annotate src/share/vm/opto/memnode.cpp @ 163:885ed790ecf0

6695810: null oop passed to encode_heap_oop_not_null Summary: fix several problems in C2 related to Escape Analysis and Compressed Oops. Reviewed-by: never, jrose
author kvn
date Wed, 21 May 2008 10:45:07 -0700
parents 723be81c1212
children c436414a719e
rev   line source
duke@0 1 /*
duke@0 2 * Copyright 1997-2007 Sun Microsystems, Inc. All Rights Reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
duke@0 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
duke@0 20 * CA 95054 USA or visit www.sun.com if you need additional information or
duke@0 21 * have any questions.
duke@0 22 *
duke@0 23 */
duke@0 24
duke@0 25 // Portions of code courtesy of Clifford Click
duke@0 26
duke@0 27 // Optimization - Graph Style
duke@0 28
duke@0 29 #include "incls/_precompiled.incl"
duke@0 30 #include "incls/_memnode.cpp.incl"
duke@0 31
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@74 94 if (tinst == NULL || !tinst->is_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@74 97 Node *prev = NULL;
kvn@74 98 Node *result = mchain;
kvn@74 99 while (prev != result) {
kvn@74 100 prev = result;
kvn@74 101 // skip over a call which does not affect this memory slice
kvn@74 102 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
kvn@74 103 Node *proj_in = result->in(0);
kvn@74 104 if (proj_in->is_Call()) {
kvn@74 105 CallNode *call = proj_in->as_Call();
kvn@74 106 if (!call->may_modify(t_adr, phase)) {
kvn@74 107 result = call->in(TypeFunc::Memory);
kvn@74 108 }
kvn@74 109 } else if (proj_in->is_Initialize()) {
kvn@74 110 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
kvn@74 111 // Stop if this is the initialization for the object instance which
kvn@74 112 // which contains this memory slice, otherwise skip over it.
kvn@74 113 if (alloc != NULL && alloc->_idx != instance_id) {
kvn@74 114 result = proj_in->in(TypeFunc::Memory);
kvn@74 115 }
kvn@74 116 } else if (proj_in->is_MemBar()) {
kvn@74 117 result = proj_in->in(TypeFunc::Memory);
kvn@74 118 }
kvn@74 119 } else if (result->is_MergeMem()) {
kvn@74 120 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
kvn@74 121 }
kvn@74 122 }
kvn@74 123 return result;
kvn@74 124 }
kvn@74 125
kvn@74 126 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
kvn@74 127 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
kvn@74 128 bool is_instance = (t_oop != NULL) && t_oop->is_instance_field();
kvn@74 129 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@74 130 Node *result = mchain;
kvn@74 131 result = optimize_simple_memory_chain(result, t_adr, phase);
kvn@74 132 if (is_instance && igvn != NULL && result->is_Phi()) {
kvn@74 133 PhiNode *mphi = result->as_Phi();
kvn@74 134 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
kvn@74 135 const TypePtr *t = mphi->adr_type();
kvn@163 136 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
kvn@163 137 t->isa_oopptr() && !t->is_oopptr()->is_instance() &&
kvn@163 138 t->is_oopptr()->cast_to_instance(t_oop->instance_id()) == t_oop) {
kvn@74 139 // clone the Phi with our address type
kvn@74 140 result = mphi->split_out_instance(t_adr, igvn);
kvn@74 141 } else {
kvn@74 142 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
kvn@74 143 }
kvn@74 144 }
kvn@74 145 return result;
kvn@74 146 }
kvn@74 147
kvn@64 148 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
kvn@64 149 uint alias_idx = phase->C->get_alias_index(tp);
kvn@64 150 Node *mem = mmem;
kvn@64 151 #ifdef ASSERT
kvn@64 152 {
kvn@64 153 // Check that current type is consistent with the alias index used during graph construction
kvn@64 154 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
kvn@64 155 bool consistent = adr_check == NULL || adr_check->empty() ||
kvn@64 156 phase->C->must_alias(adr_check, alias_idx );
kvn@64 157 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
kvn@64 158 if( !consistent && adr_check != NULL && !adr_check->empty() &&
kvn@64 159 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
kvn@64 160 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
kvn@64 161 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
kvn@64 162 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
kvn@64 163 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
kvn@64 164 // don't assert if it is dead code.
kvn@64 165 consistent = true;
kvn@64 166 }
kvn@64 167 if( !consistent ) {
kvn@64 168 st->print("alias_idx==%d, adr_check==", alias_idx);
kvn@64 169 if( adr_check == NULL ) {
kvn@64 170 st->print("NULL");
kvn@64 171 } else {
kvn@64 172 adr_check->dump();
kvn@64 173 }
kvn@64 174 st->cr();
kvn@64 175 print_alias_types();
kvn@64 176 assert(consistent, "adr_check must match alias idx");
kvn@64 177 }
kvn@64 178 }
kvn@64 179 #endif
kvn@64 180 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
kvn@64 181 // means an array I have not precisely typed yet. Do not do any
kvn@64 182 // alias stuff with it any time soon.
kvn@64 183 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@64 184 if( tp->base() != Type::AnyPtr &&
kvn@64 185 !(tinst &&
kvn@64 186 tinst->klass()->is_java_lang_Object() &&
kvn@64 187 tinst->offset() == Type::OffsetBot) ) {
kvn@64 188 // compress paths and change unreachable cycles to TOP
kvn@64 189 // If not, we can update the input infinitely along a MergeMem cycle
kvn@64 190 // Equivalent code in PhiNode::Ideal
kvn@64 191 Node* m = phase->transform(mmem);
kvn@64 192 // If tranformed to a MergeMem, get the desired slice
kvn@64 193 // Otherwise the returned node represents memory for every slice
kvn@64 194 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
kvn@64 195 // Update input if it is progress over what we have now
kvn@64 196 }
kvn@64 197 return mem;
kvn@64 198 }
kvn@64 199
duke@0 200 //--------------------------Ideal_common---------------------------------------
duke@0 201 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
duke@0 202 // Unhook non-raw memories from complete (macro-expanded) initializations.
duke@0 203 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
duke@0 204 // If our control input is a dead region, kill all below the region
duke@0 205 Node *ctl = in(MemNode::Control);
duke@0 206 if (ctl && remove_dead_region(phase, can_reshape))
duke@0 207 return this;
duke@0 208
duke@0 209 // Ignore if memory is dead, or self-loop
duke@0 210 Node *mem = in(MemNode::Memory);
duke@0 211 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 212 assert( mem != this, "dead loop in MemNode::Ideal" );
duke@0 213
duke@0 214 Node *address = in(MemNode::Address);
duke@0 215 const Type *t_adr = phase->type( address );
duke@0 216 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
duke@0 217
duke@0 218 // Avoid independent memory operations
duke@0 219 Node* old_mem = mem;
duke@0 220
kvn@36 221 // The code which unhooks non-raw memories from complete (macro-expanded)
kvn@36 222 // initializations was removed. After macro-expansion all stores catched
kvn@36 223 // by Initialize node became raw stores and there is no information
kvn@36 224 // which memory slices they modify. So it is unsafe to move any memory
kvn@36 225 // operation above these stores. Also in most cases hooked non-raw memories
kvn@36 226 // were already unhooked by using information from detect_ptr_independence()
kvn@36 227 // and find_previous_store().
duke@0 228
duke@0 229 if (mem->is_MergeMem()) {
duke@0 230 MergeMemNode* mmem = mem->as_MergeMem();
duke@0 231 const TypePtr *tp = t_adr->is_ptr();
kvn@64 232
kvn@64 233 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
duke@0 234 }
duke@0 235
duke@0 236 if (mem != old_mem) {
duke@0 237 set_req(MemNode::Memory, mem);
duke@0 238 return this;
duke@0 239 }
duke@0 240
duke@0 241 // let the subclass continue analyzing...
duke@0 242 return NULL;
duke@0 243 }
duke@0 244
duke@0 245 // Helper function for proving some simple control dominations.
kvn@119 246 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
duke@0 247 // Already assumes that 'dom' is available at 'sub', and that 'sub'
duke@0 248 // is not a constant (dominated by the method's StartNode).
duke@0 249 // Used by MemNode::find_previous_store to prove that the
duke@0 250 // control input of a memory operation predates (dominates)
duke@0 251 // an allocation it wants to look past.
kvn@119 252 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
kvn@119 253 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
kvn@119 254 return false; // Conservative answer for dead code
kvn@119 255
kvn@119 256 // Check 'dom'.
kvn@119 257 dom = dom->find_exact_control(dom);
kvn@119 258 if (dom == NULL || dom->is_top())
kvn@119 259 return false; // Conservative answer for dead code
kvn@119 260
kvn@155 261 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
kvn@119 262 return true;
kvn@119 263
kvn@119 264 // 'dom' dominates 'sub' if its control edge and control edges
kvn@119 265 // of all its inputs dominate or equal to sub's control edge.
kvn@119 266
kvn@119 267 // Currently 'sub' is either Allocate, Initialize or Start nodes.
kvn@163 268 // Or Region for the check in LoadNode::Ideal();
kvn@163 269 // 'sub' should have sub->in(0) != NULL.
kvn@163 270 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
kvn@163 271 sub->is_Region(), "expecting only these nodes");
kvn@119 272
kvn@119 273 // Get control edge of 'sub'.
kvn@119 274 sub = sub->find_exact_control(sub->in(0));
kvn@119 275 if (sub == NULL || sub->is_top())
kvn@119 276 return false; // Conservative answer for dead code
kvn@119 277
kvn@119 278 assert(sub->is_CFG(), "expecting control");
kvn@119 279
kvn@119 280 if (sub == dom)
kvn@119 281 return true;
kvn@119 282
kvn@119 283 if (sub->is_Start() || sub->is_Root())
kvn@119 284 return false;
kvn@119 285
kvn@119 286 {
kvn@119 287 // Check all control edges of 'dom'.
kvn@119 288
kvn@119 289 ResourceMark rm;
kvn@119 290 Arena* arena = Thread::current()->resource_area();
kvn@119 291 Node_List nlist(arena);
kvn@119 292 Unique_Node_List dom_list(arena);
kvn@119 293
kvn@119 294 dom_list.push(dom);
kvn@119 295 bool only_dominating_controls = false;
kvn@119 296
kvn@119 297 for (uint next = 0; next < dom_list.size(); next++) {
kvn@119 298 Node* n = dom_list.at(next);
kvn@119 299 if (!n->is_CFG() && n->pinned()) {
kvn@119 300 // Check only own control edge for pinned non-control nodes.
kvn@119 301 n = n->find_exact_control(n->in(0));
kvn@119 302 if (n == NULL || n->is_top())
kvn@119 303 return false; // Conservative answer for dead code
kvn@119 304 assert(n->is_CFG(), "expecting control");
kvn@119 305 }
kvn@155 306 if (n->is_Con() || n->is_Start() || n->is_Root()) {
kvn@119 307 only_dominating_controls = true;
kvn@119 308 } else if (n->is_CFG()) {
kvn@119 309 if (n->dominates(sub, nlist))
kvn@119 310 only_dominating_controls = true;
kvn@119 311 else
kvn@119 312 return false;
kvn@119 313 } else {
kvn@119 314 // First, own control edge.
kvn@119 315 Node* m = n->find_exact_control(n->in(0));
kvn@155 316 if (m != NULL) {
kvn@155 317 if (m->is_top())
kvn@155 318 return false; // Conservative answer for dead code
kvn@155 319 dom_list.push(m);
kvn@155 320 }
kvn@119 321 // Now, the rest of edges.
kvn@119 322 uint cnt = n->req();
kvn@119 323 for (uint i = 1; i < cnt; i++) {
kvn@119 324 m = n->find_exact_control(n->in(i));
kvn@119 325 if (m == NULL || m->is_top())
kvn@119 326 continue;
kvn@119 327 dom_list.push(m);
duke@0 328 }
duke@0 329 }
duke@0 330 }
kvn@119 331 return only_dominating_controls;
duke@0 332 }
duke@0 333 }
duke@0 334
duke@0 335 //---------------------detect_ptr_independence---------------------------------
duke@0 336 // Used by MemNode::find_previous_store to prove that two base
duke@0 337 // pointers are never equal.
duke@0 338 // The pointers are accompanied by their associated allocations,
duke@0 339 // if any, which have been previously discovered by the caller.
duke@0 340 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
duke@0 341 Node* p2, AllocateNode* a2,
duke@0 342 PhaseTransform* phase) {
duke@0 343 // Attempt to prove that these two pointers cannot be aliased.
duke@0 344 // They may both manifestly be allocations, and they should differ.
duke@0 345 // Or, if they are not both allocations, they can be distinct constants.
duke@0 346 // Otherwise, one is an allocation and the other a pre-existing value.
duke@0 347 if (a1 == NULL && a2 == NULL) { // neither an allocation
duke@0 348 return (p1 != p2) && p1->is_Con() && p2->is_Con();
duke@0 349 } else if (a1 != NULL && a2 != NULL) { // both allocations
duke@0 350 return (a1 != a2);
duke@0 351 } else if (a1 != NULL) { // one allocation a1
duke@0 352 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
kvn@119 353 return all_controls_dominate(p2, a1);
duke@0 354 } else { //(a2 != NULL) // one allocation a2
kvn@119 355 return all_controls_dominate(p1, a2);
duke@0 356 }
duke@0 357 return false;
duke@0 358 }
duke@0 359
duke@0 360
duke@0 361 // The logic for reordering loads and stores uses four steps:
duke@0 362 // (a) Walk carefully past stores and initializations which we
duke@0 363 // can prove are independent of this load.
duke@0 364 // (b) Observe that the next memory state makes an exact match
duke@0 365 // with self (load or store), and locate the relevant store.
duke@0 366 // (c) Ensure that, if we were to wire self directly to the store,
duke@0 367 // the optimizer would fold it up somehow.
duke@0 368 // (d) Do the rewiring, and return, depending on some other part of
duke@0 369 // the optimizer to fold up the load.
duke@0 370 // This routine handles steps (a) and (b). Steps (c) and (d) are
duke@0 371 // specific to loads and stores, so they are handled by the callers.
duke@0 372 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
duke@0 373 //
duke@0 374 Node* MemNode::find_previous_store(PhaseTransform* phase) {
duke@0 375 Node* ctrl = in(MemNode::Control);
duke@0 376 Node* adr = in(MemNode::Address);
duke@0 377 intptr_t offset = 0;
duke@0 378 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 379 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
duke@0 380
duke@0 381 if (offset == Type::OffsetBot)
duke@0 382 return NULL; // cannot unalias unless there are precise offsets
duke@0 383
kvn@74 384 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
kvn@74 385
duke@0 386 intptr_t size_in_bytes = memory_size();
duke@0 387
duke@0 388 Node* mem = in(MemNode::Memory); // start searching here...
duke@0 389
duke@0 390 int cnt = 50; // Cycle limiter
duke@0 391 for (;;) { // While we can dance past unrelated stores...
duke@0 392 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
duke@0 393
duke@0 394 if (mem->is_Store()) {
duke@0 395 Node* st_adr = mem->in(MemNode::Address);
duke@0 396 intptr_t st_offset = 0;
duke@0 397 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
duke@0 398 if (st_base == NULL)
duke@0 399 break; // inscrutable pointer
duke@0 400 if (st_offset != offset && st_offset != Type::OffsetBot) {
duke@0 401 const int MAX_STORE = BytesPerLong;
duke@0 402 if (st_offset >= offset + size_in_bytes ||
duke@0 403 st_offset <= offset - MAX_STORE ||
duke@0 404 st_offset <= offset - mem->as_Store()->memory_size()) {
duke@0 405 // Success: The offsets are provably independent.
duke@0 406 // (You may ask, why not just test st_offset != offset and be done?
duke@0 407 // The answer is that stores of different sizes can co-exist
duke@0 408 // in the same sequence of RawMem effects. We sometimes initialize
duke@0 409 // a whole 'tile' of array elements with a single jint or jlong.)
duke@0 410 mem = mem->in(MemNode::Memory);
duke@0 411 continue; // (a) advance through independent store memory
duke@0 412 }
duke@0 413 }
duke@0 414 if (st_base != base &&
duke@0 415 detect_ptr_independence(base, alloc,
duke@0 416 st_base,
duke@0 417 AllocateNode::Ideal_allocation(st_base, phase),
duke@0 418 phase)) {
duke@0 419 // Success: The bases are provably independent.
duke@0 420 mem = mem->in(MemNode::Memory);
duke@0 421 continue; // (a) advance through independent store memory
duke@0 422 }
duke@0 423
duke@0 424 // (b) At this point, if the bases or offsets do not agree, we lose,
duke@0 425 // since we have not managed to prove 'this' and 'mem' independent.
duke@0 426 if (st_base == base && st_offset == offset) {
duke@0 427 return mem; // let caller handle steps (c), (d)
duke@0 428 }
duke@0 429
duke@0 430 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 431 InitializeNode* st_init = mem->in(0)->as_Initialize();
duke@0 432 AllocateNode* st_alloc = st_init->allocation();
duke@0 433 if (st_alloc == NULL)
duke@0 434 break; // something degenerated
duke@0 435 bool known_identical = false;
duke@0 436 bool known_independent = false;
duke@0 437 if (alloc == st_alloc)
duke@0 438 known_identical = true;
duke@0 439 else if (alloc != NULL)
duke@0 440 known_independent = true;
kvn@119 441 else if (all_controls_dominate(this, st_alloc))
duke@0 442 known_independent = true;
duke@0 443
duke@0 444 if (known_independent) {
duke@0 445 // The bases are provably independent: Either they are
duke@0 446 // manifestly distinct allocations, or else the control
duke@0 447 // of this load dominates the store's allocation.
duke@0 448 int alias_idx = phase->C->get_alias_index(adr_type());
duke@0 449 if (alias_idx == Compile::AliasIdxRaw) {
duke@0 450 mem = st_alloc->in(TypeFunc::Memory);
duke@0 451 } else {
duke@0 452 mem = st_init->memory(alias_idx);
duke@0 453 }
duke@0 454 continue; // (a) advance through independent store memory
duke@0 455 }
duke@0 456
duke@0 457 // (b) at this point, if we are not looking at a store initializing
duke@0 458 // the same allocation we are loading from, we lose.
duke@0 459 if (known_identical) {
duke@0 460 // From caller, can_see_stored_value will consult find_captured_store.
duke@0 461 return mem; // let caller handle steps (c), (d)
duke@0 462 }
duke@0 463
kvn@74 464 } else if (addr_t != NULL && addr_t->is_instance_field()) {
kvn@74 465 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
kvn@74 466 if (mem->is_Proj() && mem->in(0)->is_Call()) {
kvn@74 467 CallNode *call = mem->in(0)->as_Call();
kvn@74 468 if (!call->may_modify(addr_t, phase)) {
kvn@74 469 mem = call->in(TypeFunc::Memory);
kvn@74 470 continue; // (a) advance through independent call memory
kvn@74 471 }
kvn@74 472 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
kvn@74 473 mem = mem->in(0)->in(TypeFunc::Memory);
kvn@74 474 continue; // (a) advance through independent MemBar memory
kvn@74 475 } else if (mem->is_MergeMem()) {
kvn@74 476 int alias_idx = phase->C->get_alias_index(adr_type());
kvn@74 477 mem = mem->as_MergeMem()->memory_at(alias_idx);
kvn@74 478 continue; // (a) advance through independent MergeMem memory
kvn@74 479 }
duke@0 480 }
duke@0 481
duke@0 482 // Unless there is an explicit 'continue', we must bail out here,
duke@0 483 // because 'mem' is an inscrutable memory state (e.g., a call).
duke@0 484 break;
duke@0 485 }
duke@0 486
duke@0 487 return NULL; // bail out
duke@0 488 }
duke@0 489
duke@0 490 //----------------------calculate_adr_type-------------------------------------
duke@0 491 // Helper function. Notices when the given type of address hits top or bottom.
duke@0 492 // Also, asserts a cross-check of the type against the expected address type.
duke@0 493 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
duke@0 494 if (t == Type::TOP) return NULL; // does not touch memory any more?
duke@0 495 #ifdef PRODUCT
duke@0 496 cross_check = NULL;
duke@0 497 #else
duke@0 498 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
duke@0 499 #endif
duke@0 500 const TypePtr* tp = t->isa_ptr();
duke@0 501 if (tp == NULL) {
duke@0 502 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
duke@0 503 return TypePtr::BOTTOM; // touches lots of memory
duke@0 504 } else {
duke@0 505 #ifdef ASSERT
duke@0 506 // %%%% [phh] We don't check the alias index if cross_check is
duke@0 507 // TypeRawPtr::BOTTOM. Needs to be investigated.
