annotate src/share/vm/opto/memnode.cpp @ 708:f2049ae95c3d

6711117: Assertion in 64bit server vm (flat != TypePtr::BOTTOM,"cannot alias-analyze an untyped ptr") Summary: Delay a memory node transformation if its control or address on IGVN worklist. Reviewed-by: never
author kvn
date Tue, 07 Apr 2009 19:04:24 -0700
parents fbde8ec322d0
children 685e959d09ea
rev   line source
duke@0 1 /*
twisti@624 2 * Copyright 1997-2009 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
kvn@74 32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
kvn@74 33
duke@0 34 //=============================================================================
duke@0 35 uint MemNode::size_of() const { return sizeof(*this); }
duke@0 36
duke@0 37 const TypePtr *MemNode::adr_type() const {
duke@0 38 Node* adr = in(Address);
duke@0 39 const TypePtr* cross_check = NULL;
duke@0 40 DEBUG_ONLY(cross_check = _adr_type);
duke@0 41 return calculate_adr_type(adr->bottom_type(), cross_check);
duke@0 42 }
duke@0 43
duke@0 44 #ifndef PRODUCT
duke@0 45 void MemNode::dump_spec(outputStream *st) const {
duke@0 46 if (in(Address) == NULL) return; // node is dead
duke@0 47 #ifndef ASSERT
duke@0 48 // fake the missing field
duke@0 49 const TypePtr* _adr_type = NULL;
duke@0 50 if (in(Address) != NULL)
duke@0 51 _adr_type = in(Address)->bottom_type()->isa_ptr();
duke@0 52 #endif
duke@0 53 dump_adr_type(this, _adr_type, st);
duke@0 54
duke@0 55 Compile* C = Compile::current();
duke@0 56 if( C->alias_type(_adr_type)->is_volatile() )
duke@0 57 st->print(" Volatile!");
duke@0 58 }
duke@0 59
duke@0 60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
duke@0 61 st->print(" @");
duke@0 62 if (adr_type == NULL) {
duke@0 63 st->print("NULL");
duke@0 64 } else {
duke@0 65 adr_type->dump_on(st);
duke@0 66 Compile* C = Compile::current();
duke@0 67 Compile::AliasType* atp = NULL;
duke@0 68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
duke@0 69 if (atp == NULL)
duke@0 70 st->print(", idx=?\?;");
duke@0 71 else if (atp->index() == Compile::AliasIdxBot)
duke@0 72 st->print(", idx=Bot;");
duke@0 73 else if (atp->index() == Compile::AliasIdxTop)
duke@0 74 st->print(", idx=Top;");
duke@0 75 else if (atp->index() == Compile::AliasIdxRaw)
duke@0 76 st->print(", idx=Raw;");
duke@0 77 else {
duke@0 78 ciField* field = atp->field();
duke@0 79 if (field) {
duke@0 80 st->print(", name=");
duke@0 81 field->print_name_on(st);
duke@0 82 }
duke@0 83 st->print(", idx=%d;", atp->index());
duke@0 84 }
duke@0 85 }
duke@0 86 }
duke@0 87
duke@0 88 extern void print_alias_types();
duke@0 89
duke@0 90 #endif
duke@0 91
kvn@74 92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
kvn@74 93 const TypeOopPtr *tinst = t_adr->isa_oopptr();
kvn@223 94 if (tinst == NULL || !tinst->is_known_instance_field())
kvn@74 95 return mchain; // don't try to optimize non-instance types
kvn@74 96 uint instance_id = tinst->instance_id();
kvn@253 97 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
kvn@74 98 Node *prev = NULL;
kvn@74 99 Node *result = mchain;
kvn@74 100 while (prev != result) {
kvn@74 101 prev = result;
kvn@253 102 if (result == start_mem)
twisti@605 103 break; // hit one of our sentinels
kvn@74 104 // skip over a call which does not affect this memory slice
kvn@74 105 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
kvn@74 106 Node *proj_in = result->in(0);
kvn@253 107 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
twisti@605 108 break; // hit one of our sentinels
kvn@253 109 } else if (proj_in->is_Call()) {
kvn@74 110 CallNode *call = proj_in->as_Call();
kvn@74 111 if (!call->may_modify(t_adr, phase)) {
kvn@74 112 result = call->in(TypeFunc::Memory);
kvn@74 113 }
kvn@74 114 } else if (proj_in->is_Initialize()) {
kvn@74 115 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
kvn@74 116 // Stop if this is the initialization for the object instance which
kvn@74 117 // which contains this memory slice, otherwise skip over it.
kvn@74 118 if (alloc != NULL && alloc->_idx != instance_id) {
kvn@74 119 result = proj_in->in(TypeFunc::Memory);
kvn@74 120 }
kvn@74 121 } else if (proj_in->is_MemBar()) {
kvn@74 122 result = proj_in->in(TypeFunc::Memory);
kvn@253 123 } else {
kvn@253 124 assert(false, "unexpected projection");
kvn@74 125 }
kvn@74 126 } else if (result->is_MergeMem()) {
kvn@74 127 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
kvn@74 128 }
kvn@74 129 }
kvn@74 130 return result;
kvn@74 131 }
kvn@74 132
kvn@74 133 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
kvn@74 134 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
kvn@223 135 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
kvn@74 136 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@74 137 Node *result = mchain;
kvn@74 138 result = optimize_simple_memory_chain(result, t_adr, phase);
kvn@74 139 if (is_instance && igvn != NULL && result->is_Phi()) {
kvn@74 140 PhiNode *mphi = result->as_Phi();
kvn@74 141 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
kvn@74 142 const TypePtr *t = mphi->adr_type();
kvn@163 143 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
kvn@223 144 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
kvn@247 145 t->is_oopptr()->cast_to_exactness(true)
kvn@247 146 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
kvn@247 147 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
kvn@74 148 // clone the Phi with our address type
kvn@74 149 result = mphi->split_out_instance(t_adr, igvn);
kvn@74 150 } else {
kvn@74 151 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
kvn@74 152 }
kvn@74 153 }
kvn@74 154 return result;
kvn@74 155 }
kvn@74 156
kvn@64 157 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
kvn@64 158 uint alias_idx = phase->C->get_alias_index(tp);
kvn@64 159 Node *mem = mmem;
kvn@64 160 #ifdef ASSERT
kvn@64 161 {
kvn@64 162 // Check that current type is consistent with the alias index used during graph construction
kvn@64 163 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
kvn@64 164 bool consistent = adr_check == NULL || adr_check->empty() ||
kvn@64 165 phase->C->must_alias(adr_check, alias_idx );
kvn@64 166 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
kvn@64 167 if( !consistent && adr_check != NULL && !adr_check->empty() &&
rasbold@169 168 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
kvn@64 169 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
kvn@64 170 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
kvn@64 171 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
kvn@64 172 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
kvn@64 173 // don't assert if it is dead code.
kvn@64 174 consistent = true;
kvn@64 175 }
kvn@64 176 if( !consistent ) {
kvn@64 177 st->print("alias_idx==%d, adr_check==", alias_idx);
kvn@64 178 if( adr_check == NULL ) {
kvn@64 179 st->print("NULL");
kvn@64 180 } else {
kvn@64 181 adr_check->dump();
kvn@64 182 }
kvn@64 183 st->cr();
kvn@64 184 print_alias_types();
kvn@64 185 assert(consistent, "adr_check must match alias idx");
kvn@64 186 }
kvn@64 187 }
kvn@64 188 #endif
kvn@64 189 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
kvn@64 190 // means an array I have not precisely typed yet. Do not do any
kvn@64 191 // alias stuff with it any time soon.
kvn@64 192 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@64 193 if( tp->base() != Type::AnyPtr &&
kvn@64 194 !(tinst &&
kvn@64 195 tinst->klass()->is_java_lang_Object() &&
kvn@64 196 tinst->offset() == Type::OffsetBot) ) {
kvn@64 197 // compress paths and change unreachable cycles to TOP
kvn@64 198 // If not, we can update the input infinitely along a MergeMem cycle
kvn@64 199 // Equivalent code in PhiNode::Ideal
kvn@64 200 Node* m = phase->transform(mmem);
twisti@605 201 // If transformed to a MergeMem, get the desired slice
kvn@64 202 // Otherwise the returned node represents memory for every slice
kvn@64 203 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
kvn@64 204 // Update input if it is progress over what we have now
kvn@64 205 }
kvn@64 206 return mem;
kvn@64 207 }
kvn@64 208
duke@0 209 //--------------------------Ideal_common---------------------------------------
duke@0 210 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
duke@0 211 // Unhook non-raw memories from complete (macro-expanded) initializations.
duke@0 212 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
duke@0 213 // If our control input is a dead region, kill all below the region
duke@0 214 Node *ctl = in(MemNode::Control);
duke@0 215 if (ctl && remove_dead_region(phase, can_reshape))
duke@0 216 return this;
kvn@305 217 ctl = in(MemNode::Control);
kvn@305 218 // Don't bother trying to transform a dead node
kvn@305 219 if( ctl && ctl->is_top() ) return NodeSentinel;
duke@0 220
kvn@708 221 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@708 222 // Wait if control on the worklist.
kvn@708 223 if (ctl && can_reshape && igvn != NULL) {
kvn@708 224 Node* bol = NULL;
kvn@708 225 Node* cmp = NULL;
kvn@708 226 if (ctl->in(0)->is_If()) {
kvn@708 227 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
kvn@708 228 bol = ctl->in(0)->in(1);
kvn@708 229 if (bol->is_Bool())
kvn@708 230 cmp = ctl->in(0)->in(1)->in(1);
kvn@708 231 }
kvn@708 232 if (igvn->_worklist.member(ctl) ||
kvn@708 233 (bol != NULL && igvn->_worklist.member(bol)) ||
kvn@708 234 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
kvn@708 235 // This control path may be dead.
kvn@708 236 // Delay this memory node transformation until the control is processed.
kvn@708 237 phase->is_IterGVN()->_worklist.push(this);
kvn@708 238 return NodeSentinel; // caller will return NULL
kvn@708 239 }
kvn@708 240 }
duke@0 241 // Ignore if memory is dead, or self-loop
duke@0 242 Node *mem = in(MemNode::Memory);
duke@0 243 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 244 assert( mem != this, "dead loop in MemNode::Ideal" );
duke@0 245
duke@0 246 Node *address = in(MemNode::Address);
duke@0 247 const Type *t_adr = phase->type( address );
duke@0 248 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 249
kvn@708 250 if( can_reshape && igvn != NULL &&
kvn@708 251 (igvn->_worklist.member(address) || phase->type(address) != adr_type()) ) {
kvn@420 252 // The address's base and type may change when the address is processed.
kvn@420 253 // Delay this mem node transformation until the address is processed.
kvn@420 254 phase->is_IterGVN()->_worklist.push(this);
kvn@420 255 return NodeSentinel; // caller will return NULL
kvn@420 256 }
kvn@420 257
kvn@708 258 #ifdef ASSERT
kvn@708 259 Node* base = NULL;
kvn@708 260 if (address->is_AddP())
kvn@708 261 base = address->in(AddPNode::Base);
kvn@708 262 assert(base == NULL || t_adr->isa_rawptr() ||
kvn@708 263 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
kvn@708 264 #endif
kvn@708 265
duke@0 266 // Avoid independent memory operations
duke@0 267 Node* old_mem = mem;
duke@0 268
kvn@36 269 // The code which unhooks non-raw memories from complete (macro-expanded)
kvn@36 270 // initializations was removed. After macro-expansion all stores catched
kvn@36 271 // by Initialize node became raw stores and there is no information
kvn@36 272 // which memory slices they modify. So it is unsafe to move any memory
kvn@36 273 // operation above these stores. Also in most cases hooked non-raw memories
kvn@36 274 // were already unhooked by using information from detect_ptr_independence()
kvn@36 275 // and find_previous_store().
duke@0 276
duke@0 277 if (mem->is_MergeMem()) {
duke@0 278 MergeMemNode* mmem = mem->as_MergeMem();
duke@0 279 const TypePtr *tp = t_adr->is_ptr();
kvn@64 280
kvn@64 281 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
duke@0 282 }
duke@0 283
duke@0 284 if (mem != old_mem) {
duke@0 285 set_req(MemNode::Memory, mem);
kvn@305 286 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
duke@0 287 return this;
duke@0 288 }
duke@0 289
duke@0 290 // let the subclass continue analyzing...
duke@0 291 return NULL;
duke@0 292 }
duke@0 293
duke@0 294 // Helper function for proving some simple control dominations.
kvn@119 295 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
duke@0 296 // Already assumes that 'dom' is available at 'sub', and that 'sub'
duke@0 297 // is not a constant (dominated by the method's StartNode).
duke@0 298 // Used by MemNode::find_previous_store to prove that the
duke@0 299 // control input of a memory operation predates (dominates)
duke@0 300 // an allocation it wants to look past.
kvn@119 301 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
kvn@119 302 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
kvn@119 303 return false; // Conservative answer for dead code
kvn@119 304
kvn@193 305 // Check 'dom'. Skip Proj and CatchProj nodes.
kvn@119 306 dom = dom->find_exact_control(dom);
kvn@119 307 if (dom == NULL || dom->is_top())
kvn@119 308 return false; // Conservative answer for dead code
kvn@119 309
kvn@193 310 if (dom == sub) {
kvn@193 311 // For the case when, for example, 'sub' is Initialize and the original
kvn@193 312 // 'dom' is Proj node of the 'sub'.
kvn@193 313 return false;
kvn@193 314 }
kvn@193 315
kvn@155 316 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
kvn@119 317 return true;
kvn@119 318
kvn@119 319 // 'dom' dominates 'sub' if its control edge and control edges
kvn@119 320 // of all its inputs dominate or equal to sub's control edge.
kvn@119 321
kvn@119 322 // Currently 'sub' is either Allocate, Initialize or Start nodes.
kvn@163 323 // Or Region for the check in LoadNode::Ideal();
kvn@163 324 // 'sub' should have sub->in(0) != NULL.
kvn@163 325 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
kvn@163 326 sub->is_Region(), "expecting only these nodes");
kvn@119 327
kvn@119 328 // Get control edge of 'sub'.
kvn@193 329 Node* orig_sub = sub;
kvn@119 330 sub = sub->find_exact_control(sub->in(0));
kvn@119 331 if (sub == NULL || sub->is_top())
kvn@119 332 return false; // Conservative answer for dead code
kvn@119 333
kvn@119 334 assert(sub->is_CFG(), "expecting control");
kvn@119 335
kvn@119 336 if (sub == dom)
kvn@119 337 return true;
kvn@119 338
kvn@119 339 if (sub->is_Start() || sub->is_Root())
kvn@119 340 return false;
kvn@119 341
kvn@119 342 {
kvn@119 343 // Check all control edges of 'dom'.
kvn@119 344
kvn@119 345 ResourceMark rm;
kvn@119 346 Arena* arena = Thread::current()->resource_area();
kvn@119 347 Node_List nlist(arena);
kvn@119 348 Unique_Node_List dom_list(arena);
kvn@119 349
kvn@119 350 dom_list.push(dom);
kvn@119 351 bool only_dominating_controls = false;
kvn@119 352
kvn@119 353 for (uint next = 0; next < dom_list.size(); next++) {
kvn@119 354 Node* n = dom_list.at(next);
kvn@193 355 if (n == orig_sub)
kvn@193 356 return false; // One of dom's inputs dominated by sub.
kvn@119 357 if (!n->is_CFG() && n->pinned()) {
kvn@119 358 // Check only own control edge for pinned non-control nodes.
kvn@119 359 n = n->find_exact_control(n->in(0));
kvn@119 360 if (n == NULL || n->is_top())
kvn@119 361 return false; // Conservative answer for dead code
kvn@119 362 assert(n->is_CFG(), "expecting control");
kvn@193 363 dom_list.push(n);
kvn@193 364 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
kvn@119 365 only_dominating_controls = true;
kvn@119 366 } else if (n->is_CFG()) {
kvn@119 367 if (n->dominates(sub, nlist))
kvn@119 368 only_dominating_controls = true;
kvn@119 369 else
kvn@119 370 return false;
kvn@119 371 } else {
kvn@119 372 // First, own control edge.
kvn@119 373 Node* m = n->find_exact_control(n->in(0));
kvn@155 374 if (m != NULL) {
kvn@155 375 if (m->is_top())
kvn@155 376 return false; // Conservative answer for dead code
kvn@155 377 dom_list.push(m);
kvn@155 378 }
kvn@119 379 // Now, the rest of edges.
kvn@119 380 uint cnt = n->req();
kvn@119 381 for (uint i = 1; i < cnt; i++) {
kvn@119 382 m = n->find_exact_control(n->in(i));
kvn@119 383 if (m == NULL || m->is_top())
kvn@119 384 continue;
kvn@119 385 dom_list.push(m);
duke@0 386 }
duke@0 387 }
duke@0 388 }
kvn@119 389 return only_dominating_controls;
duke@0 390 }
duke@0 391 }
duke@0 392
duke@0 393 //---------------------detect_ptr_independence---------------------------------
duke@0 394 // Used by MemNode::find_previous_store to prove that two base
duke@0 395 // pointers are never equal.
duke@0 396 // The pointers are accompanied by their associated allocations,
duke@0 397 // if any, which have been previously discovered by the caller.
duke@0 398 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
duke@0 399 Node* p2, AllocateNode* a2,
duke@0 400 PhaseTransform* phase) {
duke@0 401 // Attempt to prove that these two pointers cannot be aliased.
duke@0 402 // They may both manifestly be allocations, and they should differ.
duke@0 403 // Or, if they are not both allocations, they can be distinct constants.
duke@0 404 // Otherwise, one is an allocation and the other a pre-existing value.
duke@0 405 if (a1 == NULL && a2 == NULL) { // neither an allocation
duke@0 406 return (p1 != p2) && p1->is_Con() && p2->is_Con();
duke@0 407 } else if (a1 != NULL && a2 != NULL) { // both allocations
duke@0 408 return (a1 != a2);
duke@0 409 } else if (a1 != NULL) { // one allocation a1
duke@0 410 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
kvn@119 411 return all_controls_dominate(p2, a1);
duke@0 412 } else { //(a2 != NULL) // one allocation a2
kvn@119 413 return all_controls_dominate(p1, a2);
duke@0 414 }
duke@0 415 return false;
duke@0 416 }
duke@0 417
duke@0 418
duke@0 419 // The logic for reordering loads and stores uses four steps:
duke@0 420 // (a) Walk carefully past stores and initializations which we
duke@0 421 // can prove are independent of this load.
duke@0 422 // (b) Observe that the next memory state makes an exact match
duke@0 423 // with self (load or store), and locate the relevant store.
duke@0 424 // (c) Ensure that, if we were to wire self directly to the store,
duke@0 425 // the optimizer would fold it up somehow.
duke@0 426 // (d) Do the rewiring, and return, depending on some other part of
duke@0 427 // the optimizer to fold up the load.
duke@0 428 // This routine handles steps (a) and (b). Steps (c) and (d) are
duke@0 429 // specific to loads and stores, so they are handled by the callers.
duke@0 430 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
duke@0 431 //
duke@0 432 Node* MemNode::find_previous_store(PhaseTransform* phase) {
duke@0 433 Node* ctrl = in(MemNode::Control);
duke@0 434 Node* adr = in(MemNode::Address);
duke@0 435 intptr_t offset = 0;
duke@0 436 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 437 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
duke@0 438
duke@0 439 if (offset == Type::OffsetBot)
duke@0 440 return NULL; // cannot unalias unless there are precise offsets
duke@0 441
kvn@74 442 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
kvn@74 443
duke@0 444 intptr_t size_in_bytes = memory_size();
duke@0 445
duke@0 446 Node* mem = in(MemNode::Memory); // start searching here...
duke@0 447
duke@0 448 int cnt = 50; // Cycle limiter
duke@0 449 for (;;) { // While we can dance past unrelated stores...
duke@0 450 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
duke@0 451
duke@0 452 if (mem->is_Store()) {
duke@0 453 Node* st_adr = mem->in(MemNode::Address);
duke@0 454 intptr_t st_offset = 0;
duke@0 455 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
duke@0 456 if (st_base == NULL)
duke@0 457 break; // inscrutable pointer
duke@0 458 if (st_offset != offset && st_offset != Type::OffsetBot) {
duke@0 459 const int MAX_STORE = BytesPerLong;
duke@0 460 if (st_offset >= offset + size_in_bytes ||
duke@0 461 st_offset <= offset - MAX_STORE ||
duke@0 462 st_offset <= offset - mem->as_Store()->memory_size()) {
duke@0 463 // Success: The offsets are provably independent.
