Tue, 26 Nov 2013 18:38:19 -0800
8028515: PPPC64 (part 113.2): opto: Introduce LoadFence/StoreFence.
Summary: Use new nodes for loadFence/storeFence intrinsics in C2.
Reviewed-by: kvn, dholmes
1 /*
2 * Copyright (c) 1997, 2012, 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 */
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "oops/objArrayKlass.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/compile.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/machnode.hpp"
36 #include "opto/matcher.hpp"
37 #include "opto/memnode.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/phaseX.hpp"
40 #include "opto/regmask.hpp"
42 // Portions of code courtesy of Clifford Click
44 // Optimization - Graph Style
46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
48 //=============================================================================
49 uint MemNode::size_of() const { return sizeof(*this); }
51 const TypePtr *MemNode::adr_type() const {
52 Node* adr = in(Address);
53 const TypePtr* cross_check = NULL;
54 DEBUG_ONLY(cross_check = _adr_type);
55 return calculate_adr_type(adr->bottom_type(), cross_check);
56 }
58 #ifndef PRODUCT
59 void MemNode::dump_spec(outputStream *st) const {
60 if (in(Address) == NULL) return; // node is dead
61 #ifndef ASSERT
62 // fake the missing field
63 const TypePtr* _adr_type = NULL;
64 if (in(Address) != NULL)
65 _adr_type = in(Address)->bottom_type()->isa_ptr();
66 #endif
67 dump_adr_type(this, _adr_type, st);
69 Compile* C = Compile::current();
70 if( C->alias_type(_adr_type)->is_volatile() )
71 st->print(" Volatile!");
72 }
74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
75 st->print(" @");
76 if (adr_type == NULL) {
77 st->print("NULL");
78 } else {
79 adr_type->dump_on(st);
80 Compile* C = Compile::current();
81 Compile::AliasType* atp = NULL;
82 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
83 if (atp == NULL)
84 st->print(", idx=?\?;");
85 else if (atp->index() == Compile::AliasIdxBot)
86 st->print(", idx=Bot;");
87 else if (atp->index() == Compile::AliasIdxTop)
88 st->print(", idx=Top;");
89 else if (atp->index() == Compile::AliasIdxRaw)
90 st->print(", idx=Raw;");
91 else {
92 ciField* field = atp->field();
93 if (field) {
94 st->print(", name=");
95 field->print_name_on(st);
96 }
97 st->print(", idx=%d;", atp->index());
98 }
99 }
100 }
102 extern void print_alias_types();
104 #endif
106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
107 assert((t_oop != NULL), "sanity");
108 bool is_instance = t_oop->is_known_instance_field();
109 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
110 (load != NULL) && load->is_Load() &&
111 (phase->is_IterGVN() != NULL);
112 if (!(is_instance || is_boxed_value_load))
113 return mchain; // don't try to optimize non-instance types
114 uint instance_id = t_oop->instance_id();
115 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
116 Node *prev = NULL;
117 Node *result = mchain;
118 while (prev != result) {
119 prev = result;
120 if (result == start_mem)
121 break; // hit one of our sentinels
122 // skip over a call which does not affect this memory slice
123 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
124 Node *proj_in = result->in(0);
125 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
126 break; // hit one of our sentinels
127 } else if (proj_in->is_Call()) {
128 CallNode *call = proj_in->as_Call();
129 if (!call->may_modify(t_oop, phase)) { // returns false for instances
130 result = call->in(TypeFunc::Memory);
131 }
132 } else if (proj_in->is_Initialize()) {
133 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
134 // Stop if this is the initialization for the object instance which
135 // which contains this memory slice, otherwise skip over it.
136 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
137 break;
138 }
139 if (is_instance) {
140 result = proj_in->in(TypeFunc::Memory);
141 } else if (is_boxed_value_load) {
142 Node* klass = alloc->in(AllocateNode::KlassNode);
143 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
144 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
145 result = proj_in->in(TypeFunc::Memory); // not related allocation
146 }
147 }
148 } else if (proj_in->is_MemBar()) {
149 result = proj_in->in(TypeFunc::Memory);
150 } else {
151 assert(false, "unexpected projection");
152 }
153 } else if (result->is_ClearArray()) {
154 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
155 // Can not bypass initialization of the instance
156 // we are looking for.
157 break;
158 }
159 // Otherwise skip it (the call updated 'result' value).
160 } else if (result->is_MergeMem()) {
161 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
162 }
163 }
164 return result;
165 }
167 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
168 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
169 if (t_oop == NULL)
170 return mchain; // don't try to optimize non-oop types
171 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
172 bool is_instance = t_oop->is_known_instance_field();
173 PhaseIterGVN *igvn = phase->is_IterGVN();
174 if (is_instance && igvn != NULL && result->is_Phi()) {
175 PhiNode *mphi = result->as_Phi();
176 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
177 const TypePtr *t = mphi->adr_type();
178 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
179 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
180 t->is_oopptr()->cast_to_exactness(true)
181 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
182 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
183 // clone the Phi with our address type
184 result = mphi->split_out_instance(t_adr, igvn);
185 } else {
186 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
187 }
188 }
189 return result;
190 }
192 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
193 uint alias_idx = phase->C->get_alias_index(tp);
194 Node *mem = mmem;
195 #ifdef ASSERT
196 {
197 // Check that current type is consistent with the alias index used during graph construction
198 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
199 bool consistent = adr_check == NULL || adr_check->empty() ||
200 phase->C->must_alias(adr_check, alias_idx );
201 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
202 if( !consistent && adr_check != NULL && !adr_check->empty() &&
203 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
204 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
205 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
206 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
207 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
208 // don't assert if it is dead code.
209 consistent = true;
210 }
211 if( !consistent ) {
212 st->print("alias_idx==%d, adr_check==", alias_idx);
213 if( adr_check == NULL ) {
214 st->print("NULL");
215 } else {
216 adr_check->dump();
217 }
218 st->cr();
219 print_alias_types();
220 assert(consistent, "adr_check must match alias idx");
221 }
222 }
223 #endif
224 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
225 // means an array I have not precisely typed yet. Do not do any
226 // alias stuff with it any time soon.
227 const TypeOopPtr *toop = tp->isa_oopptr();
228 if( tp->base() != Type::AnyPtr &&
229 !(toop &&
230 toop->klass() != NULL &&
231 toop->klass()->is_java_lang_Object() &&
232 toop->offset() == Type::OffsetBot) ) {
233 // compress paths and change unreachable cycles to TOP
234 // If not, we can update the input infinitely along a MergeMem cycle
235 // Equivalent code in PhiNode::Ideal
236 Node* m = phase->transform(mmem);
237 // If transformed to a MergeMem, get the desired slice
238 // Otherwise the returned node represents memory for every slice
239 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
240 // Update input if it is progress over what we have now
241 }
242 return mem;
243 }
245 //--------------------------Ideal_common---------------------------------------
246 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
247 // Unhook non-raw memories from complete (macro-expanded) initializations.
248 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
249 // If our control input is a dead region, kill all below the region
250 Node *ctl = in(MemNode::Control);
251 if (ctl && remove_dead_region(phase, can_reshape))
252 return this;
253 ctl = in(MemNode::Control);
254 // Don't bother trying to transform a dead node
255 if (ctl && ctl->is_top()) return NodeSentinel;
257 PhaseIterGVN *igvn = phase->is_IterGVN();
258 // Wait if control on the worklist.
259 if (ctl && can_reshape && igvn != NULL) {
260 Node* bol = NULL;
261 Node* cmp = NULL;
262 if (ctl->in(0)->is_If()) {
263 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
264 bol = ctl->in(0)->in(1);
265 if (bol->is_Bool())
266 cmp = ctl->in(0)->in(1)->in(1);
267 }
268 if (igvn->_worklist.member(ctl) ||
269 (bol != NULL && igvn->_worklist.member(bol)) ||
270 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
271 // This control path may be dead.
272 // Delay this memory node transformation until the control is processed.
273 phase->is_IterGVN()->_worklist.push(this);
274 return NodeSentinel; // caller will return NULL
275 }
276 }
277 // Ignore if memory is dead, or self-loop
278 Node *mem = in(MemNode::Memory);
279 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
280 assert(mem != this, "dead loop in MemNode::Ideal");
282 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
283 // This memory slice may be dead.
284 // Delay this mem node transformation until the memory is processed.
285 phase->is_IterGVN()->_worklist.push(this);
286 return NodeSentinel; // caller will return NULL
287 }
289 Node *address = in(MemNode::Address);
290 const Type *t_adr = phase->type(address);
291 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
293 if (can_reshape && igvn != NULL &&
294 (igvn->_worklist.member(address) ||
295 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
296 // The address's base and type may change when the address is processed.
297 // Delay this mem node transformation until the address is processed.
298 phase->is_IterGVN()->_worklist.push(this);
299 return NodeSentinel; // caller will return NULL
300 }
302 // Do NOT remove or optimize the next lines: ensure a new alias index
303 // is allocated for an oop pointer type before Escape Analysis.
304 // Note: C++ will not remove it since the call has side effect.
305 if (t_adr->isa_oopptr()) {
306 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
307 }
309 #ifdef ASSERT
310 Node* base = NULL;
311 if (address->is_AddP())
312 base = address->in(AddPNode::Base);
313 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
314 !t_adr->isa_rawptr()) {
315 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
316 Compile* C = phase->C;
317 tty->cr();
318 tty->print_cr("===== NULL+offs not RAW address =====");
319 if (C->is_dead_node(this->_idx)) tty->print_cr("'this' is dead");
320 if ((ctl != NULL) && C->is_dead_node(ctl->_idx)) tty->print_cr("'ctl' is dead");
321 if (C->is_dead_node(mem->_idx)) tty->print_cr("'mem' is dead");
322 if (C->is_dead_node(address->_idx)) tty->print_cr("'address' is dead");
323 if (C->is_dead_node(base->_idx)) tty->print_cr("'base' is dead");
324 tty->cr();
325 base->dump(1);
326 tty->cr();
327 this->dump(2);
328 tty->print("this->adr_type(): "); adr_type()->dump(); tty->cr();
329 tty->print("phase->type(address): "); t_adr->dump(); tty->cr();
330 tty->print("phase->type(base): "); phase->type(address)->dump(); tty->cr();
331 tty->cr();
332 }
333 assert(base == NULL || t_adr->isa_rawptr() ||
334 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
335 #endif
337 // Avoid independent memory operations
338 Node* old_mem = mem;
340 // The code which unhooks non-raw memories from complete (macro-expanded)
341 // initializations was removed. After macro-expansion all stores catched
342 // by Initialize node became raw stores and there is no information
343 // which memory slices they modify. So it is unsafe to move any memory
344 // operation above these stores. Also in most cases hooked non-raw memories
345 // were already unhooked by using information from detect_ptr_independence()
346 // and find_previous_store().
348 if (mem->is_MergeMem()) {
349 MergeMemNode* mmem = mem->as_MergeMem();
350 const TypePtr *tp = t_adr->is_ptr();
352 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
353 }
355 if (mem != old_mem) {
356 set_req(MemNode::Memory, mem);
357 if (can_reshape && old_mem->outcnt() == 0) {
358 igvn->_worklist.push(old_mem);
359 }
360 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
361 return this;
362 }
364 // let the subclass continue analyzing...
365 return NULL;
366 }
368 // Helper function for proving some simple control dominations.
369 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
370 // Already assumes that 'dom' is available at 'sub', and that 'sub'
371 // is not a constant (dominated by the method's StartNode).
372 // Used by MemNode::find_previous_store to prove that the
373 // control input of a memory operation predates (dominates)
374 // an allocation it wants to look past.
375 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
376 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
377 return false; // Conservative answer for dead code
379 // Check 'dom'. Skip Proj and CatchProj nodes.
380 dom = dom->find_exact_control(dom);
381 if (dom == NULL || dom->is_top())
382 return false; // Conservative answer for dead code
384 if (dom == sub) {
385 // For the case when, for example, 'sub' is Initialize and the original
386 // 'dom' is Proj node of the 'sub'.
387 return false;
388 }
390 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
391 return true;
393 // 'dom' dominates 'sub' if its control edge and control edges
394 // of all its inputs dominate or equal to sub's control edge.
396 // Currently 'sub' is either Allocate, Initialize or Start nodes.
397 // Or Region for the check in LoadNode::Ideal();
398 // 'sub' should have sub->in(0) != NULL.
399 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
400 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
402 // Get control edge of 'sub'.
403 Node* orig_sub = sub;
404 sub = sub->find_exact_control(sub->in(0));
405 if (sub == NULL || sub->is_top())
406 return false; // Conservative answer for dead code
408 assert(sub->is_CFG(), "expecting control");
410 if (sub == dom)
411 return true;
413 if (sub->is_Start() || sub->is_Root())
414 return false;
416 {
417 // Check all control edges of 'dom'.
419 ResourceMark rm;
420 Arena* arena = Thread::current()->resource_area();
421 Node_List nlist(arena);
422 Unique_Node_List dom_list(arena);
424 dom_list.push(dom);
425 bool only_dominating_controls = false;
427 for (uint next = 0; next < dom_list.size(); next++) {
428 Node* n = dom_list.at(next);
429 if (n == orig_sub)
430 return false; // One of dom's inputs dominated by sub.
431 if (!n->is_CFG() && n->pinned()) {
432 // Check only own control edge for pinned non-control nodes.
433 n = n->find_exact_control(n->in(0));
434 if (n == NULL || n->is_top())
435 return false; // Conservative answer for dead code
436 assert(n->is_CFG(), "expecting control");
437 dom_list.push(n);
438 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
439 only_dominating_controls = true;
440 } else if (n->is_CFG()) {
441 if (n->dominates(sub, nlist))
442 only_dominating_controls = true;
443 else
444 return false;
445 } else {
446 // First, own control edge.
447 Node* m = n->find_exact_control(n->in(0));
448 if (m != NULL) {
449 if (m->is_top())
450 return false; // Conservative answer for dead code
451 dom_list.push(m);
452 }
453 // Now, the rest of edges.
454 uint cnt = n->req();
455 for (uint i = 1; i < cnt; i++) {
456 m = n->find_exact_control(n->in(i));
457 if (m == NULL || m->is_top())
458 continue;
459 dom_list.push(m);
460 }
461 }
462 }
463 return only_dominating_controls;
464 }
465 }
467 //---------------------detect_ptr_independence---------------------------------
468 // Used by MemNode::find_previous_store to prove that two base
469 // pointers are never equal.
470 // The pointers are accompanied by their associated allocations,
471 // if any, which have been previously discovered by the caller.
472 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
473 Node* p2, AllocateNode* a2,
474 PhaseTransform* phase) {
475 // Attempt to prove that these two pointers cannot be aliased.
476 // They may both manifestly be allocations, and they should differ.
477 // Or, if they are not both allocations, they can be distinct constants.
478 // Otherwise, one is an allocation and the other a pre-existing value.
479 if (a1 == NULL && a2 == NULL) { // neither an allocation
480 return (p1 != p2) && p1->is_Con() && p2->is_Con();
481 } else if (a1 != NULL && a2 != NULL) { // both allocations
482 return (a1 != a2);
483 } else if (a1 != NULL) { // one allocation a1
484 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
485 return all_controls_dominate(p2, a1);
486 } else { //(a2 != NULL) // one allocation a2
487 return all_controls_dominate(p1, a2);
488 }
489 return false;
490 }
493 // The logic for reordering loads and stores uses four steps:
494 // (a) Walk carefully past stores and initializations which we
495 // can prove are independent of this load.
496 // (b) Observe that the next memory state makes an exact match
497 // with self (load or store), and locate the relevant store.
498 // (c) Ensure that, if we were to wire self directly to the store,
499 // the optimizer would fold it up somehow.
500 // (d) Do the rewiring, and return, depending on some other part of
501 // the optimizer to fold up the load.
502 // This routine handles steps (a) and (b). Steps (c) and (d) are
503 // specific to loads and stores, so they are handled by the callers.
504 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
505 //
506 Node* MemNode::find_previous_store(PhaseTransform* phase) {
507 Node* ctrl = in(MemNode::Control);
508 Node* adr = in(MemNode::Address);
509 intptr_t offset = 0;
510 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
511 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
513 if (offset == Type::OffsetBot)
514 return NULL; // cannot unalias unless there are precise offsets
516 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
518 intptr_t size_in_bytes = memory_size();
520 Node* mem = in(MemNode::Memory); // start searching here...
522 int cnt = 50; // Cycle limiter
523 for (;;) { // While we can dance past unrelated stores...
524 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
526 if (mem->is_Store()) {
527 Node* st_adr = mem->in(MemNode::Address);
528 intptr_t st_offset = 0;
529 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
530 if (st_base == NULL)
531 break; // inscrutable pointer
532 if (st_offset != offset && st_offset != Type::OffsetBot) {
533 const int MAX_STORE = BytesPerLong;
534 if (st_offset >= offset + size_in_bytes ||
535 st_offset <= offset - MAX_STORE ||
536 st_offset <= offset - mem->as_Store()->memory_size()) {
537 // Success: The offsets are provably independent.
538 // (You may ask, why not just test st_offset != offset and be done?
539 // The answer is that stores of different sizes can co-exist
540 // in the same sequence of RawMem effects. We sometimes initialize
541 // a whole 'tile' of array elements with a single jint or jlong.)
542 mem = mem->in(MemNode::Memory);
543 continue; // (a) advance through independent store memory
544 }
545 }
546 if (st_base != base &&
547 detect_ptr_independence(base, alloc,
548 st_base,
549 AllocateNode::Ideal_allocation(st_base, phase),
550 phase)) {
551 // Success: The bases are provably independent.
552 mem = mem->in(MemNode::Memory);
553 continue; // (a) advance through independent store memory
554 }
556 // (b) At this point, if the bases or offsets do not agree, we lose,
557 // since we have not managed to prove 'this' and 'mem' independent.
558 if (st_base == base && st_offset == offset) {
559 return mem; // let caller handle steps (c), (d)
560 }
562 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
563 InitializeNode* st_init = mem->in(0)->as_Initialize();
564 AllocateNode* st_alloc = st_init->allocation();
565 if (st_alloc == NULL)
566 break; // something degenerated
567 bool known_identical = false;
568 bool known_independent = false;
569 if (alloc == st_alloc)
570 known_identical = true;
571 else if (alloc != NULL)
572 known_independent = true;
573 else if (all_controls_dominate(this, st_alloc))
574 known_independent = true;
576 if (known_independent) {
577 // The bases are provably independent: Either they are
578 // manifestly distinct allocations, or else the control
579 // of this load dominates the store's allocation.
580 int alias_idx = phase->C->get_alias_index(adr_type());
581 if (alias_idx == Compile::AliasIdxRaw) {
582 mem = st_alloc->in(TypeFunc::Memory);
583 } else {
584 mem = st_init->memory(alias_idx);
585 }
586 continue; // (a) advance through independent store memory
587 }
589 // (b) at this point, if we are not looking at a store initializing
590 // the same allocation we are loading from, we lose.
591 if (known_identical) {
592 // From caller, can_see_stored_value will consult find_captured_store.
593 return mem; // let caller handle steps (c), (d)
594 }
596 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
597 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
598 if (mem->is_Proj() && mem->in(0)->is_Call()) {
599 CallNode *call = mem->in(0)->as_Call();
600 if (!call->may_modify(addr_t, phase)) {
601 mem = call->in(TypeFunc::Memory);
602 continue; // (a) advance through independent call memory
603 }
604 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
605 mem = mem->in(0)->in(TypeFunc::Memory);
606 continue; // (a) advance through independent MemBar memory
607 } else if (mem->is_ClearArray()) {
608 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
609 // (the call updated 'mem' value)
610 continue; // (a) advance through independent allocation memory
611 } else {
612 // Can not bypass initialization of the instance
613 // we are looking for.
