annotate src/share/vm/opto/memnode.cpp @ 68:daf38130e60d

6676841: ClearArrayNode::Identity is incorrect for 64-bit Summary: ClearArrayNode::Identity should use TypeX instead of TypeInt Reviewed-by: jrose, kvn, sgoldman
author never
date Tue, 18 Mar 2008 23:44:46 -0700
parents b8f5ba577b02
children 2a9af0b9cb1c
rev   line source
duke@0 1 /*
duke@0 2 * Copyright 1997-2007 Sun Microsystems, Inc. All Rights Reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
duke@0 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
duke@0 20 * CA 95054 USA or visit www.sun.com if you need additional information or
duke@0 21 * have any questions.
duke@0 22 *
duke@0 23 */
duke@0 24
duke@0 25 // Portions of code courtesy of Clifford Click
duke@0 26
duke@0 27 // Optimization - Graph Style
duke@0 28
duke@0 29 #include "incls/_precompiled.incl"
duke@0 30 #include "incls/_memnode.cpp.incl"
duke@0 31
duke@0 32 //=============================================================================
duke@0 33 uint MemNode::size_of() const { return sizeof(*this); }
duke@0 34
duke@0 35 const TypePtr *MemNode::adr_type() const {
duke@0 36 Node* adr = in(Address);
duke@0 37 const TypePtr* cross_check = NULL;
duke@0 38 DEBUG_ONLY(cross_check = _adr_type);
duke@0 39 return calculate_adr_type(adr->bottom_type(), cross_check);
duke@0 40 }
duke@0 41
duke@0 42 #ifndef PRODUCT
duke@0 43 void MemNode::dump_spec(outputStream *st) const {
duke@0 44 if (in(Address) == NULL) return; // node is dead
duke@0 45 #ifndef ASSERT
duke@0 46 // fake the missing field
duke@0 47 const TypePtr* _adr_type = NULL;
duke@0 48 if (in(Address) != NULL)
duke@0 49 _adr_type = in(Address)->bottom_type()->isa_ptr();
duke@0 50 #endif
duke@0 51 dump_adr_type(this, _adr_type, st);
duke@0 52
duke@0 53 Compile* C = Compile::current();
duke@0 54 if( C->alias_type(_adr_type)->is_volatile() )
duke@0 55 st->print(" Volatile!");
duke@0 56 }
duke@0 57
duke@0 58 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
duke@0 59 st->print(" @");
duke@0 60 if (adr_type == NULL) {
duke@0 61 st->print("NULL");
duke@0 62 } else {
duke@0 63 adr_type->dump_on(st);
duke@0 64 Compile* C = Compile::current();
duke@0 65 Compile::AliasType* atp = NULL;
duke@0 66 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
duke@0 67 if (atp == NULL)
duke@0 68 st->print(", idx=?\?;");
duke@0 69 else if (atp->index() == Compile::AliasIdxBot)
duke@0 70 st->print(", idx=Bot;");
duke@0 71 else if (atp->index() == Compile::AliasIdxTop)
duke@0 72 st->print(", idx=Top;");
duke@0 73 else if (atp->index() == Compile::AliasIdxRaw)
duke@0 74 st->print(", idx=Raw;");
duke@0 75 else {
duke@0 76 ciField* field = atp->field();
duke@0 77 if (field) {
duke@0 78 st->print(", name=");
duke@0 79 field->print_name_on(st);
duke@0 80 }
duke@0 81 st->print(", idx=%d;", atp->index());
duke@0 82 }
duke@0 83 }
duke@0 84 }
duke@0 85
duke@0 86 extern void print_alias_types();
duke@0 87
duke@0 88 #endif
duke@0 89
kvn@64 90 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
kvn@64 91 uint alias_idx = phase->C->get_alias_index(tp);
kvn@64 92 Node *mem = mmem;
kvn@64 93 #ifdef ASSERT
kvn@64 94 {
kvn@64 95 // Check that current type is consistent with the alias index used during graph construction
kvn@64 96 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
kvn@64 97 bool consistent = adr_check == NULL || adr_check->empty() ||
kvn@64 98 phase->C->must_alias(adr_check, alias_idx );
kvn@64 99 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
kvn@64 100 if( !consistent && adr_check != NULL && !adr_check->empty() &&
kvn@64 101 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
kvn@64 102 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
kvn@64 103 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
kvn@64 104 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
kvn@64 105 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
kvn@64 106 // don't assert if it is dead code.
kvn@64 107 consistent = true;
kvn@64 108 }
kvn@64 109 if( !consistent ) {
kvn@64 110 st->print("alias_idx==%d, adr_check==", alias_idx);
kvn@64 111 if( adr_check == NULL ) {
kvn@64 112 st->print("NULL");
kvn@64 113 } else {
kvn@64 114 adr_check->dump();
kvn@64 115 }
kvn@64 116 st->cr();
kvn@64 117 print_alias_types();
kvn@64 118 assert(consistent, "adr_check must match alias idx");
kvn@64 119 }
kvn@64 120 }
kvn@64 121 #endif
kvn@64 122 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
kvn@64 123 // means an array I have not precisely typed yet. Do not do any
kvn@64 124 // alias stuff with it any time soon.
kvn@64 125 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@64 126 if( tp->base() != Type::AnyPtr &&
kvn@64 127 !(tinst &&
kvn@64 128 tinst->klass()->is_java_lang_Object() &&
kvn@64 129 tinst->offset() == Type::OffsetBot) ) {
kvn@64 130 // compress paths and change unreachable cycles to TOP
kvn@64 131 // If not, we can update the input infinitely along a MergeMem cycle
kvn@64 132 // Equivalent code in PhiNode::Ideal
kvn@64 133 Node* m = phase->transform(mmem);
kvn@64 134 // If tranformed to a MergeMem, get the desired slice
kvn@64 135 // Otherwise the returned node represents memory for every slice
kvn@64 136 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
kvn@64 137 // Update input if it is progress over what we have now
kvn@64 138 }
kvn@64 139 return mem;
kvn@64 140 }
kvn@64 141
duke@0 142 //--------------------------Ideal_common---------------------------------------
duke@0 143 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
duke@0 144 // Unhook non-raw memories from complete (macro-expanded) initializations.
duke@0 145 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
duke@0 146 // If our control input is a dead region, kill all below the region
duke@0 147 Node *ctl = in(MemNode::Control);
duke@0 148 if (ctl && remove_dead_region(phase, can_reshape))
duke@0 149 return this;
duke@0 150
duke@0 151 // Ignore if memory is dead, or self-loop
duke@0 152 Node *mem = in(MemNode::Memory);
duke@0 153 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 154 assert( mem != this, "dead loop in MemNode::Ideal" );
duke@0 155
duke@0 156 Node *address = in(MemNode::Address);
duke@0 157 const Type *t_adr = phase->type( address );
duke@0 158 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 159
duke@0 160 // Avoid independent memory operations
duke@0 161 Node* old_mem = mem;
duke@0 162
kvn@36 163 // The code which unhooks non-raw memories from complete (macro-expanded)
kvn@36 164 // initializations was removed. After macro-expansion all stores catched
kvn@36 165 // by Initialize node became raw stores and there is no information
kvn@36 166 // which memory slices they modify. So it is unsafe to move any memory
kvn@36 167 // operation above these stores. Also in most cases hooked non-raw memories
kvn@36 168 // were already unhooked by using information from detect_ptr_independence()
kvn@36 169 // and find_previous_store().
duke@0 170
duke@0 171 if (mem->is_MergeMem()) {
duke@0 172 MergeMemNode* mmem = mem->as_MergeMem();
duke@0 173 const TypePtr *tp = t_adr->is_ptr();
kvn@64 174
kvn@64 175 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
duke@0 176 }
duke@0 177
duke@0 178 if (mem != old_mem) {
duke@0 179 set_req(MemNode::Memory, mem);
duke@0 180 return this;
duke@0 181 }
duke@0 182
duke@0 183 // let the subclass continue analyzing...
duke@0 184 return NULL;
duke@0 185 }
duke@0 186
duke@0 187 // Helper function for proving some simple control dominations.
duke@0 188 // Attempt to prove that control input 'dom' dominates (or equals) 'sub'.
duke@0 189 // Already assumes that 'dom' is available at 'sub', and that 'sub'
duke@0 190 // is not a constant (dominated by the method's StartNode).
duke@0 191 // Used by MemNode::find_previous_store to prove that the
duke@0 192 // control input of a memory operation predates (dominates)
duke@0 193 // an allocation it wants to look past.
duke@0 194 bool MemNode::detect_dominating_control(Node* dom, Node* sub) {
duke@0 195 if (dom == NULL) return false;
duke@0 196 if (dom->is_Proj()) dom = dom->in(0);
duke@0 197 if (dom->is_Start()) return true; // anything inside the method
duke@0 198 if (dom->is_Root()) return true; // dom 'controls' a constant
duke@0 199 int cnt = 20; // detect cycle or too much effort
duke@0 200 while (sub != NULL) { // walk 'sub' up the chain to 'dom'
duke@0 201 if (--cnt < 0) return false; // in a cycle or too complex
duke@0 202 if (sub == dom) return true;
duke@0 203 if (sub->is_Start()) return false;
duke@0 204 if (sub->is_Root()) return false;
duke@0 205 Node* up = sub->in(0);
duke@0 206 if (sub == up && sub->is_Region()) {
duke@0 207 for (uint i = 1; i < sub->req(); i++) {
duke@0 208 Node* in = sub->in(i);
duke@0 209 if (in != NULL && !in->is_top() && in != sub) {
duke@0 210 up = in; break; // take any path on the way up to 'dom'
duke@0 211 }
duke@0 212 }
duke@0 213 }
duke@0 214 if (sub == up) return false; // some kind of tight cycle
duke@0 215 sub = up;
duke@0 216 }
duke@0 217 return false;
duke@0 218 }
duke@0 219
duke@0 220 //---------------------detect_ptr_independence---------------------------------
duke@0 221 // Used by MemNode::find_previous_store to prove that two base
duke@0 222 // pointers are never equal.
duke@0 223 // The pointers are accompanied by their associated allocations,
duke@0 224 // if any, which have been previously discovered by the caller.
duke@0 225 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
duke@0 226 Node* p2, AllocateNode* a2,
duke@0 227 PhaseTransform* phase) {
duke@0 228 // Attempt to prove that these two pointers cannot be aliased.
duke@0 229 // They may both manifestly be allocations, and they should differ.
duke@0 230 // Or, if they are not both allocations, they can be distinct constants.
duke@0 231 // Otherwise, one is an allocation and the other a pre-existing value.
duke@0 232 if (a1 == NULL && a2 == NULL) { // neither an allocation
duke@0 233 return (p1 != p2) && p1->is_Con() && p2->is_Con();
duke@0 234 } else if (a1 != NULL && a2 != NULL) { // both allocations
duke@0 235 return (a1 != a2);
duke@0 236 } else if (a1 != NULL) { // one allocation a1
duke@0 237 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
duke@0 238 return detect_dominating_control(p2->in(0), a1->in(0));
duke@0 239 } else { //(a2 != NULL) // one allocation a2
duke@0 240 return detect_dominating_control(p1->in(0), a2->in(0));
duke@0 241 }
duke@0 242 return false;
duke@0 243 }
duke@0 244
duke@0 245
duke@0 246 // The logic for reordering loads and stores uses four steps:
duke@0 247 // (a) Walk carefully past stores and initializations which we
duke@0 248 // can prove are independent of this load.
duke@0 249 // (b) Observe that the next memory state makes an exact match
duke@0 250 // with self (load or store), and locate the relevant store.
duke@0 251 // (c) Ensure that, if we were to wire self directly to the store,
duke@0 252 // the optimizer would fold it up somehow.
duke@0 253 // (d) Do the rewiring, and return, depending on some other part of
duke@0 254 // the optimizer to fold up the load.
duke@0 255 // This routine handles steps (a) and (b). Steps (c) and (d) are
duke@0 256 // specific to loads and stores, so they are handled by the callers.
duke@0 257 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
duke@0 258 //
duke@0 259 Node* MemNode::find_previous_store(PhaseTransform* phase) {
duke@0 260 Node* ctrl = in(MemNode::Control);
duke@0 261 Node* adr = in(MemNode::Address);
duke@0 262 intptr_t offset = 0;
duke@0 263 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 264 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
duke@0 265
duke@0 266 if (offset == Type::OffsetBot)
duke@0 267 return NULL; // cannot unalias unless there are precise offsets
duke@0 268
duke@0 269 intptr_t size_in_bytes = memory_size();
duke@0 270
duke@0 271 Node* mem = in(MemNode::Memory); // start searching here...
duke@0 272
duke@0 273 int cnt = 50; // Cycle limiter
duke@0 274 for (;;) { // While we can dance past unrelated stores...
duke@0 275 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
duke@0 276
duke@0 277 if (mem->is_Store()) {
duke@0 278 Node* st_adr = mem->in(MemNode::Address);
duke@0 279 intptr_t st_offset = 0;
duke@0 280 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
duke@0 281 if (st_base == NULL)
duke@0 282 break; // inscrutable pointer
duke@0 283 if (st_offset != offset && st_offset != Type::OffsetBot) {
duke@0 284 const int MAX_STORE = BytesPerLong;
duke@0 285 if (st_offset >= offset + size_in_bytes ||
duke@0 286 st_offset <= offset - MAX_STORE ||
duke@0 287 st_offset <= offset - mem->as_Store()->memory_size()) {
duke@0 288 // Success: The offsets are provably independent.
duke@0 289 // (You may ask, why not just test st_offset != offset and be done?
duke@0 290 // The answer is that stores of different sizes can co-exist
duke@0 291 // in the same sequence of RawMem effects. We sometimes initialize
duke@0 292 // a whole 'tile' of array elements with a single jint or jlong.)
duke@0 293 mem = mem->in(MemNode::Memory);
duke@0 294 continue; // (a) advance through independent store memory
duke@0 295 }
duke@0 296 }
duke@0 297 if (st_base != base &&
duke@0 298 detect_ptr_independence(base, alloc,
duke@0 299 st_base,
duke@0 300 AllocateNode::Ideal_allocation(st_base, phase),
duke@0 301 phase)) {
duke@0 302 // Success: The bases are provably independent.
duke@0 303 mem = mem->in(MemNode::Memory);
duke@0 304 continue; // (a) advance through independent store memory
duke@0 305 }
duke@0 306
duke@0 307 // (b) At this point, if the bases or offsets do not agree, we lose,
duke@0 308 // since we have not managed to prove 'this' and 'mem' independent.
duke@0 309 if (st_base == base && st_offset == offset) {
duke@0 310 return mem; // let caller handle steps (c), (d)
duke@0 311 }
duke@0 312
duke@0 313 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 314 InitializeNode* st_init = mem->in(0)->as_Initialize();
duke@0 315 AllocateNode* st_alloc = st_init->allocation();
duke@0 316 if (st_alloc == NULL)
duke@0 317 break; // something degenerated
duke@0 318 bool known_identical = false;
duke@0 319 bool known_independent = false;
duke@0 320 if (alloc == st_alloc)
duke@0 321 known_identical = true;
duke@0 322 else if (alloc != NULL)
duke@0 323 known_independent = true;
duke@0 324 else if (ctrl != NULL &&
duke@0 325 detect_dominating_control(ctrl, st_alloc->in(0)))
duke@0 326 known_independent = true;
duke@0 327
duke@0 328 if (known_independent) {
duke@0 329 // The bases are provably independent: Either they are
duke@0 330 // manifestly distinct allocations, or else the control
duke@0 331 // of this load dominates the store's allocation.
duke@0 332 int alias_idx = phase->C->get_alias_index(adr_type());
duke@0 333 if (alias_idx == Compile::AliasIdxRaw) {
duke@0 334 mem = st_alloc->in(TypeFunc::Memory);
duke@0 335 } else {
duke@0 336 mem = st_init->memory(alias_idx);
duke@0 337 }
duke@0 338 continue; // (a) advance through independent store memory
duke@0 339 }
duke@0 340
duke@0 341 // (b) at this point, if we are not looking at a store initializing
duke@0 342 // the same allocation we are loading from, we lose.
duke@0 343 if (known_identical) {
duke@0 344 // From caller, can_see_stored_value will consult find_captured_store.
duke@0 345 return mem; // let caller handle steps (c), (d)
duke@0 346 }
duke@0 347
duke@0 348 }
duke@0 349
duke@0 350 // Unless there is an explicit 'continue', we must bail out here,
duke@0 351 // because 'mem' is an inscrutable memory state (e.g., a call).
duke@0 352 break;
duke@0 353 }
duke@0 354
duke@0 355 return NULL; // bail out
duke@0 356 }
duke@0 357
duke@0 358 //----------------------calculate_adr_type-------------------------------------
duke@0 359 // Helper function. Notices when the given type of address hits top or bottom.
duke@0 360 // Also, asserts a cross-check of the type against the expected address type.
duke@0 361 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
duke@0 362 if (t == Type::TOP) return NULL; // does not touch memory any more?
duke@0 363 #ifdef PRODUCT
duke@0 364 cross_check = NULL;
duke@0 365 #else
duke@0 366 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
duke@0 367 #endif
duke@0 368 const TypePtr* tp = t->isa_ptr();
duke@0 369 if (tp == NULL) {
duke@0 370 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
duke@0 371 return TypePtr::BOTTOM; // touches lots of memory
duke@0 372 } else {
duke@0 373 #ifdef ASSERT
duke@0 374 // %%%% [phh] We don't check the alias index if cross_check is
duke@0 375 // TypeRawPtr::BOTTOM. Needs to be investigated.
duke@0 376 if (cross_check != NULL &&
duke@0 377 cross_check != TypePtr::BOTTOM &&
duke@0 378 cross_check != TypeRawPtr::BOTTOM) {
duke@0 379 // Recheck the alias index, to see if it has changed (due to a bug).
duke@0 380 Compile* C = Compile::current();
duke@0 381 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
duke@0 382 "must stay in the original alias category");
duke@0 383 // The type of the address must be contained in the adr_type,
duke@0 384 // disregarding "null"-ness.
duke@0 385 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
duke@0 386 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
duke@0 387 assert(cross_check->meet(tp_notnull) == cross_check,
duke@0 388 "real address must not escape from expected memory type");
duke@0 389 }
duke@0 390 #endif
duke@0 391 return tp;
duke@0 392 }
duke@0 393 }
duke@0 394
duke@0 395 //------------------------adr_phi_is_loop_invariant----------------------------
duke@0 396 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
duke@0 397 // loop is loop invariant. Make a quick traversal of Phi and associated
duke@0 398 // CastPP nodes, looking to see if they are a closed group within the loop.
duke@0 399 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
duke@0 400 // The idea is that the phi-nest must boil down to only CastPP nodes
duke@0 401 // with the same data. This implies that any path into the loop already
duke@0 402 // includes such a CastPP, and so the original cast, whatever its input,
duke@0 403 // must be covered by an equivalent cast, with an earlier control input.
duke@0 404 ResourceMark rm;
duke@0 405
duke@0 406 // The loop entry input of the phi should be the unique dominating
duke@0 407 // node for every Phi/CastPP in the loop.
duke@0 408 Unique_Node_List closure;
duke@0 409 closure.push(adr_phi->in(LoopNode::EntryControl));
duke@0 410
duke@0 411 // Add the phi node and the cast to the worklist.
duke@0 412 Unique_Node_List worklist;
duke@0 413 worklist.push(adr_phi);
duke@0 414 if( cast != NULL ){
duke@0 415 if( !cast->is_ConstraintCast() ) return false;
duke@0 416 worklist.push(cast);
duke@0 417 }
duke@0 418
duke@0 419 // Begin recursive walk of phi nodes.
duke@0 420 while( worklist.size() ){
duke@0 421 // Take a node off the worklist
duke@0 422 Node *n = worklist.pop();
duke@0 423 if( !closure.member(n) ){
duke@0 424 // Add it to the closure.
duke@0 425 closure.push(n);
duke@0 426 // Make a sanity check to ensure we don't waste too much time here.