duke@0 464 // (You may ask, why not just test st_offset != offset and be done?
duke@0 465 // The answer is that stores of different sizes can co-exist
duke@0 466 // in the same sequence of RawMem effects. We sometimes initialize
duke@0 467 // a whole 'tile' of array elements with a single jint or jlong.)
duke@0 468 mem = mem->in(MemNode::Memory);
duke@0 469 continue; // (a) advance through independent store memory
duke@0 470 }
duke@0 471 }
duke@0 472 if (st_base != base &&
duke@0 473 detect_ptr_independence(base, alloc,
duke@0 474 st_base,
duke@0 475 AllocateNode::Ideal_allocation(st_base, phase),
duke@0 476 phase)) {
duke@0 477 // Success: The bases are provably independent.
duke@0 478 mem = mem->in(MemNode::Memory);
duke@0 479 continue; // (a) advance through independent store memory
duke@0 480 }
duke@0 481
duke@0 482 // (b) At this point, if the bases or offsets do not agree, we lose,
duke@0 483 // since we have not managed to prove 'this' and 'mem' independent.
duke@0 484 if (st_base == base && st_offset == offset) {
duke@0 485 return mem; // let caller handle steps (c), (d)
duke@0 486 }
duke@0 487
duke@0 488 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 489 InitializeNode* st_init = mem->in(0)->as_Initialize();
duke@0 490 AllocateNode* st_alloc = st_init->allocation();
duke@0 491 if (st_alloc == NULL)
duke@0 492 break; // something degenerated
duke@0 493 bool known_identical = false;
duke@0 494 bool known_independent = false;
duke@0 495 if (alloc == st_alloc)
duke@0 496 known_identical = true;
duke@0 497 else if (alloc != NULL)
duke@0 498 known_independent = true;
kvn@119 499 else if (all_controls_dominate(this, st_alloc))
duke@0 500 known_independent = true;
duke@0 501
duke@0 502 if (known_independent) {
duke@0 503 // The bases are provably independent: Either they are
duke@0 504 // manifestly distinct allocations, or else the control
duke@0 505 // of this load dominates the store's allocation.
duke@0 506 int alias_idx = phase->C->get_alias_index(adr_type());
duke@0 507 if (alias_idx == Compile::AliasIdxRaw) {
duke@0 508 mem = st_alloc->in(TypeFunc::Memory);
duke@0 509 } else {
duke@0 510 mem = st_init->memory(alias_idx);
duke@0 511 }
duke@0 512 continue; // (a) advance through independent store memory
duke@0 513 }
duke@0 514
duke@0 515 // (b) at this point, if we are not looking at a store initializing
duke@0 516 // the same allocation we are loading from, we lose.
duke@0 517 if (known_identical) {
duke@0 518 // From caller, can_see_stored_value will consult find_captured_store.
duke@0 519 return mem; // let caller handle steps (c), (d)
duke@0 520 }
duke@0 521
kvn@223 522 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
kvn@74 523 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
kvn@74 524 if (mem->is_Proj() && mem->in(0)->is_Call()) {
kvn@74 525 CallNode *call = mem->in(0)->as_Call();
kvn@74 526 if (!call->may_modify(addr_t, phase)) {
kvn@74 527 mem = call->in(TypeFunc::Memory);
kvn@74 528 continue; // (a) advance through independent call memory
kvn@74 529 }
kvn@74 530 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
kvn@74 531 mem = mem->in(0)->in(TypeFunc::Memory);
kvn@74 532 continue; // (a) advance through independent MemBar memory
kvn@74 533 } else if (mem->is_MergeMem()) {
kvn@74 534 int alias_idx = phase->C->get_alias_index(adr_type());
kvn@74 535 mem = mem->as_MergeMem()->memory_at(alias_idx);
kvn@74 536 continue; // (a) advance through independent MergeMem memory
kvn@74 537 }
duke@0 538 }
duke@0 539
duke@0 540 // Unless there is an explicit 'continue', we must bail out here,
duke@0 541 // because 'mem' is an inscrutable memory state (e.g., a call).
duke@0 542 break;
duke@0 543 }
duke@0 544
duke@0 545 return NULL; // bail out
duke@0 546 }
duke@0 547
duke@0 548 //----------------------calculate_adr_type-------------------------------------
duke@0 549 // Helper function. Notices when the given type of address hits top or bottom.
duke@0 550 // Also, asserts a cross-check of the type against the expected address type.
duke@0 551 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
duke@0 552 if (t == Type::TOP) return NULL; // does not touch memory any more?
duke@0 553 #ifdef PRODUCT
duke@0 554 cross_check = NULL;
duke@0 555 #else
duke@0 556 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
duke@0 557 #endif
duke@0 558 const TypePtr* tp = t->isa_ptr();
duke@0 559 if (tp == NULL) {
duke@0 560 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
duke@0 561 return TypePtr::BOTTOM; // touches lots of memory
duke@0 562 } else {
duke@0 563 #ifdef ASSERT
duke@0 564 // %%%% [phh] We don't check the alias index if cross_check is
duke@0 565 // TypeRawPtr::BOTTOM. Needs to be investigated.
duke@0 566 if (cross_check != NULL &&
duke@0 567 cross_check != TypePtr::BOTTOM &&
duke@0 568 cross_check != TypeRawPtr::BOTTOM) {
duke@0 569 // Recheck the alias index, to see if it has changed (due to a bug).
duke@0 570 Compile* C = Compile::current();
duke@0 571 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
duke@0 572 "must stay in the original alias category");
duke@0 573 // The type of the address must be contained in the adr_type,
duke@0 574 // disregarding "null"-ness.
duke@0 575 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
duke@0 576 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
duke@0 577 assert(cross_check->meet(tp_notnull) == cross_check,
duke@0 578 "real address must not escape from expected memory type");
duke@0 579 }
duke@0 580 #endif
duke@0 581 return tp;
duke@0 582 }
duke@0 583 }
duke@0 584
duke@0 585 //------------------------adr_phi_is_loop_invariant----------------------------
duke@0 586 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
duke@0 587 // loop is loop invariant. Make a quick traversal of Phi and associated
duke@0 588 // CastPP nodes, looking to see if they are a closed group within the loop.
duke@0 589 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
duke@0 590 // The idea is that the phi-nest must boil down to only CastPP nodes
duke@0 591 // with the same data. This implies that any path into the loop already
duke@0 592 // includes such a CastPP, and so the original cast, whatever its input,
duke@0 593 // must be covered by an equivalent cast, with an earlier control input.
duke@0 594 ResourceMark rm;
duke@0 595
duke@0 596 // The loop entry input of the phi should be the unique dominating
duke@0 597 // node for every Phi/CastPP in the loop.
duke@0 598 Unique_Node_List closure;
duke@0 599 closure.push(adr_phi->in(LoopNode::EntryControl));
duke@0 600
duke@0 601 // Add the phi node and the cast to the worklist.
duke@0 602 Unique_Node_List worklist;
duke@0 603 worklist.push(adr_phi);
duke@0 604 if( cast != NULL ){
duke@0 605 if( !cast->is_ConstraintCast() ) return false;
duke@0 606 worklist.push(cast);
duke@0 607 }
duke@0 608
duke@0 609 // Begin recursive walk of phi nodes.
duke@0 610 while( worklist.size() ){
duke@0 611 // Take a node off the worklist
duke@0 612 Node *n = worklist.pop();
duke@0 613 if( !closure.member(n) ){
duke@0 614 // Add it to the closure.
duke@0 615 closure.push(n);
duke@0 616 // Make a sanity check to ensure we don't waste too much time here.
duke@0 617 if( closure.size() > 20) return false;
duke@0 618 // This node is OK if:
duke@0 619 // - it is a cast of an identical value
duke@0 620 // - or it is a phi node (then we add its inputs to the worklist)
duke@0 621 // Otherwise, the node is not OK, and we presume the cast is not invariant
duke@0 622 if( n->is_ConstraintCast() ){
duke@0 623 worklist.push(n->in(1));
duke@0 624 } else if( n->is_Phi() ) {
duke@0 625 for( uint i = 1; i < n->req(); i++ ) {
duke@0 626 worklist.push(n->in(i));
duke@0 627 }
duke@0 628 } else {
duke@0 629 return false;
duke@0 630 }
duke@0 631 }
duke@0 632 }
duke@0 633
duke@0 634 // Quit when the worklist is empty, and we've found no offending nodes.
duke@0 635 return true;
duke@0 636 }
duke@0 637
duke@0 638 //------------------------------Ideal_DU_postCCP-------------------------------
duke@0 639 // Find any cast-away of null-ness and keep its control. Null cast-aways are
duke@0 640 // going away in this pass and we need to make this memory op depend on the
duke@0 641 // gating null check.
kvn@163 642 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
kvn@163 643 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
kvn@163 644 }
duke@0 645
duke@0 646 // I tried to leave the CastPP's in. This makes the graph more accurate in
duke@0 647 // some sense; we get to keep around the knowledge that an oop is not-null
duke@0 648 // after some test. Alas, the CastPP's interfere with GVN (some values are
duke@0 649 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
duke@0 650 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
duke@0 651 // some of the more trivial cases in the optimizer. Removing more useless
duke@0 652 // Phi's started allowing Loads to illegally float above null checks. I gave
duke@0 653 // up on this approach. CNC 10/20/2000
kvn@163 654 // This static method may be called not from MemNode (EncodePNode calls it).
kvn@163 655 // Only the control edge of the node 'n' might be updated.
kvn@163 656 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
duke@0 657 Node *skipped_cast = NULL;
duke@0 658 // Need a null check? Regular static accesses do not because they are
duke@0 659 // from constant addresses. Array ops are gated by the range check (which
duke@0 660 // always includes a NULL check). Just check field ops.
kvn@163 661 if( n->in(MemNode::Control) == NULL ) {
duke@0 662 // Scan upwards for the highest location we can place this memory op.
duke@0 663 while( true ) {
duke@0 664 switch( adr->Opcode() ) {
duke@0 665
duke@0 666 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
duke@0 667 adr = adr->in(AddPNode::Base);
duke@0 668 continue;
duke@0 669
coleenp@113 670 case Op_DecodeN: // No change to NULL-ness, so peek thru
coleenp@113 671 adr = adr->in(1);
coleenp@113 672 continue;
coleenp@113 673
duke@0 674 case Op_CastPP:
duke@0 675 // If the CastPP is useless, just peek on through it.
duke@0 676 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
duke@0 677 // Remember the cast that we've peeked though. If we peek
duke@0 678 // through more than one, then we end up remembering the highest
duke@0 679 // one, that is, if in a loop, the one closest to the top.
duke@0 680 skipped_cast = adr;
duke@0 681 adr = adr->in(1);
duke@0 682 continue;
duke@0 683 }
duke@0 684 // CastPP is going away in this pass! We need this memory op to be
duke@0 685 // control-dependent on the test that is guarding the CastPP.
kvn@163 686 ccp->hash_delete(n);
kvn@163 687 n->set_req(MemNode::Control, adr->in(0));
kvn@163 688 ccp->hash_insert(n);
kvn@163 689 return n;
duke@0 690
duke@0 691 case Op_Phi:
duke@0 692 // Attempt to float above a Phi to some dominating point.
duke@0 693 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
duke@0 694 // If we've already peeked through a Cast (which could have set the
duke@0 695 // control), we can't float above a Phi, because the skipped Cast
duke@0 696 // may not be loop invariant.
duke@0 697 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
duke@0 698 adr = adr->in(1);
duke@0 699 continue;
duke@0 700 }
duke@0 701 }
duke@0 702
duke@0 703 // Intentional fallthrough!
duke@0 704
duke@0 705 // No obvious dominating point. The mem op is pinned below the Phi
duke@0 706 // by the Phi itself. If the Phi goes away (no true value is merged)
duke@0 707 // then the mem op can float, but not indefinitely. It must be pinned
duke@0 708 // behind the controls leading to the Phi.
duke@0 709 case Op_CheckCastPP:
duke@0 710 // These usually stick around to change address type, however a
duke@0 711 // useless one can be elided and we still need to pick up a control edge
duke@0 712 if (adr->in(0) == NULL) {
duke@0 713 // This CheckCastPP node has NO control and is likely useless. But we
duke@0 714 // need check further up the ancestor chain for a control input to keep
duke@0 715 // the node in place. 4959717.
duke@0 716 skipped_cast = adr;
duke@0 717 adr = adr->in(1);
duke@0 718 continue;
duke@0 719 }
kvn@163 720 ccp->hash_delete(n);
kvn@163 721 n->set_req(MemNode::Control, adr->in(0));
kvn@163 722 ccp->hash_insert(n);
kvn@163 723 return n;
duke@0 724
duke@0 725 // List of "safe" opcodes; those that implicitly block the memory
duke@0 726 // op below any null check.
duke@0 727 case Op_CastX2P: // no null checks on native pointers
duke@0 728 case Op_Parm: // 'this' pointer is not null
duke@0 729 case Op_LoadP: // Loading from within a klass
coleenp@113 730 case Op_LoadN: // Loading from within a klass
duke@0 731 case Op_LoadKlass: // Loading from within a klass
kvn@164 732 case Op_LoadNKlass: // Loading from within a klass
duke@0 733 case Op_ConP: // Loading from a klass
kvn@163 734 case Op_ConN: // Loading from a klass
duke@0 735 case Op_CreateEx: // Sucking up the guts of an exception oop
duke@0 736 case Op_Con: // Reading from TLS
duke@0 737 case Op_CMoveP: // CMoveP is pinned
kvn@164 738 case Op_CMoveN: // CMoveN is pinned
duke@0 739 break; // No progress
duke@0 740
duke@0 741 case Op_Proj: // Direct call to an allocation routine
duke@0 742 case Op_SCMemProj: // Memory state from store conditional ops
duke@0 743 #ifdef ASSERT
duke@0 744 {
duke@0 745 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
duke@0 746 const Node* call = adr->in(0);
kvn@163 747 if (call->is_CallJava()) {
kvn@163 748 const CallJavaNode* call_java = call->as_CallJava();
kvn@64 749 const TypeTuple *r = call_java->tf()->range();
kvn@64 750 assert(r->cnt() > TypeFunc::Parms, "must return value");
kvn@64 751 const Type* ret_type = r->field_at(TypeFunc::Parms);
kvn@64 752 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
duke@0 753 // We further presume that this is one of
duke@0 754 // new_instance_Java, new_array_Java, or
duke@0 755 // the like, but do not assert for this.
duke@0 756 } else if (call->is_Allocate()) {
duke@0 757 // similar case to new_instance_Java, etc.
duke@0 758 } else if (!call->is_CallLeaf()) {
duke@0 759 // Projections from fetch_oop (OSR) are allowed as well.
duke@0 760 ShouldNotReachHere();
duke@0 761 }
duke@0 762 }
duke@0 763 #endif
duke@0 764 break;
duke@0 765 default:
duke@0 766 ShouldNotReachHere();
duke@0 767 }
duke@0 768 break;
duke@0 769 }
duke@0 770 }
duke@0 771
duke@0 772 return NULL; // No progress
duke@0 773 }
duke@0 774
duke@0 775
duke@0 776 //=============================================================================
duke@0 777 uint LoadNode::size_of() const { return sizeof(*this); }
duke@0 778 uint LoadNode::cmp( const Node &n ) const
duke@0 779 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
duke@0 780 const Type *LoadNode::bottom_type() const { return _type; }
duke@0 781 uint LoadNode::ideal_reg() const {
duke@0 782 return Matcher::base2reg[_type->base()];
duke@0 783 }
duke@0 784
duke@0 785 #ifndef PRODUCT
duke@0 786 void LoadNode::dump_spec(outputStream *st) const {
duke@0 787 MemNode::dump_spec(st);
duke@0 788 if( !Verbose && !WizardMode ) {
duke@0 789 // standard dump does this in Verbose and WizardMode
duke@0 790 st->print(" #"); _type->dump_on(st);
duke@0 791 }
duke@0 792 }
duke@0 793 #endif
duke@0 794
duke@0 795
duke@0 796 //----------------------------LoadNode::make-----------------------------------
duke@0 797 // Polymorphic factory method:
coleenp@113 798 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
coleenp@113 799 Compile* C = gvn.C;
coleenp@113 800
duke@0 801 // sanity check the alias category against the created node type
duke@0 802 assert(!(adr_type->isa_oopptr() &&
duke@0 803 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
duke@0 804 "use LoadKlassNode instead");
duke@0 805 assert(!(adr_type->isa_aryptr() &&
duke@0 806 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
duke@0 807 "use LoadRangeNode instead");
duke@0 808 switch (bt) {
twisti@624 809 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 810 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 811 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 812 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 813 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() );
twisti@558 814 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() );
twisti@558 815 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt );
twisti@558 816 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt );
twisti@558 817 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() );
coleenp@113 818 case T_OBJECT:
coleenp@113 819 #ifdef _LP64
kvn@163 820 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
kvn@221 821 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
kvn@221 822 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
coleenp@113 823 } else
coleenp@113 824 #endif
kvn@163 825 {
kvn@163 826 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
kvn@163 827 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
kvn@163 828 }
duke@0 829 }
duke@0 830 ShouldNotReachHere();
duke@0 831 return (LoadNode*)NULL;
duke@0 832 }
duke@0 833
duke@0 834 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
duke@0 835 bool require_atomic = true;
duke@0 836 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
duke@0 837 }
duke@0 838
duke@0 839
duke@0 840
duke@0 841
duke@0 842 //------------------------------hash-------------------------------------------
duke@0 843 uint LoadNode::hash() const {
duke@0 844 // unroll addition of interesting fields
duke@0 845 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
duke@0 846 }
duke@0 847
duke@0 848 //---------------------------can_see_stored_value------------------------------
duke@0 849 // This routine exists to make sure this set of tests is done the same
duke@0 850 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
duke@0 851 // will change the graph shape in a way which makes memory alive twice at the
duke@0 852 // same time (uses the Oracle model of aliasing), then some
duke@0 853 // LoadXNode::Identity will fold things back to the equivalence-class model
duke@0 854 // of aliasing.
duke@0 855 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
duke@0 856 Node* ld_adr = in(MemNode::Address);
duke@0 857
never@17 858 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
never@17 859 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
never@17 860 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
never@17 861 atp->field() != NULL && !atp->field()->is_volatile()) {
never@17 862 uint alias_idx = atp->index();
never@17 863 bool final = atp->field()->is_final();
never@17 864 Node* result = NULL;
never@17 865 Node* current = st;
never@17 866 // Skip through chains of MemBarNodes checking the MergeMems for
never@17 867 // new states for the slice of this load. Stop once any other
never@17 868 // kind of node is encountered. Loads from final memory can skip
never@17 869 // through any kind of MemBar but normal loads shouldn't skip
never@17 870 // through MemBarAcquire since the could allow them to move out of
never@17 871 // a synchronized region.
never@17 872 while (current->is_Proj()) {
never@17 873 int opc = current->in(0)->Opcode();
never@17 874 if ((final && opc == Op_MemBarAcquire) ||
never@17 875 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
never@17 876 Node* mem = current->in(0)->in(TypeFunc::Memory);
never@17 877 if (mem->is_MergeMem()) {
never@17 878 MergeMemNode* merge = mem->as_MergeMem();
never@17 879 Node* new_st = merge->memory_at(alias_idx);
never@17 880 if (new_st == merge->base_memory()) {
never@17 881 // Keep searching
never@17 882 current = merge->base_memory();
never@17 883 continue;
never@17 884 }
never@17 885 // Save the new memory state for the slice and fall through
never@17 886 // to exit.
never@17 887 result = new_st;
never@17 888 }
never@17 889 }
never@17 890 break;
never@17 891 }
never@17 892 if (result != NULL) {
never@17 893 st = result;
never@17 894 }
never@17 895 }
never@17 896
never@17 897
duke@0 898 // Loop around twice in the case Load -> Initialize -> Store.
duke@0 899 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
duke@0 900 for (int trip = 0; trip <= 1; trip++) {
duke@0 901
duke@0 902 if (st->is_Store()) {
duke@0 903 Node* st_adr = st->in(MemNode::Address);
duke@0 904 if (!phase->eqv(st_adr, ld_adr)) {
duke@0 905 // Try harder before giving up... Match raw and non-raw pointers.
duke@0 906 intptr_t st_off = 0;
duke@0 907 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
duke@0 908 if (alloc == NULL) return NULL;
duke@0 909 intptr_t ld_off = 0;
duke@0 910 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
duke@0 911 if (alloc != allo2) return NULL;
duke@0 912 if (ld_off != st_off) return NULL;
duke@0 913 // At this point we have proven something like this setup:
duke@0 914 // A = Allocate(...)
