Wed, 24 Oct 2012 14:33:22 -0700
7184394: add intrinsics to use AES instructions
Summary: Use new x86 AES instructions for AESCrypt.
Reviewed-by: twisti, kvn, roland
Contributed-by: tom.deneau@amd.com
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 TypePtr *t_adr, PhaseGVN *phase) {
107 const TypeOopPtr *tinst = t_adr->isa_oopptr();
108 if (tinst == NULL || !tinst->is_known_instance_field())
109 return mchain; // don't try to optimize non-instance types
110 uint instance_id = tinst->instance_id();
111 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
112 Node *prev = NULL;
113 Node *result = mchain;
114 while (prev != result) {
115 prev = result;
116 if (result == start_mem)
117 break; // hit one of our sentinels
118 // skip over a call which does not affect this memory slice
119 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
120 Node *proj_in = result->in(0);
121 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
122 break; // hit one of our sentinels
123 } else if (proj_in->is_Call()) {
124 CallNode *call = proj_in->as_Call();
125 if (!call->may_modify(t_adr, phase)) {
126 result = call->in(TypeFunc::Memory);
127 }
128 } else if (proj_in->is_Initialize()) {
129 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
130 // Stop if this is the initialization for the object instance which
131 // which contains this memory slice, otherwise skip over it.
132 if (alloc != NULL && alloc->_idx != instance_id) {
133 result = proj_in->in(TypeFunc::Memory);
134 }
135 } else if (proj_in->is_MemBar()) {
136 result = proj_in->in(TypeFunc::Memory);
137 } else {
138 assert(false, "unexpected projection");
139 }
140 } else if (result->is_ClearArray()) {
141 if (!ClearArrayNode::step_through(&result, instance_id, phase)) {
142 // Can not bypass initialization of the instance
143 // we are looking for.
144 break;
145 }
146 // Otherwise skip it (the call updated 'result' value).
147 } else if (result->is_MergeMem()) {
148 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
149 }
150 }
151 return result;
152 }
154 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
155 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
156 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
157 PhaseIterGVN *igvn = phase->is_IterGVN();
158 Node *result = mchain;
159 result = optimize_simple_memory_chain(result, t_adr, phase);
160 if (is_instance && igvn != NULL && result->is_Phi()) {
161 PhiNode *mphi = result->as_Phi();
162 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
163 const TypePtr *t = mphi->adr_type();
164 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
165 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
166 t->is_oopptr()->cast_to_exactness(true)
167 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
168 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
169 // clone the Phi with our address type
170 result = mphi->split_out_instance(t_adr, igvn);
171 } else {
172 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
173 }
174 }
175 return result;
176 }
178 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
179 uint alias_idx = phase->C->get_alias_index(tp);
180 Node *mem = mmem;
181 #ifdef ASSERT
182 {
183 // Check that current type is consistent with the alias index used during graph construction
184 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
185 bool consistent = adr_check == NULL || adr_check->empty() ||
186 phase->C->must_alias(adr_check, alias_idx );
187 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
188 if( !consistent && adr_check != NULL && !adr_check->empty() &&
189 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
190 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
191 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
192 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
193 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
194 // don't assert if it is dead code.
195 consistent = true;
196 }
197 if( !consistent ) {
198 st->print("alias_idx==%d, adr_check==", alias_idx);
199 if( adr_check == NULL ) {
200 st->print("NULL");
201 } else {
202 adr_check->dump();
203 }
204 st->cr();
205 print_alias_types();
206 assert(consistent, "adr_check must match alias idx");
207 }
208 }
209 #endif
210 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
211 // means an array I have not precisely typed yet. Do not do any
212 // alias stuff with it any time soon.
213 const TypeOopPtr *toop = tp->isa_oopptr();
214 if( tp->base() != Type::AnyPtr &&
215 !(toop &&
216 toop->klass() != NULL &&
217 toop->klass()->is_java_lang_Object() &&
218 toop->offset() == Type::OffsetBot) ) {
219 // compress paths and change unreachable cycles to TOP
220 // If not, we can update the input infinitely along a MergeMem cycle
221 // Equivalent code in PhiNode::Ideal
222 Node* m = phase->transform(mmem);
223 // If transformed to a MergeMem, get the desired slice
224 // Otherwise the returned node represents memory for every slice
225 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
226 // Update input if it is progress over what we have now
227 }
228 return mem;
229 }
231 //--------------------------Ideal_common---------------------------------------
232 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
233 // Unhook non-raw memories from complete (macro-expanded) initializations.
234 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
235 // If our control input is a dead region, kill all below the region
236 Node *ctl = in(MemNode::Control);
237 if (ctl && remove_dead_region(phase, can_reshape))
238 return this;
239 ctl = in(MemNode::Control);
240 // Don't bother trying to transform a dead node
241 if( ctl && ctl->is_top() ) return NodeSentinel;
243 PhaseIterGVN *igvn = phase->is_IterGVN();
244 // Wait if control on the worklist.
245 if (ctl && can_reshape && igvn != NULL) {
246 Node* bol = NULL;
247 Node* cmp = NULL;
248 if (ctl->in(0)->is_If()) {
249 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
250 bol = ctl->in(0)->in(1);
251 if (bol->is_Bool())
252 cmp = ctl->in(0)->in(1)->in(1);
253 }
254 if (igvn->_worklist.member(ctl) ||
255 (bol != NULL && igvn->_worklist.member(bol)) ||
256 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
257 // This control path may be dead.
258 // Delay this memory node transformation until the control is processed.
259 phase->is_IterGVN()->_worklist.push(this);
260 return NodeSentinel; // caller will return NULL
261 }
262 }
263 // Ignore if memory is dead, or self-loop
264 Node *mem = in(MemNode::Memory);
265 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
266 assert( mem != this, "dead loop in MemNode::Ideal" );
268 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
269 // This memory slice may be dead.
270 // Delay this mem node transformation until the memory is processed.
271 phase->is_IterGVN()->_worklist.push(this);
272 return NodeSentinel; // caller will return NULL
273 }
275 Node *address = in(MemNode::Address);
276 const Type *t_adr = phase->type( address );
277 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
279 if( can_reshape && igvn != NULL &&
280 (igvn->_worklist.member(address) ||
281 igvn->_worklist.size() > 0 && (phase->type(address) != adr_type())) ) {
282 // The address's base and type may change when the address is processed.
283 // Delay this mem node transformation until the address is processed.
284 phase->is_IterGVN()->_worklist.push(this);
285 return NodeSentinel; // caller will return NULL
286 }
288 // Do NOT remove or optimize the next lines: ensure a new alias index
289 // is allocated for an oop pointer type before Escape Analysis.
290 // Note: C++ will not remove it since the call has side effect.
291 if ( t_adr->isa_oopptr() ) {
292 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
293 }
295 #ifdef ASSERT
296 Node* base = NULL;
297 if (address->is_AddP())
298 base = address->in(AddPNode::Base);
299 assert(base == NULL || t_adr->isa_rawptr() ||
300 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
301 #endif
303 // Avoid independent memory operations
304 Node* old_mem = mem;
306 // The code which unhooks non-raw memories from complete (macro-expanded)
307 // initializations was removed. After macro-expansion all stores catched
308 // by Initialize node became raw stores and there is no information
309 // which memory slices they modify. So it is unsafe to move any memory
310 // operation above these stores. Also in most cases hooked non-raw memories
311 // were already unhooked by using information from detect_ptr_independence()
312 // and find_previous_store().
314 if (mem->is_MergeMem()) {
315 MergeMemNode* mmem = mem->as_MergeMem();
316 const TypePtr *tp = t_adr->is_ptr();
318 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
319 }
321 if (mem != old_mem) {
322 set_req(MemNode::Memory, mem);
323 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
324 return this;
325 }
327 // let the subclass continue analyzing...
328 return NULL;
329 }
331 // Helper function for proving some simple control dominations.
332 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
333 // Already assumes that 'dom' is available at 'sub', and that 'sub'
334 // is not a constant (dominated by the method's StartNode).
335 // Used by MemNode::find_previous_store to prove that the
336 // control input of a memory operation predates (dominates)
337 // an allocation it wants to look past.
338 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
339 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
340 return false; // Conservative answer for dead code
342 // Check 'dom'. Skip Proj and CatchProj nodes.
343 dom = dom->find_exact_control(dom);
344 if (dom == NULL || dom->is_top())
345 return false; // Conservative answer for dead code
347 if (dom == sub) {
348 // For the case when, for example, 'sub' is Initialize and the original
349 // 'dom' is Proj node of the 'sub'.
350 return false;
351 }
353 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
354 return true;
356 // 'dom' dominates 'sub' if its control edge and control edges
357 // of all its inputs dominate or equal to sub's control edge.
359 // Currently 'sub' is either Allocate, Initialize or Start nodes.
360 // Or Region for the check in LoadNode::Ideal();
361 // 'sub' should have sub->in(0) != NULL.
362 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
363 sub->is_Region(), "expecting only these nodes");
365 // Get control edge of 'sub'.
366 Node* orig_sub = sub;
367 sub = sub->find_exact_control(sub->in(0));
368 if (sub == NULL || sub->is_top())
369 return false; // Conservative answer for dead code
371 assert(sub->is_CFG(), "expecting control");
373 if (sub == dom)
374 return true;
376 if (sub->is_Start() || sub->is_Root())
377 return false;
379 {
380 // Check all control edges of 'dom'.
382 ResourceMark rm;
383 Arena* arena = Thread::current()->resource_area();
384 Node_List nlist(arena);
385 Unique_Node_List dom_list(arena);
387 dom_list.push(dom);
388 bool only_dominating_controls = false;
390 for (uint next = 0; next < dom_list.size(); next++) {
391 Node* n = dom_list.at(next);
392 if (n == orig_sub)
393 return false; // One of dom's inputs dominated by sub.
394 if (!n->is_CFG() && n->pinned()) {
395 // Check only own control edge for pinned non-control nodes.
396 n = n->find_exact_control(n->in(0));
397 if (n == NULL || n->is_top())
398 return false; // Conservative answer for dead code
399 assert(n->is_CFG(), "expecting control");
400 dom_list.push(n);
401 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
402 only_dominating_controls = true;
403 } else if (n->is_CFG()) {
404 if (n->dominates(sub, nlist))
405 only_dominating_controls = true;
406 else
407 return false;
408 } else {
409 // First, own control edge.
410 Node* m = n->find_exact_control(n->in(0));
411 if (m != NULL) {
412 if (m->is_top())
413 return false; // Conservative answer for dead code
414 dom_list.push(m);
415 }
416 // Now, the rest of edges.
417 uint cnt = n->req();
418 for (uint i = 1; i < cnt; i++) {
419 m = n->find_exact_control(n->in(i));
420 if (m == NULL || m->is_top())
421 continue;
422 dom_list.push(m);
423 }
424 }
425 }
426 return only_dominating_controls;
427 }
428 }
430 //---------------------detect_ptr_independence---------------------------------
431 // Used by MemNode::find_previous_store to prove that two base
432 // pointers are never equal.
433 // The pointers are accompanied by their associated allocations,
434 // if any, which have been previously discovered by the caller.
435 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
436 Node* p2, AllocateNode* a2,
437 PhaseTransform* phase) {
438 // Attempt to prove that these two pointers cannot be aliased.
439 // They may both manifestly be allocations, and they should differ.
440 // Or, if they are not both allocations, they can be distinct constants.
441 // Otherwise, one is an allocation and the other a pre-existing value.
442 if (a1 == NULL && a2 == NULL) { // neither an allocation
443 return (p1 != p2) && p1->is_Con() && p2->is_Con();
444 } else if (a1 != NULL && a2 != NULL) { // both allocations
445 return (a1 != a2);
446 } else if (a1 != NULL) { // one allocation a1
447 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
448 return all_controls_dominate(p2, a1);
449 } else { //(a2 != NULL) // one allocation a2
450 return all_controls_dominate(p1, a2);
451 }
452 return false;
453 }
456 // The logic for reordering loads and stores uses four steps:
457 // (a) Walk carefully past stores and initializations which we
458 // can prove are independent of this load.
459 // (b) Observe that the next memory state makes an exact match
460 // with self (load or store), and locate the relevant store.
461 // (c) Ensure that, if we were to wire self directly to the store,
462 // the optimizer would fold it up somehow.
463 // (d) Do the rewiring, and return, depending on some other part of
464 // the optimizer to fold up the load.
465 // This routine handles steps (a) and (b). Steps (c) and (d) are
466 // specific to loads and stores, so they are handled by the callers.
467 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
468 //
469 Node* MemNode::find_previous_store(PhaseTransform* phase) {
470 Node* ctrl = in(MemNode::Control);
471 Node* adr = in(MemNode::Address);
472 intptr_t offset = 0;
473 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
474 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
476 if (offset == Type::OffsetBot)
477 return NULL; // cannot unalias unless there are precise offsets
479 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
481 intptr_t size_in_bytes = memory_size();
483 Node* mem = in(MemNode::Memory); // start searching here...
485 int cnt = 50; // Cycle limiter
486 for (;;) { // While we can dance past unrelated stores...
487 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
489 if (mem->is_Store()) {
490 Node* st_adr = mem->in(MemNode::Address);
491 intptr_t st_offset = 0;
492 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
493 if (st_base == NULL)
494 break; // inscrutable pointer
495 if (st_offset != offset && st_offset != Type::OffsetBot) {
496 const int MAX_STORE = BytesPerLong;
497 if (st_offset >= offset + size_in_bytes ||
498 st_offset <= offset - MAX_STORE ||
499 st_offset <= offset - mem->as_Store()->memory_size()) {
500 // Success: The offsets are provably independent.
501 // (You may ask, why not just test st_offset != offset and be done?
502 // The answer is that stores of different sizes can co-exist
503 // in the same sequence of RawMem effects. We sometimes initialize
504 // a whole 'tile' of array elements with a single jint or jlong.)
505 mem = mem->in(MemNode::Memory);
506 continue; // (a) advance through independent store memory
507 }
508 }
509 if (st_base != base &&
510 detect_ptr_independence(base, alloc,
511 st_base,
512 AllocateNode::Ideal_allocation(st_base, phase),
513 phase)) {
514 // Success: The bases are provably independent.
515 mem = mem->in(MemNode::Memory);
516 continue; // (a) advance through independent store memory
517 }
519 // (b) At this point, if the bases or offsets do not agree, we lose,
520 // since we have not managed to prove 'this' and 'mem' independent.
521 if (st_base == base && st_offset == offset) {
522 return mem; // let caller handle steps (c), (d)
523 }
525 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
526 InitializeNode* st_init = mem->in(0)->as_Initialize();
527 AllocateNode* st_alloc = st_init->allocation();
528 if (st_alloc == NULL)
529 break; // something degenerated
530 bool known_identical = false;
531 bool known_independent = false;
532 if (alloc == st_alloc)
533 known_identical = true;
534 else if (alloc != NULL)
535 known_independent = true;
536 else if (all_controls_dominate(this, st_alloc))
537 known_independent = true;
539 if (known_independent) {
540 // The bases are provably independent: Either they are
541 // manifestly distinct allocations, or else the control
542 // of this load dominates the store's allocation.
543 int alias_idx = phase->C->get_alias_index(adr_type());
544 if (alias_idx == Compile::AliasIdxRaw) {
545 mem = st_alloc->in(TypeFunc::Memory);
546 } else {
547 mem = st_init->memory(alias_idx);
548 }
549 continue; // (a) advance through independent store memory
550 }
552 // (b) at this point, if we are not looking at a store initializing
553 // the same allocation we are loading from, we lose.
554 if (known_identical) {
555 // From caller, can_see_stored_value will consult find_captured_store.
556 return mem; // let caller handle steps (c), (d)
557 }
559 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
560 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
561 if (mem->is_Proj() && mem->in(0)->is_Call()) {
562 CallNode *call = mem->in(0)->as_Call();
563 if (!call->may_modify(addr_t, phase)) {
564 mem = call->in(TypeFunc::Memory);
565 continue; // (a) advance through independent call memory
566 }
567 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
568 mem = mem->in(0)->in(TypeFunc::Memory);
569 continue; // (a) advance through independent MemBar memory
570 } else if (mem->is_ClearArray()) {
571 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
572 // (the call updated 'mem' value)
573 continue; // (a) advance through independent allocation memory
574 } else {
575 // Can not bypass initialization of the instance
576 // we are looking for.
577 return mem;
578 }
579 } else if (mem->is_MergeMem()) {
580 int alias_idx = phase->C->get_alias_index(adr_type());
581 mem = mem->as_MergeMem()->memory_at(alias_idx);
582 continue; // (a) advance through independent MergeMem memory
583 }
584 }
586 // Unless there is an explicit 'continue', we must bail out here,
587 // because 'mem' is an inscrutable memory state (e.g., a call).
588 break;
589 }
591 return NULL; // bail out
592 }
594 //----------------------calculate_adr_type-------------------------------------
595 // Helper function. Notices when the given type of address hits top or bottom.
596 // Also, asserts a cross-check of the type against the expected address type.
597 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
598 if (t == Type::TOP) return NULL; // does not touch memory any more?
599 #ifdef PRODUCT
600 cross_check = NULL;
601 #else
602 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
603 #endif
604 const TypePtr* tp = t->isa_ptr();
605 if (tp == NULL) {
606 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
607 return TypePtr::BOTTOM; // touches lots of memory
608 } else {
609 #ifdef ASSERT
610 // %%%% [phh] We don't check the alias index if cross_check is
611 // TypeRawPtr::BOTTOM. Needs to be investigated.
612 if (cross_check != NULL &&
613 cross_check != TypePtr::BOTTOM &&
614 cross_check != TypeRawPtr::BOTTOM) {
615 // Recheck the alias index, to see if it has changed (due to a bug).
616 Compile* C = Compile::current();
617 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
618 "must stay in the original alias category");
619 // The type of the address must be contained in the adr_type,
620 // disregarding "null"-ness.
