phi->in(i) == this) {
4680 phase->is_IterGVN()->_worklist.push(phi);
4681 break;
4682 }
4683 }
4684 }
4685 }
4686
4687 assert(progress || verify_sparse(), "please, no dups of base");
4688 return progress;
4689 }
4690
4691 //-------------------------set_base_memory-------------------------------------
4692 void MergeMemNode::set_base_memory(Node *new_base) {
4693 Node* empty_mem = empty_memory();
4694 set_req(Compile::AliasIdxBot, new_base);
4695 assert(memory_at(req()) == new_base, "must set default memory");
4696 // Clear out other occurrences of new_base:
4697 if (new_base != empty_mem) {
4698 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4699 if (in(i) == new_base) set_req(i, empty_mem);
4700 }
4701 }
4702 }
4703
4704 //------------------------------out_RegMask------------------------------------
4705 const RegMask &MergeMemNode::out_RegMask() const {
4706 return RegMask::Empty;
4707 }
4708
4709 //------------------------------dump_spec--------------------------------------
4710 #ifndef PRODUCT
4711 void MergeMemNode::dump_spec(outputStream *st) const {
4712 st->print(" {");
4713 Node* base_mem = base_memory();
4714 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4715 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4716 if (mem == base_mem) { st->print(" -"); continue; }
4717 st->print( " N%d:", mem->_idx );
4718 Compile::current()->get_adr_type(i)->dump_on(st);
4719 }
4720 st->print(" }");
4721 }
4722 #endif // !PRODUCT
4723
4724
4725 #ifdef ASSERT
4726 static bool might_be_same(Node* a, Node* b) {
4727 if (a == b) return true;
4728 if (!(a->is_Phi() || b->is_Phi())) return false;
4729 // phis shift around during optimization
4730 return true; // pretty stupid...
4731 }
4732
4733 // verify a narrow slice (either incoming or outgoing)
4734 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4735 if (!VerifyAliases) return; // don't bother to verify unless requested
4736 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4737 if (Node::in_dump()) return; // muzzle asserts when printing
4738 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4739 assert(n != NULL, "");
4740 // Elide intervening MergeMem's
4741 while (n->is_MergeMem()) {
4742 n = n->as_MergeMem()->memory_at(alias_idx);
4743 }
4744 Compile* C = Compile::current();
4745 const TypePtr* n_adr_type = n->adr_type();
4746 if (n == m->empty_memory()) {
4747 // Implicit copy of base_memory()
4748 } else if (n_adr_type != TypePtr::BOTTOM) {
4749 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4750 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4751 } else {
4752 // A few places like make_runtime_call "know" that VM calls are narrow,
4753 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4754 bool expected_wide_mem = false;
4755 if (n == m->base_memory()) {
4756 expected_wide_mem = true;
4757 } else if (alias_idx == Compile::AliasIdxRaw ||
4758 n == m->memory_at(Compile::AliasIdxRaw)) {
4759 expected_wide_mem = true;
4760 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4761 // memory can "leak through" calls on channels that
4762 // are write-once. Allow this also.
4763 expected_wide_mem = true;
4764 }
4765 assert(expected_wide_mem, "expected narrow slice replacement");
4766 }
4767 }
4768 #else // !ASSERT
4769 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4770 #endif
4771
4772
4773 //-----------------------------memory_at---------------------------------------
4774 Node* MergeMemNode::memory_at(uint alias_idx) const {
4775 assert(alias_idx >= Compile::AliasIdxRaw ||
4776 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4777 "must avoid base_memory and AliasIdxTop");
4778
4779 // Otherwise, it is a narrow slice.
4780 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4781 Compile *C = Compile::current();
4782 if (is_empty_memory(n)) {
4783 // the array is sparse; empty slots are the "top" node
4784 n = base_memory();
4785 assert(Node::in_dump()
4786 || n == NULL || n->bottom_type() == Type::TOP
4787 || n->adr_type() == NULL // address is TOP
4788 || n->adr_type() == TypePtr::BOTTOM
4789 || n->adr_type() == TypeRawPtr::BOTTOM
4790 || Compile::current()->AliasLevel() == 0,
4791 "must be a wide memory");
4792 // AliasLevel == 0 if we are organizing the memory states manually.
4793 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4794 } else {
4795 // make sure the stored slice is sane
4796 #ifdef ASSERT
4797 if (VMError::is_error_reported() || Node::in_dump()) {
4798 } else if (might_be_same(n, base_memory())) {
4799 // Give it a pass: It is a mostly harmless repetition of the base.
4800 // This can arise normally from node subsumption during optimization.
4801 } else {
4802 verify_memory_slice(this, alias_idx, n);
4803 }
4804 #endif
4805 }
4806 return n;
4807 }
4808
4809 //---------------------------set_memory_at-------------------------------------
4810 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4811 verify_memory_slice(this, alias_idx, n);
4812 Node* empty_mem = empty_memory();
4813 if (n == base_memory()) n = empty_mem; // collapse default
4814 uint need_req = alias_idx+1;
4815 if (req() < need_req) {
4816 if (n == empty_mem) return; // already the default, so do not grow me
4817 // grow the sparse array
4818 do {
4819 add_req(empty_mem);
4820 } while (req() < need_req);
4821 }
4822 set_req( alias_idx, n );
4823 }
4824
4825
4826
4827 //--------------------------iteration_setup------------------------------------
4828 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4829 if (other != NULL) {
4830 grow_to_match(other);
4831 // invariant: the finite support of mm2 is within mm->req()
4832 #ifdef ASSERT
4833 for (uint i = req(); i < other->req(); i++) {
4834 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4835 }
4836 #endif
4837 }
4838 // Replace spurious copies of base_memory by top.
4839 Node* base_mem = base_memory();
4840 if (base_mem != NULL && !base_mem->is_top()) {
4841 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4842 if (in(i) == base_mem)
4843 set_req(i, empty_memory());
4844 }
4845 }
4846 }
4847
4848 //---------------------------grow_to_match-------------------------------------
4849 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4850 Node* empty_mem = empty_memory();
4851 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4852 // look for the finite support of the other memory
4853 for (uint i = other->req(); --i >= req(); ) {
4854 if (other->in(i) != empty_mem) {
4855 uint new_len = i+1;
4856 while (req() < new_len) add_req(empty_mem);
4857 break;
4858 }
4859 }
4860 }
4861
4862 //---------------------------verify_sparse-------------------------------------
4863 #ifndef PRODUCT
4864 bool MergeMemNode::verify_sparse() const {
4865 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4866 Node* base_mem = base_memory();
4867 // The following can happen in degenerate cases, since empty==top.
4868 if (is_empty_memory(base_mem)) return true;
4869 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4870 assert(in(i) != NULL, "sane slice");
4871 if (in(i) == base_mem) return false; // should have been the sentinel value!
4872 }
4873 return true;
4874 }
4875
4876 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4877 Node* n;
4878 n = mm->in(idx);
4879 if (mem == n) return true; // might be empty_memory()
4880 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4881 if (mem == n) return true;
4882 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4883 if (mem == n) return true;
4884 if (n == NULL) break;
4885 }
4886 return false;
4887 }
4888 #endif // !PRODUCT