duke@0 915 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
duke@0 916 // S = StoreQ(, AddP(, A.Parm , #Off), V)
duke@0 917 // (Actually, we haven't yet proven the Q's are the same.)
duke@0 918 // In other words, we are loading from a casted version of
duke@0 919 // the same pointer-and-offset that we stored to.
duke@0 920 // Thus, we are able to replace L by V.
duke@0 921 }
duke@0 922 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
duke@0 923 if (store_Opcode() != st->Opcode())
duke@0 924 return NULL;
duke@0 925 return st->in(MemNode::ValueIn);
duke@0 926 }
duke@0 927
duke@0 928 intptr_t offset = 0; // scratch
duke@0 929
duke@0 930 // A load from a freshly-created object always returns zero.
duke@0 931 // (This can happen after LoadNode::Ideal resets the load's memory input
duke@0 932 // to find_captured_store, which returned InitializeNode::zero_memory.)
duke@0 933 if (st->is_Proj() && st->in(0)->is_Allocate() &&
duke@0 934 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
duke@0 935 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
duke@0 936 // return a zero value for the load's basic type
duke@0 937 // (This is one of the few places where a generic PhaseTransform
duke@0 938 // can create new nodes. Think of it as lazily manifesting
duke@0 939 // virtually pre-existing constants.)
duke@0 940 return phase->zerocon(memory_type());
duke@0 941 }
duke@0 942
duke@0 943 // A load from an initialization barrier can match a captured store.
duke@0 944 if (st->is_Proj() && st->in(0)->is_Initialize()) {
duke@0 945 InitializeNode* init = st->in(0)->as_Initialize();
duke@0 946 AllocateNode* alloc = init->allocation();
duke@0 947 if (alloc != NULL &&
duke@0 948 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
duke@0 949 // examine a captured store value
duke@0 950 st = init->find_captured_store(offset, memory_size(), phase);
duke@0 951 if (st != NULL)
duke@0 952 continue; // take one more trip around
duke@0 953 }
duke@0 954 }
duke@0 955
duke@0 956 break;
duke@0 957 }
duke@0 958
duke@0 959 return NULL;
duke@0 960 }
duke@0 961
kvn@64 962 //----------------------is_instance_field_load_with_local_phi------------------
kvn@64 963 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
kvn@64 964 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
kvn@64 965 in(MemNode::Address)->is_AddP() ) {
kvn@64 966 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
kvn@64 967 // Only instances.
kvn@223 968 if( t_oop != NULL && t_oop->is_known_instance_field() &&
kvn@64 969 t_oop->offset() != Type::OffsetBot &&
kvn@64 970 t_oop->offset() != Type::OffsetTop) {
kvn@64 971 return true;
kvn@64 972 }
kvn@64 973 }
kvn@64 974 return false;
kvn@64 975 }
kvn@64 976
duke@0 977 //------------------------------Identity---------------------------------------
duke@0 978 // Loads are identity if previous store is to same address
duke@0 979 Node *LoadNode::Identity( PhaseTransform *phase ) {
duke@0 980 // If the previous store-maker is the right kind of Store, and the store is
duke@0 981 // to the same address, then we are equal to the value stored.
duke@0 982 Node* mem = in(MemNode::Memory);
duke@0 983 Node* value = can_see_stored_value(mem, phase);
duke@0 984 if( value ) {
duke@0 985 // byte, short & char stores truncate naturally.
duke@0 986 // A load has to load the truncated value which requires
duke@0 987 // some sort of masking operation and that requires an
duke@0 988 // Ideal call instead of an Identity call.
duke@0 989 if (memory_size() < BytesPerInt) {
duke@0 990 // If the input to the store does not fit with the load's result type,
duke@0 991 // it must be truncated via an Ideal call.
duke@0 992 if (!phase->type(value)->higher_equal(phase->type(this)))
duke@0 993 return this;
duke@0 994 }
duke@0 995 // (This works even when value is a Con, but LoadNode::Value
duke@0 996 // usually runs first, producing the singleton type of the Con.)
duke@0 997 return value;
duke@0 998 }
kvn@64 999
kvn@64 1000 // Search for an existing data phi which was generated before for the same
twisti@605 1001 // instance's field to avoid infinite generation of phis in a loop.
kvn@64 1002 Node *region = mem->in(0);
kvn@64 1003 if (is_instance_field_load_with_local_phi(region)) {
kvn@64 1004 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
kvn@64 1005 int this_index = phase->C->get_alias_index(addr_t);
kvn@64 1006 int this_offset = addr_t->offset();
kvn@64 1007 int this_id = addr_t->is_oopptr()->instance_id();
kvn@64 1008 const Type* this_type = bottom_type();
kvn@64 1009 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
kvn@64 1010 Node* phi = region->fast_out(i);
kvn@64 1011 if (phi->is_Phi() && phi != mem &&
kvn@64 1012 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
kvn@64 1013 return phi;
kvn@64 1014 }
kvn@64 1015 }
kvn@64 1016 }
kvn@64 1017
duke@0 1018 return this;
duke@0 1019 }
duke@0 1020
never@17 1021
never@17 1022 // Returns true if the AliasType refers to the field that holds the
never@17 1023 // cached box array. Currently only handles the IntegerCache case.
never@17 1024 static bool is_autobox_cache(Compile::AliasType* atp) {
never@17 1025 if (atp != NULL && atp->field() != NULL) {
never@17 1026 ciField* field = atp->field();
never@17 1027 ciSymbol* klass = field->holder()->name();
never@17 1028 if (field->name() == ciSymbol::cache_field_name() &&
never@17 1029 field->holder()->uses_default_loader() &&
never@17 1030 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 1031 return true;
never@17 1032 }
never@17 1033 }
never@17 1034 return false;
never@17 1035 }
never@17 1036
never@17 1037 // Fetch the base value in the autobox array
never@17 1038 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
never@17 1039 if (atp != NULL && atp->field() != NULL) {
never@17 1040 ciField* field = atp->field();
never@17 1041 ciSymbol* klass = field->holder()->name();
never@17 1042 if (field->name() == ciSymbol::cache_field_name() &&
never@17 1043 field->holder()->uses_default_loader() &&
never@17 1044 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 1045 assert(field->is_constant(), "what?");
never@17 1046 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
never@17 1047 // Fetch the box object at the base of the array and get its value
never@17 1048 ciInstance* box = array->obj_at(0)->as_instance();
never@17 1049 ciInstanceKlass* ik = box->klass()->as_instance_klass();
never@17 1050 if (ik->nof_nonstatic_fields() == 1) {
never@17 1051 // This should be true nonstatic_field_at requires calling
never@17 1052 // nof_nonstatic_fields so check it anyway
never@17 1053 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
never@17 1054 cache_offset = c.as_int();
never@17 1055 }
never@17 1056 return true;
never@17 1057 }
never@17 1058 }
never@17 1059 return false;
never@17 1060 }
never@17 1061
never@17 1062 // Returns true if the AliasType refers to the value field of an
never@17 1063 // autobox object. Currently only handles Integer.
never@17 1064 static bool is_autobox_object(Compile::AliasType* atp) {
never@17 1065 if (atp != NULL && atp->field() != NULL) {
never@17 1066 ciField* field = atp->field();
never@17 1067 ciSymbol* klass = field->holder()->name();
never@17 1068 if (field->name() == ciSymbol::value_name() &&
never@17 1069 field->holder()->uses_default_loader() &&
never@17 1070 klass == ciSymbol::java_lang_Integer()) {
never@17 1071 return true;
never@17 1072 }
never@17 1073 }
never@17 1074 return false;
never@17 1075 }
never@17 1076
never@17 1077
never@17 1078 // We're loading from an object which has autobox behaviour.
never@17 1079 // If this object is result of a valueOf call we'll have a phi
never@17 1080 // merging a newly allocated object and a load from the cache.
never@17 1081 // We want to replace this load with the original incoming
never@17 1082 // argument to the valueOf call.
never@17 1083 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
never@17 1084 Node* base = in(Address)->in(AddPNode::Base);
never@17 1085 if (base->is_Phi() && base->req() == 3) {
never@17 1086 AllocateNode* allocation = NULL;
never@17 1087 int allocation_index = -1;
never@17 1088 int load_index = -1;
never@17 1089 for (uint i = 1; i < base->req(); i++) {
never@17 1090 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
never@17 1091 if (allocation != NULL) {
never@17 1092 allocation_index = i;
never@17 1093 load_index = 3 - allocation_index;
never@17 1094 break;
never@17 1095 }
never@17 1096 }
kvn@583 1097 bool has_load = ( allocation != NULL &&
kvn@583 1098 (base->in(load_index)->is_Load() ||
kvn@583 1099 base->in(load_index)->is_DecodeN() &&
kvn@583 1100 base->in(load_index)->in(1)->is_Load()) );
kvn@583 1101 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
never@17 1102 // Push the loads from the phi that comes from valueOf up
never@17 1103 // through it to allow elimination of the loads and the recovery
never@17 1104 // of the original value.
never@17 1105 Node* mem_phi = in(Memory);
never@17 1106 Node* offset = in(Address)->in(AddPNode::Offset);
never@553 1107 Node* region = base->in(0);
never@17 1108
never@17 1109 Node* in1 = clone();
never@17 1110 Node* in1_addr = in1->in(Address)->clone();
never@17 1111 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
never@17 1112 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
never@17 1113 in1_addr->set_req(AddPNode::Offset, offset);
never@553 1114 in1->set_req(0, region->in(allocation_index));
never@17 1115 in1->set_req(Address, in1_addr);
never@17 1116 in1->set_req(Memory, mem_phi->in(allocation_index));
never@17 1117
never@17 1118 Node* in2 = clone();
never@17 1119 Node* in2_addr = in2->in(Address)->clone();
never@17 1120 in2_addr->set_req(AddPNode::Base, base->in(load_index));
never@17 1121 in2_addr->set_req(AddPNode::Address, base->in(load_index));
never@17 1122 in2_addr->set_req(AddPNode::Offset, offset);
never@553 1123 in2->set_req(0, region->in(load_index));
never@17 1124 in2->set_req(Address, in2_addr);
never@17 1125 in2->set_req(Memory, mem_phi->in(load_index));
never@17 1126
never@17 1127 in1_addr = phase->transform(in1_addr);
never@17 1128 in1 = phase->transform(in1);
never@17 1129 in2_addr = phase->transform(in2_addr);
never@17 1130 in2 = phase->transform(in2);
never@17 1131
never@553 1132 PhiNode* result = PhiNode::make_blank(region, this);
never@17 1133 result->set_req(allocation_index, in1);
never@17 1134 result->set_req(load_index, in2);
never@17 1135 return result;
never@17 1136 }
kvn@583 1137 } else if (base->is_Load() ||
kvn@583 1138 base->is_DecodeN() && base->in(1)->is_Load()) {
kvn@583 1139 if (base->is_DecodeN()) {
kvn@583 1140 // Get LoadN node which loads cached Integer object
kvn@583 1141 base = base->in(1);
kvn@583 1142 }
never@17 1143 // Eliminate the load of Integer.value for integers from the cache
never@17 1144 // array by deriving the value from the index into the array.
never@17 1145 // Capture the offset of the load and then reverse the computation.
never@17 1146 Node* load_base = base->in(Address)->in(AddPNode::Base);
kvn@583 1147 if (load_base->is_DecodeN()) {
kvn@583 1148 // Get LoadN node which loads IntegerCache.cache field
kvn@583 1149 load_base = load_base->in(1);
kvn@583 1150 }
never@17 1151 if (load_base != NULL) {
never@17 1152 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
never@17 1153 intptr_t cache_offset;
never@17 1154 int shift = -1;
never@17 1155 Node* cache = NULL;
never@17 1156 if (is_autobox_cache(atp)) {
kvn@29 1157 shift = exact_log2(type2aelembytes(T_OBJECT));
never@17 1158 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
never@17 1159 }
never@17 1160 if (cache != NULL && base->in(Address)->is_AddP()) {
never@17 1161 Node* elements[4];
never@17 1162 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
never@17 1163 int cache_low;
never@17 1164 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
never@17 1165 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
never@17 1166 // Add up all the offsets making of the address of the load
never@17 1167 Node* result = elements[0];
never@17 1168 for (int i = 1; i < count; i++) {
never@17 1169 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
never@17 1170 }
never@17 1171 // Remove the constant offset from the address and then
never@17 1172 // remove the scaling of the offset to recover the original index.
never@17 1173 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
never@17 1174 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
never@17 1175 // Peel the shift off directly but wrap it in a dummy node
never@17 1176 // since Ideal can't return existing nodes
never@17 1177 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
never@17 1178 } else {
never@17 1179 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
never@17 1180 }
never@17 1181 #ifdef _LP64
never@17 1182 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
never@17 1183 #endif
never@17 1184 return result;
never@17 1185 }
never@17 1186 }
never@17 1187 }
never@17 1188 }
never@17 1189 return NULL;
never@17 1190 }
never@17 1191
kvn@163 1192 //------------------------------split_through_phi------------------------------
kvn@163 1193 // Split instance field load through Phi.
kvn@163 1194 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
kvn@163 1195 Node* mem = in(MemNode::Memory);
kvn@163 1196 Node* address = in(MemNode::Address);
kvn@163 1197 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@163 1198 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@163 1199
kvn@163 1200 assert(mem->is_Phi() && (t_oop != NULL) &&
kvn@223 1201 t_oop->is_known_instance_field(), "invalide conditions");
kvn@163 1202
kvn@163 1203 Node *region = mem->in(0);
kvn@163 1204 if (region == NULL) {
kvn@163 1205 return NULL; // Wait stable graph
kvn@163 1206 }
kvn@163 1207 uint cnt = mem->req();
kvn@163 1208 for( uint i = 1; i < cnt; i++ ) {
kvn@163 1209 Node *in = mem->in(i);
kvn@163 1210 if( in == NULL ) {
kvn@163 1211 return NULL; // Wait stable graph
kvn@163 1212 }
kvn@163 1213 }
kvn@163 1214 // Check for loop invariant.
kvn@163 1215 if (cnt == 3) {
kvn@163 1216 for( uint i = 1; i < cnt; i++ ) {
kvn@163 1217 Node *in = mem->in(i);
kvn@163 1218 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
kvn@163 1219 if (m == mem) {
kvn@163 1220 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
kvn@163 1221 return this;
kvn@163 1222 }
kvn@163 1223 }
kvn@163 1224 }
kvn@163 1225 // Split through Phi (see original code in loopopts.cpp).
kvn@163 1226 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
kvn@163 1227
kvn@163 1228 // Do nothing here if Identity will find a value
kvn@163 1229 // (to avoid infinite chain of value phis generation).
kvn@163 1230 if ( !phase->eqv(this, this->Identity(phase)) )
kvn@163 1231 return NULL;
kvn@163 1232
kvn@163 1233 // Skip the split if the region dominates some control edge of the address.
kvn@163 1234 if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
kvn@163 1235 return NULL;
kvn@163 1236
kvn@163 1237 const Type* this_type = this->bottom_type();
kvn@163 1238 int this_index = phase->C->get_alias_index(addr_t);
kvn@163 1239 int this_offset = addr_t->offset();
kvn@163 1240 int this_iid = addr_t->is_oopptr()->instance_id();
kvn@163 1241 int wins = 0;
kvn@163 1242 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@163 1243 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
kvn@163 1244 for( uint i = 1; i < region->req(); i++ ) {
kvn@163 1245 Node *x;
kvn@163 1246 Node* the_clone = NULL;
kvn@163 1247 if( region->in(i) == phase->C->top() ) {
kvn@163 1248 x = phase->C->top(); // Dead path? Use a dead data op
kvn@163 1249 } else {
kvn@163 1250 x = this->clone(); // Else clone up the data op
kvn@163 1251 the_clone = x; // Remember for possible deletion.
kvn@163 1252 // Alter data node to use pre-phi inputs
kvn@163 1253 if( this->in(0) == region ) {
kvn@163 1254 x->set_req( 0, region->in(i) );
kvn@163 1255 } else {
kvn@163 1256 x->set_req( 0, NULL );
kvn@163 1257 }
kvn@163 1258 for( uint j = 1; j < this->req(); j++ ) {
kvn@163 1259 Node *in = this->in(j);
kvn@163 1260 if( in->is_Phi() && in->in(0) == region )
kvn@163 1261 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
kvn@163 1262 }
kvn@163 1263 }
kvn@163 1264 // Check for a 'win' on some paths
kvn@163 1265 const Type *t = x->Value(igvn);
kvn@163 1266
kvn@163 1267 bool singleton = t->singleton();
kvn@163 1268
kvn@163 1269 // See comments in PhaseIdealLoop::split_thru_phi().
kvn@163 1270 if( singleton && t == Type::TOP ) {
kvn@163 1271 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
kvn@163 1272 }
kvn@163 1273
kvn@163 1274 if( singleton ) {
kvn@163 1275 wins++;
kvn@163 1276 x = igvn->makecon(t);
kvn@163 1277 } else {
kvn@163 1278 // We now call Identity to try to simplify the cloned node.
kvn@163 1279 // Note that some Identity methods call phase->type(this).
kvn@163 1280 // Make sure that the type array is big enough for
kvn@163 1281 // our new node, even though we may throw the node away.
kvn@163 1282 // (This tweaking with igvn only works because x is a new node.)
kvn@163 1283 igvn->set_type(x, t);
kvn@293 1284 // If x is a TypeNode, capture any more-precise type permanently into Node
twisti@605 1285 // otherwise it will be not updated during igvn->transform since
kvn@293 1286 // igvn->type(x) is set to x->Value() already.
kvn@293 1287 x->raise_bottom_type(t);
kvn@163 1288 Node *y = x->Identity(igvn);
kvn@163 1289 if( y != x ) {
kvn@163 1290 wins++;
kvn@163 1291 x = y;
kvn@163 1292 } else {
kvn@163 1293 y = igvn->hash_find(x);
kvn@163 1294 if( y ) {
kvn@163 1295 wins++;
kvn@163 1296 x = y;
kvn@163 1297 } else {
kvn@163 1298 // Else x is a new node we are keeping
kvn@163 1299 // We do not need register_new_node_with_optimizer
kvn@163 1300 // because set_type has already been called.
kvn@163 1301 igvn->_worklist.push(x);
kvn@163 1302 }
kvn@163 1303 }
kvn@163 1304 }
kvn@163 1305 if (x != the_clone && the_clone != NULL)
kvn@163 1306 igvn->remove_dead_node(the_clone);
kvn@163 1307 phi->set_req(i, x);
kvn@163 1308 }
kvn@163 1309 if( wins > 0 ) {
kvn@163 1310 // Record Phi
kvn@163 1311 igvn->register_new_node_with_optimizer(phi);
kvn@163 1312 return phi;
kvn@163 1313 }
kvn@163 1314 igvn->remove_dead_node(phi);
kvn@163 1315 return NULL;
kvn@163 1316 }
never@17 1317
duke@0 1318 //------------------------------Ideal------------------------------------------
duke@0 1319 // If the load is from Field memory and the pointer is non-null, we can
duke@0 1320 // zero out the control input.
duke@0 1321 // If the offset is constant and the base is an object allocation,
duke@0 1322 // try to hook me up to the exact initializing store.
duke@0 1323 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1324 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 1325 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 1326
duke@0 1327 Node* ctrl = in(MemNode::Control);
duke@0 1328 Node* address = in(MemNode::Address);
duke@0 1329
duke@0 1330 // Skip up past a SafePoint control. Cannot do this for Stores because
duke@0 1331 // pointer stores & cardmarks must stay on the same side of a SafePoint.
duke@0 1332 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
duke@0 1333 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
duke@0 1334 ctrl = ctrl->in(0);
duke@0 1335 set_req(MemNode::Control,ctrl);
duke@0 1336 }
duke@0 1337
kvn@708 1338 intptr_t ignore = 0;
kvn@708 1339 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
kvn@708 1340 if (base != NULL
kvn@708 1341 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
kvn@708 1342 // Check for useless control edge in some common special cases
kvn@708 1343 if (in(MemNode::Control) != NULL
duke@0 1344 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
kvn@119 1345 && all_controls_dominate(base, phase->C->start())) {
duke@0 1346 // A method-invariant, non-null address (constant or 'this' argument).