621 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
622 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
623 assert(cross_check->meet(tp_notnull) == cross_check,
624 "real address must not escape from expected memory type");
625 }
626 #endif
627 return tp;
628 }
629 }
631 //------------------------adr_phi_is_loop_invariant----------------------------
632 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
633 // loop is loop invariant. Make a quick traversal of Phi and associated
634 // CastPP nodes, looking to see if they are a closed group within the loop.
635 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
636 // The idea is that the phi-nest must boil down to only CastPP nodes
637 // with the same data. This implies that any path into the loop already
638 // includes such a CastPP, and so the original cast, whatever its input,
639 // must be covered by an equivalent cast, with an earlier control input.
640 ResourceMark rm;
642 // The loop entry input of the phi should be the unique dominating
643 // node for every Phi/CastPP in the loop.
644 Unique_Node_List closure;
645 closure.push(adr_phi->in(LoopNode::EntryControl));
647 // Add the phi node and the cast to the worklist.
648 Unique_Node_List worklist;
649 worklist.push(adr_phi);
650 if( cast != NULL ){
651 if( !cast->is_ConstraintCast() ) return false;
652 worklist.push(cast);
653 }
655 // Begin recursive walk of phi nodes.
656 while( worklist.size() ){
657 // Take a node off the worklist
658 Node *n = worklist.pop();
659 if( !closure.member(n) ){
660 // Add it to the closure.
661 closure.push(n);
662 // Make a sanity check to ensure we don't waste too much time here.
663 if( closure.size() > 20) return false;
664 // This node is OK if:
665 // - it is a cast of an identical value
666 // - or it is a phi node (then we add its inputs to the worklist)
667 // Otherwise, the node is not OK, and we presume the cast is not invariant
668 if( n->is_ConstraintCast() ){
669 worklist.push(n->in(1));
670 } else if( n->is_Phi() ) {
671 for( uint i = 1; i < n->req(); i++ ) {
672 worklist.push(n->in(i));
673 }
674 } else {
675 return false;
676 }
677 }
678 }
680 // Quit when the worklist is empty, and we've found no offending nodes.
681 return true;
682 }
684 //------------------------------Ideal_DU_postCCP-------------------------------
685 // Find any cast-away of null-ness and keep its control. Null cast-aways are
686 // going away in this pass and we need to make this memory op depend on the
687 // gating null check.
688 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
689 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
690 }
692 // I tried to leave the CastPP's in. This makes the graph more accurate in
693 // some sense; we get to keep around the knowledge that an oop is not-null
694 // after some test. Alas, the CastPP's interfere with GVN (some values are
695 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
696 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
697 // some of the more trivial cases in the optimizer. Removing more useless
698 // Phi's started allowing Loads to illegally float above null checks. I gave
699 // up on this approach. CNC 10/20/2000
700 // This static method may be called not from MemNode (EncodePNode calls it).
701 // Only the control edge of the node 'n' might be updated.
702 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
703 Node *skipped_cast = NULL;
704 // Need a null check? Regular static accesses do not because they are
705 // from constant addresses. Array ops are gated by the range check (which
706 // always includes a NULL check). Just check field ops.
707 if( n->in(MemNode::Control) == NULL ) {
708 // Scan upwards for the highest location we can place this memory op.
709 while( true ) {
710 switch( adr->Opcode() ) {
712 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
713 adr = adr->in(AddPNode::Base);
714 continue;
716 case Op_DecodeN: // No change to NULL-ness, so peek thru
717 case Op_DecodeNKlass:
718 adr = adr->in(1);
719 continue;
721 case Op_EncodeP:
722 case Op_EncodePKlass:
723 // EncodeP node's control edge could be set by this method
724 // when EncodeP node depends on CastPP node.
725 //
726 // Use its control edge for memory op because EncodeP may go away
727 // later when it is folded with following or preceding DecodeN node.
728 if (adr->in(0) == NULL) {
729 // Keep looking for cast nodes.
730 adr = adr->in(1);
731 continue;
732 }
733 ccp->hash_delete(n);
734 n->set_req(MemNode::Control, adr->in(0));
735 ccp->hash_insert(n);
736 return n;
738 case Op_CastPP:
739 // If the CastPP is useless, just peek on through it.
740 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
741 // Remember the cast that we've peeked though. If we peek
742 // through more than one, then we end up remembering the highest
743 // one, that is, if in a loop, the one closest to the top.
744 skipped_cast = adr;
745 adr = adr->in(1);
746 continue;
747 }
748 // CastPP is going away in this pass! We need this memory op to be
749 // control-dependent on the test that is guarding the CastPP.
750 ccp->hash_delete(n);
751 n->set_req(MemNode::Control, adr->in(0));
752 ccp->hash_insert(n);
753 return n;
755 case Op_Phi:
756 // Attempt to float above a Phi to some dominating point.
757 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
758 // If we've already peeked through a Cast (which could have set the
759 // control), we can't float above a Phi, because the skipped Cast
760 // may not be loop invariant.
761 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
762 adr = adr->in(1);
763 continue;
764 }
765 }
767 // Intentional fallthrough!
769 // No obvious dominating point. The mem op is pinned below the Phi
770 // by the Phi itself. If the Phi goes away (no true value is merged)
771 // then the mem op can float, but not indefinitely. It must be pinned
772 // behind the controls leading to the Phi.
773 case Op_CheckCastPP:
774 // These usually stick around to change address type, however a
775 // useless one can be elided and we still need to pick up a control edge
776 if (adr->in(0) == NULL) {
777 // This CheckCastPP node has NO control and is likely useless. But we
778 // need check further up the ancestor chain for a control input to keep
779 // the node in place. 4959717.
780 skipped_cast = adr;
781 adr = adr->in(1);
782 continue;
783 }
784 ccp->hash_delete(n);
785 n->set_req(MemNode::Control, adr->in(0));
786 ccp->hash_insert(n);
787 return n;
789 // List of "safe" opcodes; those that implicitly block the memory
790 // op below any null check.
791 case Op_CastX2P: // no null checks on native pointers
792 case Op_Parm: // 'this' pointer is not null
793 case Op_LoadP: // Loading from within a klass
794 case Op_LoadN: // Loading from within a klass
795 case Op_LoadKlass: // Loading from within a klass
796 case Op_LoadNKlass: // Loading from within a klass
797 case Op_ConP: // Loading from a klass
798 case Op_ConN: // Loading from a klass
799 case Op_ConNKlass: // Loading from a klass
800 case Op_CreateEx: // Sucking up the guts of an exception oop
801 case Op_Con: // Reading from TLS
802 case Op_CMoveP: // CMoveP is pinned
803 case Op_CMoveN: // CMoveN is pinned
804 break; // No progress
806 case Op_Proj: // Direct call to an allocation routine
807 case Op_SCMemProj: // Memory state from store conditional ops
808 #ifdef ASSERT
809 {
810 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
811 const Node* call = adr->in(0);
812 if (call->is_CallJava()) {
813 const CallJavaNode* call_java = call->as_CallJava();
814 const TypeTuple *r = call_java->tf()->range();
815 assert(r->cnt() > TypeFunc::Parms, "must return value");
816 const Type* ret_type = r->field_at(TypeFunc::Parms);
817 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
818 // We further presume that this is one of
819 // new_instance_Java, new_array_Java, or
820 // the like, but do not assert for this.
821 } else if (call->is_Allocate()) {
822 // similar case to new_instance_Java, etc.
823 } else if (!call->is_CallLeaf()) {
824 // Projections from fetch_oop (OSR) are allowed as well.
825 ShouldNotReachHere();
826 }
827 }
828 #endif
829 break;
830 default:
831 ShouldNotReachHere();
832 }
833 break;
834 }
835 }
837 return NULL; // No progress
838 }
841 //=============================================================================
842 uint LoadNode::size_of() const { return sizeof(*this); }
843 uint LoadNode::cmp( const Node &n ) const
844 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
845 const Type *LoadNode::bottom_type() const { return _type; }
846 uint LoadNode::ideal_reg() const {
847 return _type->ideal_reg();
848 }
850 #ifndef PRODUCT
851 void LoadNode::dump_spec(outputStream *st) const {
852 MemNode::dump_spec(st);
853 if( !Verbose && !WizardMode ) {
854 // standard dump does this in Verbose and WizardMode
855 st->print(" #"); _type->dump_on(st);
856 }
857 }
858 #endif
860 #ifdef ASSERT
861 //----------------------------is_immutable_value-------------------------------
862 // Helper function to allow a raw load without control edge for some cases
863 bool LoadNode::is_immutable_value(Node* adr) {
864 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
865 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
866 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
867 in_bytes(JavaThread::osthread_offset())));
868 }
869 #endif
871 //----------------------------LoadNode::make-----------------------------------
872 // Polymorphic factory method:
873 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
874 Compile* C = gvn.C;
876 // sanity check the alias category against the created node type
877 assert(!(adr_type->isa_oopptr() &&
878 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
879 "use LoadKlassNode instead");
880 assert(!(adr_type->isa_aryptr() &&
881 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
882 "use LoadRangeNode instead");
883 // Check control edge of raw loads
884 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
885 // oop will be recorded in oop map if load crosses safepoint
886 rt->isa_oopptr() || is_immutable_value(adr),
887 "raw memory operations should have control edge");
888 switch (bt) {
889 case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() );
890 case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() );
891 case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int() );
892 case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() );
893 case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() );
894 case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() );
895 case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt );
896 case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt );
897 case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() );
898 case T_OBJECT:
899 #ifdef _LP64
900 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
901 Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
902 return new (C) DecodeNNode(load, load->bottom_type()->make_ptr());
903 } else
904 #endif
905 {
906 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
907 return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
908 }
909 }
910 ShouldNotReachHere();
911 return (LoadNode*)NULL;
912 }
914 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
915 bool require_atomic = true;
916 return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
917 }
922 //------------------------------hash-------------------------------------------
923 uint LoadNode::hash() const {
924 // unroll addition of interesting fields
925 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
926 }
928 //---------------------------can_see_stored_value------------------------------
929 // This routine exists to make sure this set of tests is done the same
930 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
931 // will change the graph shape in a way which makes memory alive twice at the
932 // same time (uses the Oracle model of aliasing), then some
933 // LoadXNode::Identity will fold things back to the equivalence-class model
934 // of aliasing.
935 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
936 Node* ld_adr = in(MemNode::Address);
938 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
939 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
940 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
941 atp->field() != NULL && !atp->field()->is_volatile()) {
942 uint alias_idx = atp->index();
943 bool final = atp->field()->is_final();
944 Node* result = NULL;
945 Node* current = st;
946 // Skip through chains of MemBarNodes checking the MergeMems for
947 // new states for the slice of this load. Stop once any other
948 // kind of node is encountered. Loads from final memory can skip
949 // through any kind of MemBar but normal loads shouldn't skip
950 // through MemBarAcquire since the could allow them to move out of
951 // a synchronized region.
952 while (current->is_Proj()) {
953 int opc = current->in(0)->Opcode();
954 if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) ||
955 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder ||
956 opc == Op_MemBarReleaseLock) {
957 Node* mem = current->in(0)->in(TypeFunc::Memory);
958 if (mem->is_MergeMem()) {
959 MergeMemNode* merge = mem->as_MergeMem();
960 Node* new_st = merge->memory_at(alias_idx);
961 if (new_st == merge->base_memory()) {
962 // Keep searching
963 current = merge->base_memory();
964 continue;
965 }
966 // Save the new memory state for the slice and fall through
967 // to exit.
968 result = new_st;
969 }
970 }
971 break;
972 }
973 if (result != NULL) {
974 st = result;
975 }
976 }
979 // Loop around twice in the case Load -> Initialize -> Store.
980 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
981 for (int trip = 0; trip <= 1; trip++) {
983 if (st->is_Store()) {
984 Node* st_adr = st->in(MemNode::Address);
985 if (!phase->eqv(st_adr, ld_adr)) {
986 // Try harder before giving up... Match raw and non-raw pointers.
987 intptr_t st_off = 0;
988 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
989 if (alloc == NULL) return NULL;
990 intptr_t ld_off = 0;
991 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
992 if (alloc != allo2) return NULL;
993 if (ld_off != st_off) return NULL;
994 // At this point we have proven something like this setup:
995 // A = Allocate(...)
996 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
997 // S = StoreQ(, AddP(, A.Parm , #Off), V)
998 // (Actually, we haven't yet proven the Q's are the same.)
999 // In other words, we are loading from a casted version of
1000 // the same pointer-and-offset that we stored to.
1001 // Thus, we are able to replace L by V.
1002 }
1003 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1004 if (store_Opcode() != st->Opcode())
1005 return NULL;
1006 return st->in(MemNode::ValueIn);
1007 }
1009 intptr_t offset = 0; // scratch
1011 // A load from a freshly-created object always returns zero.
1012 // (This can happen after LoadNode::Ideal resets the load's memory input
1013 // to find_captured_store, which returned InitializeNode::zero_memory.)
1014 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1015 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
1016 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
1017 // return a zero value for the load's basic type
1018 // (This is one of the few places where a generic PhaseTransform
1019 // can create new nodes. Think of it as lazily manifesting
1020 // virtually pre-existing constants.)
1021 return phase->zerocon(memory_type());
1022 }
1024 // A load from an initialization barrier can match a captured store.
1025 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1026 InitializeNode* init = st->in(0)->as_Initialize();
1027 AllocateNode* alloc = init->allocation();
1028 if (alloc != NULL &&
1029 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
1030 // examine a captured store value
1031 st = init->find_captured_store(offset, memory_size(), phase);
1032 if (st != NULL)
1033 continue; // take one more trip around
1034 }
1035 }
1037 break;
1038 }
1040 return NULL;
1041 }
1043 //----------------------is_instance_field_load_with_local_phi------------------
1044 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1045 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
1046 in(MemNode::Address)->is_AddP() ) {
1047 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
1048 // Only instances.
1049 if( t_oop != NULL && t_oop->is_known_instance_field() &&
1050 t_oop->offset() != Type::OffsetBot &&
1051 t_oop->offset() != Type::OffsetTop) {
1052 return true;
1053 }
1054 }
1055 return false;
1056 }
1058 //------------------------------Identity---------------------------------------
1059 // Loads are identity if previous store is to same address
1060 Node *LoadNode::Identity( PhaseTransform *phase ) {
1061 // If the previous store-maker is the right kind of Store, and the store is
1062 // to the same address, then we are equal to the value stored.
1063 Node* mem = in(MemNode::Memory);
1064 Node* value = can_see_stored_value(mem, phase);
1065 if( value ) {
1066 // byte, short & char stores truncate naturally.
1067 // A load has to load the truncated value which requires
1068 // some sort of masking operation and that requires an
1069 // Ideal call instead of an Identity call.
1070 if (memory_size() < BytesPerInt) {
1071 // If the input to the store does not fit with the load's result type,
1072 // it must be truncated via an Ideal call.
1073 if (!phase->type(value)->higher_equal(phase->type(this)))
1074 return this;
1075 }
1076 // (This works even when value is a Con, but LoadNode::Value
1077 // usually runs first, producing the singleton type of the Con.)
1078 return value;
1079 }
1081 // Search for an existing data phi which was generated before for the same
1082 // instance's field to avoid infinite generation of phis in a loop.
1083 Node *region = mem->in(0);
1084 if (is_instance_field_load_with_local_phi(region)) {
1085 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
1086 int this_index = phase->C->get_alias_index(addr_t);
1087 int this_offset = addr_t->offset();
1088 int this_id = addr_t->is_oopptr()->instance_id();
1089 const Type* this_type = bottom_type();
1090 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1091 Node* phi = region->fast_out(i);
1092 if (phi->is_Phi() && phi != mem &&
1093 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
1094 return phi;
1095 }
1096 }
1097 }
1099 return this;
1100 }
1103 // Returns true if the AliasType refers to the field that holds the
1104 // cached box array. Currently only handles the IntegerCache case.
1105 static bool is_autobox_cache(Compile::AliasType* atp) {
1106 if (atp != NULL && atp->field() != NULL) {
1107 ciField* field = atp->field();
1108 ciSymbol* klass = field->holder()->name();
1109 if (field->name() == ciSymbol::cache_field_name() &&
1110 field->holder()->uses_default_loader() &&
1111 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1112 return true;
1113 }
1114 }
1115 return false;
1116 }
1118 // Fetch the base value in the autobox array
1119 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
1120 if (atp != NULL && atp->field() != NULL) {
1121 ciField* field = atp->field();
1122 ciSymbol* klass = field->holder()->name();
1123 if (field->name() == ciSymbol::cache_field_name() &&
1124 field->holder()->uses_default_loader() &&
1125 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1126 assert(field->is_constant(), "what?");
1127 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
1128 // Fetch the box object at the base of the array and get its value
1129 ciInstance* box = array->obj_at(0)->as_instance();
1130 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1131 if (ik->nof_nonstatic_fields() == 1) {
1132 // This should be true nonstatic_field_at requires calling
1133 // nof_nonstatic_fields so check it anyway
1134 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1135 cache_offset = c.as_int();
1136 }
1137 return true;
1138 }
1139 }
1140 return false;
1141 }
1143 // Returns true if the AliasType refers to the value field of an
1144 // autobox object. Currently only handles Integer.
1145 static bool is_autobox_object(Compile::AliasType* atp) {
1146 if (atp != NULL && atp->field() != NULL) {
1147 ciField* field = atp->field();
1148 ciSymbol* klass = field->holder()->name();
1149 if (field->name() == ciSymbol::value_name() &&
1150 field->holder()->uses_default_loader() &&
1151 klass == ciSymbol::java_lang_Integer()) {
1152 return true;
1153 }
1154 }
1155 return false;
1156 }
1159 // We're loading from an object which has autobox behaviour.
1160 // If this object is result of a valueOf call we'll have a phi
1161 // merging a newly allocated object and a load from the cache.
1162 // We want to replace this load with the original incoming
1163 // argument to the valueOf call.
1164 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1165 Node* base = in(Address)->in(AddPNode::Base);
1166 if (base->is_Phi() && base->req() == 3) {
1167 AllocateNode* allocation = NULL;
1168 int allocation_index = -1;
1169 int load_index = -1;
1170 for (uint i = 1; i < base->req(); i++) {
1171 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1172 if (allocation != NULL) {
1173 allocation_index = i;
1174 load_index = 3 - allocation_index;
1175 break;
1176 }
1177 }
1178 bool has_load = ( allocation != NULL &&
1179 (base->in(load_index)->is_Load() ||
1180 base->in(load_index)->is_DecodeN() &&
1181 base->in(load_index)->in(1)->is_Load()) );
1182 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
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.