614 return mem;
615 }
616 } else if (mem->is_MergeMem()) {
617 int alias_idx = phase->C->get_alias_index(adr_type());
618 mem = mem->as_MergeMem()->memory_at(alias_idx);
619 continue; // (a) advance through independent MergeMem memory
620 }
621 }
623 // Unless there is an explicit 'continue', we must bail out here,
624 // because 'mem' is an inscrutable memory state (e.g., a call).
625 break;
626 }
628 return NULL; // bail out
629 }
631 //----------------------calculate_adr_type-------------------------------------
632 // Helper function. Notices when the given type of address hits top or bottom.
633 // Also, asserts a cross-check of the type against the expected address type.
634 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
635 if (t == Type::TOP) return NULL; // does not touch memory any more?
636 #ifdef PRODUCT
637 cross_check = NULL;
638 #else
639 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
640 #endif
641 const TypePtr* tp = t->isa_ptr();
642 if (tp == NULL) {
643 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
644 return TypePtr::BOTTOM; // touches lots of memory
645 } else {
646 #ifdef ASSERT
647 // %%%% [phh] We don't check the alias index if cross_check is
648 // TypeRawPtr::BOTTOM. Needs to be investigated.
649 if (cross_check != NULL &&
650 cross_check != TypePtr::BOTTOM &&
651 cross_check != TypeRawPtr::BOTTOM) {
652 // Recheck the alias index, to see if it has changed (due to a bug).
653 Compile* C = Compile::current();
654 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
655 "must stay in the original alias category");
656 // The type of the address must be contained in the adr_type,
657 // disregarding "null"-ness.
658 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
659 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
660 assert(cross_check->meet(tp_notnull) == cross_check,
661 "real address must not escape from expected memory type");
662 }
663 #endif
664 return tp;
665 }
666 }
668 //------------------------adr_phi_is_loop_invariant----------------------------
669 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
670 // loop is loop invariant. Make a quick traversal of Phi and associated
671 // CastPP nodes, looking to see if they are a closed group within the loop.
672 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
673 // The idea is that the phi-nest must boil down to only CastPP nodes
674 // with the same data. This implies that any path into the loop already
675 // includes such a CastPP, and so the original cast, whatever its input,
676 // must be covered by an equivalent cast, with an earlier control input.
677 ResourceMark rm;
679 // The loop entry input of the phi should be the unique dominating
680 // node for every Phi/CastPP in the loop.
681 Unique_Node_List closure;
682 closure.push(adr_phi->in(LoopNode::EntryControl));
684 // Add the phi node and the cast to the worklist.
685 Unique_Node_List worklist;
686 worklist.push(adr_phi);
687 if( cast != NULL ){
688 if( !cast->is_ConstraintCast() ) return false;
689 worklist.push(cast);
690 }
692 // Begin recursive walk of phi nodes.
693 while( worklist.size() ){
694 // Take a node off the worklist
695 Node *n = worklist.pop();
696 if( !closure.member(n) ){
697 // Add it to the closure.
698 closure.push(n);
699 // Make a sanity check to ensure we don't waste too much time here.
700 if( closure.size() > 20) return false;
701 // This node is OK if:
702 // - it is a cast of an identical value
703 // - or it is a phi node (then we add its inputs to the worklist)
704 // Otherwise, the node is not OK, and we presume the cast is not invariant
705 if( n->is_ConstraintCast() ){
706 worklist.push(n->in(1));
707 } else if( n->is_Phi() ) {
708 for( uint i = 1; i < n->req(); i++ ) {
709 worklist.push(n->in(i));
710 }
711 } else {
712 return false;
713 }
714 }
715 }
717 // Quit when the worklist is empty, and we've found no offending nodes.
718 return true;
719 }
721 //------------------------------Ideal_DU_postCCP-------------------------------
722 // Find any cast-away of null-ness and keep its control. Null cast-aways are
723 // going away in this pass and we need to make this memory op depend on the
724 // gating null check.
725 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
726 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
727 }
729 // I tried to leave the CastPP's in. This makes the graph more accurate in
730 // some sense; we get to keep around the knowledge that an oop is not-null
731 // after some test. Alas, the CastPP's interfere with GVN (some values are
732 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
733 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
734 // some of the more trivial cases in the optimizer. Removing more useless
735 // Phi's started allowing Loads to illegally float above null checks. I gave
736 // up on this approach. CNC 10/20/2000
737 // This static method may be called not from MemNode (EncodePNode calls it).
738 // Only the control edge of the node 'n' might be updated.
739 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
740 Node *skipped_cast = NULL;
741 // Need a null check? Regular static accesses do not because they are
742 // from constant addresses. Array ops are gated by the range check (which
743 // always includes a NULL check). Just check field ops.
744 if( n->in(MemNode::Control) == NULL ) {
745 // Scan upwards for the highest location we can place this memory op.
746 while( true ) {
747 switch( adr->Opcode() ) {
749 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
750 adr = adr->in(AddPNode::Base);
751 continue;
753 case Op_DecodeN: // No change to NULL-ness, so peek thru
754 case Op_DecodeNKlass:
755 adr = adr->in(1);
756 continue;
758 case Op_EncodeP:
759 case Op_EncodePKlass:
760 // EncodeP node's control edge could be set by this method
761 // when EncodeP node depends on CastPP node.
762 //
763 // Use its control edge for memory op because EncodeP may go away
764 // later when it is folded with following or preceding DecodeN node.
765 if (adr->in(0) == NULL) {
766 // Keep looking for cast nodes.
767 adr = adr->in(1);
768 continue;
769 }
770 ccp->hash_delete(n);
771 n->set_req(MemNode::Control, adr->in(0));
772 ccp->hash_insert(n);
773 return n;
775 case Op_CastPP:
776 // If the CastPP is useless, just peek on through it.
777 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
778 // Remember the cast that we've peeked though. If we peek
779 // through more than one, then we end up remembering the highest
780 // one, that is, if in a loop, the one closest to the top.
781 skipped_cast = adr;
782 adr = adr->in(1);
783 continue;
784 }
785 // CastPP is going away in this pass! We need this memory op to be
786 // control-dependent on the test that is guarding the CastPP.
787 ccp->hash_delete(n);
788 n->set_req(MemNode::Control, adr->in(0));
789 ccp->hash_insert(n);
790 return n;
792 case Op_Phi:
793 // Attempt to float above a Phi to some dominating point.
794 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
795 // If we've already peeked through a Cast (which could have set the
796 // control), we can't float above a Phi, because the skipped Cast
797 // may not be loop invariant.
798 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
799 adr = adr->in(1);
800 continue;
801 }
802 }
804 // Intentional fallthrough!
806 // No obvious dominating point. The mem op is pinned below the Phi
807 // by the Phi itself. If the Phi goes away (no true value is merged)
808 // then the mem op can float, but not indefinitely. It must be pinned
809 // behind the controls leading to the Phi.
810 case Op_CheckCastPP:
811 // These usually stick around to change address type, however a
812 // useless one can be elided and we still need to pick up a control edge
813 if (adr->in(0) == NULL) {
814 // This CheckCastPP node has NO control and is likely useless. But we
815 // need check further up the ancestor chain for a control input to keep
816 // the node in place. 4959717.
817 skipped_cast = adr;
818 adr = adr->in(1);
819 continue;
820 }
821 ccp->hash_delete(n);
822 n->set_req(MemNode::Control, adr->in(0));
823 ccp->hash_insert(n);
824 return n;
826 // List of "safe" opcodes; those that implicitly block the memory
827 // op below any null check.
828 case Op_CastX2P: // no null checks on native pointers
829 case Op_Parm: // 'this' pointer is not null
830 case Op_LoadP: // Loading from within a klass
831 case Op_LoadN: // Loading from within a klass
832 case Op_LoadKlass: // Loading from within a klass
833 case Op_LoadNKlass: // Loading from within a klass
834 case Op_ConP: // Loading from a klass
835 case Op_ConN: // Loading from a klass
836 case Op_ConNKlass: // Loading from a klass
837 case Op_CreateEx: // Sucking up the guts of an exception oop
838 case Op_Con: // Reading from TLS
839 case Op_CMoveP: // CMoveP is pinned
840 case Op_CMoveN: // CMoveN is pinned
841 break; // No progress
843 case Op_Proj: // Direct call to an allocation routine
844 case Op_SCMemProj: // Memory state from store conditional ops
845 #ifdef ASSERT
846 {
847 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
848 const Node* call = adr->in(0);
849 if (call->is_CallJava()) {
850 const CallJavaNode* call_java = call->as_CallJava();
851 const TypeTuple *r = call_java->tf()->range();
852 assert(r->cnt() > TypeFunc::Parms, "must return value");
853 const Type* ret_type = r->field_at(TypeFunc::Parms);
854 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
855 // We further presume that this is one of
856 // new_instance_Java, new_array_Java, or
857 // the like, but do not assert for this.
858 } else if (call->is_Allocate()) {
859 // similar case to new_instance_Java, etc.
860 } else if (!call->is_CallLeaf()) {
861 // Projections from fetch_oop (OSR) are allowed as well.
862 ShouldNotReachHere();
863 }
864 }
865 #endif
866 break;
867 default:
868 ShouldNotReachHere();
869 }
870 break;
871 }
872 }
874 return NULL; // No progress
875 }
878 //=============================================================================
879 uint LoadNode::size_of() const { return sizeof(*this); }
880 uint LoadNode::cmp( const Node &n ) const
881 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
882 const Type *LoadNode::bottom_type() const { return _type; }
883 uint LoadNode::ideal_reg() const {
884 return _type->ideal_reg();
885 }
887 #ifndef PRODUCT
888 void LoadNode::dump_spec(outputStream *st) const {
889 MemNode::dump_spec(st);
890 if( !Verbose && !WizardMode ) {
891 // standard dump does this in Verbose and WizardMode
892 st->print(" #"); _type->dump_on(st);
893 }
894 }
895 #endif
897 #ifdef ASSERT
898 //----------------------------is_immutable_value-------------------------------
899 // Helper function to allow a raw load without control edge for some cases
900 bool LoadNode::is_immutable_value(Node* adr) {
901 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
902 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
903 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
904 in_bytes(JavaThread::osthread_offset())));
905 }
906 #endif
908 //----------------------------LoadNode::make-----------------------------------
909 // Polymorphic factory method:
910 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo) {
911 Compile* C = gvn.C;
913 // sanity check the alias category against the created node type
914 assert(!(adr_type->isa_oopptr() &&
915 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
916 "use LoadKlassNode instead");
917 assert(!(adr_type->isa_aryptr() &&
918 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
919 "use LoadRangeNode instead");
920 // Check control edge of raw loads
921 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
922 // oop will be recorded in oop map if load crosses safepoint
923 rt->isa_oopptr() || is_immutable_value(adr),
924 "raw memory operations should have control edge");
925 switch (bt) {
926 case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo);
927 case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo);
928 case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo);
929 case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo);
930 case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo);
931 case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo);
932 case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt, mo);
933 case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt, mo);
934 case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo);
935 case T_OBJECT:
936 #ifdef _LP64
937 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
938 Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo));
939 return new (C) DecodeNNode(load, load->bottom_type()->make_ptr());
940 } else
941 #endif
942 {
943 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
944 return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo);
945 }
946 }
947 ShouldNotReachHere();
948 return (LoadNode*)NULL;
949 }
951 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) {
952 bool require_atomic = true;
953 return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, require_atomic);
954 }
959 //------------------------------hash-------------------------------------------
960 uint LoadNode::hash() const {
961 // unroll addition of interesting fields
962 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
963 }
965 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
966 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
967 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
968 bool is_stable_ary = FoldStableValues &&
969 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
970 tp->isa_aryptr()->is_stable();
972 return (eliminate_boxing && non_volatile) || is_stable_ary;
973 }
975 return false;
976 }
978 //---------------------------can_see_stored_value------------------------------
979 // This routine exists to make sure this set of tests is done the same
980 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
981 // will change the graph shape in a way which makes memory alive twice at the
982 // same time (uses the Oracle model of aliasing), then some
983 // LoadXNode::Identity will fold things back to the equivalence-class model
984 // of aliasing.
985 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
986 Node* ld_adr = in(MemNode::Address);
987 intptr_t ld_off = 0;
988 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
989 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
990 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
991 // This is more general than load from boxing objects.
992 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
993 uint alias_idx = atp->index();
994 bool final = !atp->is_rewritable();
995 Node* result = NULL;
996 Node* current = st;
997 // Skip through chains of MemBarNodes checking the MergeMems for
998 // new states for the slice of this load. Stop once any other
999 // kind of node is encountered. Loads from final memory can skip
1000 // through any kind of MemBar but normal loads shouldn't skip
1001 // through MemBarAcquire since the could allow them to move out of
1002 // a synchronized region.
1003 while (current->is_Proj()) {
1004 int opc = current->in(0)->Opcode();
1005 if ((final && (opc == Op_MemBarAcquire ||
1006 opc == Op_MemBarAcquireLock ||
1007 opc == Op_LoadFence)) ||
1008 opc == Op_MemBarRelease ||
1009 opc == Op_StoreFence ||
1010 opc == Op_MemBarReleaseLock ||
1011 opc == Op_MemBarCPUOrder) {
1012 Node* mem = current->in(0)->in(TypeFunc::Memory);
1013 if (mem->is_MergeMem()) {
1014 MergeMemNode* merge = mem->as_MergeMem();
1015 Node* new_st = merge->memory_at(alias_idx);
1016 if (new_st == merge->base_memory()) {
1017 // Keep searching
1018 current = new_st;
1019 continue;
1020 }
1021 // Save the new memory state for the slice and fall through
1022 // to exit.
1023 result = new_st;
1024 }
1025 }
1026 break;
1027 }
1028 if (result != NULL) {
1029 st = result;
1030 }
1031 }
1033 // Loop around twice in the case Load -> Initialize -> Store.
1034 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1035 for (int trip = 0; trip <= 1; trip++) {
1037 if (st->is_Store()) {
1038 Node* st_adr = st->in(MemNode::Address);
1039 if (!phase->eqv(st_adr, ld_adr)) {
1040 // Try harder before giving up... Match raw and non-raw pointers.
1041 intptr_t st_off = 0;
1042 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1043 if (alloc == NULL) return NULL;
1044 if (alloc != ld_alloc) return NULL;
1045 if (ld_off != st_off) return NULL;
1046 // At this point we have proven something like this setup:
1047 // A = Allocate(...)
1048 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
1049 // S = StoreQ(, AddP(, A.Parm , #Off), V)
1050 // (Actually, we haven't yet proven the Q's are the same.)
1051 // In other words, we are loading from a casted version of
1052 // the same pointer-and-offset that we stored to.
1053 // Thus, we are able to replace L by V.
1054 }
1055 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1056 if (store_Opcode() != st->Opcode())
1057 return NULL;
1058 return st->in(MemNode::ValueIn);
1059 }
1061 // A load from a freshly-created object always returns zero.
1062 // (This can happen after LoadNode::Ideal resets the load's memory input
1063 // to find_captured_store, which returned InitializeNode::zero_memory.)
1064 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1065 (st->in(0) == ld_alloc) &&
1066 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1067 // return a zero value for the load's basic type
1068 // (This is one of the few places where a generic PhaseTransform
1069 // can create new nodes. Think of it as lazily manifesting
1070 // virtually pre-existing constants.)
1071 return phase->zerocon(memory_type());
1072 }
1074 // A load from an initialization barrier can match a captured store.
1075 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1076 InitializeNode* init = st->in(0)->as_Initialize();
1077 AllocateNode* alloc = init->allocation();
1078 if ((alloc != NULL) && (alloc == ld_alloc)) {
1079 // examine a captured store value
1080 st = init->find_captured_store(ld_off, memory_size(), phase);
1081 if (st != NULL)
1082 continue; // take one more trip around
1083 }
1084 }
1086 // Load boxed value from result of valueOf() call is input parameter.
1087 if (this->is_Load() && ld_adr->is_AddP() &&
1088 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1089 intptr_t ignore = 0;
1090 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1091 if (base != NULL && base->is_Proj() &&
1092 base->as_Proj()->_con == TypeFunc::Parms &&
1093 base->in(0)->is_CallStaticJava() &&
1094 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1095 return base->in(0)->in(TypeFunc::Parms);
1096 }
1097 }
1099 break;
1100 }
1102 return NULL;
1103 }
1105 //----------------------is_instance_field_load_with_local_phi------------------
1106 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1107 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1108 in(Address)->is_AddP() ) {
1109 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1110 // Only instances and boxed values.
1111 if( t_oop != NULL &&
1112 (t_oop->is_ptr_to_boxed_value() ||
1113 t_oop->is_known_instance_field()) &&
1114 t_oop->offset() != Type::OffsetBot &&
1115 t_oop->offset() != Type::OffsetTop) {
1116 return true;
1117 }
1118 }
1119 return false;
1120 }
1122 //------------------------------Identity---------------------------------------
1123 // Loads are identity if previous store is to same address
1124 Node *LoadNode::Identity( PhaseTransform *phase ) {
1125 // If the previous store-maker is the right kind of Store, and the store is
1126 // to the same address, then we are equal to the value stored.
1127 Node* mem = in(Memory);
1128 Node* value = can_see_stored_value(mem, phase);
1129 if( value ) {
1130 // byte, short & char stores truncate naturally.
1131 // A load has to load the truncated value which requires
1132 // some sort of masking operation and that requires an
1133 // Ideal call instead of an Identity call.
1134 if (memory_size() < BytesPerInt) {
1135 // If the input to the store does not fit with the load's result type,
1136 // it must be truncated via an Ideal call.
1137 if (!phase->type(value)->higher_equal(phase->type(this)))
1138 return this;
1139 }
1140 // (This works even when value is a Con, but LoadNode::Value
1141 // usually runs first, producing the singleton type of the Con.)
1142 return value;
1143 }
1145 // Search for an existing data phi which was generated before for the same
1146 // instance's field to avoid infinite generation of phis in a loop.
1147 Node *region = mem->in(0);
1148 if (is_instance_field_load_with_local_phi(region)) {
1149 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1150 int this_index = phase->C->get_alias_index(addr_t);
1151 int this_offset = addr_t->offset();
1152 int this_iid = addr_t->instance_id();
1153 if (!addr_t->is_known_instance() &&
1154 addr_t->is_ptr_to_boxed_value()) {
1155 // Use _idx of address base (could be Phi node) for boxed values.
1156 intptr_t ignore = 0;
1157 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1158 this_iid = base->_idx;
1159 }
1160 const Type* this_type = bottom_type();
1161 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1162 Node* phi = region->fast_out(i);
1163 if (phi->is_Phi() && phi != mem &&
1164 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {
1165 return phi;
1166 }
1167 }
1168 }
1170 return this;
1171 }
1173 // We're loading from an object which has autobox behaviour.
1174 // If this object is result of a valueOf call we'll have a phi
1175 // merging a newly allocated object and a load from the cache.
1176 // We want to replace this load with the original incoming
1177 // argument to the valueOf call.