duke@0 1347 set_req(MemNode::Control, NULL);
duke@0 1348 }
kvn@708 1349
kvn@708 1350 if (EliminateAutoBox && can_reshape) {
kvn@708 1351 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null");
never@17 1352 Compile::AliasType* atp = phase->C->alias_type(adr_type());
never@17 1353 if (is_autobox_object(atp)) {
never@17 1354 Node* result = eliminate_autobox(phase);
never@17 1355 if (result != NULL) return result;
never@17 1356 }
never@17 1357 }
never@17 1358 }
never@17 1359
kvn@74 1360 Node* mem = in(MemNode::Memory);
kvn@74 1361 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@74 1362
kvn@74 1363 if (addr_t != NULL) {
kvn@74 1364 // try to optimize our memory input
kvn@74 1365 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
kvn@74 1366 if (opt_mem != mem) {
kvn@74 1367 set_req(MemNode::Memory, opt_mem);
kvn@305 1368 if (phase->type( opt_mem ) == Type::TOP) return NULL;
kvn@74 1369 return this;
kvn@74 1370 }
kvn@74 1371 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@74 1372 if (can_reshape && opt_mem->is_Phi() &&
kvn@223 1373 (t_oop != NULL) && t_oop->is_known_instance_field()) {
kvn@163 1374 // Split instance field load through Phi.
kvn@163 1375 Node* result = split_through_phi(phase);
kvn@163 1376 if (result != NULL) return result;
kvn@74 1377 }
kvn@74 1378 }
kvn@74 1379
duke@0 1380 // Check for prior store with a different base or offset; make Load
duke@0 1381 // independent. Skip through any number of them. Bail out if the stores
duke@0 1382 // are in an endless dead cycle and report no progress. This is a key
duke@0 1383 // transform for Reflection. However, if after skipping through the Stores
duke@0 1384 // we can't then fold up against a prior store do NOT do the transform as
duke@0 1385 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
duke@0 1386 // array memory alive twice: once for the hoisted Load and again after the
duke@0 1387 // bypassed Store. This situation only works if EVERYBODY who does
duke@0 1388 // anti-dependence work knows how to bypass. I.e. we need all
duke@0 1389 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
duke@0 1390 // the alias index stuff. So instead, peek through Stores and IFF we can
duke@0 1391 // fold up, do so.
duke@0 1392 Node* prev_mem = find_previous_store(phase);
duke@0 1393 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 1394 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
duke@0 1395 // (c) See if we can fold up on the spot, but don't fold up here.
twisti@558 1396 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
duke@0 1397 // just return a prior value, which is done by Identity calls.
duke@0 1398 if (can_see_stored_value(prev_mem, phase)) {
duke@0 1399 // Make ready for step (d):
duke@0 1400 set_req(MemNode::Memory, prev_mem);
duke@0 1401 return this;
duke@0 1402 }
duke@0 1403 }
duke@0 1404
duke@0 1405 return NULL; // No further progress
duke@0 1406 }
duke@0 1407
duke@0 1408 // Helper to recognize certain Klass fields which are invariant across
duke@0 1409 // some group of array types (e.g., int[] or all T[] where T < Object).
duke@0 1410 const Type*
duke@0 1411 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
duke@0 1412 ciKlass* klass) const {
duke@0 1413 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1414 // The field is Klass::_modifier_flags. Return its (constant) value.
duke@0 1415 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
duke@0 1416 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
duke@0 1417 return TypeInt::make(klass->modifier_flags());
duke@0 1418 }
duke@0 1419 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1420 // The field is Klass::_access_flags. Return its (constant) value.
duke@0 1421 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
duke@0 1422 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
duke@0 1423 return TypeInt::make(klass->access_flags());
duke@0 1424 }
duke@0 1425 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1426 // The field is Klass::_layout_helper. Return its constant value if known.
duke@0 1427 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1428 return TypeInt::make(klass->layout_helper());
duke@0 1429 }
duke@0 1430
duke@0 1431 // No match.
duke@0 1432 return NULL;
duke@0 1433 }
duke@0 1434
duke@0 1435 //------------------------------Value-----------------------------------------
duke@0 1436 const Type *LoadNode::Value( PhaseTransform *phase ) const {
duke@0 1437 // Either input is TOP ==> the result is TOP
duke@0 1438 Node* mem = in(MemNode::Memory);
duke@0 1439 const Type *t1 = phase->type(mem);
duke@0 1440 if (t1 == Type::TOP) return Type::TOP;
duke@0 1441 Node* adr = in(MemNode::Address);
duke@0 1442 const TypePtr* tp = phase->type(adr)->isa_ptr();
duke@0 1443 if (tp == NULL || tp->empty()) return Type::TOP;
duke@0 1444 int off = tp->offset();
duke@0 1445 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
duke@0 1446
duke@0 1447 // Try to guess loaded type from pointer type
duke@0 1448 if (tp->base() == Type::AryPtr) {
duke@0 1449 const Type *t = tp->is_aryptr()->elem();
duke@0 1450 // Don't do this for integer types. There is only potential profit if
duke@0 1451 // the element type t is lower than _type; that is, for int types, if _type is
duke@0 1452 // more restrictive than t. This only happens here if one is short and the other
duke@0 1453 // char (both 16 bits), and in those cases we've made an intentional decision
duke@0 1454 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
duke@0 1455 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
duke@0 1456 //
duke@0 1457 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
duke@0 1458 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
duke@0 1459 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
duke@0 1460 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
duke@0 1461 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
duke@0 1462 // In fact, that could have been the original type of p1, and p1 could have
duke@0 1463 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
duke@0 1464 // expression (LShiftL quux 3) independently optimized to the constant 8.
duke@0 1465 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
kvn@293 1466 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
duke@0 1467 // t might actually be lower than _type, if _type is a unique
duke@0 1468 // concrete subclass of abstract class t.
duke@0 1469 // Make sure the reference is not into the header, by comparing
duke@0 1470 // the offset against the offset of the start of the array's data.
duke@0 1471 // Different array types begin at slightly different offsets (12 vs. 16).
duke@0 1472 // We choose T_BYTE as an example base type that is least restrictive
duke@0 1473 // as to alignment, which will therefore produce the smallest
duke@0 1474 // possible base offset.
duke@0 1475 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
duke@0 1476 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
duke@0 1477 const Type* jt = t->join(_type);
duke@0 1478 // In any case, do not allow the join, per se, to empty out the type.
duke@0 1479 if (jt->empty() && !t->empty()) {
duke@0 1480 // This can happen if a interface-typed array narrows to a class type.
duke@0 1481 jt = _type;
duke@0 1482 }
never@17 1483
kvn@708 1484 if (EliminateAutoBox && adr->is_AddP()) {
never@17 1485 // The pointers in the autobox arrays are always non-null
kvn@708 1486 Node* base = adr->in(AddPNode::Base);
kvn@708 1487 if (base != NULL &&
kvn@708 1488 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) {
never@17 1489 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
never@17 1490 if (is_autobox_cache(atp)) {
never@17 1491 return jt->join(TypePtr::NOTNULL)->is_ptr();
never@17 1492 }
never@17 1493 }
never@17 1494 }
duke@0 1495 return jt;
duke@0 1496 }
duke@0 1497 }
duke@0 1498 } else if (tp->base() == Type::InstPtr) {
duke@0 1499 assert( off != Type::OffsetBot ||
duke@0 1500 // arrays can be cast to Objects
duke@0 1501 tp->is_oopptr()->klass()->is_java_lang_Object() ||
duke@0 1502 // unsafe field access may not have a constant offset
duke@0 1503 phase->C->has_unsafe_access(),
duke@0 1504 "Field accesses must be precise" );
duke@0 1505 // For oop loads, we expect the _type to be precise
duke@0 1506 } else if (tp->base() == Type::KlassPtr) {
duke@0 1507 assert( off != Type::OffsetBot ||
duke@0 1508 // arrays can be cast to Objects
duke@0 1509 tp->is_klassptr()->klass()->is_java_lang_Object() ||
duke@0 1510 // also allow array-loading from the primary supertype
duke@0 1511 // array during subtype checks
duke@0 1512 Opcode() == Op_LoadKlass,
duke@0 1513 "Field accesses must be precise" );
duke@0 1514 // For klass/static loads, we expect the _type to be precise
duke@0 1515 }
duke@0 1516
duke@0 1517 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1518 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1519 ciKlass* klass = tkls->klass();
duke@0 1520 if (klass->is_loaded() && tkls->klass_is_exact()) {
duke@0 1521 // We are loading a field from a Klass metaobject whose identity
duke@0 1522 // is known at compile time (the type is "exact" or "precise").
duke@0 1523 // Check for fields we know are maintained as constants by the VM.
duke@0 1524 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1525 // The field is Klass::_super_check_offset. Return its (constant) value.
duke@0 1526 // (Folds up type checking code.)
duke@0 1527 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
duke@0 1528 return TypeInt::make(klass->super_check_offset());
duke@0 1529 }
duke@0 1530 // Compute index into primary_supers array
duke@0 1531 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1532 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1533 if( depth < ciKlass::primary_super_limit() ) {
duke@0 1534 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1535 // (Folds up type checking code.)
duke@0 1536 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1537 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1538 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1539 }
duke@0 1540 const Type* aift = load_array_final_field(tkls, klass);
duke@0 1541 if (aift != NULL) return aift;
duke@0 1542 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
duke@0 1543 && klass->is_array_klass()) {
duke@0 1544 // The field is arrayKlass::_component_mirror. Return its (constant) value.
duke@0 1545 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
duke@0 1546 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
duke@0 1547 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
duke@0 1548 }
duke@0 1549 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1550 // The field is Klass::_java_mirror. Return its (constant) value.
duke@0 1551 // (Folds up the 2nd indirection in anObjConstant.getClass().)
duke@0 1552 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
duke@0 1553 return TypeInstPtr::make(klass->java_mirror());
duke@0 1554 }
duke@0 1555 }
duke@0 1556
duke@0 1557 // We can still check if we are loading from the primary_supers array at a
duke@0 1558 // shallow enough depth. Even though the klass is not exact, entries less
duke@0 1559 // than or equal to its super depth are correct.
duke@0 1560 if (klass->is_loaded() ) {
duke@0 1561 ciType *inner = klass->klass();
duke@0 1562 while( inner->is_obj_array_klass() )
duke@0 1563 inner = inner->as_obj_array_klass()->base_element_type();
duke@0 1564 if( inner->is_instance_klass() &&
duke@0 1565 !inner->as_instance_klass()->flags().is_interface() ) {
duke@0 1566 // Compute index into primary_supers array
duke@0 1567 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1568 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1569 if( depth < ciKlass::primary_super_limit() &&
duke@0 1570 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
duke@0 1571 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1572 // (Folds up type checking code.)
duke@0 1573 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1574 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1575 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1576 }
duke@0 1577 }
duke@0 1578 }
duke@0 1579
duke@0 1580 // If the type is enough to determine that the thing is not an array,
duke@0 1581 // we can give the layout_helper a positive interval type.
duke@0 1582 // This will help short-circuit some reflective code.
duke@0 1583 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
duke@0 1584 && !klass->is_array_klass() // not directly typed as an array
duke@0 1585 && !klass->is_interface() // specifically not Serializable & Cloneable
duke@0 1586 && !klass->is_java_lang_Object() // not the supertype of all T[]
duke@0 1587 ) {
duke@0 1588 // Note: When interfaces are reliable, we can narrow the interface
duke@0 1589 // test to (klass != Serializable && klass != Cloneable).
duke@0 1590 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1591 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
duke@0 1592 // The key property of this type is that it folds up tests
duke@0 1593 // for array-ness, since it proves that the layout_helper is positive.
duke@0 1594 // Thus, a generic value like the basic object layout helper works fine.
duke@0 1595 return TypeInt::make(min_size, max_jint, Type::WidenMin);
duke@0 1596 }
duke@0 1597 }
duke@0 1598
duke@0 1599 // If we are loading from a freshly-allocated object, produce a zero,
duke@0 1600 // if the load is provably beyond the header of the object.
duke@0 1601 // (Also allow a variable load from a fresh array to produce zero.)
duke@0 1602 if (ReduceFieldZeroing) {
duke@0 1603 Node* value = can_see_stored_value(mem,phase);
duke@0 1604 if (value != NULL && value->is_Con())
duke@0 1605 return value->bottom_type();
duke@0 1606 }
duke@0 1607
kvn@64 1608 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@223 1609 if (tinst != NULL && tinst->is_known_instance_field()) {
kvn@64 1610 // If we have an instance type and our memory input is the
kvn@64 1611 // programs's initial memory state, there is no matching store,
kvn@64 1612 // so just return a zero of the appropriate type
kvn@64 1613 Node *mem = in(MemNode::Memory);
kvn@64 1614 if (mem->is_Parm() && mem->in(0)->is_Start()) {
kvn@64 1615 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
kvn@64 1616 return Type::get_zero_type(_type->basic_type());
kvn@64 1617 }
kvn@64 1618 }
duke@0 1619 return _type;
duke@0 1620 }
duke@0 1621
duke@0 1622 //------------------------------match_edge-------------------------------------
duke@0 1623 // Do we Match on this edge index or not? Match only the address.
duke@0 1624 uint LoadNode::match_edge(uint idx) const {
duke@0 1625 return idx == MemNode::Address;
duke@0 1626 }
duke@0 1627
duke@0 1628 //--------------------------LoadBNode::Ideal--------------------------------------
duke@0 1629 //
duke@0 1630 // If the previous store is to the same address as this load,
duke@0 1631 // and the value stored was larger than a byte, replace this load
duke@0 1632 // with the value stored truncated to a byte. If no truncation is
duke@0 1633 // needed, the replacement is done in LoadNode::Identity().
duke@0 1634 //
duke@0 1635 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1636 Node* mem = in(MemNode::Memory);
duke@0 1637 Node* value = can_see_stored_value(mem,phase);
duke@0 1638 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1639 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
duke@0 1640 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
duke@0 1641 }
duke@0 1642 // Identity call will handle the case where truncation is not needed.
duke@0 1643 return LoadNode::Ideal(phase, can_reshape);
duke@0 1644 }
duke@0 1645
twisti@624 1646 //--------------------------LoadUBNode::Ideal-------------------------------------
twisti@624 1647 //
twisti@624 1648 // If the previous store is to the same address as this load,
twisti@624 1649 // and the value stored was larger than a byte, replace this load
twisti@624 1650 // with the value stored truncated to a byte. If no truncation is
twisti@624 1651 // needed, the replacement is done in LoadNode::Identity().
twisti@624 1652 //
twisti@624 1653 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
twisti@624 1654 Node* mem = in(MemNode::Memory);
twisti@624 1655 Node* value = can_see_stored_value(mem, phase);
twisti@624 1656 if (value && !phase->type(value)->higher_equal(_type))
twisti@624 1657 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF));
twisti@624 1658 // Identity call will handle the case where truncation is not needed.
twisti@624 1659 return LoadNode::Ideal(phase, can_reshape);
twisti@624 1660 }
twisti@624 1661
twisti@558 1662 //--------------------------LoadUSNode::Ideal-------------------------------------
duke@0 1663 //
duke@0 1664 // If the previous store is to the same address as this load,
duke@0 1665 // and the value stored was larger than a char, replace this load
duke@0 1666 // with the value stored truncated to a char. If no truncation is
duke@0 1667 // needed, the replacement is done in LoadNode::Identity().
duke@0 1668 //
twisti@558 1669 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1670 Node* mem = in(MemNode::Memory);
duke@0 1671 Node* value = can_see_stored_value(mem,phase);
duke@0 1672 if( value && !phase->type(value)->higher_equal( _type ) )
duke@0 1673 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
duke@0 1674 // Identity call will handle the case where truncation is not needed.
duke@0 1675 return LoadNode::Ideal(phase, can_reshape);
duke@0 1676 }
duke@0 1677
duke@0 1678 //--------------------------LoadSNode::Ideal--------------------------------------
duke@0 1679 //
duke@0 1680 // If the previous store is to the same address as this load,
duke@0 1681 // and the value stored was larger than a short, replace this load
duke@0 1682 // with the value stored truncated to a short. If no truncation is
duke@0 1683 // needed, the replacement is done in LoadNode::Identity().
duke@0 1684 //
duke@0 1685 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1686 Node* mem = in(MemNode::Memory);
duke@0 1687 Node* value = can_see_stored_value(mem,phase);
duke@0 1688 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1689 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
duke@0 1690 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
duke@0 1691 }
duke@0 1692 // Identity call will handle the case where truncation is not needed.
duke@0 1693 return LoadNode::Ideal(phase, can_reshape);
duke@0 1694 }
duke@0 1695
duke@0 1696 //=============================================================================
kvn@164 1697 //----------------------------LoadKlassNode::make------------------------------
kvn@164 1698 // Polymorphic factory method:
kvn@164 1699 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
kvn@164 1700 Compile* C = gvn.C;
kvn@164 1701 Node *ctl = NULL;
kvn@164 1702 // sanity check the alias category against the created node type
kvn@164 1703 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
kvn@164 1704 assert(adr_type != NULL, "expecting TypeOopPtr");
kvn@164 1705 #ifdef _LP64
kvn@164 1706 if (adr_type->is_ptr_to_narrowoop()) {
kvn@221 1707 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
kvn@221 1708 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
kvn@168 1709 }
kvn@164 1710 #endif
kvn@168 1711 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
kvn@168 1712 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
kvn@164 1713 }
kvn@164 1714
duke@0 1715 //------------------------------Value------------------------------------------
duke@0 1716 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
kvn@164 1717 return klass_value_common(phase);
kvn@164 1718 }
kvn@164 1719
kvn@164 1720 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
duke@0 1721 // Either input is TOP ==> the result is TOP
duke@0 1722 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1723 if (t1 == Type::TOP) return Type::TOP;
duke@0 1724 Node *adr = in(MemNode::Address);
duke@0 1725 const Type *t2 = phase->type( adr );
duke@0 1726 if (t2 == Type::TOP) return Type::TOP;
duke@0 1727 const TypePtr *tp = t2->is_ptr();
duke@0 1728 if (TypePtr::above_centerline(tp->ptr()) ||
duke@0 1729 tp->ptr() == TypePtr::Null) return Type::TOP;
duke@0 1730
duke@0 1731 // Return a more precise klass, if possible
duke@0 1732 const TypeInstPtr *tinst = tp->isa_instptr();
duke@0 1733 if (tinst != NULL) {
duke@0 1734 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
duke@0 1735 int offset = tinst->offset();
duke@0 1736 if (ik == phase->C->env()->Class_klass()
duke@0 1737 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1738 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1739 // We are loading a special hidden field from a Class mirror object,
duke@0 1740 // the field which points to the VM's Klass metaobject.
duke@0 1741 ciType* t = tinst->java_mirror_type();
duke@0 1742 // java_mirror_type returns non-null for compile-time Class constants.
duke@0 1743 if (t != NULL) {
duke@0 1744 // constant oop => constant klass
duke@0 1745 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1746 return TypeKlassPtr::make(ciArrayKlass::make(t));
duke@0 1747 }
duke@0 1748 if (!t->is_klass()) {
duke@0 1749 // a primitive Class (e.g., int.class) has NULL for a klass field
duke@0 1750 return TypePtr::NULL_PTR;
duke@0 1751 }
duke@0 1752 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
duke@0 1753 return TypeKlassPtr::make(t->as_klass());
duke@0 1754 }
duke@0 1755 // non-constant mirror, so we can't tell what's going on
duke@0 1756 }
duke@0 1757 if( !ik->is_loaded() )
duke@0 1758 return _type; // Bail out if not loaded
duke@0 1759 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1760 if (tinst->klass_is_exact()) {
duke@0 1761 return TypeKlassPtr::make(ik);
duke@0 1762 }
duke@0 1763 // See if we can become precise: no subklasses and no interface
duke@0 1764 // (Note: We need to support verified interfaces.)
duke@0 1765 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1766 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1767 // Add a dependence; if any subclass added we need to recompile
duke@0 1768 if (!ik->is_final()) {
duke@0 1769 // %%% should use stronger assert_unique_concrete_subtype instead
duke@0 1770 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1771 }
duke@0 1772 // Return precise klass
duke@0 1773 return TypeKlassPtr::make(ik);
duke@0 1774 }
duke@0 1775
duke@0 1776 // Return root of possible klass
duke@0 1777 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
duke@0 1778 }
duke@0 1779 }
duke@0 1780
duke@0 1781 // Check for loading klass from an array
duke@0 1782 const TypeAryPtr *tary = tp->isa_aryptr();
duke@0 1783 if( tary != NULL ) {
duke@0 1784 ciKlass *tary_klass = tary->klass();
duke@0 1785 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
duke@0 1786 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
duke@0 1787 if (tary->klass_is_exact()) {
duke@0 1788 return TypeKlassPtr::make(tary_klass);
duke@0 1789 }
duke@0 1790 ciArrayKlass *ak = tary->klass()->as_array_klass();
duke@0 1791 // If the klass is an object array, we defer the question to the
duke@0 1792 // array component klass.