1186 Node* mem_phi = in(Memory);
1187 Node* offset = in(Address)->in(AddPNode::Offset);
1188 Node* region = base->in(0);
1190 Node* in1 = clone();
1191 Node* in1_addr = in1->in(Address)->clone();
1192 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1193 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1194 in1_addr->set_req(AddPNode::Offset, offset);
1195 in1->set_req(0, region->in(allocation_index));
1196 in1->set_req(Address, in1_addr);
1197 in1->set_req(Memory, mem_phi->in(allocation_index));
1199 Node* in2 = clone();
1200 Node* in2_addr = in2->in(Address)->clone();
1201 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1202 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1203 in2_addr->set_req(AddPNode::Offset, offset);
1204 in2->set_req(0, region->in(load_index));
1205 in2->set_req(Address, in2_addr);
1206 in2->set_req(Memory, mem_phi->in(load_index));
1208 in1_addr = phase->transform(in1_addr);
1209 in1 = phase->transform(in1);
1210 in2_addr = phase->transform(in2_addr);
1211 in2 = phase->transform(in2);
1213 PhiNode* result = PhiNode::make_blank(region, this);
1214 result->set_req(allocation_index, in1);
1215 result->set_req(load_index, in2);
1216 return result;
1217 }
1218 } else if (base->is_Load() ||
1219 base->is_DecodeN() && base->in(1)->is_Load()) {
1220 if (base->is_DecodeN()) {
1221 // Get LoadN node which loads cached Integer object
1222 base = base->in(1);
1223 }
1224 // Eliminate the load of Integer.value for integers from the cache
1225 // array by deriving the value from the index into the array.
1226 // Capture the offset of the load and then reverse the computation.
1227 Node* load_base = base->in(Address)->in(AddPNode::Base);
1228 if (load_base->is_DecodeN()) {
1229 // Get LoadN node which loads IntegerCache.cache field
1230 load_base = load_base->in(1);
1231 }
1232 if (load_base != NULL) {
1233 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1234 intptr_t cache_offset;
1235 int shift = -1;
1236 Node* cache = NULL;
1237 if (is_autobox_cache(atp)) {
1238 shift = exact_log2(type2aelembytes(T_OBJECT));
1239 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1240 }
1241 if (cache != NULL && base->in(Address)->is_AddP()) {
1242 Node* elements[4];
1243 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1244 int cache_low;
1245 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1246 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1247 // Add up all the offsets making of the address of the load
1248 Node* result = elements[0];
1249 for (int i = 1; i < count; i++) {
1250 result = phase->transform(new (phase->C) AddXNode(result, elements[i]));
1251 }
1252 // Remove the constant offset from the address and then
1253 // remove the scaling of the offset to recover the original index.
1254 result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-offset)));
1255 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1256 // Peel the shift off directly but wrap it in a dummy node
1257 // since Ideal can't return existing nodes
1258 result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));
1259 } else {
1260 result = new (phase->C) RShiftXNode(result, phase->intcon(shift));
1261 }
1262 #ifdef _LP64
1263 result = new (phase->C) ConvL2INode(phase->transform(result));
1264 #endif
1265 return result;
1266 }
1267 }
1268 }
1269 }
1270 return NULL;
1271 }
1273 //------------------------------split_through_phi------------------------------
1274 // Split instance field load through Phi.
1275 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1276 Node* mem = in(MemNode::Memory);
1277 Node* address = in(MemNode::Address);
1278 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1279 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1281 assert(mem->is_Phi() && (t_oop != NULL) &&
1282 t_oop->is_known_instance_field(), "invalide conditions");
1284 Node *region = mem->in(0);
1285 if (region == NULL) {
1286 return NULL; // Wait stable graph
1287 }
1288 uint cnt = mem->req();
1289 for (uint i = 1; i < cnt; i++) {
1290 Node* rc = region->in(i);
1291 if (rc == NULL || phase->type(rc) == Type::TOP)
1292 return NULL; // Wait stable graph
1293 Node *in = mem->in(i);
1294 if (in == NULL) {
1295 return NULL; // Wait stable graph
1296 }
1297 }
1298 // Check for loop invariant.
1299 if (cnt == 3) {
1300 for (uint i = 1; i < cnt; i++) {
1301 Node *in = mem->in(i);
1302 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1303 if (m == mem) {
1304 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
1305 return this;
1306 }
1307 }
1308 }
1309 // Split through Phi (see original code in loopopts.cpp).
1310 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1312 // Do nothing here if Identity will find a value
1313 // (to avoid infinite chain of value phis generation).
1314 if (!phase->eqv(this, this->Identity(phase)))
1315 return NULL;
1317 // Skip the split if the region dominates some control edge of the address.
1318 if (!MemNode::all_controls_dominate(address, region))
1319 return NULL;
1321 const Type* this_type = this->bottom_type();
1322 int this_index = phase->C->get_alias_index(addr_t);
1323 int this_offset = addr_t->offset();
1324 int this_iid = addr_t->is_oopptr()->instance_id();
1325 PhaseIterGVN *igvn = phase->is_IterGVN();
1326 Node *phi = new (igvn->C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1327 for (uint i = 1; i < region->req(); i++) {
1328 Node *x;
1329 Node* the_clone = NULL;
1330 if (region->in(i) == phase->C->top()) {
1331 x = phase->C->top(); // Dead path? Use a dead data op
1332 } else {
1333 x = this->clone(); // Else clone up the data op
1334 the_clone = x; // Remember for possible deletion.
1335 // Alter data node to use pre-phi inputs
1336 if (this->in(0) == region) {
1337 x->set_req(0, region->in(i));
1338 } else {
1339 x->set_req(0, NULL);
1340 }
1341 for (uint j = 1; j < this->req(); j++) {
1342 Node *in = this->in(j);
1343 if (in->is_Phi() && in->in(0) == region)
1344 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
1345 }
1346 }
1347 // Check for a 'win' on some paths
1348 const Type *t = x->Value(igvn);
1350 bool singleton = t->singleton();
1352 // See comments in PhaseIdealLoop::split_thru_phi().
1353 if (singleton && t == Type::TOP) {
1354 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1355 }
1357 if (singleton) {
1358 x = igvn->makecon(t);
1359 } else {
1360 // We now call Identity to try to simplify the cloned node.
1361 // Note that some Identity methods call phase->type(this).
1362 // Make sure that the type array is big enough for
1363 // our new node, even though we may throw the node away.
1364 // (This tweaking with igvn only works because x is a new node.)
1365 igvn->set_type(x, t);
1366 // If x is a TypeNode, capture any more-precise type permanently into Node
1367 // otherwise it will be not updated during igvn->transform since
1368 // igvn->type(x) is set to x->Value() already.
1369 x->raise_bottom_type(t);
1370 Node *y = x->Identity(igvn);
1371 if (y != x) {
1372 x = y;
1373 } else {
1374 y = igvn->hash_find(x);
1375 if (y) {
1376 x = y;
1377 } else {
1378 // Else x is a new node we are keeping
1379 // We do not need register_new_node_with_optimizer
1380 // because set_type has already been called.
1381 igvn->_worklist.push(x);
1382 }
1383 }
1384 }
1385 if (x != the_clone && the_clone != NULL)
1386 igvn->remove_dead_node(the_clone);
1387 phi->set_req(i, x);
1388 }
1389 // Record Phi
1390 igvn->register_new_node_with_optimizer(phi);
1391 return phi;
1392 }
1394 //------------------------------Ideal------------------------------------------
1395 // If the load is from Field memory and the pointer is non-null, we can
1396 // zero out the control input.
1397 // If the offset is constant and the base is an object allocation,
1398 // try to hook me up to the exact initializing store.
1399 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1400 Node* p = MemNode::Ideal_common(phase, can_reshape);
1401 if (p) return (p == NodeSentinel) ? NULL : p;
1403 Node* ctrl = in(MemNode::Control);
1404 Node* address = in(MemNode::Address);
1406 // Skip up past a SafePoint control. Cannot do this for Stores because
1407 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1408 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1409 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1410 ctrl = ctrl->in(0);
1411 set_req(MemNode::Control,ctrl);
1412 }
1414 intptr_t ignore = 0;
1415 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1416 if (base != NULL
1417 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1418 // Check for useless control edge in some common special cases
1419 if (in(MemNode::Control) != NULL
1420 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1421 && all_controls_dominate(base, phase->C->start())) {
1422 // A method-invariant, non-null address (constant or 'this' argument).
1423 set_req(MemNode::Control, NULL);
1424 }
1426 if (EliminateAutoBox && can_reshape) {
1427 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null");
1428 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1429 if (is_autobox_object(atp)) {
1430 Node* result = eliminate_autobox(phase);
1431 if (result != NULL) return result;
1432 }
1433 }
1434 }
1436 Node* mem = in(MemNode::Memory);
1437 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1439 if (addr_t != NULL) {
1440 // try to optimize our memory input
1441 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1442 if (opt_mem != mem) {
1443 set_req(MemNode::Memory, opt_mem);
1444 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1445 return this;
1446 }
1447 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1448 if (can_reshape && opt_mem->is_Phi() &&
1449 (t_oop != NULL) && t_oop->is_known_instance_field()) {
1450 PhaseIterGVN *igvn = phase->is_IterGVN();
1451 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1452 // Delay this transformation until memory Phi is processed.
1453 phase->is_IterGVN()->_worklist.push(this);
1454 return NULL;
1455 }
1456 // Split instance field load through Phi.
1457 Node* result = split_through_phi(phase);
1458 if (result != NULL) return result;
1459 }
1460 }
1462 // Check for prior store with a different base or offset; make Load
1463 // independent. Skip through any number of them. Bail out if the stores
1464 // are in an endless dead cycle and report no progress. This is a key
1465 // transform for Reflection. However, if after skipping through the Stores
1466 // we can't then fold up against a prior store do NOT do the transform as
1467 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1468 // array memory alive twice: once for the hoisted Load and again after the
1469 // bypassed Store. This situation only works if EVERYBODY who does
1470 // anti-dependence work knows how to bypass. I.e. we need all
1471 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1472 // the alias index stuff. So instead, peek through Stores and IFF we can
1473 // fold up, do so.
1474 Node* prev_mem = find_previous_store(phase);
1475 // Steps (a), (b): Walk past independent stores to find an exact match.
1476 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1477 // (c) See if we can fold up on the spot, but don't fold up here.
1478 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1479 // just return a prior value, which is done by Identity calls.
1480 if (can_see_stored_value(prev_mem, phase)) {
1481 // Make ready for step (d):
1482 set_req(MemNode::Memory, prev_mem);
1483 return this;
1484 }
1485 }
1487 return NULL; // No further progress
1488 }
1490 // Helper to recognize certain Klass fields which are invariant across
1491 // some group of array types (e.g., int[] or all T[] where T < Object).
1492 const Type*
1493 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1494 ciKlass* klass) const {
1495 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1496 // The field is Klass::_modifier_flags. Return its (constant) value.
1497 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1498 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1499 return TypeInt::make(klass->modifier_flags());
1500 }
1501 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1502 // The field is Klass::_access_flags. Return its (constant) value.
1503 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1504 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1505 return TypeInt::make(klass->access_flags());
1506 }
1507 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1508 // The field is Klass::_layout_helper. Return its constant value if known.
1509 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1510 return TypeInt::make(klass->layout_helper());
1511 }
1513 // No match.
1514 return NULL;
1515 }
1517 //------------------------------Value-----------------------------------------
1518 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1519 // Either input is TOP ==> the result is TOP
1520 Node* mem = in(MemNode::Memory);
1521 const Type *t1 = phase->type(mem);
1522 if (t1 == Type::TOP) return Type::TOP;
1523 Node* adr = in(MemNode::Address);
1524 const TypePtr* tp = phase->type(adr)->isa_ptr();
1525 if (tp == NULL || tp->empty()) return Type::TOP;
1526 int off = tp->offset();
1527 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1528 Compile* C = phase->C;
1530 // Try to guess loaded type from pointer type
1531 if (tp->base() == Type::AryPtr) {
1532 const Type *t = tp->is_aryptr()->elem();
1533 // Don't do this for integer types. There is only potential profit if
1534 // the element type t is lower than _type; that is, for int types, if _type is
1535 // more restrictive than t. This only happens here if one is short and the other
1536 // char (both 16 bits), and in those cases we've made an intentional decision
1537 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1538 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1539 //
1540 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1541 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1542 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1543 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1544 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1545 // In fact, that could have been the original type of p1, and p1 could have
1546 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1547 // expression (LShiftL quux 3) independently optimized to the constant 8.
1548 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1549 && (_type->isa_vect() == NULL)
1550 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1551 // t might actually be lower than _type, if _type is a unique
1552 // concrete subclass of abstract class t.
1553 // Make sure the reference is not into the header, by comparing
1554 // the offset against the offset of the start of the array's data.
1555 // Different array types begin at slightly different offsets (12 vs. 16).
1556 // We choose T_BYTE as an example base type that is least restrictive
1557 // as to alignment, which will therefore produce the smallest
1558 // possible base offset.
1559 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1560 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1561 const Type* jt = t->join(_type);
1562 // In any case, do not allow the join, per se, to empty out the type.
1563 if (jt->empty() && !t->empty()) {
1564 // This can happen if a interface-typed array narrows to a class type.
1565 jt = _type;
1566 }
1568 if (EliminateAutoBox && adr->is_AddP()) {
1569 // The pointers in the autobox arrays are always non-null
1570 Node* base = adr->in(AddPNode::Base);
1571 if (base != NULL &&
1572 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) {
1573 Compile::AliasType* atp = C->alias_type(base->adr_type());
1574 if (is_autobox_cache(atp)) {
1575 return jt->join(TypePtr::NOTNULL)->is_ptr();
1576 }
1577 }
1578 }
1579 return jt;
1580 }
1581 }
1582 } else if (tp->base() == Type::InstPtr) {
1583 ciEnv* env = C->env();
1584 const TypeInstPtr* tinst = tp->is_instptr();
1585 ciKlass* klass = tinst->klass();
1586 assert( off != Type::OffsetBot ||
1587 // arrays can be cast to Objects
1588 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1589 // unsafe field access may not have a constant offset
1590 C->has_unsafe_access(),
1591 "Field accesses must be precise" );
1592 // For oop loads, we expect the _type to be precise
1593 if (klass == env->String_klass() &&
1594 adr->is_AddP() && off != Type::OffsetBot) {
1595 // For constant Strings treat the final fields as compile time constants.
1596 Node* base = adr->in(AddPNode::Base);
1597 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1598 if (t != NULL && t->singleton()) {
1599 ciField* field = env->String_klass()->get_field_by_offset(off, false);
1600 if (field != NULL && field->is_final()) {
1601 ciObject* string = t->const_oop();
1602 ciConstant constant = string->as_instance()->field_value(field);
1603 if (constant.basic_type() == T_INT) {
1604 return TypeInt::make(constant.as_int());
1605 } else if (constant.basic_type() == T_ARRAY) {
1606 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1607 return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1608 } else {
1609 return TypeOopPtr::make_from_constant(constant.as_object(), true);
1610 }
1611 }
1612 }
1613 }
1614 }
1615 // Optimizations for constant objects
1616 ciObject* const_oop = tinst->const_oop();
1617 if (const_oop != NULL) {
1618 // For constant CallSites treat the target field as a compile time constant.
1619 if (const_oop->is_call_site()) {
1620 ciCallSite* call_site = const_oop->as_call_site();
1621 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1622 if (field != NULL && field->is_call_site_target()) {
1623 ciMethodHandle* target = call_site->get_target();
1624 if (target != NULL) { // just in case
1625 ciConstant constant(T_OBJECT, target);
1626 const Type* t;
1627 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1628 t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1629 } else {
1630 t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1631 }
1632 // Add a dependence for invalidation of the optimization.
1633 if (!call_site->is_constant_call_site()) {
1634 C->dependencies()->assert_call_site_target_value(call_site, target);
1635 }
1636 return t;
1637 }
1638 }
1639 }
1640 }
1641 } else if (tp->base() == Type::KlassPtr) {
1642 assert( off != Type::OffsetBot ||
1643 // arrays can be cast to Objects
1644 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1645 // also allow array-loading from the primary supertype
1646 // array during subtype checks
1647 Opcode() == Op_LoadKlass,
1648 "Field accesses must be precise" );
1649 // For klass/static loads, we expect the _type to be precise
1650 }
1652 const TypeKlassPtr *tkls = tp->isa_klassptr();
1653 if (tkls != NULL && !StressReflectiveCode) {
1654 ciKlass* klass = tkls->klass();
1655 if (klass->is_loaded() && tkls->klass_is_exact()) {
1656 // We are loading a field from a Klass metaobject whose identity
1657 // is known at compile time (the type is "exact" or "precise").
1658 // Check for fields we know are maintained as constants by the VM.
1659 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1660 // The field is Klass::_super_check_offset. Return its (constant) value.
1661 // (Folds up type checking code.)
1662 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1663 return TypeInt::make(klass->super_check_offset());
1664 }
1665 // Compute index into primary_supers array
1666 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1667 // Check for overflowing; use unsigned compare to handle the negative case.
1668 if( depth < ciKlass::primary_super_limit() ) {
1669 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1670 // (Folds up type checking code.)
1671 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1672 ciKlass *ss = klass->super_of_depth(depth);
1673 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1674 }
1675 const Type* aift = load_array_final_field(tkls, klass);
1676 if (aift != NULL) return aift;
1677 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())
1678 && klass->is_array_klass()) {
1679 // The field is ArrayKlass::_component_mirror. Return its (constant) value.
1680 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1681 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1682 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1683 }
1684 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1685 // The field is Klass::_java_mirror. Return its (constant) value.