1178 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1179 assert(phase->C->eliminate_boxing(), "sanity");
1180 intptr_t ignore = 0;
1181 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1182 if ((base == NULL) || base->is_Phi()) {
1183 // Push the loads from the phi that comes from valueOf up
1184 // through it to allow elimination of the loads and the recovery
1185 // of the original value. It is done in split_through_phi().
1186 return NULL;
1187 } else if (base->is_Load() ||
1188 base->is_DecodeN() && base->in(1)->is_Load()) {
1189 // Eliminate the load of boxed value for integer types from the cache
1190 // array by deriving the value from the index into the array.
1191 // Capture the offset of the load and then reverse the computation.
1193 // Get LoadN node which loads a boxing object from 'cache' array.
1194 if (base->is_DecodeN()) {
1195 base = base->in(1);
1196 }
1197 if (!base->in(Address)->is_AddP()) {
1198 return NULL; // Complex address
1199 }
1200 AddPNode* address = base->in(Address)->as_AddP();
1201 Node* cache_base = address->in(AddPNode::Base);
1202 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1203 // Get ConP node which is static 'cache' field.
1204 cache_base = cache_base->in(1);
1205 }
1206 if ((cache_base != NULL) && cache_base->is_Con()) {
1207 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1208 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1209 Node* elements[4];
1210 int shift = exact_log2(type2aelembytes(T_OBJECT));
1211 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1212 if ((count > 0) && elements[0]->is_Con() &&
1213 ((count == 1) ||
1214 (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1215 elements[1]->in(2) == phase->intcon(shift))) {
1216 ciObjArray* array = base_type->const_oop()->as_obj_array();
1217 // Fetch the box object cache[0] at the base of the array and get its value
1218 ciInstance* box = array->obj_at(0)->as_instance();
1219 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1220 assert(ik->is_box_klass(), "sanity");
1221 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1222 if (ik->nof_nonstatic_fields() == 1) {
1223 // This should be true nonstatic_field_at requires calling
1224 // nof_nonstatic_fields so check it anyway
1225 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1226 BasicType bt = c.basic_type();
1227 // Only integer types have boxing cache.
1228 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1229 bt == T_BYTE || bt == T_SHORT ||
1230 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt)));
1231 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1232 if (cache_low != (int)cache_low) {
1233 return NULL; // should not happen since cache is array indexed by value
1234 }
1235 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1236 if (offset != (int)offset) {
1237 return NULL; // should not happen since cache is array indexed by value
1238 }
1239 // Add up all the offsets making of the address of the load
1240 Node* result = elements[0];
1241 for (int i = 1; i < count; i++) {
1242 result = phase->transform(new (phase->C) AddXNode(result, elements[i]));
1243 }
1244 // Remove the constant offset from the address and then
1245 result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset)));
1246 // remove the scaling of the offset to recover the original index.
1247 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1248 // Peel the shift off directly but wrap it in a dummy node
1249 // since Ideal can't return existing nodes
1250 result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));
1251 } else if (result->is_Add() && result->in(2)->is_Con() &&
1252 result->in(1)->Opcode() == Op_LShiftX &&
1253 result->in(1)->in(2) == phase->intcon(shift)) {
1254 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1255 // but for boxing cache access we know that X<<Z will not overflow
1256 // (there is range check) so we do this optimizatrion by hand here.
1257 Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift));
1258 result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con));
1259 } else {
1260 result = new (phase->C) RShiftXNode(result, phase->intcon(shift));
1261 }
1262 #ifdef _LP64
1263 if (bt != T_LONG) {
1264 result = new (phase->C) ConvL2INode(phase->transform(result));
1265 }
1266 #else
1267 if (bt == T_LONG) {
1268 result = new (phase->C) ConvI2LNode(phase->transform(result));
1269 }
1270 #endif
1271 return result;
1272 }
1273 }
1274 }
1275 }
1276 }
1277 return NULL;
1278 }
1280 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1281 Node* region = phi->in(0);
1282 if (region == NULL) {
1283 return false; // Wait stable graph
1284 }
1285 uint cnt = phi->req();
1286 for (uint i = 1; i < cnt; i++) {
1287 Node* rc = region->in(i);
1288 if (rc == NULL || phase->type(rc) == Type::TOP)
1289 return false; // Wait stable graph
1290 Node* in = phi->in(i);
1291 if (in == NULL || phase->type(in) == Type::TOP)
1292 return false; // Wait stable graph
1293 }
1294 return true;
1295 }
1296 //------------------------------split_through_phi------------------------------
1297 // Split instance or boxed field load through Phi.
1298 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1299 Node* mem = in(Memory);
1300 Node* address = in(Address);
1301 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1303 assert((t_oop != NULL) &&
1304 (t_oop->is_known_instance_field() ||
1305 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1307 Compile* C = phase->C;
1308 intptr_t ignore = 0;
1309 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1310 bool base_is_phi = (base != NULL) && base->is_Phi();
1311 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1312 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1313 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1315 if (!((mem->is_Phi() || base_is_phi) &&
1316 (load_boxed_values || t_oop->is_known_instance_field()))) {
1317 return NULL; // memory is not Phi
1318 }
1320 if (mem->is_Phi()) {
1321 if (!stable_phi(mem->as_Phi(), phase)) {
1322 return NULL; // Wait stable graph
1323 }
1324 uint cnt = mem->req();
1325 // Check for loop invariant memory.
1326 if (cnt == 3) {
1327 for (uint i = 1; i < cnt; i++) {
1328 Node* in = mem->in(i);
1329 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1330 if (m == mem) {
1331 set_req(Memory, mem->in(cnt - i));
1332 return this; // made change
1333 }
1334 }
1335 }
1336 }
1337 if (base_is_phi) {
1338 if (!stable_phi(base->as_Phi(), phase)) {
1339 return NULL; // Wait stable graph
1340 }
1341 uint cnt = base->req();
1342 // Check for loop invariant memory.
1343 if (cnt == 3) {
1344 for (uint i = 1; i < cnt; i++) {
1345 if (base->in(i) == base) {
1346 return NULL; // Wait stable graph
1347 }
1348 }
1349 }
1350 }
1352 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1354 // Split through Phi (see original code in loopopts.cpp).
1355 assert(C->have_alias_type(t_oop), "instance should have alias type");
1357 // Do nothing here if Identity will find a value
1358 // (to avoid infinite chain of value phis generation).
1359 if (!phase->eqv(this, this->Identity(phase)))
1360 return NULL;
1362 // Select Region to split through.
1363 Node* region;
1364 if (!base_is_phi) {
1365 assert(mem->is_Phi(), "sanity");
1366 region = mem->in(0);
1367 // Skip if the region dominates some control edge of the address.
1368 if (!MemNode::all_controls_dominate(address, region))
1369 return NULL;
1370 } else if (!mem->is_Phi()) {
1371 assert(base_is_phi, "sanity");
1372 region = base->in(0);
1373 // Skip if the region dominates some control edge of the memory.
1374 if (!MemNode::all_controls_dominate(mem, region))
1375 return NULL;
1376 } else if (base->in(0) != mem->in(0)) {
1377 assert(base_is_phi && mem->is_Phi(), "sanity");
1378 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1379 region = base->in(0);
1380 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1381 region = mem->in(0);
1382 } else {
1383 return NULL; // complex graph
1384 }
1385 } else {
1386 assert(base->in(0) == mem->in(0), "sanity");
1387 region = mem->in(0);
1388 }
1390 const Type* this_type = this->bottom_type();
1391 int this_index = C->get_alias_index(t_oop);
1392 int this_offset = t_oop->offset();
1393 int this_iid = t_oop->instance_id();
1394 if (!t_oop->is_known_instance() && load_boxed_values) {
1395 // Use _idx of address base for boxed values.
1396 this_iid = base->_idx;
1397 }
1398 PhaseIterGVN* igvn = phase->is_IterGVN();
1399 Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1400 for (uint i = 1; i < region->req(); i++) {
1401 Node* x;
1402 Node* the_clone = NULL;
1403 if (region->in(i) == C->top()) {
1404 x = C->top(); // Dead path? Use a dead data op
1405 } else {
1406 x = this->clone(); // Else clone up the data op
1407 the_clone = x; // Remember for possible deletion.
1408 // Alter data node to use pre-phi inputs
1409 if (this->in(0) == region) {
1410 x->set_req(0, region->in(i));
1411 } else {
1412 x->set_req(0, NULL);
1413 }
1414 if (mem->is_Phi() && (mem->in(0) == region)) {
1415 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1416 }
1417 if (address->is_Phi() && address->in(0) == region) {
1418 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1419 }
1420 if (base_is_phi && (base->in(0) == region)) {
1421 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1422 Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1423 x->set_req(Address, adr_x);
1424 }
1425 }
1426 // Check for a 'win' on some paths
1427 const Type *t = x->Value(igvn);
1429 bool singleton = t->singleton();
1431 // See comments in PhaseIdealLoop::split_thru_phi().
1432 if (singleton && t == Type::TOP) {
1433 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1434 }
1436 if (singleton) {
1437 x = igvn->makecon(t);
1438 } else {
1439 // We now call Identity to try to simplify the cloned node.
1440 // Note that some Identity methods call phase->type(this).
1441 // Make sure that the type array is big enough for
1442 // our new node, even though we may throw the node away.
1443 // (This tweaking with igvn only works because x is a new node.)
1444 igvn->set_type(x, t);
1445 // If x is a TypeNode, capture any more-precise type permanently into Node
1446 // otherwise it will be not updated during igvn->transform since
1447 // igvn->type(x) is set to x->Value() already.
1448 x->raise_bottom_type(t);
1449 Node *y = x->Identity(igvn);
1450 if (y != x) {
1451 x = y;
1452 } else {
1453 y = igvn->hash_find_insert(x);
1454 if (y) {
1455 x = y;
1456 } else {
1457 // Else x is a new node we are keeping
1458 // We do not need register_new_node_with_optimizer
1459 // because set_type has already been called.
1460 igvn->_worklist.push(x);
1461 }
1462 }
1463 }
1464 if (x != the_clone && the_clone != NULL) {
1465 igvn->remove_dead_node(the_clone);
1466 }
1467 phi->set_req(i, x);
1468 }
1469 // Record Phi
1470 igvn->register_new_node_with_optimizer(phi);
1471 return phi;
1472 }
1474 //------------------------------Ideal------------------------------------------
1475 // If the load is from Field memory and the pointer is non-null, we can
1476 // zero out the control input.
1477 // If the offset is constant and the base is an object allocation,
1478 // try to hook me up to the exact initializing store.
1479 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1480 Node* p = MemNode::Ideal_common(phase, can_reshape);
1481 if (p) return (p == NodeSentinel) ? NULL : p;
1483 Node* ctrl = in(MemNode::Control);
1484 Node* address = in(MemNode::Address);
1486 // Skip up past a SafePoint control. Cannot do this for Stores because
1487 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1488 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1489 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1490 ctrl = ctrl->in(0);
1491 set_req(MemNode::Control,ctrl);
1492 }
1494 intptr_t ignore = 0;
1495 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1496 if (base != NULL
1497 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1498 // Check for useless control edge in some common special cases
1499 if (in(MemNode::Control) != NULL
1500 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1501 && all_controls_dominate(base, phase->C->start())) {
1502 // A method-invariant, non-null address (constant or 'this' argument).
1503 set_req(MemNode::Control, NULL);
1504 }
1505 }
1507 Node* mem = in(MemNode::Memory);
1508 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1510 if (can_reshape && (addr_t != NULL)) {
1511 // try to optimize our memory input
1512 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1513 if (opt_mem != mem) {
1514 set_req(MemNode::Memory, opt_mem);
1515 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1516 return this;
1517 }
1518 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1519 if ((t_oop != NULL) &&
1520 (t_oop->is_known_instance_field() ||
1521 t_oop->is_ptr_to_boxed_value())) {
1522 PhaseIterGVN *igvn = phase->is_IterGVN();
1523 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1524 // Delay this transformation until memory Phi is processed.
1525 phase->is_IterGVN()->_worklist.push(this);
1526 return NULL;
1527 }
1528 // Split instance field load through Phi.
1529 Node* result = split_through_phi(phase);
1530 if (result != NULL) return result;
1532 if (t_oop->is_ptr_to_boxed_value()) {
1533 Node* result = eliminate_autobox(phase);
1534 if (result != NULL) return result;
1535 }
1536 }
1537 }
1539 // Check for prior store with a different base or offset; make Load
1540 // independent. Skip through any number of them. Bail out if the stores
1541 // are in an endless dead cycle and report no progress. This is a key
1542 // transform for Reflection. However, if after skipping through the Stores
1543 // we can't then fold up against a prior store do NOT do the transform as
1544 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1545 // array memory alive twice: once for the hoisted Load and again after the
1546 // bypassed Store. This situation only works if EVERYBODY who does
1547 // anti-dependence work knows how to bypass. I.e. we need all
1548 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1549 // the alias index stuff. So instead, peek through Stores and IFF we can
1550 // fold up, do so.
1551 Node* prev_mem = find_previous_store(phase);
1552 // Steps (a), (b): Walk past independent stores to find an exact match.
1553 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1554 // (c) See if we can fold up on the spot, but don't fold up here.
1555 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1556 // just return a prior value, which is done by Identity calls.
1557 if (can_see_stored_value(prev_mem, phase)) {
1558 // Make ready for step (d):
1559 set_req(MemNode::Memory, prev_mem);
1560 return this;
1561 }
1562 }
1564 return NULL; // No further progress
1565 }
1567 // Helper to recognize certain Klass fields which are invariant across
1568 // some group of array types (e.g., int[] or all T[] where T < Object).
1569 const Type*
1570 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1571 ciKlass* klass) const {
1572 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1573 // The field is Klass::_modifier_flags. Return its (constant) value.
1574 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1575 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1576 return TypeInt::make(klass->modifier_flags());
1577 }
1578 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1579 // The field is Klass::_access_flags. Return its (constant) value.
1580 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1581 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1582 return TypeInt::make(klass->access_flags());
1583 }
1584 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1585 // The field is Klass::_layout_helper. Return its constant value if known.
1586 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1587 return TypeInt::make(klass->layout_helper());
1588 }
1590 // No match.
1591 return NULL;
1592 }
1594 // Try to constant-fold a stable array element.
1595 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) {
1596 assert(ary->is_stable(), "array should be stable");
1598 if (ary->const_oop() != NULL) {
1599 // Decode the results of GraphKit::array_element_address.
1600 ciArray* aobj = ary->const_oop()->as_array();
1601 ciConstant con = aobj->element_value_by_offset(off);
1603 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) {
1604 const Type* con_type = Type::make_from_constant(con);
1605 if (con_type != NULL) {
1606 if (con_type->isa_aryptr()) {
1607 // Join with the array element type, in case it is also stable.
1608 int dim = ary->stable_dimension();
1609 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1);
1610 }
1611 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) {
1612 con_type = con_type->make_narrowoop();
1613 }
1614 #ifndef PRODUCT
1615 if (TraceIterativeGVN) {
1616 tty->print("FoldStableValues: array element [off=%d]: con_type=", off);
1617 con_type->dump(); tty->cr();
1618 }
1619 #endif //PRODUCT
1620 return con_type;
1621 }
1622 }
1623 }
1625 return NULL;
1626 }
1628 //------------------------------Value-----------------------------------------
1629 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1630 // Either input is TOP ==> the result is TOP
1631 Node* mem = in(MemNode::Memory);
1632 const Type *t1 = phase->type(mem);
1633 if (t1 == Type::TOP) return Type::TOP;
1634 Node* adr = in(MemNode::Address);
1635 const TypePtr* tp = phase->type(adr)->isa_ptr();
1636 if (tp == NULL || tp->empty()) return Type::TOP;
1637 int off = tp->offset();
1638 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1639 Compile* C = phase->C;
1641 // Try to guess loaded type from pointer type
1642 if (tp->isa_aryptr()) {
1643 const TypeAryPtr* ary = tp->is_aryptr();
1644 const Type *t = ary->elem();
1646 // Determine whether the reference is beyond the header or not, by comparing
1647 // the offset against the offset of the start of the array's data.
1648 // Different array types begin at slightly different offsets (12 vs. 16).
1649 // We choose T_BYTE as an example base type that is least restrictive
1650 // as to alignment, which will therefore produce the smallest
1651 // possible base offset.
1652 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1653 const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1655 // Try to constant-fold a stable array element.
1656 if (FoldStableValues && ary->is_stable()) {
1657 // Make sure the reference is not into the header
1658 if (off_beyond_header && off != Type::OffsetBot) {
1659 assert(adr->is_AddP() && adr->in(AddPNode::Offset)->is_Con(), "offset is a constant");
1660 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type());
1661 if (con_type != NULL) {
1662 return con_type;
1663 }
1664 }
1665 }
1667 // Don't do this for integer types. There is only potential profit if
1668 // the element type t is lower than _type; that is, for int types, if _type is
1669 // more restrictive than t. This only happens here if one is short and the other
1670 // char (both 16 bits), and in those cases we've made an intentional decision
1671 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1672 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1673 //
1674 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1675 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1676 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1677 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1678 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1679 // In fact, that could have been the original type of p1, and p1 could have
1680 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1681 // expression (LShiftL quux 3) independently optimized to the constant 8.
1682 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1683 && (_type->isa_vect() == NULL)
1684 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1685 // t might actually be lower than _type, if _type is a unique
1686 // concrete subclass of abstract class t.
1687 if (off_beyond_header) { // is the offset beyond the header?
1688 const Type* jt = t->join(_type);
1689 // In any case, do not allow the join, per se, to empty out the type.
1690 if (jt->empty() && !t->empty()) {
1691 // This can happen if a interface-typed array narrows to a class type.
1692 jt = _type;
1693 }
1694 #ifdef ASSERT
1695 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1696 // The pointers in the autobox arrays are always non-null
1697 Node* base = adr->in(AddPNode::Base);
1698 if ((base != NULL) && base->is_DecodeN()) {
1699 // Get LoadN node which loads IntegerCache.cache field
1700 base = base->in(1);
1701 }
1702 if ((base != NULL) && base->is_Con()) {
1703 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1704 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1705 // It could be narrow oop
1706 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1707 }
1708 }
1709 }
1710 #endif
1711 return jt;
1712 }
1713 }
1714 } else if (tp->base() == Type::InstPtr) {
1715 ciEnv* env = C->env();
1716 const TypeInstPtr* tinst = tp->is_instptr();
1717 ciKlass* klass = tinst->klass();
1718 assert( off != Type::OffsetBot ||
1719 // arrays can be cast to Objects
1720 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1721 // unsafe field access may not have a constant offset
1722 C->has_unsafe_access(),
1723 "Field accesses must be precise" );
1724 // For oop loads, we expect the _type to be precise
1725 if (klass == env->String_klass() &&
1726 adr->is_AddP() && off != Type::OffsetBot) {
1727 // For constant Strings treat the final fields as compile time constants.
1728 Node* base = adr->in(AddPNode::Base);
1729 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1730 if (t != NULL && t->singleton()) {
1731 ciField* field = env->String_klass()->get_field_by_offset(off, false);
1732 if (field != NULL && field->is_final()) {
1733 ciObject* string = t->const_oop();
1734 ciConstant constant = string->as_instance()->field_value(field);
1735 if (constant.basic_type() == T_INT) {
1736 return TypeInt::make(constant.as_int());
1737 } else if (constant.basic_type() == T_ARRAY) {
1738 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1739 return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1740 } else {
1741 return TypeOopPtr::make_from_constant(constant.as_object(), true);
1742 }
1743 }
1744 }
1745 }
1746 }
1747 // Optimizations for constant objects
1748 ciObject* const_oop = tinst->const_oop();
1749 if (const_oop != NULL) {
1750 // For constant Boxed value treat the target field as a compile time constant.