duke@0 1793 if( ak->is_obj_array_klass() ) {
duke@0 1794 assert( ak->is_loaded(), "" );
duke@0 1795 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
duke@0 1796 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
duke@0 1797 ciInstanceKlass* ik = base_k->as_instance_klass();
duke@0 1798 // See if we can become precise: no subklasses and no interface
duke@0 1799 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1800 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1801 // Add a dependence; if any subclass added we need to recompile
duke@0 1802 if (!ik->is_final()) {
duke@0 1803 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1804 }
duke@0 1805 // Return precise array klass
duke@0 1806 return TypeKlassPtr::make(ak);
duke@0 1807 }
duke@0 1808 }
duke@0 1809 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@0 1810 } else { // Found a type-array?
duke@0 1811 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1812 assert( ak->is_type_array_klass(), "" );
duke@0 1813 return TypeKlassPtr::make(ak); // These are always precise
duke@0 1814 }
duke@0 1815 }
duke@0 1816 }
duke@0 1817
duke@0 1818 // Check for loading klass from an array klass
duke@0 1819 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1820 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1821 ciKlass* klass = tkls->klass();
duke@0 1822 if( !klass->is_loaded() )
duke@0 1823 return _type; // Bail out if not loaded
duke@0 1824 if( klass->is_obj_array_klass() &&
duke@0 1825 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1826 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
duke@0 1827 // // Always returning precise element type is incorrect,
duke@0 1828 // // e.g., element type could be object and array may contain strings
duke@0 1829 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
duke@0 1830
duke@0 1831 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
duke@0 1832 // according to the element type's subclassing.
duke@0 1833 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
duke@0 1834 }
duke@0 1835 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
duke@0 1836 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1837 ciKlass* sup = klass->as_instance_klass()->super();
duke@0 1838 // The field is Klass::_super. Return its (constant) value.
duke@0 1839 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
duke@0 1840 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
duke@0 1841 }
duke@0 1842 }
duke@0 1843
duke@0 1844 // Bailout case
duke@0 1845 return LoadNode::Value(phase);
duke@0 1846 }
duke@0 1847
duke@0 1848 //------------------------------Identity---------------------------------------
duke@0 1849 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
duke@0 1850 // Also feed through the klass in Allocate(...klass...)._klass.
duke@0 1851 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
kvn@164 1852 return klass_identity_common(phase);
kvn@164 1853 }
kvn@164 1854
kvn@164 1855 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
duke@0 1856 Node* x = LoadNode::Identity(phase);
duke@0 1857 if (x != this) return x;
duke@0 1858
duke@0 1859 // Take apart the address into an oop and and offset.
duke@0 1860 // Return 'this' if we cannot.
duke@0 1861 Node* adr = in(MemNode::Address);
duke@0 1862 intptr_t offset = 0;
duke@0 1863 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1864 if (base == NULL) return this;
duke@0 1865 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
duke@0 1866 if (toop == NULL) return this;
duke@0 1867
duke@0 1868 // We can fetch the klass directly through an AllocateNode.
duke@0 1869 // This works even if the klass is not constant (clone or newArray).
duke@0 1870 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1871 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
duke@0 1872 if (allocated_klass != NULL) {
duke@0 1873 return allocated_klass;
duke@0 1874 }
duke@0 1875 }
duke@0 1876
duke@0 1877 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
duke@0 1878 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
duke@0 1879 // See inline_native_Class_query for occurrences of these patterns.
duke@0 1880 // Java Example: x.getClass().isAssignableFrom(y)
duke@0 1881 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
duke@0 1882 //
duke@0 1883 // This improves reflective code, often making the Class
duke@0 1884 // mirror go completely dead. (Current exception: Class
duke@0 1885 // mirrors may appear in debug info, but we could clean them out by
duke@0 1886 // introducing a new debug info operator for klassOop.java_mirror).
duke@0 1887 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
duke@0 1888 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1889 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1890 // We are loading a special hidden field from a Class mirror,
duke@0 1891 // the field which points to its Klass or arrayKlass metaobject.
duke@0 1892 if (base->is_Load()) {
duke@0 1893 Node* adr2 = base->in(MemNode::Address);
duke@0 1894 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
duke@0 1895 if (tkls != NULL && !tkls->empty()
duke@0 1896 && (tkls->klass()->is_instance_klass() ||
duke@0 1897 tkls->klass()->is_array_klass())
duke@0 1898 && adr2->is_AddP()
duke@0 1899 ) {
duke@0 1900 int mirror_field = Klass::java_mirror_offset_in_bytes();
duke@0 1901 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1902 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
duke@0 1903 }
duke@0 1904 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
duke@0 1905 return adr2->in(AddPNode::Base);
duke@0 1906 }
duke@0 1907 }
duke@0 1908 }
duke@0 1909 }
duke@0 1910
duke@0 1911 return this;
duke@0 1912 }
duke@0 1913
kvn@164 1914
kvn@164 1915 //------------------------------Value------------------------------------------
kvn@164 1916 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
kvn@164 1917 const Type *t = klass_value_common(phase);
kvn@221 1918 if (t == Type::TOP)
kvn@221 1919 return t;
kvn@221 1920
kvn@221 1921 return t->make_narrowoop();
kvn@164 1922 }
kvn@164 1923
kvn@164 1924 //------------------------------Identity---------------------------------------
kvn@164 1925 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
kvn@164 1926 // Also feed through the klass in Allocate(...klass...)._klass.
kvn@164 1927 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
kvn@164 1928 Node *x = klass_identity_common(phase);
kvn@164 1929
kvn@164 1930 const Type *t = phase->type( x );
kvn@164 1931 if( t == Type::TOP ) return x;
kvn@164 1932 if( t->isa_narrowoop()) return x;
kvn@164 1933
kvn@221 1934 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
kvn@164 1935 }
kvn@164 1936
duke@0 1937 //------------------------------Value-----------------------------------------
duke@0 1938 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
duke@0 1939 // Either input is TOP ==> the result is TOP
duke@0 1940 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1941 if( t1 == Type::TOP ) return Type::TOP;
duke@0 1942 Node *adr = in(MemNode::Address);
duke@0 1943 const Type *t2 = phase->type( adr );
duke@0 1944 if( t2 == Type::TOP ) return Type::TOP;
duke@0 1945 const TypePtr *tp = t2->is_ptr();
duke@0 1946 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
duke@0 1947 const TypeAryPtr *tap = tp->isa_aryptr();
duke@0 1948 if( !tap ) return _type;
duke@0 1949 return tap->size();
duke@0 1950 }
duke@0 1951
rasbold@366 1952 //-------------------------------Ideal---------------------------------------
rasbold@366 1953 // Feed through the length in AllocateArray(...length...)._length.
rasbold@366 1954 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
rasbold@366 1955 Node* p = MemNode::Ideal_common(phase, can_reshape);
rasbold@366 1956 if (p) return (p == NodeSentinel) ? NULL : p;
rasbold@366 1957
rasbold@366 1958 // Take apart the address into an oop and and offset.
rasbold@366 1959 // Return 'this' if we cannot.
rasbold@366 1960 Node* adr = in(MemNode::Address);
rasbold@366 1961 intptr_t offset = 0;
rasbold@366 1962 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
rasbold@366 1963 if (base == NULL) return NULL;
rasbold@366 1964 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
rasbold@366 1965 if (tary == NULL) return NULL;
rasbold@366 1966
rasbold@366 1967 // We can fetch the length directly through an AllocateArrayNode.
rasbold@366 1968 // This works even if the length is not constant (clone or newArray).
rasbold@366 1969 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@366 1970 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@366 1971 if (alloc != NULL) {
rasbold@366 1972 Node* allocated_length = alloc->Ideal_length();
rasbold@366 1973 Node* len = alloc->make_ideal_length(tary, phase);
rasbold@366 1974 if (allocated_length != len) {
rasbold@366 1975 // New CastII improves on this.
rasbold@366 1976 return len;
rasbold@366 1977 }
rasbold@366 1978 }
rasbold@366 1979 }
rasbold@366 1980
rasbold@366 1981 return NULL;
rasbold@366 1982 }
rasbold@366 1983
duke@0 1984 //------------------------------Identity---------------------------------------
duke@0 1985 // Feed through the length in AllocateArray(...length...)._length.
duke@0 1986 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
duke@0 1987 Node* x = LoadINode::Identity(phase);
duke@0 1988 if (x != this) return x;
duke@0 1989
duke@0 1990 // Take apart the address into an oop and and offset.
duke@0 1991 // Return 'this' if we cannot.
duke@0 1992 Node* adr = in(MemNode::Address);
duke@0 1993 intptr_t offset = 0;
duke@0 1994 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1995 if (base == NULL) return this;
duke@0 1996 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
duke@0 1997 if (tary == NULL) return this;
duke@0 1998
duke@0 1999 // We can fetch the length directly through an AllocateArrayNode.
duke@0 2000 // This works even if the length is not constant (clone or newArray).
duke@0 2001 if (offset == arrayOopDesc::length_offset_in_bytes()) {
rasbold@366 2002 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
rasbold@366 2003 if (alloc != NULL) {
rasbold@366 2004 Node* allocated_length = alloc->Ideal_length();
rasbold@366 2005 // Do not allow make_ideal_length to allocate a CastII node.
rasbold@366 2006 Node* len = alloc->make_ideal_length(tary, phase, false);
rasbold@366 2007 if (allocated_length == len) {
rasbold@366 2008 // Return allocated_length only if it would not be improved by a CastII.
rasbold@366 2009 return allocated_length;
rasbold@366 2010 }
duke@0 2011 }
duke@0 2012 }
duke@0 2013
duke@0 2014 return this;
duke@0 2015
duke@0 2016 }
rasbold@366 2017
duke@0 2018 //=============================================================================
duke@0 2019 //---------------------------StoreNode::make-----------------------------------
duke@0 2020 // Polymorphic factory method:
coleenp@113 2021 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
coleenp@113 2022 Compile* C = gvn.C;
coleenp@113 2023
duke@0 2024 switch (bt) {
duke@0 2025 case T_BOOLEAN:
duke@0 2026 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
duke@0 2027 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
duke@0 2028 case T_CHAR:
duke@0 2029 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
duke@0 2030 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
duke@0 2031 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
duke@0 2032 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
duke@0 2033 case T_ADDRESS:
coleenp@113 2034 case T_OBJECT:
coleenp@113 2035 #ifdef _LP64
kvn@163 2036 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
coleenp@113 2037 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
coleenp@113 2038 adr->bottom_type()->isa_rawptr())) {
kvn@221 2039 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
kvn@221 2040 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
coleenp@113 2041 } else
coleenp@113 2042 #endif
kvn@221 2043 {
kvn@221 2044 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
kvn@221 2045 }
duke@0 2046 }
duke@0 2047 ShouldNotReachHere();
duke@0 2048 return (StoreNode*)NULL;
duke@0 2049 }
duke@0 2050
duke@0 2051 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
duke@0 2052 bool require_atomic = true;
duke@0 2053 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
duke@0 2054 }
duke@0 2055
duke@0 2056
duke@0 2057 //--------------------------bottom_type----------------------------------------
duke@0 2058 const Type *StoreNode::bottom_type() const {
duke@0 2059 return Type::MEMORY;
duke@0 2060 }
duke@0 2061
duke@0 2062 //------------------------------hash-------------------------------------------
duke@0 2063 uint StoreNode::hash() const {
duke@0 2064 // unroll addition of interesting fields
duke@0 2065 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
duke@0 2066
duke@0 2067 // Since they are not commoned, do not hash them:
duke@0 2068 return NO_HASH;
duke@0 2069 }
duke@0 2070
duke@0 2071 //------------------------------Ideal------------------------------------------
duke@0 2072 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
duke@0 2073 // When a store immediately follows a relevant allocation/initialization,
duke@0 2074 // try to capture it into the initialization, or hoist it above.
duke@0 2075 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 2076 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 2077 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 2078
duke@0 2079 Node* mem = in(MemNode::Memory);
duke@0 2080 Node* address = in(MemNode::Address);
duke@0 2081
duke@0 2082 // Back-to-back stores to same address? Fold em up.
duke@0 2083 // Generally unsafe if I have intervening uses...
duke@0 2084 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
duke@0 2085 // Looking at a dead closed cycle of memory?
duke@0 2086 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
duke@0 2087
duke@0 2088 assert(Opcode() == mem->Opcode() ||
duke@0 2089 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
duke@0 2090 "no mismatched stores, except on raw memory");
duke@0 2091
duke@0 2092 if (mem->outcnt() == 1 && // check for intervening uses
duke@0 2093 mem->as_Store()->memory_size() <= this->memory_size()) {
duke@0 2094 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
duke@0 2095 // For example, 'mem' might be the final state at a conditional return.
duke@0 2096 // Or, 'mem' might be used by some node which is live at the same time
duke@0 2097 // 'this' is live, which might be unschedulable. So, require exactly
duke@0 2098 // ONE user, the 'this' store, until such time as we clone 'mem' for
duke@0 2099 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
duke@0 2100 if (can_reshape) { // (%%% is this an anachronism?)
duke@0 2101 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
duke@0 2102 phase->is_IterGVN());
duke@0 2103 } else {
duke@0 2104 // It's OK to do this in the parser, since DU info is always accurate,
duke@0 2105 // and the parser always refers to nodes via SafePointNode maps.
duke@0 2106 set_req(MemNode::Memory, mem->in(MemNode::Memory));
duke@0 2107 }
duke@0 2108 return this;
duke@0 2109 }
duke@0 2110 }
duke@0 2111
duke@0 2112 // Capture an unaliased, unconditional, simple store into an initializer.
duke@0 2113 // Or, if it is independent of the allocation, hoist it above the allocation.
duke@0 2114 if (ReduceFieldZeroing && /*can_reshape &&*/
duke@0 2115 mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 2116 InitializeNode* init = mem->in(0)->as_Initialize();
duke@0 2117 intptr_t offset = init->can_capture_store(this, phase);
duke@0 2118 if (offset > 0) {
duke@0 2119 Node* moved = init->capture_store(this, offset, phase);
duke@0 2120 // If the InitializeNode captured me, it made a raw copy of me,
duke@0 2121 // and I need to disappear.
duke@0 2122 if (moved != NULL) {
duke@0 2123 // %%% hack to ensure that Ideal returns a new node:
duke@0 2124 mem = MergeMemNode::make(phase->C, mem);
duke@0 2125 return mem; // fold me away
duke@0 2126 }
duke@0 2127 }
duke@0 2128 }
duke@0 2129
duke@0 2130 return NULL; // No further progress
duke@0 2131 }
duke@0 2132
duke@0 2133 //------------------------------Value-----------------------------------------
duke@0 2134 const Type *StoreNode::Value( PhaseTransform *phase ) const {
duke@0 2135 // Either input is TOP ==> the result is TOP
duke@0 2136 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 2137 if( t1 == Type::TOP ) return Type::TOP;
duke@0 2138 const Type *t2 = phase->type( in(MemNode::Address) );
duke@0 2139 if( t2 == Type::TOP ) return Type::TOP;
duke@0 2140 const Type *t3 = phase->type( in(MemNode::ValueIn) );
duke@0 2141 if( t3 == Type::TOP ) return Type::TOP;
duke@0 2142 return Type::MEMORY;
duke@0 2143 }
duke@0 2144
duke@0 2145 //------------------------------Identity---------------------------------------
duke@0 2146 // Remove redundant stores:
duke@0 2147 // Store(m, p, Load(m, p)) changes to m.
duke@0 2148 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
duke@0 2149 Node *StoreNode::Identity( PhaseTransform *phase ) {
duke@0 2150 Node* mem = in(MemNode::Memory);
duke@0 2151 Node* adr = in(MemNode::Address);
duke@0 2152 Node* val = in(MemNode::ValueIn);
duke@0 2153
duke@0 2154 // Load then Store? Then the Store is useless
duke@0 2155 if (val->is_Load() &&
duke@0 2156 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
duke@0 2157 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
duke@0 2158 val->as_Load()->store_Opcode() == Opcode()) {
duke@0 2159 return mem;
duke@0 2160 }
duke@0 2161
duke@0 2162 // Two stores in a row of the same value?
duke@0 2163 if (mem->is_Store() &&
duke@0 2164 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
duke@0 2165 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
duke@0 2166 mem->Opcode() == Opcode()) {
duke@0 2167 return mem;
duke@0 2168 }
duke@0 2169
duke@0 2170 // Store of zero anywhere into a freshly-allocated object?
duke@0 2171 // Then the store is useless.
duke@0 2172 // (It must already have been captured by the InitializeNode.)
duke@0 2173 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
duke@0 2174 // a newly allocated object is already all-zeroes everywhere
duke@0 2175 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
duke@0 2176 return mem;
duke@0 2177 }
duke@0 2178
duke@0 2179 // the store may also apply to zero-bits in an earlier object
duke@0 2180 Node* prev_mem = find_previous_store(phase);
duke@0 2181 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 2182 if (prev_mem != NULL) {
duke@0 2183 Node* prev_val = can_see_stored_value(prev_mem, phase);
duke@0 2184 if (prev_val != NULL && phase->eqv(prev_val, val)) {
duke@0 2185 // prev_val and val might differ by a cast; it would be good
duke@0 2186 // to keep the more informative of the two.
duke@0 2187 return mem;
duke@0 2188 }
duke@0 2189 }
duke@0 2190 }
duke@0 2191
duke@0 2192 return this;
duke@0 2193 }
duke@0 2194
duke@0 2195 //------------------------------match_edge-------------------------------------
duke@0 2196 // Do we Match on this edge index or not? Match only memory & value
duke@0 2197 uint StoreNode::match_edge(uint idx) const {
duke@0 2198 return idx == MemNode::Address || idx == MemNode::ValueIn;
duke@0 2199 }
duke@0 2200
duke@0 2201 //------------------------------cmp--------------------------------------------
duke@0 2202 // Do not common stores up together. They generally have to be split
duke@0 2203 // back up anyways, so do not bother.
duke@0 2204 uint StoreNode::cmp( const Node &n ) const {
duke@0 2205 return (&n == this); // Always fail except on self
duke@0 2206 }
duke@0 2207
duke@0 2208 //------------------------------Ideal_masked_input-----------------------------
duke@0 2209 // Check for a useless mask before a partial-word store
duke@0 2210 // (StoreB ... (AndI valIn conIa) )
duke@0 2211 // If (conIa & mask == mask) this simplifies to
duke@0 2212 // (StoreB ... (valIn) )
duke@0 2213 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
duke@0 2214 Node *val = in(MemNode::ValueIn);
duke@0 2215 if( val->Opcode() == Op_AndI ) {
duke@0 2216 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2217 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
duke@0 2218 set_req(MemNode::ValueIn, val->in(1));
duke@0 2219 return this;
duke@0 2220 }
duke@0 2221 }
duke@0 2222 return NULL;
duke@0 2223 }
duke@0 2224
duke@0 2225
duke@0 2226 //------------------------------Ideal_sign_extended_input----------------------
duke@0 2227 // Check for useless sign-extension before a partial-word store
duke@0 2228 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
duke@0 2229 // If (conIL == conIR && conIR <= num_bits) this simplifies to
duke@0 2230 // (StoreB ... (valIn) )
duke@0 2231 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
duke@0 2232 Node *val = in(MemNode::ValueIn);
duke@0 2233 if( val->Opcode() == Op_RShiftI ) {
duke@0 2234 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2235 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
duke@0 2236 Node *shl = val->in(1);
duke@0 2237 if( shl->Opcode() == Op_LShiftI ) {
duke@0 2238 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
duke@0 2239 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
duke@0 2240 set_req(MemNode::ValueIn, shl->in(1));
duke@0 2241 return this;
duke@0 2242 }
duke@0 2243 }
duke@0 2244 }
duke@0 2245 }
duke@0 2246 return NULL;
duke@0 2247 }
duke@0 2248
duke@0 2249 //------------------------------value_never_loaded-----------------------------------
duke@0 2250 // Determine whether there are any possible loads of the value stored.
duke@0 2251 // For simplicity, we actually check if there are any loads from the
duke@0 2252 // address stored to, not just for loads of the value stored by this node.