duke@0 427 if( closure.size() > 20) return false;
duke@0 428 // This node is OK if:
duke@0 429 // - it is a cast of an identical value
duke@0 430 // - or it is a phi node (then we add its inputs to the worklist)
duke@0 431 // Otherwise, the node is not OK, and we presume the cast is not invariant
duke@0 432 if( n->is_ConstraintCast() ){
duke@0 433 worklist.push(n->in(1));
duke@0 434 } else if( n->is_Phi() ) {
duke@0 435 for( uint i = 1; i < n->req(); i++ ) {
duke@0 436 worklist.push(n->in(i));
duke@0 437 }
duke@0 438 } else {
duke@0 439 return false;
duke@0 440 }
duke@0 441 }
duke@0 442 }
duke@0 443
duke@0 444 // Quit when the worklist is empty, and we've found no offending nodes.
duke@0 445 return true;
duke@0 446 }
duke@0 447
duke@0 448 //------------------------------Ideal_DU_postCCP-------------------------------
duke@0 449 // Find any cast-away of null-ness and keep its control. Null cast-aways are
duke@0 450 // going away in this pass and we need to make this memory op depend on the
duke@0 451 // gating null check.
duke@0 452
duke@0 453 // I tried to leave the CastPP's in. This makes the graph more accurate in
duke@0 454 // some sense; we get to keep around the knowledge that an oop is not-null
duke@0 455 // after some test. Alas, the CastPP's interfere with GVN (some values are
duke@0 456 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
duke@0 457 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
duke@0 458 // some of the more trivial cases in the optimizer. Removing more useless
duke@0 459 // Phi's started allowing Loads to illegally float above null checks. I gave
duke@0 460 // up on this approach. CNC 10/20/2000
duke@0 461 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
duke@0 462 Node *ctr = in(MemNode::Control);
duke@0 463 Node *mem = in(MemNode::Memory);
duke@0 464 Node *adr = in(MemNode::Address);
duke@0 465 Node *skipped_cast = NULL;
duke@0 466 // Need a null check? Regular static accesses do not because they are
duke@0 467 // from constant addresses. Array ops are gated by the range check (which
duke@0 468 // always includes a NULL check). Just check field ops.
duke@0 469 if( !ctr ) {
duke@0 470 // Scan upwards for the highest location we can place this memory op.
duke@0 471 while( true ) {
duke@0 472 switch( adr->Opcode() ) {
duke@0 473
duke@0 474 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
duke@0 475 adr = adr->in(AddPNode::Base);
duke@0 476 continue;
duke@0 477
duke@0 478 case Op_CastPP:
duke@0 479 // If the CastPP is useless, just peek on through it.
duke@0 480 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
duke@0 481 // Remember the cast that we've peeked though. If we peek
duke@0 482 // through more than one, then we end up remembering the highest
duke@0 483 // one, that is, if in a loop, the one closest to the top.
duke@0 484 skipped_cast = adr;
duke@0 485 adr = adr->in(1);
duke@0 486 continue;
duke@0 487 }
duke@0 488 // CastPP is going away in this pass! We need this memory op to be
duke@0 489 // control-dependent on the test that is guarding the CastPP.
duke@0 490 ccp->hash_delete(this);
duke@0 491 set_req(MemNode::Control, adr->in(0));
duke@0 492 ccp->hash_insert(this);
duke@0 493 return this;
duke@0 494
duke@0 495 case Op_Phi:
duke@0 496 // Attempt to float above a Phi to some dominating point.
duke@0 497 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
duke@0 498 // If we've already peeked through a Cast (which could have set the
duke@0 499 // control), we can't float above a Phi, because the skipped Cast
duke@0 500 // may not be loop invariant.
duke@0 501 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
duke@0 502 adr = adr->in(1);
duke@0 503 continue;
duke@0 504 }
duke@0 505 }
duke@0 506
duke@0 507 // Intentional fallthrough!
duke@0 508
duke@0 509 // No obvious dominating point. The mem op is pinned below the Phi
duke@0 510 // by the Phi itself. If the Phi goes away (no true value is merged)
duke@0 511 // then the mem op can float, but not indefinitely. It must be pinned
duke@0 512 // behind the controls leading to the Phi.
duke@0 513 case Op_CheckCastPP:
duke@0 514 // These usually stick around to change address type, however a
duke@0 515 // useless one can be elided and we still need to pick up a control edge
duke@0 516 if (adr->in(0) == NULL) {
duke@0 517 // This CheckCastPP node has NO control and is likely useless. But we
duke@0 518 // need check further up the ancestor chain for a control input to keep
duke@0 519 // the node in place. 4959717.
duke@0 520 skipped_cast = adr;
duke@0 521 adr = adr->in(1);
duke@0 522 continue;
duke@0 523 }
duke@0 524 ccp->hash_delete(this);
duke@0 525 set_req(MemNode::Control, adr->in(0));
duke@0 526 ccp->hash_insert(this);
duke@0 527 return this;
duke@0 528
duke@0 529 // List of "safe" opcodes; those that implicitly block the memory
duke@0 530 // op below any null check.
duke@0 531 case Op_CastX2P: // no null checks on native pointers
duke@0 532 case Op_Parm: // 'this' pointer is not null
duke@0 533 case Op_LoadP: // Loading from within a klass
duke@0 534 case Op_LoadKlass: // Loading from within a klass
duke@0 535 case Op_ConP: // Loading from a klass
duke@0 536 case Op_CreateEx: // Sucking up the guts of an exception oop
duke@0 537 case Op_Con: // Reading from TLS
duke@0 538 case Op_CMoveP: // CMoveP is pinned
duke@0 539 break; // No progress
duke@0 540
duke@0 541 case Op_Proj: // Direct call to an allocation routine
duke@0 542 case Op_SCMemProj: // Memory state from store conditional ops
duke@0 543 #ifdef ASSERT
duke@0 544 {
duke@0 545 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
duke@0 546 const Node* call = adr->in(0);
duke@0 547 if (call->is_CallStaticJava()) {
duke@0 548 const CallStaticJavaNode* call_java = call->as_CallStaticJava();
kvn@64 549 const TypeTuple *r = call_java->tf()->range();
kvn@64 550 assert(r->cnt() > TypeFunc::Parms, "must return value");
kvn@64 551 const Type* ret_type = r->field_at(TypeFunc::Parms);
kvn@64 552 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
duke@0 553 // We further presume that this is one of
duke@0 554 // new_instance_Java, new_array_Java, or
duke@0 555 // the like, but do not assert for this.
duke@0 556 } else if (call->is_Allocate()) {
duke@0 557 // similar case to new_instance_Java, etc.
duke@0 558 } else if (!call->is_CallLeaf()) {
duke@0 559 // Projections from fetch_oop (OSR) are allowed as well.
duke@0 560 ShouldNotReachHere();
duke@0 561 }
duke@0 562 }
duke@0 563 #endif
duke@0 564 break;
duke@0 565 default:
duke@0 566 ShouldNotReachHere();
duke@0 567 }
duke@0 568 break;
duke@0 569 }
duke@0 570 }
duke@0 571
duke@0 572 return NULL; // No progress
duke@0 573 }
duke@0 574
duke@0 575
duke@0 576 //=============================================================================
duke@0 577 uint LoadNode::size_of() const { return sizeof(*this); }
duke@0 578 uint LoadNode::cmp( const Node &n ) const
duke@0 579 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
duke@0 580 const Type *LoadNode::bottom_type() const { return _type; }
duke@0 581 uint LoadNode::ideal_reg() const {
duke@0 582 return Matcher::base2reg[_type->base()];
duke@0 583 }
duke@0 584
duke@0 585 #ifndef PRODUCT
duke@0 586 void LoadNode::dump_spec(outputStream *st) const {
duke@0 587 MemNode::dump_spec(st);
duke@0 588 if( !Verbose && !WizardMode ) {
duke@0 589 // standard dump does this in Verbose and WizardMode
duke@0 590 st->print(" #"); _type->dump_on(st);
duke@0 591 }
duke@0 592 }
duke@0 593 #endif
duke@0 594
duke@0 595
duke@0 596 //----------------------------LoadNode::make-----------------------------------
duke@0 597 // Polymorphic factory method:
duke@0 598 LoadNode *LoadNode::make( Compile *C, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
duke@0 599 // sanity check the alias category against the created node type
duke@0 600 assert(!(adr_type->isa_oopptr() &&
duke@0 601 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
duke@0 602 "use LoadKlassNode instead");
duke@0 603 assert(!(adr_type->isa_aryptr() &&
duke@0 604 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
duke@0 605 "use LoadRangeNode instead");
duke@0 606 switch (bt) {
duke@0 607 case T_BOOLEAN:
duke@0 608 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 609 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 610 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 611 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 612 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
duke@0 613 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
duke@0 614 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
duke@0 615 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
duke@0 616 case T_OBJECT: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
duke@0 617 }
duke@0 618 ShouldNotReachHere();
duke@0 619 return (LoadNode*)NULL;
duke@0 620 }
duke@0 621
duke@0 622 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
duke@0 623 bool require_atomic = true;
duke@0 624 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
duke@0 625 }
duke@0 626
duke@0 627
duke@0 628
duke@0 629
duke@0 630 //------------------------------hash-------------------------------------------
duke@0 631 uint LoadNode::hash() const {
duke@0 632 // unroll addition of interesting fields
duke@0 633 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
duke@0 634 }
duke@0 635
duke@0 636 //---------------------------can_see_stored_value------------------------------
duke@0 637 // This routine exists to make sure this set of tests is done the same
duke@0 638 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
duke@0 639 // will change the graph shape in a way which makes memory alive twice at the
duke@0 640 // same time (uses the Oracle model of aliasing), then some
duke@0 641 // LoadXNode::Identity will fold things back to the equivalence-class model
duke@0 642 // of aliasing.
duke@0 643 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
duke@0 644 Node* ld_adr = in(MemNode::Address);
duke@0 645
never@17 646 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
never@17 647 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
never@17 648 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
never@17 649 atp->field() != NULL && !atp->field()->is_volatile()) {
never@17 650 uint alias_idx = atp->index();
never@17 651 bool final = atp->field()->is_final();
never@17 652 Node* result = NULL;
never@17 653 Node* current = st;
never@17 654 // Skip through chains of MemBarNodes checking the MergeMems for
never@17 655 // new states for the slice of this load. Stop once any other
never@17 656 // kind of node is encountered. Loads from final memory can skip
never@17 657 // through any kind of MemBar but normal loads shouldn't skip
never@17 658 // through MemBarAcquire since the could allow them to move out of
never@17 659 // a synchronized region.
never@17 660 while (current->is_Proj()) {
never@17 661 int opc = current->in(0)->Opcode();
never@17 662 if ((final && opc == Op_MemBarAcquire) ||
never@17 663 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
never@17 664 Node* mem = current->in(0)->in(TypeFunc::Memory);
never@17 665 if (mem->is_MergeMem()) {
never@17 666 MergeMemNode* merge = mem->as_MergeMem();
never@17 667 Node* new_st = merge->memory_at(alias_idx);
never@17 668 if (new_st == merge->base_memory()) {
never@17 669 // Keep searching
never@17 670 current = merge->base_memory();
never@17 671 continue;
never@17 672 }
never@17 673 // Save the new memory state for the slice and fall through
never@17 674 // to exit.
never@17 675 result = new_st;
never@17 676 }
never@17 677 }
never@17 678 break;
never@17 679 }
never@17 680 if (result != NULL) {
never@17 681 st = result;
never@17 682 }
never@17 683 }
never@17 684
never@17 685
duke@0 686 // Loop around twice in the case Load -> Initialize -> Store.
duke@0 687 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
duke@0 688 for (int trip = 0; trip <= 1; trip++) {
duke@0 689
duke@0 690 if (st->is_Store()) {
duke@0 691 Node* st_adr = st->in(MemNode::Address);
duke@0 692 if (!phase->eqv(st_adr, ld_adr)) {
duke@0 693 // Try harder before giving up... Match raw and non-raw pointers.
duke@0 694 intptr_t st_off = 0;
duke@0 695 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
duke@0 696 if (alloc == NULL) return NULL;
duke@0 697 intptr_t ld_off = 0;
duke@0 698 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
duke@0 699 if (alloc != allo2) return NULL;
duke@0 700 if (ld_off != st_off) return NULL;
duke@0 701 // At this point we have proven something like this setup:
duke@0 702 // A = Allocate(...)
duke@0 703 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
duke@0 704 // S = StoreQ(, AddP(, A.Parm , #Off), V)
duke@0 705 // (Actually, we haven't yet proven the Q's are the same.)
duke@0 706 // In other words, we are loading from a casted version of
duke@0 707 // the same pointer-and-offset that we stored to.
duke@0 708 // Thus, we are able to replace L by V.
duke@0 709 }
duke@0 710 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
duke@0 711 if (store_Opcode() != st->Opcode())
duke@0 712 return NULL;
duke@0 713 return st->in(MemNode::ValueIn);
duke@0 714 }
duke@0 715
duke@0 716 intptr_t offset = 0; // scratch
duke@0 717
duke@0 718 // A load from a freshly-created object always returns zero.
duke@0 719 // (This can happen after LoadNode::Ideal resets the load's memory input
duke@0 720 // to find_captured_store, which returned InitializeNode::zero_memory.)
duke@0 721 if (st->is_Proj() && st->in(0)->is_Allocate() &&
duke@0 722 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
duke@0 723 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
duke@0 724 // return a zero value for the load's basic type
duke@0 725 // (This is one of the few places where a generic PhaseTransform
duke@0 726 // can create new nodes. Think of it as lazily manifesting
duke@0 727 // virtually pre-existing constants.)
duke@0 728 return phase->zerocon(memory_type());
duke@0 729 }
duke@0 730
duke@0 731 // A load from an initialization barrier can match a captured store.
duke@0 732 if (st->is_Proj() && st->in(0)->is_Initialize()) {
duke@0 733 InitializeNode* init = st->in(0)->as_Initialize();
duke@0 734 AllocateNode* alloc = init->allocation();
duke@0 735 if (alloc != NULL &&
duke@0 736 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
duke@0 737 // examine a captured store value
duke@0 738 st = init->find_captured_store(offset, memory_size(), phase);
duke@0 739 if (st != NULL)
duke@0 740 continue; // take one more trip around
duke@0 741 }
duke@0 742 }
duke@0 743
duke@0 744 break;
duke@0 745 }
duke@0 746
duke@0 747 return NULL;
duke@0 748 }
duke@0 749
kvn@64 750 //----------------------is_instance_field_load_with_local_phi------------------
kvn@64 751 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
kvn@64 752 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
kvn@64 753 in(MemNode::Address)->is_AddP() ) {
kvn@64 754 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
kvn@64 755 // Only instances.
kvn@64 756 if( t_oop != NULL && t_oop->is_instance_field() &&
kvn@64 757 t_oop->offset() != Type::OffsetBot &&
kvn@64 758 t_oop->offset() != Type::OffsetTop) {
kvn@64 759 return true;
kvn@64 760 }
kvn@64 761 }
kvn@64 762 return false;
kvn@64 763 }
kvn@64 764
duke@0 765 //------------------------------Identity---------------------------------------
duke@0 766 // Loads are identity if previous store is to same address
duke@0 767 Node *LoadNode::Identity( PhaseTransform *phase ) {
duke@0 768 // If the previous store-maker is the right kind of Store, and the store is
duke@0 769 // to the same address, then we are equal to the value stored.
duke@0 770 Node* mem = in(MemNode::Memory);
duke@0 771 Node* value = can_see_stored_value(mem, phase);
duke@0 772 if( value ) {
duke@0 773 // byte, short & char stores truncate naturally.
duke@0 774 // A load has to load the truncated value which requires
duke@0 775 // some sort of masking operation and that requires an
duke@0 776 // Ideal call instead of an Identity call.
duke@0 777 if (memory_size() < BytesPerInt) {
duke@0 778 // If the input to the store does not fit with the load's result type,
duke@0 779 // it must be truncated via an Ideal call.
duke@0 780 if (!phase->type(value)->higher_equal(phase->type(this)))
duke@0 781 return this;
duke@0 782 }
duke@0 783 // (This works even when value is a Con, but LoadNode::Value
duke@0 784 // usually runs first, producing the singleton type of the Con.)
duke@0 785 return value;
duke@0 786 }
kvn@64 787
kvn@64 788 // Search for an existing data phi which was generated before for the same
kvn@64 789 // instance's field to avoid infinite genertion of phis in a loop.
kvn@64 790 Node *region = mem->in(0);
kvn@64 791 if (is_instance_field_load_with_local_phi(region)) {
kvn@64 792 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
kvn@64 793 int this_index = phase->C->get_alias_index(addr_t);
kvn@64 794 int this_offset = addr_t->offset();
kvn@64 795 int this_id = addr_t->is_oopptr()->instance_id();
kvn@64 796 const Type* this_type = bottom_type();
kvn@64 797 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
kvn@64 798 Node* phi = region->fast_out(i);
kvn@64 799 if (phi->is_Phi() && phi != mem &&
kvn@64 800 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
kvn@64 801 return phi;
kvn@64 802 }
kvn@64 803 }
kvn@64 804 }
kvn@64 805
duke@0 806 return this;
duke@0 807 }
duke@0 808
never@17 809
never@17 810 // Returns true if the AliasType refers to the field that holds the
never@17 811 // cached box array. Currently only handles the IntegerCache case.
never@17 812 static bool is_autobox_cache(Compile::AliasType* atp) {
never@17 813 if (atp != NULL && atp->field() != NULL) {
never@17 814 ciField* field = atp->field();
never@17 815 ciSymbol* klass = field->holder()->name();
never@17 816 if (field->name() == ciSymbol::cache_field_name() &&
never@17 817 field->holder()->uses_default_loader() &&
never@17 818 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 819 return true;
never@17 820 }
never@17 821 }
never@17 822 return false;
never@17 823 }
never@17 824
never@17 825 // Fetch the base value in the autobox array
never@17 826 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
never@17 827 if (atp != NULL && atp->field() != NULL) {
never@17 828 ciField* field = atp->field();
never@17 829 ciSymbol* klass = field->holder()->name();
never@17 830 if (field->name() == ciSymbol::cache_field_name() &&
never@17 831 field->holder()->uses_default_loader() &&
never@17 832 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 833 assert(field->is_constant(), "what?");
never@17 834 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
never@17 835 // Fetch the box object at the base of the array and get its value
never@17 836 ciInstance* box = array->obj_at(0)->as_instance();
never@17 837 ciInstanceKlass* ik = box->klass()->as_instance_klass();
never@17 838 if (ik->nof_nonstatic_fields() == 1) {
never@17 839 // This should be true nonstatic_field_at requires calling
never@17 840 // nof_nonstatic_fields so check it anyway
never@17 841 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
never@17 842 cache_offset = c.as_int();
never@17 843 }
never@17 844 return true;
never@17 845 }
never@17 846 }
never@17 847 return false;
never@17 848 }
never@17 849
never@17 850 // Returns true if the AliasType refers to the value field of an
never@17 851 // autobox object. Currently only handles Integer.
never@17 852 static bool is_autobox_object(Compile::AliasType* atp) {
never@17 853 if (atp != NULL && atp->field() != NULL) {
never@17 854 ciField* field = atp->field();
never@17 855 ciSymbol* klass = field->holder()->name();
never@17 856 if (field->name() == ciSymbol::value_name() &&
never@17 857 field->holder()->uses_default_loader() &&
never@17 858 klass == ciSymbol::java_lang_Integer()) {
never@17 859 return true;
never@17 860 }
never@17 861 }
never@17 862 return false;
never@17 863 }
never@17 864
never@17 865
never@17 866 // We're loading from an object which has autobox behaviour.
never@17 867 // If this object is result of a valueOf call we'll have a phi
never@17 868 // merging a newly allocated object and a load from the cache.
never@17 869 // We want to replace this load with the original incoming
never@17 870 // argument to the valueOf call.
never@17 871 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
never@17 872 Node* base = in(Address)->in(AddPNode::Base);
never@17 873 if (base->is_Phi() && base->req() == 3) {
never@17 874 AllocateNode* allocation = NULL;
never@17 875 int allocation_index = -1;
never@17 876 int load_index = -1;
never@17 877 for (uint i = 1; i < base->req(); i++) {
never@17 878 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
never@17 879 if (allocation != NULL) {
never@17 880 allocation_index = i;
never@17 881 load_index = 3 - allocation_index;
never@17 882 break;
never@17 883 }
never@17 884 }
never@17 885 LoadNode* load = NULL;
never@17 886 if (allocation != NULL && base->in(load_index)->is_Load()) {
never@17 887 load = base->in(load_index)->as_Load();
never@17 888 }
never@17 889 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
never@17 890 // Push the loads from the phi that comes from valueOf up
never@17 891 // through it to allow elimination of the loads and the recovery
never@17 892 // of the original value.