1751 if (tinst->is_ptr_to_boxed_value()) {
1752 return tinst->get_const_boxed_value();
1753 } else
1754 // For constant CallSites treat the target field as a compile time constant.
1755 if (const_oop->is_call_site()) {
1756 ciCallSite* call_site = const_oop->as_call_site();
1757 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1758 if (field != NULL && field->is_call_site_target()) {
1759 ciMethodHandle* target = call_site->get_target();
1760 if (target != NULL) { // just in case
1761 ciConstant constant(T_OBJECT, target);
1762 const Type* t;
1763 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1764 t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1765 } else {
1766 t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1767 }
1768 // Add a dependence for invalidation of the optimization.
1769 if (!call_site->is_constant_call_site()) {
1770 C->dependencies()->assert_call_site_target_value(call_site, target);
1771 }
1772 return t;
1773 }
1774 }
1775 }
1776 }
1777 } else if (tp->base() == Type::KlassPtr) {
1778 assert( off != Type::OffsetBot ||
1779 // arrays can be cast to Objects
1780 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1781 // also allow array-loading from the primary supertype
1782 // array during subtype checks
1783 Opcode() == Op_LoadKlass,
1784 "Field accesses must be precise" );
1785 // For klass/static loads, we expect the _type to be precise
1786 }
1788 const TypeKlassPtr *tkls = tp->isa_klassptr();
1789 if (tkls != NULL && !StressReflectiveCode) {
1790 ciKlass* klass = tkls->klass();
1791 if (klass->is_loaded() && tkls->klass_is_exact()) {
1792 // We are loading a field from a Klass metaobject whose identity
1793 // is known at compile time (the type is "exact" or "precise").
1794 // Check for fields we know are maintained as constants by the VM.
1795 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1796 // The field is Klass::_super_check_offset. Return its (constant) value.
1797 // (Folds up type checking code.)
1798 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1799 return TypeInt::make(klass->super_check_offset());
1800 }
1801 // Compute index into primary_supers array
1802 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1803 // Check for overflowing; use unsigned compare to handle the negative case.
1804 if( depth < ciKlass::primary_super_limit() ) {
1805 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1806 // (Folds up type checking code.)
1807 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1808 ciKlass *ss = klass->super_of_depth(depth);
1809 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1810 }
1811 const Type* aift = load_array_final_field(tkls, klass);
1812 if (aift != NULL) return aift;
1813 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())
1814 && klass->is_array_klass()) {
1815 // The field is ArrayKlass::_component_mirror. Return its (constant) value.
1816 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1817 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1818 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1819 }
1820 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1821 // The field is Klass::_java_mirror. Return its (constant) value.
1822 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1823 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1824 return TypeInstPtr::make(klass->java_mirror());
1825 }
1826 }
1828 // We can still check if we are loading from the primary_supers array at a
1829 // shallow enough depth. Even though the klass is not exact, entries less
1830 // than or equal to its super depth are correct.
1831 if (klass->is_loaded() ) {
1832 ciType *inner = klass;
1833 while( inner->is_obj_array_klass() )
1834 inner = inner->as_obj_array_klass()->base_element_type();
1835 if( inner->is_instance_klass() &&
1836 !inner->as_instance_klass()->flags().is_interface() ) {
1837 // Compute index into primary_supers array
1838 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1839 // Check for overflowing; use unsigned compare to handle the negative case.
1840 if( depth < ciKlass::primary_super_limit() &&
1841 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1842 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1843 // (Folds up type checking code.)
1844 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1845 ciKlass *ss = klass->super_of_depth(depth);
1846 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1847 }
1848 }
1849 }
1851 // If the type is enough to determine that the thing is not an array,
1852 // we can give the layout_helper a positive interval type.
1853 // This will help short-circuit some reflective code.
1854 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1855 && !klass->is_array_klass() // not directly typed as an array
1856 && !klass->is_interface() // specifically not Serializable & Cloneable
1857 && !klass->is_java_lang_Object() // not the supertype of all T[]
1858 ) {
1859 // Note: When interfaces are reliable, we can narrow the interface
1860 // test to (klass != Serializable && klass != Cloneable).
1861 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1862 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1863 // The key property of this type is that it folds up tests
1864 // for array-ness, since it proves that the layout_helper is positive.
1865 // Thus, a generic value like the basic object layout helper works fine.
1866 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1867 }
1868 }
1870 // If we are loading from a freshly-allocated object, produce a zero,
1871 // if the load is provably beyond the header of the object.
1872 // (Also allow a variable load from a fresh array to produce zero.)
1873 const TypeOopPtr *tinst = tp->isa_oopptr();
1874 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1875 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1876 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1877 Node* value = can_see_stored_value(mem,phase);
1878 if (value != NULL && value->is_Con()) {
1879 assert(value->bottom_type()->higher_equal(_type),"sanity");
1880 return value->bottom_type();
1881 }
1882 }
1884 if (is_instance) {
1885 // If we have an instance type and our memory input is the
1886 // programs's initial memory state, there is no matching store,
1887 // so just return a zero of the appropriate type
1888 Node *mem = in(MemNode::Memory);
1889 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1890 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1891 return Type::get_zero_type(_type->basic_type());
1892 }
1893 }
1894 return _type;
1895 }
1897 //------------------------------match_edge-------------------------------------
1898 // Do we Match on this edge index or not? Match only the address.
1899 uint LoadNode::match_edge(uint idx) const {
1900 return idx == MemNode::Address;
1901 }
1903 //--------------------------LoadBNode::Ideal--------------------------------------
1904 //
1905 // If the previous store is to the same address as this load,
1906 // and the value stored was larger than a byte, replace this load
1907 // with the value stored truncated to a byte. If no truncation is
1908 // needed, the replacement is done in LoadNode::Identity().
1909 //
1910 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1911 Node* mem = in(MemNode::Memory);
1912 Node* value = can_see_stored_value(mem,phase);
1913 if( value && !phase->type(value)->higher_equal( _type ) ) {
1914 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );
1915 return new (phase->C) RShiftINode(result, phase->intcon(24));
1916 }
1917 // Identity call will handle the case where truncation is not needed.
1918 return LoadNode::Ideal(phase, can_reshape);
1919 }
1921 const Type* LoadBNode::Value(PhaseTransform *phase) const {
1922 Node* mem = in(MemNode::Memory);
1923 Node* value = can_see_stored_value(mem,phase);
1924 if (value != NULL && value->is_Con() &&
1925 !value->bottom_type()->higher_equal(_type)) {
1926 // If the input to the store does not fit with the load's result type,
1927 // it must be truncated. We can't delay until Ideal call since
1928 // a singleton Value is needed for split_thru_phi optimization.
1929 int con = value->get_int();
1930 return TypeInt::make((con << 24) >> 24);
1931 }
1932 return LoadNode::Value(phase);
1933 }
1935 //--------------------------LoadUBNode::Ideal-------------------------------------
1936 //
1937 // If the previous store is to the same address as this load,
1938 // and the value stored was larger than a byte, replace this load
1939 // with the value stored truncated to a byte. If no truncation is
1940 // needed, the replacement is done in LoadNode::Identity().
1941 //
1942 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1943 Node* mem = in(MemNode::Memory);
1944 Node* value = can_see_stored_value(mem, phase);
1945 if (value && !phase->type(value)->higher_equal(_type))
1946 return new (phase->C) AndINode(value, phase->intcon(0xFF));
1947 // Identity call will handle the case where truncation is not needed.
1948 return LoadNode::Ideal(phase, can_reshape);
1949 }
1951 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1952 Node* mem = in(MemNode::Memory);
1953 Node* value = can_see_stored_value(mem,phase);
1954 if (value != NULL && value->is_Con() &&
1955 !value->bottom_type()->higher_equal(_type)) {
1956 // If the input to the store does not fit with the load's result type,
1957 // it must be truncated. We can't delay until Ideal call since
1958 // a singleton Value is needed for split_thru_phi optimization.
1959 int con = value->get_int();
1960 return TypeInt::make(con & 0xFF);
1961 }
1962 return LoadNode::Value(phase);
1963 }
1965 //--------------------------LoadUSNode::Ideal-------------------------------------
1966 //
1967 // If the previous store is to the same address as this load,
1968 // and the value stored was larger than a char, replace this load
1969 // with the value stored truncated to a char. If no truncation is
1970 // needed, the replacement is done in LoadNode::Identity().
1971 //
1972 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1973 Node* mem = in(MemNode::Memory);
1974 Node* value = can_see_stored_value(mem,phase);
1975 if( value && !phase->type(value)->higher_equal( _type ) )
1976 return new (phase->C) AndINode(value,phase->intcon(0xFFFF));
1977 // Identity call will handle the case where truncation is not needed.
1978 return LoadNode::Ideal(phase, can_reshape);
1979 }
1981 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1982 Node* mem = in(MemNode::Memory);
1983 Node* value = can_see_stored_value(mem,phase);
1984 if (value != NULL && value->is_Con() &&
1985 !value->bottom_type()->higher_equal(_type)) {
1986 // If the input to the store does not fit with the load's result type,
1987 // it must be truncated. We can't delay until Ideal call since
1988 // a singleton Value is needed for split_thru_phi optimization.
1989 int con = value->get_int();
1990 return TypeInt::make(con & 0xFFFF);
1991 }
1992 return LoadNode::Value(phase);
1993 }
1995 //--------------------------LoadSNode::Ideal--------------------------------------
1996 //
1997 // If the previous store is to the same address as this load,
1998 // and the value stored was larger than a short, replace this load
1999 // with the value stored truncated to a short. If no truncation is
2000 // needed, the replacement is done in LoadNode::Identity().
2001 //
2002 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2003 Node* mem = in(MemNode::Memory);
2004 Node* value = can_see_stored_value(mem,phase);
2005 if( value && !phase->type(value)->higher_equal( _type ) ) {
2006 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );
2007 return new (phase->C) RShiftINode(result, phase->intcon(16));
2008 }
2009 // Identity call will handle the case where truncation is not needed.
2010 return LoadNode::Ideal(phase, can_reshape);
2011 }
2013 const Type* LoadSNode::Value(PhaseTransform *phase) const {
2014 Node* mem = in(MemNode::Memory);
2015 Node* value = can_see_stored_value(mem,phase);
2016 if (value != NULL && value->is_Con() &&
2017 !value->bottom_type()->higher_equal(_type)) {
2018 // If the input to the store does not fit with the load's result type,
2019 // it must be truncated. We can't delay until Ideal call since
2020 // a singleton Value is needed for split_thru_phi optimization.
2021 int con = value->get_int();
2022 return TypeInt::make((con << 16) >> 16);
2023 }
2024 return LoadNode::Value(phase);
2025 }
2027 //=============================================================================
2028 //----------------------------LoadKlassNode::make------------------------------
2029 // Polymorphic factory method:
2030 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
2031 Compile* C = gvn.C;
2032 Node *ctl = NULL;
2033 // sanity check the alias category against the created node type
2034 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2035 assert(adr_type != NULL, "expecting TypeKlassPtr");
2036 #ifdef _LP64
2037 if (adr_type->is_ptr_to_narrowklass()) {
2038 assert(UseCompressedClassPointers, "no compressed klasses");
2039 Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2040 return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2041 }
2042 #endif
2043 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2044 return new (C) LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2045 }
2047 //------------------------------Value------------------------------------------
2048 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
2049 return klass_value_common(phase);
2050 }
2052 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
2053 // Either input is TOP ==> the result is TOP
2054 const Type *t1 = phase->type( in(MemNode::Memory) );
2055 if (t1 == Type::TOP) return Type::TOP;
2056 Node *adr = in(MemNode::Address);
2057 const Type *t2 = phase->type( adr );
2058 if (t2 == Type::TOP) return Type::TOP;
2059 const TypePtr *tp = t2->is_ptr();
2060 if (TypePtr::above_centerline(tp->ptr()) ||
2061 tp->ptr() == TypePtr::Null) return Type::TOP;
2063 // Return a more precise klass, if possible
2064 const TypeInstPtr *tinst = tp->isa_instptr();
2065 if (tinst != NULL) {
2066 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2067 int offset = tinst->offset();
2068 if (ik == phase->C->env()->Class_klass()
2069 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2070 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2071 // We are loading a special hidden field from a Class mirror object,
2072 // the field which points to the VM's Klass metaobject.
2073 ciType* t = tinst->java_mirror_type();
2074 // java_mirror_type returns non-null for compile-time Class constants.
2075 if (t != NULL) {
2076 // constant oop => constant klass
2077 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2078 return TypeKlassPtr::make(ciArrayKlass::make(t));
2079 }
2080 if (!t->is_klass()) {
2081 // a primitive Class (e.g., int.class) has NULL for a klass field
2082 return TypePtr::NULL_PTR;
2083 }
2084 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2085 return TypeKlassPtr::make(t->as_klass());
2086 }
2087 // non-constant mirror, so we can't tell what's going on
2088 }
2089 if( !ik->is_loaded() )
2090 return _type; // Bail out if not loaded
2091 if (offset == oopDesc::klass_offset_in_bytes()) {
2092 if (tinst->klass_is_exact()) {
2093 return TypeKlassPtr::make(ik);
2094 }
2095 // See if we can become precise: no subklasses and no interface
2096 // (Note: We need to support verified interfaces.)
2097 if (!ik->is_interface() && !ik->has_subklass()) {
2098 //assert(!UseExactTypes, "this code should be useless with exact types");
2099 // Add a dependence; if any subclass added we need to recompile
2100 if (!ik->is_final()) {
2101 // %%% should use stronger assert_unique_concrete_subtype instead
2102 phase->C->dependencies()->assert_leaf_type(ik);
2103 }
2104 // Return precise klass
2105 return TypeKlassPtr::make(ik);
2106 }
2108 // Return root of possible klass
2109 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2110 }
2111 }
2113 // Check for loading klass from an array
2114 const TypeAryPtr *tary = tp->isa_aryptr();
2115 if( tary != NULL ) {
2116 ciKlass *tary_klass = tary->klass();
2117 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2118 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2119 if (tary->klass_is_exact()) {
2120 return TypeKlassPtr::make(tary_klass);
2121 }
2122 ciArrayKlass *ak = tary->klass()->as_array_klass();
2123 // If the klass is an object array, we defer the question to the
2124 // array component klass.
2125 if( ak->is_obj_array_klass() ) {
2126 assert( ak->is_loaded(), "" );
2127 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2128 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2129 ciInstanceKlass* ik = base_k->as_instance_klass();
2130 // See if we can become precise: no subklasses and no interface
2131 if (!ik->is_interface() && !ik->has_subklass()) {
2132 //assert(!UseExactTypes, "this code should be useless with exact types");
2133 // Add a dependence; if any subclass added we need to recompile
2134 if (!ik->is_final()) {
2135 phase->C->dependencies()->assert_leaf_type(ik);
2136 }
2137 // Return precise array klass
2138 return TypeKlassPtr::make(ak);
2139 }
2140 }
2141 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2142 } else { // Found a type-array?
2143 //assert(!UseExactTypes, "this code should be useless with exact types");
2144 assert( ak->is_type_array_klass(), "" );
2145 return TypeKlassPtr::make(ak); // These are always precise
2146 }
2147 }
2148 }
2150 // Check for loading klass from an array klass
2151 const TypeKlassPtr *tkls = tp->isa_klassptr();
2152 if (tkls != NULL && !StressReflectiveCode) {
2153 ciKlass* klass = tkls->klass();
2154 if( !klass->is_loaded() )
2155 return _type; // Bail out if not loaded
2156 if( klass->is_obj_array_klass() &&
2157 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2158 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2159 // // Always returning precise element type is incorrect,
2160 // // e.g., element type could be object and array may contain strings
2161 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2163 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2164 // according to the element type's subclassing.
2165 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2166 }
2167 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2168 tkls->offset() == in_bytes(Klass::super_offset())) {
2169 ciKlass* sup = klass->as_instance_klass()->super();
2170 // The field is Klass::_super. Return its (constant) value.
2171 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2172 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2173 }
2174 }
2176 // Bailout case
2177 return LoadNode::Value(phase);
2178 }
2180 //------------------------------Identity---------------------------------------
2181 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2182 // Also feed through the klass in Allocate(...klass...)._klass.
2183 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
2184 return klass_identity_common(phase);
2185 }
2187 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
2188 Node* x = LoadNode::Identity(phase);
2189 if (x != this) return x;
2191 // Take apart the address into an oop and and offset.
2192 // Return 'this' if we cannot.
2193 Node* adr = in(MemNode::Address);
2194 intptr_t offset = 0;
2195 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2196 if (base == NULL) return this;
2197 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2198 if (toop == NULL) return this;
2200 // We can fetch the klass directly through an AllocateNode.
2201 // This works even if the klass is not constant (clone or newArray).
2202 if (offset == oopDesc::klass_offset_in_bytes()) {
2203 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2204 if (allocated_klass != NULL) {
2205 return allocated_klass;
2206 }
2207 }
2209 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2210 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.
2211 // See inline_native_Class_query for occurrences of these patterns.
2212 // Java Example: x.getClass().isAssignableFrom(y)
2213 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
2214 //
2215 // This improves reflective code, often making the Class
2216 // mirror go completely dead. (Current exception: Class
2217 // mirrors may appear in debug info, but we could clean them out by
2218 // introducing a new debug info operator for Klass*.java_mirror).
2219 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2220 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2221 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2222 // We are loading a special hidden field from a Class mirror,
2223 // the field which points to its Klass or ArrayKlass metaobject.
2224 if (base->is_Load()) {
2225 Node* adr2 = base->in(MemNode::Address);
2226 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2227 if (tkls != NULL && !tkls->empty()
2228 && (tkls->klass()->is_instance_klass() ||
2229 tkls->klass()->is_array_klass())
2230 && adr2->is_AddP()
2231 ) {
2232 int mirror_field = in_bytes(Klass::java_mirror_offset());
2233 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2234 mirror_field = in_bytes(ArrayKlass::component_mirror_offset());
2235 }
2236 if (tkls->offset() == mirror_field) {
2237 return adr2->in(AddPNode::Base);
2238 }
2239 }
2240 }
2241 }
2243 return this;
2244 }
2247 //------------------------------Value------------------------------------------
2248 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
2249 const Type *t = klass_value_common(phase);
2250 if (t == Type::TOP)
2251 return t;
2253 return t->make_narrowklass();
2254 }
2256 //------------------------------Identity---------------------------------------
2257 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2258 // Also feed through the klass in Allocate(...klass...)._klass.
2259 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2260 Node *x = klass_identity_common(phase);
2262 const Type *t = phase->type( x );
2263 if( t == Type::TOP ) return x;
2264 if( t->isa_narrowklass()) return x;
2265 assert (!t->isa_narrowoop(), "no narrow oop here");
2267 return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));
2268 }
2270 //------------------------------Value-----------------------------------------
2271 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2272 // Either input is TOP ==> the result is TOP
2273 const Type *t1 = phase->type( in(MemNode::Memory) );
2274 if( t1 == Type::TOP ) return Type::TOP;
2275 Node *adr = in(MemNode::Address);
2276 const Type *t2 = phase->type( adr );
2277 if( t2 == Type::TOP ) return Type::TOP;
2278 const TypePtr *tp = t2->is_ptr();
2279 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2280 const TypeAryPtr *tap = tp->isa_aryptr();
2281 if( !tap ) return _type;
2282 return tap->size();
2283 }
2285 //-------------------------------Ideal---------------------------------------
2286 // Feed through the length in AllocateArray(...length...)._length.