duke@0 2253 //
duke@0 2254 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
duke@0 2255 Node *adr = in(Address);
duke@0 2256 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
duke@0 2257 if (adr_oop == NULL)
duke@0 2258 return false;
kvn@223 2259 if (!adr_oop->is_known_instance_field())
duke@0 2260 return false; // if not a distinct instance, there may be aliases of the address
duke@0 2261 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
duke@0 2262 Node *use = adr->fast_out(i);
duke@0 2263 int opc = use->Opcode();
duke@0 2264 if (use->is_Load() || use->is_LoadStore()) {
duke@0 2265 return false;
duke@0 2266 }
duke@0 2267 }
duke@0 2268 return true;
duke@0 2269 }
duke@0 2270
duke@0 2271 //=============================================================================
duke@0 2272 //------------------------------Ideal------------------------------------------
duke@0 2273 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2274 // we can skip the AND operation. If the store is from a sign-extension
duke@0 2275 // (a left shift, then right shift) we can skip both.
duke@0 2276 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2277 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
duke@0 2278 if( progress != NULL ) return progress;
duke@0 2279
duke@0 2280 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
duke@0 2281 if( progress != NULL ) return progress;
duke@0 2282
duke@0 2283 // Finally check the default case
duke@0 2284 return StoreNode::Ideal(phase, can_reshape);
duke@0 2285 }
duke@0 2286
duke@0 2287 //=============================================================================
duke@0 2288 //------------------------------Ideal------------------------------------------
duke@0 2289 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2290 // we can skip the AND operation
duke@0 2291 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2292 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
duke@0 2293 if( progress != NULL ) return progress;
duke@0 2294
duke@0 2295 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
duke@0 2296 if( progress != NULL ) return progress;
duke@0 2297
duke@0 2298 // Finally check the default case
duke@0 2299 return StoreNode::Ideal(phase, can_reshape);
duke@0 2300 }
duke@0 2301
duke@0 2302 //=============================================================================
duke@0 2303 //------------------------------Identity---------------------------------------
duke@0 2304 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
duke@0 2305 // No need to card mark when storing a null ptr
duke@0 2306 Node* my_store = in(MemNode::OopStore);
duke@0 2307 if (my_store->is_Store()) {
duke@0 2308 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
duke@0 2309 if( t1 == TypePtr::NULL_PTR ) {
duke@0 2310 return in(MemNode::Memory);
duke@0 2311 }
duke@0 2312 }
duke@0 2313 return this;
duke@0 2314 }
duke@0 2315
duke@0 2316 //------------------------------Value-----------------------------------------
duke@0 2317 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
kvn@43 2318 // Either input is TOP ==> the result is TOP
kvn@43 2319 const Type *t = phase->type( in(MemNode::Memory) );
kvn@43 2320 if( t == Type::TOP ) return Type::TOP;
kvn@43 2321 t = phase->type( in(MemNode::Address) );
kvn@43 2322 if( t == Type::TOP ) return Type::TOP;
kvn@43 2323 t = phase->type( in(MemNode::ValueIn) );
kvn@43 2324 if( t == Type::TOP ) return Type::TOP;
duke@0 2325 // If extra input is TOP ==> the result is TOP
kvn@43 2326 t = phase->type( in(MemNode::OopStore) );
kvn@43 2327 if( t == Type::TOP ) return Type::TOP;
duke@0 2328
duke@0 2329 return StoreNode::Value( phase );
duke@0 2330 }
duke@0 2331
duke@0 2332
duke@0 2333 //=============================================================================
duke@0 2334 //----------------------------------SCMemProjNode------------------------------
duke@0 2335 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
duke@0 2336 {
duke@0 2337 return bottom_type();
duke@0 2338 }
duke@0 2339
duke@0 2340 //=============================================================================
duke@0 2341 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
duke@0 2342 init_req(MemNode::Control, c );
duke@0 2343 init_req(MemNode::Memory , mem);
duke@0 2344 init_req(MemNode::Address, adr);
duke@0 2345 init_req(MemNode::ValueIn, val);
duke@0 2346 init_req( ExpectedIn, ex );
duke@0 2347 init_class_id(Class_LoadStore);
duke@0 2348
duke@0 2349 }
duke@0 2350
duke@0 2351 //=============================================================================
duke@0 2352 //-------------------------------adr_type--------------------------------------
duke@0 2353 // Do we Match on this edge index or not? Do not match memory
duke@0 2354 const TypePtr* ClearArrayNode::adr_type() const {
duke@0 2355 Node *adr = in(3);
duke@0 2356 return MemNode::calculate_adr_type(adr->bottom_type());
duke@0 2357 }
duke@0 2358
duke@0 2359 //------------------------------match_edge-------------------------------------
duke@0 2360 // Do we Match on this edge index or not? Do not match memory
duke@0 2361 uint ClearArrayNode::match_edge(uint idx) const {
duke@0 2362 return idx > 1;
duke@0 2363 }
duke@0 2364
duke@0 2365 //------------------------------Identity---------------------------------------
duke@0 2366 // Clearing a zero length array does nothing
duke@0 2367 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
never@68 2368 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
duke@0 2369 }
duke@0 2370
duke@0 2371 //------------------------------Idealize---------------------------------------
duke@0 2372 // Clearing a short array is faster with stores
duke@0 2373 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2374 const int unit = BytesPerLong;
duke@0 2375 const TypeX* t = phase->type(in(2))->isa_intptr_t();
duke@0 2376 if (!t) return NULL;
duke@0 2377 if (!t->is_con()) return NULL;
duke@0 2378 intptr_t raw_count = t->get_con();
duke@0 2379 intptr_t size = raw_count;
duke@0 2380 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
duke@0 2381 // Clearing nothing uses the Identity call.
duke@0 2382 // Negative clears are possible on dead ClearArrays
duke@0 2383 // (see jck test stmt114.stmt11402.val).
duke@0 2384 if (size <= 0 || size % unit != 0) return NULL;
duke@0 2385 intptr_t count = size / unit;
duke@0 2386 // Length too long; use fast hardware clear
duke@0 2387 if (size > Matcher::init_array_short_size) return NULL;
duke@0 2388 Node *mem = in(1);
duke@0 2389 if( phase->type(mem)==Type::TOP ) return NULL;
duke@0 2390 Node *adr = in(3);
duke@0 2391 const Type* at = phase->type(adr);
duke@0 2392 if( at==Type::TOP ) return NULL;
duke@0 2393 const TypePtr* atp = at->isa_ptr();
duke@0 2394 // adjust atp to be the correct array element address type
duke@0 2395 if (atp == NULL) atp = TypePtr::BOTTOM;
duke@0 2396 else atp = atp->add_offset(Type::OffsetBot);
duke@0 2397 // Get base for derived pointer purposes
duke@0 2398 if( adr->Opcode() != Op_AddP ) Unimplemented();
duke@0 2399 Node *base = adr->in(1);
duke@0 2400
duke@0 2401 Node *zero = phase->makecon(TypeLong::ZERO);
duke@0 2402 Node *off = phase->MakeConX(BytesPerLong);
duke@0 2403 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2404 count--;
duke@0 2405 while( count-- ) {
duke@0 2406 mem = phase->transform(mem);
duke@0 2407 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
duke@0 2408 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2409 }
duke@0 2410 return mem;
duke@0 2411 }
duke@0 2412
duke@0 2413 //----------------------------clear_memory-------------------------------------
duke@0 2414 // Generate code to initialize object storage to zero.
duke@0 2415 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2416 intptr_t start_offset,
duke@0 2417 Node* end_offset,
duke@0 2418 PhaseGVN* phase) {
duke@0 2419 Compile* C = phase->C;
duke@0 2420 intptr_t offset = start_offset;
duke@0 2421
duke@0 2422 int unit = BytesPerLong;
duke@0 2423 if ((offset % unit) != 0) {
duke@0 2424 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
duke@0 2425 adr = phase->transform(adr);
duke@0 2426 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2427 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2428 mem = phase->transform(mem);
duke@0 2429 offset += BytesPerInt;
duke@0 2430 }
duke@0 2431 assert((offset % unit) == 0, "");
duke@0 2432
duke@0 2433 // Initialize the remaining stuff, if any, with a ClearArray.
duke@0 2434 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
duke@0 2435 }
duke@0 2436
duke@0 2437 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2438 Node* start_offset,
duke@0 2439 Node* end_offset,
duke@0 2440 PhaseGVN* phase) {
never@68 2441 if (start_offset == end_offset) {
never@68 2442 // nothing to do
never@68 2443 return mem;
never@68 2444 }
never@68 2445
duke@0 2446 Compile* C = phase->C;
duke@0 2447 int unit = BytesPerLong;
duke@0 2448 Node* zbase = start_offset;
duke@0 2449 Node* zend = end_offset;
duke@0 2450
duke@0 2451 // Scale to the unit required by the CPU:
duke@0 2452 if (!Matcher::init_array_count_is_in_bytes) {
duke@0 2453 Node* shift = phase->intcon(exact_log2(unit));
duke@0 2454 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
duke@0 2455 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
duke@0 2456 }
duke@0 2457
duke@0 2458 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
duke@0 2459 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
duke@0 2460
duke@0 2461 // Bulk clear double-words
duke@0 2462 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
duke@0 2463 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
duke@0 2464 return phase->transform(mem);
duke@0 2465 }
duke@0 2466
duke@0 2467 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2468 intptr_t start_offset,
duke@0 2469 intptr_t end_offset,
duke@0 2470 PhaseGVN* phase) {
never@68 2471 if (start_offset == end_offset) {
never@68 2472 // nothing to do
never@68 2473 return mem;
never@68 2474 }
never@68 2475
duke@0 2476 Compile* C = phase->C;
duke@0 2477 assert((end_offset % BytesPerInt) == 0, "odd end offset");
duke@0 2478 intptr_t done_offset = end_offset;
duke@0 2479 if ((done_offset % BytesPerLong) != 0) {
duke@0 2480 done_offset -= BytesPerInt;
duke@0 2481 }
duke@0 2482 if (done_offset > start_offset) {
duke@0 2483 mem = clear_memory(ctl, mem, dest,
duke@0 2484 start_offset, phase->MakeConX(done_offset), phase);
duke@0 2485 }
duke@0 2486 if (done_offset < end_offset) { // emit the final 32-bit store
duke@0 2487 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
duke@0 2488 adr = phase->transform(adr);
duke@0 2489 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2490 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2491 mem = phase->transform(mem);
duke@0 2492 done_offset += BytesPerInt;
duke@0 2493 }
duke@0 2494 assert(done_offset == end_offset, "");
duke@0 2495 return mem;
duke@0 2496 }
duke@0 2497
duke@0 2498 //=============================================================================
duke@0 2499 // Do we match on this edge? No memory edges
duke@0 2500 uint StrCompNode::match_edge(uint idx) const {
duke@0 2501 return idx == 5 || idx == 6;
duke@0 2502 }
duke@0 2503
duke@0 2504 //------------------------------Ideal------------------------------------------
duke@0 2505 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2506 // control copies
duke@0 2507 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2508 return remove_dead_region(phase, can_reshape) ? this : NULL;
duke@0 2509 }
duke@0 2510
cfang@681 2511 // Do we match on this edge? No memory edges
cfang@681 2512 uint StrEqualsNode::match_edge(uint idx) const {
cfang@681 2513 return idx == 5 || idx == 6;
cfang@681 2514 }
cfang@681 2515
cfang@681 2516 //------------------------------Ideal------------------------------------------
cfang@681 2517 // Return a node which is more "ideal" than the current node. Strip out
cfang@681 2518 // control copies
cfang@681 2519 Node *StrEqualsNode::Ideal(PhaseGVN *phase, bool can_reshape){
cfang@681 2520 return remove_dead_region(phase, can_reshape) ? this : NULL;
cfang@681 2521 }
cfang@681 2522
cfang@681 2523 //=============================================================================
cfang@681 2524 // Do we match on this edge? No memory edges
cfang@681 2525 uint StrIndexOfNode::match_edge(uint idx) const {
cfang@681 2526 return idx == 5 || idx == 6;
cfang@681 2527 }
cfang@681 2528
cfang@681 2529 //------------------------------Ideal------------------------------------------
cfang@681 2530 // Return a node which is more "ideal" than the current node. Strip out
cfang@681 2531 // control copies
cfang@681 2532 Node *StrIndexOfNode::Ideal(PhaseGVN *phase, bool can_reshape){
cfang@681 2533 return remove_dead_region(phase, can_reshape) ? this : NULL;
cfang@681 2534 }
cfang@681 2535
rasbold@169 2536 //------------------------------Ideal------------------------------------------
rasbold@169 2537 // Return a node which is more "ideal" than the current node. Strip out
rasbold@169 2538 // control copies
rasbold@169 2539 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){
rasbold@169 2540 return remove_dead_region(phase, can_reshape) ? this : NULL;
rasbold@169 2541 }
rasbold@169 2542
duke@0 2543 //=============================================================================
duke@0 2544 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
duke@0 2545 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
duke@0 2546 _adr_type(C->get_adr_type(alias_idx))
duke@0 2547 {
duke@0 2548 init_class_id(Class_MemBar);
duke@0 2549 Node* top = C->top();
duke@0 2550 init_req(TypeFunc::I_O,top);
duke@0 2551 init_req(TypeFunc::FramePtr,top);
duke@0 2552 init_req(TypeFunc::ReturnAdr,top);
duke@0 2553 if (precedent != NULL)
duke@0 2554 init_req(TypeFunc::Parms, precedent);
duke@0 2555 }
duke@0 2556
duke@0 2557 //------------------------------cmp--------------------------------------------
duke@0 2558 uint MemBarNode::hash() const { return NO_HASH; }
duke@0 2559 uint MemBarNode::cmp( const Node &n ) const {
duke@0 2560 return (&n == this); // Always fail except on self
duke@0 2561 }
duke@0 2562
duke@0 2563 //------------------------------make-------------------------------------------
duke@0 2564 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
duke@0 2565 int len = Precedent + (pn == NULL? 0: 1);
duke@0 2566 switch (opcode) {
duke@0 2567 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
duke@0 2568 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
duke@0 2569 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
duke@0 2570 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
duke@0 2571 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
duke@0 2572 default: ShouldNotReachHere(); return NULL;
duke@0 2573 }
duke@0 2574 }
duke@0 2575
duke@0 2576 //------------------------------Ideal------------------------------------------
duke@0 2577 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2578 // control copies
duke@0 2579 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
kvn@305 2580 return remove_dead_region(phase, can_reshape) ? this : NULL;
duke@0 2581 }
duke@0 2582
duke@0 2583 //------------------------------Value------------------------------------------
duke@0 2584 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
duke@0 2585 if( !in(0) ) return Type::TOP;
duke@0 2586 if( phase->type(in(0)) == Type::TOP )
duke@0 2587 return Type::TOP;
duke@0 2588 return TypeTuple::MEMBAR;
duke@0 2589 }
duke@0 2590
duke@0 2591 //------------------------------match------------------------------------------
duke@0 2592 // Construct projections for memory.
duke@0 2593 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
duke@0 2594 switch (proj->_con) {
duke@0 2595 case TypeFunc::Control:
duke@0 2596 case TypeFunc::Memory:
duke@0 2597 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
duke@0 2598 }
duke@0 2599 ShouldNotReachHere();
duke@0 2600 return NULL;
duke@0 2601 }
duke@0 2602
duke@0 2603 //===========================InitializeNode====================================
duke@0 2604 // SUMMARY:
duke@0 2605 // This node acts as a memory barrier on raw memory, after some raw stores.
duke@0 2606 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
duke@0 2607 // The Initialize can 'capture' suitably constrained stores as raw inits.
duke@0 2608 // It can coalesce related raw stores into larger units (called 'tiles').
duke@0 2609 // It can avoid zeroing new storage for memory units which have raw inits.
duke@0 2610 // At macro-expansion, it is marked 'complete', and does not optimize further.
duke@0 2611 //
duke@0 2612 // EXAMPLE:
duke@0 2613 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
duke@0 2614 // ctl = incoming control; mem* = incoming memory
duke@0 2615 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
duke@0 2616 // First allocate uninitialized memory and fill in the header:
duke@0 2617 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
duke@0 2618 // ctl := alloc.Control; mem* := alloc.Memory*
duke@0 2619 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
duke@0 2620 // Then initialize to zero the non-header parts of the raw memory block:
duke@0 2621 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
duke@0 2622 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
duke@0 2623 // After the initialize node executes, the object is ready for service:
duke@0 2624 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
duke@0 2625 // Suppose its body is immediately initialized as {1,2}:
duke@0 2626 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2627 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2628 // mem.SLICE(#short[*]) := store2
duke@0 2629 //
duke@0 2630 // DETAILS:
duke@0 2631 // An InitializeNode collects and isolates object initialization after
duke@0 2632 // an AllocateNode and before the next possible safepoint. As a
duke@0 2633 // memory barrier (MemBarNode), it keeps critical stores from drifting
duke@0 2634 // down past any safepoint or any publication of the allocation.
duke@0 2635 // Before this barrier, a newly-allocated object may have uninitialized bits.
duke@0 2636 // After this barrier, it may be treated as a real oop, and GC is allowed.
duke@0 2637 //
duke@0 2638 // The semantics of the InitializeNode include an implicit zeroing of
duke@0 2639 // the new object from object header to the end of the object.
duke@0 2640 // (The object header and end are determined by the AllocateNode.)
duke@0 2641 //
duke@0 2642 // Certain stores may be added as direct inputs to the InitializeNode.
duke@0 2643 // These stores must update raw memory, and they must be to addresses
duke@0 2644 // derived from the raw address produced by AllocateNode, and with
duke@0 2645 // a constant offset. They must be ordered by increasing offset.
duke@0 2646 // The first one is at in(RawStores), the last at in(req()-1).
duke@0 2647 // Unlike most memory operations, they are not linked in a chain,
duke@0 2648 // but are displayed in parallel as users of the rawmem output of
duke@0 2649 // the allocation.
duke@0 2650 //
duke@0 2651 // (See comments in InitializeNode::capture_store, which continue
duke@0 2652 // the example given above.)
duke@0 2653 //
duke@0 2654 // When the associated Allocate is macro-expanded, the InitializeNode
duke@0 2655 // may be rewritten to optimize collected stores. A ClearArrayNode
duke@0 2656 // may also be created at that point to represent any required zeroing.
duke@0 2657 // The InitializeNode is then marked 'complete', prohibiting further
duke@0 2658 // capturing of nearby memory operations.
duke@0 2659 //
duke@0 2660 // During macro-expansion, all captured initializations which store
twisti@605 2661 // constant values of 32 bits or smaller are coalesced (if advantageous)
duke@0 2662 // into larger 'tiles' 32 or 64 bits. This allows an object to be
duke@0 2663 // initialized in fewer memory operations. Memory words which are
duke@0 2664 // covered by neither tiles nor non-constant stores are pre-zeroed
duke@0 2665 // by explicit stores of zero. (The code shape happens to do all
duke@0 2666 // zeroing first, then all other stores, with both sequences occurring
duke@0 2667 // in order of ascending offsets.)
duke@0 2668 //
duke@0 2669 // Alternatively, code may be inserted between an AllocateNode and its
duke@0 2670 // InitializeNode, to perform arbitrary initialization of the new object.
duke@0 2671 // E.g., the object copying intrinsics insert complex data transfers here.
duke@0 2672 // The initialization must then be marked as 'complete' disable the
duke@0 2673 // built-in zeroing semantics and the collection of initializing stores.
duke@0 2674 //
duke@0 2675 // While an InitializeNode is incomplete, reads from the memory state
duke@0 2676 // produced by it are optimizable if they match the control edge and
duke@0 2677 // new oop address associated with the allocation/initialization.
duke@0 2678 // They return a stored value (if the offset matches) or else zero.
duke@0 2679 // A write to the memory state, if it matches control and address,
duke@0 2680 // and if it is to a constant offset, may be 'captured' by the
duke@0 2681 // InitializeNode. It is cloned as a raw memory operation and rewired
duke@0 2682 // inside the initialization, to the raw oop produced by the allocation.
duke@0 2683 // Operations on addresses which are provably distinct (e.g., to
duke@0 2684 // other AllocateNodes) are allowed to bypass the initialization.
duke@0 2685 //
duke@0 2686 // The effect of all this is to consolidate object initialization
duke@0 2687 // (both arrays and non-arrays, both piecewise and bulk) into a
duke@0 2688 // single location, where it can be optimized as a unit.
duke@0 2689 //
duke@0 2690 // Only stores with an offset less than TrackedInitializationLimit words
duke@0 2691 // will be considered for capture by an InitializeNode. This puts a
duke@0 2692 // reasonable limit on the complexity of optimized initializations.