1686 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1687 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1688 return TypeInstPtr::make(klass->java_mirror());
1689 }
1690 }
1692 // We can still check if we are loading from the primary_supers array at a
1693 // shallow enough depth. Even though the klass is not exact, entries less
1694 // than or equal to its super depth are correct.
1695 if (klass->is_loaded() ) {
1696 ciType *inner = klass;
1697 while( inner->is_obj_array_klass() )
1698 inner = inner->as_obj_array_klass()->base_element_type();
1699 if( inner->is_instance_klass() &&
1700 !inner->as_instance_klass()->flags().is_interface() ) {
1701 // Compute index into primary_supers array
1702 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1703 // Check for overflowing; use unsigned compare to handle the negative case.
1704 if( depth < ciKlass::primary_super_limit() &&
1705 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1706 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1707 // (Folds up type checking code.)
1708 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1709 ciKlass *ss = klass->super_of_depth(depth);
1710 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1711 }
1712 }
1713 }
1715 // If the type is enough to determine that the thing is not an array,
1716 // we can give the layout_helper a positive interval type.
1717 // This will help short-circuit some reflective code.
1718 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1719 && !klass->is_array_klass() // not directly typed as an array
1720 && !klass->is_interface() // specifically not Serializable & Cloneable
1721 && !klass->is_java_lang_Object() // not the supertype of all T[]
1722 ) {
1723 // Note: When interfaces are reliable, we can narrow the interface
1724 // test to (klass != Serializable && klass != Cloneable).
1725 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1726 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1727 // The key property of this type is that it folds up tests
1728 // for array-ness, since it proves that the layout_helper is positive.
1729 // Thus, a generic value like the basic object layout helper works fine.
1730 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1731 }
1732 }
1734 // If we are loading from a freshly-allocated object, produce a zero,
1735 // if the load is provably beyond the header of the object.
1736 // (Also allow a variable load from a fresh array to produce zero.)
1737 const TypeOopPtr *tinst = tp->isa_oopptr();
1738 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1739 if (ReduceFieldZeroing || is_instance) {
1740 Node* value = can_see_stored_value(mem,phase);
1741 if (value != NULL && value->is_Con()) {
1742 assert(value->bottom_type()->higher_equal(_type),"sanity");
1743 return value->bottom_type();
1744 }
1745 }
1747 if (is_instance) {
1748 // If we have an instance type and our memory input is the
1749 // programs's initial memory state, there is no matching store,
1750 // so just return a zero of the appropriate type
1751 Node *mem = in(MemNode::Memory);
1752 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1753 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1754 return Type::get_zero_type(_type->basic_type());
1755 }
1756 }
1757 return _type;
1758 }
1760 //------------------------------match_edge-------------------------------------
1761 // Do we Match on this edge index or not? Match only the address.
1762 uint LoadNode::match_edge(uint idx) const {
1763 return idx == MemNode::Address;
1764 }
1766 //--------------------------LoadBNode::Ideal--------------------------------------
1767 //
1768 // If the previous store is to the same address as this load,
1769 // and the value stored was larger than a byte, replace this load
1770 // with the value stored truncated to a byte. If no truncation is
1771 // needed, the replacement is done in LoadNode::Identity().
1772 //
1773 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1774 Node* mem = in(MemNode::Memory);
1775 Node* value = can_see_stored_value(mem,phase);
1776 if( value && !phase->type(value)->higher_equal( _type ) ) {
1777 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );
1778 return new (phase->C) RShiftINode(result, phase->intcon(24));
1779 }
1780 // Identity call will handle the case where truncation is not needed.
1781 return LoadNode::Ideal(phase, can_reshape);
1782 }
1784 const Type* LoadBNode::Value(PhaseTransform *phase) const {
1785 Node* mem = in(MemNode::Memory);
1786 Node* value = can_see_stored_value(mem,phase);
1787 if (value != NULL && value->is_Con() &&
1788 !value->bottom_type()->higher_equal(_type)) {
1789 // If the input to the store does not fit with the load's result type,
1790 // it must be truncated. We can't delay until Ideal call since
1791 // a singleton Value is needed for split_thru_phi optimization.
1792 int con = value->get_int();
1793 return TypeInt::make((con << 24) >> 24);
1794 }
1795 return LoadNode::Value(phase);
1796 }
1798 //--------------------------LoadUBNode::Ideal-------------------------------------
1799 //
1800 // If the previous store is to the same address as this load,
1801 // and the value stored was larger than a byte, replace this load
1802 // with the value stored truncated to a byte. If no truncation is
1803 // needed, the replacement is done in LoadNode::Identity().
1804 //
1805 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1806 Node* mem = in(MemNode::Memory);
1807 Node* value = can_see_stored_value(mem, phase);
1808 if (value && !phase->type(value)->higher_equal(_type))
1809 return new (phase->C) AndINode(value, phase->intcon(0xFF));
1810 // Identity call will handle the case where truncation is not needed.
1811 return LoadNode::Ideal(phase, can_reshape);
1812 }
1814 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1815 Node* mem = in(MemNode::Memory);
1816 Node* value = can_see_stored_value(mem,phase);
1817 if (value != NULL && value->is_Con() &&
1818 !value->bottom_type()->higher_equal(_type)) {
1819 // If the input to the store does not fit with the load's result type,
1820 // it must be truncated. We can't delay until Ideal call since
1821 // a singleton Value is needed for split_thru_phi optimization.
1822 int con = value->get_int();
1823 return TypeInt::make(con & 0xFF);
1824 }
1825 return LoadNode::Value(phase);
1826 }
1828 //--------------------------LoadUSNode::Ideal-------------------------------------
1829 //
1830 // If the previous store is to the same address as this load,
1831 // and the value stored was larger than a char, replace this load
1832 // with the value stored truncated to a char. If no truncation is
1833 // needed, the replacement is done in LoadNode::Identity().
1834 //
1835 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1836 Node* mem = in(MemNode::Memory);
1837 Node* value = can_see_stored_value(mem,phase);
1838 if( value && !phase->type(value)->higher_equal( _type ) )
1839 return new (phase->C) AndINode(value,phase->intcon(0xFFFF));
1840 // Identity call will handle the case where truncation is not needed.
1841 return LoadNode::Ideal(phase, can_reshape);
1842 }
1844 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1845 Node* mem = in(MemNode::Memory);
1846 Node* value = can_see_stored_value(mem,phase);
1847 if (value != NULL && value->is_Con() &&
1848 !value->bottom_type()->higher_equal(_type)) {
1849 // If the input to the store does not fit with the load's result type,
1850 // it must be truncated. We can't delay until Ideal call since
1851 // a singleton Value is needed for split_thru_phi optimization.
1852 int con = value->get_int();
1853 return TypeInt::make(con & 0xFFFF);
1854 }
1855 return LoadNode::Value(phase);
1856 }
1858 //--------------------------LoadSNode::Ideal--------------------------------------
1859 //
1860 // If the previous store is to the same address as this load,
1861 // and the value stored was larger than a short, replace this load
1862 // with the value stored truncated to a short. If no truncation is
1863 // needed, the replacement is done in LoadNode::Identity().
1864 //
1865 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1866 Node* mem = in(MemNode::Memory);
1867 Node* value = can_see_stored_value(mem,phase);
1868 if( value && !phase->type(value)->higher_equal( _type ) ) {
1869 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );
1870 return new (phase->C) RShiftINode(result, phase->intcon(16));
1871 }
1872 // Identity call will handle the case where truncation is not needed.
1873 return LoadNode::Ideal(phase, can_reshape);
1874 }
1876 const Type* LoadSNode::Value(PhaseTransform *phase) const {
1877 Node* mem = in(MemNode::Memory);
1878 Node* value = can_see_stored_value(mem,phase);
1879 if (value != NULL && value->is_Con() &&
1880 !value->bottom_type()->higher_equal(_type)) {
1881 // If the input to the store does not fit with the load's result type,
1882 // it must be truncated. We can't delay until Ideal call since
1883 // a singleton Value is needed for split_thru_phi optimization.
1884 int con = value->get_int();
1885 return TypeInt::make((con << 16) >> 16);
1886 }
1887 return LoadNode::Value(phase);
1888 }
1890 //=============================================================================
1891 //----------------------------LoadKlassNode::make------------------------------
1892 // Polymorphic factory method:
1893 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1894 Compile* C = gvn.C;
1895 Node *ctl = NULL;
1896 // sanity check the alias category against the created node type
1897 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
1898 assert(adr_type != NULL, "expecting TypeKlassPtr");
1899 #ifdef _LP64
1900 if (adr_type->is_ptr_to_narrowklass()) {
1901 assert(UseCompressedKlassPointers, "no compressed klasses");
1902 Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass()));
1903 return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
1904 }
1905 #endif
1906 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1907 return new (C) LoadKlassNode(ctl, mem, adr, at, tk);
1908 }
1910 //------------------------------Value------------------------------------------
1911 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1912 return klass_value_common(phase);
1913 }
1915 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1916 // Either input is TOP ==> the result is TOP
1917 const Type *t1 = phase->type( in(MemNode::Memory) );
1918 if (t1 == Type::TOP) return Type::TOP;
1919 Node *adr = in(MemNode::Address);
1920 const Type *t2 = phase->type( adr );
1921 if (t2 == Type::TOP) return Type::TOP;
1922 const TypePtr *tp = t2->is_ptr();
1923 if (TypePtr::above_centerline(tp->ptr()) ||
1924 tp->ptr() == TypePtr::Null) return Type::TOP;
1926 // Return a more precise klass, if possible
1927 const TypeInstPtr *tinst = tp->isa_instptr();
1928 if (tinst != NULL) {
1929 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1930 int offset = tinst->offset();
1931 if (ik == phase->C->env()->Class_klass()
1932 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1933 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1934 // We are loading a special hidden field from a Class mirror object,
1935 // the field which points to the VM's Klass metaobject.
1936 ciType* t = tinst->java_mirror_type();
1937 // java_mirror_type returns non-null for compile-time Class constants.
1938 if (t != NULL) {
1939 // constant oop => constant klass
1940 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1941 return TypeKlassPtr::make(ciArrayKlass::make(t));
1942 }
1943 if (!t->is_klass()) {
1944 // a primitive Class (e.g., int.class) has NULL for a klass field
1945 return TypePtr::NULL_PTR;
1946 }
1947 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1948 return TypeKlassPtr::make(t->as_klass());
1949 }
1950 // non-constant mirror, so we can't tell what's going on
1951 }
1952 if( !ik->is_loaded() )
1953 return _type; // Bail out if not loaded
1954 if (offset == oopDesc::klass_offset_in_bytes()) {
1955 if (tinst->klass_is_exact()) {
1956 return TypeKlassPtr::make(ik);
1957 }
1958 // See if we can become precise: no subklasses and no interface
1959 // (Note: We need to support verified interfaces.)
1960 if (!ik->is_interface() && !ik->has_subklass()) {
1961 //assert(!UseExactTypes, "this code should be useless with exact types");
1962 // Add a dependence; if any subclass added we need to recompile
1963 if (!ik->is_final()) {
1964 // %%% should use stronger assert_unique_concrete_subtype instead
1965 phase->C->dependencies()->assert_leaf_type(ik);
1966 }
1967 // Return precise klass
1968 return TypeKlassPtr::make(ik);
1969 }
1971 // Return root of possible klass
1972 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1973 }
1974 }
1976 // Check for loading klass from an array
1977 const TypeAryPtr *tary = tp->isa_aryptr();
1978 if( tary != NULL ) {
1979 ciKlass *tary_klass = tary->klass();
1980 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1981 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1982 if (tary->klass_is_exact()) {
1983 return TypeKlassPtr::make(tary_klass);
1984 }
1985 ciArrayKlass *ak = tary->klass()->as_array_klass();
1986 // If the klass is an object array, we defer the question to the
1987 // array component klass.
1988 if( ak->is_obj_array_klass() ) {
1989 assert( ak->is_loaded(), "" );
1990 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1991 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1992 ciInstanceKlass* ik = base_k->as_instance_klass();
1993 // See if we can become precise: no subklasses and no interface
1994 if (!ik->is_interface() && !ik->has_subklass()) {
1995 //assert(!UseExactTypes, "this code should be useless with exact types");
1996 // Add a dependence; if any subclass added we need to recompile
1997 if (!ik->is_final()) {
1998 phase->C->dependencies()->assert_leaf_type(ik);
1999 }
2000 // Return precise array klass
2001 return TypeKlassPtr::make(ak);
2002 }
2003 }
2004 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2005 } else { // Found a type-array?
2006 //assert(!UseExactTypes, "this code should be useless with exact types");
2007 assert( ak->is_type_array_klass(), "" );
2008 return TypeKlassPtr::make(ak); // These are always precise
2009 }
2010 }
2011 }
2013 // Check for loading klass from an array klass
2014 const TypeKlassPtr *tkls = tp->isa_klassptr();
2015 if (tkls != NULL && !StressReflectiveCode) {
2016 ciKlass* klass = tkls->klass();
2017 if( !klass->is_loaded() )
2018 return _type; // Bail out if not loaded
2019 if( klass->is_obj_array_klass() &&
2020 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2021 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2022 // // Always returning precise element type is incorrect,
2023 // // e.g., element type could be object and array may contain strings
2024 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2026 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2027 // according to the element type's subclassing.
2028 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2029 }
2030 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2031 tkls->offset() == in_bytes(Klass::super_offset())) {
2032 ciKlass* sup = klass->as_instance_klass()->super();
2033 // The field is Klass::_super. Return its (constant) value.
2034 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2035 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2036 }
2037 }
2039 // Bailout case
2040 return LoadNode::Value(phase);
2041 }
2043 //------------------------------Identity---------------------------------------
2044 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2045 // Also feed through the klass in Allocate(...klass...)._klass.
2046 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
2047 return klass_identity_common(phase);
2048 }
2050 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
2051 Node* x = LoadNode::Identity(phase);
2052 if (x != this) return x;
2054 // Take apart the address into an oop and and offset.
2055 // Return 'this' if we cannot.
2056 Node* adr = in(MemNode::Address);
2057 intptr_t offset = 0;
2058 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2059 if (base == NULL) return this;
2060 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2061 if (toop == NULL) return this;
2063 // We can fetch the klass directly through an AllocateNode.
2064 // This works even if the klass is not constant (clone or newArray).
2065 if (offset == oopDesc::klass_offset_in_bytes()) {
2066 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2067 if (allocated_klass != NULL) {
2068 return allocated_klass;
2069 }
2070 }
2072 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2073 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.
2074 // See inline_native_Class_query for occurrences of these patterns.
2075 // Java Example: x.getClass().isAssignableFrom(y)
2076 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
2077 //
2078 // This improves reflective code, often making the Class
2079 // mirror go completely dead. (Current exception: Class
2080 // mirrors may appear in debug info, but we could clean them out by
2081 // introducing a new debug info operator for Klass*.java_mirror).
2082 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2083 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2084 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2085 // We are loading a special hidden field from a Class mirror,
2086 // the field which points to its Klass or ArrayKlass metaobject.
2087 if (base->is_Load()) {
2088 Node* adr2 = base->in(MemNode::Address);
2089 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2090 if (tkls != NULL && !tkls->empty()
2091 && (tkls->klass()->is_instance_klass() ||
2092 tkls->klass()->is_array_klass())
2093 && adr2->is_AddP()
2094 ) {
2095 int mirror_field = in_bytes(Klass::java_mirror_offset());
2096 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2097 mirror_field = in_bytes(ArrayKlass::component_mirror_offset());
2098 }
2099 if (tkls->offset() == mirror_field) {
2100 return adr2->in(AddPNode::Base);
2101 }
2102 }
2103 }
2104 }
2106 return this;
2107 }
2110 //------------------------------Value------------------------------------------
2111 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
2112 const Type *t = klass_value_common(phase);
2113 if (t == Type::TOP)
2114 return t;
2116 return t->make_narrowklass();
2117 }
2119 //------------------------------Identity---------------------------------------
2120 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2121 // Also feed through the klass in Allocate(...klass...)._klass.
2122 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2123 Node *x = klass_identity_common(phase);
2125 const Type *t = phase->type( x );
2126 if( t == Type::TOP ) return x;
2127 if( t->isa_narrowklass()) return x;
2128 assert (!t->isa_narrowoop(), "no narrow oop here");
2130 return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));
2131 }
2133 //------------------------------Value-----------------------------------------
2134 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2135 // Either input is TOP ==> the result is TOP
2136 const Type *t1 = phase->type( in(MemNode::Memory) );
2137 if( t1 == Type::TOP ) return Type::TOP;
2138 Node *adr = in(MemNode::Address);
2139 const Type *t2 = phase->type( adr );
2140 if( t2 == Type::TOP ) return Type::TOP;
2141 const TypePtr *tp = t2->is_ptr();
2142 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2143 const TypeAryPtr *tap = tp->isa_aryptr();
2144 if( !tap ) return _type;
2145 return tap->size();
2146 }
2148 //-------------------------------Ideal---------------------------------------
2149 // Feed through the length in AllocateArray(...length...)._length.
2150 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2151 Node* p = MemNode::Ideal_common(phase, can_reshape);
2152 if (p) return (p == NodeSentinel) ? NULL : p;
2154 // Take apart the address into an oop and and offset.
2155 // Return 'this' if we cannot.
2156 Node* adr = in(MemNode::Address);
2157 intptr_t offset = 0;
2158 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2159 if (base == NULL) return NULL;
2160 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2161 if (tary == NULL) return NULL;
2163 // We can fetch the length directly through an AllocateArrayNode.
2164 // This works even if the length is not constant (clone or newArray).
2165 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2166 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2167 if (alloc != NULL) {
2168 Node* allocated_length = alloc->Ideal_length();
2169 Node* len = alloc->make_ideal_length(tary, phase);
2170 if (allocated_length != len) {
2171 // New CastII improves on this.
2172 return len;
2173 }
2174 }
2175 }
2177 return NULL;
2178 }
2180 //------------------------------Identity---------------------------------------
2181 // Feed through the length in AllocateArray(...length...)._length.
2182 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2183 Node* x = LoadINode::Identity(phase);
2184 if (x != this) return x;
2186 // Take apart the address into an oop and and offset.
2187 // Return 'this' if we cannot.