2287 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2288 Node* p = MemNode::Ideal_common(phase, can_reshape);
2289 if (p) return (p == NodeSentinel) ? NULL : p;
2291 // Take apart the address into an oop and and offset.
2292 // Return 'this' if we cannot.
2293 Node* adr = in(MemNode::Address);
2294 intptr_t offset = 0;
2295 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2296 if (base == NULL) return NULL;
2297 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2298 if (tary == NULL) return NULL;
2300 // We can fetch the length directly through an AllocateArrayNode.
2301 // This works even if the length is not constant (clone or newArray).
2302 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2303 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2304 if (alloc != NULL) {
2305 Node* allocated_length = alloc->Ideal_length();
2306 Node* len = alloc->make_ideal_length(tary, phase);
2307 if (allocated_length != len) {
2308 // New CastII improves on this.
2309 return len;
2310 }
2311 }
2312 }
2314 return NULL;
2315 }
2317 //------------------------------Identity---------------------------------------
2318 // Feed through the length in AllocateArray(...length...)._length.
2319 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2320 Node* x = LoadINode::Identity(phase);
2321 if (x != this) return x;
2323 // Take apart the address into an oop and and offset.
2324 // Return 'this' if we cannot.
2325 Node* adr = in(MemNode::Address);
2326 intptr_t offset = 0;
2327 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2328 if (base == NULL) return this;
2329 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2330 if (tary == NULL) return this;
2332 // We can fetch the length directly through an AllocateArrayNode.
2333 // This works even if the length is not constant (clone or newArray).
2334 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2335 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2336 if (alloc != NULL) {
2337 Node* allocated_length = alloc->Ideal_length();
2338 // Do not allow make_ideal_length to allocate a CastII node.
2339 Node* len = alloc->make_ideal_length(tary, phase, false);
2340 if (allocated_length == len) {
2341 // Return allocated_length only if it would not be improved by a CastII.
2342 return allocated_length;
2343 }
2344 }
2345 }
2347 return this;
2349 }
2351 //=============================================================================
2352 //---------------------------StoreNode::make-----------------------------------
2353 // Polymorphic factory method:
2354 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2355 assert((mo == unordered || mo == release), "unexpected");
2356 Compile* C = gvn.C;
2357 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2358 ctl != NULL, "raw memory operations should have control edge");
2360 switch (bt) {
2361 case T_BOOLEAN:
2362 case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val, mo);
2363 case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val, mo);
2364 case T_CHAR:
2365 case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val, mo);
2366 case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo);
2367 case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val, mo);
2368 case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo);
2369 case T_METADATA:
2370 case T_ADDRESS:
2371 case T_OBJECT:
2372 #ifdef _LP64
2373 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2374 val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2375 return new (C) StoreNNode(ctl, mem, adr, adr_type, val, mo);
2376 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2377 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2378 adr->bottom_type()->isa_rawptr())) {
2379 val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2380 return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2381 }
2382 #endif
2383 {
2384 return new (C) StorePNode(ctl, mem, adr, adr_type, val, mo);
2385 }
2386 }
2387 ShouldNotReachHere();
2388 return (StoreNode*)NULL;
2389 }
2391 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2392 bool require_atomic = true;
2393 return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2394 }
2397 //--------------------------bottom_type----------------------------------------
2398 const Type *StoreNode::bottom_type() const {
2399 return Type::MEMORY;
2400 }
2402 //------------------------------hash-------------------------------------------
2403 uint StoreNode::hash() const {
2404 // unroll addition of interesting fields
2405 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2407 // Since they are not commoned, do not hash them:
2408 return NO_HASH;
2409 }
2411 //------------------------------Ideal------------------------------------------
2412 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2413 // When a store immediately follows a relevant allocation/initialization,
2414 // try to capture it into the initialization, or hoist it above.
2415 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2416 Node* p = MemNode::Ideal_common(phase, can_reshape);
2417 if (p) return (p == NodeSentinel) ? NULL : p;
2419 Node* mem = in(MemNode::Memory);
2420 Node* address = in(MemNode::Address);
2422 // Back-to-back stores to same address? Fold em up. Generally
2423 // unsafe if I have intervening uses... Also disallowed for StoreCM
2424 // since they must follow each StoreP operation. Redundant StoreCMs
2425 // are eliminated just before matching in final_graph_reshape.
2426 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
2427 mem->Opcode() != Op_StoreCM) {
2428 // Looking at a dead closed cycle of memory?
2429 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2431 assert(Opcode() == mem->Opcode() ||
2432 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2433 "no mismatched stores, except on raw memory");
2435 if (mem->outcnt() == 1 && // check for intervening uses
2436 mem->as_Store()->memory_size() <= this->memory_size()) {
2437 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2438 // For example, 'mem' might be the final state at a conditional return.
2439 // Or, 'mem' might be used by some node which is live at the same time
2440 // 'this' is live, which might be unschedulable. So, require exactly
2441 // ONE user, the 'this' store, until such time as we clone 'mem' for
2442 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2443 if (can_reshape) { // (%%% is this an anachronism?)
2444 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2445 phase->is_IterGVN());
2446 } else {
2447 // It's OK to do this in the parser, since DU info is always accurate,
2448 // and the parser always refers to nodes via SafePointNode maps.
2449 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2450 }
2451 return this;
2452 }
2453 }
2455 // Capture an unaliased, unconditional, simple store into an initializer.
2456 // Or, if it is independent of the allocation, hoist it above the allocation.
2457 if (ReduceFieldZeroing && /*can_reshape &&*/
2458 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2459 InitializeNode* init = mem->in(0)->as_Initialize();
2460 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2461 if (offset > 0) {
2462 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2463 // If the InitializeNode captured me, it made a raw copy of me,
2464 // and I need to disappear.
2465 if (moved != NULL) {
2466 // %%% hack to ensure that Ideal returns a new node:
2467 mem = MergeMemNode::make(phase->C, mem);
2468 return mem; // fold me away
2469 }
2470 }
2471 }
2473 return NULL; // No further progress
2474 }
2476 //------------------------------Value-----------------------------------------
2477 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2478 // Either input is TOP ==> the result is TOP
2479 const Type *t1 = phase->type( in(MemNode::Memory) );
2480 if( t1 == Type::TOP ) return Type::TOP;
2481 const Type *t2 = phase->type( in(MemNode::Address) );
2482 if( t2 == Type::TOP ) return Type::TOP;
2483 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2484 if( t3 == Type::TOP ) return Type::TOP;
2485 return Type::MEMORY;
2486 }
2488 //------------------------------Identity---------------------------------------
2489 // Remove redundant stores:
2490 // Store(m, p, Load(m, p)) changes to m.
2491 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2492 Node *StoreNode::Identity( PhaseTransform *phase ) {
2493 Node* mem = in(MemNode::Memory);
2494 Node* adr = in(MemNode::Address);
2495 Node* val = in(MemNode::ValueIn);
2497 // Load then Store? Then the Store is useless
2498 if (val->is_Load() &&
2499 val->in(MemNode::Address)->eqv_uncast(adr) &&
2500 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2501 val->as_Load()->store_Opcode() == Opcode()) {
2502 return mem;
2503 }
2505 // Two stores in a row of the same value?
2506 if (mem->is_Store() &&
2507 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2508 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2509 mem->Opcode() == Opcode()) {
2510 return mem;
2511 }
2513 // Store of zero anywhere into a freshly-allocated object?
2514 // Then the store is useless.
2515 // (It must already have been captured by the InitializeNode.)
2516 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2517 // a newly allocated object is already all-zeroes everywhere
2518 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2519 return mem;
2520 }
2522 // the store may also apply to zero-bits in an earlier object
2523 Node* prev_mem = find_previous_store(phase);
2524 // Steps (a), (b): Walk past independent stores to find an exact match.
2525 if (prev_mem != NULL) {
2526 Node* prev_val = can_see_stored_value(prev_mem, phase);
2527 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2528 // prev_val and val might differ by a cast; it would be good
2529 // to keep the more informative of the two.
2530 return mem;
2531 }
2532 }
2533 }
2535 return this;
2536 }
2538 //------------------------------match_edge-------------------------------------
2539 // Do we Match on this edge index or not? Match only memory & value
2540 uint StoreNode::match_edge(uint idx) const {
2541 return idx == MemNode::Address || idx == MemNode::ValueIn;
2542 }
2544 //------------------------------cmp--------------------------------------------
2545 // Do not common stores up together. They generally have to be split
2546 // back up anyways, so do not bother.
2547 uint StoreNode::cmp( const Node &n ) const {
2548 return (&n == this); // Always fail except on self
2549 }
2551 //------------------------------Ideal_masked_input-----------------------------
2552 // Check for a useless mask before a partial-word store
2553 // (StoreB ... (AndI valIn conIa) )
2554 // If (conIa & mask == mask) this simplifies to
2555 // (StoreB ... (valIn) )
2556 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2557 Node *val = in(MemNode::ValueIn);
2558 if( val->Opcode() == Op_AndI ) {
2559 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2560 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2561 set_req(MemNode::ValueIn, val->in(1));
2562 return this;
2563 }
2564 }
2565 return NULL;
2566 }
2569 //------------------------------Ideal_sign_extended_input----------------------
2570 // Check for useless sign-extension before a partial-word store
2571 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2572 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2573 // (StoreB ... (valIn) )
2574 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2575 Node *val = in(MemNode::ValueIn);
2576 if( val->Opcode() == Op_RShiftI ) {
2577 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2578 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2579 Node *shl = val->in(1);
2580 if( shl->Opcode() == Op_LShiftI ) {
2581 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2582 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2583 set_req(MemNode::ValueIn, shl->in(1));
2584 return this;
2585 }
2586 }
2587 }
2588 }
2589 return NULL;
2590 }
2592 //------------------------------value_never_loaded-----------------------------------
2593 // Determine whether there are any possible loads of the value stored.
2594 // For simplicity, we actually check if there are any loads from the
2595 // address stored to, not just for loads of the value stored by this node.
2596 //
2597 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2598 Node *adr = in(Address);
2599 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2600 if (adr_oop == NULL)
2601 return false;
2602 if (!adr_oop->is_known_instance_field())
2603 return false; // if not a distinct instance, there may be aliases of the address
2604 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2605 Node *use = adr->fast_out(i);
2606 int opc = use->Opcode();
2607 if (use->is_Load() || use->is_LoadStore()) {
2608 return false;
2609 }
2610 }
2611 return true;
2612 }
2614 //=============================================================================
2615 //------------------------------Ideal------------------------------------------
2616 // If the store is from an AND mask that leaves the low bits untouched, then
2617 // we can skip the AND operation. If the store is from a sign-extension
2618 // (a left shift, then right shift) we can skip both.
2619 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2620 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2621 if( progress != NULL ) return progress;
2623 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2624 if( progress != NULL ) return progress;
2626 // Finally check the default case
2627 return StoreNode::Ideal(phase, can_reshape);
2628 }
2630 //=============================================================================
2631 //------------------------------Ideal------------------------------------------
2632 // If the store is from an AND mask that leaves the low bits untouched, then
2633 // we can skip the AND operation
2634 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2635 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2636 if( progress != NULL ) return progress;
2638 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2639 if( progress != NULL ) return progress;
2641 // Finally check the default case
2642 return StoreNode::Ideal(phase, can_reshape);
2643 }
2645 //=============================================================================
2646 //------------------------------Identity---------------------------------------
2647 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2648 // No need to card mark when storing a null ptr
2649 Node* my_store = in(MemNode::OopStore);
2650 if (my_store->is_Store()) {
2651 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2652 if( t1 == TypePtr::NULL_PTR ) {
2653 return in(MemNode::Memory);
2654 }
2655 }
2656 return this;
2657 }
2659 //=============================================================================
2660 //------------------------------Ideal---------------------------------------
2661 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2662 Node* progress = StoreNode::Ideal(phase, can_reshape);
2663 if (progress != NULL) return progress;
2665 Node* my_store = in(MemNode::OopStore);
2666 if (my_store->is_MergeMem()) {
2667 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2668 set_req(MemNode::OopStore, mem);
2669 return this;
2670 }
2672 return NULL;
2673 }
2675 //------------------------------Value-----------------------------------------
2676 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2677 // Either input is TOP ==> the result is TOP
2678 const Type *t = phase->type( in(MemNode::Memory) );
2679 if( t == Type::TOP ) return Type::TOP;
2680 t = phase->type( in(MemNode::Address) );
2681 if( t == Type::TOP ) return Type::TOP;
2682 t = phase->type( in(MemNode::ValueIn) );
2683 if( t == Type::TOP ) return Type::TOP;
2684 // If extra input is TOP ==> the result is TOP
2685 t = phase->type( in(MemNode::OopStore) );
2686 if( t == Type::TOP ) return Type::TOP;
2688 return StoreNode::Value( phase );
2689 }
2692 //=============================================================================
2693 //----------------------------------SCMemProjNode------------------------------
2694 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2695 {
2696 return bottom_type();
2697 }
2699 //=============================================================================
2700 //----------------------------------LoadStoreNode------------------------------
2701 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2702 : Node(required),
2703 _type(rt),
2704 _adr_type(at)
2705 {
2706 init_req(MemNode::Control, c );
2707 init_req(MemNode::Memory , mem);
2708 init_req(MemNode::Address, adr);
2709 init_req(MemNode::ValueIn, val);
2710 init_class_id(Class_LoadStore);
2711 }
2713 uint LoadStoreNode::ideal_reg() const {
2714 return _type->ideal_reg();
2715 }
2717 bool LoadStoreNode::result_not_used() const {
2718 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2719 Node *x = fast_out(i);
2720 if (x->Opcode() == Op_SCMemProj) continue;
2721 return false;
2722 }
2723 return true;
2724 }
2726 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2728 //=============================================================================
2729 //----------------------------------LoadStoreConditionalNode--------------------
2730 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2731 init_req(ExpectedIn, ex );
2732 }
2734 //=============================================================================
2735 //-------------------------------adr_type--------------------------------------
2736 // Do we Match on this edge index or not? Do not match memory
2737 const TypePtr* ClearArrayNode::adr_type() const {
2738 Node *adr = in(3);
2739 return MemNode::calculate_adr_type(adr->bottom_type());
2740 }
2742 //------------------------------match_edge-------------------------------------
2743 // Do we Match on this edge index or not? Do not match memory
2744 uint ClearArrayNode::match_edge(uint idx) const {
2745 return idx > 1;
2746 }
2748 //------------------------------Identity---------------------------------------
2749 // Clearing a zero length array does nothing
2750 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2751 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2752 }
2754 //------------------------------Idealize---------------------------------------
2755 // Clearing a short array is faster with stores
2756 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2757 const int unit = BytesPerLong;
2758 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2759 if (!t) return NULL;
2760 if (!t->is_con()) return NULL;
2761 intptr_t raw_count = t->get_con();
2762 intptr_t size = raw_count;
2763 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2764 // Clearing nothing uses the Identity call.
2765 // Negative clears are possible on dead ClearArrays
2766 // (see jck test stmt114.stmt11402.val).
2767 if (size <= 0 || size % unit != 0) return NULL;
2768 intptr_t count = size / unit;
2769 // Length too long; use fast hardware clear
2770 if (size > Matcher::init_array_short_size) return NULL;
2771 Node *mem = in(1);
2772 if( phase->type(mem)==Type::TOP ) return NULL;
2773 Node *adr = in(3);
2774 const Type* at = phase->type(adr);
2775 if( at==Type::TOP ) return NULL;
2776 const TypePtr* atp = at->isa_ptr();
2777 // adjust atp to be the correct array element address type
2778 if (atp == NULL) atp = TypePtr::BOTTOM;
2779 else atp = atp->add_offset(Type::OffsetBot);
2780 // Get base for derived pointer purposes
2781 if( adr->Opcode() != Op_AddP ) Unimplemented();
2782 Node *base = adr->in(1);
2784 Node *zero = phase->makecon(TypeLong::ZERO);
2785 Node *off = phase->MakeConX(BytesPerLong);
2786 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2787 count--;
2788 while( count-- ) {
2789 mem = phase->transform(mem);
2790 adr = phase->transform(new (phase->C) AddPNode(base,adr,off));
2791 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2792 }
2793 return mem;
2794 }
2796 //----------------------------step_through----------------------------------
2797 // Return allocation input memory edge if it is different instance
2798 // or itself if it is the one we are looking for.
2799 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2800 Node* n = *np;
2801 assert(n->is_ClearArray(), "sanity");
2802 intptr_t offset;
2803 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2804 // This method is called only before Allocate nodes are expanded during
2805 // macro nodes expansion. Before that ClearArray nodes are only generated
2806 // in LibraryCallKit::generate_arraycopy() which follows allocations.
2807 assert(alloc != NULL, "should have allocation");
2808 if (alloc->_idx == instance_id) {
2809 // Can not bypass initialization of the instance we are looking for.
2810 return false;
2811 }
2812 // Otherwise skip it.
2813 InitializeNode* init = alloc->initialization();
2814 if (init != NULL)
2815 *np = init->in(TypeFunc::Memory);
2816 else
2817 *np = alloc->in(TypeFunc::Memory);
2818 return true;
2819 }
2821 //----------------------------clear_memory-------------------------------------
2822 // Generate code to initialize object storage to zero.
2823 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2824 intptr_t start_offset,
2825 Node* end_offset,
2826 PhaseGVN* phase) {
2827 Compile* C = phase->C;
2828 intptr_t offset = start_offset;
2830 int unit = BytesPerLong;
2831 if ((offset % unit) != 0) {
2832 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset));
2833 adr = phase->transform(adr);
2834 const TypePtr* atp = TypeRawPtr::BOTTOM;
2835 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2836 mem = phase->transform(mem);
2837 offset += BytesPerInt;
2838 }
2839 assert((offset % unit) == 0, "");
2841 // Initialize the remaining stuff, if any, with a ClearArray.
2842 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2843 }
2845 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2846 Node* start_offset,
2847 Node* end_offset,
2848 PhaseGVN* phase) {
2849 if (start_offset == end_offset) {
2850 // nothing to do
2851 return mem;
2852 }
2854 Compile* C = phase->C;
2855 int unit = BytesPerLong;
2856 Node* zbase = start_offset;
2857 Node* zend = end_offset;
2859 // Scale to the unit required by the CPU:
2860 if (!Matcher::init_array_count_is_in_bytes) {
2861 Node* shift = phase->intcon(exact_log2(unit));
2862 zbase = phase->transform( new(C) URShiftXNode(zbase, shift) );
2863 zend = phase->transform( new(C) URShiftXNode(zend, shift) );
2864 }
2866 // Bulk clear double-words
2867 Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) );
2868 Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) );
2869 mem = new (C) ClearArrayNode(ctl, mem, zsize, adr);
2870 return phase->transform(mem);
2871 }
2873 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2874 intptr_t start_offset,
2875 intptr_t end_offset,
2876 PhaseGVN* phase) {
2877 if (start_offset == end_offset) {
2878 // nothing to do
2879 return mem;
2880 }
2882 Compile* C = phase->C;
2883 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2884 intptr_t done_offset = end_offset;
2885 if ((done_offset % BytesPerLong) != 0) {
2886 done_offset -= BytesPerInt;
2887 }
2888 if (done_offset > start_offset) {
2889 mem = clear_memory(ctl, mem, dest,
2890 start_offset, phase->MakeConX(done_offset), phase);
2891 }
2892 if (done_offset < end_offset) { // emit the final 32-bit store
2893 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset));
2894 adr = phase->transform(adr);
2895 const TypePtr* atp = TypeRawPtr::BOTTOM;
2896 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2897 mem = phase->transform(mem);
2898 done_offset += BytesPerInt;
2899 }
2900 assert(done_offset == end_offset, "");
2901 return mem;
2902 }
2904 //=============================================================================
2905 // Do not match memory edge.