never@17 893 Node* mem_phi = in(Memory);
never@17 894 Node* offset = in(Address)->in(AddPNode::Offset);
never@17 895
never@17 896 Node* in1 = clone();
never@17 897 Node* in1_addr = in1->in(Address)->clone();
never@17 898 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
never@17 899 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
never@17 900 in1_addr->set_req(AddPNode::Offset, offset);
never@17 901 in1->set_req(0, base->in(allocation_index));
never@17 902 in1->set_req(Address, in1_addr);
never@17 903 in1->set_req(Memory, mem_phi->in(allocation_index));
never@17 904
never@17 905 Node* in2 = clone();
never@17 906 Node* in2_addr = in2->in(Address)->clone();
never@17 907 in2_addr->set_req(AddPNode::Base, base->in(load_index));
never@17 908 in2_addr->set_req(AddPNode::Address, base->in(load_index));
never@17 909 in2_addr->set_req(AddPNode::Offset, offset);
never@17 910 in2->set_req(0, base->in(load_index));
never@17 911 in2->set_req(Address, in2_addr);
never@17 912 in2->set_req(Memory, mem_phi->in(load_index));
never@17 913
never@17 914 in1_addr = phase->transform(in1_addr);
never@17 915 in1 = phase->transform(in1);
never@17 916 in2_addr = phase->transform(in2_addr);
never@17 917 in2 = phase->transform(in2);
never@17 918
never@17 919 PhiNode* result = PhiNode::make_blank(base->in(0), this);
never@17 920 result->set_req(allocation_index, in1);
never@17 921 result->set_req(load_index, in2);
never@17 922 return result;
never@17 923 }
never@17 924 } else if (base->is_Load()) {
never@17 925 // Eliminate the load of Integer.value for integers from the cache
never@17 926 // array by deriving the value from the index into the array.
never@17 927 // Capture the offset of the load and then reverse the computation.
never@17 928 Node* load_base = base->in(Address)->in(AddPNode::Base);
never@17 929 if (load_base != NULL) {
never@17 930 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
never@17 931 intptr_t cache_offset;
never@17 932 int shift = -1;
never@17 933 Node* cache = NULL;
never@17 934 if (is_autobox_cache(atp)) {
kvn@29 935 shift = exact_log2(type2aelembytes(T_OBJECT));
never@17 936 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
never@17 937 }
never@17 938 if (cache != NULL && base->in(Address)->is_AddP()) {
never@17 939 Node* elements[4];
never@17 940 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
never@17 941 int cache_low;
never@17 942 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
never@17 943 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
never@17 944 // Add up all the offsets making of the address of the load
never@17 945 Node* result = elements[0];
never@17 946 for (int i = 1; i < count; i++) {
never@17 947 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
never@17 948 }
never@17 949 // Remove the constant offset from the address and then
never@17 950 // remove the scaling of the offset to recover the original index.
never@17 951 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
never@17 952 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
never@17 953 // Peel the shift off directly but wrap it in a dummy node
never@17 954 // since Ideal can't return existing nodes
never@17 955 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
never@17 956 } else {
never@17 957 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
never@17 958 }
never@17 959 #ifdef _LP64
never@17 960 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
never@17 961 #endif
never@17 962 return result;
never@17 963 }
never@17 964 }
never@17 965 }
never@17 966 }
never@17 967 return NULL;
never@17 968 }
never@17 969
never@17 970
duke@0 971 //------------------------------Ideal------------------------------------------
duke@0 972 // If the load is from Field memory and the pointer is non-null, we can
duke@0 973 // zero out the control input.
duke@0 974 // If the offset is constant and the base is an object allocation,
duke@0 975 // try to hook me up to the exact initializing store.
duke@0 976 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 977 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 978 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 979
duke@0 980 Node* ctrl = in(MemNode::Control);
duke@0 981 Node* address = in(MemNode::Address);
duke@0 982
duke@0 983 // Skip up past a SafePoint control. Cannot do this for Stores because
duke@0 984 // pointer stores & cardmarks must stay on the same side of a SafePoint.
duke@0 985 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
duke@0 986 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
duke@0 987 ctrl = ctrl->in(0);
duke@0 988 set_req(MemNode::Control,ctrl);
duke@0 989 }
duke@0 990
duke@0 991 // Check for useless control edge in some common special cases
duke@0 992 if (in(MemNode::Control) != NULL) {
duke@0 993 intptr_t ignore = 0;
duke@0 994 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
duke@0 995 if (base != NULL
duke@0 996 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
duke@0 997 && detect_dominating_control(base->in(0), phase->C->start())) {
duke@0 998 // A method-invariant, non-null address (constant or 'this' argument).
duke@0 999 set_req(MemNode::Control, NULL);
duke@0 1000 }
duke@0 1001 }
duke@0 1002
never@17 1003 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
never@17 1004 Node* base = in(Address)->in(AddPNode::Base);
never@17 1005 if (base != NULL) {
never@17 1006 Compile::AliasType* atp = phase->C->alias_type(adr_type());
never@17 1007 if (is_autobox_object(atp)) {
never@17 1008 Node* result = eliminate_autobox(phase);
never@17 1009 if (result != NULL) return result;
never@17 1010 }
never@17 1011 }
never@17 1012 }
never@17 1013
duke@0 1014 // Check for prior store with a different base or offset; make Load
duke@0 1015 // independent. Skip through any number of them. Bail out if the stores
duke@0 1016 // are in an endless dead cycle and report no progress. This is a key
duke@0 1017 // transform for Reflection. However, if after skipping through the Stores
duke@0 1018 // we can't then fold up against a prior store do NOT do the transform as
duke@0 1019 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
duke@0 1020 // array memory alive twice: once for the hoisted Load and again after the
duke@0 1021 // bypassed Store. This situation only works if EVERYBODY who does
duke@0 1022 // anti-dependence work knows how to bypass. I.e. we need all
duke@0 1023 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
duke@0 1024 // the alias index stuff. So instead, peek through Stores and IFF we can
duke@0 1025 // fold up, do so.
duke@0 1026 Node* prev_mem = find_previous_store(phase);
duke@0 1027 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 1028 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
duke@0 1029 // (c) See if we can fold up on the spot, but don't fold up here.
duke@0 1030 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
duke@0 1031 // just return a prior value, which is done by Identity calls.
duke@0 1032 if (can_see_stored_value(prev_mem, phase)) {
duke@0 1033 // Make ready for step (d):
duke@0 1034 set_req(MemNode::Memory, prev_mem);
duke@0 1035 return this;
duke@0 1036 }
duke@0 1037 }
duke@0 1038
duke@0 1039 return NULL; // No further progress
duke@0 1040 }
duke@0 1041
duke@0 1042 // Helper to recognize certain Klass fields which are invariant across
duke@0 1043 // some group of array types (e.g., int[] or all T[] where T < Object).
duke@0 1044 const Type*
duke@0 1045 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
duke@0 1046 ciKlass* klass) const {
duke@0 1047 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1048 // The field is Klass::_modifier_flags. Return its (constant) value.
duke@0 1049 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
duke@0 1050 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
duke@0 1051 return TypeInt::make(klass->modifier_flags());
duke@0 1052 }
duke@0 1053 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1054 // The field is Klass::_access_flags. Return its (constant) value.
duke@0 1055 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
duke@0 1056 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
duke@0 1057 return TypeInt::make(klass->access_flags());
duke@0 1058 }
duke@0 1059 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1060 // The field is Klass::_layout_helper. Return its constant value if known.
duke@0 1061 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1062 return TypeInt::make(klass->layout_helper());
duke@0 1063 }
duke@0 1064
duke@0 1065 // No match.
duke@0 1066 return NULL;
duke@0 1067 }
duke@0 1068
duke@0 1069 //------------------------------Value-----------------------------------------
duke@0 1070 const Type *LoadNode::Value( PhaseTransform *phase ) const {
duke@0 1071 // Either input is TOP ==> the result is TOP
duke@0 1072 Node* mem = in(MemNode::Memory);
duke@0 1073 const Type *t1 = phase->type(mem);
duke@0 1074 if (t1 == Type::TOP) return Type::TOP;
duke@0 1075 Node* adr = in(MemNode::Address);
duke@0 1076 const TypePtr* tp = phase->type(adr)->isa_ptr();
duke@0 1077 if (tp == NULL || tp->empty()) return Type::TOP;
duke@0 1078 int off = tp->offset();
duke@0 1079 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
duke@0 1080
duke@0 1081 // Try to guess loaded type from pointer type
duke@0 1082 if (tp->base() == Type::AryPtr) {
duke@0 1083 const Type *t = tp->is_aryptr()->elem();
duke@0 1084 // Don't do this for integer types. There is only potential profit if
duke@0 1085 // the element type t is lower than _type; that is, for int types, if _type is
duke@0 1086 // more restrictive than t. This only happens here if one is short and the other
duke@0 1087 // char (both 16 bits), and in those cases we've made an intentional decision
duke@0 1088 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
duke@0 1089 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
duke@0 1090 //
duke@0 1091 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
duke@0 1092 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
duke@0 1093 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
duke@0 1094 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
duke@0 1095 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
duke@0 1096 // In fact, that could have been the original type of p1, and p1 could have
duke@0 1097 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
duke@0 1098 // expression (LShiftL quux 3) independently optimized to the constant 8.
duke@0 1099 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
duke@0 1100 && Opcode() != Op_LoadKlass) {
duke@0 1101 // t might actually be lower than _type, if _type is a unique
duke@0 1102 // concrete subclass of abstract class t.
duke@0 1103 // Make sure the reference is not into the header, by comparing
duke@0 1104 // the offset against the offset of the start of the array's data.
duke@0 1105 // Different array types begin at slightly different offsets (12 vs. 16).
duke@0 1106 // We choose T_BYTE as an example base type that is least restrictive
duke@0 1107 // as to alignment, which will therefore produce the smallest
duke@0 1108 // possible base offset.
duke@0 1109 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
duke@0 1110 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
duke@0 1111 const Type* jt = t->join(_type);
duke@0 1112 // In any case, do not allow the join, per se, to empty out the type.
duke@0 1113 if (jt->empty() && !t->empty()) {
duke@0 1114 // This can happen if a interface-typed array narrows to a class type.
duke@0 1115 jt = _type;
duke@0 1116 }
never@17 1117
never@17 1118 if (EliminateAutoBox) {
never@17 1119 // The pointers in the autobox arrays are always non-null
never@17 1120 Node* base = in(Address)->in(AddPNode::Base);
never@17 1121 if (base != NULL) {
never@17 1122 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
never@17 1123 if (is_autobox_cache(atp)) {
never@17 1124 return jt->join(TypePtr::NOTNULL)->is_ptr();
never@17 1125 }
never@17 1126 }
never@17 1127 }
duke@0 1128 return jt;
duke@0 1129 }
duke@0 1130 }
duke@0 1131 } else if (tp->base() == Type::InstPtr) {
duke@0 1132 assert( off != Type::OffsetBot ||
duke@0 1133 // arrays can be cast to Objects
duke@0 1134 tp->is_oopptr()->klass()->is_java_lang_Object() ||
duke@0 1135 // unsafe field access may not have a constant offset
duke@0 1136 phase->C->has_unsafe_access(),
duke@0 1137 "Field accesses must be precise" );
duke@0 1138 // For oop loads, we expect the _type to be precise
duke@0 1139 } else if (tp->base() == Type::KlassPtr) {
duke@0 1140 assert( off != Type::OffsetBot ||
duke@0 1141 // arrays can be cast to Objects
duke@0 1142 tp->is_klassptr()->klass()->is_java_lang_Object() ||
duke@0 1143 // also allow array-loading from the primary supertype
duke@0 1144 // array during subtype checks
duke@0 1145 Opcode() == Op_LoadKlass,
duke@0 1146 "Field accesses must be precise" );
duke@0 1147 // For klass/static loads, we expect the _type to be precise
duke@0 1148 }
duke@0 1149
duke@0 1150 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1151 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1152 ciKlass* klass = tkls->klass();
duke@0 1153 if (klass->is_loaded() && tkls->klass_is_exact()) {
duke@0 1154 // We are loading a field from a Klass metaobject whose identity
duke@0 1155 // is known at compile time (the type is "exact" or "precise").
duke@0 1156 // Check for fields we know are maintained as constants by the VM.
duke@0 1157 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1158 // The field is Klass::_super_check_offset. Return its (constant) value.
duke@0 1159 // (Folds up type checking code.)
duke@0 1160 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
duke@0 1161 return TypeInt::make(klass->super_check_offset());
duke@0 1162 }
duke@0 1163 // Compute index into primary_supers array
duke@0 1164 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1165 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1166 if( depth < ciKlass::primary_super_limit() ) {
duke@0 1167 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1168 // (Folds up type checking code.)
duke@0 1169 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1170 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1171 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1172 }
duke@0 1173 const Type* aift = load_array_final_field(tkls, klass);
duke@0 1174 if (aift != NULL) return aift;
duke@0 1175 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
duke@0 1176 && klass->is_array_klass()) {
duke@0 1177 // The field is arrayKlass::_component_mirror. Return its (constant) value.
duke@0 1178 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
duke@0 1179 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
duke@0 1180 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
duke@0 1181 }
duke@0 1182 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1183 // The field is Klass::_java_mirror. Return its (constant) value.
duke@0 1184 // (Folds up the 2nd indirection in anObjConstant.getClass().)
duke@0 1185 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
duke@0 1186 return TypeInstPtr::make(klass->java_mirror());
duke@0 1187 }
duke@0 1188 }
duke@0 1189
duke@0 1190 // We can still check if we are loading from the primary_supers array at a
duke@0 1191 // shallow enough depth. Even though the klass is not exact, entries less
duke@0 1192 // than or equal to its super depth are correct.
duke@0 1193 if (klass->is_loaded() ) {
duke@0 1194 ciType *inner = klass->klass();
duke@0 1195 while( inner->is_obj_array_klass() )
duke@0 1196 inner = inner->as_obj_array_klass()->base_element_type();
duke@0 1197 if( inner->is_instance_klass() &&
duke@0 1198 !inner->as_instance_klass()->flags().is_interface() ) {
duke@0 1199 // Compute index into primary_supers array
duke@0 1200 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1201 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1202 if( depth < ciKlass::primary_super_limit() &&
duke@0 1203 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
duke@0 1204 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1205 // (Folds up type checking code.)
duke@0 1206 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1207 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1208 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1209 }
duke@0 1210 }
duke@0 1211 }
duke@0 1212
duke@0 1213 // If the type is enough to determine that the thing is not an array,
duke@0 1214 // we can give the layout_helper a positive interval type.
duke@0 1215 // This will help short-circuit some reflective code.
duke@0 1216 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
duke@0 1217 && !klass->is_array_klass() // not directly typed as an array
duke@0 1218 && !klass->is_interface() // specifically not Serializable & Cloneable
duke@0 1219 && !klass->is_java_lang_Object() // not the supertype of all T[]
duke@0 1220 ) {
duke@0 1221 // Note: When interfaces are reliable, we can narrow the interface
duke@0 1222 // test to (klass != Serializable && klass != Cloneable).
duke@0 1223 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1224 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
duke@0 1225 // The key property of this type is that it folds up tests
duke@0 1226 // for array-ness, since it proves that the layout_helper is positive.
duke@0 1227 // Thus, a generic value like the basic object layout helper works fine.
duke@0 1228 return TypeInt::make(min_size, max_jint, Type::WidenMin);
duke@0 1229 }
duke@0 1230 }
duke@0 1231
duke@0 1232 // If we are loading from a freshly-allocated object, produce a zero,
duke@0 1233 // if the load is provably beyond the header of the object.
duke@0 1234 // (Also allow a variable load from a fresh array to produce zero.)
duke@0 1235 if (ReduceFieldZeroing) {
duke@0 1236 Node* value = can_see_stored_value(mem,phase);
duke@0 1237 if (value != NULL && value->is_Con())
duke@0 1238 return value->bottom_type();
duke@0 1239 }
duke@0 1240
kvn@64 1241 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@64 1242 if (tinst != NULL && tinst->is_instance_field()) {
kvn@64 1243 // If we have an instance type and our memory input is the
kvn@64 1244 // programs's initial memory state, there is no matching store,
kvn@64 1245 // so just return a zero of the appropriate type
kvn@64 1246 Node *mem = in(MemNode::Memory);
kvn@64 1247 if (mem->is_Parm() && mem->in(0)->is_Start()) {
kvn@64 1248 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
kvn@64 1249 return Type::get_zero_type(_type->basic_type());
kvn@64 1250 }
kvn@64 1251 }
duke@0 1252 return _type;
duke@0 1253 }
duke@0 1254
duke@0 1255 //------------------------------match_edge-------------------------------------
duke@0 1256 // Do we Match on this edge index or not? Match only the address.
duke@0 1257 uint LoadNode::match_edge(uint idx) const {
duke@0 1258 return idx == MemNode::Address;
duke@0 1259 }
duke@0 1260
duke@0 1261 //--------------------------LoadBNode::Ideal--------------------------------------
duke@0 1262 //
duke@0 1263 // If the previous store is to the same address as this load,
duke@0 1264 // and the value stored was larger than a byte, replace this load
duke@0 1265 // with the value stored truncated to a byte. If no truncation is
duke@0 1266 // needed, the replacement is done in LoadNode::Identity().
duke@0 1267 //
duke@0 1268 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1269 Node* mem = in(MemNode::Memory);
duke@0 1270 Node* value = can_see_stored_value(mem,phase);
duke@0 1271 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1272 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
duke@0 1273 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
duke@0 1274 }
duke@0 1275 // Identity call will handle the case where truncation is not needed.
duke@0 1276 return LoadNode::Ideal(phase, can_reshape);
duke@0 1277 }
duke@0 1278
duke@0 1279 //--------------------------LoadCNode::Ideal--------------------------------------
duke@0 1280 //
duke@0 1281 // If the previous store is to the same address as this load,
duke@0 1282 // and the value stored was larger than a char, replace this load
duke@0 1283 // with the value stored truncated to a char. If no truncation is
duke@0 1284 // needed, the replacement is done in LoadNode::Identity().
duke@0 1285 //
duke@0 1286 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1287 Node* mem = in(MemNode::Memory);
duke@0 1288 Node* value = can_see_stored_value(mem,phase);
duke@0 1289 if( value && !phase->type(value)->higher_equal( _type ) )
duke@0 1290 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
duke@0 1291 // Identity call will handle the case where truncation is not needed.
duke@0 1292 return LoadNode::Ideal(phase, can_reshape);
duke@0 1293 }
duke@0 1294
duke@0 1295 //--------------------------LoadSNode::Ideal--------------------------------------
duke@0 1296 //
duke@0 1297 // If the previous store is to the same address as this load,
duke@0 1298 // and the value stored was larger than a short, replace this load
duke@0 1299 // with the value stored truncated to a short. If no truncation is
duke@0 1300 // needed, the replacement is done in LoadNode::Identity().
duke@0 1301 //
duke@0 1302 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1303 Node* mem = in(MemNode::Memory);
duke@0 1304 Node* value = can_see_stored_value(mem,phase);
duke@0 1305 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1306 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
duke@0 1307 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
duke@0 1308 }
duke@0 1309 // Identity call will handle the case where truncation is not needed.
duke@0 1310 return LoadNode::Ideal(phase, can_reshape);
duke@0 1311 }
duke@0 1312
duke@0 1313 //=============================================================================
duke@0 1314 //------------------------------Value------------------------------------------
duke@0 1315 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
duke@0 1316 // Either input is TOP ==> the result is TOP
duke@0 1317 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1318 if (t1 == Type::TOP) return Type::TOP;
duke@0 1319 Node *adr = in(MemNode::Address);
duke@0 1320 const Type *t2 = phase->type( adr );
duke@0 1321 if (t2 == Type::TOP) return Type::TOP;
duke@0 1322 const TypePtr *tp = t2->is_ptr();
duke@0 1323 if (TypePtr::above_centerline(tp->ptr()) ||
duke@0 1324 tp->ptr() == TypePtr::Null) return Type::TOP;
duke@0 1325
duke@0 1326 // Return a more precise klass, if possible
duke@0 1327 const TypeInstPtr *tinst = tp->isa_instptr();
duke@0 1328 if (tinst != NULL) {
duke@0 1329 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
duke@0 1330 int offset = tinst->offset();
duke@0 1331 if (ik == phase->C->env()->Class_klass()
duke@0 1332 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1333 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1334 // We are loading a special hidden field from a Class mirror object,
duke@0 1335 // the field which points to the VM's Klass metaobject.