duke@0 508 if (cross_check != NULL &&
duke@0 509 cross_check != TypePtr::BOTTOM &&
duke@0 510 cross_check != TypeRawPtr::BOTTOM) {
duke@0 511 // Recheck the alias index, to see if it has changed (due to a bug).
duke@0 512 Compile* C = Compile::current();
duke@0 513 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
duke@0 514 "must stay in the original alias category");
duke@0 515 // The type of the address must be contained in the adr_type,
duke@0 516 // disregarding "null"-ness.
duke@0 517 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
duke@0 518 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
duke@0 519 assert(cross_check->meet(tp_notnull) == cross_check,
duke@0 520 "real address must not escape from expected memory type");
duke@0 521 }
duke@0 522 #endif
duke@0 523 return tp;
duke@0 524 }
duke@0 525 }
duke@0 526
duke@0 527 //------------------------adr_phi_is_loop_invariant----------------------------
duke@0 528 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
duke@0 529 // loop is loop invariant. Make a quick traversal of Phi and associated
duke@0 530 // CastPP nodes, looking to see if they are a closed group within the loop.
duke@0 531 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
duke@0 532 // The idea is that the phi-nest must boil down to only CastPP nodes
duke@0 533 // with the same data. This implies that any path into the loop already
duke@0 534 // includes such a CastPP, and so the original cast, whatever its input,
duke@0 535 // must be covered by an equivalent cast, with an earlier control input.
duke@0 536 ResourceMark rm;
duke@0 537
duke@0 538 // The loop entry input of the phi should be the unique dominating
duke@0 539 // node for every Phi/CastPP in the loop.
duke@0 540 Unique_Node_List closure;
duke@0 541 closure.push(adr_phi->in(LoopNode::EntryControl));
duke@0 542
duke@0 543 // Add the phi node and the cast to the worklist.
duke@0 544 Unique_Node_List worklist;
duke@0 545 worklist.push(adr_phi);
duke@0 546 if( cast != NULL ){
duke@0 547 if( !cast->is_ConstraintCast() ) return false;
duke@0 548 worklist.push(cast);
duke@0 549 }
duke@0 550
duke@0 551 // Begin recursive walk of phi nodes.
duke@0 552 while( worklist.size() ){
duke@0 553 // Take a node off the worklist
duke@0 554 Node *n = worklist.pop();
duke@0 555 if( !closure.member(n) ){
duke@0 556 // Add it to the closure.
duke@0 557 closure.push(n);
duke@0 558 // Make a sanity check to ensure we don't waste too much time here.
duke@0 559 if( closure.size() > 20) return false;
duke@0 560 // This node is OK if:
duke@0 561 // - it is a cast of an identical value
duke@0 562 // - or it is a phi node (then we add its inputs to the worklist)
duke@0 563 // Otherwise, the node is not OK, and we presume the cast is not invariant
duke@0 564 if( n->is_ConstraintCast() ){
duke@0 565 worklist.push(n->in(1));
duke@0 566 } else if( n->is_Phi() ) {
duke@0 567 for( uint i = 1; i < n->req(); i++ ) {
duke@0 568 worklist.push(n->in(i));
duke@0 569 }
duke@0 570 } else {
duke@0 571 return false;
duke@0 572 }
duke@0 573 }
duke@0 574 }
duke@0 575
duke@0 576 // Quit when the worklist is empty, and we've found no offending nodes.
duke@0 577 return true;
duke@0 578 }
duke@0 579
duke@0 580 //------------------------------Ideal_DU_postCCP-------------------------------
duke@0 581 // Find any cast-away of null-ness and keep its control. Null cast-aways are
duke@0 582 // going away in this pass and we need to make this memory op depend on the
duke@0 583 // gating null check.
kvn@163 584 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
kvn@163 585 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
kvn@163 586 }
duke@0 587
duke@0 588 // I tried to leave the CastPP's in. This makes the graph more accurate in
duke@0 589 // some sense; we get to keep around the knowledge that an oop is not-null
duke@0 590 // after some test. Alas, the CastPP's interfere with GVN (some values are
duke@0 591 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
duke@0 592 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
duke@0 593 // some of the more trivial cases in the optimizer. Removing more useless
duke@0 594 // Phi's started allowing Loads to illegally float above null checks. I gave
duke@0 595 // up on this approach. CNC 10/20/2000
kvn@163 596 // This static method may be called not from MemNode (EncodePNode calls it).
kvn@163 597 // Only the control edge of the node 'n' might be updated.
kvn@163 598 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
duke@0 599 Node *skipped_cast = NULL;
duke@0 600 // Need a null check? Regular static accesses do not because they are
duke@0 601 // from constant addresses. Array ops are gated by the range check (which
duke@0 602 // always includes a NULL check). Just check field ops.
kvn@163 603 if( n->in(MemNode::Control) == NULL ) {
duke@0 604 // Scan upwards for the highest location we can place this memory op.
duke@0 605 while( true ) {
duke@0 606 switch( adr->Opcode() ) {
duke@0 607
duke@0 608 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
duke@0 609 adr = adr->in(AddPNode::Base);
duke@0 610 continue;
duke@0 611
coleenp@113 612 case Op_DecodeN: // No change to NULL-ness, so peek thru
coleenp@113 613 adr = adr->in(1);
coleenp@113 614 continue;
coleenp@113 615
duke@0 616 case Op_CastPP:
duke@0 617 // If the CastPP is useless, just peek on through it.
duke@0 618 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
duke@0 619 // Remember the cast that we've peeked though. If we peek
duke@0 620 // through more than one, then we end up remembering the highest
duke@0 621 // one, that is, if in a loop, the one closest to the top.
duke@0 622 skipped_cast = adr;
duke@0 623 adr = adr->in(1);
duke@0 624 continue;
duke@0 625 }
duke@0 626 // CastPP is going away in this pass! We need this memory op to be
duke@0 627 // control-dependent on the test that is guarding the CastPP.
kvn@163 628 ccp->hash_delete(n);
kvn@163 629 n->set_req(MemNode::Control, adr->in(0));
kvn@163 630 ccp->hash_insert(n);
kvn@163 631 return n;
duke@0 632
duke@0 633 case Op_Phi:
duke@0 634 // Attempt to float above a Phi to some dominating point.
duke@0 635 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
duke@0 636 // If we've already peeked through a Cast (which could have set the
duke@0 637 // control), we can't float above a Phi, because the skipped Cast
duke@0 638 // may not be loop invariant.
duke@0 639 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
duke@0 640 adr = adr->in(1);
duke@0 641 continue;
duke@0 642 }
duke@0 643 }
duke@0 644
duke@0 645 // Intentional fallthrough!
duke@0 646
duke@0 647 // No obvious dominating point. The mem op is pinned below the Phi
duke@0 648 // by the Phi itself. If the Phi goes away (no true value is merged)
duke@0 649 // then the mem op can float, but not indefinitely. It must be pinned
duke@0 650 // behind the controls leading to the Phi.
duke@0 651 case Op_CheckCastPP:
duke@0 652 // These usually stick around to change address type, however a
duke@0 653 // useless one can be elided and we still need to pick up a control edge
duke@0 654 if (adr->in(0) == NULL) {
duke@0 655 // This CheckCastPP node has NO control and is likely useless. But we
duke@0 656 // need check further up the ancestor chain for a control input to keep
duke@0 657 // the node in place. 4959717.
duke@0 658 skipped_cast = adr;
duke@0 659 adr = adr->in(1);
duke@0 660 continue;
duke@0 661 }
kvn@163 662 ccp->hash_delete(n);
kvn@163 663 n->set_req(MemNode::Control, adr->in(0));
kvn@163 664 ccp->hash_insert(n);
kvn@163 665 return n;
duke@0 666
duke@0 667 // List of "safe" opcodes; those that implicitly block the memory
duke@0 668 // op below any null check.
duke@0 669 case Op_CastX2P: // no null checks on native pointers
duke@0 670 case Op_Parm: // 'this' pointer is not null
duke@0 671 case Op_LoadP: // Loading from within a klass
coleenp@113 672 case Op_LoadN: // Loading from within a klass
duke@0 673 case Op_LoadKlass: // Loading from within a klass
duke@0 674 case Op_ConP: // Loading from a klass
kvn@163 675 case Op_ConN: // Loading from a klass
duke@0 676 case Op_CreateEx: // Sucking up the guts of an exception oop
duke@0 677 case Op_Con: // Reading from TLS
duke@0 678 case Op_CMoveP: // CMoveP is pinned
duke@0 679 break; // No progress
duke@0 680
duke@0 681 case Op_Proj: // Direct call to an allocation routine
duke@0 682 case Op_SCMemProj: // Memory state from store conditional ops
duke@0 683 #ifdef ASSERT
duke@0 684 {
duke@0 685 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
duke@0 686 const Node* call = adr->in(0);
kvn@163 687 if (call->is_CallJava()) {
kvn@163 688 const CallJavaNode* call_java = call->as_CallJava();
kvn@64 689 const TypeTuple *r = call_java->tf()->range();
kvn@64 690 assert(r->cnt() > TypeFunc::Parms, "must return value");
kvn@64 691 const Type* ret_type = r->field_at(TypeFunc::Parms);
kvn@64 692 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
duke@0 693 // We further presume that this is one of
duke@0 694 // new_instance_Java, new_array_Java, or
duke@0 695 // the like, but do not assert for this.
duke@0 696 } else if (call->is_Allocate()) {
duke@0 697 // similar case to new_instance_Java, etc.
duke@0 698 } else if (!call->is_CallLeaf()) {
duke@0 699 // Projections from fetch_oop (OSR) are allowed as well.
duke@0 700 ShouldNotReachHere();
duke@0 701 }
duke@0 702 }
duke@0 703 #endif
duke@0 704 break;
duke@0 705 default:
duke@0 706 ShouldNotReachHere();
duke@0 707 }
duke@0 708 break;
duke@0 709 }
duke@0 710 }
duke@0 711
duke@0 712 return NULL; // No progress
duke@0 713 }
duke@0 714
duke@0 715
duke@0 716 //=============================================================================
duke@0 717 uint LoadNode::size_of() const { return sizeof(*this); }
duke@0 718 uint LoadNode::cmp( const Node &n ) const
duke@0 719 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
duke@0 720 const Type *LoadNode::bottom_type() const { return _type; }
duke@0 721 uint LoadNode::ideal_reg() const {
duke@0 722 return Matcher::base2reg[_type->base()];
duke@0 723 }
duke@0 724
duke@0 725 #ifndef PRODUCT
duke@0 726 void LoadNode::dump_spec(outputStream *st) const {
duke@0 727 MemNode::dump_spec(st);
duke@0 728 if( !Verbose && !WizardMode ) {
duke@0 729 // standard dump does this in Verbose and WizardMode
duke@0 730 st->print(" #"); _type->dump_on(st);
duke@0 731 }
duke@0 732 }
duke@0 733 #endif
duke@0 734
duke@0 735
duke@0 736 //----------------------------LoadNode::make-----------------------------------
duke@0 737 // Polymorphic factory method:
coleenp@113 738 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
coleenp@113 739 Compile* C = gvn.C;
coleenp@113 740
duke@0 741 // sanity check the alias category against the created node type
duke@0 742 assert(!(adr_type->isa_oopptr() &&
duke@0 743 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
duke@0 744 "use LoadKlassNode instead");
duke@0 745 assert(!(adr_type->isa_aryptr() &&
duke@0 746 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
duke@0 747 "use LoadRangeNode instead");
duke@0 748 switch (bt) {
duke@0 749 case T_BOOLEAN:
duke@0 750 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 751 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 752 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 753 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
duke@0 754 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
duke@0 755 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
duke@0 756 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
duke@0 757 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
coleenp@113 758 case T_OBJECT:
coleenp@113 759 #ifdef _LP64
kvn@163 760 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
coleenp@113 761 const TypeNarrowOop* narrowtype;
coleenp@113 762 if (rt->isa_narrowoop()) {
coleenp@113 763 narrowtype = rt->is_narrowoop();
coleenp@113 764 } else {
coleenp@113 765 narrowtype = rt->is_oopptr()->make_narrowoop();
coleenp@113 766 }
coleenp@113 767 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, narrowtype));
coleenp@113 768
kvn@124 769 return DecodeNNode::decode(&gvn, load);
coleenp@113 770 } else
coleenp@113 771 #endif
kvn@163 772 {
kvn@163 773 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
kvn@163 774 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
kvn@163 775 }
duke@0 776 }
duke@0 777 ShouldNotReachHere();
duke@0 778 return (LoadNode*)NULL;
duke@0 779 }
duke@0 780
duke@0 781 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
duke@0 782 bool require_atomic = true;
duke@0 783 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
duke@0 784 }
duke@0 785
duke@0 786
duke@0 787
duke@0 788
duke@0 789 //------------------------------hash-------------------------------------------
duke@0 790 uint LoadNode::hash() const {
duke@0 791 // unroll addition of interesting fields
duke@0 792 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
duke@0 793 }
duke@0 794
duke@0 795 //---------------------------can_see_stored_value------------------------------
duke@0 796 // This routine exists to make sure this set of tests is done the same
duke@0 797 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
duke@0 798 // will change the graph shape in a way which makes memory alive twice at the
duke@0 799 // same time (uses the Oracle model of aliasing), then some
duke@0 800 // LoadXNode::Identity will fold things back to the equivalence-class model
duke@0 801 // of aliasing.
duke@0 802 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
duke@0 803 Node* ld_adr = in(MemNode::Address);
duke@0 804
never@17 805 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
never@17 806 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
never@17 807 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
never@17 808 atp->field() != NULL && !atp->field()->is_volatile()) {
never@17 809 uint alias_idx = atp->index();
never@17 810 bool final = atp->field()->is_final();
never@17 811 Node* result = NULL;
never@17 812 Node* current = st;
never@17 813 // Skip through chains of MemBarNodes checking the MergeMems for
never@17 814 // new states for the slice of this load. Stop once any other
never@17 815 // kind of node is encountered. Loads from final memory can skip
never@17 816 // through any kind of MemBar but normal loads shouldn't skip
never@17 817 // through MemBarAcquire since the could allow them to move out of
never@17 818 // a synchronized region.
never@17 819 while (current->is_Proj()) {
never@17 820 int opc = current->in(0)->Opcode();
never@17 821 if ((final && opc == Op_MemBarAcquire) ||
never@17 822 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
never@17 823 Node* mem = current->in(0)->in(TypeFunc::Memory);
never@17 824 if (mem->is_MergeMem()) {
never@17 825 MergeMemNode* merge = mem->as_MergeMem();
never@17 826 Node* new_st = merge->memory_at(alias_idx);
never@17 827 if (new_st == merge->base_memory()) {
never@17 828 // Keep searching
never@17 829 current = merge->base_memory();
never@17 830 continue;
never@17 831 }
never@17 832 // Save the new memory state for the slice and fall through
never@17 833 // to exit.
never@17 834 result = new_st;
never@17 835 }
never@17 836 }
never@17 837 break;
never@17 838 }
never@17 839 if (result != NULL) {
never@17 840 st = result;
never@17 841 }
never@17 842 }
never@17 843
never@17 844
duke@0 845 // Loop around twice in the case Load -> Initialize -> Store.
duke@0 846 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
duke@0 847 for (int trip = 0; trip <= 1; trip++) {
duke@0 848
duke@0 849 if (st->is_Store()) {
duke@0 850 Node* st_adr = st->in(MemNode::Address);
duke@0 851 if (!phase->eqv(st_adr, ld_adr)) {
duke@0 852 // Try harder before giving up... Match raw and non-raw pointers.
duke@0 853 intptr_t st_off = 0;
duke@0 854 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
duke@0 855 if (alloc == NULL) return NULL;
duke@0 856 intptr_t ld_off = 0;
duke@0 857 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
duke@0 858 if (alloc != allo2) return NULL;
duke@0 859 if (ld_off != st_off) return NULL;
duke@0 860 // At this point we have proven something like this setup:
duke@0 861 // A = Allocate(...)
duke@0 862 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
duke@0 863 // S = StoreQ(, AddP(, A.Parm , #Off), V)
duke@0 864 // (Actually, we haven't yet proven the Q's are the same.)
duke@0 865 // In other words, we are loading from a casted version of
duke@0 866 // the same pointer-and-offset that we stored to.
duke@0 867 // Thus, we are able to replace L by V.
duke@0 868 }
duke@0 869 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
duke@0 870 if (store_Opcode() != st->Opcode())
duke@0 871 return NULL;
duke@0 872 return st->in(MemNode::ValueIn);
duke@0 873 }
duke@0 874
duke@0 875 intptr_t offset = 0; // scratch
duke@0 876
duke@0 877 // A load from a freshly-created object always returns zero.
duke@0 878 // (This can happen after LoadNode::Ideal resets the load's memory input
duke@0 879 // to find_captured_store, which returned InitializeNode::zero_memory.)
duke@0 880 if (st->is_Proj() && st->in(0)->is_Allocate() &&
duke@0 881 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
duke@0 882 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
duke@0 883 // return a zero value for the load's basic type
duke@0 884 // (This is one of the few places where a generic PhaseTransform
duke@0 885 // can create new nodes. Think of it as lazily manifesting
duke@0 886 // virtually pre-existing constants.)
duke@0 887 return phase->zerocon(memory_type());
duke@0 888 }
duke@0 889
duke@0 890 // A load from an initialization barrier can match a captured store.
duke@0 891 if (st->is_Proj() && st->in(0)->is_Initialize()) {
duke@0 892 InitializeNode* init = st->in(0)->as_Initialize();
duke@0 893 AllocateNode* alloc = init->allocation();
duke@0 894 if (alloc != NULL &&
duke@0 895 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
duke@0 896 // examine a captured store value
duke@0 897 st = init->find_captured_store(offset, memory_size(), phase);
duke@0 898 if (st != NULL)
duke@0 899 continue; // take one more trip around
duke@0 900 }
duke@0 901 }
duke@0 902
duke@0 903 break;
duke@0 904 }
duke@0 905
duke@0 906 return NULL;
duke@0 907 }
duke@0 908
kvn@64 909 //----------------------is_instance_field_load_with_local_phi------------------
kvn@64 910 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
kvn@64 911 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
kvn@64 912 in(MemNode::Address)->is_AddP() ) {
kvn@64 913 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
kvn@64 914 // Only instances.
kvn@64 915 if( t_oop != NULL && t_oop->is_instance_field() &&
kvn@64 916 t_oop->offset() != Type::OffsetBot &&
kvn@64 917 t_oop->offset() != Type::OffsetTop) {
kvn@64 918 return true;
kvn@64 919 }
kvn@64 920 }
kvn@64 921 return false;
kvn@64 922 }
kvn@64 923
duke@0 924 //------------------------------Identity---------------------------------------
duke@0 925 // Loads are identity if previous store is to same address
duke@0 926 Node *LoadNode::Identity( PhaseTransform *phase ) {
duke@0 927 // If the previous store-maker is the right kind of Store, and the store is
duke@0 928 // to the same address, then we are equal to the value stored.
duke@0 929 Node* mem = in(MemNode::Memory);
duke@0 930 Node* value = can_see_stored_value(mem, phase);
duke@0 931 if( value ) {
duke@0 932 // byte, short & char stores truncate naturally.
duke@0 933 // A load has to load the truncated value which requires
duke@0 934 // some sort of masking operation and that requires an
duke@0 935 // Ideal call instead of an Identity call.
duke@0 936 if (memory_size() < BytesPerInt) {
duke@0 937 // If the input to the store does not fit with the load's result type,
duke@0 938 // it must be truncated via an Ideal call.
duke@0 939 if (!phase->type(value)->higher_equal(phase->type(this)))
duke@0 940 return this;
duke@0 941 }
duke@0 942 // (This works even when value is a Con, but LoadNode::Value
duke@0 943 // usually runs first, producing the singleton type of the Con.)
duke@0 944 return value;
duke@0 945 }
kvn@64 946
kvn@64 947 // Search for an existing data phi which was generated before for the same
kvn@64 948 // instance's field to avoid infinite genertion of phis in a loop.
kvn@64 949 Node *region = mem->in(0);
kvn@64 950 if (is_instance_field_load_with_local_phi(region)) {
kvn@64 951 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
kvn@64 952 int this_index = phase->C->get_alias_index(addr_t);
kvn@64 953 int this_offset = addr_t->offset();
kvn@64 954 int this_id = addr_t->is_oopptr()->instance_id();
kvn@64 955 const Type* this_type = bottom_type();
kvn@64 956 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
kvn@64 957 Node* phi = region->fast_out(i);
kvn@64 958 if (phi->is_Phi() && phi != mem &&
kvn@64 959 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
kvn@64 960 return phi;
kvn@64 961 }
kvn@64 962 }
kvn@64 963 }
kvn@64 964
duke@0 965 return this;
duke@0 966 }
duke@0 967
never@17 968
never@17 969 // Returns true if the AliasType refers to the field that holds the
never@17 970 // cached box array. Currently only handles the IntegerCache case.