2188 Node* adr = in(MemNode::Address);
2189 intptr_t offset = 0;
2190 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2191 if (base == NULL) return this;
2192 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2193 if (tary == NULL) return this;
2195 // We can fetch the length directly through an AllocateArrayNode.
2196 // This works even if the length is not constant (clone or newArray).
2197 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2198 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2199 if (alloc != NULL) {
2200 Node* allocated_length = alloc->Ideal_length();
2201 // Do not allow make_ideal_length to allocate a CastII node.
2202 Node* len = alloc->make_ideal_length(tary, phase, false);
2203 if (allocated_length == len) {
2204 // Return allocated_length only if it would not be improved by a CastII.
2205 return allocated_length;
2206 }
2207 }
2208 }
2210 return this;
2212 }
2214 //=============================================================================
2215 //---------------------------StoreNode::make-----------------------------------
2216 // Polymorphic factory method:
2217 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
2218 Compile* C = gvn.C;
2219 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2220 ctl != NULL, "raw memory operations should have control edge");
2222 switch (bt) {
2223 case T_BOOLEAN:
2224 case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val);
2225 case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val);
2226 case T_CHAR:
2227 case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val);
2228 case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val);
2229 case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val);
2230 case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val);
2231 case T_METADATA:
2232 case T_ADDRESS:
2233 case T_OBJECT:
2234 #ifdef _LP64
2235 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2236 val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2237 return new (C) StoreNNode(ctl, mem, adr, adr_type, val);
2238 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2239 (UseCompressedKlassPointers && val->bottom_type()->isa_klassptr() &&
2240 adr->bottom_type()->isa_rawptr())) {
2241 val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2242 return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val);
2243 }
2244 #endif
2245 {
2246 return new (C) StorePNode(ctl, mem, adr, adr_type, val);
2247 }
2248 }
2249 ShouldNotReachHere();
2250 return (StoreNode*)NULL;
2251 }
2253 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
2254 bool require_atomic = true;
2255 return new (C) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
2256 }
2259 //--------------------------bottom_type----------------------------------------
2260 const Type *StoreNode::bottom_type() const {
2261 return Type::MEMORY;
2262 }
2264 //------------------------------hash-------------------------------------------
2265 uint StoreNode::hash() const {
2266 // unroll addition of interesting fields
2267 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2269 // Since they are not commoned, do not hash them:
2270 return NO_HASH;
2271 }
2273 //------------------------------Ideal------------------------------------------
2274 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2275 // When a store immediately follows a relevant allocation/initialization,
2276 // try to capture it into the initialization, or hoist it above.
2277 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2278 Node* p = MemNode::Ideal_common(phase, can_reshape);
2279 if (p) return (p == NodeSentinel) ? NULL : p;
2281 Node* mem = in(MemNode::Memory);
2282 Node* address = in(MemNode::Address);
2284 // Back-to-back stores to same address? Fold em up. Generally
2285 // unsafe if I have intervening uses... Also disallowed for StoreCM
2286 // since they must follow each StoreP operation. Redundant StoreCMs
2287 // are eliminated just before matching in final_graph_reshape.
2288 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
2289 mem->Opcode() != Op_StoreCM) {
2290 // Looking at a dead closed cycle of memory?
2291 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2293 assert(Opcode() == mem->Opcode() ||
2294 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2295 "no mismatched stores, except on raw memory");
2297 if (mem->outcnt() == 1 && // check for intervening uses
2298 mem->as_Store()->memory_size() <= this->memory_size()) {
2299 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2300 // For example, 'mem' might be the final state at a conditional return.
2301 // Or, 'mem' might be used by some node which is live at the same time
2302 // 'this' is live, which might be unschedulable. So, require exactly
2303 // ONE user, the 'this' store, until such time as we clone 'mem' for
2304 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2305 if (can_reshape) { // (%%% is this an anachronism?)
2306 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2307 phase->is_IterGVN());
2308 } else {
2309 // It's OK to do this in the parser, since DU info is always accurate,
2310 // and the parser always refers to nodes via SafePointNode maps.
2311 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2312 }
2313 return this;
2314 }
2315 }
2317 // Capture an unaliased, unconditional, simple store into an initializer.
2318 // Or, if it is independent of the allocation, hoist it above the allocation.
2319 if (ReduceFieldZeroing && /*can_reshape &&*/
2320 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2321 InitializeNode* init = mem->in(0)->as_Initialize();
2322 intptr_t offset = init->can_capture_store(this, phase);
2323 if (offset > 0) {
2324 Node* moved = init->capture_store(this, offset, phase);
2325 // If the InitializeNode captured me, it made a raw copy of me,
2326 // and I need to disappear.
2327 if (moved != NULL) {
2328 // %%% hack to ensure that Ideal returns a new node:
2329 mem = MergeMemNode::make(phase->C, mem);
2330 return mem; // fold me away
2331 }
2332 }
2333 }
2335 return NULL; // No further progress
2336 }
2338 //------------------------------Value-----------------------------------------
2339 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2340 // Either input is TOP ==> the result is TOP
2341 const Type *t1 = phase->type( in(MemNode::Memory) );
2342 if( t1 == Type::TOP ) return Type::TOP;
2343 const Type *t2 = phase->type( in(MemNode::Address) );
2344 if( t2 == Type::TOP ) return Type::TOP;
2345 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2346 if( t3 == Type::TOP ) return Type::TOP;
2347 return Type::MEMORY;
2348 }
2350 //------------------------------Identity---------------------------------------
2351 // Remove redundant stores:
2352 // Store(m, p, Load(m, p)) changes to m.
2353 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2354 Node *StoreNode::Identity( PhaseTransform *phase ) {
2355 Node* mem = in(MemNode::Memory);
2356 Node* adr = in(MemNode::Address);
2357 Node* val = in(MemNode::ValueIn);
2359 // Load then Store? Then the Store is useless
2360 if (val->is_Load() &&
2361 val->in(MemNode::Address)->eqv_uncast(adr) &&
2362 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2363 val->as_Load()->store_Opcode() == Opcode()) {
2364 return mem;
2365 }
2367 // Two stores in a row of the same value?
2368 if (mem->is_Store() &&
2369 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2370 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2371 mem->Opcode() == Opcode()) {
2372 return mem;
2373 }
2375 // Store of zero anywhere into a freshly-allocated object?
2376 // Then the store is useless.
2377 // (It must already have been captured by the InitializeNode.)
2378 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2379 // a newly allocated object is already all-zeroes everywhere
2380 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2381 return mem;
2382 }
2384 // the store may also apply to zero-bits in an earlier object
2385 Node* prev_mem = find_previous_store(phase);
2386 // Steps (a), (b): Walk past independent stores to find an exact match.
2387 if (prev_mem != NULL) {
2388 Node* prev_val = can_see_stored_value(prev_mem, phase);
2389 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2390 // prev_val and val might differ by a cast; it would be good
2391 // to keep the more informative of the two.
2392 return mem;
2393 }
2394 }
2395 }
2397 return this;
2398 }
2400 //------------------------------match_edge-------------------------------------
2401 // Do we Match on this edge index or not? Match only memory & value
2402 uint StoreNode::match_edge(uint idx) const {
2403 return idx == MemNode::Address || idx == MemNode::ValueIn;
2404 }
2406 //------------------------------cmp--------------------------------------------
2407 // Do not common stores up together. They generally have to be split
2408 // back up anyways, so do not bother.
2409 uint StoreNode::cmp( const Node &n ) const {
2410 return (&n == this); // Always fail except on self
2411 }
2413 //------------------------------Ideal_masked_input-----------------------------
2414 // Check for a useless mask before a partial-word store
2415 // (StoreB ... (AndI valIn conIa) )
2416 // If (conIa & mask == mask) this simplifies to
2417 // (StoreB ... (valIn) )
2418 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2419 Node *val = in(MemNode::ValueIn);
2420 if( val->Opcode() == Op_AndI ) {
2421 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2422 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2423 set_req(MemNode::ValueIn, val->in(1));
2424 return this;
2425 }
2426 }
2427 return NULL;
2428 }
2431 //------------------------------Ideal_sign_extended_input----------------------
2432 // Check for useless sign-extension before a partial-word store
2433 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2434 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2435 // (StoreB ... (valIn) )
2436 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2437 Node *val = in(MemNode::ValueIn);
2438 if( val->Opcode() == Op_RShiftI ) {
2439 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2440 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2441 Node *shl = val->in(1);
2442 if( shl->Opcode() == Op_LShiftI ) {
2443 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2444 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2445 set_req(MemNode::ValueIn, shl->in(1));
2446 return this;
2447 }
2448 }
2449 }
2450 }
2451 return NULL;
2452 }
2454 //------------------------------value_never_loaded-----------------------------------
2455 // Determine whether there are any possible loads of the value stored.
2456 // For simplicity, we actually check if there are any loads from the
2457 // address stored to, not just for loads of the value stored by this node.
2458 //
2459 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2460 Node *adr = in(Address);
2461 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2462 if (adr_oop == NULL)
2463 return false;
2464 if (!adr_oop->is_known_instance_field())
2465 return false; // if not a distinct instance, there may be aliases of the address
2466 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2467 Node *use = adr->fast_out(i);
2468 int opc = use->Opcode();
2469 if (use->is_Load() || use->is_LoadStore()) {
2470 return false;
2471 }
2472 }
2473 return true;
2474 }
2476 //=============================================================================
2477 //------------------------------Ideal------------------------------------------
2478 // If the store is from an AND mask that leaves the low bits untouched, then
2479 // we can skip the AND operation. If the store is from a sign-extension
2480 // (a left shift, then right shift) we can skip both.
2481 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2482 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2483 if( progress != NULL ) return progress;
2485 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2486 if( progress != NULL ) return progress;
2488 // Finally check the default case
2489 return StoreNode::Ideal(phase, can_reshape);
2490 }
2492 //=============================================================================
2493 //------------------------------Ideal------------------------------------------
2494 // If the store is from an AND mask that leaves the low bits untouched, then
2495 // we can skip the AND operation
2496 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2497 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2498 if( progress != NULL ) return progress;
2500 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2501 if( progress != NULL ) return progress;
2503 // Finally check the default case
2504 return StoreNode::Ideal(phase, can_reshape);
2505 }
2507 //=============================================================================
2508 //------------------------------Identity---------------------------------------
2509 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2510 // No need to card mark when storing a null ptr
2511 Node* my_store = in(MemNode::OopStore);
2512 if (my_store->is_Store()) {
2513 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2514 if( t1 == TypePtr::NULL_PTR ) {
2515 return in(MemNode::Memory);
2516 }
2517 }
2518 return this;
2519 }
2521 //=============================================================================
2522 //------------------------------Ideal---------------------------------------
2523 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2524 Node* progress = StoreNode::Ideal(phase, can_reshape);
2525 if (progress != NULL) return progress;
2527 Node* my_store = in(MemNode::OopStore);
2528 if (my_store->is_MergeMem()) {
2529 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2530 set_req(MemNode::OopStore, mem);
2531 return this;
2532 }
2534 return NULL;
2535 }
2537 //------------------------------Value-----------------------------------------
2538 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2539 // Either input is TOP ==> the result is TOP
2540 const Type *t = phase->type( in(MemNode::Memory) );
2541 if( t == Type::TOP ) return Type::TOP;
2542 t = phase->type( in(MemNode::Address) );
2543 if( t == Type::TOP ) return Type::TOP;
2544 t = phase->type( in(MemNode::ValueIn) );
2545 if( t == Type::TOP ) return Type::TOP;
2546 // If extra input is TOP ==> the result is TOP
2547 t = phase->type( in(MemNode::OopStore) );
2548 if( t == Type::TOP ) return Type::TOP;
2550 return StoreNode::Value( phase );
2551 }
2554 //=============================================================================
2555 //----------------------------------SCMemProjNode------------------------------
2556 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2557 {
2558 return bottom_type();
2559 }
2561 //=============================================================================
2562 //----------------------------------LoadStoreNode------------------------------
2563 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2564 : Node(required),
2565 _type(rt),
2566 _adr_type(at)
2567 {
2568 init_req(MemNode::Control, c );
2569 init_req(MemNode::Memory , mem);
2570 init_req(MemNode::Address, adr);
2571 init_req(MemNode::ValueIn, val);
2572 init_class_id(Class_LoadStore);
2573 }
2575 uint LoadStoreNode::ideal_reg() const {
2576 return _type->ideal_reg();
2577 }
2579 bool LoadStoreNode::result_not_used() const {
2580 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2581 Node *x = fast_out(i);
2582 if (x->Opcode() == Op_SCMemProj) continue;
2583 return false;
2584 }
2585 return true;
2586 }
2588 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2590 //=============================================================================
2591 //----------------------------------LoadStoreConditionalNode--------------------
2592 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2593 init_req(ExpectedIn, ex );
2594 }
2596 //=============================================================================
2597 //-------------------------------adr_type--------------------------------------
2598 // Do we Match on this edge index or not? Do not match memory
2599 const TypePtr* ClearArrayNode::adr_type() const {
2600 Node *adr = in(3);
2601 return MemNode::calculate_adr_type(adr->bottom_type());
2602 }
2604 //------------------------------match_edge-------------------------------------
2605 // Do we Match on this edge index or not? Do not match memory
2606 uint ClearArrayNode::match_edge(uint idx) const {
2607 return idx > 1;
2608 }
2610 //------------------------------Identity---------------------------------------
2611 // Clearing a zero length array does nothing
2612 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2613 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2614 }
2616 //------------------------------Idealize---------------------------------------
2617 // Clearing a short array is faster with stores
2618 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2619 const int unit = BytesPerLong;
2620 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2621 if (!t) return NULL;
2622 if (!t->is_con()) return NULL;
2623 intptr_t raw_count = t->get_con();
2624 intptr_t size = raw_count;
2625 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2626 // Clearing nothing uses the Identity call.
2627 // Negative clears are possible on dead ClearArrays
2628 // (see jck test stmt114.stmt11402.val).
2629 if (size <= 0 || size % unit != 0) return NULL;
2630 intptr_t count = size / unit;
2631 // Length too long; use fast hardware clear
2632 if (size > Matcher::init_array_short_size) return NULL;
2633 Node *mem = in(1);
2634 if( phase->type(mem)==Type::TOP ) return NULL;
2635 Node *adr = in(3);
2636 const Type* at = phase->type(adr);
2637 if( at==Type::TOP ) return NULL;
2638 const TypePtr* atp = at->isa_ptr();
2639 // adjust atp to be the correct array element address type
2640 if (atp == NULL) atp = TypePtr::BOTTOM;
2641 else atp = atp->add_offset(Type::OffsetBot);
2642 // Get base for derived pointer purposes
2643 if( adr->Opcode() != Op_AddP ) Unimplemented();
2644 Node *base = adr->in(1);
2646 Node *zero = phase->makecon(TypeLong::ZERO);
2647 Node *off = phase->MakeConX(BytesPerLong);
2648 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero);
2649 count--;
2650 while( count-- ) {
2651 mem = phase->transform(mem);
2652 adr = phase->transform(new (phase->C) AddPNode(base,adr,off));
2653 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero);
2654 }
2655 return mem;
2656 }
2658 //----------------------------step_through----------------------------------
2659 // Return allocation input memory edge if it is different instance
2660 // or itself if it is the one we are looking for.
2661 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2662 Node* n = *np;
2663 assert(n->is_ClearArray(), "sanity");
2664 intptr_t offset;
2665 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2666 // This method is called only before Allocate nodes are expanded during
2667 // macro nodes expansion. Before that ClearArray nodes are only generated
2668 // in LibraryCallKit::generate_arraycopy() which follows allocations.
2669 assert(alloc != NULL, "should have allocation");
2670 if (alloc->_idx == instance_id) {
2671 // Can not bypass initialization of the instance we are looking for.
2672 return false;
2673 }
2674 // Otherwise skip it.
2675 InitializeNode* init = alloc->initialization();
2676 if (init != NULL)
2677 *np = init->in(TypeFunc::Memory);
2678 else
2679 *np = alloc->in(TypeFunc::Memory);
2680 return true;
2681 }
2683 //----------------------------clear_memory-------------------------------------
2684 // Generate code to initialize object storage to zero.
2685 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2686 intptr_t start_offset,
2687 Node* end_offset,
2688 PhaseGVN* phase) {
2689 Compile* C = phase->C;
2690 intptr_t offset = start_offset;
2692 int unit = BytesPerLong;
2693 if ((offset % unit) != 0) {
2694 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset));
2695 adr = phase->transform(adr);
2696 const TypePtr* atp = TypeRawPtr::BOTTOM;
2697 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2698 mem = phase->transform(mem);
2699 offset += BytesPerInt;
2700 }
2701 assert((offset % unit) == 0, "");
2703 // Initialize the remaining stuff, if any, with a ClearArray.
2704 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2705 }
2707 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2708 Node* start_offset,
2709 Node* end_offset,
2710 PhaseGVN* phase) {
2711 if (start_offset == end_offset) {
2712 // nothing to do
2713 return mem;
2714 }
2716 Compile* C = phase->C;
2717 int unit = BytesPerLong;
2718 Node* zbase = start_offset;
2719 Node* zend = end_offset;
2721 // Scale to the unit required by the CPU:
2722 if (!Matcher::init_array_count_is_in_bytes) {
2723 Node* shift = phase->intcon(exact_log2(unit));
2724 zbase = phase->transform( new(C) URShiftXNode(zbase, shift) );
2725 zend = phase->transform( new(C) URShiftXNode(zend, shift) );
2726 }
2728 Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) );
2729 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2731 // Bulk clear double-words
2732 Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) );
2733 mem = new (C) ClearArrayNode(ctl, mem, zsize, adr);
2734 return phase->transform(mem);
2735 }
2737 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2738 intptr_t start_offset,
2739 intptr_t end_offset,
2740 PhaseGVN* phase) {
2741 if (start_offset == end_offset) {
2742 // nothing to do
2743 return mem;
2744 }
2746 Compile* C = phase->C;
2747 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2748 intptr_t done_offset = end_offset;
2749 if ((done_offset % BytesPerLong) != 0) {
2750 done_offset -= BytesPerInt;
2751 }
2752 if (done_offset > start_offset) {
2753 mem = clear_memory(ctl, mem, dest,
2754 start_offset, phase->MakeConX(done_offset), phase);
2755 }
2756 if (done_offset < end_offset) { // emit the final 32-bit store
2757 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset));
2758 adr = phase->transform(adr);
2759 const TypePtr* atp = TypeRawPtr::BOTTOM;
2760 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2761 mem = phase->transform(mem);
2762 done_offset += BytesPerInt;
2763 }
2764 assert(done_offset == end_offset, "");
2765 return mem;
2766 }
2768 //=============================================================================
2769 // Do not match memory edge.