duke@0 2693
duke@0 2694 //---------------------------InitializeNode------------------------------------
duke@0 2695 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
duke@0 2696 : _is_complete(false),
duke@0 2697 MemBarNode(C, adr_type, rawoop)
duke@0 2698 {
duke@0 2699 init_class_id(Class_Initialize);
duke@0 2700
duke@0 2701 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
duke@0 2702 assert(in(RawAddress) == rawoop, "proper init");
duke@0 2703 // Note: allocation() can be NULL, for secondary initialization barriers
duke@0 2704 }
duke@0 2705
duke@0 2706 // Since this node is not matched, it will be processed by the
duke@0 2707 // register allocator. Declare that there are no constraints
duke@0 2708 // on the allocation of the RawAddress edge.
duke@0 2709 const RegMask &InitializeNode::in_RegMask(uint idx) const {
duke@0 2710 // This edge should be set to top, by the set_complete. But be conservative.
duke@0 2711 if (idx == InitializeNode::RawAddress)
duke@0 2712 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
duke@0 2713 return RegMask::Empty;
duke@0 2714 }
duke@0 2715
duke@0 2716 Node* InitializeNode::memory(uint alias_idx) {
duke@0 2717 Node* mem = in(Memory);
duke@0 2718 if (mem->is_MergeMem()) {
duke@0 2719 return mem->as_MergeMem()->memory_at(alias_idx);
duke@0 2720 } else {
duke@0 2721 // incoming raw memory is not split
duke@0 2722 return mem;
duke@0 2723 }
duke@0 2724 }
duke@0 2725
duke@0 2726 bool InitializeNode::is_non_zero() {
duke@0 2727 if (is_complete()) return false;
duke@0 2728 remove_extra_zeroes();
duke@0 2729 return (req() > RawStores);
duke@0 2730 }
duke@0 2731
duke@0 2732 void InitializeNode::set_complete(PhaseGVN* phase) {
duke@0 2733 assert(!is_complete(), "caller responsibility");
duke@0 2734 _is_complete = true;
duke@0 2735
duke@0 2736 // After this node is complete, it contains a bunch of
duke@0 2737 // raw-memory initializations. There is no need for
duke@0 2738 // it to have anything to do with non-raw memory effects.
duke@0 2739 // Therefore, tell all non-raw users to re-optimize themselves,
duke@0 2740 // after skipping the memory effects of this initialization.
duke@0 2741 PhaseIterGVN* igvn = phase->is_IterGVN();
duke@0 2742 if (igvn) igvn->add_users_to_worklist(this);
duke@0 2743 }
duke@0 2744
duke@0 2745 // convenience function
duke@0 2746 // return false if the init contains any stores already
duke@0 2747 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
duke@0 2748 InitializeNode* init = initialization();
duke@0 2749 if (init == NULL || init->is_complete()) return false;
duke@0 2750 init->remove_extra_zeroes();
duke@0 2751 // for now, if this allocation has already collected any inits, bail:
duke@0 2752 if (init->is_non_zero()) return false;
duke@0 2753 init->set_complete(phase);
duke@0 2754 return true;
duke@0 2755 }
duke@0 2756
duke@0 2757 void InitializeNode::remove_extra_zeroes() {
duke@0 2758 if (req() == RawStores) return;
duke@0 2759 Node* zmem = zero_memory();
duke@0 2760 uint fill = RawStores;
duke@0 2761 for (uint i = fill; i < req(); i++) {
duke@0 2762 Node* n = in(i);
duke@0 2763 if (n->is_top() || n == zmem) continue; // skip
duke@0 2764 if (fill < i) set_req(fill, n); // compact
duke@0 2765 ++fill;
duke@0 2766 }
duke@0 2767 // delete any empty spaces created:
duke@0 2768 while (fill < req()) {
duke@0 2769 del_req(fill);
duke@0 2770 }
duke@0 2771 }
duke@0 2772
duke@0 2773 // Helper for remembering which stores go with which offsets.
duke@0 2774 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
duke@0 2775 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
duke@0 2776 intptr_t offset = -1;
duke@0 2777 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
duke@0 2778 phase, offset);
duke@0 2779 if (base == NULL) return -1; // something is dead,
duke@0 2780 if (offset < 0) return -1; // dead, dead
duke@0 2781 return offset;
duke@0 2782 }
duke@0 2783
duke@0 2784 // Helper for proving that an initialization expression is
duke@0 2785 // "simple enough" to be folded into an object initialization.
duke@0 2786 // Attempts to prove that a store's initial value 'n' can be captured
duke@0 2787 // within the initialization without creating a vicious cycle, such as:
duke@0 2788 // { Foo p = new Foo(); p.next = p; }
duke@0 2789 // True for constants and parameters and small combinations thereof.
duke@0 2790 bool InitializeNode::detect_init_independence(Node* n,
duke@0 2791 bool st_is_pinned,
duke@0 2792 int& count) {
duke@0 2793 if (n == NULL) return true; // (can this really happen?)
duke@0 2794 if (n->is_Proj()) n = n->in(0);
duke@0 2795 if (n == this) return false; // found a cycle
duke@0 2796 if (n->is_Con()) return true;
duke@0 2797 if (n->is_Start()) return true; // params, etc., are OK
duke@0 2798 if (n->is_Root()) return true; // even better
duke@0 2799
duke@0 2800 Node* ctl = n->in(0);
duke@0 2801 if (ctl != NULL && !ctl->is_top()) {
duke@0 2802 if (ctl->is_Proj()) ctl = ctl->in(0);
duke@0 2803 if (ctl == this) return false;
duke@0 2804
duke@0 2805 // If we already know that the enclosing memory op is pinned right after
duke@0 2806 // the init, then any control flow that the store has picked up
duke@0 2807 // must have preceded the init, or else be equal to the init.
duke@0 2808 // Even after loop optimizations (which might change control edges)
duke@0 2809 // a store is never pinned *before* the availability of its inputs.
kvn@119 2810 if (!MemNode::all_controls_dominate(n, this))
duke@0 2811 return false; // failed to prove a good control
duke@0 2812
duke@0 2813 }
duke@0 2814
duke@0 2815 // Check data edges for possible dependencies on 'this'.
duke@0 2816 if ((count += 1) > 20) return false; // complexity limit
duke@0 2817 for (uint i = 1; i < n->req(); i++) {
duke@0 2818 Node* m = n->in(i);
duke@0 2819 if (m == NULL || m == n || m->is_top()) continue;
duke@0 2820 uint first_i = n->find_edge(m);
duke@0 2821 if (i != first_i) continue; // process duplicate edge just once
duke@0 2822 if (!detect_init_independence(m, st_is_pinned, count)) {
duke@0 2823 return false;
duke@0 2824 }
duke@0 2825 }
duke@0 2826
duke@0 2827 return true;
duke@0 2828 }
duke@0 2829
duke@0 2830 // Here are all the checks a Store must pass before it can be moved into
duke@0 2831 // an initialization. Returns zero if a check fails.
duke@0 2832 // On success, returns the (constant) offset to which the store applies,
duke@0 2833 // within the initialized memory.
duke@0 2834 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
duke@0 2835 const int FAIL = 0;
duke@0 2836 if (st->req() != MemNode::ValueIn + 1)
duke@0 2837 return FAIL; // an inscrutable StoreNode (card mark?)
duke@0 2838 Node* ctl = st->in(MemNode::Control);
duke@0 2839 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
duke@0 2840 return FAIL; // must be unconditional after the initialization
duke@0 2841 Node* mem = st->in(MemNode::Memory);
duke@0 2842 if (!(mem->is_Proj() && mem->in(0) == this))
duke@0 2843 return FAIL; // must not be preceded by other stores
duke@0 2844 Node* adr = st->in(MemNode::Address);
duke@0 2845 intptr_t offset;
duke@0 2846 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
duke@0 2847 if (alloc == NULL)
duke@0 2848 return FAIL; // inscrutable address
duke@0 2849 if (alloc != allocation())
duke@0 2850 return FAIL; // wrong allocation! (store needs to float up)
duke@0 2851 Node* val = st->in(MemNode::ValueIn);
duke@0 2852 int complexity_count = 0;
duke@0 2853 if (!detect_init_independence(val, true, complexity_count))
duke@0 2854 return FAIL; // stored value must be 'simple enough'
duke@0 2855
duke@0 2856 return offset; // success
duke@0 2857 }
duke@0 2858
duke@0 2859 // Find the captured store in(i) which corresponds to the range
duke@0 2860 // [start..start+size) in the initialized object.
duke@0 2861 // If there is one, return its index i. If there isn't, return the
duke@0 2862 // negative of the index where it should be inserted.
duke@0 2863 // Return 0 if the queried range overlaps an initialization boundary
duke@0 2864 // or if dead code is encountered.
duke@0 2865 // If size_in_bytes is zero, do not bother with overlap checks.
duke@0 2866 int InitializeNode::captured_store_insertion_point(intptr_t start,
duke@0 2867 int size_in_bytes,
duke@0 2868 PhaseTransform* phase) {
duke@0 2869 const int FAIL = 0, MAX_STORE = BytesPerLong;
duke@0 2870
duke@0 2871 if (is_complete())
duke@0 2872 return FAIL; // arraycopy got here first; punt
duke@0 2873
duke@0 2874 assert(allocation() != NULL, "must be present");
duke@0 2875
duke@0 2876 // no negatives, no header fields:
coleenp@113 2877 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
duke@0 2878
duke@0 2879 // after a certain size, we bail out on tracking all the stores:
duke@0 2880 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2881 if (start >= ti_limit) return FAIL;
duke@0 2882
duke@0 2883 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
duke@0 2884 if (i >= limit) return -(int)i; // not found; here is where to put it
duke@0 2885
duke@0 2886 Node* st = in(i);
duke@0 2887 intptr_t st_off = get_store_offset(st, phase);
duke@0 2888 if (st_off < 0) {
duke@0 2889 if (st != zero_memory()) {
duke@0 2890 return FAIL; // bail out if there is dead garbage
duke@0 2891 }
duke@0 2892 } else if (st_off > start) {
duke@0 2893 // ...we are done, since stores are ordered
duke@0 2894 if (st_off < start + size_in_bytes) {
duke@0 2895 return FAIL; // the next store overlaps
duke@0 2896 }
duke@0 2897 return -(int)i; // not found; here is where to put it
duke@0 2898 } else if (st_off < start) {
duke@0 2899 if (size_in_bytes != 0 &&
duke@0 2900 start < st_off + MAX_STORE &&
duke@0 2901 start < st_off + st->as_Store()->memory_size()) {
duke@0 2902 return FAIL; // the previous store overlaps
duke@0 2903 }
duke@0 2904 } else {
duke@0 2905 if (size_in_bytes != 0 &&
duke@0 2906 st->as_Store()->memory_size() != size_in_bytes) {
duke@0 2907 return FAIL; // mismatched store size
duke@0 2908 }
duke@0 2909 return i;
duke@0 2910 }
duke@0 2911
duke@0 2912 ++i;
duke@0 2913 }
duke@0 2914 }
duke@0 2915
duke@0 2916 // Look for a captured store which initializes at the offset 'start'
duke@0 2917 // with the given size. If there is no such store, and no other
duke@0 2918 // initialization interferes, then return zero_memory (the memory
duke@0 2919 // projection of the AllocateNode).
duke@0 2920 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
duke@0 2921 PhaseTransform* phase) {
duke@0 2922 assert(stores_are_sane(phase), "");
duke@0 2923 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2924 if (i == 0) {
duke@0 2925 return NULL; // something is dead
duke@0 2926 } else if (i < 0) {
duke@0 2927 return zero_memory(); // just primordial zero bits here
duke@0 2928 } else {
duke@0 2929 Node* st = in(i); // here is the store at this position
duke@0 2930 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
duke@0 2931 return st;
duke@0 2932 }
duke@0 2933 }
duke@0 2934
duke@0 2935 // Create, as a raw pointer, an address within my new object at 'offset'.
duke@0 2936 Node* InitializeNode::make_raw_address(intptr_t offset,
duke@0 2937 PhaseTransform* phase) {
duke@0 2938 Node* addr = in(RawAddress);
duke@0 2939 if (offset != 0) {
duke@0 2940 Compile* C = phase->C;
duke@0 2941 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
duke@0 2942 phase->MakeConX(offset)) );
duke@0 2943 }
duke@0 2944 return addr;
duke@0 2945 }
duke@0 2946
duke@0 2947 // Clone the given store, converting it into a raw store
duke@0 2948 // initializing a field or element of my new object.
duke@0 2949 // Caller is responsible for retiring the original store,
duke@0 2950 // with subsume_node or the like.
duke@0 2951 //
duke@0 2952 // From the example above InitializeNode::InitializeNode,
duke@0 2953 // here are the old stores to be captured:
duke@0 2954 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2955 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2956 //
duke@0 2957 // Here is the changed code; note the extra edges on init:
duke@0 2958 // alloc = (Allocate ...)
duke@0 2959 // rawoop = alloc.RawAddress
duke@0 2960 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
duke@0 2961 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
duke@0 2962 // init = (Initialize alloc.Control alloc.Memory rawoop
duke@0 2963 // rawstore1 rawstore2)
duke@0 2964 //
duke@0 2965 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
duke@0 2966 PhaseTransform* phase) {
duke@0 2967 assert(stores_are_sane(phase), "");
duke@0 2968
duke@0 2969 if (start < 0) return NULL;
duke@0 2970 assert(can_capture_store(st, phase) == start, "sanity");
duke@0 2971
duke@0 2972 Compile* C = phase->C;
duke@0 2973 int size_in_bytes = st->memory_size();
duke@0 2974 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2975 if (i == 0) return NULL; // bail out
duke@0 2976 Node* prev_mem = NULL; // raw memory for the captured store
duke@0 2977 if (i > 0) {
duke@0 2978 prev_mem = in(i); // there is a pre-existing store under this one
duke@0 2979 set_req(i, C->top()); // temporarily disconnect it
duke@0 2980 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
duke@0 2981 } else {
duke@0 2982 i = -i; // no pre-existing store
duke@0 2983 prev_mem = zero_memory(); // a slice of the newly allocated object
duke@0 2984 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
duke@0 2985 set_req(--i, C->top()); // reuse this edge; it has been folded away
duke@0 2986 else
duke@0 2987 ins_req(i, C->top()); // build a new edge
duke@0 2988 }
duke@0 2989 Node* new_st = st->clone();
duke@0 2990 new_st->set_req(MemNode::Control, in(Control));
duke@0 2991 new_st->set_req(MemNode::Memory, prev_mem);
duke@0 2992 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
duke@0 2993 new_st = phase->transform(new_st);
duke@0 2994
duke@0 2995 // At this point, new_st might have swallowed a pre-existing store
duke@0 2996 // at the same offset, or perhaps new_st might have disappeared,
duke@0 2997 // if it redundantly stored the same value (or zero to fresh memory).
duke@0 2998
duke@0 2999 // In any case, wire it in:
duke@0 3000 set_req(i, new_st);
duke@0 3001
duke@0 3002 // The caller may now kill the old guy.
duke@0 3003 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
duke@0 3004 assert(check_st == new_st || check_st == NULL, "must be findable");
duke@0 3005 assert(!is_complete(), "");
duke@0 3006 return new_st;
duke@0 3007 }
duke@0 3008
duke@0 3009 static bool store_constant(jlong* tiles, int num_tiles,
duke@0 3010 intptr_t st_off, int st_size,
duke@0 3011 jlong con) {
duke@0 3012 if ((st_off & (st_size-1)) != 0)
duke@0 3013 return false; // strange store offset (assume size==2**N)
duke@0 3014 address addr = (address)tiles + st_off;
duke@0 3015 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
duke@0 3016 switch (st_size) {
duke@0 3017 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
duke@0 3018 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
duke@0 3019 case sizeof(jint): *(jint*) addr = (jint) con; break;
duke@0 3020 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
duke@0 3021 default: return false; // strange store size (detect size!=2**N here)
duke@0 3022 }
duke@0 3023 return true; // return success to caller
duke@0 3024 }
duke@0 3025
duke@0 3026 // Coalesce subword constants into int constants and possibly
duke@0 3027 // into long constants. The goal, if the CPU permits,
duke@0 3028 // is to initialize the object with a small number of 64-bit tiles.
duke@0 3029 // Also, convert floating-point constants to bit patterns.
duke@0 3030 // Non-constants are not relevant to this pass.
duke@0 3031 //
duke@0 3032 // In terms of the running example on InitializeNode::InitializeNode
duke@0 3033 // and InitializeNode::capture_store, here is the transformation
duke@0 3034 // of rawstore1 and rawstore2 into rawstore12:
duke@0 3035 // alloc = (Allocate ...)
duke@0 3036 // rawoop = alloc.RawAddress
duke@0 3037 // tile12 = 0x00010002
duke@0 3038 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
duke@0 3039 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
duke@0 3040 //
duke@0 3041 void
duke@0 3042 InitializeNode::coalesce_subword_stores(intptr_t header_size,
duke@0 3043 Node* size_in_bytes,
duke@0 3044 PhaseGVN* phase) {
duke@0 3045 Compile* C = phase->C;
duke@0 3046
duke@0 3047 assert(stores_are_sane(phase), "");
duke@0 3048 // Note: After this pass, they are not completely sane,
duke@0 3049 // since there may be some overlaps.
duke@0 3050
duke@0 3051 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
duke@0 3052
duke@0 3053 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 3054 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
duke@0 3055 size_limit = MIN2(size_limit, ti_limit);
duke@0 3056 size_limit = align_size_up(size_limit, BytesPerLong);
duke@0 3057 int num_tiles = size_limit / BytesPerLong;
duke@0 3058
duke@0 3059 // allocate space for the tile map:
duke@0 3060 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
duke@0 3061 jlong tiles_buf[small_len];
duke@0 3062 Node* nodes_buf[small_len];
duke@0 3063 jlong inits_buf[small_len];
duke@0 3064 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
duke@0 3065 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 3066 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
duke@0 3067 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
duke@0 3068 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
duke@0 3069 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 3070 // tiles: exact bitwise model of all primitive constants
duke@0 3071 // nodes: last constant-storing node subsumed into the tiles model
duke@0 3072 // inits: which bytes (in each tile) are touched by any initializations
duke@0 3073
duke@0 3074 //// Pass A: Fill in the tile model with any relevant stores.
duke@0 3075
duke@0 3076 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
duke@0 3077 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
duke@0 3078 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
duke@0 3079 Node* zmem = zero_memory(); // initially zero memory state
duke@0 3080 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 3081 Node* st = in(i);
duke@0 3082 intptr_t st_off = get_store_offset(st, phase);
duke@0 3083
duke@0 3084 // Figure out the store's offset and constant value:
duke@0 3085 if (st_off < header_size) continue; //skip (ignore header)
duke@0 3086 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
duke@0 3087 int st_size = st->as_Store()->memory_size();
duke@0 3088 if (st_off + st_size > size_limit) break;
duke@0 3089
duke@0 3090 // Record which bytes are touched, whether by constant or not.
duke@0 3091 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
duke@0 3092 continue; // skip (strange store size)
duke@0 3093
duke@0 3094 const Type* val = phase->type(st->in(MemNode::ValueIn));
duke@0 3095 if (!val->singleton()) continue; //skip (non-con store)
duke@0 3096 BasicType type = val->basic_type();
duke@0 3097
duke@0 3098 jlong con = 0;
duke@0 3099 switch (type) {
duke@0 3100 case T_INT: con = val->is_int()->get_con(); break;
duke@0 3101 case T_LONG: con = val->is_long()->get_con(); break;
duke@0 3102 case T_FLOAT: con = jint_cast(val->getf()); break;
duke@0 3103 case T_DOUBLE: con = jlong_cast(val->getd()); break;
duke@0 3104 default: continue; //skip (odd store type)
duke@0 3105 }
duke@0 3106
duke@0 3107 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
duke@0 3108 st->Opcode() == Op_StoreL) {
duke@0 3109 continue; // This StoreL is already optimal.
duke@0 3110 }
duke@0 3111
duke@0 3112 // Store down the constant.
duke@0 3113 store_constant(tiles, num_tiles, st_off, st_size, con);
duke@0 3114
duke@0 3115 intptr_t j = st_off >> LogBytesPerLong;
duke@0 3116
duke@0 3117 if (type == T_INT && st_size == BytesPerInt
duke@0 3118 && (st_off & BytesPerInt) == BytesPerInt) {
duke@0 3119 jlong lcon = tiles[j];
duke@0 3120 if (!Matcher::isSimpleConstant64(lcon) &&
duke@0 3121 st->Opcode() == Op_StoreI) {
duke@0 3122 // This StoreI is already optimal by itself.