never@17 971 static bool is_autobox_cache(Compile::AliasType* atp) {
never@17 972 if (atp != NULL && atp->field() != NULL) {
never@17 973 ciField* field = atp->field();
never@17 974 ciSymbol* klass = field->holder()->name();
never@17 975 if (field->name() == ciSymbol::cache_field_name() &&
never@17 976 field->holder()->uses_default_loader() &&
never@17 977 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 978 return true;
never@17 979 }
never@17 980 }
never@17 981 return false;
never@17 982 }
never@17 983
never@17 984 // Fetch the base value in the autobox array
never@17 985 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
never@17 986 if (atp != NULL && atp->field() != NULL) {
never@17 987 ciField* field = atp->field();
never@17 988 ciSymbol* klass = field->holder()->name();
never@17 989 if (field->name() == ciSymbol::cache_field_name() &&
never@17 990 field->holder()->uses_default_loader() &&
never@17 991 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
never@17 992 assert(field->is_constant(), "what?");
never@17 993 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
never@17 994 // Fetch the box object at the base of the array and get its value
never@17 995 ciInstance* box = array->obj_at(0)->as_instance();
never@17 996 ciInstanceKlass* ik = box->klass()->as_instance_klass();
never@17 997 if (ik->nof_nonstatic_fields() == 1) {
never@17 998 // This should be true nonstatic_field_at requires calling
never@17 999 // nof_nonstatic_fields so check it anyway
never@17 1000 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
never@17 1001 cache_offset = c.as_int();
never@17 1002 }
never@17 1003 return true;
never@17 1004 }
never@17 1005 }
never@17 1006 return false;
never@17 1007 }
never@17 1008
never@17 1009 // Returns true if the AliasType refers to the value field of an
never@17 1010 // autobox object. Currently only handles Integer.
never@17 1011 static bool is_autobox_object(Compile::AliasType* atp) {
never@17 1012 if (atp != NULL && atp->field() != NULL) {
never@17 1013 ciField* field = atp->field();
never@17 1014 ciSymbol* klass = field->holder()->name();
never@17 1015 if (field->name() == ciSymbol::value_name() &&
never@17 1016 field->holder()->uses_default_loader() &&
never@17 1017 klass == ciSymbol::java_lang_Integer()) {
never@17 1018 return true;
never@17 1019 }
never@17 1020 }
never@17 1021 return false;
never@17 1022 }
never@17 1023
never@17 1024
never@17 1025 // We're loading from an object which has autobox behaviour.
never@17 1026 // If this object is result of a valueOf call we'll have a phi
never@17 1027 // merging a newly allocated object and a load from the cache.
never@17 1028 // We want to replace this load with the original incoming
never@17 1029 // argument to the valueOf call.
never@17 1030 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
never@17 1031 Node* base = in(Address)->in(AddPNode::Base);
never@17 1032 if (base->is_Phi() && base->req() == 3) {
never@17 1033 AllocateNode* allocation = NULL;
never@17 1034 int allocation_index = -1;
never@17 1035 int load_index = -1;
never@17 1036 for (uint i = 1; i < base->req(); i++) {
never@17 1037 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
never@17 1038 if (allocation != NULL) {
never@17 1039 allocation_index = i;
never@17 1040 load_index = 3 - allocation_index;
never@17 1041 break;
never@17 1042 }
never@17 1043 }
never@17 1044 LoadNode* load = NULL;
never@17 1045 if (allocation != NULL && base->in(load_index)->is_Load()) {
never@17 1046 load = base->in(load_index)->as_Load();
never@17 1047 }
never@17 1048 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
never@17 1049 // Push the loads from the phi that comes from valueOf up
never@17 1050 // through it to allow elimination of the loads and the recovery
never@17 1051 // of the original value.
never@17 1052 Node* mem_phi = in(Memory);
never@17 1053 Node* offset = in(Address)->in(AddPNode::Offset);
never@17 1054
never@17 1055 Node* in1 = clone();
never@17 1056 Node* in1_addr = in1->in(Address)->clone();
never@17 1057 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
never@17 1058 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
never@17 1059 in1_addr->set_req(AddPNode::Offset, offset);
never@17 1060 in1->set_req(0, base->in(allocation_index));
never@17 1061 in1->set_req(Address, in1_addr);
never@17 1062 in1->set_req(Memory, mem_phi->in(allocation_index));
never@17 1063
never@17 1064 Node* in2 = clone();
never@17 1065 Node* in2_addr = in2->in(Address)->clone();
never@17 1066 in2_addr->set_req(AddPNode::Base, base->in(load_index));
never@17 1067 in2_addr->set_req(AddPNode::Address, base->in(load_index));
never@17 1068 in2_addr->set_req(AddPNode::Offset, offset);
never@17 1069 in2->set_req(0, base->in(load_index));
never@17 1070 in2->set_req(Address, in2_addr);
never@17 1071 in2->set_req(Memory, mem_phi->in(load_index));
never@17 1072
never@17 1073 in1_addr = phase->transform(in1_addr);
never@17 1074 in1 = phase->transform(in1);
never@17 1075 in2_addr = phase->transform(in2_addr);
never@17 1076 in2 = phase->transform(in2);
never@17 1077
never@17 1078 PhiNode* result = PhiNode::make_blank(base->in(0), this);
never@17 1079 result->set_req(allocation_index, in1);
never@17 1080 result->set_req(load_index, in2);
never@17 1081 return result;
never@17 1082 }
never@17 1083 } else if (base->is_Load()) {
never@17 1084 // Eliminate the load of Integer.value for integers from the cache
never@17 1085 // array by deriving the value from the index into the array.
never@17 1086 // Capture the offset of the load and then reverse the computation.
never@17 1087 Node* load_base = base->in(Address)->in(AddPNode::Base);
never@17 1088 if (load_base != NULL) {
never@17 1089 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
never@17 1090 intptr_t cache_offset;
never@17 1091 int shift = -1;
never@17 1092 Node* cache = NULL;
never@17 1093 if (is_autobox_cache(atp)) {
kvn@29 1094 shift = exact_log2(type2aelembytes(T_OBJECT));
never@17 1095 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
never@17 1096 }
never@17 1097 if (cache != NULL && base->in(Address)->is_AddP()) {
never@17 1098 Node* elements[4];
never@17 1099 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
never@17 1100 int cache_low;
never@17 1101 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
never@17 1102 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
never@17 1103 // Add up all the offsets making of the address of the load
never@17 1104 Node* result = elements[0];
never@17 1105 for (int i = 1; i < count; i++) {
never@17 1106 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
never@17 1107 }
never@17 1108 // Remove the constant offset from the address and then
never@17 1109 // remove the scaling of the offset to recover the original index.
never@17 1110 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
never@17 1111 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
never@17 1112 // Peel the shift off directly but wrap it in a dummy node
never@17 1113 // since Ideal can't return existing nodes
never@17 1114 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
never@17 1115 } else {
never@17 1116 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
never@17 1117 }
never@17 1118 #ifdef _LP64
never@17 1119 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
never@17 1120 #endif
never@17 1121 return result;
never@17 1122 }
never@17 1123 }
never@17 1124 }
never@17 1125 }
never@17 1126 return NULL;
never@17 1127 }
never@17 1128
kvn@163 1129 //------------------------------split_through_phi------------------------------
kvn@163 1130 // Split instance field load through Phi.
kvn@163 1131 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
kvn@163 1132 Node* mem = in(MemNode::Memory);
kvn@163 1133 Node* address = in(MemNode::Address);
kvn@163 1134 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@163 1135 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@163 1136
kvn@163 1137 assert(mem->is_Phi() && (t_oop != NULL) &&
kvn@163 1138 t_oop->is_instance_field(), "invalide conditions");
kvn@163 1139
kvn@163 1140 Node *region = mem->in(0);
kvn@163 1141 if (region == NULL) {
kvn@163 1142 return NULL; // Wait stable graph
kvn@163 1143 }
kvn@163 1144 uint cnt = mem->req();
kvn@163 1145 for( uint i = 1; i < cnt; i++ ) {
kvn@163 1146 Node *in = mem->in(i);
kvn@163 1147 if( in == NULL ) {
kvn@163 1148 return NULL; // Wait stable graph
kvn@163 1149 }
kvn@163 1150 }
kvn@163 1151 // Check for loop invariant.
kvn@163 1152 if (cnt == 3) {
kvn@163 1153 for( uint i = 1; i < cnt; i++ ) {
kvn@163 1154 Node *in = mem->in(i);
kvn@163 1155 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
kvn@163 1156 if (m == mem) {
kvn@163 1157 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
kvn@163 1158 return this;
kvn@163 1159 }
kvn@163 1160 }
kvn@163 1161 }
kvn@163 1162 // Split through Phi (see original code in loopopts.cpp).
kvn@163 1163 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
kvn@163 1164
kvn@163 1165 // Do nothing here if Identity will find a value
kvn@163 1166 // (to avoid infinite chain of value phis generation).
kvn@163 1167 if ( !phase->eqv(this, this->Identity(phase)) )
kvn@163 1168 return NULL;
kvn@163 1169
kvn@163 1170 // Skip the split if the region dominates some control edge of the address.
kvn@163 1171 if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
kvn@163 1172 return NULL;
kvn@163 1173
kvn@163 1174 const Type* this_type = this->bottom_type();
kvn@163 1175 int this_index = phase->C->get_alias_index(addr_t);
kvn@163 1176 int this_offset = addr_t->offset();
kvn@163 1177 int this_iid = addr_t->is_oopptr()->instance_id();
kvn@163 1178 int wins = 0;
kvn@163 1179 PhaseIterGVN *igvn = phase->is_IterGVN();
kvn@163 1180 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
kvn@163 1181 for( uint i = 1; i < region->req(); i++ ) {
kvn@163 1182 Node *x;
kvn@163 1183 Node* the_clone = NULL;
kvn@163 1184 if( region->in(i) == phase->C->top() ) {
kvn@163 1185 x = phase->C->top(); // Dead path? Use a dead data op
kvn@163 1186 } else {
kvn@163 1187 x = this->clone(); // Else clone up the data op
kvn@163 1188 the_clone = x; // Remember for possible deletion.
kvn@163 1189 // Alter data node to use pre-phi inputs
kvn@163 1190 if( this->in(0) == region ) {
kvn@163 1191 x->set_req( 0, region->in(i) );
kvn@163 1192 } else {
kvn@163 1193 x->set_req( 0, NULL );
kvn@163 1194 }
kvn@163 1195 for( uint j = 1; j < this->req(); j++ ) {
kvn@163 1196 Node *in = this->in(j);
kvn@163 1197 if( in->is_Phi() && in->in(0) == region )
kvn@163 1198 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
kvn@163 1199 }
kvn@163 1200 }
kvn@163 1201 // Check for a 'win' on some paths
kvn@163 1202 const Type *t = x->Value(igvn);
kvn@163 1203
kvn@163 1204 bool singleton = t->singleton();
kvn@163 1205
kvn@163 1206 // See comments in PhaseIdealLoop::split_thru_phi().
kvn@163 1207 if( singleton && t == Type::TOP ) {
kvn@163 1208 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
kvn@163 1209 }
kvn@163 1210
kvn@163 1211 if( singleton ) {
kvn@163 1212 wins++;
kvn@163 1213 x = igvn->makecon(t);
kvn@163 1214 } else {
kvn@163 1215 // We now call Identity to try to simplify the cloned node.
kvn@163 1216 // Note that some Identity methods call phase->type(this).
kvn@163 1217 // Make sure that the type array is big enough for
kvn@163 1218 // our new node, even though we may throw the node away.
kvn@163 1219 // (This tweaking with igvn only works because x is a new node.)
kvn@163 1220 igvn->set_type(x, t);
kvn@163 1221 Node *y = x->Identity(igvn);
kvn@163 1222 if( y != x ) {
kvn@163 1223 wins++;
kvn@163 1224 x = y;
kvn@163 1225 } else {
kvn@163 1226 y = igvn->hash_find(x);
kvn@163 1227 if( y ) {
kvn@163 1228 wins++;
kvn@163 1229 x = y;
kvn@163 1230 } else {
kvn@163 1231 // Else x is a new node we are keeping
kvn@163 1232 // We do not need register_new_node_with_optimizer
kvn@163 1233 // because set_type has already been called.
kvn@163 1234 igvn->_worklist.push(x);
kvn@163 1235 }
kvn@163 1236 }
kvn@163 1237 }
kvn@163 1238 if (x != the_clone && the_clone != NULL)
kvn@163 1239 igvn->remove_dead_node(the_clone);
kvn@163 1240 phi->set_req(i, x);
kvn@163 1241 }
kvn@163 1242 if( wins > 0 ) {
kvn@163 1243 // Record Phi
kvn@163 1244 igvn->register_new_node_with_optimizer(phi);
kvn@163 1245 return phi;
kvn@163 1246 }
kvn@163 1247 igvn->remove_dead_node(phi);
kvn@163 1248 return NULL;
kvn@163 1249 }
never@17 1250
duke@0 1251 //------------------------------Ideal------------------------------------------
duke@0 1252 // If the load is from Field memory and the pointer is non-null, we can
duke@0 1253 // zero out the control input.
duke@0 1254 // If the offset is constant and the base is an object allocation,
duke@0 1255 // try to hook me up to the exact initializing store.
duke@0 1256 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1257 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 1258 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 1259
duke@0 1260 Node* ctrl = in(MemNode::Control);
duke@0 1261 Node* address = in(MemNode::Address);
duke@0 1262
duke@0 1263 // Skip up past a SafePoint control. Cannot do this for Stores because
duke@0 1264 // pointer stores & cardmarks must stay on the same side of a SafePoint.
duke@0 1265 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
duke@0 1266 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
duke@0 1267 ctrl = ctrl->in(0);
duke@0 1268 set_req(MemNode::Control,ctrl);
duke@0 1269 }
duke@0 1270
duke@0 1271 // Check for useless control edge in some common special cases
duke@0 1272 if (in(MemNode::Control) != NULL) {
duke@0 1273 intptr_t ignore = 0;
duke@0 1274 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
duke@0 1275 if (base != NULL
duke@0 1276 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
kvn@119 1277 && all_controls_dominate(base, phase->C->start())) {
duke@0 1278 // A method-invariant, non-null address (constant or 'this' argument).
duke@0 1279 set_req(MemNode::Control, NULL);
duke@0 1280 }
duke@0 1281 }
duke@0 1282
never@17 1283 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
never@17 1284 Node* base = in(Address)->in(AddPNode::Base);
never@17 1285 if (base != NULL) {
never@17 1286 Compile::AliasType* atp = phase->C->alias_type(adr_type());
never@17 1287 if (is_autobox_object(atp)) {
never@17 1288 Node* result = eliminate_autobox(phase);
never@17 1289 if (result != NULL) return result;
never@17 1290 }
never@17 1291 }
never@17 1292 }
never@17 1293
kvn@74 1294 Node* mem = in(MemNode::Memory);
kvn@74 1295 const TypePtr *addr_t = phase->type(address)->isa_ptr();
kvn@74 1296
kvn@74 1297 if (addr_t != NULL) {
kvn@74 1298 // try to optimize our memory input
kvn@74 1299 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
kvn@74 1300 if (opt_mem != mem) {
kvn@74 1301 set_req(MemNode::Memory, opt_mem);
kvn@74 1302 return this;
kvn@74 1303 }
kvn@74 1304 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
kvn@74 1305 if (can_reshape && opt_mem->is_Phi() &&
kvn@74 1306 (t_oop != NULL) && t_oop->is_instance_field()) {
kvn@163 1307 // Split instance field load through Phi.
kvn@163 1308 Node* result = split_through_phi(phase);
kvn@163 1309 if (result != NULL) return result;
kvn@74 1310 }
kvn@74 1311 }
kvn@74 1312
duke@0 1313 // Check for prior store with a different base or offset; make Load
duke@0 1314 // independent. Skip through any number of them. Bail out if the stores
duke@0 1315 // are in an endless dead cycle and report no progress. This is a key
duke@0 1316 // transform for Reflection. However, if after skipping through the Stores
duke@0 1317 // we can't then fold up against a prior store do NOT do the transform as
duke@0 1318 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
duke@0 1319 // array memory alive twice: once for the hoisted Load and again after the
duke@0 1320 // bypassed Store. This situation only works if EVERYBODY who does
duke@0 1321 // anti-dependence work knows how to bypass. I.e. we need all
duke@0 1322 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
duke@0 1323 // the alias index stuff. So instead, peek through Stores and IFF we can
duke@0 1324 // fold up, do so.
duke@0 1325 Node* prev_mem = find_previous_store(phase);
duke@0 1326 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 1327 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
duke@0 1328 // (c) See if we can fold up on the spot, but don't fold up here.
duke@0 1329 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
duke@0 1330 // just return a prior value, which is done by Identity calls.
duke@0 1331 if (can_see_stored_value(prev_mem, phase)) {
duke@0 1332 // Make ready for step (d):
duke@0 1333 set_req(MemNode::Memory, prev_mem);
duke@0 1334 return this;
duke@0 1335 }
duke@0 1336 }
duke@0 1337
duke@0 1338 return NULL; // No further progress
duke@0 1339 }
duke@0 1340
duke@0 1341 // Helper to recognize certain Klass fields which are invariant across
duke@0 1342 // some group of array types (e.g., int[] or all T[] where T < Object).
duke@0 1343 const Type*
duke@0 1344 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
duke@0 1345 ciKlass* klass) const {
duke@0 1346 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1347 // The field is Klass::_modifier_flags. Return its (constant) value.
duke@0 1348 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
duke@0 1349 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
duke@0 1350 return TypeInt::make(klass->modifier_flags());
duke@0 1351 }
duke@0 1352 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1353 // The field is Klass::_access_flags. Return its (constant) value.
duke@0 1354 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
duke@0 1355 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
duke@0 1356 return TypeInt::make(klass->access_flags());
duke@0 1357 }
duke@0 1358 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1359 // The field is Klass::_layout_helper. Return its constant value if known.
duke@0 1360 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1361 return TypeInt::make(klass->layout_helper());
duke@0 1362 }
duke@0 1363
duke@0 1364 // No match.
duke@0 1365 return NULL;
duke@0 1366 }
duke@0 1367
duke@0 1368 //------------------------------Value-----------------------------------------
duke@0 1369 const Type *LoadNode::Value( PhaseTransform *phase ) const {
duke@0 1370 // Either input is TOP ==> the result is TOP
duke@0 1371 Node* mem = in(MemNode::Memory);
duke@0 1372 const Type *t1 = phase->type(mem);
duke@0 1373 if (t1 == Type::TOP) return Type::TOP;
duke@0 1374 Node* adr = in(MemNode::Address);
duke@0 1375 const TypePtr* tp = phase->type(adr)->isa_ptr();
duke@0 1376 if (tp == NULL || tp->empty()) return Type::TOP;
duke@0 1377 int off = tp->offset();
duke@0 1378 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
duke@0 1379
duke@0 1380 // Try to guess loaded type from pointer type
duke@0 1381 if (tp->base() == Type::AryPtr) {
duke@0 1382 const Type *t = tp->is_aryptr()->elem();
duke@0 1383 // Don't do this for integer types. There is only potential profit if
duke@0 1384 // the element type t is lower than _type; that is, for int types, if _type is
duke@0 1385 // more restrictive than t. This only happens here if one is short and the other
duke@0 1386 // char (both 16 bits), and in those cases we've made an intentional decision
duke@0 1387 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
duke@0 1388 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
duke@0 1389 //
duke@0 1390 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
duke@0 1391 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
duke@0 1392 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
duke@0 1393 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
duke@0 1394 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
duke@0 1395 // In fact, that could have been the original type of p1, and p1 could have
duke@0 1396 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
duke@0 1397 // expression (LShiftL quux 3) independently optimized to the constant 8.
duke@0 1398 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
duke@0 1399 && Opcode() != Op_LoadKlass) {
duke@0 1400 // t might actually be lower than _type, if _type is a unique
duke@0 1401 // concrete subclass of abstract class t.
duke@0 1402 // Make sure the reference is not into the header, by comparing
duke@0 1403 // the offset against the offset of the start of the array's data.
duke@0 1404 // Different array types begin at slightly different offsets (12 vs. 16).
duke@0 1405 // We choose T_BYTE as an example base type that is least restrictive
duke@0 1406 // as to alignment, which will therefore produce the smallest
duke@0 1407 // possible base offset.
duke@0 1408 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
duke@0 1409 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
duke@0 1410 const Type* jt = t->join(_type);
duke@0 1411 // In any case, do not allow the join, per se, to empty out the type.
duke@0 1412 if (jt->empty() && !t->empty()) {
duke@0 1413 // This can happen if a interface-typed array narrows to a class type.