2770 uint StrIntrinsicNode::match_edge(uint idx) const {
2771 return idx == 2 || idx == 3;
2772 }
2774 //------------------------------Ideal------------------------------------------
2775 // Return a node which is more "ideal" than the current node. Strip out
2776 // control copies
2777 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2778 if (remove_dead_region(phase, can_reshape)) return this;
2779 // Don't bother trying to transform a dead node
2780 if (in(0) && in(0)->is_top()) return NULL;
2782 if (can_reshape) {
2783 Node* mem = phase->transform(in(MemNode::Memory));
2784 // If transformed to a MergeMem, get the desired slice
2785 uint alias_idx = phase->C->get_alias_index(adr_type());
2786 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem;
2787 if (mem != in(MemNode::Memory)) {
2788 set_req(MemNode::Memory, mem);
2789 return this;
2790 }
2791 }
2792 return NULL;
2793 }
2795 //------------------------------Value------------------------------------------
2796 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const {
2797 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
2798 return bottom_type();
2799 }
2801 //=============================================================================
2802 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2803 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2804 _adr_type(C->get_adr_type(alias_idx))
2805 {
2806 init_class_id(Class_MemBar);
2807 Node* top = C->top();
2808 init_req(TypeFunc::I_O,top);
2809 init_req(TypeFunc::FramePtr,top);
2810 init_req(TypeFunc::ReturnAdr,top);
2811 if (precedent != NULL)
2812 init_req(TypeFunc::Parms, precedent);
2813 }
2815 //------------------------------cmp--------------------------------------------
2816 uint MemBarNode::hash() const { return NO_HASH; }
2817 uint MemBarNode::cmp( const Node &n ) const {
2818 return (&n == this); // Always fail except on self
2819 }
2821 //------------------------------make-------------------------------------------
2822 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2823 switch (opcode) {
2824 case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn);
2825 case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn);
2826 case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn);
2827 case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn);
2828 case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn);
2829 case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn);
2830 case Op_Initialize: return new(C) InitializeNode(C, atp, pn);
2831 case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn);
2832 default: ShouldNotReachHere(); return NULL;
2833 }
2834 }
2836 //------------------------------Ideal------------------------------------------
2837 // Return a node which is more "ideal" than the current node. Strip out
2838 // control copies
2839 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2840 if (remove_dead_region(phase, can_reshape)) return this;
2841 // Don't bother trying to transform a dead node
2842 if (in(0) && in(0)->is_top()) return NULL;
2844 // Eliminate volatile MemBars for scalar replaced objects.
2845 if (can_reshape && req() == (Precedent+1) &&
2846 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) {
2847 // Volatile field loads and stores.
2848 Node* my_mem = in(MemBarNode::Precedent);
2849 if (my_mem != NULL && my_mem->is_Mem()) {
2850 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2851 // Check for scalar replaced object reference.
2852 if( t_oop != NULL && t_oop->is_known_instance_field() &&
2853 t_oop->offset() != Type::OffsetBot &&
2854 t_oop->offset() != Type::OffsetTop) {
2855 // Replace MemBar projections by its inputs.
2856 PhaseIterGVN* igvn = phase->is_IterGVN();
2857 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2858 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2859 // Must return either the original node (now dead) or a new node
2860 // (Do not return a top here, since that would break the uniqueness of top.)
2861 return new (phase->C) ConINode(TypeInt::ZERO);
2862 }
2863 }
2864 }
2865 return NULL;
2866 }
2868 //------------------------------Value------------------------------------------
2869 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2870 if( !in(0) ) return Type::TOP;
2871 if( phase->type(in(0)) == Type::TOP )
2872 return Type::TOP;
2873 return TypeTuple::MEMBAR;
2874 }
2876 //------------------------------match------------------------------------------
2877 // Construct projections for memory.
2878 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2879 switch (proj->_con) {
2880 case TypeFunc::Control:
2881 case TypeFunc::Memory:
2882 return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2883 }
2884 ShouldNotReachHere();
2885 return NULL;
2886 }
2888 //===========================InitializeNode====================================
2889 // SUMMARY:
2890 // This node acts as a memory barrier on raw memory, after some raw stores.
2891 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2892 // The Initialize can 'capture' suitably constrained stores as raw inits.
2893 // It can coalesce related raw stores into larger units (called 'tiles').
2894 // It can avoid zeroing new storage for memory units which have raw inits.
2895 // At macro-expansion, it is marked 'complete', and does not optimize further.
2896 //
2897 // EXAMPLE:
2898 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2899 // ctl = incoming control; mem* = incoming memory
2900 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2901 // First allocate uninitialized memory and fill in the header:
2902 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2903 // ctl := alloc.Control; mem* := alloc.Memory*
2904 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2905 // Then initialize to zero the non-header parts of the raw memory block:
2906 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2907 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2908 // After the initialize node executes, the object is ready for service:
2909 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2910 // Suppose its body is immediately initialized as {1,2}:
2911 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2912 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2913 // mem.SLICE(#short[*]) := store2
2914 //
2915 // DETAILS:
2916 // An InitializeNode collects and isolates object initialization after
2917 // an AllocateNode and before the next possible safepoint. As a
2918 // memory barrier (MemBarNode), it keeps critical stores from drifting
2919 // down past any safepoint or any publication of the allocation.
2920 // Before this barrier, a newly-allocated object may have uninitialized bits.
2921 // After this barrier, it may be treated as a real oop, and GC is allowed.
2922 //
2923 // The semantics of the InitializeNode include an implicit zeroing of
2924 // the new object from object header to the end of the object.
2925 // (The object header and end are determined by the AllocateNode.)
2926 //
2927 // Certain stores may be added as direct inputs to the InitializeNode.
2928 // These stores must update raw memory, and they must be to addresses
2929 // derived from the raw address produced by AllocateNode, and with
2930 // a constant offset. They must be ordered by increasing offset.
2931 // The first one is at in(RawStores), the last at in(req()-1).
2932 // Unlike most memory operations, they are not linked in a chain,
2933 // but are displayed in parallel as users of the rawmem output of
2934 // the allocation.
2935 //
2936 // (See comments in InitializeNode::capture_store, which continue
2937 // the example given above.)
2938 //
2939 // When the associated Allocate is macro-expanded, the InitializeNode
2940 // may be rewritten to optimize collected stores. A ClearArrayNode
2941 // may also be created at that point to represent any required zeroing.
2942 // The InitializeNode is then marked 'complete', prohibiting further
2943 // capturing of nearby memory operations.
2944 //
2945 // During macro-expansion, all captured initializations which store
2946 // constant values of 32 bits or smaller are coalesced (if advantageous)
2947 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2948 // initialized in fewer memory operations. Memory words which are
2949 // covered by neither tiles nor non-constant stores are pre-zeroed
2950 // by explicit stores of zero. (The code shape happens to do all
2951 // zeroing first, then all other stores, with both sequences occurring
2952 // in order of ascending offsets.)
2953 //
2954 // Alternatively, code may be inserted between an AllocateNode and its
2955 // InitializeNode, to perform arbitrary initialization of the new object.
2956 // E.g., the object copying intrinsics insert complex data transfers here.
2957 // The initialization must then be marked as 'complete' disable the
2958 // built-in zeroing semantics and the collection of initializing stores.
2959 //
2960 // While an InitializeNode is incomplete, reads from the memory state
2961 // produced by it are optimizable if they match the control edge and
2962 // new oop address associated with the allocation/initialization.
2963 // They return a stored value (if the offset matches) or else zero.
2964 // A write to the memory state, if it matches control and address,
2965 // and if it is to a constant offset, may be 'captured' by the
2966 // InitializeNode. It is cloned as a raw memory operation and rewired
2967 // inside the initialization, to the raw oop produced by the allocation.
2968 // Operations on addresses which are provably distinct (e.g., to
2969 // other AllocateNodes) are allowed to bypass the initialization.
2970 //
2971 // The effect of all this is to consolidate object initialization
2972 // (both arrays and non-arrays, both piecewise and bulk) into a
2973 // single location, where it can be optimized as a unit.
2974 //
2975 // Only stores with an offset less than TrackedInitializationLimit words
2976 // will be considered for capture by an InitializeNode. This puts a
2977 // reasonable limit on the complexity of optimized initializations.
2979 //---------------------------InitializeNode------------------------------------
2980 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2981 : _is_complete(Incomplete), _does_not_escape(false),
2982 MemBarNode(C, adr_type, rawoop)
2983 {
2984 init_class_id(Class_Initialize);
2986 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2987 assert(in(RawAddress) == rawoop, "proper init");
2988 // Note: allocation() can be NULL, for secondary initialization barriers
2989 }
2991 // Since this node is not matched, it will be processed by the
2992 // register allocator. Declare that there are no constraints
2993 // on the allocation of the RawAddress edge.
2994 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2995 // This edge should be set to top, by the set_complete. But be conservative.
2996 if (idx == InitializeNode::RawAddress)
2997 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2998 return RegMask::Empty;
2999 }
3001 Node* InitializeNode::memory(uint alias_idx) {
3002 Node* mem = in(Memory);
3003 if (mem->is_MergeMem()) {
3004 return mem->as_MergeMem()->memory_at(alias_idx);
3005 } else {
3006 // incoming raw memory is not split
3007 return mem;
3008 }
3009 }
3011 bool InitializeNode::is_non_zero() {
3012 if (is_complete()) return false;
3013 remove_extra_zeroes();
3014 return (req() > RawStores);
3015 }
3017 void InitializeNode::set_complete(PhaseGVN* phase) {
3018 assert(!is_complete(), "caller responsibility");
3019 _is_complete = Complete;
3021 // After this node is complete, it contains a bunch of
3022 // raw-memory initializations. There is no need for
3023 // it to have anything to do with non-raw memory effects.
3024 // Therefore, tell all non-raw users to re-optimize themselves,
3025 // after skipping the memory effects of this initialization.
3026 PhaseIterGVN* igvn = phase->is_IterGVN();
3027 if (igvn) igvn->add_users_to_worklist(this);
3028 }
3030 // convenience function
3031 // return false if the init contains any stores already
3032 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3033 InitializeNode* init = initialization();
3034 if (init == NULL || init->is_complete()) return false;
3035 init->remove_extra_zeroes();
3036 // for now, if this allocation has already collected any inits, bail:
3037 if (init->is_non_zero()) return false;
3038 init->set_complete(phase);
3039 return true;
3040 }
3042 void InitializeNode::remove_extra_zeroes() {
3043 if (req() == RawStores) return;
3044 Node* zmem = zero_memory();
3045 uint fill = RawStores;
3046 for (uint i = fill; i < req(); i++) {
3047 Node* n = in(i);
3048 if (n->is_top() || n == zmem) continue; // skip
3049 if (fill < i) set_req(fill, n); // compact
3050 ++fill;
3051 }
3052 // delete any empty spaces created:
3053 while (fill < req()) {
3054 del_req(fill);
3055 }
3056 }
3058 // Helper for remembering which stores go with which offsets.
3059 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3060 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3061 intptr_t offset = -1;
3062 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3063 phase, offset);
3064 if (base == NULL) return -1; // something is dead,
3065 if (offset < 0) return -1; // dead, dead
3066 return offset;
3067 }
3069 // Helper for proving that an initialization expression is
3070 // "simple enough" to be folded into an object initialization.
3071 // Attempts to prove that a store's initial value 'n' can be captured
3072 // within the initialization without creating a vicious cycle, such as:
3073 // { Foo p = new Foo(); p.next = p; }
3074 // True for constants and parameters and small combinations thereof.
3075 bool InitializeNode::detect_init_independence(Node* n,
3076 bool st_is_pinned,
3077 int& count) {
3078 if (n == NULL) return true; // (can this really happen?)
3079 if (n->is_Proj()) n = n->in(0);
3080 if (n == this) return false; // found a cycle
3081 if (n->is_Con()) return true;
3082 if (n->is_Start()) return true; // params, etc., are OK
3083 if (n->is_Root()) return true; // even better
3085 Node* ctl = n->in(0);
3086 if (ctl != NULL && !ctl->is_top()) {
3087 if (ctl->is_Proj()) ctl = ctl->in(0);
3088 if (ctl == this) return false;
3090 // If we already know that the enclosing memory op is pinned right after
3091 // the init, then any control flow that the store has picked up
3092 // must have preceded the init, or else be equal to the init.
3093 // Even after loop optimizations (which might change control edges)
3094 // a store is never pinned *before* the availability of its inputs.
3095 if (!MemNode::all_controls_dominate(n, this))
3096 return false; // failed to prove a good control
3098 }
3100 // Check data edges for possible dependencies on 'this'.
3101 if ((count += 1) > 20) return false; // complexity limit
3102 for (uint i = 1; i < n->req(); i++) {
3103 Node* m = n->in(i);
3104 if (m == NULL || m == n || m->is_top()) continue;
3105 uint first_i = n->find_edge(m);
3106 if (i != first_i) continue; // process duplicate edge just once
3107 if (!detect_init_independence(m, st_is_pinned, count)) {
3108 return false;
3109 }
3110 }
3112 return true;
3113 }
3115 // Here are all the checks a Store must pass before it can be moved into
3116 // an initialization. Returns zero if a check fails.
3117 // On success, returns the (constant) offset to which the store applies,
3118 // within the initialized memory.
3119 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
3120 const int FAIL = 0;
3121 if (st->req() != MemNode::ValueIn + 1)
3122 return FAIL; // an inscrutable StoreNode (card mark?)
3123 Node* ctl = st->in(MemNode::Control);
3124 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3125 return FAIL; // must be unconditional after the initialization
3126 Node* mem = st->in(MemNode::Memory);
3127 if (!(mem->is_Proj() && mem->in(0) == this))
3128 return FAIL; // must not be preceded by other stores
3129 Node* adr = st->in(MemNode::Address);
3130 intptr_t offset;
3131 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3132 if (alloc == NULL)
3133 return FAIL; // inscrutable address
3134 if (alloc != allocation())
3135 return FAIL; // wrong allocation! (store needs to float up)
3136 Node* val = st->in(MemNode::ValueIn);
3137 int complexity_count = 0;
3138 if (!detect_init_independence(val, true, complexity_count))
3139 return FAIL; // stored value must be 'simple enough'
3141 return offset; // success
3142 }
3144 // Find the captured store in(i) which corresponds to the range
3145 // [start..start+size) in the initialized object.
3146 // If there is one, return its index i. If there isn't, return the
3147 // negative of the index where it should be inserted.
3148 // Return 0 if the queried range overlaps an initialization boundary
3149 // or if dead code is encountered.
3150 // If size_in_bytes is zero, do not bother with overlap checks.
3151 int InitializeNode::captured_store_insertion_point(intptr_t start,
3152 int size_in_bytes,
3153 PhaseTransform* phase) {
3154 const int FAIL = 0, MAX_STORE = BytesPerLong;
3156 if (is_complete())
3157 return FAIL; // arraycopy got here first; punt
3159 assert(allocation() != NULL, "must be present");
3161 // no negatives, no header fields:
3162 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3164 // after a certain size, we bail out on tracking all the stores:
3165 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3166 if (start >= ti_limit) return FAIL;
3168 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3169 if (i >= limit) return -(int)i; // not found; here is where to put it
3171 Node* st = in(i);
3172 intptr_t st_off = get_store_offset(st, phase);
3173 if (st_off < 0) {
3174 if (st != zero_memory()) {
3175 return FAIL; // bail out if there is dead garbage
3176 }
3177 } else if (st_off > start) {
3178 // ...we are done, since stores are ordered
3179 if (st_off < start + size_in_bytes) {
3180 return FAIL; // the next store overlaps
3181 }
3182 return -(int)i; // not found; here is where to put it
3183 } else if (st_off < start) {
3184 if (size_in_bytes != 0 &&
3185 start < st_off + MAX_STORE &&
3186 start < st_off + st->as_Store()->memory_size()) {
3187 return FAIL; // the previous store overlaps
3188 }
3189 } else {
3190 if (size_in_bytes != 0 &&
3191 st->as_Store()->memory_size() != size_in_bytes) {
3192 return FAIL; // mismatched store size
3193 }
3194 return i;
3195 }
3197 ++i;
3198 }
3199 }
3201 // Look for a captured store which initializes at the offset 'start'
3202 // with the given size. If there is no such store, and no other
3203 // initialization interferes, then return zero_memory (the memory
3204 // projection of the AllocateNode).
3205 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3206 PhaseTransform* phase) {
3207 assert(stores_are_sane(phase), "");
3208 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3209 if (i == 0) {
3210 return NULL; // something is dead
3211 } else if (i < 0) {
3212 return zero_memory(); // just primordial zero bits here
3213 } else {
3214 Node* st = in(i); // here is the store at this position
3215 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3216 return st;
3217 }
3218 }
3220 // Create, as a raw pointer, an address within my new object at 'offset'.
3221 Node* InitializeNode::make_raw_address(intptr_t offset,
3222 PhaseTransform* phase) {
3223 Node* addr = in(RawAddress);
3224 if (offset != 0) {
3225 Compile* C = phase->C;
3226 addr = phase->transform( new (C) AddPNode(C->top(), addr,
3227 phase->MakeConX(offset)) );
3228 }
3229 return addr;
3230 }
3232 // Clone the given store, converting it into a raw store
3233 // initializing a field or element of my new object.
3234 // Caller is responsible for retiring the original store,
3235 // with subsume_node or the like.