duke@0 1336 ciType* t = tinst->java_mirror_type();
duke@0 1337 // java_mirror_type returns non-null for compile-time Class constants.
duke@0 1338 if (t != NULL) {
duke@0 1339 // constant oop => constant klass
duke@0 1340 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1341 return TypeKlassPtr::make(ciArrayKlass::make(t));
duke@0 1342 }
duke@0 1343 if (!t->is_klass()) {
duke@0 1344 // a primitive Class (e.g., int.class) has NULL for a klass field
duke@0 1345 return TypePtr::NULL_PTR;
duke@0 1346 }
duke@0 1347 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
duke@0 1348 return TypeKlassPtr::make(t->as_klass());
duke@0 1349 }
duke@0 1350 // non-constant mirror, so we can't tell what's going on
duke@0 1351 }
duke@0 1352 if( !ik->is_loaded() )
duke@0 1353 return _type; // Bail out if not loaded
duke@0 1354 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1355 if (tinst->klass_is_exact()) {
duke@0 1356 return TypeKlassPtr::make(ik);
duke@0 1357 }
duke@0 1358 // See if we can become precise: no subklasses and no interface
duke@0 1359 // (Note: We need to support verified interfaces.)
duke@0 1360 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1361 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1362 // Add a dependence; if any subclass added we need to recompile
duke@0 1363 if (!ik->is_final()) {
duke@0 1364 // %%% should use stronger assert_unique_concrete_subtype instead
duke@0 1365 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1366 }
duke@0 1367 // Return precise klass
duke@0 1368 return TypeKlassPtr::make(ik);
duke@0 1369 }
duke@0 1370
duke@0 1371 // Return root of possible klass
duke@0 1372 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
duke@0 1373 }
duke@0 1374 }
duke@0 1375
duke@0 1376 // Check for loading klass from an array
duke@0 1377 const TypeAryPtr *tary = tp->isa_aryptr();
duke@0 1378 if( tary != NULL ) {
duke@0 1379 ciKlass *tary_klass = tary->klass();
duke@0 1380 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
duke@0 1381 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
duke@0 1382 if (tary->klass_is_exact()) {
duke@0 1383 return TypeKlassPtr::make(tary_klass);
duke@0 1384 }
duke@0 1385 ciArrayKlass *ak = tary->klass()->as_array_klass();
duke@0 1386 // If the klass is an object array, we defer the question to the
duke@0 1387 // array component klass.
duke@0 1388 if( ak->is_obj_array_klass() ) {
duke@0 1389 assert( ak->is_loaded(), "" );
duke@0 1390 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
duke@0 1391 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
duke@0 1392 ciInstanceKlass* ik = base_k->as_instance_klass();
duke@0 1393 // See if we can become precise: no subklasses and no interface
duke@0 1394 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1395 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1396 // Add a dependence; if any subclass added we need to recompile
duke@0 1397 if (!ik->is_final()) {
duke@0 1398 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1399 }
duke@0 1400 // Return precise array klass
duke@0 1401 return TypeKlassPtr::make(ak);
duke@0 1402 }
duke@0 1403 }
duke@0 1404 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@0 1405 } else { // Found a type-array?
duke@0 1406 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1407 assert( ak->is_type_array_klass(), "" );
duke@0 1408 return TypeKlassPtr::make(ak); // These are always precise
duke@0 1409 }
duke@0 1410 }
duke@0 1411 }
duke@0 1412
duke@0 1413 // Check for loading klass from an array klass
duke@0 1414 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1415 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1416 ciKlass* klass = tkls->klass();
duke@0 1417 if( !klass->is_loaded() )
duke@0 1418 return _type; // Bail out if not loaded
duke@0 1419 if( klass->is_obj_array_klass() &&
duke@0 1420 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1421 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
duke@0 1422 // // Always returning precise element type is incorrect,
duke@0 1423 // // e.g., element type could be object and array may contain strings
duke@0 1424 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
duke@0 1425
duke@0 1426 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
duke@0 1427 // according to the element type's subclassing.
duke@0 1428 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
duke@0 1429 }
duke@0 1430 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
duke@0 1431 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1432 ciKlass* sup = klass->as_instance_klass()->super();
duke@0 1433 // The field is Klass::_super. Return its (constant) value.
duke@0 1434 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
duke@0 1435 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
duke@0 1436 }
duke@0 1437 }
duke@0 1438
duke@0 1439 // Bailout case
duke@0 1440 return LoadNode::Value(phase);
duke@0 1441 }
duke@0 1442
duke@0 1443 //------------------------------Identity---------------------------------------
duke@0 1444 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
duke@0 1445 // Also feed through the klass in Allocate(...klass...)._klass.
duke@0 1446 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
duke@0 1447 Node* x = LoadNode::Identity(phase);
duke@0 1448 if (x != this) return x;
duke@0 1449
duke@0 1450 // Take apart the address into an oop and and offset.
duke@0 1451 // Return 'this' if we cannot.
duke@0 1452 Node* adr = in(MemNode::Address);
duke@0 1453 intptr_t offset = 0;
duke@0 1454 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1455 if (base == NULL) return this;
duke@0 1456 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
duke@0 1457 if (toop == NULL) return this;
duke@0 1458
duke@0 1459 // We can fetch the klass directly through an AllocateNode.
duke@0 1460 // This works even if the klass is not constant (clone or newArray).
duke@0 1461 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1462 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
duke@0 1463 if (allocated_klass != NULL) {
duke@0 1464 return allocated_klass;
duke@0 1465 }
duke@0 1466 }
duke@0 1467
duke@0 1468 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
duke@0 1469 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
duke@0 1470 // See inline_native_Class_query for occurrences of these patterns.
duke@0 1471 // Java Example: x.getClass().isAssignableFrom(y)
duke@0 1472 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
duke@0 1473 //
duke@0 1474 // This improves reflective code, often making the Class
duke@0 1475 // mirror go completely dead. (Current exception: Class
duke@0 1476 // mirrors may appear in debug info, but we could clean them out by
duke@0 1477 // introducing a new debug info operator for klassOop.java_mirror).
duke@0 1478 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
duke@0 1479 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1480 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1481 // We are loading a special hidden field from a Class mirror,
duke@0 1482 // the field which points to its Klass or arrayKlass metaobject.
duke@0 1483 if (base->is_Load()) {
duke@0 1484 Node* adr2 = base->in(MemNode::Address);
duke@0 1485 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
duke@0 1486 if (tkls != NULL && !tkls->empty()
duke@0 1487 && (tkls->klass()->is_instance_klass() ||
duke@0 1488 tkls->klass()->is_array_klass())
duke@0 1489 && adr2->is_AddP()
duke@0 1490 ) {
duke@0 1491 int mirror_field = Klass::java_mirror_offset_in_bytes();
duke@0 1492 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1493 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
duke@0 1494 }
duke@0 1495 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
duke@0 1496 return adr2->in(AddPNode::Base);
duke@0 1497 }
duke@0 1498 }
duke@0 1499 }
duke@0 1500 }
duke@0 1501
duke@0 1502 return this;
duke@0 1503 }
duke@0 1504
duke@0 1505 //------------------------------Value-----------------------------------------
duke@0 1506 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
duke@0 1507 // Either input is TOP ==> the result is TOP
duke@0 1508 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1509 if( t1 == Type::TOP ) return Type::TOP;
duke@0 1510 Node *adr = in(MemNode::Address);
duke@0 1511 const Type *t2 = phase->type( adr );
duke@0 1512 if( t2 == Type::TOP ) return Type::TOP;
duke@0 1513 const TypePtr *tp = t2->is_ptr();
duke@0 1514 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
duke@0 1515 const TypeAryPtr *tap = tp->isa_aryptr();
duke@0 1516 if( !tap ) return _type;
duke@0 1517 return tap->size();
duke@0 1518 }
duke@0 1519
duke@0 1520 //------------------------------Identity---------------------------------------
duke@0 1521 // Feed through the length in AllocateArray(...length...)._length.
duke@0 1522 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
duke@0 1523 Node* x = LoadINode::Identity(phase);
duke@0 1524 if (x != this) return x;
duke@0 1525
duke@0 1526 // Take apart the address into an oop and and offset.
duke@0 1527 // Return 'this' if we cannot.
duke@0 1528 Node* adr = in(MemNode::Address);
duke@0 1529 intptr_t offset = 0;
duke@0 1530 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1531 if (base == NULL) return this;
duke@0 1532 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
duke@0 1533 if (tary == NULL) return this;
duke@0 1534
duke@0 1535 // We can fetch the length directly through an AllocateArrayNode.
duke@0 1536 // This works even if the length is not constant (clone or newArray).
duke@0 1537 if (offset == arrayOopDesc::length_offset_in_bytes()) {
duke@0 1538 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
duke@0 1539 if (allocated_length != NULL) {
duke@0 1540 return allocated_length;
duke@0 1541 }
duke@0 1542 }
duke@0 1543
duke@0 1544 return this;
duke@0 1545
duke@0 1546 }
duke@0 1547 //=============================================================================
duke@0 1548 //---------------------------StoreNode::make-----------------------------------
duke@0 1549 // Polymorphic factory method:
duke@0 1550 StoreNode* StoreNode::make( Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
duke@0 1551 switch (bt) {
duke@0 1552 case T_BOOLEAN:
duke@0 1553 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
duke@0 1554 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
duke@0 1555 case T_CHAR:
duke@0 1556 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
duke@0 1557 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
duke@0 1558 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
duke@0 1559 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
duke@0 1560 case T_ADDRESS:
duke@0 1561 case T_OBJECT: return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
duke@0 1562 }
duke@0 1563 ShouldNotReachHere();
duke@0 1564 return (StoreNode*)NULL;
duke@0 1565 }
duke@0 1566
duke@0 1567 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
duke@0 1568 bool require_atomic = true;
duke@0 1569 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
duke@0 1570 }
duke@0 1571
duke@0 1572
duke@0 1573 //--------------------------bottom_type----------------------------------------
duke@0 1574 const Type *StoreNode::bottom_type() const {
duke@0 1575 return Type::MEMORY;
duke@0 1576 }
duke@0 1577
duke@0 1578 //------------------------------hash-------------------------------------------
duke@0 1579 uint StoreNode::hash() const {
duke@0 1580 // unroll addition of interesting fields
duke@0 1581 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
duke@0 1582
duke@0 1583 // Since they are not commoned, do not hash them:
duke@0 1584 return NO_HASH;
duke@0 1585 }
duke@0 1586
duke@0 1587 //------------------------------Ideal------------------------------------------
duke@0 1588 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
duke@0 1589 // When a store immediately follows a relevant allocation/initialization,
duke@0 1590 // try to capture it into the initialization, or hoist it above.
duke@0 1591 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1592 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 1593 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 1594
duke@0 1595 Node* mem = in(MemNode::Memory);
duke@0 1596 Node* address = in(MemNode::Address);
duke@0 1597
duke@0 1598 // Back-to-back stores to same address? Fold em up.
duke@0 1599 // Generally unsafe if I have intervening uses...
duke@0 1600 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
duke@0 1601 // Looking at a dead closed cycle of memory?
duke@0 1602 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
duke@0 1603
duke@0 1604 assert(Opcode() == mem->Opcode() ||
duke@0 1605 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
duke@0 1606 "no mismatched stores, except on raw memory");
duke@0 1607
duke@0 1608 if (mem->outcnt() == 1 && // check for intervening uses
duke@0 1609 mem->as_Store()->memory_size() <= this->memory_size()) {
duke@0 1610 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
duke@0 1611 // For example, 'mem' might be the final state at a conditional return.
duke@0 1612 // Or, 'mem' might be used by some node which is live at the same time
duke@0 1613 // 'this' is live, which might be unschedulable. So, require exactly
duke@0 1614 // ONE user, the 'this' store, until such time as we clone 'mem' for
duke@0 1615 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
duke@0 1616 if (can_reshape) { // (%%% is this an anachronism?)
duke@0 1617 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
duke@0 1618 phase->is_IterGVN());
duke@0 1619 } else {
duke@0 1620 // It's OK to do this in the parser, since DU info is always accurate,
duke@0 1621 // and the parser always refers to nodes via SafePointNode maps.
duke@0 1622 set_req(MemNode::Memory, mem->in(MemNode::Memory));
duke@0 1623 }
duke@0 1624 return this;
duke@0 1625 }
duke@0 1626 }
duke@0 1627
duke@0 1628 // Capture an unaliased, unconditional, simple store into an initializer.
duke@0 1629 // Or, if it is independent of the allocation, hoist it above the allocation.
duke@0 1630 if (ReduceFieldZeroing && /*can_reshape &&*/
duke@0 1631 mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 1632 InitializeNode* init = mem->in(0)->as_Initialize();
duke@0 1633 intptr_t offset = init->can_capture_store(this, phase);
duke@0 1634 if (offset > 0) {
duke@0 1635 Node* moved = init->capture_store(this, offset, phase);
duke@0 1636 // If the InitializeNode captured me, it made a raw copy of me,
duke@0 1637 // and I need to disappear.
duke@0 1638 if (moved != NULL) {
duke@0 1639 // %%% hack to ensure that Ideal returns a new node:
duke@0 1640 mem = MergeMemNode::make(phase->C, mem);
duke@0 1641 return mem; // fold me away
duke@0 1642 }
duke@0 1643 }
duke@0 1644 }
duke@0 1645
duke@0 1646 return NULL; // No further progress
duke@0 1647 }
duke@0 1648
duke@0 1649 //------------------------------Value-----------------------------------------
duke@0 1650 const Type *StoreNode::Value( PhaseTransform *phase ) const {
duke@0 1651 // Either input is TOP ==> the result is TOP
duke@0 1652 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1653 if( t1 == Type::TOP ) return Type::TOP;
duke@0 1654 const Type *t2 = phase->type( in(MemNode::Address) );
duke@0 1655 if( t2 == Type::TOP ) return Type::TOP;
duke@0 1656 const Type *t3 = phase->type( in(MemNode::ValueIn) );
duke@0 1657 if( t3 == Type::TOP ) return Type::TOP;
duke@0 1658 return Type::MEMORY;
duke@0 1659 }
duke@0 1660
duke@0 1661 //------------------------------Identity---------------------------------------
duke@0 1662 // Remove redundant stores:
duke@0 1663 // Store(m, p, Load(m, p)) changes to m.
duke@0 1664 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
duke@0 1665 Node *StoreNode::Identity( PhaseTransform *phase ) {
duke@0 1666 Node* mem = in(MemNode::Memory);
duke@0 1667 Node* adr = in(MemNode::Address);
duke@0 1668 Node* val = in(MemNode::ValueIn);
duke@0 1669
duke@0 1670 // Load then Store? Then the Store is useless
duke@0 1671 if (val->is_Load() &&
duke@0 1672 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
duke@0 1673 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
duke@0 1674 val->as_Load()->store_Opcode() == Opcode()) {
duke@0 1675 return mem;
duke@0 1676 }
duke@0 1677
duke@0 1678 // Two stores in a row of the same value?
duke@0 1679 if (mem->is_Store() &&
duke@0 1680 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
duke@0 1681 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
duke@0 1682 mem->Opcode() == Opcode()) {
duke@0 1683 return mem;
duke@0 1684 }
duke@0 1685
duke@0 1686 // Store of zero anywhere into a freshly-allocated object?
duke@0 1687 // Then the store is useless.
duke@0 1688 // (It must already have been captured by the InitializeNode.)
duke@0 1689 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
duke@0 1690 // a newly allocated object is already all-zeroes everywhere
duke@0 1691 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
duke@0 1692 return mem;
duke@0 1693 }
duke@0 1694
duke@0 1695 // the store may also apply to zero-bits in an earlier object
duke@0 1696 Node* prev_mem = find_previous_store(phase);
duke@0 1697 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 1698 if (prev_mem != NULL) {
duke@0 1699 Node* prev_val = can_see_stored_value(prev_mem, phase);
duke@0 1700 if (prev_val != NULL && phase->eqv(prev_val, val)) {
duke@0 1701 // prev_val and val might differ by a cast; it would be good
duke@0 1702 // to keep the more informative of the two.
duke@0 1703 return mem;
duke@0 1704 }
duke@0 1705 }
duke@0 1706 }
duke@0 1707
duke@0 1708 return this;
duke@0 1709 }
duke@0 1710
duke@0 1711 //------------------------------match_edge-------------------------------------
duke@0 1712 // Do we Match on this edge index or not? Match only memory & value
duke@0 1713 uint StoreNode::match_edge(uint idx) const {
duke@0 1714 return idx == MemNode::Address || idx == MemNode::ValueIn;
duke@0 1715 }
duke@0 1716
duke@0 1717 //------------------------------cmp--------------------------------------------
duke@0 1718 // Do not common stores up together. They generally have to be split
duke@0 1719 // back up anyways, so do not bother.
duke@0 1720 uint StoreNode::cmp( const Node &n ) const {
duke@0 1721 return (&n == this); // Always fail except on self
duke@0 1722 }
duke@0 1723
duke@0 1724 //------------------------------Ideal_masked_input-----------------------------
duke@0 1725 // Check for a useless mask before a partial-word store
duke@0 1726 // (StoreB ... (AndI valIn conIa) )
duke@0 1727 // If (conIa & mask == mask) this simplifies to
duke@0 1728 // (StoreB ... (valIn) )
duke@0 1729 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
duke@0 1730 Node *val = in(MemNode::ValueIn);
duke@0 1731 if( val->Opcode() == Op_AndI ) {
duke@0 1732 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 1733 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
duke@0 1734 set_req(MemNode::ValueIn, val->in(1));
duke@0 1735 return this;
duke@0 1736 }
duke@0 1737 }
duke@0 1738 return NULL;
duke@0 1739 }
duke@0 1740
duke@0 1741
duke@0 1742 //------------------------------Ideal_sign_extended_input----------------------
duke@0 1743 // Check for useless sign-extension before a partial-word store
duke@0 1744 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
duke@0 1745 // If (conIL == conIR && conIR <= num_bits) this simplifies to
duke@0 1746 // (StoreB ... (valIn) )
duke@0 1747 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
duke@0 1748 Node *val = in(MemNode::ValueIn);
duke@0 1749 if( val->Opcode() == Op_RShiftI ) {
duke@0 1750 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 1751 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
duke@0 1752 Node *shl = val->in(1);
duke@0 1753 if( shl->Opcode() == Op_LShiftI ) {
duke@0 1754 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
duke@0 1755 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
duke@0 1756 set_req(MemNode::ValueIn, shl->in(1));
duke@0 1757 return this;
duke@0 1758 }
duke@0 1759 }
duke@0 1760 }
duke@0 1761 }
duke@0 1762 return NULL;
duke@0 1763 }
duke@0 1764
duke@0 1765 //------------------------------value_never_loaded-----------------------------------
duke@0 1766 // Determine whether there are any possible loads of the value stored.
duke@0 1767 // For simplicity, we actually check if there are any loads from the
duke@0 1768 // address stored to, not just for loads of the value stored by this node.
duke@0 1769 //
duke@0 1770 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
duke@0 1771 Node *adr = in(Address);
duke@0 1772 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
duke@0 1773 if (adr_oop == NULL)
duke@0 1774 return false;
kvn@64 1775 if (!adr_oop->is_instance_field())
duke@0 1776 return false; // if not a distinct instance, there may be aliases of the address
duke@0 1777 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
duke@0 1778 Node *use = adr->fast_out(i);
duke@0 1779 int opc = use->Opcode();
duke@0 1780 if (use->is_Load() || use->is_LoadStore()) {
duke@0 1781 return false;
duke@0 1782 }
duke@0 1783 }
duke@0 1784 return true;
duke@0 1785 }
duke@0 1786
duke@0 1787 //=============================================================================
duke@0 1788 //------------------------------Ideal------------------------------------------
duke@0 1789 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 1790 // we can skip the AND operation. If the store is from a sign-extension
duke@0 1791 // (a left shift, then right shift) we can skip both.