2906 uint StrIntrinsicNode::match_edge(uint idx) const {
2907 return idx == 2 || idx == 3;
2908 }
2910 //------------------------------Ideal------------------------------------------
2911 // Return a node which is more "ideal" than the current node. Strip out
2912 // control copies
2913 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2914 if (remove_dead_region(phase, can_reshape)) return this;
2915 // Don't bother trying to transform a dead node
2916 if (in(0) && in(0)->is_top()) return NULL;
2918 if (can_reshape) {
2919 Node* mem = phase->transform(in(MemNode::Memory));
2920 // If transformed to a MergeMem, get the desired slice
2921 uint alias_idx = phase->C->get_alias_index(adr_type());
2922 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem;
2923 if (mem != in(MemNode::Memory)) {
2924 set_req(MemNode::Memory, mem);
2925 return this;
2926 }
2927 }
2928 return NULL;
2929 }
2931 //------------------------------Value------------------------------------------
2932 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const {
2933 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
2934 return bottom_type();
2935 }
2937 //=============================================================================
2938 //------------------------------match_edge-------------------------------------
2939 // Do not match memory edge
2940 uint EncodeISOArrayNode::match_edge(uint idx) const {
2941 return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len)
2942 }
2944 //------------------------------Ideal------------------------------------------
2945 // Return a node which is more "ideal" than the current node. Strip out
2946 // control copies
2947 Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2948 return remove_dead_region(phase, can_reshape) ? this : NULL;
2949 }
2951 //------------------------------Value------------------------------------------
2952 const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const {
2953 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
2954 return bottom_type();
2955 }
2957 //=============================================================================
2958 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2959 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2960 _adr_type(C->get_adr_type(alias_idx))
2961 {
2962 init_class_id(Class_MemBar);
2963 Node* top = C->top();
2964 init_req(TypeFunc::I_O,top);
2965 init_req(TypeFunc::FramePtr,top);
2966 init_req(TypeFunc::ReturnAdr,top);
2967 if (precedent != NULL)
2968 init_req(TypeFunc::Parms, precedent);
2969 }
2971 //------------------------------cmp--------------------------------------------
2972 uint MemBarNode::hash() const { return NO_HASH; }
2973 uint MemBarNode::cmp( const Node &n ) const {
2974 return (&n == this); // Always fail except on self
2975 }
2977 //------------------------------make-------------------------------------------
2978 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2979 switch (opcode) {
2980 case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn);
2981 case Op_LoadFence: return new(C) LoadFenceNode(C, atp, pn);
2982 case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn);
2983 case Op_StoreFence: return new(C) StoreFenceNode(C, atp, pn);
2984 case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn);
2985 case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn);
2986 case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn);
2987 case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn);
2988 case Op_Initialize: return new(C) InitializeNode(C, atp, pn);
2989 case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn);
2990 default: ShouldNotReachHere(); return NULL;
2991 }
2992 }
2994 //------------------------------Ideal------------------------------------------
2995 // Return a node which is more "ideal" than the current node. Strip out
2996 // control copies
2997 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2998 if (remove_dead_region(phase, can_reshape)) return this;
2999 // Don't bother trying to transform a dead node
3000 if (in(0) && in(0)->is_top()) {
3001 return NULL;
3002 }
3004 // Eliminate volatile MemBars for scalar replaced objects.
3005 if (can_reshape && req() == (Precedent+1)) {
3006 bool eliminate = false;
3007 int opc = Opcode();
3008 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3009 // Volatile field loads and stores.
3010 Node* my_mem = in(MemBarNode::Precedent);
3011 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3012 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3013 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3014 // replace this Precedent (decodeN) with the Load instead.
3015 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3016 Node* load_node = my_mem->in(1);
3017 set_req(MemBarNode::Precedent, load_node);
3018 phase->is_IterGVN()->_worklist.push(my_mem);
3019 my_mem = load_node;
3020 } else {
3021 assert(my_mem->unique_out() == this, "sanity");
3022 del_req(Precedent);
3023 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3024 my_mem = NULL;
3025 }
3026 }
3027 if (my_mem != NULL && my_mem->is_Mem()) {
3028 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3029 // Check for scalar replaced object reference.
3030 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3031 t_oop->offset() != Type::OffsetBot &&
3032 t_oop->offset() != Type::OffsetTop) {
3033 eliminate = true;
3034 }
3035 }
3036 } else if (opc == Op_MemBarRelease) {
3037 // Final field stores.
3038 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3039 if ((alloc != NULL) && alloc->is_Allocate() &&
3040 alloc->as_Allocate()->_is_non_escaping) {
3041 // The allocated object does not escape.
3042 eliminate = true;
3043 }
3044 }
3045 if (eliminate) {
3046 // Replace MemBar projections by its inputs.
3047 PhaseIterGVN* igvn = phase->is_IterGVN();
3048 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3049 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3050 // Must return either the original node (now dead) or a new node
3051 // (Do not return a top here, since that would break the uniqueness of top.)
3052 return new (phase->C) ConINode(TypeInt::ZERO);
3053 }
3054 }
3055 return NULL;
3056 }
3058 //------------------------------Value------------------------------------------
3059 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
3060 if( !in(0) ) return Type::TOP;
3061 if( phase->type(in(0)) == Type::TOP )
3062 return Type::TOP;
3063 return TypeTuple::MEMBAR;
3064 }
3066 //------------------------------match------------------------------------------
3067 // Construct projections for memory.
3068 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3069 switch (proj->_con) {
3070 case TypeFunc::Control:
3071 case TypeFunc::Memory:
3072 return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3073 }
3074 ShouldNotReachHere();
3075 return NULL;
3076 }
3078 //===========================InitializeNode====================================
3079 // SUMMARY:
3080 // This node acts as a memory barrier on raw memory, after some raw stores.
3081 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3082 // The Initialize can 'capture' suitably constrained stores as raw inits.
3083 // It can coalesce related raw stores into larger units (called 'tiles').
3084 // It can avoid zeroing new storage for memory units which have raw inits.
3085 // At macro-expansion, it is marked 'complete', and does not optimize further.
3086 //
3087 // EXAMPLE:
3088 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3089 // ctl = incoming control; mem* = incoming memory
3090 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3091 // First allocate uninitialized memory and fill in the header:
3092 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3093 // ctl := alloc.Control; mem* := alloc.Memory*
3094 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3095 // Then initialize to zero the non-header parts of the raw memory block:
3096 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3097 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3098 // After the initialize node executes, the object is ready for service:
3099 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3100 // Suppose its body is immediately initialized as {1,2}:
3101 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3102 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3103 // mem.SLICE(#short[*]) := store2
3104 //
3105 // DETAILS:
3106 // An InitializeNode collects and isolates object initialization after
3107 // an AllocateNode and before the next possible safepoint. As a
3108 // memory barrier (MemBarNode), it keeps critical stores from drifting
3109 // down past any safepoint or any publication of the allocation.
3110 // Before this barrier, a newly-allocated object may have uninitialized bits.
3111 // After this barrier, it may be treated as a real oop, and GC is allowed.
3112 //
3113 // The semantics of the InitializeNode include an implicit zeroing of
3114 // the new object from object header to the end of the object.
3115 // (The object header and end are determined by the AllocateNode.)
3116 //
3117 // Certain stores may be added as direct inputs to the InitializeNode.
3118 // These stores must update raw memory, and they must be to addresses
3119 // derived from the raw address produced by AllocateNode, and with
3120 // a constant offset. They must be ordered by increasing offset.
3121 // The first one is at in(RawStores), the last at in(req()-1).
3122 // Unlike most memory operations, they are not linked in a chain,
3123 // but are displayed in parallel as users of the rawmem output of
3124 // the allocation.
3125 //
3126 // (See comments in InitializeNode::capture_store, which continue
3127 // the example given above.)
3128 //
3129 // When the associated Allocate is macro-expanded, the InitializeNode
3130 // may be rewritten to optimize collected stores. A ClearArrayNode
3131 // may also be created at that point to represent any required zeroing.
3132 // The InitializeNode is then marked 'complete', prohibiting further
3133 // capturing of nearby memory operations.
3134 //
3135 // During macro-expansion, all captured initializations which store
3136 // constant values of 32 bits or smaller are coalesced (if advantageous)
3137 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3138 // initialized in fewer memory operations. Memory words which are
3139 // covered by neither tiles nor non-constant stores are pre-zeroed
3140 // by explicit stores of zero. (The code shape happens to do all
3141 // zeroing first, then all other stores, with both sequences occurring
3142 // in order of ascending offsets.)
3143 //
3144 // Alternatively, code may be inserted between an AllocateNode and its
3145 // InitializeNode, to perform arbitrary initialization of the new object.
3146 // E.g., the object copying intrinsics insert complex data transfers here.
3147 // The initialization must then be marked as 'complete' disable the
3148 // built-in zeroing semantics and the collection of initializing stores.
3149 //
3150 // While an InitializeNode is incomplete, reads from the memory state
3151 // produced by it are optimizable if they match the control edge and
3152 // new oop address associated with the allocation/initialization.
3153 // They return a stored value (if the offset matches) or else zero.
3154 // A write to the memory state, if it matches control and address,
3155 // and if it is to a constant offset, may be 'captured' by the
3156 // InitializeNode. It is cloned as a raw memory operation and rewired
3157 // inside the initialization, to the raw oop produced by the allocation.
3158 // Operations on addresses which are provably distinct (e.g., to
3159 // other AllocateNodes) are allowed to bypass the initialization.
3160 //
3161 // The effect of all this is to consolidate object initialization
3162 // (both arrays and non-arrays, both piecewise and bulk) into a
3163 // single location, where it can be optimized as a unit.
3164 //
3165 // Only stores with an offset less than TrackedInitializationLimit words
3166 // will be considered for capture by an InitializeNode. This puts a
3167 // reasonable limit on the complexity of optimized initializations.
3169 //---------------------------InitializeNode------------------------------------
3170 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3171 : _is_complete(Incomplete), _does_not_escape(false),
3172 MemBarNode(C, adr_type, rawoop)
3173 {
3174 init_class_id(Class_Initialize);
3176 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3177 assert(in(RawAddress) == rawoop, "proper init");
3178 // Note: allocation() can be NULL, for secondary initialization barriers
3179 }
3181 // Since this node is not matched, it will be processed by the
3182 // register allocator. Declare that there are no constraints
3183 // on the allocation of the RawAddress edge.
3184 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3185 // This edge should be set to top, by the set_complete. But be conservative.
3186 if (idx == InitializeNode::RawAddress)
3187 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3188 return RegMask::Empty;
3189 }
3191 Node* InitializeNode::memory(uint alias_idx) {
3192 Node* mem = in(Memory);
3193 if (mem->is_MergeMem()) {
3194 return mem->as_MergeMem()->memory_at(alias_idx);
3195 } else {
3196 // incoming raw memory is not split
3197 return mem;
3198 }
3199 }
3201 bool InitializeNode::is_non_zero() {
3202 if (is_complete()) return false;
3203 remove_extra_zeroes();
3204 return (req() > RawStores);
3205 }
3207 void InitializeNode::set_complete(PhaseGVN* phase) {
3208 assert(!is_complete(), "caller responsibility");
3209 _is_complete = Complete;
3211 // After this node is complete, it contains a bunch of
3212 // raw-memory initializations. There is no need for
3213 // it to have anything to do with non-raw memory effects.
3214 // Therefore, tell all non-raw users to re-optimize themselves,
3215 // after skipping the memory effects of this initialization.
3216 PhaseIterGVN* igvn = phase->is_IterGVN();
3217 if (igvn) igvn->add_users_to_worklist(this);
3218 }
3220 // convenience function
3221 // return false if the init contains any stores already
3222 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3223 InitializeNode* init = initialization();
3224 if (init == NULL || init->is_complete()) return false;
3225 init->remove_extra_zeroes();
3226 // for now, if this allocation has already collected any inits, bail:
3227 if (init->is_non_zero()) return false;
3228 init->set_complete(phase);
3229 return true;
3230 }
3232 void InitializeNode::remove_extra_zeroes() {
3233 if (req() == RawStores) return;
3234 Node* zmem = zero_memory();
3235 uint fill = RawStores;
3236 for (uint i = fill; i < req(); i++) {
3237 Node* n = in(i);
3238 if (n->is_top() || n == zmem) continue; // skip
3239 if (fill < i) set_req(fill, n); // compact
3240 ++fill;
3241 }
3242 // delete any empty spaces created:
3243 while (fill < req()) {
3244 del_req(fill);
3245 }
3246 }
3248 // Helper for remembering which stores go with which offsets.
3249 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3250 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3251 intptr_t offset = -1;
3252 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3253 phase, offset);
3254 if (base == NULL) return -1; // something is dead,
3255 if (offset < 0) return -1; // dead, dead
3256 return offset;
3257 }
3259 // Helper for proving that an initialization expression is
3260 // "simple enough" to be folded into an object initialization.
3261 // Attempts to prove that a store's initial value 'n' can be captured
3262 // within the initialization without creating a vicious cycle, such as:
3263 // { Foo p = new Foo(); p.next = p; }
3264 // True for constants and parameters and small combinations thereof.
3265 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3266 if (n == NULL) return true; // (can this really happen?)
3267 if (n->is_Proj()) n = n->in(0);
3268 if (n == this) return false; // found a cycle
3269 if (n->is_Con()) return true;
3270 if (n->is_Start()) return true; // params, etc., are OK
3271 if (n->is_Root()) return true; // even better
3273 Node* ctl = n->in(0);
3274 if (ctl != NULL && !ctl->is_top()) {
3275 if (ctl->is_Proj()) ctl = ctl->in(0);
3276 if (ctl == this) return false;
3278 // If we already know that the enclosing memory op is pinned right after
3279 // the init, then any control flow that the store has picked up
3280 // must have preceded the init, or else be equal to the init.
3281 // Even after loop optimizations (which might change control edges)
3282 // a store is never pinned *before* the availability of its inputs.
3283 if (!MemNode::all_controls_dominate(n, this))
3284 return false; // failed to prove a good control
3285 }
3287 // Check data edges for possible dependencies on 'this'.
3288 if ((count += 1) > 20) return false; // complexity limit
3289 for (uint i = 1; i < n->req(); i++) {
3290 Node* m = n->in(i);
3291 if (m == NULL || m == n || m->is_top()) continue;
3292 uint first_i = n->find_edge(m);
3293 if (i != first_i) continue; // process duplicate edge just once
3294 if (!detect_init_independence(m, count)) {
3295 return false;
3296 }
3297 }
3299 return true;
3300 }
3302 // Here are all the checks a Store must pass before it can be moved into
3303 // an initialization. Returns zero if a check fails.
3304 // On success, returns the (constant) offset to which the store applies,
3305 // within the initialized memory.
3306 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3307 const int FAIL = 0;
3308 if (st->req() != MemNode::ValueIn + 1)
3309 return FAIL; // an inscrutable StoreNode (card mark?)
3310 Node* ctl = st->in(MemNode::Control);
3311 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3312 return FAIL; // must be unconditional after the initialization
3313 Node* mem = st->in(MemNode::Memory);
3314 if (!(mem->is_Proj() && mem->in(0) == this))
3315 return FAIL; // must not be preceded by other stores
3316 Node* adr = st->in(MemNode::Address);
3317 intptr_t offset;
3318 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3319 if (alloc == NULL)
3320 return FAIL; // inscrutable address
3321 if (alloc != allocation())
3322 return FAIL; // wrong allocation! (store needs to float up)
3323 Node* val = st->in(MemNode::ValueIn);
3324 int complexity_count = 0;
3325 if (!detect_init_independence(val, complexity_count))
3326 return FAIL; // stored value must be 'simple enough'
3328 // The Store can be captured only if nothing after the allocation
3329 // and before the Store is using the memory location that the store
3330 // overwrites.
3331 bool failed = false;
3332 // If is_complete_with_arraycopy() is true the shape of the graph is
3333 // well defined and is safe so no need for extra checks.
3334 if (!is_complete_with_arraycopy()) {
3335 // We are going to look at each use of the memory state following
3336 // the allocation to make sure nothing reads the memory that the
3337 // Store writes.
3338 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3339 int alias_idx = phase->C->get_alias_index(t_adr);
3340 ResourceMark rm;
3341 Unique_Node_List mems;
3342 mems.push(mem);
3343 Node* unique_merge = NULL;
3344 for (uint next = 0; next < mems.size(); ++next) {
3345 Node *m = mems.at(next);
3346 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3347 Node *n = m->fast_out(j);
3348 if (n->outcnt() == 0) {
3349 continue;
3350 }
3351 if (n == st) {
3352 continue;
3353 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3354 // If the control of this use is different from the control
3355 // of the Store which is right after the InitializeNode then
3356 // this node cannot be between the InitializeNode and the
3357 // Store.
3358 continue;
3359 } else if (n->is_MergeMem()) {
3360 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3361 // We can hit a MergeMemNode (that will likely go away
3362 // later) that is a direct use of the memory state
3363 // following the InitializeNode on the same slice as the
3364 // store node that we'd like to capture. We need to check
3365 // the uses of the MergeMemNode.
3366 mems.push(n);
3367 }
3368 } else if (n->is_Mem()) {
3369 Node* other_adr = n->in(MemNode::Address);
3370 if (other_adr == adr) {
3371 failed = true;
3372 break;
3373 } else {
3374 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3375 if (other_t_adr != NULL) {
3376 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3377 if (other_alias_idx == alias_idx) {
3378 // A load from the same memory slice as the store right
3379 // after the InitializeNode. We check the control of the
3380 // object/array that is loaded from. If it's the same as
3381 // the store control then we cannot capture the store.
3382 assert(!n->is_Store(), "2 stores to same slice on same control?");
3383 Node* base = other_adr;
3384 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
3385 base = base->in(AddPNode::Base);
3386 if (base != NULL) {
3387 base = base->uncast();
3388 if (base->is_Proj() && base->in(0) == alloc) {
3389 failed = true;
3390 break;
3391 }
3392 }
3393 }
3394 }
3395 }
3396 } else {
3397 failed = true;
3398 break;
3399 }
3400 }
3401 }
3402 }
3403 if (failed) {
3404 if (!can_reshape) {
3405 // We decided we couldn't capture the store during parsing. We
3406 // should try again during the next IGVN once the graph is
3407 // cleaner.
3408 phase->C->record_for_igvn(st);
3409 }
3410 return FAIL;
3411 }
3413 return offset; // success
3414 }
3416 // Find the captured store in(i) which corresponds to the range
3417 // [start..start+size) in the initialized object.
3418 // If there is one, return its index i. If there isn't, return the
3419 // negative of the index where it should be inserted.
3420 // Return 0 if the queried range overlaps an initialization boundary
3421 // or if dead code is encountered.
3422 // If size_in_bytes is zero, do not bother with overlap checks.