duke@0 1414 jt = _type;
duke@0 1415 }
never@17 1416
never@17 1417 if (EliminateAutoBox) {
never@17 1418 // The pointers in the autobox arrays are always non-null
never@17 1419 Node* base = in(Address)->in(AddPNode::Base);
never@17 1420 if (base != NULL) {
never@17 1421 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
never@17 1422 if (is_autobox_cache(atp)) {
never@17 1423 return jt->join(TypePtr::NOTNULL)->is_ptr();
never@17 1424 }
never@17 1425 }
never@17 1426 }
duke@0 1427 return jt;
duke@0 1428 }
duke@0 1429 }
duke@0 1430 } else if (tp->base() == Type::InstPtr) {
duke@0 1431 assert( off != Type::OffsetBot ||
duke@0 1432 // arrays can be cast to Objects
duke@0 1433 tp->is_oopptr()->klass()->is_java_lang_Object() ||
duke@0 1434 // unsafe field access may not have a constant offset
duke@0 1435 phase->C->has_unsafe_access(),
duke@0 1436 "Field accesses must be precise" );
duke@0 1437 // For oop loads, we expect the _type to be precise
duke@0 1438 } else if (tp->base() == Type::KlassPtr) {
duke@0 1439 assert( off != Type::OffsetBot ||
duke@0 1440 // arrays can be cast to Objects
duke@0 1441 tp->is_klassptr()->klass()->is_java_lang_Object() ||
duke@0 1442 // also allow array-loading from the primary supertype
duke@0 1443 // array during subtype checks
duke@0 1444 Opcode() == Op_LoadKlass,
duke@0 1445 "Field accesses must be precise" );
duke@0 1446 // For klass/static loads, we expect the _type to be precise
duke@0 1447 }
duke@0 1448
duke@0 1449 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1450 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1451 ciKlass* klass = tkls->klass();
duke@0 1452 if (klass->is_loaded() && tkls->klass_is_exact()) {
duke@0 1453 // We are loading a field from a Klass metaobject whose identity
duke@0 1454 // is known at compile time (the type is "exact" or "precise").
duke@0 1455 // Check for fields we know are maintained as constants by the VM.
duke@0 1456 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1457 // The field is Klass::_super_check_offset. Return its (constant) value.
duke@0 1458 // (Folds up type checking code.)
duke@0 1459 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
duke@0 1460 return TypeInt::make(klass->super_check_offset());
duke@0 1461 }
duke@0 1462 // Compute index into primary_supers array
duke@0 1463 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1464 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1465 if( depth < ciKlass::primary_super_limit() ) {
duke@0 1466 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1467 // (Folds up type checking code.)
duke@0 1468 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1469 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1470 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1471 }
duke@0 1472 const Type* aift = load_array_final_field(tkls, klass);
duke@0 1473 if (aift != NULL) return aift;
duke@0 1474 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
duke@0 1475 && klass->is_array_klass()) {
duke@0 1476 // The field is arrayKlass::_component_mirror. Return its (constant) value.
duke@0 1477 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
duke@0 1478 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
duke@0 1479 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
duke@0 1480 }
duke@0 1481 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
duke@0 1482 // The field is Klass::_java_mirror. Return its (constant) value.
duke@0 1483 // (Folds up the 2nd indirection in anObjConstant.getClass().)
duke@0 1484 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
duke@0 1485 return TypeInstPtr::make(klass->java_mirror());
duke@0 1486 }
duke@0 1487 }
duke@0 1488
duke@0 1489 // We can still check if we are loading from the primary_supers array at a
duke@0 1490 // shallow enough depth. Even though the klass is not exact, entries less
duke@0 1491 // than or equal to its super depth are correct.
duke@0 1492 if (klass->is_loaded() ) {
duke@0 1493 ciType *inner = klass->klass();
duke@0 1494 while( inner->is_obj_array_klass() )
duke@0 1495 inner = inner->as_obj_array_klass()->base_element_type();
duke@0 1496 if( inner->is_instance_klass() &&
duke@0 1497 !inner->as_instance_klass()->flags().is_interface() ) {
duke@0 1498 // Compute index into primary_supers array
duke@0 1499 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
duke@0 1500 // Check for overflowing; use unsigned compare to handle the negative case.
duke@0 1501 if( depth < ciKlass::primary_super_limit() &&
duke@0 1502 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
duke@0 1503 // The field is an element of Klass::_primary_supers. Return its (constant) value.
duke@0 1504 // (Folds up type checking code.)
duke@0 1505 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
duke@0 1506 ciKlass *ss = klass->super_of_depth(depth);
duke@0 1507 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
duke@0 1508 }
duke@0 1509 }
duke@0 1510 }
duke@0 1511
duke@0 1512 // If the type is enough to determine that the thing is not an array,
duke@0 1513 // we can give the layout_helper a positive interval type.
duke@0 1514 // This will help short-circuit some reflective code.
duke@0 1515 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
duke@0 1516 && !klass->is_array_klass() // not directly typed as an array
duke@0 1517 && !klass->is_interface() // specifically not Serializable & Cloneable
duke@0 1518 && !klass->is_java_lang_Object() // not the supertype of all T[]
duke@0 1519 ) {
duke@0 1520 // Note: When interfaces are reliable, we can narrow the interface
duke@0 1521 // test to (klass != Serializable && klass != Cloneable).
duke@0 1522 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
duke@0 1523 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
duke@0 1524 // The key property of this type is that it folds up tests
duke@0 1525 // for array-ness, since it proves that the layout_helper is positive.
duke@0 1526 // Thus, a generic value like the basic object layout helper works fine.
duke@0 1527 return TypeInt::make(min_size, max_jint, Type::WidenMin);
duke@0 1528 }
duke@0 1529 }
duke@0 1530
duke@0 1531 // If we are loading from a freshly-allocated object, produce a zero,
duke@0 1532 // if the load is provably beyond the header of the object.
duke@0 1533 // (Also allow a variable load from a fresh array to produce zero.)
duke@0 1534 if (ReduceFieldZeroing) {
duke@0 1535 Node* value = can_see_stored_value(mem,phase);
duke@0 1536 if (value != NULL && value->is_Con())
duke@0 1537 return value->bottom_type();
duke@0 1538 }
duke@0 1539
kvn@64 1540 const TypeOopPtr *tinst = tp->isa_oopptr();
kvn@64 1541 if (tinst != NULL && tinst->is_instance_field()) {
kvn@64 1542 // If we have an instance type and our memory input is the
kvn@64 1543 // programs's initial memory state, there is no matching store,
kvn@64 1544 // so just return a zero of the appropriate type
kvn@64 1545 Node *mem = in(MemNode::Memory);
kvn@64 1546 if (mem->is_Parm() && mem->in(0)->is_Start()) {
kvn@64 1547 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
kvn@64 1548 return Type::get_zero_type(_type->basic_type());
kvn@64 1549 }
kvn@64 1550 }
duke@0 1551 return _type;
duke@0 1552 }
duke@0 1553
duke@0 1554 //------------------------------match_edge-------------------------------------
duke@0 1555 // Do we Match on this edge index or not? Match only the address.
duke@0 1556 uint LoadNode::match_edge(uint idx) const {
duke@0 1557 return idx == MemNode::Address;
duke@0 1558 }
duke@0 1559
duke@0 1560 //--------------------------LoadBNode::Ideal--------------------------------------
duke@0 1561 //
duke@0 1562 // If the previous store is to the same address as this load,
duke@0 1563 // and the value stored was larger than a byte, replace this load
duke@0 1564 // with the value stored truncated to a byte. If no truncation is
duke@0 1565 // needed, the replacement is done in LoadNode::Identity().
duke@0 1566 //
duke@0 1567 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1568 Node* mem = in(MemNode::Memory);
duke@0 1569 Node* value = can_see_stored_value(mem,phase);
duke@0 1570 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1571 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
duke@0 1572 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
duke@0 1573 }
duke@0 1574 // Identity call will handle the case where truncation is not needed.
duke@0 1575 return LoadNode::Ideal(phase, can_reshape);
duke@0 1576 }
duke@0 1577
duke@0 1578 //--------------------------LoadCNode::Ideal--------------------------------------
duke@0 1579 //
duke@0 1580 // If the previous store is to the same address as this load,
duke@0 1581 // and the value stored was larger than a char, replace this load
duke@0 1582 // with the value stored truncated to a char. If no truncation is
duke@0 1583 // needed, the replacement is done in LoadNode::Identity().
duke@0 1584 //
duke@0 1585 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1586 Node* mem = in(MemNode::Memory);
duke@0 1587 Node* value = can_see_stored_value(mem,phase);
duke@0 1588 if( value && !phase->type(value)->higher_equal( _type ) )
duke@0 1589 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
duke@0 1590 // Identity call will handle the case where truncation is not needed.
duke@0 1591 return LoadNode::Ideal(phase, can_reshape);
duke@0 1592 }
duke@0 1593
duke@0 1594 //--------------------------LoadSNode::Ideal--------------------------------------
duke@0 1595 //
duke@0 1596 // If the previous store is to the same address as this load,
duke@0 1597 // and the value stored was larger than a short, replace this load
duke@0 1598 // with the value stored truncated to a short. If no truncation is
duke@0 1599 // needed, the replacement is done in LoadNode::Identity().
duke@0 1600 //
duke@0 1601 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1602 Node* mem = in(MemNode::Memory);
duke@0 1603 Node* value = can_see_stored_value(mem,phase);
duke@0 1604 if( value && !phase->type(value)->higher_equal( _type ) ) {
duke@0 1605 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
duke@0 1606 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
duke@0 1607 }
duke@0 1608 // Identity call will handle the case where truncation is not needed.
duke@0 1609 return LoadNode::Ideal(phase, can_reshape);
duke@0 1610 }
duke@0 1611
duke@0 1612 //=============================================================================
duke@0 1613 //------------------------------Value------------------------------------------
duke@0 1614 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
duke@0 1615 // Either input is TOP ==> the result is TOP
duke@0 1616 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1617 if (t1 == Type::TOP) return Type::TOP;
duke@0 1618 Node *adr = in(MemNode::Address);
duke@0 1619 const Type *t2 = phase->type( adr );
duke@0 1620 if (t2 == Type::TOP) return Type::TOP;
duke@0 1621 const TypePtr *tp = t2->is_ptr();
duke@0 1622 if (TypePtr::above_centerline(tp->ptr()) ||
duke@0 1623 tp->ptr() == TypePtr::Null) return Type::TOP;
duke@0 1624
duke@0 1625 // Return a more precise klass, if possible
duke@0 1626 const TypeInstPtr *tinst = tp->isa_instptr();
duke@0 1627 if (tinst != NULL) {
duke@0 1628 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
duke@0 1629 int offset = tinst->offset();
duke@0 1630 if (ik == phase->C->env()->Class_klass()
duke@0 1631 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1632 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1633 // We are loading a special hidden field from a Class mirror object,
duke@0 1634 // the field which points to the VM's Klass metaobject.
duke@0 1635 ciType* t = tinst->java_mirror_type();
duke@0 1636 // java_mirror_type returns non-null for compile-time Class constants.
duke@0 1637 if (t != NULL) {
duke@0 1638 // constant oop => constant klass
duke@0 1639 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1640 return TypeKlassPtr::make(ciArrayKlass::make(t));
duke@0 1641 }
duke@0 1642 if (!t->is_klass()) {
duke@0 1643 // a primitive Class (e.g., int.class) has NULL for a klass field
duke@0 1644 return TypePtr::NULL_PTR;
duke@0 1645 }
duke@0 1646 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
duke@0 1647 return TypeKlassPtr::make(t->as_klass());
duke@0 1648 }
duke@0 1649 // non-constant mirror, so we can't tell what's going on
duke@0 1650 }
duke@0 1651 if( !ik->is_loaded() )
duke@0 1652 return _type; // Bail out if not loaded
duke@0 1653 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1654 if (tinst->klass_is_exact()) {
duke@0 1655 return TypeKlassPtr::make(ik);
duke@0 1656 }
duke@0 1657 // See if we can become precise: no subklasses and no interface
duke@0 1658 // (Note: We need to support verified interfaces.)
duke@0 1659 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1660 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1661 // Add a dependence; if any subclass added we need to recompile
duke@0 1662 if (!ik->is_final()) {
duke@0 1663 // %%% should use stronger assert_unique_concrete_subtype instead
duke@0 1664 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1665 }
duke@0 1666 // Return precise klass
duke@0 1667 return TypeKlassPtr::make(ik);
duke@0 1668 }
duke@0 1669
duke@0 1670 // Return root of possible klass
duke@0 1671 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
duke@0 1672 }
duke@0 1673 }
duke@0 1674
duke@0 1675 // Check for loading klass from an array
duke@0 1676 const TypeAryPtr *tary = tp->isa_aryptr();
duke@0 1677 if( tary != NULL ) {
duke@0 1678 ciKlass *tary_klass = tary->klass();
duke@0 1679 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
duke@0 1680 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
duke@0 1681 if (tary->klass_is_exact()) {
duke@0 1682 return TypeKlassPtr::make(tary_klass);
duke@0 1683 }
duke@0 1684 ciArrayKlass *ak = tary->klass()->as_array_klass();
duke@0 1685 // If the klass is an object array, we defer the question to the
duke@0 1686 // array component klass.
duke@0 1687 if( ak->is_obj_array_klass() ) {
duke@0 1688 assert( ak->is_loaded(), "" );
duke@0 1689 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
duke@0 1690 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
duke@0 1691 ciInstanceKlass* ik = base_k->as_instance_klass();
duke@0 1692 // See if we can become precise: no subklasses and no interface
duke@0 1693 if (!ik->is_interface() && !ik->has_subklass()) {
duke@0 1694 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1695 // Add a dependence; if any subclass added we need to recompile
duke@0 1696 if (!ik->is_final()) {
duke@0 1697 phase->C->dependencies()->assert_leaf_type(ik);
duke@0 1698 }
duke@0 1699 // Return precise array klass
duke@0 1700 return TypeKlassPtr::make(ak);
duke@0 1701 }
duke@0 1702 }
duke@0 1703 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@0 1704 } else { // Found a type-array?
duke@0 1705 //assert(!UseExactTypes, "this code should be useless with exact types");
duke@0 1706 assert( ak->is_type_array_klass(), "" );
duke@0 1707 return TypeKlassPtr::make(ak); // These are always precise
duke@0 1708 }
duke@0 1709 }
duke@0 1710 }
duke@0 1711
duke@0 1712 // Check for loading klass from an array klass
duke@0 1713 const TypeKlassPtr *tkls = tp->isa_klassptr();
duke@0 1714 if (tkls != NULL && !StressReflectiveCode) {
duke@0 1715 ciKlass* klass = tkls->klass();
duke@0 1716 if( !klass->is_loaded() )
duke@0 1717 return _type; // Bail out if not loaded
duke@0 1718 if( klass->is_obj_array_klass() &&
duke@0 1719 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1720 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
duke@0 1721 // // Always returning precise element type is incorrect,
duke@0 1722 // // e.g., element type could be object and array may contain strings
duke@0 1723 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
duke@0 1724
duke@0 1725 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
duke@0 1726 // according to the element type's subclassing.
duke@0 1727 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
duke@0 1728 }
duke@0 1729 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
duke@0 1730 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
duke@0 1731 ciKlass* sup = klass->as_instance_klass()->super();
duke@0 1732 // The field is Klass::_super. Return its (constant) value.
duke@0 1733 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
duke@0 1734 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
duke@0 1735 }
duke@0 1736 }
duke@0 1737
duke@0 1738 // Bailout case
duke@0 1739 return LoadNode::Value(phase);
duke@0 1740 }
duke@0 1741
duke@0 1742 //------------------------------Identity---------------------------------------
duke@0 1743 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
duke@0 1744 // Also feed through the klass in Allocate(...klass...)._klass.
duke@0 1745 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
duke@0 1746 Node* x = LoadNode::Identity(phase);
duke@0 1747 if (x != this) return x;
duke@0 1748
duke@0 1749 // Take apart the address into an oop and and offset.
duke@0 1750 // Return 'this' if we cannot.
duke@0 1751 Node* adr = in(MemNode::Address);
duke@0 1752 intptr_t offset = 0;
duke@0 1753 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1754 if (base == NULL) return this;
duke@0 1755 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
duke@0 1756 if (toop == NULL) return this;
duke@0 1757
duke@0 1758 // We can fetch the klass directly through an AllocateNode.
duke@0 1759 // This works even if the klass is not constant (clone or newArray).
duke@0 1760 if (offset == oopDesc::klass_offset_in_bytes()) {
duke@0 1761 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
duke@0 1762 if (allocated_klass != NULL) {
duke@0 1763 return allocated_klass;
duke@0 1764 }
duke@0 1765 }
duke@0 1766
duke@0 1767 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
duke@0 1768 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
duke@0 1769 // See inline_native_Class_query for occurrences of these patterns.
duke@0 1770 // Java Example: x.getClass().isAssignableFrom(y)
duke@0 1771 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
duke@0 1772 //
duke@0 1773 // This improves reflective code, often making the Class
duke@0 1774 // mirror go completely dead. (Current exception: Class
duke@0 1775 // mirrors may appear in debug info, but we could clean them out by
duke@0 1776 // introducing a new debug info operator for klassOop.java_mirror).
duke@0 1777 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
duke@0 1778 && (offset == java_lang_Class::klass_offset_in_bytes() ||
duke@0 1779 offset == java_lang_Class::array_klass_offset_in_bytes())) {
duke@0 1780 // We are loading a special hidden field from a Class mirror,
duke@0 1781 // the field which points to its Klass or arrayKlass metaobject.
duke@0 1782 if (base->is_Load()) {
duke@0 1783 Node* adr2 = base->in(MemNode::Address);
duke@0 1784 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
duke@0 1785 if (tkls != NULL && !tkls->empty()
duke@0 1786 && (tkls->klass()->is_instance_klass() ||
duke@0 1787 tkls->klass()->is_array_klass())
duke@0 1788 && adr2->is_AddP()
duke@0 1789 ) {
duke@0 1790 int mirror_field = Klass::java_mirror_offset_in_bytes();
duke@0 1791 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
duke@0 1792 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
duke@0 1793 }
duke@0 1794 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
duke@0 1795 return adr2->in(AddPNode::Base);
duke@0 1796 }
duke@0 1797 }
duke@0 1798 }
duke@0 1799 }
duke@0 1800
duke@0 1801 return this;
duke@0 1802 }
duke@0 1803
duke@0 1804 //------------------------------Value-----------------------------------------
duke@0 1805 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
duke@0 1806 // Either input is TOP ==> the result is TOP
duke@0 1807 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1808 if( t1 == Type::TOP ) return Type::TOP;
duke@0 1809 Node *adr = in(MemNode::Address);
duke@0 1810 const Type *t2 = phase->type( adr );
duke@0 1811 if( t2 == Type::TOP ) return Type::TOP;
duke@0 1812 const TypePtr *tp = t2->is_ptr();
duke@0 1813 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
duke@0 1814 const TypeAryPtr *tap = tp->isa_aryptr();
duke@0 1815 if( !tap ) return _type;
duke@0 1816 return tap->size();
duke@0 1817 }
duke@0 1818
duke@0 1819 //------------------------------Identity---------------------------------------
duke@0 1820 // Feed through the length in AllocateArray(...length...)._length.
duke@0 1821 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
duke@0 1822 Node* x = LoadINode::Identity(phase);
duke@0 1823 if (x != this) return x;
duke@0 1824
duke@0 1825 // Take apart the address into an oop and and offset.
duke@0 1826 // Return 'this' if we cannot.
duke@0 1827 Node* adr = in(MemNode::Address);
duke@0 1828 intptr_t offset = 0;
duke@0 1829 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
duke@0 1830 if (base == NULL) return this;
duke@0 1831 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
duke@0 1832 if (tary == NULL) return this;
duke@0 1833
duke@0 1834 // We can fetch the length directly through an AllocateArrayNode.
duke@0 1835 // This works even if the length is not constant (clone or newArray).