duke@0 1792 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 1793 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
duke@0 1794 if( progress != NULL ) return progress;
duke@0 1795
duke@0 1796 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
duke@0 1797 if( progress != NULL ) return progress;
duke@0 1798
duke@0 1799 // Finally check the default case
duke@0 1800 return StoreNode::Ideal(phase, can_reshape);
duke@0 1801 }
duke@0 1802
duke@0 1803 //=============================================================================
duke@0 1804 //------------------------------Ideal------------------------------------------
duke@0 1805 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 1806 // we can skip the AND operation
duke@0 1807 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 1808 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
duke@0 1809 if( progress != NULL ) return progress;
duke@0 1810
duke@0 1811 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
duke@0 1812 if( progress != NULL ) return progress;
duke@0 1813
duke@0 1814 // Finally check the default case
duke@0 1815 return StoreNode::Ideal(phase, can_reshape);
duke@0 1816 }
duke@0 1817
duke@0 1818 //=============================================================================
duke@0 1819 //------------------------------Identity---------------------------------------
duke@0 1820 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
duke@0 1821 // No need to card mark when storing a null ptr
duke@0 1822 Node* my_store = in(MemNode::OopStore);
duke@0 1823 if (my_store->is_Store()) {
duke@0 1824 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
duke@0 1825 if( t1 == TypePtr::NULL_PTR ) {
duke@0 1826 return in(MemNode::Memory);
duke@0 1827 }
duke@0 1828 }
duke@0 1829 return this;
duke@0 1830 }
duke@0 1831
duke@0 1832 //------------------------------Value-----------------------------------------
duke@0 1833 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
kvn@43 1834 // Either input is TOP ==> the result is TOP
kvn@43 1835 const Type *t = phase->type( in(MemNode::Memory) );
kvn@43 1836 if( t == Type::TOP ) return Type::TOP;
kvn@43 1837 t = phase->type( in(MemNode::Address) );
kvn@43 1838 if( t == Type::TOP ) return Type::TOP;
kvn@43 1839 t = phase->type( in(MemNode::ValueIn) );
kvn@43 1840 if( t == Type::TOP ) return Type::TOP;
duke@0 1841 // If extra input is TOP ==> the result is TOP
kvn@43 1842 t = phase->type( in(MemNode::OopStore) );
kvn@43 1843 if( t == Type::TOP ) return Type::TOP;
duke@0 1844
duke@0 1845 return StoreNode::Value( phase );
duke@0 1846 }
duke@0 1847
duke@0 1848
duke@0 1849 //=============================================================================
duke@0 1850 //----------------------------------SCMemProjNode------------------------------
duke@0 1851 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
duke@0 1852 {
duke@0 1853 return bottom_type();
duke@0 1854 }
duke@0 1855
duke@0 1856 //=============================================================================
duke@0 1857 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
duke@0 1858 init_req(MemNode::Control, c );
duke@0 1859 init_req(MemNode::Memory , mem);
duke@0 1860 init_req(MemNode::Address, adr);
duke@0 1861 init_req(MemNode::ValueIn, val);
duke@0 1862 init_req( ExpectedIn, ex );
duke@0 1863 init_class_id(Class_LoadStore);
duke@0 1864
duke@0 1865 }
duke@0 1866
duke@0 1867 //=============================================================================
duke@0 1868 //-------------------------------adr_type--------------------------------------
duke@0 1869 // Do we Match on this edge index or not? Do not match memory
duke@0 1870 const TypePtr* ClearArrayNode::adr_type() const {
duke@0 1871 Node *adr = in(3);
duke@0 1872 return MemNode::calculate_adr_type(adr->bottom_type());
duke@0 1873 }
duke@0 1874
duke@0 1875 //------------------------------match_edge-------------------------------------
duke@0 1876 // Do we Match on this edge index or not? Do not match memory
duke@0 1877 uint ClearArrayNode::match_edge(uint idx) const {
duke@0 1878 return idx > 1;
duke@0 1879 }
duke@0 1880
duke@0 1881 //------------------------------Identity---------------------------------------
duke@0 1882 // Clearing a zero length array does nothing
duke@0 1883 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
never@68 1884 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
duke@0 1885 }
duke@0 1886
duke@0 1887 //------------------------------Idealize---------------------------------------
duke@0 1888 // Clearing a short array is faster with stores
duke@0 1889 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 1890 const int unit = BytesPerLong;
duke@0 1891 const TypeX* t = phase->type(in(2))->isa_intptr_t();
duke@0 1892 if (!t) return NULL;
duke@0 1893 if (!t->is_con()) return NULL;
duke@0 1894 intptr_t raw_count = t->get_con();
duke@0 1895 intptr_t size = raw_count;
duke@0 1896 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
duke@0 1897 // Clearing nothing uses the Identity call.
duke@0 1898 // Negative clears are possible on dead ClearArrays
duke@0 1899 // (see jck test stmt114.stmt11402.val).
duke@0 1900 if (size <= 0 || size % unit != 0) return NULL;
duke@0 1901 intptr_t count = size / unit;
duke@0 1902 // Length too long; use fast hardware clear
duke@0 1903 if (size > Matcher::init_array_short_size) return NULL;
duke@0 1904 Node *mem = in(1);
duke@0 1905 if( phase->type(mem)==Type::TOP ) return NULL;
duke@0 1906 Node *adr = in(3);
duke@0 1907 const Type* at = phase->type(adr);
duke@0 1908 if( at==Type::TOP ) return NULL;
duke@0 1909 const TypePtr* atp = at->isa_ptr();
duke@0 1910 // adjust atp to be the correct array element address type
duke@0 1911 if (atp == NULL) atp = TypePtr::BOTTOM;
duke@0 1912 else atp = atp->add_offset(Type::OffsetBot);
duke@0 1913 // Get base for derived pointer purposes
duke@0 1914 if( adr->Opcode() != Op_AddP ) Unimplemented();
duke@0 1915 Node *base = adr->in(1);
duke@0 1916
duke@0 1917 Node *zero = phase->makecon(TypeLong::ZERO);
duke@0 1918 Node *off = phase->MakeConX(BytesPerLong);
duke@0 1919 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 1920 count--;
duke@0 1921 while( count-- ) {
duke@0 1922 mem = phase->transform(mem);
duke@0 1923 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
duke@0 1924 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 1925 }
duke@0 1926 return mem;
duke@0 1927 }
duke@0 1928
duke@0 1929 //----------------------------clear_memory-------------------------------------
duke@0 1930 // Generate code to initialize object storage to zero.
duke@0 1931 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 1932 intptr_t start_offset,
duke@0 1933 Node* end_offset,
duke@0 1934 PhaseGVN* phase) {
duke@0 1935 Compile* C = phase->C;
duke@0 1936 intptr_t offset = start_offset;
duke@0 1937
duke@0 1938 int unit = BytesPerLong;
duke@0 1939 if ((offset % unit) != 0) {
duke@0 1940 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
duke@0 1941 adr = phase->transform(adr);
duke@0 1942 const TypePtr* atp = TypeRawPtr::BOTTOM;
duke@0 1943 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 1944 mem = phase->transform(mem);
duke@0 1945 offset += BytesPerInt;
duke@0 1946 }
duke@0 1947 assert((offset % unit) == 0, "");
duke@0 1948
duke@0 1949 // Initialize the remaining stuff, if any, with a ClearArray.
duke@0 1950 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
duke@0 1951 }
duke@0 1952
duke@0 1953 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 1954 Node* start_offset,
duke@0 1955 Node* end_offset,
duke@0 1956 PhaseGVN* phase) {
never@68 1957 if (start_offset == end_offset) {
never@68 1958 // nothing to do
never@68 1959 return mem;
never@68 1960 }
never@68 1961
duke@0 1962 Compile* C = phase->C;
duke@0 1963 int unit = BytesPerLong;
duke@0 1964 Node* zbase = start_offset;
duke@0 1965 Node* zend = end_offset;
duke@0 1966
duke@0 1967 // Scale to the unit required by the CPU:
duke@0 1968 if (!Matcher::init_array_count_is_in_bytes) {
duke@0 1969 Node* shift = phase->intcon(exact_log2(unit));
duke@0 1970 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
duke@0 1971 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
duke@0 1972 }
duke@0 1973
duke@0 1974 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
duke@0 1975 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
duke@0 1976
duke@0 1977 // Bulk clear double-words
duke@0 1978 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
duke@0 1979 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
duke@0 1980 return phase->transform(mem);
duke@0 1981 }
duke@0 1982
duke@0 1983 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 1984 intptr_t start_offset,
duke@0 1985 intptr_t end_offset,
duke@0 1986 PhaseGVN* phase) {
never@68 1987 if (start_offset == end_offset) {
never@68 1988 // nothing to do
never@68 1989 return mem;
never@68 1990 }
never@68 1991
duke@0 1992 Compile* C = phase->C;
duke@0 1993 assert((end_offset % BytesPerInt) == 0, "odd end offset");
duke@0 1994 intptr_t done_offset = end_offset;
duke@0 1995 if ((done_offset % BytesPerLong) != 0) {
duke@0 1996 done_offset -= BytesPerInt;
duke@0 1997 }
duke@0 1998 if (done_offset > start_offset) {
duke@0 1999 mem = clear_memory(ctl, mem, dest,
duke@0 2000 start_offset, phase->MakeConX(done_offset), phase);
duke@0 2001 }
duke@0 2002 if (done_offset < end_offset) { // emit the final 32-bit store
duke@0 2003 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
duke@0 2004 adr = phase->transform(adr);
duke@0 2005 const TypePtr* atp = TypeRawPtr::BOTTOM;
duke@0 2006 mem = StoreNode::make(C, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2007 mem = phase->transform(mem);
duke@0 2008 done_offset += BytesPerInt;
duke@0 2009 }
duke@0 2010 assert(done_offset == end_offset, "");
duke@0 2011 return mem;
duke@0 2012 }
duke@0 2013
duke@0 2014 //=============================================================================
duke@0 2015 // Do we match on this edge? No memory edges
duke@0 2016 uint StrCompNode::match_edge(uint idx) const {
duke@0 2017 return idx == 5 || idx == 6;
duke@0 2018 }
duke@0 2019
duke@0 2020 //------------------------------Ideal------------------------------------------
duke@0 2021 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2022 // control copies
duke@0 2023 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2024 return remove_dead_region(phase, can_reshape) ? this : NULL;
duke@0 2025 }
duke@0 2026
duke@0 2027
duke@0 2028 //=============================================================================
duke@0 2029 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
duke@0 2030 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
duke@0 2031 _adr_type(C->get_adr_type(alias_idx))
duke@0 2032 {
duke@0 2033 init_class_id(Class_MemBar);
duke@0 2034 Node* top = C->top();
duke@0 2035 init_req(TypeFunc::I_O,top);
duke@0 2036 init_req(TypeFunc::FramePtr,top);
duke@0 2037 init_req(TypeFunc::ReturnAdr,top);
duke@0 2038 if (precedent != NULL)
duke@0 2039 init_req(TypeFunc::Parms, precedent);
duke@0 2040 }
duke@0 2041
duke@0 2042 //------------------------------cmp--------------------------------------------
duke@0 2043 uint MemBarNode::hash() const { return NO_HASH; }
duke@0 2044 uint MemBarNode::cmp( const Node &n ) const {
duke@0 2045 return (&n == this); // Always fail except on self
duke@0 2046 }
duke@0 2047
duke@0 2048 //------------------------------make-------------------------------------------
duke@0 2049 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
duke@0 2050 int len = Precedent + (pn == NULL? 0: 1);
duke@0 2051 switch (opcode) {
duke@0 2052 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
duke@0 2053 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
duke@0 2054 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
duke@0 2055 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
duke@0 2056 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
duke@0 2057 default: ShouldNotReachHere(); return NULL;
duke@0 2058 }
duke@0 2059 }
duke@0 2060
duke@0 2061 //------------------------------Ideal------------------------------------------
duke@0 2062 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2063 // control copies
duke@0 2064 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 2065 if (remove_dead_region(phase, can_reshape)) return this;
duke@0 2066 return NULL;
duke@0 2067 }
duke@0 2068
duke@0 2069 //------------------------------Value------------------------------------------
duke@0 2070 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
duke@0 2071 if( !in(0) ) return Type::TOP;
duke@0 2072 if( phase->type(in(0)) == Type::TOP )
duke@0 2073 return Type::TOP;
duke@0 2074 return TypeTuple::MEMBAR;
duke@0 2075 }
duke@0 2076
duke@0 2077 //------------------------------match------------------------------------------
duke@0 2078 // Construct projections for memory.
duke@0 2079 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
duke@0 2080 switch (proj->_con) {
duke@0 2081 case TypeFunc::Control:
duke@0 2082 case TypeFunc::Memory:
duke@0 2083 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
duke@0 2084 }
duke@0 2085 ShouldNotReachHere();
duke@0 2086 return NULL;
duke@0 2087 }
duke@0 2088
duke@0 2089 //===========================InitializeNode====================================
duke@0 2090 // SUMMARY:
duke@0 2091 // This node acts as a memory barrier on raw memory, after some raw stores.
duke@0 2092 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
duke@0 2093 // The Initialize can 'capture' suitably constrained stores as raw inits.
duke@0 2094 // It can coalesce related raw stores into larger units (called 'tiles').
duke@0 2095 // It can avoid zeroing new storage for memory units which have raw inits.
duke@0 2096 // At macro-expansion, it is marked 'complete', and does not optimize further.
duke@0 2097 //
duke@0 2098 // EXAMPLE:
duke@0 2099 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
duke@0 2100 // ctl = incoming control; mem* = incoming memory
duke@0 2101 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
duke@0 2102 // First allocate uninitialized memory and fill in the header:
duke@0 2103 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
duke@0 2104 // ctl := alloc.Control; mem* := alloc.Memory*
duke@0 2105 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
duke@0 2106 // Then initialize to zero the non-header parts of the raw memory block:
duke@0 2107 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
duke@0 2108 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
duke@0 2109 // After the initialize node executes, the object is ready for service:
duke@0 2110 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
duke@0 2111 // Suppose its body is immediately initialized as {1,2}:
duke@0 2112 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2113 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2114 // mem.SLICE(#short[*]) := store2
duke@0 2115 //
duke@0 2116 // DETAILS:
duke@0 2117 // An InitializeNode collects and isolates object initialization after
duke@0 2118 // an AllocateNode and before the next possible safepoint. As a
duke@0 2119 // memory barrier (MemBarNode), it keeps critical stores from drifting
duke@0 2120 // down past any safepoint or any publication of the allocation.
duke@0 2121 // Before this barrier, a newly-allocated object may have uninitialized bits.
duke@0 2122 // After this barrier, it may be treated as a real oop, and GC is allowed.
duke@0 2123 //
duke@0 2124 // The semantics of the InitializeNode include an implicit zeroing of
duke@0 2125 // the new object from object header to the end of the object.
duke@0 2126 // (The object header and end are determined by the AllocateNode.)
duke@0 2127 //
duke@0 2128 // Certain stores may be added as direct inputs to the InitializeNode.
duke@0 2129 // These stores must update raw memory, and they must be to addresses
duke@0 2130 // derived from the raw address produced by AllocateNode, and with
duke@0 2131 // a constant offset. They must be ordered by increasing offset.
duke@0 2132 // The first one is at in(RawStores), the last at in(req()-1).
duke@0 2133 // Unlike most memory operations, they are not linked in a chain,
duke@0 2134 // but are displayed in parallel as users of the rawmem output of
duke@0 2135 // the allocation.
duke@0 2136 //
duke@0 2137 // (See comments in InitializeNode::capture_store, which continue
duke@0 2138 // the example given above.)
duke@0 2139 //
duke@0 2140 // When the associated Allocate is macro-expanded, the InitializeNode
duke@0 2141 // may be rewritten to optimize collected stores. A ClearArrayNode
duke@0 2142 // may also be created at that point to represent any required zeroing.
duke@0 2143 // The InitializeNode is then marked 'complete', prohibiting further
duke@0 2144 // capturing of nearby memory operations.
duke@0 2145 //
duke@0 2146 // During macro-expansion, all captured initializations which store
duke@0 2147 // constant values of 32 bits or smaller are coalesced (if advantagous)
duke@0 2148 // into larger 'tiles' 32 or 64 bits. This allows an object to be
duke@0 2149 // initialized in fewer memory operations. Memory words which are
duke@0 2150 // covered by neither tiles nor non-constant stores are pre-zeroed
duke@0 2151 // by explicit stores of zero. (The code shape happens to do all
duke@0 2152 // zeroing first, then all other stores, with both sequences occurring
duke@0 2153 // in order of ascending offsets.)
duke@0 2154 //
duke@0 2155 // Alternatively, code may be inserted between an AllocateNode and its
duke@0 2156 // InitializeNode, to perform arbitrary initialization of the new object.
duke@0 2157 // E.g., the object copying intrinsics insert complex data transfers here.
duke@0 2158 // The initialization must then be marked as 'complete' disable the
duke@0 2159 // built-in zeroing semantics and the collection of initializing stores.
duke@0 2160 //
duke@0 2161 // While an InitializeNode is incomplete, reads from the memory state
duke@0 2162 // produced by it are optimizable if they match the control edge and
duke@0 2163 // new oop address associated with the allocation/initialization.
duke@0 2164 // They return a stored value (if the offset matches) or else zero.
duke@0 2165 // A write to the memory state, if it matches control and address,
duke@0 2166 // and if it is to a constant offset, may be 'captured' by the
duke@0 2167 // InitializeNode. It is cloned as a raw memory operation and rewired
duke@0 2168 // inside the initialization, to the raw oop produced by the allocation.
duke@0 2169 // Operations on addresses which are provably distinct (e.g., to
duke@0 2170 // other AllocateNodes) are allowed to bypass the initialization.
duke@0 2171 //
duke@0 2172 // The effect of all this is to consolidate object initialization
duke@0 2173 // (both arrays and non-arrays, both piecewise and bulk) into a
duke@0 2174 // single location, where it can be optimized as a unit.
duke@0 2175 //
duke@0 2176 // Only stores with an offset less than TrackedInitializationLimit words
duke@0 2177 // will be considered for capture by an InitializeNode. This puts a
duke@0 2178 // reasonable limit on the complexity of optimized initializations.
duke@0 2179
duke@0 2180 //---------------------------InitializeNode------------------------------------
duke@0 2181 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
duke@0 2182 : _is_complete(false),
duke@0 2183 MemBarNode(C, adr_type, rawoop)
duke@0 2184 {
duke@0 2185 init_class_id(Class_Initialize);
duke@0 2186
duke@0 2187 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
duke@0 2188 assert(in(RawAddress) == rawoop, "proper init");
duke@0 2189 // Note: allocation() can be NULL, for secondary initialization barriers
duke@0 2190 }
duke@0 2191
duke@0 2192 // Since this node is not matched, it will be processed by the
duke@0 2193 // register allocator. Declare that there are no constraints
duke@0 2194 // on the allocation of the RawAddress edge.
duke@0 2195 const RegMask &InitializeNode::in_RegMask(uint idx) const {
duke@0 2196 // This edge should be set to top, by the set_complete. But be conservative.
duke@0 2197 if (idx == InitializeNode::RawAddress)
duke@0 2198 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
duke@0 2199 return RegMask::Empty;
duke@0 2200 }
duke@0 2201
duke@0 2202 Node* InitializeNode::memory(uint alias_idx) {
duke@0 2203 Node* mem = in(Memory);
duke@0 2204 if (mem->is_MergeMem()) {
duke@0 2205 return mem->as_MergeMem()->memory_at(alias_idx);
duke@0 2206 } else {
duke@0 2207 // incoming raw memory is not split
duke@0 2208 return mem;
duke@0 2209 }
duke@0 2210 }
duke@0 2211
duke@0 2212 bool InitializeNode::is_non_zero() {
duke@0 2213 if (is_complete()) return false;
duke@0 2214 remove_extra_zeroes();
duke@0 2215 return (req() > RawStores);
duke@0 2216 }
duke@0 2217
duke@0 2218 void InitializeNode::set_complete(PhaseGVN* phase) {
duke@0 2219 assert(!is_complete(), "caller responsibility");
duke@0 2220 _is_complete = true;
duke@0 2221
duke@0 2222 // After this node is complete, it contains a bunch of
duke@0 2223 // raw-memory initializations. There is no need for
duke@0 2224 // it to have anything to do with non-raw memory effects.
duke@0 2225 // Therefore, tell all non-raw users to re-optimize themselves,
duke@0 2226 // after skipping the memory effects of this initialization.