3423 int InitializeNode::captured_store_insertion_point(intptr_t start,
3424 int size_in_bytes,
3425 PhaseTransform* phase) {
3426 const int FAIL = 0, MAX_STORE = BytesPerLong;
3428 if (is_complete())
3429 return FAIL; // arraycopy got here first; punt
3431 assert(allocation() != NULL, "must be present");
3433 // no negatives, no header fields:
3434 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3436 // after a certain size, we bail out on tracking all the stores:
3437 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3438 if (start >= ti_limit) return FAIL;
3440 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3441 if (i >= limit) return -(int)i; // not found; here is where to put it
3443 Node* st = in(i);
3444 intptr_t st_off = get_store_offset(st, phase);
3445 if (st_off < 0) {
3446 if (st != zero_memory()) {
3447 return FAIL; // bail out if there is dead garbage
3448 }
3449 } else if (st_off > start) {
3450 // ...we are done, since stores are ordered
3451 if (st_off < start + size_in_bytes) {
3452 return FAIL; // the next store overlaps
3453 }
3454 return -(int)i; // not found; here is where to put it
3455 } else if (st_off < start) {
3456 if (size_in_bytes != 0 &&
3457 start < st_off + MAX_STORE &&
3458 start < st_off + st->as_Store()->memory_size()) {
3459 return FAIL; // the previous store overlaps
3460 }
3461 } else {
3462 if (size_in_bytes != 0 &&
3463 st->as_Store()->memory_size() != size_in_bytes) {
3464 return FAIL; // mismatched store size
3465 }
3466 return i;
3467 }
3469 ++i;
3470 }
3471 }
3473 // Look for a captured store which initializes at the offset 'start'
3474 // with the given size. If there is no such store, and no other
3475 // initialization interferes, then return zero_memory (the memory
3476 // projection of the AllocateNode).
3477 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3478 PhaseTransform* phase) {
3479 assert(stores_are_sane(phase), "");
3480 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3481 if (i == 0) {
3482 return NULL; // something is dead
3483 } else if (i < 0) {
3484 return zero_memory(); // just primordial zero bits here
3485 } else {
3486 Node* st = in(i); // here is the store at this position
3487 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3488 return st;
3489 }
3490 }
3492 // Create, as a raw pointer, an address within my new object at 'offset'.
3493 Node* InitializeNode::make_raw_address(intptr_t offset,
3494 PhaseTransform* phase) {
3495 Node* addr = in(RawAddress);
3496 if (offset != 0) {
3497 Compile* C = phase->C;
3498 addr = phase->transform( new (C) AddPNode(C->top(), addr,
3499 phase->MakeConX(offset)) );
3500 }
3501 return addr;
3502 }
3504 // Clone the given store, converting it into a raw store
3505 // initializing a field or element of my new object.
3506 // Caller is responsible for retiring the original store,
3507 // with subsume_node or the like.
3508 //
3509 // From the example above InitializeNode::InitializeNode,
3510 // here are the old stores to be captured:
3511 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3512 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3513 //
3514 // Here is the changed code; note the extra edges on init:
3515 // alloc = (Allocate ...)
3516 // rawoop = alloc.RawAddress
3517 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3518 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3519 // init = (Initialize alloc.Control alloc.Memory rawoop
3520 // rawstore1 rawstore2)
3521 //
3522 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3523 PhaseTransform* phase, bool can_reshape) {
3524 assert(stores_are_sane(phase), "");
3526 if (start < 0) return NULL;
3527 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3529 Compile* C = phase->C;
3530 int size_in_bytes = st->memory_size();
3531 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3532 if (i == 0) return NULL; // bail out
3533 Node* prev_mem = NULL; // raw memory for the captured store
3534 if (i > 0) {
3535 prev_mem = in(i); // there is a pre-existing store under this one
3536 set_req(i, C->top()); // temporarily disconnect it
3537 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3538 } else {
3539 i = -i; // no pre-existing store
3540 prev_mem = zero_memory(); // a slice of the newly allocated object
3541 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3542 set_req(--i, C->top()); // reuse this edge; it has been folded away
3543 else
3544 ins_req(i, C->top()); // build a new edge
3545 }
3546 Node* new_st = st->clone();
3547 new_st->set_req(MemNode::Control, in(Control));
3548 new_st->set_req(MemNode::Memory, prev_mem);
3549 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3550 new_st = phase->transform(new_st);
3552 // At this point, new_st might have swallowed a pre-existing store
3553 // at the same offset, or perhaps new_st might have disappeared,
3554 // if it redundantly stored the same value (or zero to fresh memory).
3556 // In any case, wire it in:
3557 set_req(i, new_st);
3559 // The caller may now kill the old guy.
3560 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3561 assert(check_st == new_st || check_st == NULL, "must be findable");
3562 assert(!is_complete(), "");
3563 return new_st;
3564 }
3566 static bool store_constant(jlong* tiles, int num_tiles,
3567 intptr_t st_off, int st_size,
3568 jlong con) {
3569 if ((st_off & (st_size-1)) != 0)
3570 return false; // strange store offset (assume size==2**N)
3571 address addr = (address)tiles + st_off;
3572 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3573 switch (st_size) {
3574 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3575 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3576 case sizeof(jint): *(jint*) addr = (jint) con; break;
3577 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3578 default: return false; // strange store size (detect size!=2**N here)
3579 }
3580 return true; // return success to caller
3581 }
3583 // Coalesce subword constants into int constants and possibly
3584 // into long constants. The goal, if the CPU permits,
3585 // is to initialize the object with a small number of 64-bit tiles.
3586 // Also, convert floating-point constants to bit patterns.
3587 // Non-constants are not relevant to this pass.
3588 //
3589 // In terms of the running example on InitializeNode::InitializeNode
3590 // and InitializeNode::capture_store, here is the transformation
3591 // of rawstore1 and rawstore2 into rawstore12:
3592 // alloc = (Allocate ...)
3593 // rawoop = alloc.RawAddress
3594 // tile12 = 0x00010002
3595 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3596 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3597 //
3598 void
3599 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3600 Node* size_in_bytes,
3601 PhaseGVN* phase) {
3602 Compile* C = phase->C;
3604 assert(stores_are_sane(phase), "");
3605 // Note: After this pass, they are not completely sane,
3606 // since there may be some overlaps.
3608 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3610 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3611 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3612 size_limit = MIN2(size_limit, ti_limit);
3613 size_limit = align_size_up(size_limit, BytesPerLong);
3614 int num_tiles = size_limit / BytesPerLong;
3616 // allocate space for the tile map:
3617 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3618 jlong tiles_buf[small_len];
3619 Node* nodes_buf[small_len];
3620 jlong inits_buf[small_len];
3621 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3622 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3623 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3624 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3625 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3626 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3627 // tiles: exact bitwise model of all primitive constants
3628 // nodes: last constant-storing node subsumed into the tiles model
3629 // inits: which bytes (in each tile) are touched by any initializations
3631 //// Pass A: Fill in the tile model with any relevant stores.
3633 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3634 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3635 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3636 Node* zmem = zero_memory(); // initially zero memory state
3637 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3638 Node* st = in(i);
3639 intptr_t st_off = get_store_offset(st, phase);
3641 // Figure out the store's offset and constant value:
3642 if (st_off < header_size) continue; //skip (ignore header)
3643 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3644 int st_size = st->as_Store()->memory_size();
3645 if (st_off + st_size > size_limit) break;
3647 // Record which bytes are touched, whether by constant or not.
3648 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3649 continue; // skip (strange store size)
3651 const Type* val = phase->type(st->in(MemNode::ValueIn));
3652 if (!val->singleton()) continue; //skip (non-con store)
3653 BasicType type = val->basic_type();
3655 jlong con = 0;
3656 switch (type) {
3657 case T_INT: con = val->is_int()->get_con(); break;
3658 case T_LONG: con = val->is_long()->get_con(); break;
3659 case T_FLOAT: con = jint_cast(val->getf()); break;
3660 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3661 default: continue; //skip (odd store type)
3662 }
3664 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3665 st->Opcode() == Op_StoreL) {
3666 continue; // This StoreL is already optimal.
3667 }
3669 // Store down the constant.
3670 store_constant(tiles, num_tiles, st_off, st_size, con);
3672 intptr_t j = st_off >> LogBytesPerLong;
3674 if (type == T_INT && st_size == BytesPerInt
3675 && (st_off & BytesPerInt) == BytesPerInt) {
3676 jlong lcon = tiles[j];
3677 if (!Matcher::isSimpleConstant64(lcon) &&
3678 st->Opcode() == Op_StoreI) {
3679 // This StoreI is already optimal by itself.
3680 jint* intcon = (jint*) &tiles[j];
3681 intcon[1] = 0; // undo the store_constant()
3683 // If the previous store is also optimal by itself, back up and
3684 // undo the action of the previous loop iteration... if we can.
3685 // But if we can't, just let the previous half take care of itself.
3686 st = nodes[j];
3687 st_off -= BytesPerInt;
3688 con = intcon[0];
3689 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3690 assert(st_off >= header_size, "still ignoring header");
3691 assert(get_store_offset(st, phase) == st_off, "must be");
3692 assert(in(i-1) == zmem, "must be");
3693 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3694 assert(con == tcon->is_int()->get_con(), "must be");
3695 // Undo the effects of the previous loop trip, which swallowed st:
3696 intcon[0] = 0; // undo store_constant()
3697 set_req(i-1, st); // undo set_req(i, zmem)
3698 nodes[j] = NULL; // undo nodes[j] = st
3699 --old_subword; // undo ++old_subword
3700 }
3701 continue; // This StoreI is already optimal.
3702 }
3703 }
3705 // This store is not needed.
3706 set_req(i, zmem);
3707 nodes[j] = st; // record for the moment
3708 if (st_size < BytesPerLong) // something has changed
3709 ++old_subword; // includes int/float, but who's counting...
3710 else ++old_long;
3711 }
3713 if ((old_subword + old_long) == 0)
3714 return; // nothing more to do
3716 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3717 // Be sure to insert them before overlapping non-constant stores.
3718 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3719 for (int j = 0; j < num_tiles; j++) {
3720 jlong con = tiles[j];
3721 jlong init = inits[j];
3722 if (con == 0) continue;
3723 jint con0, con1; // split the constant, address-wise
3724 jint init0, init1; // split the init map, address-wise
3725 { union { jlong con; jint intcon[2]; } u;
3726 u.con = con;
3727 con0 = u.intcon[0];
3728 con1 = u.intcon[1];
3729 u.con = init;
3730 init0 = u.intcon[0];
3731 init1 = u.intcon[1];
3732 }
3734 Node* old = nodes[j];
3735 assert(old != NULL, "need the prior store");
3736 intptr_t offset = (j * BytesPerLong);
3738 bool split = !Matcher::isSimpleConstant64(con);
3740 if (offset < header_size) {
3741 assert(offset + BytesPerInt >= header_size, "second int counts");
3742 assert(*(jint*)&tiles[j] == 0, "junk in header");
3743 split = true; // only the second word counts
3744 // Example: int a[] = { 42 ... }
3745 } else if (con0 == 0 && init0 == -1) {
3746 split = true; // first word is covered by full inits
3747 // Example: int a[] = { ... foo(), 42 ... }
3748 } else if (con1 == 0 && init1 == -1) {
3749 split = true; // second word is covered by full inits
3750 // Example: int a[] = { ... 42, foo() ... }
3751 }
3753 // Here's a case where init0 is neither 0 nor -1:
3754 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3755 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3756 // In this case the tile is not split; it is (jlong)42.
3757 // The big tile is stored down, and then the foo() value is inserted.
3758 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3760 Node* ctl = old->in(MemNode::Control);
3761 Node* adr = make_raw_address(offset, phase);
3762 const TypePtr* atp = TypeRawPtr::BOTTOM;
3764 // One or two coalesced stores to plop down.
3765 Node* st[2];
3766 intptr_t off[2];
3767 int nst = 0;
3768 if (!split) {
3769 ++new_long;
3770 off[nst] = offset;
3771 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3772 phase->longcon(con), T_LONG, MemNode::unordered);
3773 } else {
3774 // Omit either if it is a zero.
3775 if (con0 != 0) {
3776 ++new_int;
3777 off[nst] = offset;
3778 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3779 phase->intcon(con0), T_INT, MemNode::unordered);
3780 }
3781 if (con1 != 0) {
3782 ++new_int;
3783 offset += BytesPerInt;
3784 adr = make_raw_address(offset, phase);
3785 off[nst] = offset;
3786 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3787 phase->intcon(con1), T_INT, MemNode::unordered);
3788 }
3789 }
3791 // Insert second store first, then the first before the second.
3792 // Insert each one just before any overlapping non-constant stores.
3793 while (nst > 0) {
3794 Node* st1 = st[--nst];
3795 C->copy_node_notes_to(st1, old);
3796 st1 = phase->transform(st1);
3797 offset = off[nst];
3798 assert(offset >= header_size, "do not smash header");
3799 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3800 guarantee(ins_idx != 0, "must re-insert constant store");
3801 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3802 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3803 set_req(--ins_idx, st1);
3804 else
3805 ins_req(ins_idx, st1);
3806 }
3807 }
3809 if (PrintCompilation && WizardMode)
3810 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3811 old_subword, old_long, new_int, new_long);
3812 if (C->log() != NULL)
3813 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3814 old_subword, old_long, new_int, new_long);
3816 // Clean up any remaining occurrences of zmem:
3817 remove_extra_zeroes();
3818 }
3820 // Explore forward from in(start) to find the first fully initialized
3821 // word, and return its offset. Skip groups of subword stores which
3822 // together initialize full words. If in(start) is itself part of a
3823 // fully initialized word, return the offset of in(start). If there
3824 // are no following full-word stores, or if something is fishy, return
3825 // a negative value.
3826 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3827 int int_map = 0;
3828 intptr_t int_map_off = 0;
3829 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3831 for (uint i = start, limit = req(); i < limit; i++) {
3832 Node* st = in(i);
3834 intptr_t st_off = get_store_offset(st, phase);
3835 if (st_off < 0) break; // return conservative answer
3837 int st_size = st->as_Store()->memory_size();
3838 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3839 return st_off; // we found a complete word init
3840 }
3842 // update the map:
3844 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3845 if (this_int_off != int_map_off) {
3846 // reset the map:
3847 int_map = 0;
3848 int_map_off = this_int_off;
3849 }
3851 int subword_off = st_off - this_int_off;
3852 int_map |= right_n_bits(st_size) << subword_off;
3853 if ((int_map & FULL_MAP) == FULL_MAP) {
3854 return this_int_off; // we found a complete word init
3855 }
3857 // Did this store hit or cross the word boundary?
3858 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3859 if (next_int_off == this_int_off + BytesPerInt) {
3860 // We passed the current int, without fully initializing it.
3861 int_map_off = next_int_off;
3862 int_map >>= BytesPerInt;
3863 } else if (next_int_off > this_int_off + BytesPerInt) {
3864 // We passed the current and next int.
3865 return this_int_off + BytesPerInt;
3866 }
3867 }
3869 return -1;
3870 }
3873 // Called when the associated AllocateNode is expanded into CFG.
3874 // At this point, we may perform additional optimizations.
3875 // Linearize the stores by ascending offset, to make memory
3876 // activity as coherent as possible.
3877 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3878 intptr_t header_size,
3879 Node* size_in_bytes,
3880 PhaseGVN* phase) {
3881 assert(!is_complete(), "not already complete");
3882 assert(stores_are_sane(phase), "");
3883 assert(allocation() != NULL, "must be present");
3885 remove_extra_zeroes();
3887 if (ReduceFieldZeroing || ReduceBulkZeroing)
3888 // reduce instruction count for common initialization patterns
3889 coalesce_subword_stores(header_size, size_in_bytes, phase);
3891 Node* zmem = zero_memory(); // initially zero memory state
3892 Node* inits = zmem; // accumulating a linearized chain of inits
3893 #ifdef ASSERT
3894 intptr_t first_offset = allocation()->minimum_header_size();
3895 intptr_t last_init_off = first_offset; // previous init offset
3896 intptr_t last_init_end = first_offset; // previous init offset+size
3897 intptr_t last_tile_end = first_offset; // previous tile offset+size
3898 #endif
3899 intptr_t zeroes_done = header_size;
3901 bool do_zeroing = true; // we might give up if inits are very sparse
3902 int big_init_gaps = 0; // how many large gaps have we seen?
3904 if (ZeroTLAB) do_zeroing = false;
3905 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3907 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3908 Node* st = in(i);
3909 intptr_t st_off = get_store_offset(st, phase);
3910 if (st_off < 0)
3911 break; // unknown junk in the inits
3912 if (st->in(MemNode::Memory) != zmem)
3913 break; // complicated store chains somehow in list
3915 int st_size = st->as_Store()->memory_size();
3916 intptr_t next_init_off = st_off + st_size;
3918 if (do_zeroing && zeroes_done < next_init_off) {
3919 // See if this store needs a zero before it or under it.
3920 intptr_t zeroes_needed = st_off;
3922 if (st_size < BytesPerInt) {
3923 // Look for subword stores which only partially initialize words.
3924 // If we find some, we must lay down some word-level zeroes first,
3925 // underneath the subword stores.
3926 //
3927 // Examples:
3928 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3929 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3930 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3931 //
3932 // Note: coalesce_subword_stores may have already done this,
3933 // if it was prompted by constant non-zero subword initializers.
3934 // But this case can still arise with non-constant stores.
3936 intptr_t next_full_store = find_next_fullword_store(i, phase);
3938 // In the examples above:
3939 // in(i) p q r s x y z
3940 // st_off 12 13 14 15 12 13 14
3941 // st_size 1 1 1 1 1 1 1
3942 // next_full_s. 12 16 16 16 16 16 16
3943 // z's_done 12 16 16 16 12 16 12
3944 // z's_needed 12 16 16 16 16 16 16
3945 // zsize 0 0 0 0 4 0 4
3946 if (next_full_store < 0) {
3947 // Conservative tack: Zero to end of current word.
3948 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3949 } else {
3950 // Zero to beginning of next fully initialized word.
3951 // Or, don't zero at all, if we are already in that word.
3952 assert(next_full_store >= zeroes_needed, "must go forward");
3953 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3954 zeroes_needed = next_full_store;
3955 }
3956 }
3958 if (zeroes_needed > zeroes_done) {
3959 intptr_t zsize = zeroes_needed - zeroes_done;
3960 // Do some incremental zeroing on rawmem, in parallel with inits.
3961 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3962 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3963 zeroes_done, zeroes_needed,
3964 phase);
3965 zeroes_done = zeroes_needed;
3966 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3967 do_zeroing = false; // leave the hole, next time
3968 }
3969 }
3971 // Collect the store and move on:
3972 st->set_req(MemNode::Memory, inits);
3973 inits = st; // put it on the linearized chain
3974 set_req(i, zmem); // unhook from previous position
3976 if (zeroes_done == st_off)
3977 zeroes_done = next_init_off;
3979 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3981 #ifdef ASSERT
3982 // Various order invariants. Weaker than stores_are_sane because
3983 // a large constant tile can be filled in by smaller non-constant stores.
3984 assert(st_off >= last_init_off, "inits do not reverse");
3985 last_init_off = st_off;
3986 const Type* val = NULL;
3987 if (st_size >= BytesPerInt &&
3988 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3989 (int)val->basic_type() < (int)T_OBJECT) {
3990 assert(st_off >= last_tile_end, "tiles do not overlap");
3991 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3992 last_tile_end = MAX2(last_tile_end, next_init_off);
3993 } else {
3994 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3995 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3996 assert(st_off >= last_init_end, "inits do not overlap");
3997 last_init_end = next_init_off; // it's a non-tile
3998 }
3999 #endif //ASSERT
4000 }
4002 remove_extra_zeroes(); // clear out all the zmems left over
4003 add_req(inits);
4005 if (!ZeroTLAB) {
4006 // If anything remains to be zeroed, zero it all now.