duke@0 1836 if (offset == arrayOopDesc::length_offset_in_bytes()) {
duke@0 1837 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
duke@0 1838 if (allocated_length != NULL) {
duke@0 1839 return allocated_length;
duke@0 1840 }
duke@0 1841 }
duke@0 1842
duke@0 1843 return this;
duke@0 1844
duke@0 1845 }
duke@0 1846 //=============================================================================
duke@0 1847 //---------------------------StoreNode::make-----------------------------------
duke@0 1848 // Polymorphic factory method:
coleenp@113 1849 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
coleenp@113 1850 Compile* C = gvn.C;
coleenp@113 1851
duke@0 1852 switch (bt) {
duke@0 1853 case T_BOOLEAN:
duke@0 1854 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
duke@0 1855 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
duke@0 1856 case T_CHAR:
duke@0 1857 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
duke@0 1858 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
duke@0 1859 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
duke@0 1860 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
duke@0 1861 case T_ADDRESS:
coleenp@113 1862 case T_OBJECT:
coleenp@113 1863 #ifdef _LP64
kvn@163 1864 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
coleenp@113 1865 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
coleenp@113 1866 adr->bottom_type()->isa_rawptr())) {
coleenp@113 1867 const TypePtr* type = val->bottom_type()->is_ptr();
kvn@124 1868 Node* cp = EncodePNode::encode(&gvn, val);
coleenp@113 1869 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, cp);
coleenp@113 1870 } else
coleenp@113 1871 #endif
coleenp@113 1872 {
coleenp@113 1873 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
coleenp@113 1874 }
duke@0 1875 }
duke@0 1876 ShouldNotReachHere();
duke@0 1877 return (StoreNode*)NULL;
duke@0 1878 }
duke@0 1879
duke@0 1880 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
duke@0 1881 bool require_atomic = true;
duke@0 1882 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
duke@0 1883 }
duke@0 1884
duke@0 1885
duke@0 1886 //--------------------------bottom_type----------------------------------------
duke@0 1887 const Type *StoreNode::bottom_type() const {
duke@0 1888 return Type::MEMORY;
duke@0 1889 }
duke@0 1890
duke@0 1891 //------------------------------hash-------------------------------------------
duke@0 1892 uint StoreNode::hash() const {
duke@0 1893 // unroll addition of interesting fields
duke@0 1894 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
duke@0 1895
duke@0 1896 // Since they are not commoned, do not hash them:
duke@0 1897 return NO_HASH;
duke@0 1898 }
duke@0 1899
duke@0 1900 //------------------------------Ideal------------------------------------------
duke@0 1901 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
duke@0 1902 // When a store immediately follows a relevant allocation/initialization,
duke@0 1903 // try to capture it into the initialization, or hoist it above.
duke@0 1904 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 1905 Node* p = MemNode::Ideal_common(phase, can_reshape);
duke@0 1906 if (p) return (p == NodeSentinel) ? NULL : p;
duke@0 1907
duke@0 1908 Node* mem = in(MemNode::Memory);
duke@0 1909 Node* address = in(MemNode::Address);
duke@0 1910
duke@0 1911 // Back-to-back stores to same address? Fold em up.
duke@0 1912 // Generally unsafe if I have intervening uses...
duke@0 1913 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
duke@0 1914 // Looking at a dead closed cycle of memory?
duke@0 1915 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
duke@0 1916
duke@0 1917 assert(Opcode() == mem->Opcode() ||
duke@0 1918 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
duke@0 1919 "no mismatched stores, except on raw memory");
duke@0 1920
duke@0 1921 if (mem->outcnt() == 1 && // check for intervening uses
duke@0 1922 mem->as_Store()->memory_size() <= this->memory_size()) {
duke@0 1923 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
duke@0 1924 // For example, 'mem' might be the final state at a conditional return.
duke@0 1925 // Or, 'mem' might be used by some node which is live at the same time
duke@0 1926 // 'this' is live, which might be unschedulable. So, require exactly
duke@0 1927 // ONE user, the 'this' store, until such time as we clone 'mem' for
duke@0 1928 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
duke@0 1929 if (can_reshape) { // (%%% is this an anachronism?)
duke@0 1930 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
duke@0 1931 phase->is_IterGVN());
duke@0 1932 } else {
duke@0 1933 // It's OK to do this in the parser, since DU info is always accurate,
duke@0 1934 // and the parser always refers to nodes via SafePointNode maps.
duke@0 1935 set_req(MemNode::Memory, mem->in(MemNode::Memory));
duke@0 1936 }
duke@0 1937 return this;
duke@0 1938 }
duke@0 1939 }
duke@0 1940
duke@0 1941 // Capture an unaliased, unconditional, simple store into an initializer.
duke@0 1942 // Or, if it is independent of the allocation, hoist it above the allocation.
duke@0 1943 if (ReduceFieldZeroing && /*can_reshape &&*/
duke@0 1944 mem->is_Proj() && mem->in(0)->is_Initialize()) {
duke@0 1945 InitializeNode* init = mem->in(0)->as_Initialize();
duke@0 1946 intptr_t offset = init->can_capture_store(this, phase);
duke@0 1947 if (offset > 0) {
duke@0 1948 Node* moved = init->capture_store(this, offset, phase);
duke@0 1949 // If the InitializeNode captured me, it made a raw copy of me,
duke@0 1950 // and I need to disappear.
duke@0 1951 if (moved != NULL) {
duke@0 1952 // %%% hack to ensure that Ideal returns a new node:
duke@0 1953 mem = MergeMemNode::make(phase->C, mem);
duke@0 1954 return mem; // fold me away
duke@0 1955 }
duke@0 1956 }
duke@0 1957 }
duke@0 1958
duke@0 1959 return NULL; // No further progress
duke@0 1960 }
duke@0 1961
duke@0 1962 //------------------------------Value-----------------------------------------
duke@0 1963 const Type *StoreNode::Value( PhaseTransform *phase ) const {
duke@0 1964 // Either input is TOP ==> the result is TOP
duke@0 1965 const Type *t1 = phase->type( in(MemNode::Memory) );
duke@0 1966 if( t1 == Type::TOP ) return Type::TOP;
duke@0 1967 const Type *t2 = phase->type( in(MemNode::Address) );
duke@0 1968 if( t2 == Type::TOP ) return Type::TOP;
duke@0 1969 const Type *t3 = phase->type( in(MemNode::ValueIn) );
duke@0 1970 if( t3 == Type::TOP ) return Type::TOP;
duke@0 1971 return Type::MEMORY;
duke@0 1972 }
duke@0 1973
duke@0 1974 //------------------------------Identity---------------------------------------
duke@0 1975 // Remove redundant stores:
duke@0 1976 // Store(m, p, Load(m, p)) changes to m.
duke@0 1977 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
duke@0 1978 Node *StoreNode::Identity( PhaseTransform *phase ) {
duke@0 1979 Node* mem = in(MemNode::Memory);
duke@0 1980 Node* adr = in(MemNode::Address);
duke@0 1981 Node* val = in(MemNode::ValueIn);
duke@0 1982
duke@0 1983 // Load then Store? Then the Store is useless
duke@0 1984 if (val->is_Load() &&
duke@0 1985 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
duke@0 1986 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
duke@0 1987 val->as_Load()->store_Opcode() == Opcode()) {
duke@0 1988 return mem;
duke@0 1989 }
duke@0 1990
duke@0 1991 // Two stores in a row of the same value?
duke@0 1992 if (mem->is_Store() &&
duke@0 1993 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
duke@0 1994 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
duke@0 1995 mem->Opcode() == Opcode()) {
duke@0 1996 return mem;
duke@0 1997 }
duke@0 1998
duke@0 1999 // Store of zero anywhere into a freshly-allocated object?
duke@0 2000 // Then the store is useless.
duke@0 2001 // (It must already have been captured by the InitializeNode.)
duke@0 2002 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
duke@0 2003 // a newly allocated object is already all-zeroes everywhere
duke@0 2004 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
duke@0 2005 return mem;
duke@0 2006 }
duke@0 2007
duke@0 2008 // the store may also apply to zero-bits in an earlier object
duke@0 2009 Node* prev_mem = find_previous_store(phase);
duke@0 2010 // Steps (a), (b): Walk past independent stores to find an exact match.
duke@0 2011 if (prev_mem != NULL) {
duke@0 2012 Node* prev_val = can_see_stored_value(prev_mem, phase);
duke@0 2013 if (prev_val != NULL && phase->eqv(prev_val, val)) {
duke@0 2014 // prev_val and val might differ by a cast; it would be good
duke@0 2015 // to keep the more informative of the two.
duke@0 2016 return mem;
duke@0 2017 }
duke@0 2018 }
duke@0 2019 }
duke@0 2020
duke@0 2021 return this;
duke@0 2022 }
duke@0 2023
duke@0 2024 //------------------------------match_edge-------------------------------------
duke@0 2025 // Do we Match on this edge index or not? Match only memory & value
duke@0 2026 uint StoreNode::match_edge(uint idx) const {
duke@0 2027 return idx == MemNode::Address || idx == MemNode::ValueIn;
duke@0 2028 }
duke@0 2029
duke@0 2030 //------------------------------cmp--------------------------------------------
duke@0 2031 // Do not common stores up together. They generally have to be split
duke@0 2032 // back up anyways, so do not bother.
duke@0 2033 uint StoreNode::cmp( const Node &n ) const {
duke@0 2034 return (&n == this); // Always fail except on self
duke@0 2035 }
duke@0 2036
duke@0 2037 //------------------------------Ideal_masked_input-----------------------------
duke@0 2038 // Check for a useless mask before a partial-word store
duke@0 2039 // (StoreB ... (AndI valIn conIa) )
duke@0 2040 // If (conIa & mask == mask) this simplifies to
duke@0 2041 // (StoreB ... (valIn) )
duke@0 2042 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
duke@0 2043 Node *val = in(MemNode::ValueIn);
duke@0 2044 if( val->Opcode() == Op_AndI ) {
duke@0 2045 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2046 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
duke@0 2047 set_req(MemNode::ValueIn, val->in(1));
duke@0 2048 return this;
duke@0 2049 }
duke@0 2050 }
duke@0 2051 return NULL;
duke@0 2052 }
duke@0 2053
duke@0 2054
duke@0 2055 //------------------------------Ideal_sign_extended_input----------------------
duke@0 2056 // Check for useless sign-extension before a partial-word store
duke@0 2057 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
duke@0 2058 // If (conIL == conIR && conIR <= num_bits) this simplifies to
duke@0 2059 // (StoreB ... (valIn) )
duke@0 2060 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
duke@0 2061 Node *val = in(MemNode::ValueIn);
duke@0 2062 if( val->Opcode() == Op_RShiftI ) {
duke@0 2063 const TypeInt *t = phase->type( val->in(2) )->isa_int();
duke@0 2064 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
duke@0 2065 Node *shl = val->in(1);
duke@0 2066 if( shl->Opcode() == Op_LShiftI ) {
duke@0 2067 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
duke@0 2068 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
duke@0 2069 set_req(MemNode::ValueIn, shl->in(1));
duke@0 2070 return this;
duke@0 2071 }
duke@0 2072 }
duke@0 2073 }
duke@0 2074 }
duke@0 2075 return NULL;
duke@0 2076 }
duke@0 2077
duke@0 2078 //------------------------------value_never_loaded-----------------------------------
duke@0 2079 // Determine whether there are any possible loads of the value stored.
duke@0 2080 // For simplicity, we actually check if there are any loads from the
duke@0 2081 // address stored to, not just for loads of the value stored by this node.
duke@0 2082 //
duke@0 2083 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
duke@0 2084 Node *adr = in(Address);
duke@0 2085 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
duke@0 2086 if (adr_oop == NULL)
duke@0 2087 return false;
kvn@64 2088 if (!adr_oop->is_instance_field())
duke@0 2089 return false; // if not a distinct instance, there may be aliases of the address
duke@0 2090 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
duke@0 2091 Node *use = adr->fast_out(i);
duke@0 2092 int opc = use->Opcode();
duke@0 2093 if (use->is_Load() || use->is_LoadStore()) {
duke@0 2094 return false;
duke@0 2095 }
duke@0 2096 }
duke@0 2097 return true;
duke@0 2098 }
duke@0 2099
duke@0 2100 //=============================================================================
duke@0 2101 //------------------------------Ideal------------------------------------------
duke@0 2102 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2103 // we can skip the AND operation. If the store is from a sign-extension
duke@0 2104 // (a left shift, then right shift) we can skip both.
duke@0 2105 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2106 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
duke@0 2107 if( progress != NULL ) return progress;
duke@0 2108
duke@0 2109 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
duke@0 2110 if( progress != NULL ) return progress;
duke@0 2111
duke@0 2112 // Finally check the default case
duke@0 2113 return StoreNode::Ideal(phase, can_reshape);
duke@0 2114 }
duke@0 2115
duke@0 2116 //=============================================================================
duke@0 2117 //------------------------------Ideal------------------------------------------
duke@0 2118 // If the store is from an AND mask that leaves the low bits untouched, then
duke@0 2119 // we can skip the AND operation
duke@0 2120 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2121 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
duke@0 2122 if( progress != NULL ) return progress;
duke@0 2123
duke@0 2124 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
duke@0 2125 if( progress != NULL ) return progress;
duke@0 2126
duke@0 2127 // Finally check the default case
duke@0 2128 return StoreNode::Ideal(phase, can_reshape);
duke@0 2129 }
duke@0 2130
duke@0 2131 //=============================================================================
duke@0 2132 //------------------------------Identity---------------------------------------
duke@0 2133 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
duke@0 2134 // No need to card mark when storing a null ptr
duke@0 2135 Node* my_store = in(MemNode::OopStore);
duke@0 2136 if (my_store->is_Store()) {
duke@0 2137 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
duke@0 2138 if( t1 == TypePtr::NULL_PTR ) {
duke@0 2139 return in(MemNode::Memory);
duke@0 2140 }
duke@0 2141 }
duke@0 2142 return this;
duke@0 2143 }
duke@0 2144
duke@0 2145 //------------------------------Value-----------------------------------------
duke@0 2146 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
kvn@43 2147 // Either input is TOP ==> the result is TOP
kvn@43 2148 const Type *t = phase->type( in(MemNode::Memory) );
kvn@43 2149 if( t == Type::TOP ) return Type::TOP;
kvn@43 2150 t = phase->type( in(MemNode::Address) );
kvn@43 2151 if( t == Type::TOP ) return Type::TOP;
kvn@43 2152 t = phase->type( in(MemNode::ValueIn) );
kvn@43 2153 if( t == Type::TOP ) return Type::TOP;
duke@0 2154 // If extra input is TOP ==> the result is TOP
kvn@43 2155 t = phase->type( in(MemNode::OopStore) );
kvn@43 2156 if( t == Type::TOP ) return Type::TOP;
duke@0 2157
duke@0 2158 return StoreNode::Value( phase );
duke@0 2159 }
duke@0 2160
duke@0 2161
duke@0 2162 //=============================================================================
duke@0 2163 //----------------------------------SCMemProjNode------------------------------
duke@0 2164 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
duke@0 2165 {
duke@0 2166 return bottom_type();
duke@0 2167 }
duke@0 2168
duke@0 2169 //=============================================================================
duke@0 2170 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
duke@0 2171 init_req(MemNode::Control, c );
duke@0 2172 init_req(MemNode::Memory , mem);
duke@0 2173 init_req(MemNode::Address, adr);
duke@0 2174 init_req(MemNode::ValueIn, val);
duke@0 2175 init_req( ExpectedIn, ex );
duke@0 2176 init_class_id(Class_LoadStore);
duke@0 2177
duke@0 2178 }
duke@0 2179
duke@0 2180 //=============================================================================
duke@0 2181 //-------------------------------adr_type--------------------------------------
duke@0 2182 // Do we Match on this edge index or not? Do not match memory
duke@0 2183 const TypePtr* ClearArrayNode::adr_type() const {
duke@0 2184 Node *adr = in(3);
duke@0 2185 return MemNode::calculate_adr_type(adr->bottom_type());
duke@0 2186 }
duke@0 2187
duke@0 2188 //------------------------------match_edge-------------------------------------
duke@0 2189 // Do we Match on this edge index or not? Do not match memory
duke@0 2190 uint ClearArrayNode::match_edge(uint idx) const {
duke@0 2191 return idx > 1;
duke@0 2192 }
duke@0 2193
duke@0 2194 //------------------------------Identity---------------------------------------
duke@0 2195 // Clearing a zero length array does nothing
duke@0 2196 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
never@68 2197 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
duke@0 2198 }
duke@0 2199
duke@0 2200 //------------------------------Idealize---------------------------------------
duke@0 2201 // Clearing a short array is faster with stores
duke@0 2202 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2203 const int unit = BytesPerLong;
duke@0 2204 const TypeX* t = phase->type(in(2))->isa_intptr_t();
duke@0 2205 if (!t) return NULL;
duke@0 2206 if (!t->is_con()) return NULL;
duke@0 2207 intptr_t raw_count = t->get_con();
duke@0 2208 intptr_t size = raw_count;
duke@0 2209 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
duke@0 2210 // Clearing nothing uses the Identity call.
duke@0 2211 // Negative clears are possible on dead ClearArrays
duke@0 2212 // (see jck test stmt114.stmt11402.val).
duke@0 2213 if (size <= 0 || size % unit != 0) return NULL;
duke@0 2214 intptr_t count = size / unit;
duke@0 2215 // Length too long; use fast hardware clear
duke@0 2216 if (size > Matcher::init_array_short_size) return NULL;
duke@0 2217 Node *mem = in(1);
duke@0 2218 if( phase->type(mem)==Type::TOP ) return NULL;
duke@0 2219 Node *adr = in(3);
duke@0 2220 const Type* at = phase->type(adr);
duke@0 2221 if( at==Type::TOP ) return NULL;
duke@0 2222 const TypePtr* atp = at->isa_ptr();
duke@0 2223 // adjust atp to be the correct array element address type
duke@0 2224 if (atp == NULL) atp = TypePtr::BOTTOM;
duke@0 2225 else atp = atp->add_offset(Type::OffsetBot);
duke@0 2226 // Get base for derived pointer purposes
duke@0 2227 if( adr->Opcode() != Op_AddP ) Unimplemented();
duke@0 2228 Node *base = adr->in(1);
duke@0 2229
duke@0 2230 Node *zero = phase->makecon(TypeLong::ZERO);
duke@0 2231 Node *off = phase->MakeConX(BytesPerLong);
duke@0 2232 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2233 count--;
duke@0 2234 while( count-- ) {
duke@0 2235 mem = phase->transform(mem);
duke@0 2236 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
duke@0 2237 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
duke@0 2238 }
duke@0 2239 return mem;
duke@0 2240 }
duke@0 2241
duke@0 2242 //----------------------------clear_memory-------------------------------------
duke@0 2243 // Generate code to initialize object storage to zero.
duke@0 2244 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2245 intptr_t start_offset,
duke@0 2246 Node* end_offset,
duke@0 2247 PhaseGVN* phase) {
duke@0 2248 Compile* C = phase->C;
duke@0 2249 intptr_t offset = start_offset;
duke@0 2250
duke@0 2251 int unit = BytesPerLong;
duke@0 2252 if ((offset % unit) != 0) {
duke@0 2253 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
duke@0 2254 adr = phase->transform(adr);
duke@0 2255 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2256 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2257 mem = phase->transform(mem);
duke@0 2258 offset += BytesPerInt;
duke@0 2259 }
duke@0 2260 assert((offset % unit) == 0, "");
duke@0 2261
duke@0 2262 // Initialize the remaining stuff, if any, with a ClearArray.