duke@0 2227 PhaseIterGVN* igvn = phase->is_IterGVN();
duke@0 2228 if (igvn) igvn->add_users_to_worklist(this);
duke@0 2229 }
duke@0 2230
duke@0 2231 // convenience function
duke@0 2232 // return false if the init contains any stores already
duke@0 2233 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
duke@0 2234 InitializeNode* init = initialization();
duke@0 2235 if (init == NULL || init->is_complete()) return false;
duke@0 2236 init->remove_extra_zeroes();
duke@0 2237 // for now, if this allocation has already collected any inits, bail:
duke@0 2238 if (init->is_non_zero()) return false;
duke@0 2239 init->set_complete(phase);
duke@0 2240 return true;
duke@0 2241 }
duke@0 2242
duke@0 2243 void InitializeNode::remove_extra_zeroes() {
duke@0 2244 if (req() == RawStores) return;
duke@0 2245 Node* zmem = zero_memory();
duke@0 2246 uint fill = RawStores;
duke@0 2247 for (uint i = fill; i < req(); i++) {
duke@0 2248 Node* n = in(i);
duke@0 2249 if (n->is_top() || n == zmem) continue; // skip
duke@0 2250 if (fill < i) set_req(fill, n); // compact
duke@0 2251 ++fill;
duke@0 2252 }
duke@0 2253 // delete any empty spaces created:
duke@0 2254 while (fill < req()) {
duke@0 2255 del_req(fill);
duke@0 2256 }
duke@0 2257 }
duke@0 2258
duke@0 2259 // Helper for remembering which stores go with which offsets.
duke@0 2260 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
duke@0 2261 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
duke@0 2262 intptr_t offset = -1;
duke@0 2263 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
duke@0 2264 phase, offset);
duke@0 2265 if (base == NULL) return -1; // something is dead,
duke@0 2266 if (offset < 0) return -1; // dead, dead
duke@0 2267 return offset;
duke@0 2268 }
duke@0 2269
duke@0 2270 // Helper for proving that an initialization expression is
duke@0 2271 // "simple enough" to be folded into an object initialization.
duke@0 2272 // Attempts to prove that a store's initial value 'n' can be captured
duke@0 2273 // within the initialization without creating a vicious cycle, such as:
duke@0 2274 // { Foo p = new Foo(); p.next = p; }
duke@0 2275 // True for constants and parameters and small combinations thereof.
duke@0 2276 bool InitializeNode::detect_init_independence(Node* n,
duke@0 2277 bool st_is_pinned,
duke@0 2278 int& count) {
duke@0 2279 if (n == NULL) return true; // (can this really happen?)
duke@0 2280 if (n->is_Proj()) n = n->in(0);
duke@0 2281 if (n == this) return false; // found a cycle
duke@0 2282 if (n->is_Con()) return true;
duke@0 2283 if (n->is_Start()) return true; // params, etc., are OK
duke@0 2284 if (n->is_Root()) return true; // even better
duke@0 2285
duke@0 2286 Node* ctl = n->in(0);
duke@0 2287 if (ctl != NULL && !ctl->is_top()) {
duke@0 2288 if (ctl->is_Proj()) ctl = ctl->in(0);
duke@0 2289 if (ctl == this) return false;
duke@0 2290
duke@0 2291 // If we already know that the enclosing memory op is pinned right after
duke@0 2292 // the init, then any control flow that the store has picked up
duke@0 2293 // must have preceded the init, or else be equal to the init.
duke@0 2294 // Even after loop optimizations (which might change control edges)
duke@0 2295 // a store is never pinned *before* the availability of its inputs.
duke@0 2296 if (!MemNode::detect_dominating_control(ctl, this->in(0)))
duke@0 2297 return false; // failed to prove a good control
duke@0 2298
duke@0 2299 }
duke@0 2300
duke@0 2301 // Check data edges for possible dependencies on 'this'.
duke@0 2302 if ((count += 1) > 20) return false; // complexity limit
duke@0 2303 for (uint i = 1; i < n->req(); i++) {
duke@0 2304 Node* m = n->in(i);
duke@0 2305 if (m == NULL || m == n || m->is_top()) continue;
duke@0 2306 uint first_i = n->find_edge(m);
duke@0 2307 if (i != first_i) continue; // process duplicate edge just once
duke@0 2308 if (!detect_init_independence(m, st_is_pinned, count)) {
duke@0 2309 return false;
duke@0 2310 }
duke@0 2311 }
duke@0 2312
duke@0 2313 return true;
duke@0 2314 }
duke@0 2315
duke@0 2316 // Here are all the checks a Store must pass before it can be moved into
duke@0 2317 // an initialization. Returns zero if a check fails.
duke@0 2318 // On success, returns the (constant) offset to which the store applies,
duke@0 2319 // within the initialized memory.
duke@0 2320 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
duke@0 2321 const int FAIL = 0;
duke@0 2322 if (st->req() != MemNode::ValueIn + 1)
duke@0 2323 return FAIL; // an inscrutable StoreNode (card mark?)
duke@0 2324 Node* ctl = st->in(MemNode::Control);
duke@0 2325 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
duke@0 2326 return FAIL; // must be unconditional after the initialization
duke@0 2327 Node* mem = st->in(MemNode::Memory);
duke@0 2328 if (!(mem->is_Proj() && mem->in(0) == this))
duke@0 2329 return FAIL; // must not be preceded by other stores
duke@0 2330 Node* adr = st->in(MemNode::Address);
duke@0 2331 intptr_t offset;
duke@0 2332 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
duke@0 2333 if (alloc == NULL)
duke@0 2334 return FAIL; // inscrutable address
duke@0 2335 if (alloc != allocation())
duke@0 2336 return FAIL; // wrong allocation! (store needs to float up)
duke@0 2337 Node* val = st->in(MemNode::ValueIn);
duke@0 2338 int complexity_count = 0;
duke@0 2339 if (!detect_init_independence(val, true, complexity_count))
duke@0 2340 return FAIL; // stored value must be 'simple enough'
duke@0 2341
duke@0 2342 return offset; // success
duke@0 2343 }
duke@0 2344
duke@0 2345 // Find the captured store in(i) which corresponds to the range
duke@0 2346 // [start..start+size) in the initialized object.
duke@0 2347 // If there is one, return its index i. If there isn't, return the
duke@0 2348 // negative of the index where it should be inserted.
duke@0 2349 // Return 0 if the queried range overlaps an initialization boundary
duke@0 2350 // or if dead code is encountered.
duke@0 2351 // If size_in_bytes is zero, do not bother with overlap checks.
duke@0 2352 int InitializeNode::captured_store_insertion_point(intptr_t start,
duke@0 2353 int size_in_bytes,
duke@0 2354 PhaseTransform* phase) {
duke@0 2355 const int FAIL = 0, MAX_STORE = BytesPerLong;
duke@0 2356
duke@0 2357 if (is_complete())
duke@0 2358 return FAIL; // arraycopy got here first; punt
duke@0 2359
duke@0 2360 assert(allocation() != NULL, "must be present");
duke@0 2361
duke@0 2362 // no negatives, no header fields:
duke@0 2363 if (start < (intptr_t) sizeof(oopDesc)) return FAIL;
duke@0 2364 if (start < (intptr_t) sizeof(arrayOopDesc) &&
duke@0 2365 start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
duke@0 2366
duke@0 2367 // after a certain size, we bail out on tracking all the stores:
duke@0 2368 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2369 if (start >= ti_limit) return FAIL;
duke@0 2370
duke@0 2371 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
duke@0 2372 if (i >= limit) return -(int)i; // not found; here is where to put it
duke@0 2373
duke@0 2374 Node* st = in(i);
duke@0 2375 intptr_t st_off = get_store_offset(st, phase);
duke@0 2376 if (st_off < 0) {
duke@0 2377 if (st != zero_memory()) {
duke@0 2378 return FAIL; // bail out if there is dead garbage
duke@0 2379 }
duke@0 2380 } else if (st_off > start) {
duke@0 2381 // ...we are done, since stores are ordered
duke@0 2382 if (st_off < start + size_in_bytes) {
duke@0 2383 return FAIL; // the next store overlaps
duke@0 2384 }
duke@0 2385 return -(int)i; // not found; here is where to put it
duke@0 2386 } else if (st_off < start) {
duke@0 2387 if (size_in_bytes != 0 &&
duke@0 2388 start < st_off + MAX_STORE &&
duke@0 2389 start < st_off + st->as_Store()->memory_size()) {
duke@0 2390 return FAIL; // the previous store overlaps
duke@0 2391 }
duke@0 2392 } else {
duke@0 2393 if (size_in_bytes != 0 &&
duke@0 2394 st->as_Store()->memory_size() != size_in_bytes) {
duke@0 2395 return FAIL; // mismatched store size
duke@0 2396 }
duke@0 2397 return i;
duke@0 2398 }
duke@0 2399
duke@0 2400 ++i;
duke@0 2401 }
duke@0 2402 }
duke@0 2403
duke@0 2404 // Look for a captured store which initializes at the offset 'start'
duke@0 2405 // with the given size. If there is no such store, and no other
duke@0 2406 // initialization interferes, then return zero_memory (the memory
duke@0 2407 // projection of the AllocateNode).
duke@0 2408 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
duke@0 2409 PhaseTransform* phase) {
duke@0 2410 assert(stores_are_sane(phase), "");
duke@0 2411 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2412 if (i == 0) {
duke@0 2413 return NULL; // something is dead
duke@0 2414 } else if (i < 0) {
duke@0 2415 return zero_memory(); // just primordial zero bits here
duke@0 2416 } else {
duke@0 2417 Node* st = in(i); // here is the store at this position
duke@0 2418 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
duke@0 2419 return st;
duke@0 2420 }
duke@0 2421 }
duke@0 2422
duke@0 2423 // Create, as a raw pointer, an address within my new object at 'offset'.
duke@0 2424 Node* InitializeNode::make_raw_address(intptr_t offset,
duke@0 2425 PhaseTransform* phase) {
duke@0 2426 Node* addr = in(RawAddress);
duke@0 2427 if (offset != 0) {
duke@0 2428 Compile* C = phase->C;
duke@0 2429 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
duke@0 2430 phase->MakeConX(offset)) );
duke@0 2431 }
duke@0 2432 return addr;
duke@0 2433 }
duke@0 2434
duke@0 2435 // Clone the given store, converting it into a raw store
duke@0 2436 // initializing a field or element of my new object.
duke@0 2437 // Caller is responsible for retiring the original store,
duke@0 2438 // with subsume_node or the like.
duke@0 2439 //
duke@0 2440 // From the example above InitializeNode::InitializeNode,
duke@0 2441 // here are the old stores to be captured:
duke@0 2442 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2443 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2444 //
duke@0 2445 // Here is the changed code; note the extra edges on init:
duke@0 2446 // alloc = (Allocate ...)
duke@0 2447 // rawoop = alloc.RawAddress
duke@0 2448 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
duke@0 2449 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
duke@0 2450 // init = (Initialize alloc.Control alloc.Memory rawoop
duke@0 2451 // rawstore1 rawstore2)
duke@0 2452 //
duke@0 2453 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
duke@0 2454 PhaseTransform* phase) {
duke@0 2455 assert(stores_are_sane(phase), "");
duke@0 2456
duke@0 2457 if (start < 0) return NULL;
duke@0 2458 assert(can_capture_store(st, phase) == start, "sanity");
duke@0 2459
duke@0 2460 Compile* C = phase->C;
duke@0 2461 int size_in_bytes = st->memory_size();
duke@0 2462 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2463 if (i == 0) return NULL; // bail out
duke@0 2464 Node* prev_mem = NULL; // raw memory for the captured store
duke@0 2465 if (i > 0) {
duke@0 2466 prev_mem = in(i); // there is a pre-existing store under this one
duke@0 2467 set_req(i, C->top()); // temporarily disconnect it
duke@0 2468 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
duke@0 2469 } else {
duke@0 2470 i = -i; // no pre-existing store
duke@0 2471 prev_mem = zero_memory(); // a slice of the newly allocated object
duke@0 2472 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
duke@0 2473 set_req(--i, C->top()); // reuse this edge; it has been folded away
duke@0 2474 else
duke@0 2475 ins_req(i, C->top()); // build a new edge
duke@0 2476 }
duke@0 2477 Node* new_st = st->clone();
duke@0 2478 new_st->set_req(MemNode::Control, in(Control));
duke@0 2479 new_st->set_req(MemNode::Memory, prev_mem);
duke@0 2480 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
duke@0 2481 new_st = phase->transform(new_st);
duke@0 2482
duke@0 2483 // At this point, new_st might have swallowed a pre-existing store
duke@0 2484 // at the same offset, or perhaps new_st might have disappeared,
duke@0 2485 // if it redundantly stored the same value (or zero to fresh memory).
duke@0 2486
duke@0 2487 // In any case, wire it in:
duke@0 2488 set_req(i, new_st);
duke@0 2489
duke@0 2490 // The caller may now kill the old guy.
duke@0 2491 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
duke@0 2492 assert(check_st == new_st || check_st == NULL, "must be findable");
duke@0 2493 assert(!is_complete(), "");
duke@0 2494 return new_st;
duke@0 2495 }
duke@0 2496
duke@0 2497 static bool store_constant(jlong* tiles, int num_tiles,
duke@0 2498 intptr_t st_off, int st_size,
duke@0 2499 jlong con) {
duke@0 2500 if ((st_off & (st_size-1)) != 0)
duke@0 2501 return false; // strange store offset (assume size==2**N)
duke@0 2502 address addr = (address)tiles + st_off;
duke@0 2503 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
duke@0 2504 switch (st_size) {
duke@0 2505 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
duke@0 2506 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
duke@0 2507 case sizeof(jint): *(jint*) addr = (jint) con; break;
duke@0 2508 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
duke@0 2509 default: return false; // strange store size (detect size!=2**N here)
duke@0 2510 }
duke@0 2511 return true; // return success to caller
duke@0 2512 }
duke@0 2513
duke@0 2514 // Coalesce subword constants into int constants and possibly
duke@0 2515 // into long constants. The goal, if the CPU permits,
duke@0 2516 // is to initialize the object with a small number of 64-bit tiles.
duke@0 2517 // Also, convert floating-point constants to bit patterns.
duke@0 2518 // Non-constants are not relevant to this pass.
duke@0 2519 //
duke@0 2520 // In terms of the running example on InitializeNode::InitializeNode
duke@0 2521 // and InitializeNode::capture_store, here is the transformation
duke@0 2522 // of rawstore1 and rawstore2 into rawstore12:
duke@0 2523 // alloc = (Allocate ...)
duke@0 2524 // rawoop = alloc.RawAddress
duke@0 2525 // tile12 = 0x00010002
duke@0 2526 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
duke@0 2527 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
duke@0 2528 //
duke@0 2529 void
duke@0 2530 InitializeNode::coalesce_subword_stores(intptr_t header_size,
duke@0 2531 Node* size_in_bytes,
duke@0 2532 PhaseGVN* phase) {
duke@0 2533 Compile* C = phase->C;
duke@0 2534
duke@0 2535 assert(stores_are_sane(phase), "");
duke@0 2536 // Note: After this pass, they are not completely sane,
duke@0 2537 // since there may be some overlaps.
duke@0 2538
duke@0 2539 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
duke@0 2540
duke@0 2541 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2542 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
duke@0 2543 size_limit = MIN2(size_limit, ti_limit);
duke@0 2544 size_limit = align_size_up(size_limit, BytesPerLong);
duke@0 2545 int num_tiles = size_limit / BytesPerLong;
duke@0 2546
duke@0 2547 // allocate space for the tile map:
duke@0 2548 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
duke@0 2549 jlong tiles_buf[small_len];
duke@0 2550 Node* nodes_buf[small_len];
duke@0 2551 jlong inits_buf[small_len];
duke@0 2552 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
duke@0 2553 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 2554 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
duke@0 2555 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
duke@0 2556 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
duke@0 2557 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 2558 // tiles: exact bitwise model of all primitive constants
duke@0 2559 // nodes: last constant-storing node subsumed into the tiles model
duke@0 2560 // inits: which bytes (in each tile) are touched by any initializations
duke@0 2561
duke@0 2562 //// Pass A: Fill in the tile model with any relevant stores.
duke@0 2563
duke@0 2564 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
duke@0 2565 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
duke@0 2566 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
duke@0 2567 Node* zmem = zero_memory(); // initially zero memory state
duke@0 2568 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 2569 Node* st = in(i);
duke@0 2570 intptr_t st_off = get_store_offset(st, phase);
duke@0 2571
duke@0 2572 // Figure out the store's offset and constant value:
duke@0 2573 if (st_off < header_size) continue; //skip (ignore header)
duke@0 2574 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
duke@0 2575 int st_size = st->as_Store()->memory_size();
duke@0 2576 if (st_off + st_size > size_limit) break;
duke@0 2577
duke@0 2578 // Record which bytes are touched, whether by constant or not.
duke@0 2579 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
duke@0 2580 continue; // skip (strange store size)
duke@0 2581
duke@0 2582 const Type* val = phase->type(st->in(MemNode::ValueIn));
duke@0 2583 if (!val->singleton()) continue; //skip (non-con store)
duke@0 2584 BasicType type = val->basic_type();
duke@0 2585
duke@0 2586 jlong con = 0;
duke@0 2587 switch (type) {
duke@0 2588 case T_INT: con = val->is_int()->get_con(); break;
duke@0 2589 case T_LONG: con = val->is_long()->get_con(); break;
duke@0 2590 case T_FLOAT: con = jint_cast(val->getf()); break;
duke@0 2591 case T_DOUBLE: con = jlong_cast(val->getd()); break;
duke@0 2592 default: continue; //skip (odd store type)
duke@0 2593 }
duke@0 2594
duke@0 2595 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
duke@0 2596 st->Opcode() == Op_StoreL) {
duke@0 2597 continue; // This StoreL is already optimal.
duke@0 2598 }
duke@0 2599
duke@0 2600 // Store down the constant.
duke@0 2601 store_constant(tiles, num_tiles, st_off, st_size, con);
duke@0 2602
duke@0 2603 intptr_t j = st_off >> LogBytesPerLong;
duke@0 2604
duke@0 2605 if (type == T_INT && st_size == BytesPerInt
duke@0 2606 && (st_off & BytesPerInt) == BytesPerInt) {
duke@0 2607 jlong lcon = tiles[j];
duke@0 2608 if (!Matcher::isSimpleConstant64(lcon) &&
duke@0 2609 st->Opcode() == Op_StoreI) {
duke@0 2610 // This StoreI is already optimal by itself.
duke@0 2611 jint* intcon = (jint*) &tiles[j];
duke@0 2612 intcon[1] = 0; // undo the store_constant()
duke@0 2613
duke@0 2614 // If the previous store is also optimal by itself, back up and
duke@0 2615 // undo the action of the previous loop iteration... if we can.
duke@0 2616 // But if we can't, just let the previous half take care of itself.
duke@0 2617 st = nodes[j];
duke@0 2618 st_off -= BytesPerInt;
duke@0 2619 con = intcon[0];
duke@0 2620 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
duke@0 2621 assert(st_off >= header_size, "still ignoring header");
duke@0 2622 assert(get_store_offset(st, phase) == st_off, "must be");
duke@0 2623 assert(in(i-1) == zmem, "must be");
duke@0 2624 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
duke@0 2625 assert(con == tcon->is_int()->get_con(), "must be");
duke@0 2626 // Undo the effects of the previous loop trip, which swallowed st:
duke@0 2627 intcon[0] = 0; // undo store_constant()
duke@0 2628 set_req(i-1, st); // undo set_req(i, zmem)
duke@0 2629 nodes[j] = NULL; // undo nodes[j] = st
duke@0 2630 --old_subword; // undo ++old_subword
duke@0 2631 }
duke@0 2632 continue; // This StoreI is already optimal.
duke@0 2633 }
duke@0 2634 }
duke@0 2635
duke@0 2636 // This store is not needed.
duke@0 2637 set_req(i, zmem);
duke@0 2638 nodes[j] = st; // record for the moment
duke@0 2639 if (st_size < BytesPerLong) // something has changed
duke@0 2640 ++old_subword; // includes int/float, but who's counting...
duke@0 2641 else ++old_long;
duke@0 2642 }
duke@0 2643
duke@0 2644 if ((old_subword + old_long) == 0)
duke@0 2645 return; // nothing more to do
duke@0 2646
duke@0 2647 //// Pass B: Convert any non-zero tiles into optimal constant stores.
duke@0 2648 // Be sure to insert them before overlapping non-constant stores.