4007 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
4008 // if it is the last unused 4 bytes of an instance, forget about it
4009 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4010 if (zeroes_done + BytesPerLong >= size_limit) {
4011 assert(allocation() != NULL, "");
4012 if (allocation()->Opcode() == Op_Allocate) {
4013 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
4014 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4015 if (zeroes_done == k->layout_helper())
4016 zeroes_done = size_limit;
4017 }
4018 }
4019 if (zeroes_done < size_limit) {
4020 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4021 zeroes_done, size_in_bytes, phase);
4022 }
4023 }
4025 set_complete(phase);
4026 return rawmem;
4027 }
4030 #ifdef ASSERT
4031 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4032 if (is_complete())
4033 return true; // stores could be anything at this point
4034 assert(allocation() != NULL, "must be present");
4035 intptr_t last_off = allocation()->minimum_header_size();
4036 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4037 Node* st = in(i);
4038 intptr_t st_off = get_store_offset(st, phase);
4039 if (st_off < 0) continue; // ignore dead garbage
4040 if (last_off > st_off) {
4041 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
4042 this->dump(2);
4043 assert(false, "ascending store offsets");
4044 return false;
4045 }
4046 last_off = st_off + st->as_Store()->memory_size();
4047 }
4048 return true;
4049 }
4050 #endif //ASSERT
4055 //============================MergeMemNode=====================================
4056 //
4057 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4058 // contributing store or call operations. Each contributor provides the memory
4059 // state for a particular "alias type" (see Compile::alias_type). For example,
4060 // if a MergeMem has an input X for alias category #6, then any memory reference
4061 // to alias category #6 may use X as its memory state input, as an exact equivalent
4062 // to using the MergeMem as a whole.
4063 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4064 //
4065 // (Here, the <N> notation gives the index of the relevant adr_type.)
4066 //
4067 // In one special case (and more cases in the future), alias categories overlap.
4068 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4069 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4070 // it is exactly equivalent to that state W:
4071 // MergeMem(<Bot>: W) <==> W
4072 //
4073 // Usually, the merge has more than one input. In that case, where inputs
4074 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4075 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4076 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4077 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4078 //
4079 // A merge can take a "wide" memory state as one of its narrow inputs.
4080 // This simply means that the merge observes out only the relevant parts of
4081 // the wide input. That is, wide memory states arriving at narrow merge inputs
4082 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4083 //
4084 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4085 // and that memory slices "leak through":
4086 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4087 //
4088 // But, in such a cascade, repeated memory slices can "block the leak":
4089 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4090 //
4091 // In the last example, Y is not part of the combined memory state of the
4092 // outermost MergeMem. The system must, of course, prevent unschedulable
4093 // memory states from arising, so you can be sure that the state Y is somehow
4094 // a precursor to state Y'.
4095 //
4096 //
4097 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4098 // of each MergeMemNode array are exactly the numerical alias indexes, including
4099 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4100 // Compile::alias_type (and kin) produce and manage these indexes.
4101 //
4102 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4103 // (Note that this provides quick access to the top node inside MergeMem methods,
4104 // without the need to reach out via TLS to Compile::current.)
4105 //
4106 // As a consequence of what was just described, a MergeMem that represents a full
4107 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4108 // containing all alias categories.
4109 //
4110 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4111 //
4112 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4113 // a memory state for the alias type <N>, or else the top node, meaning that
4114 // there is no particular input for that alias type. Note that the length of
4115 // a MergeMem is variable, and may be extended at any time to accommodate new
4116 // memory states at larger alias indexes. When merges grow, they are of course
4117 // filled with "top" in the unused in() positions.
4118 //
4119 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4120 // (Top was chosen because it works smoothly with passes like GCM.)
4121 //
4122 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4123 // the type of random VM bits like TLS references.) Since it is always the
4124 // first non-Bot memory slice, some low-level loops use it to initialize an
4125 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4126 //
4127 //
4128 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4129 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4130 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4131 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4132 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4133 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4134 //
4135 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4136 // really that different from the other memory inputs. An abbreviation called
4137 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4138 //
4139 //
4140 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4141 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4142 // that "emerges though" the base memory will be marked as excluding the alias types
4143 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4144 //
4145 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4146 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4147 //
4148 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4149 // (It is currently unimplemented.) As you can see, the resulting merge is
4150 // actually a disjoint union of memory states, rather than an overlay.
4151 //
4153 //------------------------------MergeMemNode-----------------------------------
4154 Node* MergeMemNode::make_empty_memory() {
4155 Node* empty_memory = (Node*) Compile::current()->top();
4156 assert(empty_memory->is_top(), "correct sentinel identity");
4157 return empty_memory;
4158 }
4160 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4161 init_class_id(Class_MergeMem);
4162 // all inputs are nullified in Node::Node(int)
4163 // set_input(0, NULL); // no control input
4165 // Initialize the edges uniformly to top, for starters.
4166 Node* empty_mem = make_empty_memory();
4167 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4168 init_req(i,empty_mem);
4169 }
4170 assert(empty_memory() == empty_mem, "");
4172 if( new_base != NULL && new_base->is_MergeMem() ) {
4173 MergeMemNode* mdef = new_base->as_MergeMem();
4174 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4175 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4176 mms.set_memory(mms.memory2());
4177 }
4178 assert(base_memory() == mdef->base_memory(), "");
4179 } else {
4180 set_base_memory(new_base);
4181 }
4182 }
4184 // Make a new, untransformed MergeMem with the same base as 'mem'.
4185 // If mem is itself a MergeMem, populate the result with the same edges.
4186 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
4187 return new(C) MergeMemNode(mem);
4188 }
4190 //------------------------------cmp--------------------------------------------
4191 uint MergeMemNode::hash() const { return NO_HASH; }
4192 uint MergeMemNode::cmp( const Node &n ) const {
4193 return (&n == this); // Always fail except on self
4194 }
4196 //------------------------------Identity---------------------------------------
4197 Node* MergeMemNode::Identity(PhaseTransform *phase) {
4198 // Identity if this merge point does not record any interesting memory
4199 // disambiguations.
4200 Node* base_mem = base_memory();
4201 Node* empty_mem = empty_memory();
4202 if (base_mem != empty_mem) { // Memory path is not dead?
4203 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4204 Node* mem = in(i);
4205 if (mem != empty_mem && mem != base_mem) {
4206 return this; // Many memory splits; no change
4207 }
4208 }
4209 }
4210 return base_mem; // No memory splits; ID on the one true input
4211 }
4213 //------------------------------Ideal------------------------------------------
4214 // This method is invoked recursively on chains of MergeMem nodes
4215 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4216 // Remove chain'd MergeMems
4217 //
4218 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4219 // relative to the "in(Bot)". Since we are patching both at the same time,
4220 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4221 // but rewrite each "in(i)" relative to the new "in(Bot)".
4222 Node *progress = NULL;
4225 Node* old_base = base_memory();
4226 Node* empty_mem = empty_memory();
4227 if (old_base == empty_mem)
4228 return NULL; // Dead memory path.
4230 MergeMemNode* old_mbase;
4231 if (old_base != NULL && old_base->is_MergeMem())
4232 old_mbase = old_base->as_MergeMem();
4233 else
4234 old_mbase = NULL;
4235 Node* new_base = old_base;
4237 // simplify stacked MergeMems in base memory
4238 if (old_mbase) new_base = old_mbase->base_memory();
4240 // the base memory might contribute new slices beyond my req()
4241 if (old_mbase) grow_to_match(old_mbase);
4243 // Look carefully at the base node if it is a phi.
4244 PhiNode* phi_base;
4245 if (new_base != NULL && new_base->is_Phi())
4246 phi_base = new_base->as_Phi();
4247 else
4248 phi_base = NULL;
4250 Node* phi_reg = NULL;
4251 uint phi_len = (uint)-1;
4252 if (phi_base != NULL && !phi_base->is_copy()) {
4253 // do not examine phi if degraded to a copy
4254 phi_reg = phi_base->region();
4255 phi_len = phi_base->req();
4256 // see if the phi is unfinished
4257 for (uint i = 1; i < phi_len; i++) {
4258 if (phi_base->in(i) == NULL) {
4259 // incomplete phi; do not look at it yet!
4260 phi_reg = NULL;
4261 phi_len = (uint)-1;
4262 break;
4263 }
4264 }
4265 }
4267 // Note: We do not call verify_sparse on entry, because inputs
4268 // can normalize to the base_memory via subsume_node or similar
4269 // mechanisms. This method repairs that damage.
4271 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4273 // Look at each slice.
4274 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4275 Node* old_in = in(i);
4276 // calculate the old memory value
4277 Node* old_mem = old_in;
4278 if (old_mem == empty_mem) old_mem = old_base;
4279 assert(old_mem == memory_at(i), "");
4281 // maybe update (reslice) the old memory value
4283 // simplify stacked MergeMems
4284 Node* new_mem = old_mem;
4285 MergeMemNode* old_mmem;
4286 if (old_mem != NULL && old_mem->is_MergeMem())
4287 old_mmem = old_mem->as_MergeMem();
4288 else
4289 old_mmem = NULL;
4290 if (old_mmem == this) {
4291 // This can happen if loops break up and safepoints disappear.
4292 // A merge of BotPtr (default) with a RawPtr memory derived from a
4293 // safepoint can be rewritten to a merge of the same BotPtr with
4294 // the BotPtr phi coming into the loop. If that phi disappears
4295 // also, we can end up with a self-loop of the mergemem.
4296 // In general, if loops degenerate and memory effects disappear,
4297 // a mergemem can be left looking at itself. This simply means
4298 // that the mergemem's default should be used, since there is
4299 // no longer any apparent effect on this slice.
4300 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4301 // from start. Update the input to TOP.
4302 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4303 }
4304 else if (old_mmem != NULL) {
4305 new_mem = old_mmem->memory_at(i);
4306 }
4307 // else preceding memory was not a MergeMem
4309 // replace equivalent phis (unfortunately, they do not GVN together)
4310 if (new_mem != NULL && new_mem != new_base &&
4311 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4312 if (new_mem->is_Phi()) {
4313 PhiNode* phi_mem = new_mem->as_Phi();
4314 for (uint i = 1; i < phi_len; i++) {
4315 if (phi_base->in(i) != phi_mem->in(i)) {
4316 phi_mem = NULL;
4317 break;
4318 }
4319 }
4320 if (phi_mem != NULL) {
4321 // equivalent phi nodes; revert to the def
4322 new_mem = new_base;
4323 }
4324 }
4325 }
4327 // maybe store down a new value
4328 Node* new_in = new_mem;
4329 if (new_in == new_base) new_in = empty_mem;
4331 if (new_in != old_in) {
4332 // Warning: Do not combine this "if" with the previous "if"
4333 // A memory slice might have be be rewritten even if it is semantically
4334 // unchanged, if the base_memory value has changed.
4335 set_req(i, new_in);
4336 progress = this; // Report progress
4337 }
4338 }
4340 if (new_base != old_base) {
4341 set_req(Compile::AliasIdxBot, new_base);
4342 // Don't use set_base_memory(new_base), because we need to update du.
4343 assert(base_memory() == new_base, "");
4344 progress = this;
4345 }
4347 if( base_memory() == this ) {
4348 // a self cycle indicates this memory path is dead
4349 set_req(Compile::AliasIdxBot, empty_mem);
4350 }
4352 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4353 // Recursion must occur after the self cycle check above
4354 if( base_memory()->is_MergeMem() ) {
4355 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4356 Node *m = phase->transform(new_mbase); // Rollup any cycles
4357 if( m != NULL && (m->is_top() ||
4358 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4359 // propagate rollup of dead cycle to self
4360 set_req(Compile::AliasIdxBot, empty_mem);
4361 }
4362 }
4364 if( base_memory() == empty_mem ) {
4365 progress = this;
4366 // Cut inputs during Parse phase only.
4367 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4368 if( !can_reshape ) {
4369 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4370 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4371 }
4372 }
4373 }
4375 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4376 // Check if PhiNode::Ideal's "Split phis through memory merges"
4377 // transform should be attempted. Look for this->phi->this cycle.
4378 uint merge_width = req();
4379 if (merge_width > Compile::AliasIdxRaw) {
4380 PhiNode* phi = base_memory()->as_Phi();
4381 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4382 if (phi->in(i) == this) {
4383 phase->is_IterGVN()->_worklist.push(phi);
4384 break;
4385 }
4386 }
4387 }
4388 }
4390 assert(progress || verify_sparse(), "please, no dups of base");
4391 return progress;
4392 }
4394 //-------------------------set_base_memory-------------------------------------
4395 void MergeMemNode::set_base_memory(Node *new_base) {
4396 Node* empty_mem = empty_memory();
4397 set_req(Compile::AliasIdxBot, new_base);
4398 assert(memory_at(req()) == new_base, "must set default memory");
4399 // Clear out other occurrences of new_base:
4400 if (new_base != empty_mem) {
4401 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4402 if (in(i) == new_base) set_req(i, empty_mem);
4403 }
4404 }
4405 }
4407 //------------------------------out_RegMask------------------------------------
4408 const RegMask &MergeMemNode::out_RegMask() const {
4409 return RegMask::Empty;
4410 }
4412 //------------------------------dump_spec--------------------------------------
4413 #ifndef PRODUCT
4414 void MergeMemNode::dump_spec(outputStream *st) const {
4415 st->print(" {");
4416 Node* base_mem = base_memory();
4417 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4418 Node* mem = memory_at(i);
4419 if (mem == base_mem) { st->print(" -"); continue; }
4420 st->print( " N%d:", mem->_idx );
4421 Compile::current()->get_adr_type(i)->dump_on(st);
4422 }
4423 st->print(" }");
4424 }
4425 #endif // !PRODUCT
4428 #ifdef ASSERT
4429 static bool might_be_same(Node* a, Node* b) {
4430 if (a == b) return true;
4431 if (!(a->is_Phi() || b->is_Phi())) return false;
4432 // phis shift around during optimization
4433 return true; // pretty stupid...
4434 }
4436 // verify a narrow slice (either incoming or outgoing)
4437 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4438 if (!VerifyAliases) return; // don't bother to verify unless requested
4439 if (is_error_reported()) return; // muzzle asserts when debugging an error
4440 if (Node::in_dump()) return; // muzzle asserts when printing
4441 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4442 assert(n != NULL, "");
4443 // Elide intervening MergeMem's
4444 while (n->is_MergeMem()) {
4445 n = n->as_MergeMem()->memory_at(alias_idx);
4446 }
4447 Compile* C = Compile::current();
4448 const TypePtr* n_adr_type = n->adr_type();
4449 if (n == m->empty_memory()) {
4450 // Implicit copy of base_memory()
4451 } else if (n_adr_type != TypePtr::BOTTOM) {
4452 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4453 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4454 } else {
4455 // A few places like make_runtime_call "know" that VM calls are narrow,
4456 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4457 bool expected_wide_mem = false;
4458 if (n == m->base_memory()) {
4459 expected_wide_mem = true;
4460 } else if (alias_idx == Compile::AliasIdxRaw ||
4461 n == m->memory_at(Compile::AliasIdxRaw)) {
4462 expected_wide_mem = true;
4463 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4464 // memory can "leak through" calls on channels that
4465 // are write-once. Allow this also.
4466 expected_wide_mem = true;
4467 }
4468 assert(expected_wide_mem, "expected narrow slice replacement");
4469 }
4470 }
4471 #else // !ASSERT
4472 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4473 #endif
4476 //-----------------------------memory_at---------------------------------------
4477 Node* MergeMemNode::memory_at(uint alias_idx) const {
4478 assert(alias_idx >= Compile::AliasIdxRaw ||
4479 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4480 "must avoid base_memory and AliasIdxTop");
4482 // Otherwise, it is a narrow slice.
4483 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4484 Compile *C = Compile::current();
4485 if (is_empty_memory(n)) {
4486 // the array is sparse; empty slots are the "top" node
4487 n = base_memory();
4488 assert(Node::in_dump()
4489 || n == NULL || n->bottom_type() == Type::TOP
4490 || n->adr_type() == NULL // address is TOP
4491 || n->adr_type() == TypePtr::BOTTOM
4492 || n->adr_type() == TypeRawPtr::BOTTOM
4493 || Compile::current()->AliasLevel() == 0,
4494 "must be a wide memory");
4495 // AliasLevel == 0 if we are organizing the memory states manually.
4496 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4497 } else {
4498 // make sure the stored slice is sane
4499 #ifdef ASSERT
4500 if (is_error_reported() || Node::in_dump()) {
4501 } else if (might_be_same(n, base_memory())) {
4502 // Give it a pass: It is a mostly harmless repetition of the base.
4503 // This can arise normally from node subsumption during optimization.
4504 } else {
4505 verify_memory_slice(this, alias_idx, n);
4506 }
4507 #endif
4508 }
4509 return n;
4510 }
4512 //---------------------------set_memory_at-------------------------------------
4513 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4514 verify_memory_slice(this, alias_idx, n);
4515 Node* empty_mem = empty_memory();
4516 if (n == base_memory()) n = empty_mem; // collapse default
4517 uint need_req = alias_idx+1;
4518 if (req() < need_req) {
4519 if (n == empty_mem) return; // already the default, so do not grow me
4520 // grow the sparse array
4521 do {
4522 add_req(empty_mem);
4523 } while (req() < need_req);
4524 }
4525 set_req( alias_idx, n );
4526 }
4530 //--------------------------iteration_setup------------------------------------
4531 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4532 if (other != NULL) {
4533 grow_to_match(other);
4534 // invariant: the finite support of mm2 is within mm->req()
4535 #ifdef ASSERT
4536 for (uint i = req(); i < other->req(); i++) {
4537 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4538 }
4539 #endif
4540 }
4541 // Replace spurious copies of base_memory by top.
4542 Node* base_mem = base_memory();
4543 if (base_mem != NULL && !base_mem->is_top()) {
4544 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4545 if (in(i) == base_mem)
4546 set_req(i, empty_memory());
4547 }
4548 }
4549 }
4551 //---------------------------grow_to_match-------------------------------------
4552 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4553 Node* empty_mem = empty_memory();
4554 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4555 // look for the finite support of the other memory
4556 for (uint i = other->req(); --i >= req(); ) {
4557 if (other->in(i) != empty_mem) {
4558 uint new_len = i+1;
4559 while (req() < new_len) add_req(empty_mem);
4560 break;
4561 }
4562 }
4563 }
4565 //---------------------------verify_sparse-------------------------------------
4566 #ifndef PRODUCT
4567 bool MergeMemNode::verify_sparse() const {
4568 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4569 Node* base_mem = base_memory();
4570 // The following can happen in degenerate cases, since empty==top.
4571 if (is_empty_memory(base_mem)) return true;
4572 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4573 assert(in(i) != NULL, "sane slice");
4574 if (in(i) == base_mem) return false; // should have been the sentinel value!
4575 }
4576 return true;
4577 }
4579 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4580 Node* n;
4581 n = mm->in(idx);
4582 if (mem == n) return true; // might be empty_memory()
4583 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4584 if (mem == n) return true;
4585 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4586 if (mem == n) return true;
4587 if (n == NULL) break;
4588 }
4589 return false;
4590 }
4591 #endif // !PRODUCT