duke@0 2263 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
duke@0 2264 }
duke@0 2265
duke@0 2266 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2267 Node* start_offset,
duke@0 2268 Node* end_offset,
duke@0 2269 PhaseGVN* phase) {
never@68 2270 if (start_offset == end_offset) {
never@68 2271 // nothing to do
never@68 2272 return mem;
never@68 2273 }
never@68 2274
duke@0 2275 Compile* C = phase->C;
duke@0 2276 int unit = BytesPerLong;
duke@0 2277 Node* zbase = start_offset;
duke@0 2278 Node* zend = end_offset;
duke@0 2279
duke@0 2280 // Scale to the unit required by the CPU:
duke@0 2281 if (!Matcher::init_array_count_is_in_bytes) {
duke@0 2282 Node* shift = phase->intcon(exact_log2(unit));
duke@0 2283 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
duke@0 2284 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
duke@0 2285 }
duke@0 2286
duke@0 2287 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
duke@0 2288 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
duke@0 2289
duke@0 2290 // Bulk clear double-words
duke@0 2291 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
duke@0 2292 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
duke@0 2293 return phase->transform(mem);
duke@0 2294 }
duke@0 2295
duke@0 2296 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
duke@0 2297 intptr_t start_offset,
duke@0 2298 intptr_t end_offset,
duke@0 2299 PhaseGVN* phase) {
never@68 2300 if (start_offset == end_offset) {
never@68 2301 // nothing to do
never@68 2302 return mem;
never@68 2303 }
never@68 2304
duke@0 2305 Compile* C = phase->C;
duke@0 2306 assert((end_offset % BytesPerInt) == 0, "odd end offset");
duke@0 2307 intptr_t done_offset = end_offset;
duke@0 2308 if ((done_offset % BytesPerLong) != 0) {
duke@0 2309 done_offset -= BytesPerInt;
duke@0 2310 }
duke@0 2311 if (done_offset > start_offset) {
duke@0 2312 mem = clear_memory(ctl, mem, dest,
duke@0 2313 start_offset, phase->MakeConX(done_offset), phase);
duke@0 2314 }
duke@0 2315 if (done_offset < end_offset) { // emit the final 32-bit store
duke@0 2316 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
duke@0 2317 adr = phase->transform(adr);
duke@0 2318 const TypePtr* atp = TypeRawPtr::BOTTOM;
coleenp@113 2319 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
duke@0 2320 mem = phase->transform(mem);
duke@0 2321 done_offset += BytesPerInt;
duke@0 2322 }
duke@0 2323 assert(done_offset == end_offset, "");
duke@0 2324 return mem;
duke@0 2325 }
duke@0 2326
duke@0 2327 //=============================================================================
duke@0 2328 // Do we match on this edge? No memory edges
duke@0 2329 uint StrCompNode::match_edge(uint idx) const {
duke@0 2330 return idx == 5 || idx == 6;
duke@0 2331 }
duke@0 2332
duke@0 2333 //------------------------------Ideal------------------------------------------
duke@0 2334 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2335 // control copies
duke@0 2336 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
duke@0 2337 return remove_dead_region(phase, can_reshape) ? this : NULL;
duke@0 2338 }
duke@0 2339
duke@0 2340
duke@0 2341 //=============================================================================
duke@0 2342 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
duke@0 2343 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
duke@0 2344 _adr_type(C->get_adr_type(alias_idx))
duke@0 2345 {
duke@0 2346 init_class_id(Class_MemBar);
duke@0 2347 Node* top = C->top();
duke@0 2348 init_req(TypeFunc::I_O,top);
duke@0 2349 init_req(TypeFunc::FramePtr,top);
duke@0 2350 init_req(TypeFunc::ReturnAdr,top);
duke@0 2351 if (precedent != NULL)
duke@0 2352 init_req(TypeFunc::Parms, precedent);
duke@0 2353 }
duke@0 2354
duke@0 2355 //------------------------------cmp--------------------------------------------
duke@0 2356 uint MemBarNode::hash() const { return NO_HASH; }
duke@0 2357 uint MemBarNode::cmp( const Node &n ) const {
duke@0 2358 return (&n == this); // Always fail except on self
duke@0 2359 }
duke@0 2360
duke@0 2361 //------------------------------make-------------------------------------------
duke@0 2362 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
duke@0 2363 int len = Precedent + (pn == NULL? 0: 1);
duke@0 2364 switch (opcode) {
duke@0 2365 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
duke@0 2366 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
duke@0 2367 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
duke@0 2368 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
duke@0 2369 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
duke@0 2370 default: ShouldNotReachHere(); return NULL;
duke@0 2371 }
duke@0 2372 }
duke@0 2373
duke@0 2374 //------------------------------Ideal------------------------------------------
duke@0 2375 // Return a node which is more "ideal" than the current node. Strip out
duke@0 2376 // control copies
duke@0 2377 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 2378 if (remove_dead_region(phase, can_reshape)) return this;
duke@0 2379 return NULL;
duke@0 2380 }
duke@0 2381
duke@0 2382 //------------------------------Value------------------------------------------
duke@0 2383 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
duke@0 2384 if( !in(0) ) return Type::TOP;
duke@0 2385 if( phase->type(in(0)) == Type::TOP )
duke@0 2386 return Type::TOP;
duke@0 2387 return TypeTuple::MEMBAR;
duke@0 2388 }
duke@0 2389
duke@0 2390 //------------------------------match------------------------------------------
duke@0 2391 // Construct projections for memory.
duke@0 2392 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
duke@0 2393 switch (proj->_con) {
duke@0 2394 case TypeFunc::Control:
duke@0 2395 case TypeFunc::Memory:
duke@0 2396 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
duke@0 2397 }
duke@0 2398 ShouldNotReachHere();
duke@0 2399 return NULL;
duke@0 2400 }
duke@0 2401
duke@0 2402 //===========================InitializeNode====================================
duke@0 2403 // SUMMARY:
duke@0 2404 // This node acts as a memory barrier on raw memory, after some raw stores.
duke@0 2405 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
duke@0 2406 // The Initialize can 'capture' suitably constrained stores as raw inits.
duke@0 2407 // It can coalesce related raw stores into larger units (called 'tiles').
duke@0 2408 // It can avoid zeroing new storage for memory units which have raw inits.
duke@0 2409 // At macro-expansion, it is marked 'complete', and does not optimize further.
duke@0 2410 //
duke@0 2411 // EXAMPLE:
duke@0 2412 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
duke@0 2413 // ctl = incoming control; mem* = incoming memory
duke@0 2414 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
duke@0 2415 // First allocate uninitialized memory and fill in the header:
duke@0 2416 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
duke@0 2417 // ctl := alloc.Control; mem* := alloc.Memory*
duke@0 2418 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
duke@0 2419 // Then initialize to zero the non-header parts of the raw memory block:
duke@0 2420 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
duke@0 2421 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
duke@0 2422 // After the initialize node executes, the object is ready for service:
duke@0 2423 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
duke@0 2424 // Suppose its body is immediately initialized as {1,2}:
duke@0 2425 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2426 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2427 // mem.SLICE(#short[*]) := store2
duke@0 2428 //
duke@0 2429 // DETAILS:
duke@0 2430 // An InitializeNode collects and isolates object initialization after
duke@0 2431 // an AllocateNode and before the next possible safepoint. As a
duke@0 2432 // memory barrier (MemBarNode), it keeps critical stores from drifting
duke@0 2433 // down past any safepoint or any publication of the allocation.
duke@0 2434 // Before this barrier, a newly-allocated object may have uninitialized bits.
duke@0 2435 // After this barrier, it may be treated as a real oop, and GC is allowed.
duke@0 2436 //
duke@0 2437 // The semantics of the InitializeNode include an implicit zeroing of
duke@0 2438 // the new object from object header to the end of the object.
duke@0 2439 // (The object header and end are determined by the AllocateNode.)
duke@0 2440 //
duke@0 2441 // Certain stores may be added as direct inputs to the InitializeNode.
duke@0 2442 // These stores must update raw memory, and they must be to addresses
duke@0 2443 // derived from the raw address produced by AllocateNode, and with
duke@0 2444 // a constant offset. They must be ordered by increasing offset.
duke@0 2445 // The first one is at in(RawStores), the last at in(req()-1).
duke@0 2446 // Unlike most memory operations, they are not linked in a chain,
duke@0 2447 // but are displayed in parallel as users of the rawmem output of
duke@0 2448 // the allocation.
duke@0 2449 //
duke@0 2450 // (See comments in InitializeNode::capture_store, which continue
duke@0 2451 // the example given above.)
duke@0 2452 //
duke@0 2453 // When the associated Allocate is macro-expanded, the InitializeNode
duke@0 2454 // may be rewritten to optimize collected stores. A ClearArrayNode
duke@0 2455 // may also be created at that point to represent any required zeroing.
duke@0 2456 // The InitializeNode is then marked 'complete', prohibiting further
duke@0 2457 // capturing of nearby memory operations.
duke@0 2458 //
duke@0 2459 // During macro-expansion, all captured initializations which store
duke@0 2460 // constant values of 32 bits or smaller are coalesced (if advantagous)
duke@0 2461 // into larger 'tiles' 32 or 64 bits. This allows an object to be
duke@0 2462 // initialized in fewer memory operations. Memory words which are
duke@0 2463 // covered by neither tiles nor non-constant stores are pre-zeroed
duke@0 2464 // by explicit stores of zero. (The code shape happens to do all
duke@0 2465 // zeroing first, then all other stores, with both sequences occurring
duke@0 2466 // in order of ascending offsets.)
duke@0 2467 //
duke@0 2468 // Alternatively, code may be inserted between an AllocateNode and its
duke@0 2469 // InitializeNode, to perform arbitrary initialization of the new object.
duke@0 2470 // E.g., the object copying intrinsics insert complex data transfers here.
duke@0 2471 // The initialization must then be marked as 'complete' disable the
duke@0 2472 // built-in zeroing semantics and the collection of initializing stores.
duke@0 2473 //
duke@0 2474 // While an InitializeNode is incomplete, reads from the memory state
duke@0 2475 // produced by it are optimizable if they match the control edge and
duke@0 2476 // new oop address associated with the allocation/initialization.
duke@0 2477 // They return a stored value (if the offset matches) or else zero.
duke@0 2478 // A write to the memory state, if it matches control and address,
duke@0 2479 // and if it is to a constant offset, may be 'captured' by the
duke@0 2480 // InitializeNode. It is cloned as a raw memory operation and rewired
duke@0 2481 // inside the initialization, to the raw oop produced by the allocation.
duke@0 2482 // Operations on addresses which are provably distinct (e.g., to
duke@0 2483 // other AllocateNodes) are allowed to bypass the initialization.
duke@0 2484 //
duke@0 2485 // The effect of all this is to consolidate object initialization
duke@0 2486 // (both arrays and non-arrays, both piecewise and bulk) into a
duke@0 2487 // single location, where it can be optimized as a unit.
duke@0 2488 //
duke@0 2489 // Only stores with an offset less than TrackedInitializationLimit words
duke@0 2490 // will be considered for capture by an InitializeNode. This puts a
duke@0 2491 // reasonable limit on the complexity of optimized initializations.
duke@0 2492
duke@0 2493 //---------------------------InitializeNode------------------------------------
duke@0 2494 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
duke@0 2495 : _is_complete(false),
duke@0 2496 MemBarNode(C, adr_type, rawoop)
duke@0 2497 {
duke@0 2498 init_class_id(Class_Initialize);
duke@0 2499
duke@0 2500 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
duke@0 2501 assert(in(RawAddress) == rawoop, "proper init");
duke@0 2502 // Note: allocation() can be NULL, for secondary initialization barriers
duke@0 2503 }
duke@0 2504
duke@0 2505 // Since this node is not matched, it will be processed by the
duke@0 2506 // register allocator. Declare that there are no constraints
duke@0 2507 // on the allocation of the RawAddress edge.
duke@0 2508 const RegMask &InitializeNode::in_RegMask(uint idx) const {
duke@0 2509 // This edge should be set to top, by the set_complete. But be conservative.
duke@0 2510 if (idx == InitializeNode::RawAddress)
duke@0 2511 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
duke@0 2512 return RegMask::Empty;
duke@0 2513 }
duke@0 2514
duke@0 2515 Node* InitializeNode::memory(uint alias_idx) {
duke@0 2516 Node* mem = in(Memory);
duke@0 2517 if (mem->is_MergeMem()) {
duke@0 2518 return mem->as_MergeMem()->memory_at(alias_idx);
duke@0 2519 } else {
duke@0 2520 // incoming raw memory is not split
duke@0 2521 return mem;
duke@0 2522 }
duke@0 2523 }
duke@0 2524
duke@0 2525 bool InitializeNode::is_non_zero() {
duke@0 2526 if (is_complete()) return false;
duke@0 2527 remove_extra_zeroes();
duke@0 2528 return (req() > RawStores);
duke@0 2529 }
duke@0 2530
duke@0 2531 void InitializeNode::set_complete(PhaseGVN* phase) {
duke@0 2532 assert(!is_complete(), "caller responsibility");
duke@0 2533 _is_complete = true;
duke@0 2534
duke@0 2535 // After this node is complete, it contains a bunch of
duke@0 2536 // raw-memory initializations. There is no need for
duke@0 2537 // it to have anything to do with non-raw memory effects.
duke@0 2538 // Therefore, tell all non-raw users to re-optimize themselves,
duke@0 2539 // after skipping the memory effects of this initialization.
duke@0 2540 PhaseIterGVN* igvn = phase->is_IterGVN();
duke@0 2541 if (igvn) igvn->add_users_to_worklist(this);
duke@0 2542 }
duke@0 2543
duke@0 2544 // convenience function
duke@0 2545 // return false if the init contains any stores already
duke@0 2546 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
duke@0 2547 InitializeNode* init = initialization();
duke@0 2548 if (init == NULL || init->is_complete()) return false;
duke@0 2549 init->remove_extra_zeroes();
duke@0 2550 // for now, if this allocation has already collected any inits, bail:
duke@0 2551 if (init->is_non_zero()) return false;
duke@0 2552 init->set_complete(phase);
duke@0 2553 return true;
duke@0 2554 }
duke@0 2555
duke@0 2556 void InitializeNode::remove_extra_zeroes() {
duke@0 2557 if (req() == RawStores) return;
duke@0 2558 Node* zmem = zero_memory();
duke@0 2559 uint fill = RawStores;
duke@0 2560 for (uint i = fill; i < req(); i++) {
duke@0 2561 Node* n = in(i);
duke@0 2562 if (n->is_top() || n == zmem) continue; // skip
duke@0 2563 if (fill < i) set_req(fill, n); // compact
duke@0 2564 ++fill;
duke@0 2565 }
duke@0 2566 // delete any empty spaces created:
duke@0 2567 while (fill < req()) {
duke@0 2568 del_req(fill);
duke@0 2569 }
duke@0 2570 }
duke@0 2571
duke@0 2572 // Helper for remembering which stores go with which offsets.
duke@0 2573 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
duke@0 2574 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
duke@0 2575 intptr_t offset = -1;
duke@0 2576 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
duke@0 2577 phase, offset);
duke@0 2578 if (base == NULL) return -1; // something is dead,
duke@0 2579 if (offset < 0) return -1; // dead, dead
duke@0 2580 return offset;
duke@0 2581 }
duke@0 2582
duke@0 2583 // Helper for proving that an initialization expression is
duke@0 2584 // "simple enough" to be folded into an object initialization.
duke@0 2585 // Attempts to prove that a store's initial value 'n' can be captured
duke@0 2586 // within the initialization without creating a vicious cycle, such as:
duke@0 2587 // { Foo p = new Foo(); p.next = p; }
duke@0 2588 // True for constants and parameters and small combinations thereof.
duke@0 2589 bool InitializeNode::detect_init_independence(Node* n,
duke@0 2590 bool st_is_pinned,
duke@0 2591 int& count) {
duke@0 2592 if (n == NULL) return true; // (can this really happen?)
duke@0 2593 if (n->is_Proj()) n = n->in(0);
duke@0 2594 if (n == this) return false; // found a cycle
duke@0 2595 if (n->is_Con()) return true;
duke@0 2596 if (n->is_Start()) return true; // params, etc., are OK
duke@0 2597 if (n->is_Root()) return true; // even better
duke@0 2598
duke@0 2599 Node* ctl = n->in(0);
duke@0 2600 if (ctl != NULL && !ctl->is_top()) {
duke@0 2601 if (ctl->is_Proj()) ctl = ctl->in(0);
duke@0 2602 if (ctl == this) return false;
duke@0 2603
duke@0 2604 // If we already know that the enclosing memory op is pinned right after
duke@0 2605 // the init, then any control flow that the store has picked up
duke@0 2606 // must have preceded the init, or else be equal to the init.
duke@0 2607 // Even after loop optimizations (which might change control edges)
duke@0 2608 // a store is never pinned *before* the availability of its inputs.
kvn@119 2609 if (!MemNode::all_controls_dominate(n, this))
duke@0 2610 return false; // failed to prove a good control
duke@0 2611
duke@0 2612 }
duke@0 2613
duke@0 2614 // Check data edges for possible dependencies on 'this'.
duke@0 2615 if ((count += 1) > 20) return false; // complexity limit
duke@0 2616 for (uint i = 1; i < n->req(); i++) {
duke@0 2617 Node* m = n->in(i);
duke@0 2618 if (m == NULL || m == n || m->is_top()) continue;
duke@0 2619 uint first_i = n->find_edge(m);
duke@0 2620 if (i != first_i) continue; // process duplicate edge just once
duke@0 2621 if (!detect_init_independence(m, st_is_pinned, count)) {
duke@0 2622 return false;
duke@0 2623 }
duke@0 2624 }
duke@0 2625
duke@0 2626 return true;
duke@0 2627 }
duke@0 2628
duke@0 2629 // Here are all the checks a Store must pass before it can be moved into
duke@0 2630 // an initialization. Returns zero if a check fails.
duke@0 2631 // On success, returns the (constant) offset to which the store applies,
duke@0 2632 // within the initialized memory.
duke@0 2633 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
duke@0 2634 const int FAIL = 0;
duke@0 2635 if (st->req() != MemNode::ValueIn + 1)
duke@0 2636 return FAIL; // an inscrutable StoreNode (card mark?)
duke@0 2637 Node* ctl = st->in(MemNode::Control);
duke@0 2638 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
duke@0 2639 return FAIL; // must be unconditional after the initialization
duke@0 2640 Node* mem = st->in(MemNode::Memory);
duke@0 2641 if (!(mem->is_Proj() && mem->in(0) == this))
duke@0 2642 return FAIL; // must not be preceded by other stores
duke@0 2643 Node* adr = st->in(MemNode::Address);
duke@0 2644 intptr_t offset;
duke@0 2645 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
duke@0 2646 if (alloc == NULL)
duke@0 2647 return FAIL; // inscrutable address
duke@0 2648 if (alloc != allocation())
duke@0 2649 return FAIL; // wrong allocation! (store needs to float up)
duke@0 2650 Node* val = st->in(MemNode::ValueIn);
duke@0 2651 int complexity_count = 0;
duke@0 2652 if (!detect_init_independence(val, true, complexity_count))
duke@0 2653 return FAIL; // stored value must be 'simple enough'
duke@0 2654
duke@0 2655 return offset; // success
duke@0 2656 }
duke@0 2657
duke@0 2658 // Find the captured store in(i) which corresponds to the range
duke@0 2659 // [start..start+size) in the initialized object.
duke@0 2660 // If there is one, return its index i. If there isn't, return the
duke@0 2661 // negative of the index where it should be inserted.
duke@0 2662 // Return 0 if the queried range overlaps an initialization boundary
duke@0 2663 // or if dead code is encountered.
duke@0 2664 // If size_in_bytes is zero, do not bother with overlap checks.
duke@0 2665 int InitializeNode::captured_store_insertion_point(intptr_t start,
duke@0 2666 int size_in_bytes,
duke@0 2667 PhaseTransform* phase) {
duke@0 2668 const int FAIL = 0, MAX_STORE = BytesPerLong;
duke@0 2669
duke@0 2670 if (is_complete())
duke@0 2671 return FAIL; // arraycopy got here first; punt
duke@0 2672
duke@0 2673 assert(allocation() != NULL, "must be present");
duke@0 2674
duke@0 2675 // no negatives, no header fields:
coleenp@113 2676 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
duke@0 2677
duke@0 2678 // after a certain size, we bail out on tracking all the stores:
duke@0 2679 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2680 if (start >= ti_limit) return FAIL;
duke@0 2681
duke@0 2682 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
duke@0 2683 if (i >= limit) return -(int)i; // not found; here is where to put it
duke@0 2684
duke@0 2685 Node* st = in(i);
duke@0 2686 intptr_t st_off = get_store_offset(st, phase);
duke@0 2687 if (st_off < 0) {
duke@0 2688 if (st != zero_memory()) {
duke@0 2689 return FAIL; // bail out if there is dead garbage
duke@0 2690 }
duke@0 2691 } else if (st_off > start) {
duke@0 2692 // ...we are done, since stores are ordered
duke@0 2693 if (st_off < start + size_in_bytes) {
duke@0 2694 return FAIL; // the next store overlaps
duke@0 2695 }
duke@0 2696 return -(int)i; // not found; here is where to put it
duke@0 2697 } else if (st_off < start) {
duke@0 2698 if (size_in_bytes != 0 &&
duke@0 2699 start < st_off + MAX_STORE &&
duke@0 2700 start < st_off + st->as_Store()->memory_size()) {
duke@0 2701 return FAIL; // the previous store overlaps
duke@0 2702 }
duke@0 2703 } else {
duke@0 2704 if (size_in_bytes != 0 &&
duke@0 2705 st->as_Store()->memory_size() != size_in_bytes) {
duke@0 2706 return FAIL; // mismatched store size
duke@0 2707 }
duke@0 2708 return i;
duke@0 2709 }
duke@0 2710
duke@0 2711 ++i;
duke@0 2712 }
duke@0 2713 }
duke@0 2714
duke@0 2715 // Look for a captured store which initializes at the offset 'start'
duke@0 2716 // with the given size. If there is no such store, and no other
duke@0 2717 // initialization interferes, then return zero_memory (the memory
duke@0 2718 // projection of the AllocateNode).
duke@0 2719 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
duke@0 2720 PhaseTransform* phase) {
duke@0 2721 assert(stores_are_sane(phase), "");
duke@0 2722 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2723 if (i == 0) {
duke@0 2724 return NULL; // something is dead
duke@0 2725 } else if (i < 0) {
duke@0 2726 return zero_memory(); // just primordial zero bits here
duke@0 2727 } else {
duke@0 2728 Node* st = in(i); // here is the store at this position
duke@0 2729 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
duke@0 2730 return st;
duke@0 2731 }
duke@0 2732 }
duke@0 2733
duke@0 2734 // Create, as a raw pointer, an address within my new object at 'offset'.