duke@0 2649 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
duke@0 2650 for (int j = 0; j < num_tiles; j++) {
duke@0 2651 jlong con = tiles[j];
duke@0 2652 jlong init = inits[j];
duke@0 2653 if (con == 0) continue;
duke@0 2654 jint con0, con1; // split the constant, address-wise
duke@0 2655 jint init0, init1; // split the init map, address-wise
duke@0 2656 { union { jlong con; jint intcon[2]; } u;
duke@0 2657 u.con = con;
duke@0 2658 con0 = u.intcon[0];
duke@0 2659 con1 = u.intcon[1];
duke@0 2660 u.con = init;
duke@0 2661 init0 = u.intcon[0];
duke@0 2662 init1 = u.intcon[1];
duke@0 2663 }
duke@0 2664
duke@0 2665 Node* old = nodes[j];
duke@0 2666 assert(old != NULL, "need the prior store");
duke@0 2667 intptr_t offset = (j * BytesPerLong);
duke@0 2668
duke@0 2669 bool split = !Matcher::isSimpleConstant64(con);
duke@0 2670
duke@0 2671 if (offset < header_size) {
duke@0 2672 assert(offset + BytesPerInt >= header_size, "second int counts");
duke@0 2673 assert(*(jint*)&tiles[j] == 0, "junk in header");
duke@0 2674 split = true; // only the second word counts
duke@0 2675 // Example: int a[] = { 42 ... }
duke@0 2676 } else if (con0 == 0 && init0 == -1) {
duke@0 2677 split = true; // first word is covered by full inits
duke@0 2678 // Example: int a[] = { ... foo(), 42 ... }
duke@0 2679 } else if (con1 == 0 && init1 == -1) {
duke@0 2680 split = true; // second word is covered by full inits
duke@0 2681 // Example: int a[] = { ... 42, foo() ... }
duke@0 2682 }
duke@0 2683
duke@0 2684 // Here's a case where init0 is neither 0 nor -1:
duke@0 2685 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
duke@0 2686 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
duke@0 2687 // In this case the tile is not split; it is (jlong)42.
duke@0 2688 // The big tile is stored down, and then the foo() value is inserted.
duke@0 2689 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
duke@0 2690
duke@0 2691 Node* ctl = old->in(MemNode::Control);
duke@0 2692 Node* adr = make_raw_address(offset, phase);
duke@0 2693 const TypePtr* atp = TypeRawPtr::BOTTOM;
duke@0 2694
duke@0 2695 // One or two coalesced stores to plop down.
duke@0 2696 Node* st[2];
duke@0 2697 intptr_t off[2];
duke@0 2698 int nst = 0;
duke@0 2699 if (!split) {
duke@0 2700 ++new_long;
duke@0 2701 off[nst] = offset;
duke@0 2702 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
duke@0 2703 phase->longcon(con), T_LONG);
duke@0 2704 } else {
duke@0 2705 // Omit either if it is a zero.
duke@0 2706 if (con0 != 0) {
duke@0 2707 ++new_int;
duke@0 2708 off[nst] = offset;
duke@0 2709 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
duke@0 2710 phase->intcon(con0), T_INT);
duke@0 2711 }
duke@0 2712 if (con1 != 0) {
duke@0 2713 ++new_int;
duke@0 2714 offset += BytesPerInt;
duke@0 2715 adr = make_raw_address(offset, phase);
duke@0 2716 off[nst] = offset;
duke@0 2717 st[nst++] = StoreNode::make(C, ctl, zmem, adr, atp,
duke@0 2718 phase->intcon(con1), T_INT);
duke@0 2719 }
duke@0 2720 }
duke@0 2721
duke@0 2722 // Insert second store first, then the first before the second.
duke@0 2723 // Insert each one just before any overlapping non-constant stores.
duke@0 2724 while (nst > 0) {
duke@0 2725 Node* st1 = st[--nst];
duke@0 2726 C->copy_node_notes_to(st1, old);
duke@0 2727 st1 = phase->transform(st1);
duke@0 2728 offset = off[nst];
duke@0 2729 assert(offset >= header_size, "do not smash header");
duke@0 2730 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
duke@0 2731 guarantee(ins_idx != 0, "must re-insert constant store");
duke@0 2732 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
duke@0 2733 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
duke@0 2734 set_req(--ins_idx, st1);
duke@0 2735 else
duke@0 2736 ins_req(ins_idx, st1);
duke@0 2737 }
duke@0 2738 }
duke@0 2739
duke@0 2740 if (PrintCompilation && WizardMode)
duke@0 2741 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
duke@0 2742 old_subword, old_long, new_int, new_long);
duke@0 2743 if (C->log() != NULL)
duke@0 2744 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
duke@0 2745 old_subword, old_long, new_int, new_long);
duke@0 2746
duke@0 2747 // Clean up any remaining occurrences of zmem:
duke@0 2748 remove_extra_zeroes();
duke@0 2749 }
duke@0 2750
duke@0 2751 // Explore forward from in(start) to find the first fully initialized
duke@0 2752 // word, and return its offset. Skip groups of subword stores which
duke@0 2753 // together initialize full words. If in(start) is itself part of a
duke@0 2754 // fully initialized word, return the offset of in(start). If there
duke@0 2755 // are no following full-word stores, or if something is fishy, return
duke@0 2756 // a negative value.
duke@0 2757 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
duke@0 2758 int int_map = 0;
duke@0 2759 intptr_t int_map_off = 0;
duke@0 2760 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
duke@0 2761
duke@0 2762 for (uint i = start, limit = req(); i < limit; i++) {
duke@0 2763 Node* st = in(i);
duke@0 2764
duke@0 2765 intptr_t st_off = get_store_offset(st, phase);
duke@0 2766 if (st_off < 0) break; // return conservative answer
duke@0 2767
duke@0 2768 int st_size = st->as_Store()->memory_size();
duke@0 2769 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
duke@0 2770 return st_off; // we found a complete word init
duke@0 2771 }
duke@0 2772
duke@0 2773 // update the map:
duke@0 2774
duke@0 2775 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
duke@0 2776 if (this_int_off != int_map_off) {
duke@0 2777 // reset the map:
duke@0 2778 int_map = 0;
duke@0 2779 int_map_off = this_int_off;
duke@0 2780 }
duke@0 2781
duke@0 2782 int subword_off = st_off - this_int_off;
duke@0 2783 int_map |= right_n_bits(st_size) << subword_off;
duke@0 2784 if ((int_map & FULL_MAP) == FULL_MAP) {
duke@0 2785 return this_int_off; // we found a complete word init
duke@0 2786 }
duke@0 2787
duke@0 2788 // Did this store hit or cross the word boundary?
duke@0 2789 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
duke@0 2790 if (next_int_off == this_int_off + BytesPerInt) {
duke@0 2791 // We passed the current int, without fully initializing it.
duke@0 2792 int_map_off = next_int_off;
duke@0 2793 int_map >>= BytesPerInt;
duke@0 2794 } else if (next_int_off > this_int_off + BytesPerInt) {
duke@0 2795 // We passed the current and next int.
duke@0 2796 return this_int_off + BytesPerInt;
duke@0 2797 }
duke@0 2798 }
duke@0 2799
duke@0 2800 return -1;
duke@0 2801 }
duke@0 2802
duke@0 2803
duke@0 2804 // Called when the associated AllocateNode is expanded into CFG.
duke@0 2805 // At this point, we may perform additional optimizations.
duke@0 2806 // Linearize the stores by ascending offset, to make memory
duke@0 2807 // activity as coherent as possible.
duke@0 2808 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
duke@0 2809 intptr_t header_size,
duke@0 2810 Node* size_in_bytes,
duke@0 2811 PhaseGVN* phase) {
duke@0 2812 assert(!is_complete(), "not already complete");
duke@0 2813 assert(stores_are_sane(phase), "");
duke@0 2814 assert(allocation() != NULL, "must be present");
duke@0 2815
duke@0 2816 remove_extra_zeroes();
duke@0 2817
duke@0 2818 if (ReduceFieldZeroing || ReduceBulkZeroing)
duke@0 2819 // reduce instruction count for common initialization patterns
duke@0 2820 coalesce_subword_stores(header_size, size_in_bytes, phase);
duke@0 2821
duke@0 2822 Node* zmem = zero_memory(); // initially zero memory state
duke@0 2823 Node* inits = zmem; // accumulating a linearized chain of inits
duke@0 2824 #ifdef ASSERT
duke@0 2825 intptr_t last_init_off = sizeof(oopDesc); // previous init offset
duke@0 2826 intptr_t last_init_end = sizeof(oopDesc); // previous init offset+size
duke@0 2827 intptr_t last_tile_end = sizeof(oopDesc); // previous tile offset+size
duke@0 2828 #endif
duke@0 2829 intptr_t zeroes_done = header_size;
duke@0 2830
duke@0 2831 bool do_zeroing = true; // we might give up if inits are very sparse
duke@0 2832 int big_init_gaps = 0; // how many large gaps have we seen?
duke@0 2833
duke@0 2834 if (ZeroTLAB) do_zeroing = false;
duke@0 2835 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
duke@0 2836
duke@0 2837 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 2838 Node* st = in(i);
duke@0 2839 intptr_t st_off = get_store_offset(st, phase);
duke@0 2840 if (st_off < 0)
duke@0 2841 break; // unknown junk in the inits
duke@0 2842 if (st->in(MemNode::Memory) != zmem)
duke@0 2843 break; // complicated store chains somehow in list
duke@0 2844
duke@0 2845 int st_size = st->as_Store()->memory_size();
duke@0 2846 intptr_t next_init_off = st_off + st_size;
duke@0 2847
duke@0 2848 if (do_zeroing && zeroes_done < next_init_off) {
duke@0 2849 // See if this store needs a zero before it or under it.
duke@0 2850 intptr_t zeroes_needed = st_off;
duke@0 2851
duke@0 2852 if (st_size < BytesPerInt) {
duke@0 2853 // Look for subword stores which only partially initialize words.
duke@0 2854 // If we find some, we must lay down some word-level zeroes first,
duke@0 2855 // underneath the subword stores.
duke@0 2856 //
duke@0 2857 // Examples:
duke@0 2858 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
duke@0 2859 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
duke@0 2860 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
duke@0 2861 //
duke@0 2862 // Note: coalesce_subword_stores may have already done this,
duke@0 2863 // if it was prompted by constant non-zero subword initializers.
duke@0 2864 // But this case can still arise with non-constant stores.
duke@0 2865
duke@0 2866 intptr_t next_full_store = find_next_fullword_store(i, phase);
duke@0 2867
duke@0 2868 // In the examples above:
duke@0 2869 // in(i) p q r s x y z
duke@0 2870 // st_off 12 13 14 15 12 13 14
duke@0 2871 // st_size 1 1 1 1 1 1 1
duke@0 2872 // next_full_s. 12 16 16 16 16 16 16
duke@0 2873 // z's_done 12 16 16 16 12 16 12
duke@0 2874 // z's_needed 12 16 16 16 16 16 16
duke@0 2875 // zsize 0 0 0 0 4 0 4
duke@0 2876 if (next_full_store < 0) {
duke@0 2877 // Conservative tack: Zero to end of current word.
duke@0 2878 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
duke@0 2879 } else {
duke@0 2880 // Zero to beginning of next fully initialized word.
duke@0 2881 // Or, don't zero at all, if we are already in that word.
duke@0 2882 assert(next_full_store >= zeroes_needed, "must go forward");
duke@0 2883 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
duke@0 2884 zeroes_needed = next_full_store;
duke@0 2885 }
duke@0 2886 }
duke@0 2887
duke@0 2888 if (zeroes_needed > zeroes_done) {
duke@0 2889 intptr_t zsize = zeroes_needed - zeroes_done;
duke@0 2890 // Do some incremental zeroing on rawmem, in parallel with inits.
duke@0 2891 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 2892 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 2893 zeroes_done, zeroes_needed,
duke@0 2894 phase);
duke@0 2895 zeroes_done = zeroes_needed;
duke@0 2896 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
duke@0 2897 do_zeroing = false; // leave the hole, next time
duke@0 2898 }
duke@0 2899 }
duke@0 2900
duke@0 2901 // Collect the store and move on:
duke@0 2902 st->set_req(MemNode::Memory, inits);
duke@0 2903 inits = st; // put it on the linearized chain
duke@0 2904 set_req(i, zmem); // unhook from previous position
duke@0 2905
duke@0 2906 if (zeroes_done == st_off)
duke@0 2907 zeroes_done = next_init_off;
duke@0 2908
duke@0 2909 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
duke@0 2910
duke@0 2911 #ifdef ASSERT
duke@0 2912 // Various order invariants. Weaker than stores_are_sane because
duke@0 2913 // a large constant tile can be filled in by smaller non-constant stores.
duke@0 2914 assert(st_off >= last_init_off, "inits do not reverse");
duke@0 2915 last_init_off = st_off;
duke@0 2916 const Type* val = NULL;
duke@0 2917 if (st_size >= BytesPerInt &&
duke@0 2918 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
duke@0 2919 (int)val->basic_type() < (int)T_OBJECT) {
duke@0 2920 assert(st_off >= last_tile_end, "tiles do not overlap");
duke@0 2921 assert(st_off >= last_init_end, "tiles do not overwrite inits");
duke@0 2922 last_tile_end = MAX2(last_tile_end, next_init_off);
duke@0 2923 } else {
duke@0 2924 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
duke@0 2925 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
duke@0 2926 assert(st_off >= last_init_end, "inits do not overlap");
duke@0 2927 last_init_end = next_init_off; // it's a non-tile
duke@0 2928 }
duke@0 2929 #endif //ASSERT
duke@0 2930 }
duke@0 2931
duke@0 2932 remove_extra_zeroes(); // clear out all the zmems left over
duke@0 2933 add_req(inits);
duke@0 2934
duke@0 2935 if (!ZeroTLAB) {
duke@0 2936 // If anything remains to be zeroed, zero it all now.
duke@0 2937 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 2938 // if it is the last unused 4 bytes of an instance, forget about it
duke@0 2939 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
duke@0 2940 if (zeroes_done + BytesPerLong >= size_limit) {
duke@0 2941 assert(allocation() != NULL, "");
duke@0 2942 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
duke@0 2943 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
duke@0 2944 if (zeroes_done == k->layout_helper())
duke@0 2945 zeroes_done = size_limit;
duke@0 2946 }
duke@0 2947 if (zeroes_done < size_limit) {
duke@0 2948 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 2949 zeroes_done, size_in_bytes, phase);
duke@0 2950 }
duke@0 2951 }
duke@0 2952
duke@0 2953 set_complete(phase);
duke@0 2954 return rawmem;
duke@0 2955 }
duke@0 2956
duke@0 2957
duke@0 2958 #ifdef ASSERT
duke@0 2959 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
duke@0 2960 if (is_complete())
duke@0 2961 return true; // stores could be anything at this point
duke@0 2962 intptr_t last_off = sizeof(oopDesc);
duke@0 2963 for (uint i = InitializeNode::RawStores; i < req(); i++) {
duke@0 2964 Node* st = in(i);
duke@0 2965 intptr_t st_off = get_store_offset(st, phase);
duke@0 2966 if (st_off < 0) continue; // ignore dead garbage
duke@0 2967 if (last_off > st_off) {
duke@0 2968 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
duke@0 2969 this->dump(2);
duke@0 2970 assert(false, "ascending store offsets");
duke@0 2971 return false;
duke@0 2972 }
duke@0 2973 last_off = st_off + st->as_Store()->memory_size();
duke@0 2974 }
duke@0 2975 return true;
duke@0 2976 }
duke@0 2977 #endif //ASSERT
duke@0 2978
duke@0 2979
duke@0 2980
duke@0 2981
duke@0 2982 //============================MergeMemNode=====================================
duke@0 2983 //
duke@0 2984 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
duke@0 2985 // contributing store or call operations. Each contributor provides the memory
duke@0 2986 // state for a particular "alias type" (see Compile::alias_type). For example,
duke@0 2987 // if a MergeMem has an input X for alias category #6, then any memory reference
duke@0 2988 // to alias category #6 may use X as its memory state input, as an exact equivalent
duke@0 2989 // to using the MergeMem as a whole.
duke@0 2990 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
duke@0 2991 //
duke@0 2992 // (Here, the <N> notation gives the index of the relevant adr_type.)
duke@0 2993 //
duke@0 2994 // In one special case (and more cases in the future), alias categories overlap.
duke@0 2995 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
duke@0 2996 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
duke@0 2997 // it is exactly equivalent to that state W:
duke@0 2998 // MergeMem(<Bot>: W) <==> W
duke@0 2999 //
duke@0 3000 // Usually, the merge has more than one input. In that case, where inputs
duke@0 3001 // overlap (i.e., one is Bot), the narrower alias type determines the memory
duke@0 3002 // state for that type, and the wider alias type (Bot) fills in everywhere else:
duke@0 3003 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
duke@0 3004 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
duke@0 3005 //
duke@0 3006 // A merge can take a "wide" memory state as one of its narrow inputs.
duke@0 3007 // This simply means that the merge observes out only the relevant parts of
duke@0 3008 // the wide input. That is, wide memory states arriving at narrow merge inputs
duke@0 3009 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
duke@0 3010 //
duke@0 3011 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
duke@0 3012 // and that memory slices "leak through":
duke@0 3013 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
duke@0 3014 //
duke@0 3015 // But, in such a cascade, repeated memory slices can "block the leak":
duke@0 3016 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
duke@0 3017 //
duke@0 3018 // In the last example, Y is not part of the combined memory state of the
duke@0 3019 // outermost MergeMem. The system must, of course, prevent unschedulable
duke@0 3020 // memory states from arising, so you can be sure that the state Y is somehow
duke@0 3021 // a precursor to state Y'.
duke@0 3022 //
duke@0 3023 //
duke@0 3024 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
duke@0 3025 // of each MergeMemNode array are exactly the numerical alias indexes, including
duke@0 3026 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
duke@0 3027 // Compile::alias_type (and kin) produce and manage these indexes.
duke@0 3028 //
duke@0 3029 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
duke@0 3030 // (Note that this provides quick access to the top node inside MergeMem methods,
duke@0 3031 // without the need to reach out via TLS to Compile::current.)
duke@0 3032 //
duke@0 3033 // As a consequence of what was just described, a MergeMem that represents a full
duke@0 3034 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
duke@0 3035 // containing all alias categories.
duke@0 3036 //
duke@0 3037 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
duke@0 3038 //
duke@0 3039 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
duke@0 3040 // a memory state for the alias type <N>, or else the top node, meaning that
duke@0 3041 // there is no particular input for that alias type. Note that the length of
duke@0 3042 // a MergeMem is variable, and may be extended at any time to accommodate new
duke@0 3043 // memory states at larger alias indexes. When merges grow, they are of course
duke@0 3044 // filled with "top" in the unused in() positions.
duke@0 3045 //
duke@0 3046 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
duke@0 3047 // (Top was chosen because it works smoothly with passes like GCM.)
duke@0 3048 //
duke@0 3049 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
duke@0 3050 // the type of random VM bits like TLS references.) Since it is always the
duke@0 3051 // first non-Bot memory slice, some low-level loops use it to initialize an
duke@0 3052 // index variable: for (i = AliasIdxRaw; i < req(); i++).
duke@0 3053 //
duke@0 3054 //
duke@0 3055 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
duke@0 3056 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
duke@0 3057 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
duke@0 3058 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
duke@0 3059 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
duke@0 3060 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
duke@0 3061 //
duke@0 3062 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
duke@0 3063 // really that different from the other memory inputs. An abbreviation called
duke@0 3064 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
duke@0 3065 //
duke@0 3066 //
duke@0 3067 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
duke@0 3068 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
duke@0 3069 // that "emerges though" the base memory will be marked as excluding the alias types
duke@0 3070 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
duke@0 3071 //
duke@0 3072 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
duke@0 3073 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
duke@0 3074 //
duke@0 3075 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
duke@0 3076 // (It is currently unimplemented.) As you can see, the resulting merge is
duke@0 3077 // actually a disjoint union of memory states, rather than an overlay.
duke@0 3078 //
duke@0 3079
duke@0 3080 //------------------------------MergeMemNode-----------------------------------
duke@0 3081 Node* MergeMemNode::make_empty_memory() {
duke@0 3082 Node* empty_memory = (Node*) Compile::current()->top();
duke@0 3083 assert(empty_memory->is_top(), "correct sentinel identity");
duke@0 3084 return empty_memory;
duke@0 3085 }
duke@0 3086
duke@0 3087 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
duke@0 3088 init_class_id(Class_MergeMem);
duke@0 3089 // all inputs are nullified in Node::Node(int)
duke@0 3090 // set_input(0, NULL); // no control input
duke@0 3091
duke@0 3092 // Initialize the edges uniformly to top, for starters.