3236 //
3237 // From the example above InitializeNode::InitializeNode,
3238 // here are the old stores to be captured:
3239 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3240 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3241 //
3242 // Here is the changed code; note the extra edges on init:
3243 // alloc = (Allocate ...)
3244 // rawoop = alloc.RawAddress
3245 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3246 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3247 // init = (Initialize alloc.Control alloc.Memory rawoop
3248 // rawstore1 rawstore2)
3249 //
3250 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3251 PhaseTransform* phase) {
3252 assert(stores_are_sane(phase), "");
3254 if (start < 0) return NULL;
3255 assert(can_capture_store(st, phase) == start, "sanity");
3257 Compile* C = phase->C;
3258 int size_in_bytes = st->memory_size();
3259 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3260 if (i == 0) return NULL; // bail out
3261 Node* prev_mem = NULL; // raw memory for the captured store
3262 if (i > 0) {
3263 prev_mem = in(i); // there is a pre-existing store under this one
3264 set_req(i, C->top()); // temporarily disconnect it
3265 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3266 } else {
3267 i = -i; // no pre-existing store
3268 prev_mem = zero_memory(); // a slice of the newly allocated object
3269 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3270 set_req(--i, C->top()); // reuse this edge; it has been folded away
3271 else
3272 ins_req(i, C->top()); // build a new edge
3273 }
3274 Node* new_st = st->clone();
3275 new_st->set_req(MemNode::Control, in(Control));
3276 new_st->set_req(MemNode::Memory, prev_mem);
3277 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3278 new_st = phase->transform(new_st);
3280 // At this point, new_st might have swallowed a pre-existing store
3281 // at the same offset, or perhaps new_st might have disappeared,
3282 // if it redundantly stored the same value (or zero to fresh memory).
3284 // In any case, wire it in:
3285 set_req(i, new_st);
3287 // The caller may now kill the old guy.
3288 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3289 assert(check_st == new_st || check_st == NULL, "must be findable");
3290 assert(!is_complete(), "");
3291 return new_st;
3292 }
3294 static bool store_constant(jlong* tiles, int num_tiles,
3295 intptr_t st_off, int st_size,
3296 jlong con) {
3297 if ((st_off & (st_size-1)) != 0)
3298 return false; // strange store offset (assume size==2**N)
3299 address addr = (address)tiles + st_off;
3300 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3301 switch (st_size) {
3302 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3303 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3304 case sizeof(jint): *(jint*) addr = (jint) con; break;
3305 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3306 default: return false; // strange store size (detect size!=2**N here)
3307 }
3308 return true; // return success to caller
3309 }
3311 // Coalesce subword constants into int constants and possibly
3312 // into long constants. The goal, if the CPU permits,
3313 // is to initialize the object with a small number of 64-bit tiles.
3314 // Also, convert floating-point constants to bit patterns.
3315 // Non-constants are not relevant to this pass.
3316 //
3317 // In terms of the running example on InitializeNode::InitializeNode
3318 // and InitializeNode::capture_store, here is the transformation
3319 // of rawstore1 and rawstore2 into rawstore12:
3320 // alloc = (Allocate ...)
3321 // rawoop = alloc.RawAddress
3322 // tile12 = 0x00010002
3323 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3324 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3325 //
3326 void
3327 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3328 Node* size_in_bytes,
3329 PhaseGVN* phase) {
3330 Compile* C = phase->C;
3332 assert(stores_are_sane(phase), "");
3333 // Note: After this pass, they are not completely sane,
3334 // since there may be some overlaps.
3336 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3338 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3339 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3340 size_limit = MIN2(size_limit, ti_limit);
3341 size_limit = align_size_up(size_limit, BytesPerLong);
3342 int num_tiles = size_limit / BytesPerLong;
3344 // allocate space for the tile map:
3345 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3346 jlong tiles_buf[small_len];
3347 Node* nodes_buf[small_len];
3348 jlong inits_buf[small_len];
3349 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3350 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3351 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3352 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3353 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3354 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3355 // tiles: exact bitwise model of all primitive constants
3356 // nodes: last constant-storing node subsumed into the tiles model
3357 // inits: which bytes (in each tile) are touched by any initializations
3359 //// Pass A: Fill in the tile model with any relevant stores.
3361 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3362 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3363 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3364 Node* zmem = zero_memory(); // initially zero memory state
3365 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3366 Node* st = in(i);
3367 intptr_t st_off = get_store_offset(st, phase);
3369 // Figure out the store's offset and constant value:
3370 if (st_off < header_size) continue; //skip (ignore header)
3371 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3372 int st_size = st->as_Store()->memory_size();
3373 if (st_off + st_size > size_limit) break;
3375 // Record which bytes are touched, whether by constant or not.
3376 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3377 continue; // skip (strange store size)
3379 const Type* val = phase->type(st->in(MemNode::ValueIn));
3380 if (!val->singleton()) continue; //skip (non-con store)
3381 BasicType type = val->basic_type();
3383 jlong con = 0;
3384 switch (type) {
3385 case T_INT: con = val->is_int()->get_con(); break;
3386 case T_LONG: con = val->is_long()->get_con(); break;
3387 case T_FLOAT: con = jint_cast(val->getf()); break;
3388 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3389 default: continue; //skip (odd store type)
3390 }
3392 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3393 st->Opcode() == Op_StoreL) {
3394 continue; // This StoreL is already optimal.
3395 }
3397 // Store down the constant.
3398 store_constant(tiles, num_tiles, st_off, st_size, con);
3400 intptr_t j = st_off >> LogBytesPerLong;
3402 if (type == T_INT && st_size == BytesPerInt
3403 && (st_off & BytesPerInt) == BytesPerInt) {
3404 jlong lcon = tiles[j];
3405 if (!Matcher::isSimpleConstant64(lcon) &&
3406 st->Opcode() == Op_StoreI) {
3407 // This StoreI is already optimal by itself.
3408 jint* intcon = (jint*) &tiles[j];
3409 intcon[1] = 0; // undo the store_constant()
3411 // If the previous store is also optimal by itself, back up and
3412 // undo the action of the previous loop iteration... if we can.
3413 // But if we can't, just let the previous half take care of itself.
3414 st = nodes[j];
3415 st_off -= BytesPerInt;
3416 con = intcon[0];
3417 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3418 assert(st_off >= header_size, "still ignoring header");
3419 assert(get_store_offset(st, phase) == st_off, "must be");
3420 assert(in(i-1) == zmem, "must be");
3421 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3422 assert(con == tcon->is_int()->get_con(), "must be");
3423 // Undo the effects of the previous loop trip, which swallowed st:
3424 intcon[0] = 0; // undo store_constant()
3425 set_req(i-1, st); // undo set_req(i, zmem)
3426 nodes[j] = NULL; // undo nodes[j] = st
3427 --old_subword; // undo ++old_subword
3428 }
3429 continue; // This StoreI is already optimal.
3430 }
3431 }
3433 // This store is not needed.
3434 set_req(i, zmem);
3435 nodes[j] = st; // record for the moment
3436 if (st_size < BytesPerLong) // something has changed
3437 ++old_subword; // includes int/float, but who's counting...
3438 else ++old_long;
3439 }
3441 if ((old_subword + old_long) == 0)
3442 return; // nothing more to do
3444 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3445 // Be sure to insert them before overlapping non-constant stores.
3446 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3447 for (int j = 0; j < num_tiles; j++) {
3448 jlong con = tiles[j];
3449 jlong init = inits[j];
3450 if (con == 0) continue;
3451 jint con0, con1; // split the constant, address-wise
3452 jint init0, init1; // split the init map, address-wise
3453 { union { jlong con; jint intcon[2]; } u;
3454 u.con = con;
3455 con0 = u.intcon[0];
3456 con1 = u.intcon[1];
3457 u.con = init;
3458 init0 = u.intcon[0];
3459 init1 = u.intcon[1];
3460 }
3462 Node* old = nodes[j];
3463 assert(old != NULL, "need the prior store");
3464 intptr_t offset = (j * BytesPerLong);
3466 bool split = !Matcher::isSimpleConstant64(con);
3468 if (offset < header_size) {
3469 assert(offset + BytesPerInt >= header_size, "second int counts");
3470 assert(*(jint*)&tiles[j] == 0, "junk in header");
3471 split = true; // only the second word counts
3472 // Example: int a[] = { 42 ... }
3473 } else if (con0 == 0 && init0 == -1) {
3474 split = true; // first word is covered by full inits
3475 // Example: int a[] = { ... foo(), 42 ... }
3476 } else if (con1 == 0 && init1 == -1) {
3477 split = true; // second word is covered by full inits
3478 // Example: int a[] = { ... 42, foo() ... }
3479 }
3481 // Here's a case where init0 is neither 0 nor -1:
3482 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3483 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3484 // In this case the tile is not split; it is (jlong)42.
3485 // The big tile is stored down, and then the foo() value is inserted.
3486 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3488 Node* ctl = old->in(MemNode::Control);
3489 Node* adr = make_raw_address(offset, phase);
3490 const TypePtr* atp = TypeRawPtr::BOTTOM;
3492 // One or two coalesced stores to plop down.
3493 Node* st[2];
3494 intptr_t off[2];
3495 int nst = 0;
3496 if (!split) {
3497 ++new_long;
3498 off[nst] = offset;
3499 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3500 phase->longcon(con), T_LONG);
3501 } else {
3502 // Omit either if it is a zero.
3503 if (con0 != 0) {
3504 ++new_int;
3505 off[nst] = offset;
3506 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3507 phase->intcon(con0), T_INT);
3508 }
3509 if (con1 != 0) {
3510 ++new_int;
3511 offset += BytesPerInt;
3512 adr = make_raw_address(offset, phase);
3513 off[nst] = offset;
3514 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3515 phase->intcon(con1), T_INT);
3516 }
3517 }
3519 // Insert second store first, then the first before the second.
3520 // Insert each one just before any overlapping non-constant stores.
3521 while (nst > 0) {
3522 Node* st1 = st[--nst];
3523 C->copy_node_notes_to(st1, old);
3524 st1 = phase->transform(st1);
3525 offset = off[nst];
3526 assert(offset >= header_size, "do not smash header");
3527 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3528 guarantee(ins_idx != 0, "must re-insert constant store");
3529 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3530 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3531 set_req(--ins_idx, st1);
3532 else
3533 ins_req(ins_idx, st1);
3534 }
3535 }
3537 if (PrintCompilation && WizardMode)
3538 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3539 old_subword, old_long, new_int, new_long);
3540 if (C->log() != NULL)
3541 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3542 old_subword, old_long, new_int, new_long);
3544 // Clean up any remaining occurrences of zmem:
3545 remove_extra_zeroes();
3546 }
3548 // Explore forward from in(start) to find the first fully initialized
3549 // word, and return its offset. Skip groups of subword stores which
3550 // together initialize full words. If in(start) is itself part of a
3551 // fully initialized word, return the offset of in(start). If there
3552 // are no following full-word stores, or if something is fishy, return
3553 // a negative value.
3554 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3555 int int_map = 0;
3556 intptr_t int_map_off = 0;
3557 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3559 for (uint i = start, limit = req(); i < limit; i++) {
3560 Node* st = in(i);
3562 intptr_t st_off = get_store_offset(st, phase);
3563 if (st_off < 0) break; // return conservative answer
3565 int st_size = st->as_Store()->memory_size();
3566 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3567 return st_off; // we found a complete word init
3568 }
3570 // update the map:
3572 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3573 if (this_int_off != int_map_off) {
3574 // reset the map:
3575 int_map = 0;
3576 int_map_off = this_int_off;
3577 }
3579 int subword_off = st_off - this_int_off;
3580 int_map |= right_n_bits(st_size) << subword_off;
3581 if ((int_map & FULL_MAP) == FULL_MAP) {
3582 return this_int_off; // we found a complete word init
3583 }
3585 // Did this store hit or cross the word boundary?
3586 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3587 if (next_int_off == this_int_off + BytesPerInt) {
3588 // We passed the current int, without fully initializing it.
3589 int_map_off = next_int_off;
3590 int_map >>= BytesPerInt;
3591 } else if (next_int_off > this_int_off + BytesPerInt) {
3592 // We passed the current and next int.
3593 return this_int_off + BytesPerInt;
3594 }
3595 }
3597 return -1;
3598 }
3601 // Called when the associated AllocateNode is expanded into CFG.
3602 // At this point, we may perform additional optimizations.
3603 // Linearize the stores by ascending offset, to make memory
3604 // activity as coherent as possible.
3605 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3606 intptr_t header_size,
3607 Node* size_in_bytes,
3608 PhaseGVN* phase) {
3609 assert(!is_complete(), "not already complete");
3610 assert(stores_are_sane(phase), "");
3611 assert(allocation() != NULL, "must be present");
3613 remove_extra_zeroes();
3615 if (ReduceFieldZeroing || ReduceBulkZeroing)
3616 // reduce instruction count for common initialization patterns
3617 coalesce_subword_stores(header_size, size_in_bytes, phase);
3619 Node* zmem = zero_memory(); // initially zero memory state
3620 Node* inits = zmem; // accumulating a linearized chain of inits
3621 #ifdef ASSERT
3622 intptr_t first_offset = allocation()->minimum_header_size();
3623 intptr_t last_init_off = first_offset; // previous init offset
3624 intptr_t last_init_end = first_offset; // previous init offset+size
3625 intptr_t last_tile_end = first_offset; // previous tile offset+size
3626 #endif
3627 intptr_t zeroes_done = header_size;
3629 bool do_zeroing = true; // we might give up if inits are very sparse
3630 int big_init_gaps = 0; // how many large gaps have we seen?
3632 if (ZeroTLAB) do_zeroing = false;
3633 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3635 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3636 Node* st = in(i);
3637 intptr_t st_off = get_store_offset(st, phase);
3638 if (st_off < 0)
3639 break; // unknown junk in the inits
3640 if (st->in(MemNode::Memory) != zmem)
3641 break; // complicated store chains somehow in list
3643 int st_size = st->as_Store()->memory_size();
3644 intptr_t next_init_off = st_off + st_size;
3646 if (do_zeroing && zeroes_done < next_init_off) {
3647 // See if this store needs a zero before it or under it.
3648 intptr_t zeroes_needed = st_off;
3650 if (st_size < BytesPerInt) {
3651 // Look for subword stores which only partially initialize words.
3652 // If we find some, we must lay down some word-level zeroes first,
3653 // underneath the subword stores.
3654 //
3655 // Examples:
3656 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3657 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3658 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3659 //
3660 // Note: coalesce_subword_stores may have already done this,
3661 // if it was prompted by constant non-zero subword initializers.
3662 // But this case can still arise with non-constant stores.
3664 intptr_t next_full_store = find_next_fullword_store(i, phase);
3666 // In the examples above:
3667 // in(i) p q r s x y z
3668 // st_off 12 13 14 15 12 13 14
3669 // st_size 1 1 1 1 1 1 1
3670 // next_full_s. 12 16 16 16 16 16 16
3671 // z's_done 12 16 16 16 12 16 12
3672 // z's_needed 12 16 16 16 16 16 16
3673 // zsize 0 0 0 0 4 0 4
3674 if (next_full_store < 0) {
3675 // Conservative tack: Zero to end of current word.
3676 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3677 } else {
3678 // Zero to beginning of next fully initialized word.
3679 // Or, don't zero at all, if we are already in that word.
3680 assert(next_full_store >= zeroes_needed, "must go forward");
3681 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3682 zeroes_needed = next_full_store;
3683 }
3684 }
3686 if (zeroes_needed > zeroes_done) {
3687 intptr_t zsize = zeroes_needed - zeroes_done;
3688 // Do some incremental zeroing on rawmem, in parallel with inits.
3689 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3690 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3691 zeroes_done, zeroes_needed,
3692 phase);
3693 zeroes_done = zeroes_needed;
3694 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3695 do_zeroing = false; // leave the hole, next time
3696 }
3697 }
3699 // Collect the store and move on:
3700 st->set_req(MemNode::Memory, inits);
3701 inits = st; // put it on the linearized chain
3702 set_req(i, zmem); // unhook from previous position
3704 if (zeroes_done == st_off)
3705 zeroes_done = next_init_off;
3707 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3709 #ifdef ASSERT
3710 // Various order invariants. Weaker than stores_are_sane because
3711 // a large constant tile can be filled in by smaller non-constant stores.
3712 assert(st_off >= last_init_off, "inits do not reverse");
3713 last_init_off = st_off;
3714 const Type* val = NULL;
3715 if (st_size >= BytesPerInt &&
3716 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3717 (int)val->basic_type() < (int)T_OBJECT) {
3718 assert(st_off >= last_tile_end, "tiles do not overlap");
3719 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3720 last_tile_end = MAX2(last_tile_end, next_init_off);
3721 } else {
3722 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3723 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3724 assert(st_off >= last_init_end, "inits do not overlap");
3725 last_init_end = next_init_off; // it's a non-tile
3726 }
3727 #endif //ASSERT
3728 }
3730 remove_extra_zeroes(); // clear out all the zmems left over
3731 add_req(inits);
3733 if (!ZeroTLAB) {
3734 // If anything remains to be zeroed, zero it all now.
3735 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3736 // if it is the last unused 4 bytes of an instance, forget about it
3737 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3738 if (zeroes_done + BytesPerLong >= size_limit) {
3739 assert(allocation() != NULL, "");
3740 if (allocation()->Opcode() == Op_Allocate) {
3741 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3742 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3743 if (zeroes_done == k->layout_helper())
3744 zeroes_done = size_limit;
3745 }
3746 }
3747 if (zeroes_done < size_limit) {
3748 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3749 zeroes_done, size_in_bytes, phase);
3750 }
3751 }
3753 set_complete(phase);
3754 return rawmem;
3755 }
3758 #ifdef ASSERT
3759 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3760 if (is_complete())
3761 return true; // stores could be anything at this point
3762 assert(allocation() != NULL, "must be present");
3763 intptr_t last_off = allocation()->minimum_header_size();
3764 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3765 Node* st = in(i);
3766 intptr_t st_off = get_store_offset(st, phase);
3767 if (st_off < 0) continue; // ignore dead garbage
3768 if (last_off > st_off) {
3769 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3770 this->dump(2);
3771 assert(false, "ascending store offsets");
3772 return false;
3773 }
3774 last_off = st_off + st->as_Store()->memory_size();
3775 }
3776 return true;
3777 }
3778 #endif //ASSERT
3783 //============================MergeMemNode=====================================
3784 //
3785 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3786 // contributing store or call operations. Each contributor provides the memory
3787 // state for a particular "alias type" (see Compile::alias_type). For example,
3788 // if a MergeMem has an input X for alias category #6, then any memory reference
3789 // to alias category #6 may use X as its memory state input, as an exact equivalent
3790 // to using the MergeMem as a whole.