duke@0 2735 Node* InitializeNode::make_raw_address(intptr_t offset,
duke@0 2736 PhaseTransform* phase) {
duke@0 2737 Node* addr = in(RawAddress);
duke@0 2738 if (offset != 0) {
duke@0 2739 Compile* C = phase->C;
duke@0 2740 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
duke@0 2741 phase->MakeConX(offset)) );
duke@0 2742 }
duke@0 2743 return addr;
duke@0 2744 }
duke@0 2745
duke@0 2746 // Clone the given store, converting it into a raw store
duke@0 2747 // initializing a field or element of my new object.
duke@0 2748 // Caller is responsible for retiring the original store,
duke@0 2749 // with subsume_node or the like.
duke@0 2750 //
duke@0 2751 // From the example above InitializeNode::InitializeNode,
duke@0 2752 // here are the old stores to be captured:
duke@0 2753 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
duke@0 2754 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
duke@0 2755 //
duke@0 2756 // Here is the changed code; note the extra edges on init:
duke@0 2757 // alloc = (Allocate ...)
duke@0 2758 // rawoop = alloc.RawAddress
duke@0 2759 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
duke@0 2760 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
duke@0 2761 // init = (Initialize alloc.Control alloc.Memory rawoop
duke@0 2762 // rawstore1 rawstore2)
duke@0 2763 //
duke@0 2764 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
duke@0 2765 PhaseTransform* phase) {
duke@0 2766 assert(stores_are_sane(phase), "");
duke@0 2767
duke@0 2768 if (start < 0) return NULL;
duke@0 2769 assert(can_capture_store(st, phase) == start, "sanity");
duke@0 2770
duke@0 2771 Compile* C = phase->C;
duke@0 2772 int size_in_bytes = st->memory_size();
duke@0 2773 int i = captured_store_insertion_point(start, size_in_bytes, phase);
duke@0 2774 if (i == 0) return NULL; // bail out
duke@0 2775 Node* prev_mem = NULL; // raw memory for the captured store
duke@0 2776 if (i > 0) {
duke@0 2777 prev_mem = in(i); // there is a pre-existing store under this one
duke@0 2778 set_req(i, C->top()); // temporarily disconnect it
duke@0 2779 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
duke@0 2780 } else {
duke@0 2781 i = -i; // no pre-existing store
duke@0 2782 prev_mem = zero_memory(); // a slice of the newly allocated object
duke@0 2783 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
duke@0 2784 set_req(--i, C->top()); // reuse this edge; it has been folded away
duke@0 2785 else
duke@0 2786 ins_req(i, C->top()); // build a new edge
duke@0 2787 }
duke@0 2788 Node* new_st = st->clone();
duke@0 2789 new_st->set_req(MemNode::Control, in(Control));
duke@0 2790 new_st->set_req(MemNode::Memory, prev_mem);
duke@0 2791 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
duke@0 2792 new_st = phase->transform(new_st);
duke@0 2793
duke@0 2794 // At this point, new_st might have swallowed a pre-existing store
duke@0 2795 // at the same offset, or perhaps new_st might have disappeared,
duke@0 2796 // if it redundantly stored the same value (or zero to fresh memory).
duke@0 2797
duke@0 2798 // In any case, wire it in:
duke@0 2799 set_req(i, new_st);
duke@0 2800
duke@0 2801 // The caller may now kill the old guy.
duke@0 2802 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
duke@0 2803 assert(check_st == new_st || check_st == NULL, "must be findable");
duke@0 2804 assert(!is_complete(), "");
duke@0 2805 return new_st;
duke@0 2806 }
duke@0 2807
duke@0 2808 static bool store_constant(jlong* tiles, int num_tiles,
duke@0 2809 intptr_t st_off, int st_size,
duke@0 2810 jlong con) {
duke@0 2811 if ((st_off & (st_size-1)) != 0)
duke@0 2812 return false; // strange store offset (assume size==2**N)
duke@0 2813 address addr = (address)tiles + st_off;
duke@0 2814 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
duke@0 2815 switch (st_size) {
duke@0 2816 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
duke@0 2817 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
duke@0 2818 case sizeof(jint): *(jint*) addr = (jint) con; break;
duke@0 2819 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
duke@0 2820 default: return false; // strange store size (detect size!=2**N here)
duke@0 2821 }
duke@0 2822 return true; // return success to caller
duke@0 2823 }
duke@0 2824
duke@0 2825 // Coalesce subword constants into int constants and possibly
duke@0 2826 // into long constants. The goal, if the CPU permits,
duke@0 2827 // is to initialize the object with a small number of 64-bit tiles.
duke@0 2828 // Also, convert floating-point constants to bit patterns.
duke@0 2829 // Non-constants are not relevant to this pass.
duke@0 2830 //
duke@0 2831 // In terms of the running example on InitializeNode::InitializeNode
duke@0 2832 // and InitializeNode::capture_store, here is the transformation
duke@0 2833 // of rawstore1 and rawstore2 into rawstore12:
duke@0 2834 // alloc = (Allocate ...)
duke@0 2835 // rawoop = alloc.RawAddress
duke@0 2836 // tile12 = 0x00010002
duke@0 2837 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
duke@0 2838 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
duke@0 2839 //
duke@0 2840 void
duke@0 2841 InitializeNode::coalesce_subword_stores(intptr_t header_size,
duke@0 2842 Node* size_in_bytes,
duke@0 2843 PhaseGVN* phase) {
duke@0 2844 Compile* C = phase->C;
duke@0 2845
duke@0 2846 assert(stores_are_sane(phase), "");
duke@0 2847 // Note: After this pass, they are not completely sane,
duke@0 2848 // since there may be some overlaps.
duke@0 2849
duke@0 2850 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
duke@0 2851
duke@0 2852 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
duke@0 2853 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
duke@0 2854 size_limit = MIN2(size_limit, ti_limit);
duke@0 2855 size_limit = align_size_up(size_limit, BytesPerLong);
duke@0 2856 int num_tiles = size_limit / BytesPerLong;
duke@0 2857
duke@0 2858 // allocate space for the tile map:
duke@0 2859 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
duke@0 2860 jlong tiles_buf[small_len];
duke@0 2861 Node* nodes_buf[small_len];
duke@0 2862 jlong inits_buf[small_len];
duke@0 2863 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
duke@0 2864 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 2865 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
duke@0 2866 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
duke@0 2867 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
duke@0 2868 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
duke@0 2869 // tiles: exact bitwise model of all primitive constants
duke@0 2870 // nodes: last constant-storing node subsumed into the tiles model
duke@0 2871 // inits: which bytes (in each tile) are touched by any initializations
duke@0 2872
duke@0 2873 //// Pass A: Fill in the tile model with any relevant stores.
duke@0 2874
duke@0 2875 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
duke@0 2876 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
duke@0 2877 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
duke@0 2878 Node* zmem = zero_memory(); // initially zero memory state
duke@0 2879 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 2880 Node* st = in(i);
duke@0 2881 intptr_t st_off = get_store_offset(st, phase);
duke@0 2882
duke@0 2883 // Figure out the store's offset and constant value:
duke@0 2884 if (st_off < header_size) continue; //skip (ignore header)
duke@0 2885 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
duke@0 2886 int st_size = st->as_Store()->memory_size();
duke@0 2887 if (st_off + st_size > size_limit) break;
duke@0 2888
duke@0 2889 // Record which bytes are touched, whether by constant or not.
duke@0 2890 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
duke@0 2891 continue; // skip (strange store size)
duke@0 2892
duke@0 2893 const Type* val = phase->type(st->in(MemNode::ValueIn));
duke@0 2894 if (!val->singleton()) continue; //skip (non-con store)
duke@0 2895 BasicType type = val->basic_type();
duke@0 2896
duke@0 2897 jlong con = 0;
duke@0 2898 switch (type) {
duke@0 2899 case T_INT: con = val->is_int()->get_con(); break;
duke@0 2900 case T_LONG: con = val->is_long()->get_con(); break;
duke@0 2901 case T_FLOAT: con = jint_cast(val->getf()); break;
duke@0 2902 case T_DOUBLE: con = jlong_cast(val->getd()); break;
duke@0 2903 default: continue; //skip (odd store type)
duke@0 2904 }
duke@0 2905
duke@0 2906 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
duke@0 2907 st->Opcode() == Op_StoreL) {
duke@0 2908 continue; // This StoreL is already optimal.
duke@0 2909 }
duke@0 2910
duke@0 2911 // Store down the constant.
duke@0 2912 store_constant(tiles, num_tiles, st_off, st_size, con);
duke@0 2913
duke@0 2914 intptr_t j = st_off >> LogBytesPerLong;
duke@0 2915
duke@0 2916 if (type == T_INT && st_size == BytesPerInt
duke@0 2917 && (st_off & BytesPerInt) == BytesPerInt) {
duke@0 2918 jlong lcon = tiles[j];
duke@0 2919 if (!Matcher::isSimpleConstant64(lcon) &&
duke@0 2920 st->Opcode() == Op_StoreI) {
duke@0 2921 // This StoreI is already optimal by itself.
duke@0 2922 jint* intcon = (jint*) &tiles[j];
duke@0 2923 intcon[1] = 0; // undo the store_constant()
duke@0 2924
duke@0 2925 // If the previous store is also optimal by itself, back up and
duke@0 2926 // undo the action of the previous loop iteration... if we can.
duke@0 2927 // But if we can't, just let the previous half take care of itself.
duke@0 2928 st = nodes[j];
duke@0 2929 st_off -= BytesPerInt;
duke@0 2930 con = intcon[0];
duke@0 2931 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
duke@0 2932 assert(st_off >= header_size, "still ignoring header");
duke@0 2933 assert(get_store_offset(st, phase) == st_off, "must be");
duke@0 2934 assert(in(i-1) == zmem, "must be");
duke@0 2935 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
duke@0 2936 assert(con == tcon->is_int()->get_con(), "must be");
duke@0 2937 // Undo the effects of the previous loop trip, which swallowed st:
duke@0 2938 intcon[0] = 0; // undo store_constant()
duke@0 2939 set_req(i-1, st); // undo set_req(i, zmem)
duke@0 2940 nodes[j] = NULL; // undo nodes[j] = st
duke@0 2941 --old_subword; // undo ++old_subword
duke@0 2942 }
duke@0 2943 continue; // This StoreI is already optimal.
duke@0 2944 }
duke@0 2945 }
duke@0 2946
duke@0 2947 // This store is not needed.
duke@0 2948 set_req(i, zmem);
duke@0 2949 nodes[j] = st; // record for the moment
duke@0 2950 if (st_size < BytesPerLong) // something has changed
duke@0 2951 ++old_subword; // includes int/float, but who's counting...
duke@0 2952 else ++old_long;
duke@0 2953 }
duke@0 2954
duke@0 2955 if ((old_subword + old_long) == 0)
duke@0 2956 return; // nothing more to do
duke@0 2957
duke@0 2958 //// Pass B: Convert any non-zero tiles into optimal constant stores.
duke@0 2959 // Be sure to insert them before overlapping non-constant stores.
duke@0 2960 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
duke@0 2961 for (int j = 0; j < num_tiles; j++) {
duke@0 2962 jlong con = tiles[j];
duke@0 2963 jlong init = inits[j];
duke@0 2964 if (con == 0) continue;
duke@0 2965 jint con0, con1; // split the constant, address-wise
duke@0 2966 jint init0, init1; // split the init map, address-wise
duke@0 2967 { union { jlong con; jint intcon[2]; } u;
duke@0 2968 u.con = con;
duke@0 2969 con0 = u.intcon[0];
duke@0 2970 con1 = u.intcon[1];
duke@0 2971 u.con = init;
duke@0 2972 init0 = u.intcon[0];
duke@0 2973 init1 = u.intcon[1];
duke@0 2974 }
duke@0 2975
duke@0 2976 Node* old = nodes[j];
duke@0 2977 assert(old != NULL, "need the prior store");
duke@0 2978 intptr_t offset = (j * BytesPerLong);
duke@0 2979
duke@0 2980 bool split = !Matcher::isSimpleConstant64(con);
duke@0 2981
duke@0 2982 if (offset < header_size) {
duke@0 2983 assert(offset + BytesPerInt >= header_size, "second int counts");
duke@0 2984 assert(*(jint*)&tiles[j] == 0, "junk in header");
duke@0 2985 split = true; // only the second word counts
duke@0 2986 // Example: int a[] = { 42 ... }
duke@0 2987 } else if (con0 == 0 && init0 == -1) {
duke@0 2988 split = true; // first word is covered by full inits
duke@0 2989 // Example: int a[] = { ... foo(), 42 ... }
duke@0 2990 } else if (con1 == 0 && init1 == -1) {
duke@0 2991 split = true; // second word is covered by full inits
duke@0 2992 // Example: int a[] = { ... 42, foo() ... }
duke@0 2993 }
duke@0 2994
duke@0 2995 // Here's a case where init0 is neither 0 nor -1:
duke@0 2996 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
duke@0 2997 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
duke@0 2998 // In this case the tile is not split; it is (jlong)42.
duke@0 2999 // The big tile is stored down, and then the foo() value is inserted.
duke@0 3000 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
duke@0 3001
duke@0 3002 Node* ctl = old->in(MemNode::Control);
duke@0 3003 Node* adr = make_raw_address(offset, phase);
duke@0 3004 const TypePtr* atp = TypeRawPtr::BOTTOM;
duke@0 3005
duke@0 3006 // One or two coalesced stores to plop down.
duke@0 3007 Node* st[2];
duke@0 3008 intptr_t off[2];
duke@0 3009 int nst = 0;
duke@0 3010 if (!split) {
duke@0 3011 ++new_long;
duke@0 3012 off[nst] = offset;
coleenp@113 3013 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3014 phase->longcon(con), T_LONG);
duke@0 3015 } else {
duke@0 3016 // Omit either if it is a zero.
duke@0 3017 if (con0 != 0) {
duke@0 3018 ++new_int;
duke@0 3019 off[nst] = offset;
coleenp@113 3020 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3021 phase->intcon(con0), T_INT);
duke@0 3022 }
duke@0 3023 if (con1 != 0) {
duke@0 3024 ++new_int;
duke@0 3025 offset += BytesPerInt;
duke@0 3026 adr = make_raw_address(offset, phase);
duke@0 3027 off[nst] = offset;
coleenp@113 3028 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
duke@0 3029 phase->intcon(con1), T_INT);
duke@0 3030 }
duke@0 3031 }
duke@0 3032
duke@0 3033 // Insert second store first, then the first before the second.
duke@0 3034 // Insert each one just before any overlapping non-constant stores.
duke@0 3035 while (nst > 0) {
duke@0 3036 Node* st1 = st[--nst];
duke@0 3037 C->copy_node_notes_to(st1, old);
duke@0 3038 st1 = phase->transform(st1);
duke@0 3039 offset = off[nst];
duke@0 3040 assert(offset >= header_size, "do not smash header");
duke@0 3041 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
duke@0 3042 guarantee(ins_idx != 0, "must re-insert constant store");
duke@0 3043 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
duke@0 3044 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
duke@0 3045 set_req(--ins_idx, st1);
duke@0 3046 else
duke@0 3047 ins_req(ins_idx, st1);
duke@0 3048 }
duke@0 3049 }
duke@0 3050
duke@0 3051 if (PrintCompilation && WizardMode)
duke@0 3052 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
duke@0 3053 old_subword, old_long, new_int, new_long);
duke@0 3054 if (C->log() != NULL)
duke@0 3055 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
duke@0 3056 old_subword, old_long, new_int, new_long);
duke@0 3057
duke@0 3058 // Clean up any remaining occurrences of zmem:
duke@0 3059 remove_extra_zeroes();
duke@0 3060 }
duke@0 3061
duke@0 3062 // Explore forward from in(start) to find the first fully initialized
duke@0 3063 // word, and return its offset. Skip groups of subword stores which
duke@0 3064 // together initialize full words. If in(start) is itself part of a
duke@0 3065 // fully initialized word, return the offset of in(start). If there
duke@0 3066 // are no following full-word stores, or if something is fishy, return
duke@0 3067 // a negative value.
duke@0 3068 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
duke@0 3069 int int_map = 0;
duke@0 3070 intptr_t int_map_off = 0;
duke@0 3071 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
duke@0 3072
duke@0 3073 for (uint i = start, limit = req(); i < limit; i++) {
duke@0 3074 Node* st = in(i);
duke@0 3075
duke@0 3076 intptr_t st_off = get_store_offset(st, phase);
duke@0 3077 if (st_off < 0) break; // return conservative answer
duke@0 3078
duke@0 3079 int st_size = st->as_Store()->memory_size();
duke@0 3080 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
duke@0 3081 return st_off; // we found a complete word init
duke@0 3082 }
duke@0 3083
duke@0 3084 // update the map:
duke@0 3085
duke@0 3086 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
duke@0 3087 if (this_int_off != int_map_off) {
duke@0 3088 // reset the map:
duke@0 3089 int_map = 0;
duke@0 3090 int_map_off = this_int_off;
duke@0 3091 }
duke@0 3092
duke@0 3093 int subword_off = st_off - this_int_off;
duke@0 3094 int_map |= right_n_bits(st_size) << subword_off;
duke@0 3095 if ((int_map & FULL_MAP) == FULL_MAP) {
duke@0 3096 return this_int_off; // we found a complete word init
duke@0 3097 }
duke@0 3098
duke@0 3099 // Did this store hit or cross the word boundary?
duke@0 3100 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
duke@0 3101 if (next_int_off == this_int_off + BytesPerInt) {
duke@0 3102 // We passed the current int, without fully initializing it.
duke@0 3103 int_map_off = next_int_off;
duke@0 3104 int_map >>= BytesPerInt;
duke@0 3105 } else if (next_int_off > this_int_off + BytesPerInt) {
duke@0 3106 // We passed the current and next int.
duke@0 3107 return this_int_off + BytesPerInt;
duke@0 3108 }
duke@0 3109 }
duke@0 3110
duke@0 3111 return -1;
duke@0 3112 }
duke@0 3113
duke@0 3114
duke@0 3115 // Called when the associated AllocateNode is expanded into CFG.
duke@0 3116 // At this point, we may perform additional optimizations.
duke@0 3117 // Linearize the stores by ascending offset, to make memory
duke@0 3118 // activity as coherent as possible.
duke@0 3119 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
duke@0 3120 intptr_t header_size,
duke@0 3121 Node* size_in_bytes,
duke@0 3122 PhaseGVN* phase) {
duke@0 3123 assert(!is_complete(), "not already complete");
duke@0 3124 assert(stores_are_sane(phase), "");
duke@0 3125 assert(allocation() != NULL, "must be present");
duke@0 3126
duke@0 3127 remove_extra_zeroes();
duke@0 3128
duke@0 3129 if (ReduceFieldZeroing || ReduceBulkZeroing)
duke@0 3130 // reduce instruction count for common initialization patterns
duke@0 3131 coalesce_subword_stores(header_size, size_in_bytes, phase);
duke@0 3132
duke@0 3133 Node* zmem = zero_memory(); // initially zero memory state
duke@0 3134 Node* inits = zmem; // accumulating a linearized chain of inits
duke@0 3135 #ifdef ASSERT
coleenp@113 3136 intptr_t first_offset = allocation()->minimum_header_size();
coleenp@113 3137 intptr_t last_init_off = first_offset; // previous init offset
coleenp@113 3138 intptr_t last_init_end = first_offset; // previous init offset+size
coleenp@113 3139 intptr_t last_tile_end = first_offset; // previous tile offset+size
duke@0 3140 #endif
duke@0 3141 intptr_t zeroes_done = header_size;
duke@0 3142
duke@0 3143 bool do_zeroing = true; // we might give up if inits are very sparse
duke@0 3144 int big_init_gaps = 0; // how many large gaps have we seen?
duke@0 3145
duke@0 3146 if (ZeroTLAB) do_zeroing = false;
duke@0 3147 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
duke@0 3148
duke@0 3149 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
duke@0 3150 Node* st = in(i);
duke@0 3151 intptr_t st_off = get_store_offset(st, phase);
duke@0 3152 if (st_off < 0)
duke@0 3153 break; // unknown junk in the inits
duke@0 3154 if (st->in(MemNode::Memory) != zmem)
duke@0 3155 break; // complicated store chains somehow in list
duke@0 3156
duke@0 3157 int st_size = st->as_Store()->memory_size();
duke@0 3158 intptr_t next_init_off = st_off + st_size;
duke@0 3159
duke@0 3160 if (do_zeroing && zeroes_done < next_init_off) {
duke@0 3161 // See if this store needs a zero before it or under it.
duke@0 3162 intptr_t zeroes_needed = st_off;
duke@0 3163
duke@0 3164 if (st_size < BytesPerInt) {
duke@0 3165 // Look for subword stores which only partially initialize words.