duke@0 3093 Node* empty_mem = make_empty_memory();
duke@0 3094 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
duke@0 3095 init_req(i,empty_mem);
duke@0 3096 }
duke@0 3097 assert(empty_memory() == empty_mem, "");
duke@0 3098
duke@0 3099 if( new_base != NULL && new_base->is_MergeMem() ) {
duke@0 3100 MergeMemNode* mdef = new_base->as_MergeMem();
duke@0 3101 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
duke@0 3102 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
duke@0 3103 mms.set_memory(mms.memory2());
duke@0 3104 }
duke@0 3105 assert(base_memory() == mdef->base_memory(), "");
duke@0 3106 } else {
duke@0 3107 set_base_memory(new_base);
duke@0 3108 }
duke@0 3109 }
duke@0 3110
duke@0 3111 // Make a new, untransformed MergeMem with the same base as 'mem'.
duke@0 3112 // If mem is itself a MergeMem, populate the result with the same edges.
duke@0 3113 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
duke@0 3114 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
duke@0 3115 }
duke@0 3116
duke@0 3117 //------------------------------cmp--------------------------------------------
duke@0 3118 uint MergeMemNode::hash() const { return NO_HASH; }
duke@0 3119 uint MergeMemNode::cmp( const Node &n ) const {
duke@0 3120 return (&n == this); // Always fail except on self
duke@0 3121 }
duke@0 3122
duke@0 3123 //------------------------------Identity---------------------------------------
duke@0 3124 Node* MergeMemNode::Identity(PhaseTransform *phase) {
duke@0 3125 // Identity if this merge point does not record any interesting memory
duke@0 3126 // disambiguations.
duke@0 3127 Node* base_mem = base_memory();
duke@0 3128 Node* empty_mem = empty_memory();
duke@0 3129 if (base_mem != empty_mem) { // Memory path is not dead?
duke@0 3130 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3131 Node* mem = in(i);
duke@0 3132 if (mem != empty_mem && mem != base_mem) {
duke@0 3133 return this; // Many memory splits; no change
duke@0 3134 }
duke@0 3135 }
duke@0 3136 }
duke@0 3137 return base_mem; // No memory splits; ID on the one true input
duke@0 3138 }
duke@0 3139
duke@0 3140 //------------------------------Ideal------------------------------------------
duke@0 3141 // This method is invoked recursively on chains of MergeMem nodes
duke@0 3142 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 3143 // Remove chain'd MergeMems
duke@0 3144 //
duke@0 3145 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
duke@0 3146 // relative to the "in(Bot)". Since we are patching both at the same time,
duke@0 3147 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
duke@0 3148 // but rewrite each "in(i)" relative to the new "in(Bot)".
duke@0 3149 Node *progress = NULL;
duke@0 3150
duke@0 3151
duke@0 3152 Node* old_base = base_memory();
duke@0 3153 Node* empty_mem = empty_memory();
duke@0 3154 if (old_base == empty_mem)
duke@0 3155 return NULL; // Dead memory path.
duke@0 3156
duke@0 3157 MergeMemNode* old_mbase;
duke@0 3158 if (old_base != NULL && old_base->is_MergeMem())
duke@0 3159 old_mbase = old_base->as_MergeMem();
duke@0 3160 else
duke@0 3161 old_mbase = NULL;
duke@0 3162 Node* new_base = old_base;
duke@0 3163
duke@0 3164 // simplify stacked MergeMems in base memory
duke@0 3165 if (old_mbase) new_base = old_mbase->base_memory();
duke@0 3166
duke@0 3167 // the base memory might contribute new slices beyond my req()
duke@0 3168 if (old_mbase) grow_to_match(old_mbase);
duke@0 3169
duke@0 3170 // Look carefully at the base node if it is a phi.
duke@0 3171 PhiNode* phi_base;
duke@0 3172 if (new_base != NULL && new_base->is_Phi())
duke@0 3173 phi_base = new_base->as_Phi();
duke@0 3174 else
duke@0 3175 phi_base = NULL;
duke@0 3176
duke@0 3177 Node* phi_reg = NULL;
duke@0 3178 uint phi_len = (uint)-1;
duke@0 3179 if (phi_base != NULL && !phi_base->is_copy()) {
duke@0 3180 // do not examine phi if degraded to a copy
duke@0 3181 phi_reg = phi_base->region();
duke@0 3182 phi_len = phi_base->req();
duke@0 3183 // see if the phi is unfinished
duke@0 3184 for (uint i = 1; i < phi_len; i++) {
duke@0 3185 if (phi_base->in(i) == NULL) {
duke@0 3186 // incomplete phi; do not look at it yet!
duke@0 3187 phi_reg = NULL;
duke@0 3188 phi_len = (uint)-1;
duke@0 3189 break;
duke@0 3190 }
duke@0 3191 }
duke@0 3192 }
duke@0 3193
duke@0 3194 // Note: We do not call verify_sparse on entry, because inputs
duke@0 3195 // can normalize to the base_memory via subsume_node or similar
duke@0 3196 // mechanisms. This method repairs that damage.
duke@0 3197
duke@0 3198 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
duke@0 3199
duke@0 3200 // Look at each slice.
duke@0 3201 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3202 Node* old_in = in(i);
duke@0 3203 // calculate the old memory value
duke@0 3204 Node* old_mem = old_in;
duke@0 3205 if (old_mem == empty_mem) old_mem = old_base;
duke@0 3206 assert(old_mem == memory_at(i), "");
duke@0 3207
duke@0 3208 // maybe update (reslice) the old memory value
duke@0 3209
duke@0 3210 // simplify stacked MergeMems
duke@0 3211 Node* new_mem = old_mem;
duke@0 3212 MergeMemNode* old_mmem;
duke@0 3213 if (old_mem != NULL && old_mem->is_MergeMem())
duke@0 3214 old_mmem = old_mem->as_MergeMem();
duke@0 3215 else
duke@0 3216 old_mmem = NULL;
duke@0 3217 if (old_mmem == this) {
duke@0 3218 // This can happen if loops break up and safepoints disappear.
duke@0 3219 // A merge of BotPtr (default) with a RawPtr memory derived from a
duke@0 3220 // safepoint can be rewritten to a merge of the same BotPtr with
duke@0 3221 // the BotPtr phi coming into the loop. If that phi disappears
duke@0 3222 // also, we can end up with a self-loop of the mergemem.
duke@0 3223 // In general, if loops degenerate and memory effects disappear,
duke@0 3224 // a mergemem can be left looking at itself. This simply means
duke@0 3225 // that the mergemem's default should be used, since there is
duke@0 3226 // no longer any apparent effect on this slice.
duke@0 3227 // Note: If a memory slice is a MergeMem cycle, it is unreachable
duke@0 3228 // from start. Update the input to TOP.
duke@0 3229 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
duke@0 3230 }
duke@0 3231 else if (old_mmem != NULL) {
duke@0 3232 new_mem = old_mmem->memory_at(i);
duke@0 3233 }
duke@0 3234 // else preceeding memory was not a MergeMem
duke@0 3235
duke@0 3236 // replace equivalent phis (unfortunately, they do not GVN together)
duke@0 3237 if (new_mem != NULL && new_mem != new_base &&
duke@0 3238 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
duke@0 3239 if (new_mem->is_Phi()) {
duke@0 3240 PhiNode* phi_mem = new_mem->as_Phi();
duke@0 3241 for (uint i = 1; i < phi_len; i++) {
duke@0 3242 if (phi_base->in(i) != phi_mem->in(i)) {
duke@0 3243 phi_mem = NULL;
duke@0 3244 break;
duke@0 3245 }
duke@0 3246 }
duke@0 3247 if (phi_mem != NULL) {
duke@0 3248 // equivalent phi nodes; revert to the def
duke@0 3249 new_mem = new_base;
duke@0 3250 }
duke@0 3251 }
duke@0 3252 }
duke@0 3253
duke@0 3254 // maybe store down a new value
duke@0 3255 Node* new_in = new_mem;
duke@0 3256 if (new_in == new_base) new_in = empty_mem;
duke@0 3257
duke@0 3258 if (new_in != old_in) {
duke@0 3259 // Warning: Do not combine this "if" with the previous "if"
duke@0 3260 // A memory slice might have be be rewritten even if it is semantically
duke@0 3261 // unchanged, if the base_memory value has changed.
duke@0 3262 set_req(i, new_in);
duke@0 3263 progress = this; // Report progress
duke@0 3264 }
duke@0 3265 }
duke@0 3266
duke@0 3267 if (new_base != old_base) {
duke@0 3268 set_req(Compile::AliasIdxBot, new_base);
duke@0 3269 // Don't use set_base_memory(new_base), because we need to update du.
duke@0 3270 assert(base_memory() == new_base, "");
duke@0 3271 progress = this;
duke@0 3272 }
duke@0 3273
duke@0 3274 if( base_memory() == this ) {
duke@0 3275 // a self cycle indicates this memory path is dead
duke@0 3276 set_req(Compile::AliasIdxBot, empty_mem);
duke@0 3277 }
duke@0 3278
duke@0 3279 // Resolve external cycles by calling Ideal on a MergeMem base_memory
duke@0 3280 // Recursion must occur after the self cycle check above
duke@0 3281 if( base_memory()->is_MergeMem() ) {
duke@0 3282 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
duke@0 3283 Node *m = phase->transform(new_mbase); // Rollup any cycles
duke@0 3284 if( m != NULL && (m->is_top() ||
duke@0 3285 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
duke@0 3286 // propagate rollup of dead cycle to self
duke@0 3287 set_req(Compile::AliasIdxBot, empty_mem);
duke@0 3288 }
duke@0 3289 }
duke@0 3290
duke@0 3291 if( base_memory() == empty_mem ) {
duke@0 3292 progress = this;
duke@0 3293 // Cut inputs during Parse phase only.
duke@0 3294 // During Optimize phase a dead MergeMem node will be subsumed by Top.
duke@0 3295 if( !can_reshape ) {
duke@0 3296 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3297 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
duke@0 3298 }
duke@0 3299 }
duke@0 3300 }
duke@0 3301
duke@0 3302 if( !progress && base_memory()->is_Phi() && can_reshape ) {
duke@0 3303 // Check if PhiNode::Ideal's "Split phis through memory merges"
duke@0 3304 // transform should be attempted. Look for this->phi->this cycle.
duke@0 3305 uint merge_width = req();
duke@0 3306 if (merge_width > Compile::AliasIdxRaw) {
duke@0 3307 PhiNode* phi = base_memory()->as_Phi();
duke@0 3308 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
duke@0 3309 if (phi->in(i) == this) {
duke@0 3310 phase->is_IterGVN()->_worklist.push(phi);
duke@0 3311 break;
duke@0 3312 }
duke@0 3313 }
duke@0 3314 }
duke@0 3315 }
duke@0 3316
kvn@64 3317 assert(progress || verify_sparse(), "please, no dups of base");
duke@0 3318 return progress;
duke@0 3319 }
duke@0 3320
duke@0 3321 //-------------------------set_base_memory-------------------------------------
duke@0 3322 void MergeMemNode::set_base_memory(Node *new_base) {
duke@0 3323 Node* empty_mem = empty_memory();
duke@0 3324 set_req(Compile::AliasIdxBot, new_base);
duke@0 3325 assert(memory_at(req()) == new_base, "must set default memory");
duke@0 3326 // Clear out other occurrences of new_base:
duke@0 3327 if (new_base != empty_mem) {
duke@0 3328 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3329 if (in(i) == new_base) set_req(i, empty_mem);
duke@0 3330 }
duke@0 3331 }
duke@0 3332 }
duke@0 3333
duke@0 3334 //------------------------------out_RegMask------------------------------------
duke@0 3335 const RegMask &MergeMemNode::out_RegMask() const {
duke@0 3336 return RegMask::Empty;
duke@0 3337 }
duke@0 3338
duke@0 3339 //------------------------------dump_spec--------------------------------------
duke@0 3340 #ifndef PRODUCT
duke@0 3341 void MergeMemNode::dump_spec(outputStream *st) const {
duke@0 3342 st->print(" {");
duke@0 3343 Node* base_mem = base_memory();
duke@0 3344 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
duke@0 3345 Node* mem = memory_at(i);
duke@0 3346 if (mem == base_mem) { st->print(" -"); continue; }
duke@0 3347 st->print( " N%d:", mem->_idx );
duke@0 3348 Compile::current()->get_adr_type(i)->dump_on(st);
duke@0 3349 }
duke@0 3350 st->print(" }");
duke@0 3351 }
duke@0 3352 #endif // !PRODUCT
duke@0 3353
duke@0 3354
duke@0 3355 #ifdef ASSERT
duke@0 3356 static bool might_be_same(Node* a, Node* b) {
duke@0 3357 if (a == b) return true;
duke@0 3358 if (!(a->is_Phi() || b->is_Phi())) return false;
duke@0 3359 // phis shift around during optimization
duke@0 3360 return true; // pretty stupid...
duke@0 3361 }
duke@0 3362
duke@0 3363 // verify a narrow slice (either incoming or outgoing)
duke@0 3364 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
duke@0 3365 if (!VerifyAliases) return; // don't bother to verify unless requested
duke@0 3366 if (is_error_reported()) return; // muzzle asserts when debugging an error
duke@0 3367 if (Node::in_dump()) return; // muzzle asserts when printing
duke@0 3368 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
duke@0 3369 assert(n != NULL, "");
duke@0 3370 // Elide intervening MergeMem's
duke@0 3371 while (n->is_MergeMem()) {
duke@0 3372 n = n->as_MergeMem()->memory_at(alias_idx);
duke@0 3373 }
duke@0 3374 Compile* C = Compile::current();
duke@0 3375 const TypePtr* n_adr_type = n->adr_type();
duke@0 3376 if (n == m->empty_memory()) {
duke@0 3377 // Implicit copy of base_memory()
duke@0 3378 } else if (n_adr_type != TypePtr::BOTTOM) {
duke@0 3379 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
duke@0 3380 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
duke@0 3381 } else {
duke@0 3382 // A few places like make_runtime_call "know" that VM calls are narrow,
duke@0 3383 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
duke@0 3384 bool expected_wide_mem = false;
duke@0 3385 if (n == m->base_memory()) {
duke@0 3386 expected_wide_mem = true;
duke@0 3387 } else if (alias_idx == Compile::AliasIdxRaw ||
duke@0 3388 n == m->memory_at(Compile::AliasIdxRaw)) {
duke@0 3389 expected_wide_mem = true;
duke@0 3390 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
duke@0 3391 // memory can "leak through" calls on channels that
duke@0 3392 // are write-once. Allow this also.
duke@0 3393 expected_wide_mem = true;
duke@0 3394 }
duke@0 3395 assert(expected_wide_mem, "expected narrow slice replacement");
duke@0 3396 }
duke@0 3397 }
duke@0 3398 #else // !ASSERT
duke@0 3399 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
duke@0 3400 #endif
duke@0 3401
duke@0 3402
duke@0 3403 //-----------------------------memory_at---------------------------------------
duke@0 3404 Node* MergeMemNode::memory_at(uint alias_idx) const {
duke@0 3405 assert(alias_idx >= Compile::AliasIdxRaw ||
duke@0 3406 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
duke@0 3407 "must avoid base_memory and AliasIdxTop");
duke@0 3408
duke@0 3409 // Otherwise, it is a narrow slice.
duke@0 3410 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
duke@0 3411 Compile *C = Compile::current();
duke@0 3412 if (is_empty_memory(n)) {
duke@0 3413 // the array is sparse; empty slots are the "top" node
duke@0 3414 n = base_memory();
duke@0 3415 assert(Node::in_dump()
duke@0 3416 || n == NULL || n->bottom_type() == Type::TOP
duke@0 3417 || n->adr_type() == TypePtr::BOTTOM
duke@0 3418 || n->adr_type() == TypeRawPtr::BOTTOM
duke@0 3419 || Compile::current()->AliasLevel() == 0,
duke@0 3420 "must be a wide memory");
duke@0 3421 // AliasLevel == 0 if we are organizing the memory states manually.
duke@0 3422 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
duke@0 3423 } else {
duke@0 3424 // make sure the stored slice is sane
duke@0 3425 #ifdef ASSERT
duke@0 3426 if (is_error_reported() || Node::in_dump()) {
duke@0 3427 } else if (might_be_same(n, base_memory())) {
duke@0 3428 // Give it a pass: It is a mostly harmless repetition of the base.
duke@0 3429 // This can arise normally from node subsumption during optimization.
duke@0 3430 } else {
duke@0 3431 verify_memory_slice(this, alias_idx, n);
duke@0 3432 }
duke@0 3433 #endif
duke@0 3434 }
duke@0 3435 return n;
duke@0 3436 }
duke@0 3437
duke@0 3438 //---------------------------set_memory_at-------------------------------------
duke@0 3439 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
duke@0 3440 verify_memory_slice(this, alias_idx, n);
duke@0 3441 Node* empty_mem = empty_memory();
duke@0 3442 if (n == base_memory()) n = empty_mem; // collapse default
duke@0 3443 uint need_req = alias_idx+1;
duke@0 3444 if (req() < need_req) {
duke@0 3445 if (n == empty_mem) return; // already the default, so do not grow me
duke@0 3446 // grow the sparse array
duke@0 3447 do {
duke@0 3448 add_req(empty_mem);
duke@0 3449 } while (req() < need_req);
duke@0 3450 }
duke@0 3451 set_req( alias_idx, n );
duke@0 3452 }
duke@0 3453
duke@0 3454
duke@0 3455
duke@0 3456 //--------------------------iteration_setup------------------------------------
duke@0 3457 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
duke@0 3458 if (other != NULL) {
duke@0 3459 grow_to_match(other);
duke@0 3460 // invariant: the finite support of mm2 is within mm->req()
duke@0 3461 #ifdef ASSERT
duke@0 3462 for (uint i = req(); i < other->req(); i++) {
duke@0 3463 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
duke@0 3464 }
duke@0 3465 #endif
duke@0 3466 }
duke@0 3467 // Replace spurious copies of base_memory by top.
duke@0 3468 Node* base_mem = base_memory();
duke@0 3469 if (base_mem != NULL && !base_mem->is_top()) {
duke@0 3470 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
duke@0 3471 if (in(i) == base_mem)
duke@0 3472 set_req(i, empty_memory());
duke@0 3473 }
duke@0 3474 }
duke@0 3475 }
duke@0 3476
duke@0 3477 //---------------------------grow_to_match-------------------------------------
duke@0 3478 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
duke@0 3479 Node* empty_mem = empty_memory();
duke@0 3480 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
duke@0 3481 // look for the finite support of the other memory
duke@0 3482 for (uint i = other->req(); --i >= req(); ) {
duke@0 3483 if (other->in(i) != empty_mem) {
duke@0 3484 uint new_len = i+1;
duke@0 3485 while (req() < new_len) add_req(empty_mem);
duke@0 3486 break;
duke@0 3487 }
duke@0 3488 }
duke@0 3489 }
duke@0 3490
duke@0 3491 //---------------------------verify_sparse-------------------------------------
duke@0 3492 #ifndef PRODUCT
duke@0 3493 bool MergeMemNode::verify_sparse() const {
duke@0 3494 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
duke@0 3495 Node* base_mem = base_memory();
duke@0 3496 // The following can happen in degenerate cases, since empty==top.
duke@0 3497 if (is_empty_memory(base_mem)) return true;
duke@0 3498 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3499 assert(in(i) != NULL, "sane slice");
duke@0 3500 if (in(i) == base_mem) return false; // should have been the sentinel value!
duke@0 3501 }
duke@0 3502 return true;
duke@0 3503 }
duke@0 3504
duke@0 3505 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
duke@0 3506 Node* n;
duke@0 3507 n = mm->in(idx);
duke@0 3508 if (mem == n) return true; // might be empty_memory()
duke@0 3509 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
duke@0 3510 if (mem == n) return true;
duke@0 3511 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
duke@0 3512 if (mem == n) return true;
duke@0 3513 if (n == NULL) break;
duke@0 3514 }
duke@0 3515 return false;
duke@0 3516 }
duke@0 3517 #endif // !PRODUCT