3791 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3792 //
3793 // (Here, the <N> notation gives the index of the relevant adr_type.)
3794 //
3795 // In one special case (and more cases in the future), alias categories overlap.
3796 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3797 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3798 // it is exactly equivalent to that state W:
3799 // MergeMem(<Bot>: W) <==> W
3800 //
3801 // Usually, the merge has more than one input. In that case, where inputs
3802 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3803 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3804 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3805 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3806 //
3807 // A merge can take a "wide" memory state as one of its narrow inputs.
3808 // This simply means that the merge observes out only the relevant parts of
3809 // the wide input. That is, wide memory states arriving at narrow merge inputs
3810 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3811 //
3812 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3813 // and that memory slices "leak through":
3814 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3815 //
3816 // But, in such a cascade, repeated memory slices can "block the leak":
3817 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3818 //
3819 // In the last example, Y is not part of the combined memory state of the
3820 // outermost MergeMem. The system must, of course, prevent unschedulable
3821 // memory states from arising, so you can be sure that the state Y is somehow
3822 // a precursor to state Y'.
3823 //
3824 //
3825 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3826 // of each MergeMemNode array are exactly the numerical alias indexes, including
3827 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3828 // Compile::alias_type (and kin) produce and manage these indexes.
3829 //
3830 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3831 // (Note that this provides quick access to the top node inside MergeMem methods,
3832 // without the need to reach out via TLS to Compile::current.)
3833 //
3834 // As a consequence of what was just described, a MergeMem that represents a full
3835 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3836 // containing all alias categories.
3837 //
3838 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3839 //
3840 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3841 // a memory state for the alias type <N>, or else the top node, meaning that
3842 // there is no particular input for that alias type. Note that the length of
3843 // a MergeMem is variable, and may be extended at any time to accommodate new
3844 // memory states at larger alias indexes. When merges grow, they are of course
3845 // filled with "top" in the unused in() positions.
3846 //
3847 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3848 // (Top was chosen because it works smoothly with passes like GCM.)
3849 //
3850 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3851 // the type of random VM bits like TLS references.) Since it is always the
3852 // first non-Bot memory slice, some low-level loops use it to initialize an
3853 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3854 //
3855 //
3856 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3857 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3858 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3859 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3860 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3861 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3862 //
3863 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3864 // really that different from the other memory inputs. An abbreviation called
3865 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3866 //
3867 //
3868 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3869 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3870 // that "emerges though" the base memory will be marked as excluding the alias types
3871 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3872 //
3873 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3874 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3875 //
3876 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3877 // (It is currently unimplemented.) As you can see, the resulting merge is
3878 // actually a disjoint union of memory states, rather than an overlay.
3879 //
3881 //------------------------------MergeMemNode-----------------------------------
3882 Node* MergeMemNode::make_empty_memory() {
3883 Node* empty_memory = (Node*) Compile::current()->top();
3884 assert(empty_memory->is_top(), "correct sentinel identity");
3885 return empty_memory;
3886 }
3888 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3889 init_class_id(Class_MergeMem);
3890 // all inputs are nullified in Node::Node(int)
3891 // set_input(0, NULL); // no control input
3893 // Initialize the edges uniformly to top, for starters.
3894 Node* empty_mem = make_empty_memory();
3895 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3896 init_req(i,empty_mem);
3897 }
3898 assert(empty_memory() == empty_mem, "");
3900 if( new_base != NULL && new_base->is_MergeMem() ) {
3901 MergeMemNode* mdef = new_base->as_MergeMem();
3902 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3903 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3904 mms.set_memory(mms.memory2());
3905 }
3906 assert(base_memory() == mdef->base_memory(), "");
3907 } else {
3908 set_base_memory(new_base);
3909 }
3910 }
3912 // Make a new, untransformed MergeMem with the same base as 'mem'.
3913 // If mem is itself a MergeMem, populate the result with the same edges.
3914 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3915 return new(C) MergeMemNode(mem);
3916 }
3918 //------------------------------cmp--------------------------------------------
3919 uint MergeMemNode::hash() const { return NO_HASH; }
3920 uint MergeMemNode::cmp( const Node &n ) const {
3921 return (&n == this); // Always fail except on self
3922 }
3924 //------------------------------Identity---------------------------------------
3925 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3926 // Identity if this merge point does not record any interesting memory
3927 // disambiguations.
3928 Node* base_mem = base_memory();
3929 Node* empty_mem = empty_memory();
3930 if (base_mem != empty_mem) { // Memory path is not dead?
3931 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3932 Node* mem = in(i);
3933 if (mem != empty_mem && mem != base_mem) {
3934 return this; // Many memory splits; no change
3935 }
3936 }
3937 }
3938 return base_mem; // No memory splits; ID on the one true input
3939 }
3941 //------------------------------Ideal------------------------------------------
3942 // This method is invoked recursively on chains of MergeMem nodes
3943 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3944 // Remove chain'd MergeMems
3945 //
3946 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3947 // relative to the "in(Bot)". Since we are patching both at the same time,
3948 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3949 // but rewrite each "in(i)" relative to the new "in(Bot)".
3950 Node *progress = NULL;
3953 Node* old_base = base_memory();
3954 Node* empty_mem = empty_memory();
3955 if (old_base == empty_mem)
3956 return NULL; // Dead memory path.
3958 MergeMemNode* old_mbase;
3959 if (old_base != NULL && old_base->is_MergeMem())
3960 old_mbase = old_base->as_MergeMem();
3961 else
3962 old_mbase = NULL;
3963 Node* new_base = old_base;
3965 // simplify stacked MergeMems in base memory
3966 if (old_mbase) new_base = old_mbase->base_memory();
3968 // the base memory might contribute new slices beyond my req()
3969 if (old_mbase) grow_to_match(old_mbase);
3971 // Look carefully at the base node if it is a phi.
3972 PhiNode* phi_base;
3973 if (new_base != NULL && new_base->is_Phi())
3974 phi_base = new_base->as_Phi();
3975 else
3976 phi_base = NULL;
3978 Node* phi_reg = NULL;
3979 uint phi_len = (uint)-1;
3980 if (phi_base != NULL && !phi_base->is_copy()) {
3981 // do not examine phi if degraded to a copy
3982 phi_reg = phi_base->region();
3983 phi_len = phi_base->req();
3984 // see if the phi is unfinished
3985 for (uint i = 1; i < phi_len; i++) {
3986 if (phi_base->in(i) == NULL) {
3987 // incomplete phi; do not look at it yet!
3988 phi_reg = NULL;
3989 phi_len = (uint)-1;
3990 break;
3991 }
3992 }
3993 }
3995 // Note: We do not call verify_sparse on entry, because inputs
3996 // can normalize to the base_memory via subsume_node or similar
3997 // mechanisms. This method repairs that damage.
3999 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4001 // Look at each slice.
4002 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4003 Node* old_in = in(i);
4004 // calculate the old memory value
4005 Node* old_mem = old_in;
4006 if (old_mem == empty_mem) old_mem = old_base;
4007 assert(old_mem == memory_at(i), "");
4009 // maybe update (reslice) the old memory value
4011 // simplify stacked MergeMems
4012 Node* new_mem = old_mem;
4013 MergeMemNode* old_mmem;
4014 if (old_mem != NULL && old_mem->is_MergeMem())
4015 old_mmem = old_mem->as_MergeMem();
4016 else
4017 old_mmem = NULL;
4018 if (old_mmem == this) {
4019 // This can happen if loops break up and safepoints disappear.
4020 // A merge of BotPtr (default) with a RawPtr memory derived from a
4021 // safepoint can be rewritten to a merge of the same BotPtr with
4022 // the BotPtr phi coming into the loop. If that phi disappears
4023 // also, we can end up with a self-loop of the mergemem.
4024 // In general, if loops degenerate and memory effects disappear,
4025 // a mergemem can be left looking at itself. This simply means
4026 // that the mergemem's default should be used, since there is
4027 // no longer any apparent effect on this slice.
4028 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4029 // from start. Update the input to TOP.
4030 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4031 }
4032 else if (old_mmem != NULL) {
4033 new_mem = old_mmem->memory_at(i);
4034 }
4035 // else preceding memory was not a MergeMem
4037 // replace equivalent phis (unfortunately, they do not GVN together)
4038 if (new_mem != NULL && new_mem != new_base &&
4039 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4040 if (new_mem->is_Phi()) {
4041 PhiNode* phi_mem = new_mem->as_Phi();
4042 for (uint i = 1; i < phi_len; i++) {
4043 if (phi_base->in(i) != phi_mem->in(i)) {
4044 phi_mem = NULL;
4045 break;
4046 }
4047 }
4048 if (phi_mem != NULL) {
4049 // equivalent phi nodes; revert to the def
4050 new_mem = new_base;
4051 }
4052 }
4053 }
4055 // maybe store down a new value
4056 Node* new_in = new_mem;
4057 if (new_in == new_base) new_in = empty_mem;
4059 if (new_in != old_in) {
4060 // Warning: Do not combine this "if" with the previous "if"
4061 // A memory slice might have be be rewritten even if it is semantically
4062 // unchanged, if the base_memory value has changed.
4063 set_req(i, new_in);
4064 progress = this; // Report progress
4065 }
4066 }
4068 if (new_base != old_base) {
4069 set_req(Compile::AliasIdxBot, new_base);
4070 // Don't use set_base_memory(new_base), because we need to update du.
4071 assert(base_memory() == new_base, "");
4072 progress = this;
4073 }
4075 if( base_memory() == this ) {
4076 // a self cycle indicates this memory path is dead
4077 set_req(Compile::AliasIdxBot, empty_mem);
4078 }
4080 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4081 // Recursion must occur after the self cycle check above
4082 if( base_memory()->is_MergeMem() ) {
4083 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4084 Node *m = phase->transform(new_mbase); // Rollup any cycles
4085 if( m != NULL && (m->is_top() ||
4086 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4087 // propagate rollup of dead cycle to self
4088 set_req(Compile::AliasIdxBot, empty_mem);
4089 }
4090 }
4092 if( base_memory() == empty_mem ) {
4093 progress = this;
4094 // Cut inputs during Parse phase only.
4095 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4096 if( !can_reshape ) {
4097 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4098 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4099 }
4100 }
4101 }
4103 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4104 // Check if PhiNode::Ideal's "Split phis through memory merges"
4105 // transform should be attempted. Look for this->phi->this cycle.
4106 uint merge_width = req();
4107 if (merge_width > Compile::AliasIdxRaw) {
4108 PhiNode* phi = base_memory()->as_Phi();
4109 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4110 if (phi->in(i) == this) {
4111 phase->is_IterGVN()->_worklist.push(phi);
4112 break;
4113 }
4114 }
4115 }
4116 }
4118 assert(progress || verify_sparse(), "please, no dups of base");
4119 return progress;
4120 }
4122 //-------------------------set_base_memory-------------------------------------
4123 void MergeMemNode::set_base_memory(Node *new_base) {
4124 Node* empty_mem = empty_memory();
4125 set_req(Compile::AliasIdxBot, new_base);
4126 assert(memory_at(req()) == new_base, "must set default memory");
4127 // Clear out other occurrences of new_base:
4128 if (new_base != empty_mem) {
4129 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4130 if (in(i) == new_base) set_req(i, empty_mem);
4131 }
4132 }
4133 }
4135 //------------------------------out_RegMask------------------------------------
4136 const RegMask &MergeMemNode::out_RegMask() const {
4137 return RegMask::Empty;
4138 }
4140 //------------------------------dump_spec--------------------------------------
4141 #ifndef PRODUCT
4142 void MergeMemNode::dump_spec(outputStream *st) const {
4143 st->print(" {");
4144 Node* base_mem = base_memory();
4145 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4146 Node* mem = memory_at(i);
4147 if (mem == base_mem) { st->print(" -"); continue; }
4148 st->print( " N%d:", mem->_idx );
4149 Compile::current()->get_adr_type(i)->dump_on(st);
4150 }
4151 st->print(" }");
4152 }
4153 #endif // !PRODUCT
4156 #ifdef ASSERT
4157 static bool might_be_same(Node* a, Node* b) {
4158 if (a == b) return true;
4159 if (!(a->is_Phi() || b->is_Phi())) return false;
4160 // phis shift around during optimization
4161 return true; // pretty stupid...
4162 }
4164 // verify a narrow slice (either incoming or outgoing)
4165 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4166 if (!VerifyAliases) return; // don't bother to verify unless requested
4167 if (is_error_reported()) return; // muzzle asserts when debugging an error
4168 if (Node::in_dump()) return; // muzzle asserts when printing
4169 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4170 assert(n != NULL, "");
4171 // Elide intervening MergeMem's
4172 while (n->is_MergeMem()) {
4173 n = n->as_MergeMem()->memory_at(alias_idx);
4174 }
4175 Compile* C = Compile::current();
4176 const TypePtr* n_adr_type = n->adr_type();
4177 if (n == m->empty_memory()) {
4178 // Implicit copy of base_memory()
4179 } else if (n_adr_type != TypePtr::BOTTOM) {
4180 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4181 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4182 } else {
4183 // A few places like make_runtime_call "know" that VM calls are narrow,
4184 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4185 bool expected_wide_mem = false;
4186 if (n == m->base_memory()) {
4187 expected_wide_mem = true;
4188 } else if (alias_idx == Compile::AliasIdxRaw ||
4189 n == m->memory_at(Compile::AliasIdxRaw)) {
4190 expected_wide_mem = true;
4191 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4192 // memory can "leak through" calls on channels that
4193 // are write-once. Allow this also.
4194 expected_wide_mem = true;
4195 }
4196 assert(expected_wide_mem, "expected narrow slice replacement");
4197 }
4198 }
4199 #else // !ASSERT
4200 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
4201 #endif
4204 //-----------------------------memory_at---------------------------------------
4205 Node* MergeMemNode::memory_at(uint alias_idx) const {
4206 assert(alias_idx >= Compile::AliasIdxRaw ||
4207 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4208 "must avoid base_memory and AliasIdxTop");
4210 // Otherwise, it is a narrow slice.
4211 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4212 Compile *C = Compile::current();
4213 if (is_empty_memory(n)) {
4214 // the array is sparse; empty slots are the "top" node
4215 n = base_memory();
4216 assert(Node::in_dump()
4217 || n == NULL || n->bottom_type() == Type::TOP
4218 || n->adr_type() == NULL // address is TOP
4219 || n->adr_type() == TypePtr::BOTTOM
4220 || n->adr_type() == TypeRawPtr::BOTTOM
4221 || Compile::current()->AliasLevel() == 0,
4222 "must be a wide memory");
4223 // AliasLevel == 0 if we are organizing the memory states manually.
4224 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4225 } else {
4226 // make sure the stored slice is sane
4227 #ifdef ASSERT
4228 if (is_error_reported() || Node::in_dump()) {
4229 } else if (might_be_same(n, base_memory())) {
4230 // Give it a pass: It is a mostly harmless repetition of the base.
4231 // This can arise normally from node subsumption during optimization.
4232 } else {
4233 verify_memory_slice(this, alias_idx, n);
4234 }
4235 #endif
4236 }
4237 return n;
4238 }
4240 //---------------------------set_memory_at-------------------------------------
4241 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4242 verify_memory_slice(this, alias_idx, n);
4243 Node* empty_mem = empty_memory();
4244 if (n == base_memory()) n = empty_mem; // collapse default
4245 uint need_req = alias_idx+1;
4246 if (req() < need_req) {
4247 if (n == empty_mem) return; // already the default, so do not grow me
4248 // grow the sparse array
4249 do {
4250 add_req(empty_mem);
4251 } while (req() < need_req);
4252 }
4253 set_req( alias_idx, n );
4254 }
4258 //--------------------------iteration_setup------------------------------------
4259 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4260 if (other != NULL) {
4261 grow_to_match(other);
4262 // invariant: the finite support of mm2 is within mm->req()
4263 #ifdef ASSERT
4264 for (uint i = req(); i < other->req(); i++) {
4265 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4266 }
4267 #endif
4268 }
4269 // Replace spurious copies of base_memory by top.
4270 Node* base_mem = base_memory();
4271 if (base_mem != NULL && !base_mem->is_top()) {
4272 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4273 if (in(i) == base_mem)
4274 set_req(i, empty_memory());
4275 }
4276 }
4277 }
4279 //---------------------------grow_to_match-------------------------------------
4280 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4281 Node* empty_mem = empty_memory();
4282 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4283 // look for the finite support of the other memory
4284 for (uint i = other->req(); --i >= req(); ) {
4285 if (other->in(i) != empty_mem) {
4286 uint new_len = i+1;
4287 while (req() < new_len) add_req(empty_mem);
4288 break;
4289 }
4290 }
4291 }
4293 //---------------------------verify_sparse-------------------------------------
4294 #ifndef PRODUCT
4295 bool MergeMemNode::verify_sparse() const {
4296 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4297 Node* base_mem = base_memory();
4298 // The following can happen in degenerate cases, since empty==top.
4299 if (is_empty_memory(base_mem)) return true;
4300 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4301 assert(in(i) != NULL, "sane slice");
4302 if (in(i) == base_mem) return false; // should have been the sentinel value!
4303 }
4304 return true;
4305 }
4307 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4308 Node* n;
4309 n = mm->in(idx);
4310 if (mem == n) return true; // might be empty_memory()
4311 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4312 if (mem == n) return true;
4313 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4314 if (mem == n) return true;
4315 if (n == NULL) break;
4316 }
4317 return false;
4318 }
4319 #endif // !PRODUCT