duke@0 3166 // If we find some, we must lay down some word-level zeroes first,
duke@0 3167 // underneath the subword stores.
duke@0 3168 //
duke@0 3169 // Examples:
duke@0 3170 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
duke@0 3171 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
duke@0 3172 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
duke@0 3173 //
duke@0 3174 // Note: coalesce_subword_stores may have already done this,
duke@0 3175 // if it was prompted by constant non-zero subword initializers.
duke@0 3176 // But this case can still arise with non-constant stores.
duke@0 3177
duke@0 3178 intptr_t next_full_store = find_next_fullword_store(i, phase);
duke@0 3179
duke@0 3180 // In the examples above:
duke@0 3181 // in(i) p q r s x y z
duke@0 3182 // st_off 12 13 14 15 12 13 14
duke@0 3183 // st_size 1 1 1 1 1 1 1
duke@0 3184 // next_full_s. 12 16 16 16 16 16 16
duke@0 3185 // z's_done 12 16 16 16 12 16 12
duke@0 3186 // z's_needed 12 16 16 16 16 16 16
duke@0 3187 // zsize 0 0 0 0 4 0 4
duke@0 3188 if (next_full_store < 0) {
duke@0 3189 // Conservative tack: Zero to end of current word.
duke@0 3190 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
duke@0 3191 } else {
duke@0 3192 // Zero to beginning of next fully initialized word.
duke@0 3193 // Or, don't zero at all, if we are already in that word.
duke@0 3194 assert(next_full_store >= zeroes_needed, "must go forward");
duke@0 3195 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
duke@0 3196 zeroes_needed = next_full_store;
duke@0 3197 }
duke@0 3198 }
duke@0 3199
duke@0 3200 if (zeroes_needed > zeroes_done) {
duke@0 3201 intptr_t zsize = zeroes_needed - zeroes_done;
duke@0 3202 // Do some incremental zeroing on rawmem, in parallel with inits.
duke@0 3203 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3204 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3205 zeroes_done, zeroes_needed,
duke@0 3206 phase);
duke@0 3207 zeroes_done = zeroes_needed;
duke@0 3208 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
duke@0 3209 do_zeroing = false; // leave the hole, next time
duke@0 3210 }
duke@0 3211 }
duke@0 3212
duke@0 3213 // Collect the store and move on:
duke@0 3214 st->set_req(MemNode::Memory, inits);
duke@0 3215 inits = st; // put it on the linearized chain
duke@0 3216 set_req(i, zmem); // unhook from previous position
duke@0 3217
duke@0 3218 if (zeroes_done == st_off)
duke@0 3219 zeroes_done = next_init_off;
duke@0 3220
duke@0 3221 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
duke@0 3222
duke@0 3223 #ifdef ASSERT
duke@0 3224 // Various order invariants. Weaker than stores_are_sane because
duke@0 3225 // a large constant tile can be filled in by smaller non-constant stores.
duke@0 3226 assert(st_off >= last_init_off, "inits do not reverse");
duke@0 3227 last_init_off = st_off;
duke@0 3228 const Type* val = NULL;
duke@0 3229 if (st_size >= BytesPerInt &&
duke@0 3230 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
duke@0 3231 (int)val->basic_type() < (int)T_OBJECT) {
duke@0 3232 assert(st_off >= last_tile_end, "tiles do not overlap");
duke@0 3233 assert(st_off >= last_init_end, "tiles do not overwrite inits");
duke@0 3234 last_tile_end = MAX2(last_tile_end, next_init_off);
duke@0 3235 } else {
duke@0 3236 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
duke@0 3237 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
duke@0 3238 assert(st_off >= last_init_end, "inits do not overlap");
duke@0 3239 last_init_end = next_init_off; // it's a non-tile
duke@0 3240 }
duke@0 3241 #endif //ASSERT
duke@0 3242 }
duke@0 3243
duke@0 3244 remove_extra_zeroes(); // clear out all the zmems left over
duke@0 3245 add_req(inits);
duke@0 3246
duke@0 3247 if (!ZeroTLAB) {
duke@0 3248 // If anything remains to be zeroed, zero it all now.
duke@0 3249 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
duke@0 3250 // if it is the last unused 4 bytes of an instance, forget about it
duke@0 3251 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
duke@0 3252 if (zeroes_done + BytesPerLong >= size_limit) {
duke@0 3253 assert(allocation() != NULL, "");
duke@0 3254 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
duke@0 3255 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
duke@0 3256 if (zeroes_done == k->layout_helper())
duke@0 3257 zeroes_done = size_limit;
duke@0 3258 }
duke@0 3259 if (zeroes_done < size_limit) {
duke@0 3260 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
duke@0 3261 zeroes_done, size_in_bytes, phase);
duke@0 3262 }
duke@0 3263 }
duke@0 3264
duke@0 3265 set_complete(phase);
duke@0 3266 return rawmem;
duke@0 3267 }
duke@0 3268
duke@0 3269
duke@0 3270 #ifdef ASSERT
duke@0 3271 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
duke@0 3272 if (is_complete())
duke@0 3273 return true; // stores could be anything at this point
coleenp@113 3274 assert(allocation() != NULL, "must be present");
coleenp@113 3275 intptr_t last_off = allocation()->minimum_header_size();
duke@0 3276 for (uint i = InitializeNode::RawStores; i < req(); i++) {
duke@0 3277 Node* st = in(i);
duke@0 3278 intptr_t st_off = get_store_offset(st, phase);
duke@0 3279 if (st_off < 0) continue; // ignore dead garbage
duke@0 3280 if (last_off > st_off) {
duke@0 3281 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
duke@0 3282 this->dump(2);
duke@0 3283 assert(false, "ascending store offsets");
duke@0 3284 return false;
duke@0 3285 }
duke@0 3286 last_off = st_off + st->as_Store()->memory_size();
duke@0 3287 }
duke@0 3288 return true;
duke@0 3289 }
duke@0 3290 #endif //ASSERT
duke@0 3291
duke@0 3292
duke@0 3293
duke@0 3294
duke@0 3295 //============================MergeMemNode=====================================
duke@0 3296 //
duke@0 3297 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
duke@0 3298 // contributing store or call operations. Each contributor provides the memory
duke@0 3299 // state for a particular "alias type" (see Compile::alias_type). For example,
duke@0 3300 // if a MergeMem has an input X for alias category #6, then any memory reference
duke@0 3301 // to alias category #6 may use X as its memory state input, as an exact equivalent
duke@0 3302 // to using the MergeMem as a whole.
duke@0 3303 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
duke@0 3304 //
duke@0 3305 // (Here, the <N> notation gives the index of the relevant adr_type.)
duke@0 3306 //
duke@0 3307 // In one special case (and more cases in the future), alias categories overlap.
duke@0 3308 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
duke@0 3309 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
duke@0 3310 // it is exactly equivalent to that state W:
duke@0 3311 // MergeMem(<Bot>: W) <==> W
duke@0 3312 //
duke@0 3313 // Usually, the merge has more than one input. In that case, where inputs
duke@0 3314 // overlap (i.e., one is Bot), the narrower alias type determines the memory
duke@0 3315 // state for that type, and the wider alias type (Bot) fills in everywhere else:
duke@0 3316 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
duke@0 3317 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
duke@0 3318 //
duke@0 3319 // A merge can take a "wide" memory state as one of its narrow inputs.
duke@0 3320 // This simply means that the merge observes out only the relevant parts of
duke@0 3321 // the wide input. That is, wide memory states arriving at narrow merge inputs
duke@0 3322 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
duke@0 3323 //
duke@0 3324 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
duke@0 3325 // and that memory slices "leak through":
duke@0 3326 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
duke@0 3327 //
duke@0 3328 // But, in such a cascade, repeated memory slices can "block the leak":
duke@0 3329 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
duke@0 3330 //
duke@0 3331 // In the last example, Y is not part of the combined memory state of the
duke@0 3332 // outermost MergeMem. The system must, of course, prevent unschedulable
duke@0 3333 // memory states from arising, so you can be sure that the state Y is somehow
duke@0 3334 // a precursor to state Y'.
duke@0 3335 //
duke@0 3336 //
duke@0 3337 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
duke@0 3338 // of each MergeMemNode array are exactly the numerical alias indexes, including
duke@0 3339 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
duke@0 3340 // Compile::alias_type (and kin) produce and manage these indexes.
duke@0 3341 //
duke@0 3342 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
duke@0 3343 // (Note that this provides quick access to the top node inside MergeMem methods,
duke@0 3344 // without the need to reach out via TLS to Compile::current.)
duke@0 3345 //
duke@0 3346 // As a consequence of what was just described, a MergeMem that represents a full
duke@0 3347 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
duke@0 3348 // containing all alias categories.
duke@0 3349 //
duke@0 3350 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
duke@0 3351 //
duke@0 3352 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
duke@0 3353 // a memory state for the alias type <N>, or else the top node, meaning that
duke@0 3354 // there is no particular input for that alias type. Note that the length of
duke@0 3355 // a MergeMem is variable, and may be extended at any time to accommodate new
duke@0 3356 // memory states at larger alias indexes. When merges grow, they are of course
duke@0 3357 // filled with "top" in the unused in() positions.
duke@0 3358 //
duke@0 3359 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
duke@0 3360 // (Top was chosen because it works smoothly with passes like GCM.)
duke@0 3361 //
duke@0 3362 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
duke@0 3363 // the type of random VM bits like TLS references.) Since it is always the
duke@0 3364 // first non-Bot memory slice, some low-level loops use it to initialize an
duke@0 3365 // index variable: for (i = AliasIdxRaw; i < req(); i++).
duke@0 3366 //
duke@0 3367 //
duke@0 3368 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
duke@0 3369 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
duke@0 3370 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
duke@0 3371 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
duke@0 3372 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
duke@0 3373 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
duke@0 3374 //
duke@0 3375 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
duke@0 3376 // really that different from the other memory inputs. An abbreviation called
duke@0 3377 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
duke@0 3378 //
duke@0 3379 //
duke@0 3380 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
duke@0 3381 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
duke@0 3382 // that "emerges though" the base memory will be marked as excluding the alias types
duke@0 3383 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
duke@0 3384 //
duke@0 3385 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
duke@0 3386 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
duke@0 3387 //
duke@0 3388 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
duke@0 3389 // (It is currently unimplemented.) As you can see, the resulting merge is
duke@0 3390 // actually a disjoint union of memory states, rather than an overlay.
duke@0 3391 //
duke@0 3392
duke@0 3393 //------------------------------MergeMemNode-----------------------------------
duke@0 3394 Node* MergeMemNode::make_empty_memory() {
duke@0 3395 Node* empty_memory = (Node*) Compile::current()->top();
duke@0 3396 assert(empty_memory->is_top(), "correct sentinel identity");
duke@0 3397 return empty_memory;
duke@0 3398 }
duke@0 3399
duke@0 3400 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
duke@0 3401 init_class_id(Class_MergeMem);
duke@0 3402 // all inputs are nullified in Node::Node(int)
duke@0 3403 // set_input(0, NULL); // no control input
duke@0 3404
duke@0 3405 // Initialize the edges uniformly to top, for starters.
duke@0 3406 Node* empty_mem = make_empty_memory();
duke@0 3407 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
duke@0 3408 init_req(i,empty_mem);
duke@0 3409 }
duke@0 3410 assert(empty_memory() == empty_mem, "");
duke@0 3411
duke@0 3412 if( new_base != NULL && new_base->is_MergeMem() ) {
duke@0 3413 MergeMemNode* mdef = new_base->as_MergeMem();
duke@0 3414 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
duke@0 3415 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
duke@0 3416 mms.set_memory(mms.memory2());
duke@0 3417 }
duke@0 3418 assert(base_memory() == mdef->base_memory(), "");
duke@0 3419 } else {
duke@0 3420 set_base_memory(new_base);
duke@0 3421 }
duke@0 3422 }
duke@0 3423
duke@0 3424 // Make a new, untransformed MergeMem with the same base as 'mem'.
duke@0 3425 // If mem is itself a MergeMem, populate the result with the same edges.
duke@0 3426 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
duke@0 3427 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
duke@0 3428 }
duke@0 3429
duke@0 3430 //------------------------------cmp--------------------------------------------
duke@0 3431 uint MergeMemNode::hash() const { return NO_HASH; }
duke@0 3432 uint MergeMemNode::cmp( const Node &n ) const {
duke@0 3433 return (&n == this); // Always fail except on self
duke@0 3434 }
duke@0 3435
duke@0 3436 //------------------------------Identity---------------------------------------
duke@0 3437 Node* MergeMemNode::Identity(PhaseTransform *phase) {
duke@0 3438 // Identity if this merge point does not record any interesting memory
duke@0 3439 // disambiguations.
duke@0 3440 Node* base_mem = base_memory();
duke@0 3441 Node* empty_mem = empty_memory();
duke@0 3442 if (base_mem != empty_mem) { // Memory path is not dead?
duke@0 3443 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3444 Node* mem = in(i);
duke@0 3445 if (mem != empty_mem && mem != base_mem) {
duke@0 3446 return this; // Many memory splits; no change
duke@0 3447 }
duke@0 3448 }
duke@0 3449 }
duke@0 3450 return base_mem; // No memory splits; ID on the one true input
duke@0 3451 }
duke@0 3452
duke@0 3453 //------------------------------Ideal------------------------------------------
duke@0 3454 // This method is invoked recursively on chains of MergeMem nodes
duke@0 3455 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
duke@0 3456 // Remove chain'd MergeMems
duke@0 3457 //
duke@0 3458 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
duke@0 3459 // relative to the "in(Bot)". Since we are patching both at the same time,
duke@0 3460 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
duke@0 3461 // but rewrite each "in(i)" relative to the new "in(Bot)".
duke@0 3462 Node *progress = NULL;
duke@0 3463
duke@0 3464
duke@0 3465 Node* old_base = base_memory();
duke@0 3466 Node* empty_mem = empty_memory();
duke@0 3467 if (old_base == empty_mem)
duke@0 3468 return NULL; // Dead memory path.
duke@0 3469
duke@0 3470 MergeMemNode* old_mbase;
duke@0 3471 if (old_base != NULL && old_base->is_MergeMem())
duke@0 3472 old_mbase = old_base->as_MergeMem();
duke@0 3473 else
duke@0 3474 old_mbase = NULL;
duke@0 3475 Node* new_base = old_base;
duke@0 3476
duke@0 3477 // simplify stacked MergeMems in base memory
duke@0 3478 if (old_mbase) new_base = old_mbase->base_memory();
duke@0 3479
duke@0 3480 // the base memory might contribute new slices beyond my req()
duke@0 3481 if (old_mbase) grow_to_match(old_mbase);
duke@0 3482
duke@0 3483 // Look carefully at the base node if it is a phi.
duke@0 3484 PhiNode* phi_base;
duke@0 3485 if (new_base != NULL && new_base->is_Phi())
duke@0 3486 phi_base = new_base->as_Phi();
duke@0 3487 else
duke@0 3488 phi_base = NULL;
duke@0 3489
duke@0 3490 Node* phi_reg = NULL;
duke@0 3491 uint phi_len = (uint)-1;
duke@0 3492 if (phi_base != NULL && !phi_base->is_copy()) {
duke@0 3493 // do not examine phi if degraded to a copy
duke@0 3494 phi_reg = phi_base->region();
duke@0 3495 phi_len = phi_base->req();
duke@0 3496 // see if the phi is unfinished
duke@0 3497 for (uint i = 1; i < phi_len; i++) {
duke@0 3498 if (phi_base->in(i) == NULL) {
duke@0 3499 // incomplete phi; do not look at it yet!
duke@0 3500 phi_reg = NULL;
duke@0 3501 phi_len = (uint)-1;
duke@0 3502 break;
duke@0 3503 }
duke@0 3504 }
duke@0 3505 }
duke@0 3506
duke@0 3507 // Note: We do not call verify_sparse on entry, because inputs
duke@0 3508 // can normalize to the base_memory via subsume_node or similar
duke@0 3509 // mechanisms. This method repairs that damage.
duke@0 3510
duke@0 3511 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
duke@0 3512
duke@0 3513 // Look at each slice.
duke@0 3514 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
duke@0 3515 Node* old_in = in(i);
duke@0 3516 // calculate the old memory value
duke@0 3517 Node* old_mem = old_in;
duke@0 3518 if (old_mem == empty_mem) old_mem = old_base;
duke@0 3519 assert(old_mem == memory_at(i), "");
duke@0 3520
duke@0 3521 // maybe update (reslice) the old memory value
duke@0 3522
duke@0 3523 // simplify stacked MergeMems
duke@0 3524 Node* new_mem = old_mem;
duke@0 3525 MergeMemNode* old_mmem;
duke@0 3526 if (old_mem != NULL && old_mem->is_MergeMem())
duke@0 3527 old_mmem = old_mem->as_MergeMem();
duke@0 3528 else
duke@0 3529 old_mmem = NULL;
duke@0 3530 if (old_mmem == this) {
duke@0 3531 // This can happen if loops break up and safepoints disappear.
duke@0 3532 // A merge of BotPtr (default) with a RawPtr memory derived from a
duke@0 3533 // safepoint can be rewritten to a merge of the same BotPtr with
duke@0 3534 // the BotPtr phi coming into the loop. If that phi disappears
duke@0 3535 // also, we can end up with a self-loop of the mergemem.
duke@0 3536 // In general, if loops degenerate and memory effects disappear,
duke@0 3537 // a mergemem can be left looking at itself. This simply means
duke@0 3538 // that the mergemem's default should be used, since there is
duke@0 3539 // no longer any apparent effect on this slice.
duke@0 3540 // Note: If a memory slice is a MergeMem cycle, it is unreachable
duke@0 3541 // from start. Update the input to TOP.
duke@0 3542 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
duke@0 3543 }
duke@0 3544 else if (old_mmem != NULL) {
duke@0 3545 new_mem = old_mmem->memory_at(i);
duke@0 3546 }
duke@0 3547 // else preceeding memory was not a MergeMem
duke@0 3548
duke@0 3549 // replace equivalent phis (unfortunately, they do not GVN together)
duke@0 3550 if (new_mem != NULL && new_mem != new_base &&
duke@0 3551 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
duke@0 3552 if (new_mem->is_Phi()) {
duke@0 3553 PhiNode* phi_mem = new_mem->as_Phi();
duke@0 3554 for (uint i = 1; i < phi_len; i++) {
duke@0 3555 if (phi_base->in(i) != phi_mem->in(i)) {
duke@0 3556 phi_mem = NULL;
duke@0 3557 break;
duke@0 3558 }
duke@0 3559 }
duke@0 3560 if (phi_mem != NULL) {
duke@0 3561 // equivalent phi nodes; revert to the def
duke@0 3562 new_mem = new_base;
duke@0 3563 }
duke@0 3564 }
duke@0 3565 }
duke@0 3566
duke@0 3567 // maybe store down a new value
duke@0 3568 Node* new_in = new_mem;
duke@0 3569 if (new_in == new_base) new_in = empty_mem;
duke@0 3570
duke@0 3571 if (new_in != old_in) {
duke@0 3572 // Warning: Do not combine this "if" with the previous "if"
duke@0 3573 // A memory slice might have be be rewritten even if it is semantically
duke@0 3574 // unchanged, if the base_memory value has changed.
duke@0 3575 set_req(i, new_in);
duke@0 3576 progress = this; // Report progress
duke@0 3577 }
duke@0 3578 }
duke@0 3579
duke@0 3580 if (new_base != old_base) {
duke@0 3581 set_req(Compile::AliasIdxBot, new_base);
duke@0 3582 // Don't use set_base_memory(new_base), because we need to update du.
duke@0 3583 assert(base_memory() == new_base, "");
duke@0 3584 progress = this;
duke@0 3585 }
duke@0 3586
duke@0 3587 if( base_memory() == this ) {
duke@0 3588 // a self cycle indicates this memory path is dead
duke@0 3589 set_req(Compile::AliasIdxBot, empty_mem);
duke@0 3590 }
duke@0 3591
duke@0 3592 // Resolve external cycles by calling Ideal on a MergeMem base_memory
duke@0 3593 // Recursion must occur after the self cycle check above
duke@0 3594 if( base_memory()->is_MergeMem() ) {
duke@0 3595 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
duke@0 3596 Node *m = phase->transform(new_mbase); // Rollup any cycles
duke@0 3597 if( m != NULL && (m->is_top() ||
duke@0 3598 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
duke@0 3599 // propagate rollup of dead cycle to self
duke@0 3600 set_req(Compile::AliasIdxBot, empty_mem);
duke@0 3601 }
duke@0 3602 }
duke@0 3603
duke@0 3604 if( base_memory() == empty_mem ) {
duke@0 3605 progress = this;
duke@0 3606 // Cut inputs during Parse phase only.