duke@0 3123 jint* intcon = (jint*) &tiles[j];
duke@0 3124 intcon[1] = 0; // undo the store_constant()
duke@0 3125
duke@0 3126 // If the previous store is also optimal by itself, back up and
duke@0 3127 // undo the action of the previous loop iteration... if we can.
duke@0 3128 // But if we can't, just let the previous half take care of itself.
duke@0 3129 st = nodes[j];
duke@0 3130 st_off -= BytesPerInt;
duke@0 3131 con = intcon[0];
duke@0 3132 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
duke@0 3133 assert(st_off >= header_size, "still ignoring header");
duke@0 3134 assert(get_store_offset(st, phase) == st_off, "must be");
duke@0 3135 assert(in(i-1) == zmem, "must be");
duke@0 3136 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
duke@0 3137 assert(con == tcon->is_int()->get_con(), "must be");
duke@0 3138 // Undo the effects of the previous loop trip, which swallowed st:
duke@0 3139 intcon[0] = 0; // undo store_constant()
duke@0 3140 set_req(i-1, st); // undo set_req(i, zmem)
duke@0 3141 nodes[j] = NULL; // undo nodes[j] = st
duke@0 3142 --old_subword; // undo ++old_subword
duke@0 3143 }
duke@0 3144 continue; // This StoreI is already optimal.
duke@0 3145 }
duke@0 3146 }
duke@0 3147
duke@0 3148 // This store is not needed.
duke@0 3149 set_req(i, zmem);
duke@0 3150 nodes[j] = st; // record for the moment
duke@0 3151 if (st_size < BytesPerLong) // something has changed
duke@0 3152 ++old_subword; // includes int/float, but who's counting...
duke@0 3153 else ++old_long;
duke@0 3154 }
duke@0 3155
duke@0 3156 if ((old_subword + old_long) == 0)
duke@0 3157 return; // nothing more to do
duke@0 3158
duke@0 3159 //// Pass B: Convert any non-zero tiles into optimal constant stores.
duke@0 3160 // Be sure to insert them before overlapping non-constant stores.
duke@0 3161 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
duke@0 3162 for (int j = 0; j < num_tiles; j++) {
duke@0 3163 jlong con = tiles[j];
duke@0 3164 jlong init = inits[j];
duke@0 3165 if (con == 0) continue;
duke@0 3166 jint con0, con1; // split the constant, address-wise
duke@0 3167 jint init0, init1; // split the init map, address-wise
duke@0 3168 { union { jlong con; jint intcon[2]; } u;
duke@0 3169 u.con = con;
duke@0 3170 con0 = u.intcon[0];
duke@0 3171 con1 = u.intcon[1];
duke@0 3172 u.con = init;
duke@0 3173 init0 = u.intcon[0];
duke@0 3174 init1 = u.intcon[1];
duke@0 3175 }
duke@0 3176
duke@0 3177 Node* old = nodes[j];
duke@0 3178 assert(old != NULL, "need the prior store");
duke@0 3179 intptr_t offset = (j * BytesPerLong);
duke@0 3180
duke@0 3181 bool split = !Matcher::isSimpleConstant64(con);
duke@0 3182
duke@0 3183 if (offset < header_size) {
duke@0 3184 assert(offset + BytesPerInt >= header_size, "second int counts");
duke@0 3185 assert(*(jint*)&tiles[j] == 0, "junk in header");
duke@0 3186 split = true; // only the second word counts
duke@0 3187 // Example: int a[] = { 42 ... }
duke@0 3188 } else if (con0 == 0 && init0 == -1) {
duke@0 3189 split = true; // first word is covered by full inits
duke@0 3190 // Example: int a[] = { ... foo(), 42 ... }
duke@0 3191 } else if (con1 == 0 && init1 == -1) {
duke@0 3192 split = true; // second word is covered by full inits
duke@0 3193 // Example: int a[] = { ... 42, foo() ... }
duke@0 3194 }
duke@0 3195
duke@0 3196 // Here's a case where init0 is neither 0 nor -1:
duke@0 3197 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
duke@0 3198 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
duke@0 3199 // In this case the tile is not split; it is (jlong)42.
duke@0 3200 // The big tile is stored down, and then the foo() value is inserted.
duke@0 3201 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
duke@0 3202
duke@0 3203 Node* ctl = old->in(MemNode::Control);
duke@0 3204 Node* adr = make_raw_address(offset, phase);
duke@0 3205 const TypePtr* atp = TypeRawPtr::BOTTOM;
duke@0 3206
duke@0 3207 // One or two coalesced stores to plop down.
duke@0 3208 Node* st[2];
duke@0 3209 intptr_t off[2];
duke@0 3210 int nst = 0;
duke@0 3211 if (!split) {
duke@0 3212 ++new_long;
duke@0 3213 off[nst] = offset;
coleenp@113 3214 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3215 phase->longcon(con), T_LONG);
duke@0 3216 } else {
duke@0 3217 // Omit either if it is a zero.
duke@0 3218 if (con0 != 0) {
duke@0 3219 ++new_int;
duke@0 3220 off[nst] = offset;
coleenp@113 3221 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3222 phase->intcon(con0), T_INT);
duke@0 3223 }
duke@0 3224 if (con1 != 0) {
duke@0 3225 ++new_int;
duke@0 3226 offset += BytesPerInt;
duke@0 3227 adr = make_raw_address(offset, phase);
duke@0 3228 off[nst] = offset;
coleenp@113 3229 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3230 phase->intcon(con1), T_INT);
duke@0 3231 }
duke@0 3232 }
duke@0 3233
duke@0 3234 // Insert second store first, then the first before the second.
duke@0 3235 // Insert each one just before any overlapping non-constant stores.
duke@0 3236 while (nst > 0) {
duke@0 3237 Node* st1 = st[--nst];
duke@0 3238 C->copy_node_notes_to(st1, old);
duke@0 3239 st1 = phase->transform(st1);
duke@0 3240 offset = off[nst];
duke@0 3241 assert(offset >= header_size, "do not smash header");
duke@0 3242 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
duke@0 3243 guarantee(ins_idx != 0, "must re-insert constant store");
duke@0 3244 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
duke@0 3245 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
duke@0 3246 set_req(--ins_idx, st1);
duke@0 3247 else
duke@0 3248 ins_req(ins_idx, st1);
duke@0 3249 }
duke@0 3250 }
duke@0 3251
duke@0 3252 if (PrintCompilation && WizardMode)
duke@0 3253 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
duke@0 3254 old_subword, old_long, new_int, new_long);
duke@0 3255 if (C->log() != NULL)
duke@0 3256 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
duke@0 3257 old_subword, old_long, new_int, new_long);
duke@0 3258
duke@0 3259 // Clean up any remaining occurrences of zmem:
duke@0 3260 remove_extra_zeroes();
duke@0 3261 }
duke@0 3262
duke@0 3263 // Explore forward from in(start) to find the first fully initialized
duke@0 3264 // word, and return its offset. Skip groups of subword stores which
duke@0 3265 // together initialize full words. If in(start) is itself part of a
duke@0 3266 // fully initialized word, return the offset of in(start). If there
duke@0 3267 // are no following full-word stores, or if something is fishy, return
duke@0 3268 // a negative value.
duke@0 3269 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
duke@0 3270 int int_map = 0;
duke@0 3271 intptr_t int_map_off = 0;
duke@0 3272 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
duke@0 3273
duke@0 3274 for (uint i = start, limit = req(); i < limit; i++) {
duke@0 3275 Node* st = in(i);
duke@0 3276
duke@0 3277 intptr_t st_off = get_store_offset(st, phase);
duke@0 3278 if (st_off < 0) break; // return conservative answer
duke@0 3279
duke@0 3280 int st_size = st->as_Store()->memory_size();
duke@0 3281 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
duke@0 3282 return st_off; // we found a complete word init
duke@0 3283 }
duke@0 3284
duke@0 3285 // update the map:
duke@0 3286
duke@0 3287 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
duke@0 3288 if (this_int_off != int_map_off) {
duke@0 3289 // reset the map:
duke@0 3290 int_map = 0;
duke@0 3291 int_map_off = this_int_off;
duke@0 3292 }
duke@0 3293
duke@0 3294 int subword_off = st_off - this_int_off;
duke@0 3295 int_map |= right_n_bits(st_size) << subword_off;
duke@0 3296 if ((int_map & FULL_MAP) == FULL_MAP) {
duke@0 3297 return this_int_off; // we found a complete word init
duke@0 3298 }
duke@0 3299
duke@0 3300 // Did this store hit or cross the word boundary?
duke@0 3301 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
duke@0 3302 if (next_int_off == this_int_off + BytesPerInt) {
duke@0 3303 // We passed the current int, without fully initializing it.
duke@0 3304 int_map_off = next_int_off;
duke@0 3305 int_map >>= BytesPerInt;
duke@0 3306 } else if (next_int_off > this_int_off + BytesPerInt) {
duke@0 3307 // We passed the current and next int.
duke@0 3308 return this_int_off + BytesPerInt;
duke@0 3309 }
duke@0 3310 }
duke@0 3311
duke@0 3312 return -1;
duke@0 3313 }
duke@0 3314
duke@0 3315
duke@0 3316 // Called when the associated AllocateNode is expanded into CFG.
duke@0 3317 // At this point, we may perform additional optimizations.
duke@0 3318 // Linearize the stores by ascending offset, to make memory
duke@0 3319 // activity as coherent as possible.
duke@0 3320 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
duke@0 3321 intptr_t header_size,
duke@0 3322 Node* size_in_bytes,
duke@0 3323 PhaseGVN* phase) {
duke@0 3324 assert(!is_complete(), "not already complete");
duke@0 3325 assert(stores_are_sane(phase), "");
duke@0 3326 assert(allocation() != NULL, "must be present");
duke@0 3327
duke@0 3328 remove_extra_zeroes();
duke@0 3329
duke@0 3330 if (ReduceFieldZeroing || ReduceBulkZeroing)
duke@0 3331 // reduce instruction count for common initialization patterns
duke@0 3332 coalesce_subword_stores(header_size, size_in_bytes, phase);
duke@0 3333
duke@0 3334 Node* zmem = zero_memory(); // initially zero memory state
duke@0 3335 Node* inits = zmem; // accumulating a linearized chain of inits
duke@0 3336 #ifdef ASSERT
coleenp@113 3337 intptr_t first_offset = allocation()->minimum_header_size();
coleenp@113 3338 intptr_t last_init_off = first_offset; // previous init offset
coleenp@113 3339 intptr_t last_init_end = first_offset; // previous init offset+size
coleenp@113 3340 intptr_t last_tile_end = first_offset; // previous tile offset+size
duke@0 3341 #endif
duke@0 3342 intptr_t zeroes_done = header_size;
duke@0 3343
duke@0 3344 bool do_zeroing = true; // we might give up if inits are very sparse
duke@0 3345 int big_init_gaps = 0; // how many large gaps have we seen?
duke@0 3346
duke@0 3347 if (ZeroTLAB) do_zeroing = false;
duke@0 3348 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
duke@0 3349
duke@0 3350 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 3351 Node* st = in(i);
duke@0 3352 intptr_t st_off = get_store_offset(st, phase);
duke@0 3353 if (st_off < 0)
duke@0 3354 break; // unknown junk in the inits
duke@0 3355 if (st->in(MemNode::Memory) != zmem)
duke@0 3356 break; // complicated store chains somehow in list
duke@0 3357
duke@0 3358 int st_size = st->as_Store()->memory_size();
duke@0 3359 intptr_t next_init_off = st_off + st_size;
duke@0 3360
duke@0 3361 if (do_zeroing && zeroes_done < next_init_off) {
duke@0 3362 // See if this store needs a zero before it or under it.
duke@0 3363 intptr_t zeroes_needed = st_off;
duke@0 3364
duke@0 3365 if (st_size < BytesPerInt) {
duke@0 3366 // Look for subword stores which only partially initialize words.
duke@0 3367 // If we find some, we must lay down some word-level zeroes first,
duke@0 3368 // underneath the subword stores.
duke@0 3369 //
duke@0 3370 // Examples:
duke@0 3371 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
duke@0 3372 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
duke@0 3373 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
duke@0 3374 //
duke@0 3375 // Note: coalesce_subword_stores may have already done this,
duke@0 3376 // if it was prompted by constant non-zero subword initializers.
duke@0 3377 // But this case can still arise with non-constant stores.
duke@0 3378
duke@0 3379 intptr_t next_full_store = find_next_fullword_store(i, phase);
duke@0 3380
duke@0 3381 // In the examples above:
duke@0 3382 // in(i) p q r s x y z
duke@0 3383 // st_off 12 13 14 15 12 13 14
duke@0 3384 // st_size 1 1 1 1 1 1 1
duke@0 3385 // next_full_s. 12 16 16 16 16 16 16
duke@0 3386 // z's_done 12 16 16 16 12 16 12
duke@0 3387 // z's_needed 12 16 16 16 16 16 16
duke@0 3388 // zsize 0 0 0 0 4 0 4
duke@0 3389 if (next_full_store < 0) {
duke@0 3390 // Conservative tack: Zero to end of current word.
duke@0 3391 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
duke@0 3392 } else {
duke@0 3393 // Zero to beginning of next fully initialized word.
duke@0 3394 // Or, don't zero at all, if we are already in that word.
duke@0 3395 assert(next_full_store >= zeroes_needed, "must go forward");
duke@0 3396 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
duke@0 3397 zeroes_needed = next_full_store;
duke@0 3398 }
duke@0 3399 }
duke@0 3400
duke@0 3401 if (zeroes_needed > zeroes_done) {
duke@0 3402 intptr_t zsize = zeroes_needed - zeroes_done;
duke@0 3403 // Do some incremental zeroing on rawmem, in parallel with inits.
duke@0 3404 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3405 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3406 zeroes_done, zeroes_needed,
duke@0 3407 phase);
duke@0 3408 zeroes_done = zeroes_needed;
duke@0 3409 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
duke@0 3410 do_zeroing = false; // leave the hole, next time
duke@0 3411 }
duke@0 3412 }
duke@0 3413
duke@0 3414 // Collect the store and move on:
duke@0 3415 st->set_req(MemNode::Memory, inits);
duke@0 3416 inits = st; // put it on the linearized chain
duke@0 3417 set_req(i, zmem); // unhook from previous position
duke@0 3418
duke@0 3419 if (zeroes_done == st_off)
duke@0 3420 zeroes_done = next_init_off;
duke@0 3421
duke@0 3422 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
duke@0 3423
duke@0 3424 #ifdef ASSERT
duke@0 3425 // Various order invariants. Weaker than stores_are_sane because
duke@0 3426 // a large constant tile can be filled in by smaller non-constant stores.
duke@0 3427 assert(st_off >= last_init_off, "inits do not reverse");
duke@0 3428 last_init_off = st_off;
duke@0 3429 const Type* val = NULL;
duke@0 3430 if (st_size >= BytesPerInt &&
duke@0 3431 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
duke@0 3432 (int)val->basic_type() < (int)T_OBJECT) {
duke@0 3433 assert(st_off >= last_tile_end, "tiles do not overlap");
duke@0 3434 assert(st_off >= last_init_end, "tiles do not overwrite inits");
duke@0 3435 last_tile_end = MAX2(last_tile_end, next_init_off);
duke@0 3436 } else {
duke@0 3437 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
duke@0 3438 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
duke@0 3439 assert(st_off >= last_init_end, "inits do not overlap");
duke@0 3440 last_init_end = next_init_off; // it's a non-tile
duke@0 3441 }
duke@0 3442 #endif //ASSERT
duke@0 3443 }
duke@0 3444
duke@0 3445 remove_extra_zeroes(); // clear out all the zmems left over
duke@0 3446 add_req(inits);
duke@0 3447
duke@0 3448 if (!ZeroTLAB) {
duke@0 3449 // If anything remains to be zeroed, zero it all now.
duke@0 3450 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3451 // if it is the last unused 4 bytes of an instance, forget about it
duke@0 3452 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
duke@0 3453 if (zeroes_done + BytesPerLong >= size_limit) {
duke@0 3454 assert(allocation() != NULL, "");
duke@0 3455 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
duke@0 3456 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
duke@0 3457 if (zeroes_done == k->layout_helper())
duke@0 3458 zeroes_done = size_limit;
duke@0 3459 }
duke@0 3460 if (zeroes_done < size_limit) {
duke@0 3461 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3462 zeroes_done, size_in_bytes, phase);
duke@0 3463 }
duke@0 3464 }
duke@0 3465
duke@0 3466 set_complete(phase);
duke@0 3467 return rawmem;
duke@0 3468 }
duke@0 3469
duke@0 3470
duke@0 3471 #ifdef ASSERT
duke@0 3472 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
duke@0 3473 if (is_complete())
duke@0 3474 return true; // stores could be anything at this point
coleenp@113 3475 assert(allocation() != NULL, "must be present");
coleenp@113 3476 intptr_t last_off = allocation()->minimum_header_size();
duke@0 3477 for (uint i = InitializeNode::RawStores; i < req(); i++) {
duke@0 3478 Node* st = in(i);
duke@0 3479 intptr_t st_off = get_store_offset(st, phase);
duke@0 3480 if (st_off < 0) continue; // ignore dead garbage
duke@0 3481 if (last_off > st_off) {
duke@0 3482 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
duke@0 3483 this->dump(2);
duke@0 3484 assert(false, "ascending store offsets");
duke@0 3485 return false;
duke@0 3486 }
duke@0 3487 last_off = st_off + st->as_Store()->memory_size();
duke@0 3488 }
duke@0 3489 return true;
duke@0 3490 }
duke@0 3491 #endif //ASSERT
duke@0 3492
duke@0 3493
duke@0 3494
duke@0 3495
duke@0 3496 //============================MergeMemNode=====================================
duke@0 3497 //
duke@0 3498 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
duke@0 3499 // contributing store or call operations. Each contributor provides the memory
duke@0 3500 // state for a particular "alias type" (see Compile::alias_type). For example,
duke@0 3501 // if a MergeMem has an input X for alias category #6, then any memory reference
duke@0 3502 // to alias category #6 may use X as its memory state input, as an exact equivalent
duke@0 3503 // to using the MergeMem as a whole.
duke@0 3504 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
duke@0 3505 //
duke@0 3506 // (Here, the <N> notation gives the index of the relevant adr_type.)
duke@0 3507 //
duke@0 3508 // In one special case (and more cases in the future), alias categories overlap.
duke@0 3509 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
duke@0 3510 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
duke@0 3511 // it is exactly equivalent to that state W:
duke@0 3512 // MergeMem(<Bot>: W) <==> W
duke@0 3513 //
duke@0 3514 // Usually, the merge has more than one input. In that case, where inputs
duke@0 3515 // overlap (i.e., one is Bot), the narrower alias type determines the memory
duke@0 3516 // state for that type, and the wider alias type (Bot) fills in everywhere else:
duke@0 3517 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
duke@0 3518 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
duke@0 3519 //
duke@0 3520 // A merge can take a "wide" memory state as one of its narrow inputs.
duke@0 3521 // This simply means that the merge observes out only the relevant parts of
duke@0 3522 // the wide input. That is, wide memory states arriving at narrow merge inputs
duke@0 3523 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
duke@0 3524 //
duke@0 3525 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
duke@0 3526 // and that memory slices "leak through":
duke@0 3527 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
duke@0 3528 //
duke@0 3529 // But, in such a cascade, repeated memory slices can "block the leak":
duke@0 3530 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
duke@0 3531 //
duke@0 3532 // In the last example, Y is not part of the combined memory state of the
duke@0 3533 // outermost MergeMem. The system must, of course, prevent unschedulable
duke@0 3534 // memory states from arising, so you can be sure that the state Y is somehow
duke@0 3535 // a precursor to state Y'.
duke@0 3536 //
duke@0 3537 //
duke@0 3538 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
duke@0 3539 // of each MergeMemNode array are exactly the numerical alias indexes, including
duke@0 3540 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
duke@0 3541 // Compile::alias_type (and kin) produce and manage these indexes.
duke@0 3542 //
duke@0 3543 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
duke@0 3544 // (Note that this provides quick access to the top node inside MergeMem methods,
duke@0 3545 // without the need to reach out via TLS to Compile::current.)
duke@0 3546 //
duke@0 3547 // As a consequence of what was just described, a MergeMem that represents a full
duke@0 3548 // memory state has an edge in(